MGY-005

TECHNIQUES IN REMOTE
Indira Gandhi National Open University
School of Sciences SENSING AND DIGITAL
IMAGE PROCESSING

Thematic Information Extraction Volume 2

and Introduction to R Programming
 MGY-005
TECHNIQUES IN REMOTE
Indira Gandhi National Open University
School of Sciences SENSING AND DIGITAL
IMAGE PROCESSING

Volume

2
THEMATIC INFORMATION EXTRACTION AND
INTRODUCTION TO R PROGRAMMING
BLOCK 3
IMAGE CLASSIFICATION AND CHANGE DETECTION TECHNIQUES 7

BLOCK 4
INTRODUCTION TO R PROGRAMMING 113
MGY-005: TECHNIQUES IN REMOTE SENSING AND DIGITAL
IMAGE PROCESSING
Programme Design Committee
Prof. Sujatha Verma Dr. I. M. Bahuguna Mr. Manish Parmar
Former Director Deputy Director (Rtd.) Scientist
School of Sciences Space Applications Centre Space Applications Centre
IGNOU, New Delhi Indian Space Research Organisation (ISRO) Ahmedabad, Gujarat
Dr. Shailesh Nayak (ISRO), Ahmedabad, Gujarat Dr. Akella V.S. Aswani
Director Prof. Shamita Kumar ESRI India Technologies Pvt.
National Institute of Advanced Studies Institute of Environment Education and Ltd.
Bangaluru, Karnataka Research Hyderabad, Telangana
Dr. P.S. Acharya Bharati Vidyapeeth University Dr. O.M. Murali
Head, NRDMS, NSDI Division Pune, Maharashtra GIS Consultant
Department of Science and Technology Ms. Asima Misra Chennai, Tamil Nadu
Ministry of Science & Technology Associate Director
New Delhi Prof. Manish Trivedi
ES & e-Governance Group School of Sciences
Dr. Debapriya Dutta Centre for Development of Advanced
Scientist ‘G’ & Associate Head IGNOU, New Delhi
Computing (C-DAC)
National Geospatial Programme Ministry of Electronics and Information Dr. Rajesh Kaliraman
Department of Science and Technology Technology (MeitY) School of Sciences
Ministry of Science & Technology Pune, Maharashtra IGNOU, New Delhi
New Delhi Dr. V. Venkat Ramanan
Dr. Sameer Saran
Dr. L.K. Sinha Head School of Inter-Disciplinary and
Former Director Geoinformatics Department Trans-Disciplinary Studies
Defence Terrain Research Lab. (DTRL), Indian Institute of Remote Sensing IGNOU, New Delhi
Delhi & Defence Geoinformatics Dehradun, U.K.
Research Establishment (DGRE) Faculty of Geology Discipline
Defence R&D Organisation (DRDO) Prof. Daljeet Singh School of Sciences, IGNOU
Chandigarh Department of Geography Prof. Meenal Mishra
Prof. P.K. Garg Swami Shraddhanand College Prof. Benidhar Deshmukh
Civil Engineering Department University of Delhi, New Delhi Prof. R. Baskar
IIT Roorkee, Roorkee, U.K. Dr. D. R. Rajak Dr. M. Prashanth
Prof. P.K. Verma Scientist Dr. Kakoli Gogoi
School of Studies in Earth Science Space Applications Centre (ISRO)
Ahmedabad, Gujarat Dr. Omkar Verma
Vikram University
Ujjain, M.P.
Course Design Committee
Prof. Shamita Kumar Dr. Dharmendra G. Shah Dr. Neha Garg
Institute of Environment Education and Department of Botany School of Sciences
Research MS University of Baroda IGNOU, New Delhi
Bharati Vidyapeeth University Vadodara, Gujarat Dr. Rajesh Kaliraman
Pune, Maharashtra Dr. Sadhana Jain School of Sciences
Prof. R. Jaishanker Regional Remote Sensing Centre IGNOU, New Delhi
CV Raman Laboratory of Ecological (RRSC), ISRO Dr. V. Venkat Ramanan
Informatics Nagpur, Maharashtra School of Inter-Disciplinary and
Digital University Kerala Dr. Neeti Trans-Disciplinary Studies
(formerly IIITM-K) Centre for Climate Change and IGNOU, New Delhi
Thiruvananthapuram, Kerala Sustainability, Azim Premji University
Ms. Asima Misra Faculty of Geology Discipline
Bengaluru, Karnataka
ES & e-Governance Group School of Sciences, IGNOU
Prof. P. V. K. Sasidhar
Centre for Development of Advanced School of Extension and Development Prof. Meenal Mishra
Computing (C-DAC), MeitY Studies, IGNOU, New Delhi Prof. Benidhar Deshmukh
Pune, Maharashtra
Prof. Nehal Farooqi Prof. R. Baskar
Dr. Amit Kumar
School of Extension and Development Dr. M. Prashanth
Environmental Technology Division
Studies, IGNOU, New Delhi Dr. Kakoli Gogoi
CSIR-Institute of Himalayan
Bioresource Technology Prof. Deepika Dr. Omkar Verma
Palampur, H.P. School of Sciences
IGNOU, New Delhi
Programme Coordinators: Prof. Benidhar Deshmukh and Prof. Meenal Mishra

2
 MGY-005: TECHNIQUES IN REMOTE SENSING AND
DIGITAL IMAGE PROCESSING
Preparation Team of VOLUME 2: Thematic Information Extraction and Introduction
to R Programming
Course Contributors
Dr. Sourish Chatterjee Dr. Sapana B. Chavan Ms. Leena K.
(Units 13 and 14) (Unit 15) (Unit 15)
Centre for Climate Change and Senior Software Engineer Tower-B, Noida One
Sustainability, Tech Mahindra Sector-62, Noida
Azim Premji University Hyderabad Uttar Pradesh
Bengaluru, Karnataka Telangana
Dr. Anupam Anand Dr. Neeti Prof. Benidhar Deshmukh
(Units 14) (Units 17, 18, 19, 20 and 21) (Units 15 and 16)
Centre for Climate Change and School of Sciences
Sustainability IGNOU, New Delhi
Azim Premji University
Bengaluru, Karnataka

Content Editor

Prof. Benidhar Deshmukh

School of Sciences
Indira Gandhi National Open University
New Delhi

Course Coordinators: Prof. Benidhar Deshmukh and Dr. Omkar Verma

Transformation and Formatting: Prof. Benidhar Deshmukh

Programme Coordinators: Prof. Benidhar Deshmukh and Prof. Meenal Mishra

Volume Production
Mr. Rajiv Girdhar Mr. Tilak Raj
A.R. (P), MPDD, IGNOU S.O. (P), MPDD, IGNOU

Acknowledgement: Ms. Savita Sharma for assistance in the preparation of CRC and some of the figures.
Cover Page Design: Prof. Benidhar Deshmukh
April, 2024
© Indira Gandhi National Open University, 2024
Disclaimer: Any materials adapted from web-based resources in this module are being used for educational purposes
only and not for commercial purposes.
All rights reserved. No part of this work may be reproduced in any form, by mimeograph or any other means, without
permission in writing from the Indira Gandhi National Open University.
Further information on the Indira Gandhi National Open University courses may be obtained from the University’s office
at Maidan Garhi, New Delhi-110 068 or visit University official website http://www.ignou.ac.in.
Printed and published on behalf of Indira Gandhi National Open University, New Delhi by the Registrar,
MPDD, IGNOU.

3
 Volume 2: Thematic Information Extraction and Introduction
to R Programming
The course MGY-005: Techniques in Remote Sensing and Digital Image Processing consists
of four blocks, which have been packaged in two volumes. The Volume 1 deals with
techniques in remote sensing and consists of two blocks namely, Remote Sensing Techniques
and Image Pre-classification Techniques. In the first block you have been introduced to various
types of remote sensing techniques such as aerial photography, multispectral, thermal,
hyperspectral, microwave, LiDAR and UAV based remote sensing, and in the second block
with image statistics, and various pre-classification techniques such as image corrections,
image enhancement and transformation including image fusion and principal component
analysis.
The Volume 2 covers some other aspects of digital image processing viz. image classification
and post-classification techniques and introduction to R programming. It comprises two blocks
namely, Image Classification and Change Detection Techniques, and Introduction to R
Programming.
The first block of this volume, Block 3: Image Classification and Change Detection
Techniques introduces you with techniques of unsupervised and supervised image
classification, change detection and also accuracy assessment, which is an important step
after extraction of thematic information from remote sensing images.
The second block of this volume, Block 2: Introduction to R Programming, introduces you
to fundamentals of computer programming and with using R package for data exploration,
image processing and data plotting.
Expected Learning Outcomes
After studying this volume, you should be able to:
 discuss various image classification techniques and commonly used classification
algorithms;
 describe the techniques employed for change detection analysis in remote sensing;
 explain accuracy assessment approaches suitable to various types of classification
outputs;
 write about R programming, its potential in geospatial data exploration, processing and
creation of plots.
After studying this volume, you will be equipped with the basic knowledge of image
classification and post-classification techniques and R programming for geospatial data
exploration and analysis.
We wish you all success in this endeavour!

4
MGY-005: Techniques in Remote Sensing and Digital
Image Processing
Block 1 : Remote Sensing Techniques
Unit 1 : Aerial Photography and Photogrammetry
Unit 2 : Multispectral and Thermal Remote Sensing
Unit 3 : Hyperspectral Remote Sensing
Unit 4 : Microwave Remote Sensing I
Unit 5 : Microwave Remote Sensing II
Unit 6 : LiDAR Remote Sensing and UAV based Remote Sensing
Block 2 : Image Pre-classification Techniques
Unit 7 : Image Statistics
Unit 8 : Radiometric Image Corrections
Unit 9 : Geometric Image Corrections
Unit 10 : Image Enhancement
Unit 11 : Image Filtering and Band Ratioing
Unit 12 : Image Fusion and Principal Component Analysis
Block 3 : Image Classification and Change Detection Techniques
Unit 13 : Unsupervised Classification
Unit 14 : Supervised Classification
Unit 15 : Change Detection Techniques
Unit 16 : Accuracy Assessment of Thematic Maps
Block 4 : Introduction to R Programming
Unit 17 : Fundamentals of Computer Programming
Unit 18 : Basics of R programming
Unit 19 : Using R for Basic Data Exploration
Unit 20 : Using R for Spatial Data Exploration
Unit 21 : Using R for Basic Image Processing

5
6
 MGY-005
TECHNIQUES IN REMOTE
Indira Gandhi National Open University
School of Sciences
SENSING AND DIGITAL
IMAGE PROCESSING

Block

3
IMAGE CLASSIFICATION AND CHANGE DETECTION
TECHNIQUES
UNIT 13
Unsupervised Classification 09
UNIT 14
Supervised Classification 39
UNIT 15
Change Detection Techniques 61
UNIT 16
Accuracy Assessment of Thematic Maps 75

Glossary 107

7
 BLOCK 3: IMAGE CLASSIFICATION AND CHANGE
DETECTION TECHNIQUES
You have learnt about various kinds of remote sensing techniques in Block-1. In Block-2, you
have studied that there are several pre-classification techniques that are required for
radiometric and geometric correction of digital images and also enhancement and
transformation techniques employed to enhance certain features of interests.
There are several techniques that are employed to extract thematic information from remote
sensing images. Image classification techniques such as the unsupervised and supervised are
the ones generally used for the purpose. We also apply change detection techniques to
understand changes taken place in an area. Once we have extracted thematic information
through some thematic information extraction techniques, the outputs are subjected to post-
classification techniques including accuracy assessment. It is required to assess accuracy of
the derived maps or information to know how close or far it is from the reality. It is also
important due to the fact such maps/information can be used with confidence for some
planning and decision making purpose with some confidence. This block covers following four
units related to these aspects:
Unit 13 “Unsupervised Classification” introduces you to image classification, its types and
specifically unsupervised mode of classification, its requirements and various methods used for
the purpose.
Unit 14 “Supervised Classification” deals specifically with supervised classification and
discusses various types of approaches used along with giving their comparison.
Unit 15 “Change Detection Techniques” focuses on the aspects related to change detection
such as requirement, pre and post classification approaches and various methods used for this.
Unit 16 “Accuracy Assessment of Thematic Maps” builds upon the accuracy assessment
concept that you have learnt in MGY-102 and introduces you to the sources of errors,
accuracy assessment approaches used for variety of outputs along with challenges and recent
developments in the field.
Expected Learning Outcomes
After studying this block, you should be able to:
 discuss about unsupervised classification technique and various algorithms used;
 describe supervised classification technique along with a brief on use of artificial intelligence
and machine learning in image classification;
 write different types of change detection techniques applicable for different types of data and
stages; and
 recognise various approaches of accuracy assessment suitable for different kinds of image
classification outputs and data types.

We wish you all the best and hope you will enjoy reading this course.

8
 UNIT 13

UNSUPERVISED CLASSIFICATION
Structure______________________________________________
13.1 Introduction ISODATA Clustering
Expected Learning Outcomes Hierarchical clustering
13.2 Image Classification Self Organising Maps
Approaches Fuzzy C-Means Clustering
Stages 13.6 Some Other Approaches
13.3 Types of Classification Gaussian Mixture Models (GMMs)
Unsupervised and Supervised Density-Based Spatial Clustering of
Classification Applications with Noise (DBSCAN)
Hard/Crisp and Soft/Fuzzy Classification 13.7 Challenges and Recent Developments
Pixel and Object based Classification 13.8 Summary
Parametric and Non-parametric 13.9 Terminal Questions
Classification 13.10 References
13.4 Steps in Unsupervised Classification 13.11 Further/Suggested Readings
13.5 Commonly Used Approaches 13.12 Answers
K-means Clustering

13.1 INTRODUCTION
In Block-2 of this course you have learnt about various pre-classification techniques. This unit and
the next one deal with a very key aspect of digital image processing and thematic information
extraction known as image classification. As you have read about image classification in the course
MGY-102, you know that it involves conversion of raster data into finite set of classes that represent
surface types i.e. a class in the theme of your interest in the imagery. This fundamental task
involves either training a model to categorise images into predefined classes (i.e. supervised
classification) or it is achieved by using algorithms that autonomously identify patterns and
similarities within an image dataset without explicit class labels (i.e. unsupervised classification).
The goal of this process is to enable machines to recognise and interpret visual content, mimicking

9 Contributor: Dr. Sourish Chatterjee

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
human visual perception. In this unit, you will learn more about image
classification with focus on unsupervised classification, starting with a brief
introduction to image classification, its typology based of various factors, and
followed by various unsupervised approaches commonly used for thematic
information extraction from remote sensing data. Supervised classification is
covered in the next unit.

Expected Learning Outcomes__________________________

After studying this unit, you should be able to:
 define image classification;
 discuss various types of image classification approaches;
 differentiate between pixel and object based classification, hard and soft
classification, parameteric and non-parametric classification;
 identify broad steps in image classification in general, and unsupervised
classification in particular;
 describe the commonly used approaches of unsupervised image classifi-
cation; and
 write about some other approaches useful for unsupervised classification.

13.2 IMAGE CLASSIFICATION

You have learnt that image classification is a process through which pixels in
the image is grouped into various classes/objects based on their spectral
signatures or reflectance properties. It may be noted that image classification is
used for several societal applications namely, land use/land cover analysis,
agriculture, urban planning, natural resource management, surveillance,
updating geographic maps and also for disaster mitigation.
Image classification is defined as a process of assigning land cover
classes/themes to pixels in an image (Lillesand and Keifer, 1994). Some of the
classes comprise built-up area, urban, forest, grassland, agriculture, water,
shadow, rocky areas, bare soil and cloud. Image classification usually
represents object of the analysis and generates a map-like image in the form of
final product/output. It is an important tool for studying digital images. There are
several image classification methods and terminologies used such as,
supervised, unsupervised, per-pixel, object-based, hard, soft, parametric, non-
parametric, spectral, contextual, etc.
13.2.1 Approaches
After following image correction, enhancement and transformation stages,
image classification begins with both or either of the following two general
approaches:
 Unsupervised Classification: It is the process of automatic identification of
natural groups or structures within a remotely sensed image, and
 Supervised Classification: It is the process of identification of classes
within a remotely sensed image with inputs from and as directed by the user
in the form of training data.
10 Contributor: Dr. Sourish Chatterjee
 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
Both the classification approaches use spectral signatures of pixels and ground
information of the area of the study for assigning each pixel a typical land cover
type or providing training samples for that. The unsupervised and supervised
image classifications differ from each other in the way the classification is
performed. For example, spectral and/or information alone is used in case of
unsupervised classification and ground truth is also required for training the
supervised classifier and testing the final result.
Some classifiers that combine both labelled and unlabelled data to train the
classifier thereby reducing the need for extensive labelled training samples can
be categorised into semi-supervised classifiers. Such classifiers are useful
when labelled data is scarce but large amounts of unlabelled data are available.
This type of classification is useful for the images of remote or inaccessible
areas where finding training sites is challenging. Advantage of such
classification is that it balances between supervised and unsupervised
approaches and reduces dependency on large training datasets. However, it
requires a good balance between labelled and unlabelled data and its
performance depends on the quality of initial labelled data.
Detailed typologies of various types of classification are discussed further in
subsection 13.2.3.
13.2.2 Stages
You have read that the image classification process broadly consists of
following three stages: training, signature evaluation and decision making.
Training is the process of generating spectral signature of each class. Training
can be carried out either by an image analyst with guidance from his
experience or knowledge in case of supervised training or by some statistical
clustering techniques requiring little input from image analysts in case of
unsupervised training. The training data selection is an important task and it
must be ensured that the location of each training sample and its thematic class
are correctly recorded, else the classification result can be erroneous.
Signature Evaluation is the checking of spectral signatures for their
representativeness of the class they attempt to describe and also to ensure as
small spectral overlap between signatures of different classes as possible.
Decision Making is the process of assigning all the image pixels into thematic
classes using evaluated signatures which is achieved using algorithms, known
as decision rules that set certain criteria. When signature of a candidate pixel
passes the criteria set for a particular class, it is assigned to that class.
You may note here that the training and signature evaluation steps are
essential steps in supervised classification whereas in the case of unsupervised
classification, training and signature evaluation is not involved and the focus is
on assigning a thematic class to the classes generated by the computer
through minimum input by the human analyst.

13.3 TYPES OF CLASSIFICATION

There are several types of image classification such as hard and soft
classifiers, per-pixel and object-based classification, parametric and non-

11
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
parametric classification, as well as statistical, ensemble, and machine learning
classification. Some broad typology of the classification is given in the Table
13.1.
Table 13.1: Typology of image classification (modified from Lu and Weng,
2007).
Basis of
Classifica- Type Characteristics Suitability Advantage Limitation
tion
Image is partitioned into
spectral classes based
Labelling by

Unsupervised
No training
on statistical information When field analyst
using clustering information is not data required,
required after
available, used automatic
algorithms; analyst classification;
for exploratory grouping of
merges and labels the analysis sensitive to
pixels parameters
spectral classes into
information classes
Use of
training Classifier classifies
samples spectral data into
Information classes When
SuperviSed

based on available field/ sufficient High Labour-intensive;

reference data through field or accuracy, classification
the reference analyst has dependent on
Signatures generated data is control training data
from available
the training samples
used

Simple, widely
Assigns each pixel or Binary decision Does not handle
Hard /
crisp

used, clear
object to a single, boundaries, crisp mixed pixels or
class
discrete class classification uncertainty well
Output type boundaries
or
assignment Handles
Better handles
of class Provides probabilistic or uncertainty;
mixed pixels, Computationally
Soft / fuzzy

membership fuzzy outputs, assigning provides

captures complex, difficult
to each pixel pixels to multiple membership
uncertainty, to interpret for
classes scores;
more some
with varying more
information applications
probabilities realistic
-rich output
outputs
Easy to
Classifies each pixel Ignores spatial
Simple, pixel- implement, fast
Per-Pixel

individually based on its information,

centric approach, processing,
spectral properties prone to salt-
uses spectral well-suited for
without considering and-pepper
values homogeneous
spatial context noise
Nature of areas
pixel The spectral value of
information each pixel is assumed
used to be a linear or non-
Useful for mixed- Computationally
linear combination of
Sub-pixel

pixel areas, high- Handles mixed intensive;

defined pure materials resolution, or pixels; provides requires complex
(or endmembers), fractional detailed class modelling;
providing proportional mapping of land proportions difficult to
membership of each covers interpret.
pixel to each
endmember

12 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
Basis of
Classifica- Type Characteristics Suitability Advantage Limitation
tion
Segments image into
meaningful objects and
Object-oriented classifies them based Computationally
Uses spatial
on spectral, shape, and High-resolution demanding;
context;
contextual data; data, complex sensitive to
reduces noise;
landscapes with segmentation
classification is accurate in
clear object quality; requires
conducted based on the heterogeneous
boundaries parameter
objects, instead of an landscapes
tuning.
individual pixel. No
vector data is used
Assumptions
may not hold
true; poor
performance with
non-normal data;
difficult to
Assumes that data
integrate
follows a known
statistical distribution Computationall ancillary
data, spatial and
Parametric

(e.g., Gaussian); Requires y efficient,

Usage of parameters (e.g. mean statistical interpretable, contextual
parameters vector and covariance assumptions well- attributes, and
such as matrix) are often about data established non-statistical
mean vector methods information into a
generated from training
and classification
covariance samples
procedure; often
matrix
produces ‘noisy’
results in
complex
landscape
Handles
Parametric

Flexible, no complex and Computationally

Does not assume any
Non-

assumptions non-linear intensive,

specific statistical
about data data, adaptable requires large
distribution of data
distribution. to various data training data
types
Pure spectral
information is used in
image classification. A
Fast and easy
‘noisy’ classification Images with high Ignores spatial
Usage of to implement;
Spectral

result is often produced spectral context; may

spectral and well-suited for
due to the high variation resolution and struggle with
spatial basic
in the spatial distribution clear class spectrally similar
information classification
of the same class. separations classes
tasks
Maximum likelihood,
minimum distance,
artificial neural network
Improved
Computationally
accuracy,
complex,
High spatial noise
parameter-
resolution reduction,
The spatially sensitive,
Contextual

imagery, enhanced
neighbouring pixel requires high-
heterogeneous object
quality data,
information is used in landscapes, recognition,
interpretation
image classification urban and better
challenges,
agricultural boundary
struggles with
monitoring delineation,
fragmented
handles mixed
landscapes
pixels

13
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
Basis of
Classifica- Type Characteristics Suitability Advantage Limitation
tion
Spectral and spatial Improves
information is used in classification

Spectral contextual
(/spectral-spatial)
Classification; High- accuracy;
parametric or non- resolution reduces
Computationally
parametric classifiers imagery misclassificatio
intensive;
are used to generate where spatial n of
requires careful
initial classification features provide spectrally
feature selection
images and then valuable similar classes;
contextual classifiers information preserves
are implemented in the spatial
classified images structure

So, you have learnt about different types of classification. Let us now discuss
some of these types of classifiers in some detail.
Let us first learn about the unsupervised and supervised classification.

13.3.1 Unsupervised and Supervised Classification

You have already learnt about the difference between unsupervised and
supervised classification in the course MGY-101 but let us recollect them here
in table 13.2
Table 13.2: Comparison of unsupervised and supervised classification.
Aspect Unsupervised Classification Supervised Classification
Automatic classification by algorithm Classification by algorithm based on user
by identifying natural groupings labelled training data, which involves
based on spectral similarities without selecting representative samples for each
Description prior knowledge or labelled data; class manually;
classes are determined after classes are predefined before
classification by interpreting the classification based on user knowledge
clusters formed by the algorithm and training data
High; to define the classes,
User
Minimal; select training samples, and
involvement
just to set the number of clusters validate the classification
output

Best for applications where

Suitable for exploratory analysis,
predefined classes are known,
Suitability areas with unknown land cover
such as agricultural mapping,
types, or when no training data is
urban planning, and
available
environmental monitoring

Requires no prior knowledge or

High accuracy;
Advantage training data;
analyst has control
automatic grouping of pixels
Labour-intensive;
classification dependent on
Labelling by analyst required after
Limitation training data, which involves selecting
classification; sensitive to parameters
representative samples for each class
manually
Produces clusters that need to be Produces classified maps with clear class
Output
interpreted and labelled by the user labels as defined by the training data

13.3.2 Hard/Crisp and Soft/Fuzzy Classification

Let us now learn the difference between hard and soft classifications. Table
13.3 compares both types of classification.

14 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
Table 13.3: Comparison of hard/crisp and soft/fuzzy classification.
Aspect Hard Classification Soft Classification
Assigns pixels to more than one
Description Assigns a single class to each pixel
class with probabilities
Class Gradual, overlapping, and fuzzy
Clear, distinct, crisp boundaries
boundaries boundaries
Clear-cut land covers, distinct Complex landscapes, mixed or
Suitability classes e.g. urban-rural mapping, transitional areas, e.g. agricultural
forest types fields, urban areas with mixed pixels
Simple, clear results, Handles mixed pixels, reflects real-
Advantage
computationally efficient world complexity
Complex interpretation;
Misclassification of mixed pixels;
computationally demanding; requires
Limitation overly simplified; cannot handle
additional analysis to convert to hard
uncertainty
classes, if needed

13.3.3 Pixel and Object based Classification

Let us now learn the difference between pixel based and object based
classifications. Table 13.4 compares both types of classification.
Table 13.4: Comparison of pixel based and object based classification.
Object-Based Classification
Aspect Pixel-Based Classification
(OBIA)
Classifies groups of pixels (objects)
Classifies individual pixels based
that are segmented based on
on their spectral information
spectral and spatial (shape, texture,
alone;
Description size) characteristics;
treats each pixel as an
segments the image into
independent unit without
meaningful objects or regions
considering spatial context
before classification
Classification Each pixel is classified Objects (groups of pixels) are
unit independently classified as a single unit
Suitable for coarse to moderate
Ideal for high-resolution imagery
Suitability resolution imagery where pixel
where spatial detail is significant
homogeneity is high
Utilises spatial context, improving
Simpler and faster to implement; classification accuracy;
directly uses spectral data, reduces salt-and-pepper effect
Advantage requiring less preprocessing; (noisy classifications);
effective for homogeneous more natural representation of real-
classes world features (e.g., roads,
buildings)
Prone to salt-and-pepper noise
Computationally intensive due to
due to pixel-level classification;
segmentation and feature
ignores spatial context, which
extraction;
can lead to misclassification of
Limitation requires careful tuning of
spectrally similar classes;
segmentation parameters;
not suitable for high-resolution
classification quality depends on the
imagery where spatial
segmentation accuracy
information is crucial
Examples of Maximum Likelihood, Minimum Multi-resolution segmentation,

15
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
algorithms Distance, K-means clustering, GEOBIA, Region Growing,
Spectral Angle Mapper., etc. Classification and Regression Trees
(CART)
Easier to interpret but less More realistic and easier to interpret
Interpretation
realistic for complex scenes in complex environments

While understanding about object based classification, it is also important to

learn here about the object detection and understand the difference between
object based image classification (OBIA) and object detection. These two are
two different types of approaches, each serving specific purposes and utilising
different methodologies. As you have read earlier in this unit, object based
classification segments an image into meaningful objects (groups of pixels) and
classifies them whereas object detection identifies and locates specific objects
(e.g., buildings, vehicles) in an image.
The purpose of OBIA is to classify entire regions or segments of an image
based on spectral, spatial, and contextual information whereas that of object
detection is to detect and precisely locate individual objects of interest within
an image. OBIA approach has high spatial context as it considers the
relationship between neighbouring pixels and object whereas object detection
has limited spatial context as it focuses mainly on the immediate surroundings
of the object being detected. Output in OBIA is classified regions or objects with
thematic labels (e.g., forest, water, urban) whereas that in object detection is
either the bounding boxes, masks, or point locations indicating the presence
and position of the objects of interest. OBIA uses both spectral information
(colour, reflectance) and spatial features (size, shape, texture) of segments
whereas object detection extracts features like edges, shapes, and textures
specific to target objects (e.g., cars, buildings).
While the OBIA works at a broader scale, classifying objects as regions rather
than identifying individual items, object detection approach aims to classify
segmented groups of pixels into thematic classes. It is focused on high level of
detail such as locating specific, discrete objects within an image. Both the
approaches have their unique strengths, making them suitable for different
kinds of remote sensing tasks.
13.3.4 Parametric and Non-parametric Classification
Let us now learn the difference between parametric and non-parametric
classifications.
Parametric and non-parametric classifications are two major approaches used
in remote sensing for classifying image data. The primary difference between
them lies in the assumptions they make about the data distribution and the
methods they use to classify pixels into different land cover classes. Both the
types of classification are compared in Table 13.5.

Table 13.5: Comparison of parametric and non-parametric classifications.

Aspect Parametric Classification Non-Parametric Classification
Requires less training data;
Requires more training data to
Training Data assumptions about data
effectively model class boundaries
requirements distribution reduce data
without distribution assumptions
dependency

16 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
Simpler, relies on statistical More complex; relies on flexible
Complexity parameters such as mean and algorithms that can model
covariance. complex relationships
Less flexible; performance drops Highly flexible; can adapt to
Flexibility if data deviates from assumed various data types and complex
distribution class boundaries
Best for simple, well-separated Suitable for complex, high-
classes with normally distributed dimensional data, non-Gaussian
data; effective when data fits distributions, and classes with
statistical assumptions; e.g. overlapping features; e.g. complex
Suitability
mapping vegetation types, water urban classifications,
bodies, or other classes with heterogeneous landscapes, and
distinct spectral signatures and cases where class boundaries are
normal distributions not clear-cut
Better at handling mixed pixels
Limited ability to handle mixed
Handling of and ambiguous class boundaries;
pixels due to rigid boundaries
mixed pixels can model complex decision
defined by statistical parameters
surfaces
Easier to interpret since it uses Often considered a "black-box,"
Interpretability statistical rules; decisions based making interpretation of decision-
on defined parameters making processes more difficult

Handles complex, non-linear

Simple implementation;
relationships; no assumptions
Advantage fast computation; effective when
about data distribution;
assumptions hold
high accuracy

Assumes data follows specific

Computationally demanding;
distributions; performance
requires a lot of training data;
Limitation degrades with non-Gaussian
less transparent decision-making
data; not suitable for complex or
process
overlapping classes

In the following sections, we will discuss the steps and commonly used
approaches of unsupervised classification in some detail.

13.4 STEPS IN UNSUPERVISED CLASSIFICATION

You have learnt earlier that as its name implies, this form of classification is
done without interpretive guidance from an analyst. Unsupervised image
classification is a fundamental approach in remote sensing data analysis that
involves automatically grouping pixels in an image into clusters or classes
without prior knowledge of the classes' characteristics. Unlike supervised
classification, where labelled training data is required, unsupervised
classification relies purely on the intrinsic spectral properties of the image data.
This method is particularly useful when labelled data is scarce, expensive, or
unavailable, making it a popular choice for exploratory data analysis in remote
sensing. An algorithm automatically organises similar pixel values into groups
that become the basis for different classes. This is entirely based on the
statistics of the image data distribution, and is often called clustering.

17
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………

Fig. 13.1: Generalised steps in unsupervised classification.

Unsupervised image classification groups pixels into clusters based on their

spectral similarities using algorithms that identify natural groupings within the
data. Each cluster represents a distinct land cover type, though the exact
nature of each class (e.g., forest, water, urban) is determined through post-
classification labelling. The process is automatically optimised according to
cluster statistics without the use of any knowledge-based control (i.e. ground
referenced data). The method is, therefore, objective and entirely data driven. It
is particularly suited to images of targets or areas where there is no ground
knowledge. Even for a well-mapped area, unsupervised classification may
reveal some spectral features which were not apparent beforehand. The basic
generalised steps of unsupervised classification are shown in Fig. 13.1.

The result of an unsupervised classification is an image of statistical clusters,

where the classified image still needs interpretation based on knowledge of
thematic contents of the clusters. There are hundreds of clustering algorithms
available for unsupervised classification and their use varies by the efficiency
and purpose. K-means and ISODATA are the two widely used algorithms which
are discussed here.

13.5 COMMONLY USED APPROACHES

In contrast to conventional supervised methods, unsupervised image
classification does not require labelled training data, making it an intriguing area
in computer vision. This approach does not require explicit class labelling;
instead, algorithms recognise patterns and similarities on their own inside an
image dataset. One prominent technique in unsupervised image classification is
clustering, where the algorithm groups similar pixels or regions together based
on inherent features (Fig. 13.2).

18 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…

Fig. 13.2: Steps in a typical unsupervised classification. In this diagram a FCC of

Delhi-NCR region created using Landsat 8 OLI images (Band 5: Band 3:
Band 2 = R:G:B) is undergone classification using a K-means clustering
algorithm (k=3).

Clustering algorithms, such as K-means or hierarchical clustering, enable

machines to uncover hidden structures within the data. By iteratively organising
pixels into clusters, the algorithm discerns patterns that may represent distinct
objects or textures. This approach is particularly valuable when dealing with
vast amounts of unlabeled imagery, as it allows for exploratory analysis and
pattern discovery without the manual annotation of training samples. It is
particularly suited to images of targets or areas where there is no ground
knowledge. Even for a well-mapped area, unsupervised classification may
reveal some spectral features which were not apparent beforehand.
Unsupervised image classification finds applications in scenarios where
obtaining labeled data is challenging or expensive. It aids in image
segmentation, anomaly detection, and uncovering latent patterns in diverse
datasets. As technology advances, unsupervised methods contribute to
unlocking insights from unannotated visual information, pushing the boundaries
of computer vision's capacity to understand and interpret complex visual data.
The result of an unsupervised classification is an image of statistical clusters,
where the classified image still needs interpretation based on knowledge of
thematic contents of the clusters. There are hundreds of clustering algorithms
available for unsupervised classification and their use varies by the efficiency
and purpose. K-means and ISODATA are the widely used algorithms which are
discussed here.

13.5.1 K-means Clustering

You have already learnt about the k-means clustering in the course MGY-102.
Let us recall it again here. You know that k-means is one of the most widely
used unsupervised classification algorithms. It partitions the image into `k`
clusters based on the nearest mean value of pixel intensities in a multi-
19
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
dimensional spectral space. The K-means clustering algorithm is a popular
unsupervised machine learning technique used for partitioning a dataset into k
distinct, non-overlapping subsets or clusters. The goal is to group data points
based on their spectral similarity, with k representing the predefined number of
clusters.
Here is a step-by-step description of how the K-means algorithm works:
1. Initialisation:
a. Choose the number of clusters (k) that you want to identify in the
dataset.
b. Randomly select k data points from the dataset as the initial cluster
centroids.
2. Assignment:
a. For each data point in the dataset, calculate the distance to each of the
k centroids. Common distance metrics include Euclidean distance or
Manhattan distance.
b. Assign the data point to the cluster whose centroid is closest, forming k
clusters.
3. Updating Centroids:
a. Recalculate the centroid of each cluster by taking the mean of all the
data points assigned to that cluster.
b. The new centroid becomes the representative point for that cluster.
4. Iteration: Repeat the assignment and centroid update steps iteratively until
convergence. Convergence occurs when the centroids no longer change
significantly or when a predefined number of iterations is reached.
5. Result: The final centroids and the assignments represent the k clusters in
the dataset.
Let us now learn the advantages and limitations of this algorithm.
Advantage
 The main advantage of this algorithm is its simplicity and speed which
allows it to run on large datasets.
Limitation
 It does not yield the same result with each run, since the resulting clusters
depend on the initial random assignments (Fig. 13.3).
 It is sensitive to outliers, so, for such datasets k-medians clustering is used.
 One of the main disadvantages to k-means is the fact that one must specify
the number of clusters as an input to algorithm.

20 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…

(a)

(b) (c)
Fig. 13.3: On the left is the cropped ROI from FCC (identical to Fig 13.1) of
Landsat 8 OLI image. The next two images are results of K-means
performed on two separate tries. Inside the yellow box one can
observe how the K-means provide different classification on different
runs, despite having the same parameters.

13.5.2 ISODATA Clustering

The ISODATA (Iterative Self-Organising Data Analysis Technique) algorithm is
an unsupervised clustering method designed to partition a dataset into a
predefined number of clusters based on the data's inherent characteristics.
Developed by Stuart Lloyd in 1982, ISODATA draws inspiration from the K-
means clustering algorithm but introduces adaptive mechanisms to handle
varying cluster shapes and sizes during the clustering process. The key
components of ISODATA algorithm are as follows:
1. Initialisation: ISODATA starts by selecting a set of initial cluster centroids.
This can be achieved through random selection or other techniques. The
algorithm also requires initial values for parameters such as the minimum
and maximum number of clusters, threshold values for cluster splitting and
merging, and maximum iteration count.
2. Cluster Assignment: In the cluster assignment step, each data point is
assigned to the cluster with the nearest centroid based on distance metrics
21
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
such as Euclidean distance. This process creates an initial clustering of the
dataset.
3. Update Cluster Centroids: After assigning data points to clusters, the
algorithm computes new centroids for each cluster by taking the mean of
the data points within that cluster.
4. Merge and Split Clusters: ISODATA introduces dynamic cluster merging
and splitting to adapt to changes in the dataset. If a cluster has too few data
points (below a specified threshold), it may be split into two clusters.
Conversely, if two clusters are deemed too similar (based on a similarity
criterion), they may be merged.
5. Update Cluster Statistics: The algorithm updates cluster statistics,
including the mean and variance, after merging or splitting clusters.
6. Iterative Process: Steps 2-5 are repeated iteratively until convergence or
until a predetermined maximum number of iterations is reached.

(a)

(b) (c)
Fig. 13.4: On the top is the cropped ROI from the FCC (identical to Fig 13.2) of
Landsat 8 OLI image (a). The two images on the lower panel are the
results of ISODATA clustering performed with initial k clusters 3 (b)
and 10 (c). It can be noted that despite specifying 10 clusters the final
result was produced in 5 clusters as seen in (c), this shows the
adaptive nature of ISODATA to cluster variability.

22 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
Let us now learn the advantages and limitations of this algorithm.
Advantage
1. Adaptive to Cluster Variability: ISODATA's ability to dynamically merge
and split clusters makes it more adaptive to varying cluster shapes, sizes,
and densities in the dataset (Fig. 13.3). This adaptability is particularly
useful when dealing with complex and heterogeneous data.
2. Automatic Determination of Cluster Number: Unlike K-means, ISODATA
does not require the user to specify the number of clusters beforehand. The
algorithm can automatically adjust the number of clusters based on the
characteristics of the data.
3. Handles Noisy Data: ISODATA is relatively robust to noise and outliers
due to its iterative nature and cluster merging/splitting mechanisms. Outliers
may be isolated into separate clusters during the process.
Limitation
 Sensitive to Initialisation: The performance of ISODATA can be sensitive
to the initial choice of cluster centroids, and different initializations may lead
to different results.
 Dependent on Parameters: The algorithm's effectiveness depends on
appropriately setting parameters such as the minimum and maximum
number of clusters, cluster splitting and merging thresholds, and
convergence criteria.
 Computational Complexity: ISODATA can be computationally intensive,
especially with large datasets, due to its iterative nature and dynamic cluster
operations.
Let us spend 5 minutes to check your progress.

SAQ I
a) What are the types of image classification?

b) Write the generalised steps of k-means clustering.

c) What are the limitations of ISODATA clustering?

13.5.3 Hierarchical Clustering

Hierarchical clustering is a commonly used technique for unsupervised
classification of remote sensing images. It groups pixels into clusters based on
their attributes (i.e. spectral similarity) without prior knowledge of the class
labels. Hierarchical clustering organises data into a tree-like structure called a
dendrogram, which represents the nested grouping of pixels or objects and
their similarities. This method is particularly useful for classifying satellite
images where predefined training data is unavailable.
Hierarchical clustering involves two main approaches i.e. agglomerative
(bottom-up) clustering and divisive (top-down) clustering. Agglomerative
(bottom-up) clustering is the most common in remote sensing. It starts with
23
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
each pixel or object as its own cluster and iteratively merges the closest cluster
based on a distance metric until all pixels are grouped into a single cluster.
Divisive (top-down) clustering begins with a single cluster containing all pixels
and recursively splits clusters into smaller clusters until each pixel is in its own
cluster. However, this approach is less commonly used due to its computational
complexity.
Following are the major steps in hierarchical clustering:
1. Calculation of Distance Matrix:
After preparing the data to be used for classification on a chosen theme, a
distance matrix is created using a distance measure such as Euclidean
distance, Manhattan distance, or cosine similarity. This matrix defines the
dissimilarities between each pair of pixels or objects. A distance or similarity
metric is computed between all pairs of pixels or data points based on their
RGB or reflectance values. For example, if there are three bands being used for
classification, the distance between pixel 1 (representing class1 having values
of 33, 87, 126 in the three bands, respectively) and pixel 2 (representing class2
having values of 121, 187, 56) would be calculated as:
Distance = sqrt[(100−120)2+(150−180)2+(50−60)2]
= sqrt[(−20)2+(−30)2+(−10)2]
= sqrt[400+900+100] = sqrt[1400]
= 37.42
Cluster Initialisation:
As you have learnt earlier in this subsection, agglomerative approach is the
most common hierarchical clustering method. It starts with each pixel as its own
cluster (singleton). In each iteration, the closest pair of clusters is merged to
form a new cluster. For example, in iteration1, it finds the closest pair of clusters
(initially individual pixels) and merges them. In our example, pixels 1 and 2
might be the closest. In iteration 2, it updates distances based on new cluster
centroids and finds the next closest pair of clusters to merge. This process
continues iteratively until all pixels are grouped into a single cluster or until a
stopping criterion is met.
In case of the divisive approach, it begins with all pixels in a single cluster and
iteratively splits the cluster into smaller clusters until each pixel is in its own
cluster or another stopping criterion is reached.
2. Merge Clusters:
Clusters are merged iteratively based on their proximity. The merging criteria
depends on the linkage method used, such as Single Linkage (Nearest
Neighbour), which merges clusters based on the smallest distance between
points; Complete Linkage (Farthest Neighbour) merges clusters based on the
maximum distance between points; Average Linkage merges clusters based on
the average distance between all pairs of points in the clusters and Ward’s
Linkage minimises the variance within clusters during merging, often yielding
more compact clusters.

24 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
As you merge clusters, you build a dendrogram showing how clusters combine.
Initially, the dendrogram will show individual pixels merging into small clusters,
which then merge into larger clusters.
3. Dendrogram Creation:
A dendrogram is constructed to visualise the hierarchy of the cluster merges.
The vertical axis shows the clusters, and the horizontal axis represents the
distance or dissimilarity at which clusters are merged. At the beginning, the
dendrogram has each pixel as its own branch. As merging progresses, the
branches for pixels 1 and 2 join, forming a branch for a class (assuming both
class1 and class2 are subclasses of a broad class) cluster. Similarly, pixels 3
and 4 form a branch for the same class cluster, and Pixels 5 and 6 form a
branch for the another cluster. In the end, you will have a single tree structure
with three main branches corresponding to the three different land cover types.
4. Cluster Selection/Extraction:
A cut-off point is chosen on the dendrogram to select/extact the desired number
of clusters, which are then used to classify the image. For example, if you
decide on the number of clusters or the similarity threshold as 3 at which to cut
the dendrogram. The cutting the dendrogram at a point where three main
branches are clearly separated will give you three clusters: one for class A, one
for class B, and one for class C.
The dendrogram graphically represents the hierarchical relationships among
clusters. It would have vertical lines representing clusters at different levels of
similarity, horizontal lines indicating the merging of clusters, with the position on
the distance scale showing the similarity level at which clusters are merged,
and the cut-off line which is a horizontal line that can be drawn across the
dendrogram to select clusters based on the desired level of similarity.
5. Post-Processing:
At the post-processing stage, you need to analyse the clusters to interpret the
land use land cover types they represent. In this case, the clusters align with
the known three land cover types. It is also required to validate the clusters with
reference data or ground truth, if available.
Let us now learn the advantages and limitations of this algorithm.
Advantage
 It does not require training data or initial class labels, thus making it
suitable for exploratory analysis.
 It produces dendrograms that offers a visual representation of data
hierarchy and clustering structure, aiding in understanding data
relationships.
 It is flexible and intuitive and easily handles different types of distance
metrics and linkage criteria, allowing customisation based on data
characteristics.
 It can reveal nested or hierarchical structures in data that other
clustering methods may not.

25
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
Limitation
 The method can be computationally expensive, especially for large
datasets, as it requires the calculation of all pairwise distances and
repeated merging operations.
 It is sensitive to outliers, which can significantly affect cluster formation.
 The approach is less practical for very large datasets due to memory
and processing time constraints.
 The final clustering result depends mainly on the chosen linkage
method, which may not always capture the most meaningful groupings.

Hierarchical clustering is useful in land use land cover mapping without

predefined classes; change detection to identify changes in land cover over
time by clustering multi-temporal images; vegetation mapping to differentiate
various vegetation types based on spectral reflectance patterns, etc.
13.5.4 Self-Organising Maps
Developed by Teuvo Kohonen in the 1980s Self-Organising Maps (SOM) also
known as Kohonen maps, are a type of artificial neural networks used largely in
unsupervised classification of remotely sensed data. SOMs are useful for
clustering remote sensing data based on inherent similarities without prior
knowledge of the classes, making them ideal for exploring complex, high-
dimensional remote sensing datasets. They are designed to reduce the
dimensionality of data while preserving the topological and metric relationships
of the input space, making them suitable for feature extraction and clustering in
remote sensing data based analysis.
It is useful to perform clustering and visualisation of high-dimensional data by
projecting them onto a low-dimensional (usually 2D) grid. It consists of neurons
(also called nodes or units), organised in a grid where each neuron represents
a prototype vector or weight, corresponding to a particular feature pattern.
Following are the major steps in self-organising maps based clustering:
1. Initialisation:
The SOM grid is initialised with random weight vectors, each representing a
neuron on the grid. Neurons are the nodes in the grid, each having a weight
vector of the same dimension as the input vector. Input data is the vectors
representing the features of each pixel or region in the remote sensing image,
such as spectral bands.
2. Training Process:
Training of SOMs involves adjusting the weight vectors of neurons to map the
input data onto the map, preserving the spatial relationships and clustering
similar data points together. It is carried out in the following steps:
 Initialisation: Weight vectors are initialised randomly or using some
heuristic, such as small values close to zero.

26 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
 Input Presentation/Selection: Each pixel or data point from the image is
presented to the SOM. For each training iteration, an input vector is
randomly selected from the dataset.
 Best Matching Unit (BMU): For each input vector, the neuron whose weight
vector is closest (based on a distance metric like Euclidean distance) to the
input vector is identified as the BMU.
 Weight Update: The BMU and its neighbouring neurons update their
weights to move closer to the input vector. The extent of the update depends
on a learning rate and a neighbourhood function, which decreases over time
and distance.
 Iteration and convergence: The process is repeated for all input data
points over multiple epochs, gradually fine-tuning the SOM.
3. Clustering and Output:
After training, similar input vectors are mapped to neighbouring neurons,
effectively clustering the data in a way that preserves the topological structure
of the input space.
The graphical output of a SOM typically consists of a 2D grid (often hexagonal
or rectangular) where each cell represents a cluster or class. The grid is
coloured or shaded based on the weights of the neurons, depicting patterns
and relationships among data points. New input data can be mapped onto the
trained SOM by finding the BMU for the new input and using the map to classify
or analyse the data based on its location on the map.
Let us now learn the advantages and limitations of this algorithm.
Advantage
 It does not require training data thus making it suitable for exploratory
analysis and scenarios where labels are unavailable or costly to obtain.
 It is useful in dimensionality reduction as it can reduce high dimensitonal
data into a lower-dimensional thus making it easier to visualise and interpret
multiple spectral bands.
 It is flexible and can be applied to a range of data types.
 It is effective at identifying clusters and patterns hence for classification and
feature extraction.
 It preserves the topological relationships of data in the input space thus
allowing for meaningful clustering and analysis.
 It is relatively robust to noise and outliers because neighborhood-based
updating mechanism smooths the influence of individual noisy data points.
Limitation
 The method can be computationally demanding and time consuming,
especially for large datasets, as it involves multiple iterations over the data,
which can be slow.

27
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 It is sensitive to the choice of parameters such as learning rate,
neighborhood function, and map size, which can lead to suboptimal
clustering and representation.
 The network structure (e.g., grid size and shape) is fixed before training. If
the chosen structure does not match the complexity of the data, the
resulting map may not effectively capture the data’s relationships.
 Interpreting the resultant map and understanding the specific meaning of
clusters can be complex, especially for high-dimensional or abstract data.
Visualisation of the map may not always clearly convey the clustering
results.
 Outputs of SOMs can vary due to converge to local minima, depending on
initialisation and training conditions.
13.5.5 Fuzzy C-Means Clustering
It is a type of clustering algorithm that is widely used for soft classification. In
hard classification, each data point belongs to exactly one cluster, whereas
FCM allows each data point to belong to multiple clusters with varying degrees
of membership. This approach is particularly useful where land cover types
often have mixed characteristics, making it difficult to classify them into discrete
categories. FCM assigns a membership value to each data point for all clusters.
These membership values range between 0 and 1, with the sum of membership
values for each data point equal to 1. In a typical graphical output of FCM
clustering in remote sensing, each pixel in an image is represented by a colour
or shade corresponding to its membership values across different clusters. In
FCM clustering, the cluster centers are represented as distinct points in the
feature space, typically marked by different colours. Each pixel in the image is
coloured based on its highest membership value, with the intensity indicating
the degree of membership. Further, unlike the hard clustering approaches you
have studied till now, in FCM clustering, boundaries between clusters are not
sharp, rather, there are smooth transitions indicating the fuzziness.
The FCM is particularly suitable for remote sensing applications where land
cover types are not clearly separable. It is used in case of land use land cover
classification, where boundaries between classes such as urban, vegetation,
water, and barren land, may be unclear. Also, it is used in case of vegetation
and soil mapping when different types of vegetation and soil types have
overlapping spectral signatures as there may be gradual transition from one
type to another. Further, it is also useful in change detection studies where
images from different time periods show gradual changes in the classes of
interest.
Following are the major steps in FCM clustering:
1. Initialisation:
The first step is to select the number of clusters, the fuzziness parameter
(typically >1), and the maximum number of iterations or convergence criterion.
Then the membership matrix is randomly initialised, where each element of the
matrix should satisfy certain criteria.
28 Contributor: Dr. Sourish Chatterjee
 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
2. Computation of Cluster Centers:
Then cluster centers are calculated using the current membership matrix. The
cluster center for a cluster is computed satisfying criteria for data point and total
number of data points.
3. Updating of Membership Matrix:
The next step is updating membership values or degrees based on the distance
of each pixel to the centroids.
4. Checking Convergence:
Next step is evaluation of convergence criterion to check if the membership
matrix or cluster centers have converged. The convergence can be determined
by either change in membership (by assessing if the change in membership
degrees between iterations is below a predefined threshold) or change in
cluster centres (by assessing if the change in cluster centers between iterations
is below a predefined threshold). If convergence criteria are met or the
maximum number of iterations is reached then the algorithm is terminated
otherwise the process continues until convergence is achieved.
5. Output Clusters:
Once convergence is achieved, the final cluster centers and membership matrix
represent the clustering results. Each data point will have membership values
for each cluster, indicating the degree to which it belongs to each cluster.
Let us now learn the advantages and limitations of this algorithm.
Advantage
 It does soft classification and captures fuzziness in thematic classes,
providing more realistic representation for mixed pixels.
 It is effective in areas where class boundaries are not clear or well-defined.
 It is flexible and allows for degree of class membership thereby
accommodating complex scenarios.
Limitation
 The method can be computationally demanding and time consuming, as
compared to hard classifiers.
 The algorithm may converge to local minima, depending on the initial cluster
centers, which can affect the final outcome.
 It is sensitive to the initial membership value and also choice of fuzziness
parameter.
 Although it is robust to within-class variations, it can still be sensitive to
noise, particularly in the cases, when the noise levels are high relative to the
actual signal.
You have become familiar with several commonly used algorithms for
unsupervised classification. Let us now see their comparison in Table 13.6.
Table 13.6: Comparison of various types of unsupervised classifiers.

29
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
Fuzzy C-
Feature / K-Means ISODATA Hierarchical Self-Organising
Means
Method Clustering Clustering Clustering Maps (SOM)
(FCM)
Algorithm Partition- Neural Network- Partition-
Partition-based Hierarchical
Type based based based
Arbitrary
Arbitrary
Cluster Spherical / Spherical / (varies with Spherical /
(depends on
Shape Convex Convex distance Convex
map grid)
metric)
Dynamic
Dynamic (can Fixed
Number of Fixed (depends on Fixed
merge/split (predefined map
Clusters (predefined) dendrogram (predefined)
clusters) size)
cut)
Initialisati Random or Random or Random or Random or
Not applicable
on heuristic heuristic heuristic heuristic
Can be slow
Can be slow and and
Can be slow
Good for large Good for large memory- memory-
Scalability for large
datasets datasets intensive for intensive for
datasets
large datasets large
datasets
Minimises
Convergence of Minimises
Converge Minimises within-cluster Based on
weights and weighted
nce within-cluster variance, allows dendrogram
membership within-cluster
Criteria variance for cut-off
matrix variance
merging/splitting
Cluster centers Dendrogram, Membership
Cluster
Cluster and assignment, cluster degrees for
assignments
Output centers and plus assignments each data
and visualization
assignment merging/splitting at various point in each
on 2D map
information levels cluster
Flexible (soft
Highly flexible
Less flexible More flexible Flexible (visual clustering
(varies with
Flexibility (fixed number (dynamic cluster representation with
dendrogram
of clusters) number) and clustering) membership
cut)
degrees)
Clustering
Well- Hierarchical
Complex data Visualising and with
separated relationships,
with varying clustering high- overlapping
Suitability clusters, fixed variable
cluster shapes dimensional data or
number of number of
and sizes data ambiguous
clusters clusters
boundaries

13.6 SOME OTHER APPROACHES

There are some other not so common approaches for unsupervised image
classification that you will be getting introduced to here.
13.6.1 Gaussian Mixture Models (GMMs)
Gaussian Mixture Models (GMMs) are probabilistic models used in statistics
and machine learning to represent complex data distributions. Comprising a
mixture of Gaussian (normal) distributions, GMMs can model diverse patterns
within a dataset. Each Gaussian component represents a potential cluster,
allowing GMMs to effectively capture intricate structures. Widely applied in
clustering, density estimation, and pattern recognition, GMMs excel in
scenarios where data exhibits multifaceted characteristics. Their flexibility and
ability to express uncertainty make GMMs valuable in various fields, including
geospatial data analysis (Sekhar et al., 2002; Vatsavai et al., 2011), image
processing, and speech recognition.
GMM-based clustering is particularly effective in scenarios where the
underlying data distribution is complex and may not adhere to a simple, linear
separation (Fig. 13.4). In the context of GMMs, the algorithm identifies clusters
30 Contributor: Dr. Sourish Chatterjee
 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
by modeling the data as a mixture of multiple Gaussian distributions (Bishop
2006).
Each Gaussian distribution represents a potential cluster, and the algorithm
assigns data points to these clusters based on the probability of belonging to
each distribution. This unsupervised approach allows for the discovery of
hidden patterns and groupings within the data without the need for labeled
examples.
Let us now learn the advantages and limitations of this algorithm.
Advantage
 Flexibility and Expressiveness: GMMs are versatile and can model
complex data distributions with multiple components. This flexibility makes
them well-suited for capturing intricate patterns and structures in the data,
especially when the underlying distribution is not easily characterised by a
single Gaussian.
 Probabilistic Output: GMMs provide probabilistic output, assigning each
data point a probability of belonging to different clusters. This allows for a
more nuanced representation of uncertainty and provides a richer
understanding of the data distribution.
 Soft Clustering: Unlike some traditional clustering algorithms that assign
each data point to a single cluster, GMMs perform soft clustering. This
means that data points can belong to multiple clusters simultaneously,
reflecting the uncertainty inherent in many real-world datasets.
 Effective Handling of Elliptical Clusters: GMMs can model clusters with
elliptical shapes, making them suitable for datasets where clusters have
varying orientations and sizes. This is an advantage over methods like k-
means, which assumes spherical clusters.
 Handling Mixed Distributions: GMMs are capable of capturing mixed
distributions within a dataset. This is particularly useful in scenarios where
subpopulations with distinct characteristics exist, and a single clustering
approach might not be sufficient.
 Robustness to Noise: GMMs are less sensitive to outliers compared to
some other clustering methods. The probabilistic nature of GMMs helps
mitigate the impact of noise by considering the overall distribution rather than
relying on individual data points.
13.6.2 Density-Based Spatial Clustering of Applications
with Noise (DBSCAN)
DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a
clustering algorithm that works by grouping data points based on their density in
the feature space. DBSCAN is particularly useful for discovering clusters of
arbitrary shapes and handling noise in datasets. Its ability to automatically
determine the number of clusters without prior knowledge makes it
advantageous for various applications, especially in geospatial data analysis
(Wang and Wang, 2007). Here's a concise explanation of how DBSCAN
operates:

31
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 Core Points: DBSCAN identifies core points as those with a minimum
number of neighboring points within a specified distance.
 Density-Reachability: Points within the defined distance of a core point are
considered part of the same cluster. This creates a region of connected
points that are dense enough to form a cluster.
 Border Points: Points within the specified distance of a core point but do not
meet the minimum density criteria become border points. They are part of
the cluster but not considered as influential as core points.
 Noise Points: Points that are not core points and do not have enough
neighboring points within the distance threshold are considered noise. These
points do not belong to any cluster.
 Cluster Formation: The algorithm iteratively expands clusters by examining
the density-reachability relationship. It continues this process until all points
are assigned to a cluster or labeled as noise.
 Parameter Tuning: DBSCAN requires two main parameters - the minimum
number of points and the distance threshold. The effectiveness of DBSCAN
can be sensitive to the appropriate selection of these parameters, and they
need to be chosen based on the characteristics of the data.
 Result: The final result is a set of clusters, each containing core and border
points. Noise points are treated as outliers.
Let us now learn the advantages and limitations of this algorithm.
Advantage
 Flexibility in Cluster Shape: DBSCAN is capable of identifying clusters with
irregular shapes, making it suitable for datasets where clusters may not
conform to traditional geometric shapes. Unlike algorithms like k-means,
which assume spherical clusters, DBSCAN excels at discovering clusters of
varying shapes and sizes, providing more flexibility in capturing complex
patterns within the data.
 Noise Handling: DBSCAN effectively handles noise and outliers in the data.
It classifies data points that do not belong to any cluster as noise, allowing
for a more robust clustering result. This feature is particularly valuable in
real-world datasets, which often contain irregularities or anomalies that can
distort the clustering outcomes of other algorithms.
 Automatic Determination of Cluster Number: DBSCAN does not require
the user to specify the number of clusters in advance, as opposed to
algorithms like k-means that rely on this input. The algorithm dynamically
identifies clusters based on the density of data points, making it well-suited
for situations where the optimal number of clusters is unknown or may vary
across different parts of the dataset. This adaptability simplifies the
clustering process and is especially advantageous in exploratory data
analysis.
Let us spend 5 minutes to check your progress.

32 Contributor: Dr. Sourish Chatterjee

 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…

SAQ II
a) What is Gaussian mixture model and it usefulness in thematic information
extraction?
b) What are the limitations of density-based spatial clustering?

13.7 CHALLENGES AND RECENT

DEVELOPMENTS
You have learnt that unsupervised classification is useful in certain cases when
you do not have field data and would like to know natural groupings present in a
data. You have also learnt that it is useful in generating initial exploratory
analysis particularly in unexplored or rapidly changing environments. It is also
useful in change detection analysis as unsupervised classification can highlight
changes in land use land cover classes over time. It may also be helpful in
geological surveys to identify areas of interest for mineral prospecting.
Despite these advantages, there are certain challenges in unsupervised
classification such as in interpreting in assigning informational class or labelling
the clusters in a meaningful way. Another challenge is the spectral similarity on
which basis the classification has been carried out by an algorithm and the
clusters generated, may not correspond well to real-world classes, if those
classes have similar spectral properties. Sensitivity to the parameter could be
another issue as algorithms require careful tuning of parameters that may
influence the outcome. Scalability is another issue as some algorithms may
struggle to perform due to them being computationally intensive.
There are certain developments which are driving advancements in the field.
The focus is on handling big data, improving robustness, and enhancing
interpretability. These trends are enhancing the accuracy and efficiency of
clustering and classification processes, making it possible to extract more
valuable insights from complex remote sensing datasets. Some of the recent
developments in the field include the following:
 Integration with Machine Learning: Hybrid approaches combine
unsupervised classification with supervised learning to refine results.
 Deep Learning Adaptations: Convolutional Neural Networks (CNNs) and
autoencoders are being adapted to perform unsupervised feature extraction
and clustering. Convolutional Neural Networks (CNNs) analyse both spectral
and spatial information in hyperspectral images, improving classification
accuracy by leveraging spatial context. Autoencoders are being used for
feature extraction and dimensionality reduction, as they learn compact
representations of high-dimensional data. These are increasingly used for
multispectral and hyperspectral images. Generative Adversarial Networks
(GANs) are being used for data augmentation and synthetic data generation,
which can improve the quality and quantity of training data for unsupervised
models. Further, deep clustering models combining deep learning with
clustering algorithms are being used to integrate feature learning and
clustering into a unified framework.

33
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 High-Dimensional Data Handling: Advanced algorithms are being used
alongside clustering to manage high-dimensional hyperspectral data.
 Multi-Scale and Multi-Resolution Approaches: Techniques that combine
information from multiple scales or resolutions to capture more detailed and
accurate features in remote sensing data are being used.
 Multi-Source Data Integration and Feature Fusion: Combining data from
different sources (e.g., optical, radar, LiDAR) to improve clustering outcomes
is another trend. Data fusion methods integrate diverse datasets to provide
richer and more accurate information for unsupervised classification. Also,
feature fusion i.e. merging features from different types of data (e.g.,
spectral, spatial, temporal) is being done to enhance the clustering process
and extract more meaningful patterns.
 Hybrid Models: Hybrid models such as combining traditional unsupervised
clustering methods with deep learning features are being used. For example,
using SOMs or FCM with features extracted by deep learning models to
enhance clustering performance.
 Handling Big Data: Development of algorithms that handle large volumes of
remote sensing data efficiently is another trend. Techniques such as parallel
processing and distributed computing are used to manage and analyse big
data in remote sensing applications. Cloud computing platforms is also being
leveraged for scalable storage and processing of remote sensing data,
thereby allowing for more extensive and complex analyses.
 Dimensionality Reduction: Advanced techniques are being used for
reducing the dimensionality of high-dimensional remote sensing data while
preserving its structure.
 Adaptive and Dynamic Clustering methods: Algorithms that can adjust
cluster parameters dynamically based on the data characteristics are being
used, which improve the adaptability of clustering algorithms to varying data
distributions and complexities. Further, techniques that evolve clusters over
time or as more data becomes available, useful for monitoring changes in
remote sensing data are also being devleloped.
 Uncertainty and Robustness: Handling uncertainty is a major issue in the
field so methods are being developed that can quantify and manage
uncertainty in clustering results. Also, the algorithms are being designed to
be less sensitive to outliers and noise, thereby improving the reliability of
clustering results in challenging environments.

13.8 SUMMARY
Let us now summarise what we studied in this unit:
 Image classification is the process of partitioning image in certain groups of
information classes based on their spectral characteristics. There are
broadly two approaches of image classification i.e. unsupervised and
supervised.
 Unsupervised image classification is the process of image classification in
which user input is minimum and the process is guided by the spectral
34 Contributor: Dr. Sourish Chatterjee
 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
similarity of the objects present in the image. This unit essentially describes
the fundamentals of unsupervised classification approaches
 The commonly used algorithms in unsupervised image classification are k-
means and ISODATA clustering.
 There are some other algorithms used in unsupervised image classification.
 However, each algorithm has its own advantages and limitations and choice
of the algorithm is guided by the nature of data being used, number and
nature of features present in the image and also computational resources.

13.9 TERMINAL QUESTIONS

1. What are the generic steps in unsupervised image classification?
2. Write comparison of unsupervised and supervised classification.
3. Compare various algorithms of unsupervised classification.

13.10 REFERENCES
 Banerjee, B., Bovolo, F., Bhattacharya, A., Bruzzone, L., Chaudhuri, S., &
Mohan, B. K. (2014). A new self-training-based unsupervised satellite image
classification technique using cluster ensemble strategy. IEEE Geoscience
and Remote Sensing Letters, 12(4), 741-745.
 Bezdek, J. C. (1981). Pattern Recognition with Fuzzy Objective Function
Algorithms. Springer.
 Bishop, C.M. (2006). Pattern recognition and machine learning by
Christopher M. Bishop. Springer Science+ Business Media, LLC.
 Comaniciu, D., & Meer, P. (2002). Mean shift: A robust approach toward
feature space analysis. *IEEE Transactions on Pattern Analysis and
Machine Intelligence.
 Ester, M., Kriegel, H. P., Sander, J., & Xu, X. (1996). A density-based
algorithm for discovering clusters in large spatial databases with noise.
Proceedings of the 2nd International Conference on Knowledge Discovery
and Data Mining.
 Foody, G. M. (1996). Fuzzy modelling of vegetation from remotely sensed
imagery. Ecological Modelling.
 Foody, G. M. (1999). Applications of the self-organizing feature map neural
network in community data analysis. Ecological Modelling.
 Jain, A. K. & Dubes, R. C. (1988). Algorithms for clustering data. Prentice
Hall.
 Jolliffe, I. T. (1986). Principal Component Analysis. Springer-Verlag.
 Kaufman, L., & Rousseeuw, P. J. (1990). Finding Groups in Data: An
Introduction to Cluster Analysis. Wiley-Interscience.
 Kohonen, T. (2001). Self-Organizing maps. Springer Series in Information
Sciences.

35
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 Lee, T. W., & Lewicki, M. S. (2002). Unsupervised image classification,
segmentation, and enhancement using ICA mixture models. IEEE
Transactions on Image Processing, 11(3), 270-279.
 Lillesand, T. M., Kiefer, R. W. & Chipman, J. W. (2015). Remote Sensing
and Image Interpretation. John Wiley & Sons.
 Liu, J., & Mason, P. J. (2009) Essential Image Processing and GIS for
Remote Sensing. John Wiley & Sons.
 Lloyd, S. (1982). Least squares quantization in PCM. IEEE transactions on
information theory, 28(2), 129-137.
 Lu, D., & Weng, Q. (2007) A survey of image classification methods and
techniques for improving classification performance. International journal of
Remote sensing, 28(5), 823-870.
 Mather, P. M. & Tso, B. (2016). Classification Methods for Remotely Sensed
Data. CRC Press.
 Mather, P. M., & Tso, B. (2009) Classification Methods for Remotely Sensed
Data. CRC Press.
 Ng, A. Y., Jordan, M. I., & Weiss, Y. (2002). On spectral clustering: Analysis
and an algorithm. Advances in Neural Information Processing Systems.
 Pal, N. R., & Bezdek, J. C. (1995). On cluster validity for the fuzzy c-means
model. IEEE Transactions on Fuzzy Systems.
 Salah, M. (2017). A survey of modern classification techniques in remote
sensing for improved image classification. Journal of Geomatics, 11(1), 21.
 Schmarje, L., Santarossa, M., Schröder, S. M., & Koch, R. (2021). A survey
on semi-, self-and unsupervised learning for image classification. IEEE
Access, 9, 82146-82168.
 Shekhar, S., Schrater, P. R., Vatsavai, R.R., Wu, W. and Chawla, S. (2002).
Spatial contextual classification and prediction models for mining geospatial
data. IEEE Transactions on Multimedia, 4(2), 174-188.
 Vatsavai, R.R., Symons, C.T., Chandola, V. and Jun, G. (2011). GX-Means:
A model-based divide and merge algorithm for geospatial image clustering.
Procedia Computer Science, 4, 186-195.
 Wang, Q., Li, Q., Liu, H., Wang, Y., & Zhu, J. (2014, October). An improved
ISODATA algorithm for hyperspectral image classification. In 2014 7th
International Congress on Image and Signal Processing (pp. 660-664).
IEEE.
 Wang, X. and Wang, J. (2010). Using clustering methods in geospatial
information systems. Geomatica, 64(3), 347-361.

13.11 FURTHER/SUGGESTED READINGS

 Abburu, S. & Golla, S. B. (2015). Satellite image classification methods and
techniques: A review. International journal of computer applications, 119(8).
 Ball, G. H. & Hall, D. J. (1967). A clustering technique for summarizing
multivariate data. Behavioral Science.
36 Contributor: Dr. Sourish Chatterjee
 Unit 13 Unsupervised Classification
…………………………….…………………………………….……………………………………………….…
 Bezdek, J. C. (1981). Pattern Recognition with Fuzzy Objective Function
Algorithms. Springer.
 Dhingra, S. & Kumar, D. (2019). A review of remotely sensed satellite image
classification. International Journal of Electrical and Computer Engineering,
9(3), 1720.
 Foody, G. M. (1996). Fuzzy modelling of vegetation from remotely sensed
imagery. Ecological Modelling.
 Foody, G. M. (1999). Applications of the self-organizing feature map neural
network in community data analysis. Ecological Modelling.
 Jain, A. K. & Dubes, R. C. (1988). Algorithms for clustering data. Prentice
Hall.
 Kohonen, T. (2001). Self-Organizing Maps. Springer Series in Information
Sciences.
 Lillesand, T. M., Kiefer, R. W. & Chipman, J. W. (2015). Remote Sensing
and Image Interpretation. John Wiley & Sons.
 Liu, J., & Mason, P. J. (2009) Essential Image Processing and GIS for
Remote Sensing. John Wiley & Sons.
 Lu, D., & Weng, Q. (2007) A survey of image classification methods and
techniques for improving classification performance. International journal of
Remote sensing, 28(5), 823-870.
 Mather, P. M. & Tso, B. (2016). Classification Methods for Remotely Sensed
Data. CRC Press.
 Pal, N. R. & Bezdek, J. C. (1995). On cluster validity for the fuzzy c-means
model. IEEE Transactions on Fuzzy Systems.
 Richards, J. A. & Jia, X. (2006). Remote Sensing Digital Image Analysis: An
Introduction. Springer.
 Salah, M. (2017). A survey of modern classification techniques in remote
sensing for improved image classification. Journal of Geomatics, 11(1), 21.
 Tou, J. T. & Gonzalez, R. C. (1974). Pattern Recognition Principles.
Addison-Wesley.

13.12 ANSWERS
SAQ I
a) Please refer to section 13.1.
b) Please refer to subsection 13.2.1.
c) Please refer to subsection 13.2.2.

SAQ II
a) Gaussian Mixture Models (GMMs) are probabilistic models used in statistics
and machine learning to represent complex data distributions.

37
Contributor: Dr. Sourish Chatterjee
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
b) Comprising a mixture of Gaussian (normal) distributions, GMMs can model
diverse patterns within a dataset. Each Gaussian component represents a
potential cluster, allowing GMMs to effectively capture intricate structures.
c) Please refer to subsection 13.3.2.

Terminal Questions
1. Refer to section 13.4.
2. You answer should major include the difference in terms of mode,
requirement of field information, stage of analysis, etc.
3. Refer to sections 13.4 and 13.5.

38 Contributor: Dr. Sourish Chatterjee

 UNIT 14

SUPERVISED CLASSIFICATION
Structure______________________________________________
14.1 Introduction Parallelepiped Classifier

Expected Learning Outcomes Maximum Likelihood Classifier

14.2 Overview of Supervised Image 14.5 Overview of Other Classification

Classification Methods
Classification Scheme 14.6 Role of AI and ML in Image
Classification
Stages
14.7 Selection of an Appropriate
14.3 Selection of Training Site and
Classification Method
Signature Evaluation
14.8 Recent Trends
Spectral Signature
14.9 Summary
Training Site Selection
14.10 Terminal Questions
Ways of Signature Evaluation
14.11 References
Selection of Optimum Number of Bands
14.12 Further/Suggested Readings
14.4 Approaches of Supervised
Classification 14.13 Answers

14.1 INTRODUCTION
You have learnt that image classification involves mapping of the digital image data into a finite set
of classes that represent surface types in the imagery. It may be used to identify vegetation types,
anthropogenic structures, mineral resources, etc. or transient changes in any of these features.
Additionally, a classified image that can be considered to a raster layer in a geographic information
system. In the previous unit, you have learnt about image classification, typology of image
classification and more about unsupervised image classification and its various approaches. In this
unit, we will learn about the other generic type of image classification i.e. supervised image
classification.

39 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
In this unit, we will discuss different image classification methods, signature
evaluation, and the guidelines for selecting a classification method.

Expected Learning Outcomes__________________

After studying this unit, you should be able to:
 define supervised image classification;
 describe different supervised image classification algorithms used in
remote sensing;
 discuss relative advantages and limitations of commonly used supervised
classification algorithms; and
 explain how to evaluate spectral signatures.

14.2 OVERVIEW OF SUPERVISED IMAGE

CLASSIFICATION
You are aware that remote sensing images can be recorded from various
platforms such as satellites, drones and airplanes. After performing image
correction, enhancement and transformation operations on them, the next step
is to recognise or classify the pixels in the image into various classes/objects
/themes present in the scene based on their spectral signatures or reflectance
and remittance properties. This is usually done by a process known as image
classification. It may be noted that image classification is used for several
societal applications namely, land use/land cover analysis, agriculture, urban
planning, natural resource management, surveillance, object detection,
updating geographic maps and disaster mitigation.
Image classification can be defined as a process of assigning land cover
classes/themes to pixels in an image (Lillesand and Keifer, 1994). Some of the
classes comprise built-up area, urban, forest, grassland, agriculture, water,
shadow, rocky areas, bare soil and cloud. Image classification usually
represents object of the analysis and generates a map-like image in the form of
final product/output. It is an important tool for studying digital images. There are
several image classification methods available namely, supervised,
unsupervised, per-pixel, object-based, hard, soft, parametric, non-parametric,
etc.
14.2.1 Classification Scheme
When we go for image classification, the first thing that is important to consider
is the classification scheme. It is important because it determines the level of
details that is going to be there in the classified output. You have read about
LULC classification scheme in Block-3 of MGY-101 and also in MGY-102. You
will read about the LULC classification scheme in some more detail in the Unit 5
of the course MGY-007. LULC classification scheme contains taxonomically
correct definition of class information that is organised according to a logical
criterion. If a class satisfies certain criteria, it will be classified to that class.
Defining the criteria helps the analyst, irrespective of the user, and therefore
helps in maintaining consistency in classification. A good classification scheme

40 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
should avoid overlapping of classes through its definition. For example, there
should be a crisp definition to classify an area as agriculture. A scheme
designed for classification needs to be exhaustive, such that all features in an
area are correctly classified into class, while also taking precaution of excluding
the classes that do not occur in a region. For example, an inland area you need
not have a class called ‘inter-tidal region’, which is available only in coastal
areas, signifying the land exposed during high-tide and low-tide, but are
extremely important from the point of view of biodiversity and conservation of
aquatic species.
So, you need to choose a classification scheme based on the theme you are
interested in and also the level of classification. It will also determine the choice
of data that is suitable for mapping the classes of your interest.
Let us now recall the stages of classification.
14.2.2 Stages
You have read in the previous Unit 13 that as the name implies, unsupervised
classification is carried out without interpretive guidance from an analyst. An
algorithm automatically organises pixels having similar spectral properties into
groups that become the basis for different classes. This is entirely based on the
statistics of the image data distribution, and is often called clustering. The
process is automatically optimised according to cluster statistics without the use
of any knowledge-based control (i.e. ground referenced data). The method is,
therefore, objective and entirely data driven. It is particularly suited to images of
targets or areas where there is no ground knowledge. Even for a well-mapped
area, unsupervised classification may reveal some spectral features which were
not apparent beforehand.
Supervised classification, as the name implies, requires human guidance. An
analyst selects groups of contiguous pixels from the input image known as
training areas that defines DN values in each channel for each class. A
classification algorithm computes certain properties (i.e. data attributes) of the
training pixels, e.g. mean DN for each channel. Then, DN values of each pixel
in the image are compared with the attributes of the training set. This is based
on the statistics of training areas representing different ground selected
subjectively by users on the basis of their own knowledge or experience.
Classification is controlled by users’ knowledge but, on the other hand, is
constrained and may even be biased by their subjective view. Classification
can, therefore, be misguided by inappropriate or inaccurate training area
information and/or incomplete user knowledge. A standard approach for
carrying out supervised image classification is given in Fig. 14.1.

14.3 SELECTION OF TRAINING SITES AND

SIGNATURE EVALUATION
As you have read earlier, evaluation of signatures is an important step in
classification, which is carried out before decision making stage. In this stage,
signatures of different classes obtained through training sites from image are
checked for their representativeness of class they attempt to describe and also

Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 41

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
to ensure their uniqueness from other classes. However, let us first review the
concept of spectral signature and then about the ways of signature evaluation.

Fig. 14.1: Typical stages in the process of supervised image classification.

14.3.1 Spectral Signature

The wavelength of any given material determines the amount of solar radiation
it reflects, absorbs, transmits, or emits. So, when the amount of solar radiation
reflected, absorbed, transmitted, or emitted (usually measured in intensity, as a
percent of maximum) by the material is plotted over a range of wavelengths, the
connected points produce a curve called the material’s spectral signature.
The percent reflectance values of similar objects at a selected wavelength will
be similar while it will vary for different objects or landscape features. These
values can be plotted in a graph and compared. Such plots are called spectral
response curves or spectral signatures. Spectral signatures of like features
have similar shapes, for example, concrete will have similar spectral signatures
while the spectral signatures of grass and concrete will vary. Differences among
spectral signatures are used to classify remotely sensed images into classes of
landscape features.
Greater details of recorded spectral information allow for greater information to
be extracted from spectral signatures. This important property of matter makes
it possible to identify different substances or classes and also to separate them
by their individual spectral signatures (Fig. 14.2).
For example, at some wavelengths, soil reflects more energy (absorbs less)
than green vegetation but at other wavelengths it absorbs more (like clayey
soil) than does the vegetation. These differences in reflectance from various
kinds of surface materials make it possible to differentiate them from one
another.
42 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee
 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…

Fig. 14.2: Spectral signature of various features.

Spectral response curves of some of the materials are discussed here.

Vegetation contains water, cellulose (tissues and fibres), constituent of wood,
lignin (non-carbohydrate nitrogen), chlorophyll (green pigments) and
anthocyanin (water-soluble pigments). Depending on how ‘active’ (i.e. kinds of
chlorophyll) a green vegetation is, the combination of transmittance,
absorbance and reflectance vary in different bands of the spectrum. Here is a
general example of a reflectance plot for a vegetation type, with the dominating
factor influencing each interval of the curve so indicated; note downturns of the
curve that result from selective absorption (Fig. 14.2).
Chlorophyll strongly absorbs radiation in the red and blue wavelengths but
reflects green wavelengths. Leaves appear “greenest” to us in the summer,
when chlorophyll content is at its maximum. In autumn, there is less chlorophyll
in the leaves, so there is less absorption and proportionately more reflection of
the red wavelengths, making leaves appear red or yellow (yellow is a
combination of red and green wavelengths). The internal structure of healthy
leaves acts as excellent diffuse reflectors of NIR wavelengths. If our eyes were
sensitive to NIR, trees would appear extremely bright to us at these
wavelengths.
Overall, factors such as leaf damage, sun and shade, leaf water content; leaf
air spaces and salinity and nutrient levels can affect spectral response of the
leaf.
The spectral response of vegetation canopies is a little different from that of
leaves. Transmittance of leaves, amount and arrangement of leaves, structural
characteristics such as stalks, trunks, limbs; background (soil, leaf litter, etc.);
solar zenith angle; viewing angle and azimuth angle influence the spectral
response.
Soil tends to have reflection properties that increase approximately
monotonically with wavelength. They tend to have high reflectance in all bands.
This is dependent on factors such as colour, constituents and especially the
moisture content. As described above, water is a relatively strong absorber of
all wavelengths, particularly those longer than the red part of visible spectrum
Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 43
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
(Fig. 14.3). Therefore, as soil moisture content increases, the overall
reflectance of that soil tends to decrease. Soils rich in iron oxide reflect
proportionally more of the red than other visible wavelengths and therefore
appear red (rust colour) to the human eye. A sandy soil on the other hand tends
to appear bright white in imagery because visible wavelengths are more or less
equally reflected; when slightly less blue wavelengths are reflected this results
in a yellow colour. In a nutshell, spectral response curves of soil and rocks are
influenced by soil colour, mineral content, inter-molecular vibration of the
molecules, organic matter (influences soil colour and moisture), particle size,
reflectance and thermal diffusivity and moisture (Fig. 14.3).

Fig. 14.3: Generalised spectral signatures for some of the features.

Water absorbs much longer wavelength at visible and NIR radiation than
shorter visible wavelengths. Thus, water typically looks blue or blue-green due
to stronger reflectance at these shorter wavelengths and darker if viewed at red
or NIR wavelengths. If there is suspended sediment present in the upper layers
of a water body, then this will allow better reflectivity and a brighter appearance
of the water. The apparent colour of the water will show a slight shift to longer
wavelengths. Suspended sediment can be easily confused with shallow (but
clear) water, since these two phenomena appear very similar. Chlorophyll in
algae absorbs more of the blue wavelengths and reflects green, making water
appear greener in colour when algae are present. The topography of the water
surface such as rough, smooth or floating materials can also lead to
complications for water related interpretation due to potential problems of
specular reflection and other influences on colour and brightness.
Spectral signatures, however, are not always “pure” which means the sensor
might record some signatures that may be emitted by surrounding objects.
“Pure” spectral signature for individual materials or classes can be determined
best under laboratory conditions, where the sensor is placed very close to the
target (Fig. 14.4). There is no interference in a closed and controlled
environment such as a laboratory. Agencies such as ISRO, US Department of
Agriculture and several universities maintain large repositories of spectral

44 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
signatures. Moreover, many image analysis tools have such spectral libraries
bundled with the software.

Fig. 14.4: Spectral curves from a spectral library. (Source: Spectral library of US
Geological Survey)

You have leant that spectral signature is the unique pattern of reflectance or
radiance values of an object or surface across various wavelengths of the
electromagnetic spectrum. Understanding spectral signature is important
because they are used to differentiate between different classes or materials in
remote sensing images. Accurate signatures are crucial for distinguishing
between similar classes and improving classification accuracy.
Let us now study about training site selection.
14.3.2 Training Site Selection
After selection of suitable classification scheme and the data for classification,
the supervised image classification process consists of three important stages:
training, signature evaluation and decision making as you have read earlier
(Fig. 14.5).

Fig. 14.5: Typical stages in the process of supervised classification.

Training is the process of generating spectral signature of each class. For

example, a forest class may be defined by minimum and maximum pixel values
in different image bands, thus defining a spectral envelope for it.
This simple statistical description of the spectral envelope is known as
signature. Training can be carried out either by an image analyst with guidance
from his experience or knowledge (i.e. supervised training) or by some
statistical clustering techniques requiring little input from image analysts (i.e.
unsupervised training).

Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 45

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
Selecting training sites for supervised image classification involves strategically
choosing specific regions or samples from the image data to build a model that
can accurately classify various classes within the image. Prior to the selection
of training sites, you need to clearly know the classes or categories that you
want to classify. For example, in land cover classification, classes could be
forest, water, urban areas, etc. You also need to have understanding of the
characteristics of each class in terms of their visual image interpretation so that
you can identify them in the image. You also need to perform appropriate
preprocessing steps such as radiometric correction, atmospheric correction,
and normalisation to ensure the data is suitable for classification. Domain
knowledge is important to correctly identify the representative areas (for the
features of interest) based on visual inspection in the image and select initial
training sites. The sites are manually selected which involves drawing polygons
or selecting points that represent different classes.
The training data selection is strongly governed by the theory of sampling that
is a part of the field of statistics. Small classes or highly homogeneous classes
can have a small training dataset while a large or heterogenous class requires
a larger training dataset. The training dataset is expected to capture all the
intra-class variation given the pixel vectors within each class. If you take very
less number of training sites then it may be difficult to obtain a spectral
signature which truly represents that class. It must be ensured that the location
of each training sample and its thematic class are correctly recorded, else the
classification result can be erroneous.
We have read that signature Evaluation is the checking of spectral signatures
for their representativeness of the class they attempt to describe and also to
ensure as small spectral overlap between signatures of different classes as
possible.
Decision making is the process of assigning all the image pixels into thematic
classes using evaluated signatures. It is achieved using algorithms, which are
known as decision rules. The decision rules set certain criteria. When signature
of a candidate pixel passes the criteria set for a particular class, it is assigned to
that class. Pixels failing to satisfy criteria of any of the classes remain
unclassified. We shall discuss in detail the decision rules in the next two
sections.
You may note here that the training and signature evaluation steps are
essential steps in supervised classification whereas in the case of unsupervised
classification, training and signature evaluation is not involved and the focus is
on assigning a thematic class to the classes generated by the computer
through minimum input by the human analyst.
14.3.3 Ways of Signature Evaluation
One of the most common techniques for feature identification is spectral
evaluation. Most of the image analysis software provides an interface to plot
spectral signature. Fig. 14.6 shows an example of how a spectral image is
plotted using an image analysis tool. With knowledge about the spectral profile
for a given feature, we can go back and change band combinations to make
that feature show up more clearly on the image.

46 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…

Fig. 14.6: Spectral plots from a satellite image.

Spectral signatures are evaluated in the following three ways:

 classification is performed on the pixels within the training samples for each
class and is compared with classes as recorded in the field data on those
location. Ideally, all pixels in a training sample should classify correctly.
However, you can expect high percentages of correctly classified pixels if
the signatures taken are appropriate.
 measuring spectral distance, i.e. separability by computing divergence,
transformed divergence or the Jeffries-Matusita distance. You can find
mathematics behind computation of these in a book by Swain and Davis
(1978). However, it is important to ensure that there is high separability
between signatures from different types of training samples and low
separability among signatures from the training samples of a particular class
and
 statistical analysis: statistical measures that you have studied in Block-1
such as mean and standard deviation of each signature are used to plot
ellipse diagrams in two or more dimensions. The plotting allows the analyst
to identify similar signatures and hence the classes which are likely to suffer
most from misclassification. If the amount of overlap between a pair of
signatures is large then those classes are not separable using that image
data.
You should note that some of the training samples whose signatures have
negative effect on the classification outcome need to be either renamed or
merged or deleted.
14.3.4 Selection of Optimum Number of Bands
Since, remote sensing images contain multiple spectral bands, each capturing
different spectral properties of the features present in the study area, choosing
the optimum number of bands for image classification is crucial for improving
the accuracy and efficiency of the classification process. Selecting the right
Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 47
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
number of bands involves balancing between capturing sufficient information for
accurate classification and avoiding redundancy and computational inefficiency.
Selection of the optimum number of bands is determined by the following
factors:
Context and objectives: You may select the bands suitable for mapping the
features based on the specific objectives of the classification task such as
mapping land use land cover classes, vegetation, waterbodies, or
minerals/rocks. It would be guided by the characteristics of the data being used.
Information contained in the bands: Since, each band provides unique
information, and some may be more relevant to the classification task than
others depending upon their characteristics. Further, the correlating bands may
be avoided and highly correlated bands might provide redundant information.
Correlation of the spectral bands can be calculated using correlation analysis or
Principal Component Analysis (PCA). PCA is used to transform the data into
principal components and select the number of components that explain a
significant portion of the variance. This helps in reducing dimensionality while
retaining important information.
Computational efficiency: Higher number of bands mean more computational
resources and processing time. You can choose a number of bands that
balances classification accuracy with computational efficiency and also the
storage requirements and manageability of the dataset with different numbers
of bands.
Class separability: Band combinations that enhance class separability should
be prioritised. Effect of inclusion or exclusion of certain bands on separability
between classes may be assessed and accordingly the optimum number of
bands may be selected.
Choosing the optimum number of bands for image classification involves
evaluating the relevance and redundancy of each spectral band, applying
feature selection techniques, and balancing classification performance with
computational efficiency. By systematically analysing band correlations,
applying dimensionality reduction methods, and testing various band
combinations, you can identify the most effective set of bands for classification.
In the following sections, we will discuss about approaches of supervised
classifications.

14.4 APPROACHES OF SUPERVISED

CLASSIFICATION
As you have read earlier, supervised classification, as the name implies,
requires human guidance. An analyst selects groups of contiguous pixels from
the input image known as training areas that defines DN values in each channel
for each class. A classification algorithm computes certain properties (data
attributes) of the training pixels, for example, mean DN for each channel (Fig.
14.7). Then, DN values of each pixel in the image are compared with the
attributes of the training set.

48 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…

Fig. 14.7: Using supervised classification, pixels are classified into different
categories.

This is based on the statistics of training areas representing different ground

objects (Fig. 14.8) selected subjectively by users on the basis of their own
knowledge or experience. Classification is controlled by users’ knowledge but,
on the other hand, is constrained and may even be biased by their subjective
view. Classification can, therefore, be misguided by inappropriate or inaccurate
training area information and/or incomplete user knowledge.

Fig.14.8: Locations of the training data collected for supervised classification.

In the following subsections, we will discuss parallelepiped and maximum

likelihood algorithms of supervised image classification.
14.4.1 Parallelepiped Classifier
Parallelepiped classifier uses the class limits stored in each class signature to
determine if a given pixel falls within the class or not. The class limits specify
the dimensions (in standard deviation units) of each side of a parallelepiped

Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 49

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
surrounding mean of the class in feature space. If pixel falls inside the
parallelepiped, it is assigned to the class. However, if pixel falls within more
than one class, it is put in the overlap class. If pixel does not fall inside any
class, it is assigned to the null class i.e., left unclassified. Such pixels are taken
up for classification at a later stage by examining the classes assigned to their
immediate neighbours.

Fig. 14.9: Steps involved in supervised classification.

In parallelepiped classifiers, an n-dimensional imaginary box is constructed

around pixels within each category of interest (Fig. 14.10). The n-dimensional
space defined by the parallelepiped delimits different categories.

Fig. 14.10: Using the parallelepiped approach, pixel 1 is classified as forest and
pixel 2 is classified as urban.

Classification using this classifier is carried out in the following steps:

Step 1: Define the range of values in each training area and use these ranges
to construct an n-dimensional box (a parallelepiped) around each class.
Step 2: Use multi-dimensional ranges to create different surface categories.
Notice that there can be overlap between the categories when simple method is
used. One solution to this problem is to use a stepped decision region
boundary.
Advantages
 it is a simple and computationally inexpensive method and
 it does not assume a class statistical distribution and includes class
variance.
Limitations
 it is least accurate method
50 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee
 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
 it does not adapt well to elongated (high-covariance) clusters
 it often produces overlapping classes, requiring a second classification step
 it also becomes more cumbersome with increasing number of channels,
and
 pixels falling outside the defined parallelepiped remain unclassified.
14.4.2 Maximum Likelihood Classifier
Maximum likelihood (MXL) classifier is one of the most widely used classifiers
in the remote sensing. In this method, a pixel is assigned to the class for which
it has maximum likelihood of membership. This classification algorithm uses
training data to estimate mean vectors and band-wise variances of the classes,
and pair-wise band to band covariances which are then used to estimate
probabilities of pixels to belong to different classes. Maximum likelihood
classification considers not only mean or average values in assigning
classification but also the variability of brightness values in each class around
the mean. It is the most powerful of the classification algorithms as long as
accurate training data is provided and certain assumptions regarding the
distributions of classes are valid.
An advantage of this algorithm is that it provides an estimate of overlap areas
based on statistics. This method is different from parallelepiped in that it uses
only maximum and minimum pixel values. The distribution of data in each
training set is described by a mean vector and a covariance matrix. Pixels are
assigned a posteriori probability of belonging to a given class and placed in the
most ‘‘likely’’ class. This is the only algorithm in this list that takes into account
the shape of the training set distribution.
For mathematical tractability, Maximum likelihood classifiers assume the
conditional probability density of the pixel data vectors given the class (from the
training data set) to be Gaussian (normal) distribution. With the help of an
estimate of the proportion of each class within the study area, and the class
conditional density function of the data vectors in the class, the probability of a
class given the pixel data vector is computed and then as mentioned above, the
pixel data vector is assigned to the most likely class whose conditional
probability given the data vector is the highest.
The basis for the Gaussian assumption used above is that plotting the number
of pixels with any given DN value yields a histogram or distribution of DN values
within a particular band. Studies have shown that for most surfaces DN values
from visible or near-infrared (NIR) region of the electromagnetic (EM) spectrum
have a normal probability distribution. It means we can define curves based on
the mean and standard deviation of the sample that describe the normal
probability distribution by selecting category that has the highest statistical
probability for each pixel. These concentric circles, called equi-probability
contours, are derived from an assumed normal distribution around each training
site. Equi-probability contours define the level of statistical confidence in the
classification accuracy. Smaller the contour, higher is the statistical confidence.
Advantages
 it is one of the most accurate methods
Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 51
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 it overcomes unclassified pixel problem (subject to threshold values)
 it provides a consistent way to separate pixels in overlap zones between
classes and
 assignment of pixels to classes can be weighted by prior knowledge of the
likelihood that a class is correct.
Limitations
 cluster distributions are assumed to be Gaussian in each class and band.
Algorithm requires enough pixels in each training area to describe a normal
population and assumes class covariance matrices are similar
 classes not assigned to training sets tend to be misclassified – a particular
problem for mixtures
 it is reliant on the accuracy of training data. Changes in training set of any
one class can affect the decision boundaries with other classes
 it is relatively computationally expensive and
 it is also not practical with imaging spectrometer data.
Let us spend 5 minutes to check your progress.

SAQ I
a) What is supervised image classification?

b) What is training?

14.5 OVERVIEW OF OTHER CLASSIFICATION

METHODS
As mentioned in the above sections, apart from the supervised and
unsupervised classification there exists a third kind of classification method
known as hybrid classification. This method uses both the afore-mentioned
methods (supervised and unsupervised classifications) and is primarily applied
to improve accuracy and efficiency of classification results. The most common
example of hybrid classification is the use of unsupervised classification to
delineate classes prior to supervised classification in order to aid the analyst in
identifying numerous spectral classes. Guided clustering is another such
method which is useful in analysis involving complex variability in spectral
response for each land cover. In this method, analysis delineates numerous
supervised training sets for each land cover. These training sets do not have to
be homogeneous as opposed to the regular supervised classification. Data
from all these training sets are used for supervised classification. The analyst
uses his discretion while selecting final spectral classes, so all of the redundant
classes are merged or discarded as per need.
There are number of other classification methods such as contextual, decision
tree, neural network, etc. Contextual classifiers incorporate spatial or temporal
information along with the spectral signatures while deciding the information
classes. Decision tree classifiers are knowledge-based classifiers which classify

52 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
in steps, where the classifier is able to distinguish between two or more classes
at each step. In this method, various types of classifiers as deemed appropriate
can be combined. Neural network classifiers do not require any assumption
about the statistical distribution of data and use machine learning techniques to
classify image pixels.
Some classification methods deal with the classification of mixed pixels, which
refer to pixels having classes corresponding to more than one land cover.
These methods, referred as spectral mixture analysis, are based on physical
models providing information on discrete spectral signatures rather than
statistical methods.
Linear mixture methods consider spectral classes from one pixel to be linear
mixture of all the land cover classes. Fuzzy classification methods account for
the transition between various land cover classes, known as fuzzy regions in
between two classes. Fuzzy classification does not have definite boundaries
and one pixel may belong to more than one class.

14.6 ROLE OF AI AND ML IN IMAGE

CLASSIFICATION
You have often come across the terms like artificial intelligence, machine
learning and deep learning, which are at times used interchangeably. However,
there is some difference. Artificial intelligence is used to classify machines that
mimic human intelligence and human cognitive functions like problem-solving
and learning. Artificial intelligence is an umbrella term under which the machine
learning (that allows for optimisation) and deep learning come. While machine
learning is a subset of artificial intelligence, deep learning is a subfield of
machine learning. Neural networks form the backbone of deep learning
algorithms. Deep learning has much higher number of node layers, or depth, of
neural networks than a single neural network.
Let us now try to understand the role and importance of AI and ML in image
classification. It may be useful for the following tasks:
 Enhanced Accuracy and Performance: these are good in recognising
complex patterns and features in images and can automatically learn details
and subtle differences between classes, leading to higher classification
accuracy.
 Adaptability to Various Data Types: the machine learning models can be
trained to handle different types of image data, including multispectral,
hyperspectral, and radar images ensuring that classification tasks are
accurate across various types of images acquired from different types of
sensors.
 Handling High-Dimensional Data: these techniques can manage and
process vast amount of spectral bands efficiently without requiring manual
feature extraction.
 Automated Feature Extraction: while traditional methods often require
manual feature extraction, which can be time-consuming and error-prone,

Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 53

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
the AI and ML can be used to automate feature extraction, making the
classification process faster and more consistent.
 Scalability: AI and ML algorithms can be scaled to handle large volumes of
datasets quickly, facilitating real-time and large-scale image classification.
 Robustness to Variability: ML models are robust to variations in data such
as changes in illumination, sensor differences, and also varying spatial
resolutions and can generalise well across diverse datasets, reducing the
impact of such variations on classification accuracy.
 Probabilistic and Soft Classification: AI techniques like ensemble
methods and Bayesian networks provide probabilistic outputs, offering
insights into the uncertainty associated with classification results which is
particularly useful for decision-making for disaster management and
environmental monitoring.
 Object Detection: AI methods, particularly object detection algorithms like
YOLO (You Only Look Once) and Faster R-CNN, can identify and localise
objects within an image, which is an important capability for applications
such as urban planning, infrastructure monitoring, and wildlife tracking and
even for military applications.
 Change Detection: these can detect and analyse changes over time by
comparing classified images from different periods which can be useful for
monitoring deforestation, urban sprawl, and other dynamic processes.
 Continuous Improvement and Learning: these models can be updated
and retrained as new data becomes available, allowing them to adapt to
changes in data distributions and improve over time. This continuous
learning process ensures that the models remain effective as conditions
evolve.
AI and ML have revolutionised image classification by providing advanced
techniques for accurate, scalable, and efficient analysis of complex image data.
They automate feature extraction, handle high-dimensional data, and adapt to
varying conditions, making them indispensable in modern remote sensing and
computer vision applications. As technology continues to advance, the
capabilities and applications of AI and ML in image classification are expected
to expand further, offering even more powerful tools for interpreting and
understanding various kinds of images.

14.7 SELECTION OF AN APPROPRIATE

CLASSIFICATION METHOD
Classification process involves translating pixel values in a remote sensing
image into meaningful categories. In case of land cover classification, these
categories comprise different types of land cover defined by the classification
scheme that is being implemented. There are number of classification methods
that can be used to group image pixels into meaningful categories.
Unfortunately, there is not a single best approach to image classification. The
choice made depends a lot on the algorithms that are available with the image
processing software used and also familiarity and experience with different
54 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee
 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
methods. The choice of classification method is dependent upon many factors,
accuracy being one of the most important criteria. Some of the ways to evaluate
efficacy and accuracy of classification methods are discussed in next unit. It
has generally been found that in areas of complex terrain, the unsupervised
approach is preferable over the supervised one. In such conditions, if the
supervised approach is used, the user will have difficulty in selecting training
sites because of the variability of spectral response within each class.
Consequently, a prior ground data collection can be very time consuming. Also,
the supervised approach is subjective in the sense that the analyst tries to
classify information categories, which are often composed of several spectral
classes whereas spectrally distinguishable classes will be revealed by the
unsupervised approach, and hence ground data collection requirements may
be reduced. Additionally, unsupervised approach has the potential advantage of
revealing distinguishable classes unknown from previous work. However, when
definition of representative training areas is possible and statistical information
classes show a close correspondence, the results of supervised classification
will be superior to unsupervised classification.
Fig. 14.11 shows a Landsat scene of Washington DC, USA which has been
classified by supervised as well as unsupervised methods. We can observe that
there are many similarities between outputs of supervised and unsupervised
methods. However, in this example the outcome of the supervised classification
method has more generalised classes than that of the unsupervised method.

Fig. 14.11: Processing of images: a) Supervised classification; and b)

Unsupervised classification with five classes.

This is because of the fact that while we use spectral information to create
classes in the unsupervised classes, performance of the supervised classes is
largely dependent on the training samples. The training samples used for the
above classification (Fig. 14.11) is not sufficient to cover the entire spectrum of
a particular class and, therefore, we get a generalised image. The results of the
supervised classification can be further improved by collecting more training
samples which would further help to reduce the differences between classes
which may be due to mixtures within each pixel e.g. grass and forest.
It is not easy to answer the question which classification method is suitable for
a study because different classification results can be obtained from different
methods and each method has its own merits and demerits. However, for a
general guideline it can be said that factors such as spatial resolution of the
remote sensing data, source of data, classification scheme and availability of
Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 55
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
classification software must be taken into account while selecting a
classification method for use.

SAQ I
a) What is signature?

b) What is signature evaluation?

14.8 RECENT TRENDS

Recent trends and developments in remote sensing image classification have
been driven by advancements in computational capabilities, machine learning
algorithms, data availability, and also sensor technologies. These innovations
have significantly enhanced the accuracy, efficiency, and applicability of
classification methods across various fields. Some of these are listed here:
 Deep Learning and Convolutional Neural Networks (CNNs) for automatically
learning hierarchical features from data. Multi-Scale CNNs is an important
networks which can capture information at different scales..
 Hybrid models combining deep learning with other machine learning
techniques (e.g., Random Forest, SVM) are being developed to enhance
classification accuracy and overcome individual method limitations.
 Integration of Object-Based Image Analysis (OBIA) and Geospatial AI for
improved object recognition and classification.
 Multi-Source and Multi-Temporal Data Fusion to integrate various data
types to enhance classification by leveraging complementary information,
reducing uncertainty, and improving robustness.
 Use of cloud platforms such as Google Earth Engine (GEE) and AWS are
increasingly being used as they provide powerful tools for processing and
analysing large-scale remote sensing data.
 Automated and Self-Supervised Learning Approaches are being developed
for large-scale classification, anomaly detection, and urban feature
extraction.
 Real-Time and Near-Real-Time Classification is another emerging trend
which could be useful for disaster response, traffic monitoring, and dynamic
environmental assessments.

14.8 SUMMARY
Let us now summarise what you have studied in this unit:
 There are two approaches to image classification, i.e. unsupervised and
supervised. Unsupervised classification is useful for complex terrains and it
can significantly reduce the cost of ground data collection than the
supervised classification.

56 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
 Spectral signatures are unique for each material, which allows us to
distinguish objects from one another and is the basis of classification in
remote sensing.
 Methods of supervised image classification include Maximum likelihood,
Parallelepiped, etc.
 Maximum likelihood method is the most powerful of the classification
methods as long as accurate training data is provided and normal
distribution of classes is justified. Sometimes, for better image classification,
both supervised and unsupervised methods may be used which is known as
a hybrid approach.
 Selection of an appropriate image classification method is a challenging
task in image classification because there are so many classification
methods available. However, the choice for a particular classification
method depends on the availability of image processing software along with
familiarity and working experience with other methods.
 There are several important developments in the field.

14.9 TERMINAL QUESTIONS

1. What is classification scheme?
2. What is training data?
3. What are spectral signatures?
4. What are the recent trends in supervised classification?
5. What is the role of AI and ML in image classification?

14.10 REFERENCES
 Abburu, S. & Golla, S. B. (2015). Satellite image classification methods and
techniques: A review. International journal of computer applications, 119(8).
 Bishop, C.M. (2006). Pattern recognition and machine learning by
Christopher M. Bishop. Springer Science+ Business Media, LLC.
 Dhingra, S. & Kumar, D. (2019). A review of remotely sensed satellite image
classification. International Journal of Electrical and Computer Engineering,
9(3), 1720.
 ERDAS Field Guide, http://www.gis.usu.edu/unix/imagine/FieldGuide.pdf
(data retrieved in February, 2012).
 Jensen, J.R. (1986) Introductory Digital Image Processing: A Remote
Sensing Perspective, Prentice-Hall, New Jersey.
 Lillesand, T. M., Kiefer, R. W. & Chipman, J. W. (2015). Remote Sensing
and Image Interpretation. John Wiley & Sons.
 Lillesand, T.M. and Kiefer, R.W. (1994) Remote Sensing and Image
Interpretation, John Wiley and Sons, Hoboken.
 Liu, J., & Mason, P. J. (2009) Essential Image Processing and GIS for
Remote Sensing. John Wiley & Sons.

Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 57

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 Lu, D., & Weng, Q. (2007) A survey of image classification methods and
techniques for improving classification performance. International journal of
Remote sensing, 28(5), 823-870.
 Mather, P. M. & Tso, B. (2016). Classification Methods for Remotely
Sensed Data. CRC Press.
 Richards, J. A. & Jia, X. (2006). Remote Sensing Digital Image Analysis: An
Introduction. Springer.
 Salah, M. (2017). A survey of modern classification techniques in remote
sensing for improved image classification. Journal of Geomatics, 11(1), 21.
 Swain, P.H. and Davis, S.M. (1978) Remote Sensing: The Quantitative
Approach, McGraw-Hill International Book Co., New York.
 US Geological Survey Spectral Library, http://speclab.cr.usgs.gov/spectral-
lib (data retrieved in September, 2011).
 Wilkie, D.S. and Finn, J.T. (1996) Remote Sensing Imagery for Natural
Resources Monitoring – A Gide for First-Time Users, Columbia University
Press, New York.

14.11 FURTHER/SUGGESTED READINGS

 Campbell, J.B. (2006) Introduction to Remote Sensing, Taylor and Francis,
London.
 Lillesand, T. M., Kiefer, R. W. & Chipman, J. W. (2015). Remote Sensing
and Image Interpretation. John Wiley & Sons.
 Lillesand, T.M. and Kiefer, R. (2007) Remote Sensing Image Interpretation,
John Wiley, New York.
 Lillesand, T.M. and Kiefer, R.W. (1994) Remote Sensing and Image
Interpretation, John Wiley and Sons, Hoboken.
 Liu, J., & Mason, P. J. (2009) Essential Image Processing and GIS for
Remote Sensing. John Wiley & Sons.
 Lu, D., & Weng, Q. (2007) A survey of image classification methods and
techniques for improving classification performance. International journal of
Remote sensing, 28(5), 823-870.
 Mather, P. M. & Tso, B. (2016). Classification Methods for Remotely
Sensed Data. CRC Press.
 Richards, J. A. & Jia, X. (2006). Remote Sensing Digital Image Analysis: An
Introduction. Springer.
 Salah, M. (2017). A survey of modern classification techniques in remote
sensing for improved image classification. Journal of Geomatics, 11(1), 21.
 Swain, P.H. and Davis, S.M. (1978) Remote Sensing: The Quantitative
Approach, McGraw-Hill International Book Co., New York.

14.12 ANSWERS
SAQ I
58 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee
 Unit 14 Supervised Classification
…………………………….…………………………………….……………………………………………….…
a) Image classification can be defined as a process of assigning land cover
classes/themes to pixels in an image.
b) Training is the process of generating spectral signature of each class.

SAQ II
a) Please refer to subsection 14.3.1.
b) Please refer to section 14.3.

Terminal Questions
1. Please refer to section 14.2.1.
2. Please refer to section 14.3.2.
3. Please refer to section 14.3.1.
4. Please refer to section 14.8.
5. Please refer to section 14.6.

Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee 59

Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………

60 Contributors: Dr. Anupam Anand and Dr. Sourish Chatterjee

 UNIT 15

CHANGE DETECTION TECHNIQUES

Structure______________________________________________
15.1 Introduction 15.4 Applications of GIS Change Detection
Expected Learning Outcomes Land Use/Land Cover Application

15.2 Change Detection Coastal Conservation/Management

Why Is Change Detection Necessary? Forest Fire Mapping

What does Image Processing's Change Change Detection using UAV Technology
Detection Entail?
Machine Learning for Change Detection
The algorithm for Change Detection
15.5 Summary
Factors for Implementing Change Detection
15.6 Terminal Question
Operation of GIS-Based Change Detection
15.7 References
Change Agent
15.8 Further/Suggested Readings
15.3 Technique for Spectral Change
15.9 Answers
Detection

15.1 INTRODUCTION
This unit mainly deals with various aspects of change detection techniques and factors responsible
to understand the changes. In this unit, we will discuss about change detection necessity, their
change agents, requirement of image processing change detection, steps involved for spectral
change detection, and applications of change detection. You will also comprehend the significance
of change detection.

61
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………

Expected Learning Outcomes___________________________

After studying this unit, you should be able to:
 define change detection and write the need of detecting changes;
 discuss basic factors/ agents responsible for change detection;
 discuss how the change detection techniques operate;
 describe various spectral change detection techniques; and
 explain the various application of change detection techniques.

15.2 CHANGE DETECTION

Human activity, sudden natural events, or long-term climatological or
environmental trends can all lead to change on the earth's surface. One of the
core uses of images and remote sensing is the detection of that change. In
order to identify the kind, size, and location of a change, numerous raster
datasets that were typically collected for the same area at various dates are
compared. This could be done by detecting changes from one image to the
next, across a stack of images, or throughout the course of an image time
series.
Change detection is useful in many industries, but it also has non-commercial
uses. This can be used, in particular, to monitor the progression and effects of
flooding, forest fires, ongoing droughts, and other disaster and weather extreme
phenomena. Finding differences between two satellite photos taken before and
after an event is the foundation of remote sensing and GIS change detection
approaches. GIS methods for change detection compare the spatial
representation of two points in time and assess variations in the relevant
variables.
Due to a number of issues, including the growing population, the destruction of
natural resources, environmental pollution, land use planning, and others,
environmental protection faces serious challenges. Unplanned changes in land
usage are currently a significant issue. The majority of land use changes take
place without a precise, rational strategy or any consideration of how they will
affect the environment. Major flooding, air pollution in big cities, deforestation,
urban expansion, soil erosion, and desertification are all effects of poor
planning that didn't take environmental effects of development plans into
account. A common result of incorrect land use change is desertification.
Remotely detected images are used in numerous applications nowadays for a
variety of reasons. One of them is moving using high-resolution satellite
symbolism of city boundaries. Over the past 50 years, rapid urbanization with
changes in land use and spread has taken place in a number of urban centers
around the world. The most important difficulty in this particular situation is the
correlation between the extraction findings from these images and the pre-
existing vector data. According to all indications, the availability of high-quality
optical symbols appears to be exciting for geo-spatial database applications,
particularly for the collection and maintenance of geodata.

62 Contributor: Dr. Sapana B. Chavan

 Unit 15 Change Detection Techniques
…………………………….…………………………………….……………………………………………….…
Remote sensing and image processing technology advancements in particular
have made it possible to determine vast areas in detail and, in this regard,
produce expanded and trustworthy recent data swiftly. As a result, it is possible
to monitor the metropolitan regions' rapid development and establish methods
to guide it. In this regard, automatic object extraction methods are now required
for large-scale topographic mapping from the photos, identifying topographical
changes, and updating the map data. Automatic object-based image analysis
has been increasingly popular in recent years for remote sensing applications
such as mapping from high resolution photos or building and updating GIS
databases. Additionally, because the results of automatic object-based
extractions are GIS-based, they may be included into GIS, queried, and
subjected to a variety of strategic analyses.

Fig. 15.1: Change Detection in Satellite Dataset.

Definition: Identifying, describing, and quantifying variations between images

of the same scene taken at different times or under various circumstances are
the goal of change detection analysis, which includes a wide range of
techniques.
15.2.1 Why Is Change Detection Necessary?
Understanding the linkages and interactions between human and natural
events is crucial for better decision-making, and this requires timely and
accurate change detection of Earth's surface features. In recent decades,

63
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
remote sensing data have been employed widely as primary resource for
change detection.
15.2.2 What does image processing's change detection
entail?
Analyzing changes in satellite photos collected at two separate times is a
process called "change detection." Different methods can be used to calculate
the changes in these multi-temporal satellite photos.
15.2.3 The algorithm for change detection
Utilizing several temporal data sets, change detection entails quantifying
temporal impacts. Satellite data is frequently employed when someone is
interested in tracking changes over vast areas and at frequent intervals. The
algorithms utilized have a significant impact on the outcomes of the digital
analysis.
15.2.4 Factors for implementing change detection
The following prerequisites must be met before change detection analysis is put
into practice: (1) accurate radiometric and atmospheric calibration or
normalization between multi-temporal images; (2) precise multi-temporal image
registration; (3) identical phenological states between multi-temporal images;
and (4) detection of change
15.2.5 Operation of GIS-Based Change Detection
To derive insightful information, GIS software mixes spatial data with statistical
data. Equipment like drones, unmanned aerial vehicles, or satellites are used to
remotely perceive and collect the aforementioned spatial data. Numerous
sources can be used to gather statistical data, and satellites, UAVs, and drones
used for remote sensing can acquire geospatial data. Due to open data access,
satellite change detection is becoming more and more popular today and is
frequently the quickest and least expensive choice.
Now, if the data is gathered and examined over time, this provides us with a
change detection dashboard to identify a given feature's aspect through time.
This dashboard can then be used in a variety of ways, such as to comprehend
changes in ice sheets and forest cover, among other things. It can recognize a
certain feature's aspect in two distinct time frames. For instance, change
detection can be used to track retail businesses to find any differences between
the quantity of stores that opened five years ago and currently. Change
detection can also be used to monitor the size, shape, and movements of a
particular feature.
The applications of geospatial change detection are numerous. The technique
is used to monitor changes in crop status, land usage, urban growth, vehicle
mobility, glacier cracking, and other aspects of the environment. The discovery
of manmade climate change in the world's oceans aids in understanding the
extent of the issue and developing a successful response strategy.
GIS examines statistical and spatial data for change detection.

64 Contributor: Dr. Sapana B. Chavan

 Unit 15 Change Detection Techniques
…………………………….…………………………………….……………………………………………….…
15.2.6 Change Agent
Change detection also benefits from knowing the "change agent," or what
causes the change. The agent uses a variety of natural occurrences, like fire,
insect infestation, storm, flooding, and drought, to describe natural changes.
The term "agent" for anthropogenic changes refers to human-induced change,
which is typically tied to land use and includes activities like urban
development, farming, logging, and mining.
Urbanization, agriculture, insect attacks, fire, logging, mining, storms, and
petroleum exploration and production are only a few examples of the change
agents that have been the subject of most studies.

SAQ I
a) What is the goal of change detection?
b) Which analysis is popular in recent years for remote sensing applications
such as mapping from high resolution photos or building and updating GIS
databases?
c) What are the anthropogenic change agents?

15.3 TECHNIQUE FOR SPECTRAL CHANGE

DETECTION
This section will help you comprehend how spectral change detection is
implemented.
Images from two different dates are combined to create a new single-band or
multi-band image that contains the spectral alterations in spectral change
detection. To attribute the changes to certain land cover types, the resulting
image needs to be further processed. These techniques are dependent on
accurate picture registration and co-registration because they are based on
pixel-wise or scene-wise procedures. The most critical aspect of these
approaches' effectiveness is their ability to distinguish between change and no-
change pixels. Use of statistical threshold is a typical technique for
discrimination. To distinguish the area of change from the no-change area in
this method, threshold borders must be placed carefully.
Following are some techniques for detecting spectral changes:
1. Image Differencing: Using this technique, a new change image between
two dates is created by subtracting two co-registered image dates pixel by
pixel in each band.
2. Image Ratioing: Using the same two co-registered image dates, each band
is ratioed pixel by pixel. Ratio values that are close to 1 define the no-
change area. Areas that have changed between two dates will have values
that are higher or lower depending on the type of changes.
3. Image Regression: This technique takes the position that pixels at time t1
are linear features of pixels at time t2. It takes into account variations in
pixel values from two dates' means and variances.

65
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
4. Change Vector Analysis: A change vector of a pixel is the vector
difference between the multi-band digital vector of the pixel on two different
dates. The amplitude and direction of change from date one to date two are
described by a spectral change vector. The output consists of two images,
one of which shows the change vector's magnitude and the other its
direction. If the magnitude of changes exceeds a certain threshold, the
change is determined, and the type of change is indicated by the direction
of the change vector.
5. Vegetation Index Differencing: In studies of vegetation, a ratio
(sometimes referred to as a vegetation index) is employed to emphasize
spectral contrasts between vegetation's strong near-infrared reflectance and
the chlorophyll-absorption band (red portion) of the spectrum. Ratio
Vegetation Index, Normalized Vegetation Index, and Transformed
Vegetation Index are examples of common vegetation indices.
6. Multi-date Principal Component Analysis: In this method, two images
from the same location taken at different times are superimposed and
analyzed as a single image. The local alterations are revealed by the minor
component images, while the major component images display the
reflectance and radiometric variances (minor changes). This method is
more useful than the post classification method for tracking rapid changes in
land use and urban growth. On two photos from various dates, they
performed a principal component analysis, and on the compressed PCA
image, an interactive supervised categorization of land-use change was
performed. The new method and the traditional post-classification
methodology were compared using photos from two.
7. Post-classification Technique: In the post-classification method, each
image is assigned a unique categorization and label. The area of changes is
then extracted after a direct comparison of the categorization findings. This
method employs both supervised and unsupervised classifications. The
difficulty of adjusting for atmospheric and sensor changes between two
dates is reduced to a minimum by individual classification of two image
dates.
The key drawback of this approach is the results of the categorization are
accuracy dependent. Individual categorization errors result in the
propagation of uncertainties in the change map, which produces erroneous
information about changes in land use. The categorization methods, source
image error, and change determination as the sources of uncertainty in
change detection.
They outlined the following three primary error factors for classification-
based change detection using Maximum Likelihood (ML):
a. Subjective data collection is utilized to gather training data;
b. the ML classifier presumes that the probability distribution of each class
is normal; and
c. the method used to assess changes is not objective (based on number
of uncertainties). The usual way to express the uncertainty in a
categorized remote sensing image is as a confusion matrix (error).
66 Contributor: Dr. Sapana B. Chavan
 Unit 15 Change Detection Techniques
…………………………….…………………………………….……………………………………………….…
Confusion matrix can be used to derive a number of error indicators,
including error of commission.
8. On-screen Digitization: This technique is typically applied to scanned aerial
photos and high-resolution data from distant sensors. This technique uses
on-screen picture interpretation of high-resolution images to update
inaccurate government urban infrastructure records.

Fig. 15.2: Techniques for Spectral Change Detection

SAQ II
a) What is the Image ratioing?
b) Explain the Multi-date Principal Component Analysis.
c) What is the Change vector?
d) ............., …………….., and ………………. are examples of common
vegetation indices.

15.4 APPLICATIONS OF GIS CHANGE

DETECTION
In this section, you shall understand various application for which change
detection techniques are applied with the help of different GIS /RS software.
15.4.1 Land Use/Land Cover Application
The world's land resources are mostly inventoried via land use/land cover
mapping. Remote sensing provides a method for quickly gathering and
displaying data on land cover, whether it be on a regional or small scale. In
recent years, geographic information systems and remote sensing have grown
in significance as crucial instruments for the study of change detection at the
district and city level. In order to maintain a sustainable ecosystem, it is
required to monitor and identify changes in land use and land cover, which are
crucial for understanding how human activities interact with the environment.
Understanding the impact of man's activities on his natural resource base
through time and from space is now essential. Data from Earth sensing
67
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
satellites has grown increasingly important in recent years for mapping the
Earth's features and infrastructures, managing natural resources, and
researching environmental change.
An inventory level of data indicating the location, nature, and extent of change
is provided by the satellite land cover change information. This information must
be collected and incorporated into a database that allows for the identification
and evaluation of the processes, outcomes, and interactions of change with the
environment in order to effectively contribute to sustainable development
initiatives.

Fig. 15.3: Change Detection for Land Use/ Land Cover Application.

15.4.2 Coastal Conservation/Management

Coastal regions are susceptible to a variety of intricate natural processes, which
inevitably result in both long- and short-term changes. These alterations can be
categorized as shoreline retreat, sediment migration, water quality degradation,
and coastal expansion. The degradation of the coastal socioeconomic value is
directly impacted by the transformation of the coastal ecosystem, along with
human life, infrastructure, property, and coastal land resources. It is necessary
to continuously monitor environmental protection and sustainable development
in coastal areas by gathering bathymetric data. The collection of bathymetric
data is regarded as a key role in the monitoring system. Since the beginning of
time, coastal regions have been crucial to humans. The majority of major cities
are located along coastlines. Living by the coast accounts for about one-third of
all human habitation.
Urbanization and population growth are increasing quickly in coastal areas due
to the abundance of natural resources. Numerous construction initiatives in
coastal areas have resulted in a variety of coastal dangers, such as soil
erosion, seawater intrusion, coral bleaching, shoreline alteration, sedimentation,
etc. Coastal landscapes are constantly changing. They are continuously
changed by both natural and artificial processes. Understanding various coastal
68 Contributor: Dr. Sapana B. Chavan
 Unit 15 Change Detection Techniques
…………………………….…………………………………….……………………………………………….…
processes requires accurate coastline demarcation and surveillance. When
employing conventional great surveying techniques, shoreline delineation for
the entire coastal system is challenging, time-consuming, and occasionally
impossible. Remote sensing and geographic information system tools have
improved coastal geo-morphological investigations during the past few
decades.
An important undertaking, shoreline change extraction and change detection
analysis has applications in a variety of areas, including setback planning,
hazard zoning, erosion-accretion investigations, regional sediment budgets,
and conceptual or predictive modelling of coastal morphodynamics. Using
conventional ground survey methods, shoreline delineation for the entire
coastal system is challenging, time-consuming, and occasionally impossible.
The challenges of isolating coastline position and detecting shoreline changes
are being overcome by recent developments in remote sensing and GIS
techniques. The most effective and dependable tools for mapping shoreline
change today are those that have been developed via the use of remote
sensing and GIS technologies.
The line of land-water contact is referred to as the shoreline.
15.4.3 Forest Fire Mapping
For a wide range of uses, spatially and temporally explicit knowledge about
forest ecosystems is crucial, and Earth observation has developed into a crucial
tool for managing forests and keeping track of the dynamics of their cover. With
a heavy emphasis on forests, forest change detection seeks to discover
significant changes in the time series signal (e.g., illegal deforestation, wind
throw, fire). The monitoring of forest fires, risk mapping, and the identification of
potential zones have all benefited greatly from the application of EO data and
other RS methodologies. Furthermore, it has become increasingly important to
precisely analyze and track the health state of forests using high-resolution
satellite photography. Sensors on EO satellites have been used to track
changes in energy emission since the 1970s. The breadth and frequency of
forest fires are now better monitored thanks to a new generation of satellite
sensors and Unmanned Air Vehicle (UAV) technology, which has improved the
synergy of existing and upcoming RS technologies. Satellite data sets are
useful for near-real-time fire detection, monitoring, and the assessment of the
burned areas due to their large-area repeating coverage and inexpensive cost.

69
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………

Fig. 15.4: Forest Fire.

15.4.4 Change Detection using UAV technology

Even while satellite imagery's spatial resolution has substantially improved over
the past ten years, the data still may not be sufficient to detect moderate to
minor coastal changes. Unmanned aerial vehicles (UAVs or drones), on the
other hand, can deliver extremely high-resolution photos for a small region at a
reasonable cost. A mapping tool for environmental monitoring has been using
drones in recent years due to their agility and high-quality image capabilities.
UAVs are a feasible alternative for gathering data from distant sensing for a
variety of real-world uses. They offer digital images with the spatial and
temporal resolution needed to get beyond some of satellite imagery and aerial
photography's drawbacks.
Datasets generated by UAV remote sensing have such high spatial resolution
(1–5 cm), allowing for highly precise mapping of landscape properties in two
(2D) and three (3D) dimensions (3D). Instead, at the spatial resolutions typically
attainable by manned aircraft (10-100 cm) and satellite systems (>50 cm), such
minute changes cannot be distinguished. Additionally, the UAV systems' simple
deployment and low operating costs enable frequent missions, producing on-
demand information with extremely high spatial and temporal resolution.
Remote sensing is still one of the most effective methods for identifying and
keeping an eye on coastlines, and it plays a significant role.
15.4.5 Machine learning for Change Detection
In recent years, artificial intelligence has advanced greatly, sometimes
approaching human precision. There are many options now that weren't there
before thanks to the merging of AI and GIS. Agriculture, law enforcement, and
storm forecasting are just a few of the fields where artificial intelligence,
machine learning, and deep learning are assisting in understanding and
managing. In its simplest form, artificial intelligence is the capacity of a machine

70 Contributor: Dr. Sapana B. Chavan

 Unit 15 Change Detection Techniques
…………………………….…………………………………….……………………………………………….…
to carry out operations that ordinarily call for human intelligence. This procedure
is one that can be carried out using machine learning.
It employs algorithms to learn from the data and provide us with the necessary
response. Spatial analysis may now be done at a higher level by employing
Deep Learning tools on ArcGIS Pro, even if Machine Learning has long been a
crucial component of GIS software in work tools like Classification, Clustering,
Geographically Weighted Regression, etc. Commercial drone use has
increased significantly in recent years, ushering in a new era of
photogrammetry marked by great precision and a sharp decline in the cost of
gathering airborne data. With this unexpected influx of data, we can now do
novel and precise analytics on topics of interest by merging Machine Learning
techniques with GIS technology.

15.5 TERMINAL QUESTIONS

1. What variables are offered by satellite data on land cover change?
2. What is referred as line of land-water contact?
3. …….. are a feasible alternative for gathering data from distant sensing for a
variety of real-world uses.

15.6 SUMMARY
In this Unit, you have learned the following:
 Identifying, describing, and quantifying variations between images of the
same scene taken at different times or under various circumstances are the
goal of change detection analysis.
 Remote sensing and image processing technology advancements in
particular have made it possible to determine vast areas in detail and, in this
regard, produce expanded and trustworthy recent data swiftly
 Analyzing changes in satellite photos collected at two separate times is a
process called "change detection." Different methods can be used to
calculate the changes in these multi-temporal satellite photos.
 GIS examines statistical and spatial data for change detection.
 Urbanization, agriculture, insect attacks, fire, logging, mining, storms, and
petroleum exploration and production are only a few examples of the
change agents that have been the subject of most studies.
 Images from two different dates are combined to create a new single-band
or multi-band image that contains the spectral alterations in spectral change
detection.
 Image differencing, Image ratioing, Image regression, change vector
analysis, Vegetation index differencing, Multi date principal component
analysis, Post classification techniques and on-screen digitization are the
techniques of spectral change detection
 Remote sensing provides a method for quickly gathering and displaying
data on land cover, whether it be on a regional or small scale

71
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 Various application of change detection such as Land use/Land cover
analysis, Coastal area management, Forest fire, change detection using
UAV technology, Machine learning for change detection.

15.7 REFERENCES
 Asokan, A., & Anitha, J. J. E. S. I. (2019). Change detection techniques for
remote sensing applications: A survey. Earth Science Informatics, 12, 143-
160.
 Jianya, G., Haigang, S., Guorui, M., & Qiming, Z. (2008). A review of multi-
temporal remote sensing data change detection algorithms. The
International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences, 37(B7), 757-762.
 Bai, T., Wang, L., Yin, D., Sun, K., Chen, Y., Li, W., & Li, D. (2023). Deep
learning for change detection in remote sensing: a review. Geo-spatial
Information Science, 26(3), 262-288.
 Lu, D., Mausel, P., Brondizio, E., & Moran, E. (2004). Change detection
techniques. International journal of remote sensing, 25(12), 2365-2401.
 Shafique, A., Cao, G., Khan, Z., Asad, M., & Aslam, M. (2022). Deep
learning-based change detection in remote sensing images: A review.
Remote Sensing, 14(4), 871.
 Coppin, P. R., & Bauer, M. E. (1996). Digital change detection in forest
ecosystems with remote sensing imagery. Remote sensing reviews, 13(3-
4), 207-234.
 Singh, A. (1989). Review article digital change detection techniques using
remotely-sensed data. International journal of remote sensing, 10(6), 989-
1003.
 Coppin, P., Jonckheere, I., Nackaerts, K., Muys, B., & Lambin, E. (2004).
Review ArticleDigital change detection methods in ecosystem monitoring: a
review. International journal of remote sensing, 25(9), 1565-1596.
 Hussain, M., Chen, D., Cheng, A., Wei, H., & Stanley, D. (2013). Change
detection from remotely sensed images: From pixel-based to object-based
approaches. ISPRS Journal of photogrammetry and remote sensing, 80,
91-106.
 Willis, K. S. (2015). Remote sensing change detection for ecological
monitoring in United States protected areas. Biological Conservation, 182,
233-242.
 Coppin, P. R., & Bauer, M. E. (1996). Digital change detection in forest
ecosystems with remote sensing imagery. Remote sensing reviews, 13(3-
4), 207-234.
 Wen, D., Huang, X., Bovolo, F., Li, J., Ke, X., Zhang, A., & Benediktsson, J.
A. (2021). Change detection from very-high-spatial-resolution optical remote
sensing images: Methods, applications, and future directions. IEEE
Geoscience and Remote Sensing Magazine, 9(4), 68-101.

72 Contributor: Dr. Sapana B. Chavan

 Unit 15 Change Detection Techniques
…………………………….…………………………………….……………………………………………….…
 Chen, G., Hay, G. J., Carvalho, L. M., & Wulder, M. A. (2012). Object-based
change detection. International Journal of Remote Sensing, 33(14), 4434-
4457.
 Shi, W., Zhang, M., Zhang, R., Chen, S., & Zhan, Z. (2020). Change
detection based on artificial intelligence: State-of-the-art and challenges.
Remote Sensing, 12(10), 1688.
 Qin, R., Tian, J., & Reinartz, P. (2016). 3D change detection–approaches
and applications. ISPRS Journal of Photogrammetry and Remote Sensing,
122, 41-56.
 Hemati, M., Hasanlou, M., Mahdianpari, M., & Mohammadimanesh, F.
(2021). A systematic review of landsat data for change detection
applications: 50 years of monitoring the earth. Remote sensing, 13(15),
2869.

15.8 FURTHER/SUGGESTED READINGS

 Asokan, A., & Anitha, J. J. E. S. I. (2019). Change detection techniques for
remote sensing applications: A survey. Earth Science Informatics, 12, 143-
160.
 Bai, T., Wang, L., Yin, D., Sun, K., Chen, Y., Li, W., & Li, D. (2023). Deep
learning for change detection in remote sensing: a review. Geo-spatial
Information Science, 26(3), 262-288.
 Lu, D., Mausel, P., Brondizio, E., & Moran, E. (2004). Change detection
techniques. International journal of remote sensing, 25(12), 2365-2401.
 Shafique, A., Cao, G., Khan, Z., Asad, M., & Aslam, M. (2022). Deep
learning-based change detection in remote sensing images: A review.
Remote Sensing, 14(4), 871.
 Coppin, P., Jonckheere, I., Nackaerts, K., Muys, B., & Lambin, E. (2004).
Review ArticleDigital change detection methods in ecosystem monitoring: a
review. International journal of remote sensing, 25(9), 1565-1596.
 Hussain, M., Chen, D., Cheng, A., Wei, H., & Stanley, D. (2013). Change
detection from remotely sensed images: From pixel-based to object-based
approaches. ISPRS Journal of photogrammetry and remote sensing, 80,
91-106.
 Willis, K. S. (2015). Remote sensing change detection for ecological
monitoring in United States protected areas. Biological Conservation, 182,
233-242.
 Wen, D., Huang, X., Bovolo, F., Li, J., Ke, X., Zhang, A., & Benediktsson, J.
A. (2021). Change detection from very-high-spatial-resolution optical remote
sensing images: Methods, applications, and future directions. IEEE
Geoscience and Remote Sensing Magazine, 9(4), 68-101.
 Chen, G., Hay, G. J., Carvalho, L. M., & Wulder, M. A. (2012). Object-based
change detection. International Journal of Remote Sensing, 33(14), 4434-
4457.

73
Contributor: Dr. Sapana B. Chavan
Block 3 Image Classification and Change Detection Techniques
…………………………………………………………………….…………………………………………………
 Shi, W., Zhang, M., Zhang, R., Chen, S., & Zhan, Z. (2020). Change
detection based on artificial intelligence: State-of-the-art and challenges.
Remote Sensing, 12(10), 1688.
 Qin, R., Tian, J., & Reinartz, P. (2016). 3D change detection–approaches
and applications. ISPRS Journal of Photogrammetry and Remote Sensing,
122, 41-56.
 Hemati, M., Hasanlou, M., Mahdianpari, M., & Mohammadimanesh, F.
(2021). A systematic review of landsat data for change detection
applications: 50 years of monitoring the earth. Remote sensing, 13(15),
2869.

15.9 ANSWERS
SAQ I
a) Identifying, describing, and quantifying variations between images of the
same scene taken at different times or under various circumstances are the
goal of change detection analysis, which includes a wide range of
techniques.
b) Automatic object-based image analysis has been increasingly popular in
recent years for remote sensing applications such as mapping from high
resolution photos or building and updating GIS databases.
c) The agents responsible for anthropogenic changes refers to human-induced
change, which is typically tied to land use and includes activities like urban
development, farming, logging, and mining.

SAQ II
a) Image ratioing is utilization of the same two co-registered image dates, each
band is rationed pixel by pixel.
b) Multi-date Principal Component Analysis is analysis where two images from
the same location taken at different times are superimposed and analyzed
as a single image. The local alterations are revealed by the minor
component images, while the major component images display the albedo
(reflectance) and radiometric variances (minor changes).
c) A change vector of a pixel is the vector difference between the multi-band
digital vector of the pixel on two different dates.
d) Ratio Vegetation Index, Normalized Vegetation Index, and Transformed
Vegetation Index are examples of common vegetation indices.

Terminal Questions
1. An inventory level of data indicating the location, nature, and extent of
change is provided by the satellite land cover change information
2. The line of land-water contact is referred to as the shoreline.
3. UAVs are a feasible alternative for gathering data from distant sensing for a
variety of real-world uses.

74 Contributor: Dr. Sapana B. Chavan

 UNIT 16

ACCURACY ASSESSMENT
OF THEMATIC MAPS
Structure_______________________________________________
16.1 Introduction For Hard and Soft Classification Outputs
Expected Learning Outcomes 16.6 Approaches to Accuracy Assessment for
other Data/Outputs
16.2 Accuracy Assessment
Change Detection Outputs
Purpose
Digital Elevation Models
Types
Modelled Outputs
Metrics
Vector Data
Steps
16.7 Challenges and Recent Trends
16.3 Historical Development
Challenges
16.4 Considerations for Accuracy
Assessment Approaches to Improving Accuracy
Sources of Errors in Classification Outputs Recent Trends and Developments
Factors Affecting Accuracy Assessment 16.8 Summary
Sampling Design Consideration 16.9 Terminal Questions
Determining Sample Size 16.10 References
16.5 Approaches to Accuracy Assessment 16.11 Further/Suggested Readings
For Per-pixel and Object based 16.12 Answers
Classification Outputs

16.1 INTRODUCTION
You have studied in the course MGY-102 that digital image processing involves three broad
image functions namely, image pre- processing, image processing and image post-processing. In
the Block 2 of this course, i.e. MGY-005, you have studied in detail about the image pre-
processing techniques that include image correction, image enhancement and transformation
techniques.

Contributor: Prof. Benidhar Deshmukh

Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
You have also studied in detail about thematic information extraction such as
image classification techniques and also change detection techniques. The final
step is the post-classification, which involves generation of output of final
thematic map/image and evaluation of its accuracy.
In the first two units of this block, you have studied about unsupervised and
supervised image classification techniques through which you create thematic
maps. You have also learnt about the advantages and limitations of some of the
commonly used classification algorithms. Both supervised and unsupervised
classification approaches need direct and indirect information about the
characteristics of the objects present in the study area. For example, for
unsupervised classification, the user labels the classes based on prior
information of the ground features, and in case of supervised classification, it is
based on the training sites. Quality and quantity of training samples, therefore,
have considerable implication on the accuracy of the classification results.
Once you have a classified image, the obvious step is that you would want to
know how much accurate those outputs are because inaccuracies in outputs
will have their bearing on the map’s utility and users would have greater
confidence in utilising the map if its accuracy is good and acceptable. Assessing
accuracy of the generated thematic maps falls under the post-classification
step. Post-processing of the classified image/map is a very important part of
the interpretation as it not only tells you about quality of maps generated or
classified images but also provides you with a benchmark to compare different
interpretation and classification methods. You have already been introduced to
accuracy assessment in the course MGY-102. In this unit, you will learn more
about accuracy assessment and various related aspects.

Expected Learning Outcomes__________________

After studying this unit, you should be able to:
 recall and define the terms associated with accuracy assessment;
 discuss the historical development of accuracy assessment;
 list various sources of errors in classification outputs and the factors
affecting accuracy assessment;
 describe the considerations regarding sampling design and sample size;
 elaborate on the approaches to accuracy assessment for various types of
image classification and object detection outputs;
 write about the approaches useful for accuracy assessment of various
other kinds of data/outputs such as change detection, DEM, modelled
outputs and vector data;
 discuss the challenges to accuracy, accuracy assessment, and the
approaches to improve accuracy; and
 highlight recent trends and development in the realm of accuracy
assessment..

76
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..

16.2 ACCURACY ASSESSMENT

You have read earlier in the course MGY-102 that accuracy assessment is the
last step in image processing tasks. In the context of remote sensing, accuracy
assessment is a critical step that evaluates the quality of outputs generated
through image classification, ensuring that the information derived from satellite
or aerial imagery accurately reflects the real-world conditions. Assessment of
accuracy is critical because accuracy of an output/data can significantly
influence its use in decisions related to environmental monitoring, land-use
planning, agriculture, forestry, and other fields.
Let us recall and discuss in some detail about the purpose, types, metrics used
and steps in accuracy assessment under different subsections.
16.2.1 Purpose
As you have read earlier, accuracy assessment is an integral part of any
mapping project. The purpose of accuracy assessment is to compare the
classified results generated from remote sensing data with reference data (i.e.
reality), which is usually collected from ground truthing, existing map, or from
higher spatial resolution satellite remote sensing images/ aerial photographs, or
expert knowledge. The user community needs to know accuracy of the
image/map being used. Different projects may have different accuracy
requirement and only those images/maps which are above a certain level of
accuracy can be useful. Accuracy becomes a critical issue while working in a
GIS framework where you use several layers of remotely sensed data and other
thematic layers. In such cases, it would be very important to know the overall
accuracy which is dependent upon knowing the accuracy of each of data layers.
Assessment of accuracy is important for the following reasons:
a) it allows self-evaluation and to learn from mistakes in the classification
process. It helps to identify areas of misclassification and validates reliability
of the classification
b) it provides quantitative comparison of various methods, algorithms and
analysts,
c) It improves future classification processes by highlighting shortcomings
related to algorithm or data. and
d) it also ensures greater reliability of the resulting maps/spatial information to
use in decision-making process.
The need for accuracy assessment is emphasised in literature as well as in
anecdotal evidence. For example, there could be maps on a specific theme
generated by various agencies/departments using techniques that includes
satellite images, aerial photographs and field data. Simply comparing the
various maps would yield little agreement about location, size and extent of the
features contained. In the absence of a valid accuracy assessment you may
never know which of these maps to use.
Accuracy assessment is essential for validating remote sensing products in
various applications, such as land use land cover mapping for verifying the
77
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
classification of different land cover types such as forests, water bodies, urban
areas, etc. It is useful in environmental monitoring to ensure accurate detection
of changes in land use, deforestation, or urban sprawl. It is also useful in the
event of disasters to evaluate accuracy of flood or wildfire extent maps derived
from remote sensing data. Accuracy assessment is indispensable for ensuring
the reliability of remote sensing data. By systematically evaluating classification
results against reference data, this process helps to improve the overall quality
of remote sensing analyses, leading to better-informed decisions in a wide
range of fields.
16.2.2 Types
Accuracy assessment can be broadly divided into two types. It can be either
qualitative or quantitative. Quantitative assessment uses statistical methods to
compare classified data with reference data. In qualitative assessment, you
determine if a map ‘looks right’ by comparing what you see in the map or image
with what you see on the ground. Qualitative assessment involves expert
evaluation of the classification results, often used when statistical methods are
not feasible. Quantitative assessment attempts to identify and measure remote
sensing based map errors. In such assessments, you compare map data with
ground truth data, which is assumed to be 100% correct.
You have already earlier in MGY-102 that accuracy is categorised into site-
specific accuracy and non-site-specific accuracy. Non-site-specific accuracy is a
simple method of comparison of two maps in which the total area assigned to
each class in both maps and the overall figures are compared. Site-specific
accuracy is based on comparison of two maps/images at specific locations (i.e.
individual pixels in two digital images), in which the degree to which pixels in
one image spatially align with pixels in the second image reflects accuracy
assessment.
16.2.3 Metrics
Accuracy of image classification is most often reported as a percentage correct.
As you have learnt earlier, commonly used metrics in accuracy assessment are
the following:
The producer’s accuracy (PA) or Omission error measures how well the
classification process identifies pixels of a particular class. It is a probability that
a reference pixel has been correctly classified and shows what percentage of a
particular ground class was correctly classified. It is calculated by dividing the
number of correctly classified pixels of a class by the total number of reference
pixels of that class. It informs the image analyst of the number of pixels correctly
classified in a particular category as a percentage of the total number of pixels
actually belonging to that category in the image. Producer’s accuracy measures
errors of omission. It is used when the same is viewed from analyst’s
perspective.
User’s or Consumer’s accuracy (CA) or Commission error measures the
reliability of a classification result, indicating the likelihood that a pixel classified
into a given class actually represents that class on the ground. It is the
probability that the class of a pixel actually represents that same class on the
ground. It is a measure of the reliability of a map generated from a
78
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
classification. It is calculated by dividing the number of correctly classified pixels
by the total number of pixels classified as that class. It is computed using the
number of correctly classified pixels to the total number of pixels assigned to a
particular category. It takes errors of commission into account by telling the
user/consumer that, for all areas identified as category X, a certain percentage
are actually correct. It is used when a classified image is examined from the
user’s point of view.
The two metrics discussed here are the individual class errors that report error
levels for an individual information class. At times we would also be interested
to know error levels for the overall map, which is the error levels averaged for all
information classes. Overall Accuracy is the proportion of correctly classified
pixels (both true positives and true negatives) out of the total number of pixels.
It gives a general view of the classification accuracy. These accuracy measures
are calculated from an error matrix. Overall accuracy is the commonly cited
measure of mapping accuracy which is the number of correctly classified pixels
(sum of major diagonal cells in the error matrix) divided by total number of pixels
checked. Though, overall accuracy is a measure of accuracy for the entire
image across all classes, it ignores off-diagonal elements (i.e. errors of
omission and commission). Further, it is difficult to compare different overall
accuracy values if different number of accuracy sites were used. The other two
accuracies such as producer’s and consumer’s accuracies are also calculated
from error matrix. The producer’s accuracy is a measure of how well a certain
area is classified. The consumer’s or user’s accuracy is a measurement of
reliability of the classification or probability that a pixel on a map actually
represents the category on the ground.
All these “naïve” accuracy measures can produce results due to classification of
pixels by chance, therefore do not provide avenues to compare accuracy
statistically. This paves way for use of other accuracy assessment methods.
Another commonly used method known is Cohen’s Kappa statistics, in which
off-diagonal elements of an error matrix are incorporated as a product of the
row and column marginal totals. It is a discrete multivariate technique used to
assess classification accuracy from an error matrix. Kappa analysis generates a
kappa coefficient or Khat statistics, the value of which ranges between 0 and 1.
Through Kappa statistics we know how well the classification performed in
comparison to randomly assigning the pixels to a specific class.
Kappa coefficient (Khat) is a measure of the agreement between two maps
taking into account all elements of error matrix. It is defined in terms of error
matrix as given here:
Khat = (Obs – Exp) / (1 – Exp)
where,
Obs = Observed correct, it represents accuracy reported in error matrix (overall
accuracy)
Exp = Expected correct, it represents correct classification
Kappa coefficient is a statistical measure that compares the observed accuracy
with an expected accuracy (random chance). It provides a measure of
agreement between the classified data and the reference data, taking into
79
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
account chance agreement. Values of Kappa range from -1 to 1, where values
closer to 1 indicate high accuracy. In other words, it is a measure of difference
between the observed agreement between two maps and the agreement that
might be attained solely by chance matching of the two maps:
Khat = (Observed Accuracy - Chance Agreement) / (1 - Chance Agreement)
According to Landis and Coch (1977), Kappa indicates the following:
i) If the value is > 0.8 (>80%) it reflects strong agreement
For perfect agreement (i.e. highest accuracy), the kappa value is 1. And, a value
of 0.82 implies that the classification process was avoiding 82% of the errors
that a completely random classification would generate
ii) the value between 0.4 and 0.8 indicates moderate agreement
iii) and the value < 0.4 indicates poor agreement
It is suggested that Kappa analysis can also be used to compare two
classifications of same area made from different image dates and/or different
algorithms and also by different image analysts.
However, some research including recently by Foody (2020) suggest that Kappa
is not suitable for accuracy assessment as it has its own limitations. Read the
article to learn more about it.
There are some other metrics which are used to measure accuracy. You will
learn about some of them later in this unit.
Error and uncertainty analysis is carried out in some cases, which identifies and
quantifies the types and sources of errors, such as spectral overlap between
classes, sensor limitations, and human error in reference data collection. It is
also discussed in the context of modelling. You will be introduced to these terms
in the course MGY-006.
16.2.4 Steps
You have learnt about carrying out accuracy assessment in the course MGY-
102. The major steps in accuracy assessment are shown in the Fig. 16.1:

80
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
Fig. 16.1: Broad steps in accuracy assessment.

Usually, the first step is the collection of reference data either from the ground or
from a higher resolution data based on the carefully designed sampling strategy
considering number of type of approach employed for thematic information
extraction, classes present, characteristics of the features, accessibility on the
ground, etc. Care should be taken that the chosen or collected ground truth data
accurately represents the classes in the study area for the land use land cover
or any other theme as per the thematic map prepared.
After collecting the reference data, the second step is to create error matrix
(also called confusion matrix) to compare the classification results with the
reference data. The third step is calculation of accuracy metrics such as overall
accuracy, producer’s and user’s accuracy, and the Kappa coefficient. Next step
is to analyse the metrics to understand the strengths and weaknesses of the
classification. Further, at the last step the findings are documented and are used
to refine classification algorithms and data processing techniques.

16.3 HISTORICAL DEVELOPMENT

Accuracy assessment in the context of remote sensing has evolved significantly
over the years. The concept of accuracy assessment in remote sensing has
evolved significantly over the decades, driven by advancements in technology,
increasing complexity of remote sensing applications, classification algorithms,
and the growing need for reliable data in environmental and spatial analysis.
This historical development can be traced through several key phases,
highlighting how the approaches and methodologies have been refined over
time.
Early Beginnings (1930s – 1960s): The concept of accuracy assessment in
remote sensing can be traced back to the aerial photography era, which was
the earliest form of remote sensing, primarily used during World War II and the
post-war period for reconnaissance and mapping. During this time, there was
little formalised accuracy assessment, and validation was primarily qualitative,
involving manual comparison of aerial photographs with ground surveys.
However, need was felt to have some standards. One of the earliest efforts
towards this may be attributed to the setting up of a committee by American
Society of Photogrammetry (Now the American Society for Photogrammetry
and Remote Sensing [ASPRS]) in 1937, to draft spatial accuracy standards for
maps prepared from remotely sensed data and two early publications in 1941 of
US National Map Accuracy Standards (NMAS) and in 1947 of National Map
Accuracy Standard by U.S. Bureau of Budget which included aspects of
horizontal and vertical accuracies among others (Congalton and Green, 2019).
This marked the critical step towards implementing consistency in positional
accuracy however it lacked the procedures for measuring accuracy (Congalton
and Green, 2019). Further, the Manual of Photo Interpretation published in 1960
by ASPRS also recognised the need to train and test photo interpreters.
Emergence of Satellite Remote Sensing (1970s – 1980s): The launch of the
first Earth observation satellites, such as Landsat-1 in 1972, marked a
significant milestone in the history of satellite remote sensing. As remote
sensing technologies advanced and the need for reliable data became more
81
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
apparent, the concept of accuracy assessment was introduced in second half of
1970s in three publications during 1976 to 1979 proposing criteria for testing
overall map accuracy. However, accuracy assessment was still rudimentary,
relying heavily on visual interpretation and limited quantitative measures. Early
assessments were often subjective, involving the comparison of classified maps
with ground truth surveys. However, standardised method for accuracy
evaluation was lacking, and error analysis was minimal. The early 1980s saw
proposal of new techniques through some publications.
During this time the error matrix (i.e. confusion matrix) was introduced by Story
and Congalton (1986), which became the standard tool for assessing
classification accuracy in terms of overall accuracy, producer’s accuracy and
user’s accuracy. This period marked a shift towards more structured and
quantitative approaches, providing a way to systematically compare classified
data with reference data. Researchers began to adopt statistical measures such
as overall accuracy, user’s accuracy, and producer’s accuracy, derived from the
error matrix. The Kappa coefficient, introduced in remote sensing became a
popular metric for evaluating agreement beyond chance, offering a more robust
assessment of classification performance.
With advent and adoption of digital image processing automated classification
of satellite images commenced. This led to an increased need for objective and
reproducible accuracy assessments approach, driving the development of
standardised approaches.
Standardisation and Methodological Refinement (1990s): During the 1990s,
organisations such as the American Society for Photogrammetry and Remote
Sensing (ASPRS) and the International Society for Photogrammetry and
Remote Sensing (ISPRS) began to publish guidelines and best practices for
accuracy assessment, to further standardise accuracy assessment methods
across the field. One of the key publications includes ASPRS Interim Accuracy
Standards for Large-Scale Maps (ASPRS, 1990) that included hardcopy as well
as digital maps and provided guidance on sampling and reference points.
Another notable publication is the U.S. National Cartographic Standards for
Spatial Accuracy (NCSSA) by Federal Geographic Data Committee (FGDC) in
1998, which established standards for medium and small scale maps. It was
revised as National Standard for Spatial Data Accuracy (NSSDA) (FGDC, 1998)
to adopt positional accuracy assessment procedures in lieu of accuracy
assessment standard and recommended reporting accuracy in ground distance
units at the “95% confidence level” and also recognising variation in map user’s
requirements and to publish their own standards (Congalton and Green, 2019).
It was the accepted standard for the next about two decades and used in
conjunction with the APPRS large scale map standards (Congalton and Green,
2019). While the NSSDA provided standardised processes for assessment of
positional accuracy, the APPRS standards set the maximum allowable errors
maps of different scales (Congalton and Green, 2019).
The importance of sampling design was recognised, which led to the
developments of more rigorous sampling strategies, including stratified random
sampling and systematic sampling. These designs aimed to ensure that the
reference data were representative of the entire study area. This period also
saw the development of error budgeting approaches that aimed to quantify and
82
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
partition the sources of errors in remote sensing classification, such as sensor
noise, atmospheric effects, and classification algorithm errors.
Integration of Advanced Statistical Methods (2000s): Traditional hard
classification methods were found to be inadequate for complex landscapes
with mixed land cover types. This led to the introduction of fuzzy logic and soft
classification methods, which accounted for the uncertainty and gradual
transitions between classes. Accuracy assessment methods were adapted to
evaluate these soft classifications using fuzzy error matrices.
The rise of machine learning and data-driven approaches, such as Support
Vector Machines (SVM), Random Forests, and Neural Networks, introduced
new complexities in accuracy assessment. These algorithms often required
advanced validation techniques like cross-validation and bootstrapping to
assess model performance. New accuracy metrics, such as spatial accuracy
and thematic accuracy (confusion among similar classes), were developed to
provide a more comprehensive evaluation of classification results.
Modern Approaches and Big Data Era (2010s - Present): With the advent of
high-resolution imagery, LiDAR, UAVs, and hyperspectral sensors, accuracy
assessment methods have had to evolve to handle more complex datasets.
Multi-source data integration required new approaches to validate
classifications across different data types and resolutions. Spatial cross-
validation technique was introduced for geospatial machine learning outputs,
which reduces spatial bias from spatially autocorrelated samples.
The rise of deep learning, particularly Convolutional Neural Networks (CNNs),
has transformed remote sensing classification, leading to more sophisticated
accuracy assessment methods that often involve complex validation
techniques, including ensemble methods and probabilistic assessments.
A document named, ASPRS Positional Accuracy Standards for Digital Data
came out in 2014 which provides the most comprehensive discussion of
positional accuracy developed so far and established standards for maps of
different quality and scale (Congalton and Green, 2019). In 2022, ASPRS
established a Positional Accuracy Standards Working Group under the
Standards Committee to evaluate user comments and consider technology
advancements to implement appropriate changes to the standards and the
document has recently been revised to its 2nd Edition in 2023 incorporating
some important changes. The changes include relaxed accuracy requirement
for ground control and checkpoints, elimination of references to 95% confidence
level, required inclusion of survey checkpoint accuracy when computing
accuracy of final product, removal of pass/fail requirement for vegetated vertical
accuracy for lidar data, increased minimum number of checkpoints required for
product accuracy assessment from 20 to 30, limited minimum number of
checkpoints for large projects to 120, measure for horizontal accuracy for
elevation data, introduction of the new 3D accuracy measure among others and
also addition of five Addenda on best practices and guidelines (Abdullah, 2023;
ASPRS, 2023). Version 2 of the 2nd edition is also expected to come in 2024.
Recent advances include the development of AI-driven accuracy assessment
tools that automate the process, reducing human bias and improving the

83
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
efficiency of validation. These systems can dynamically update and refine
reference datasets as new data becomes available.
The use of crowdsourced data and citizen science initiatives, such as through
platforms like Google Earth Engine, has expanded the availability of reference
data. This has led to the development of new methods to assess and validate
classifications using data from non-traditional sources.
The evolution of accuracy assessment in remote sensing reflects the broader
technological advancements in the field. From simple qualitative comparisons in
the early days to sophisticated statistical and AI-driven assessments today, the
field continues to develop new methods that enhance the reliability of remote
sensing data. This historical progression underscores the importance of
accuracy assessment as a dynamic and essential component of remote sensing
analysis, adapting to the increasing complexity and demands of modern
geospatial applications.
Let us spend 5 minutes to check your progress.

SAQ I
a) What are the types of accuracy assessment?

b) Write the major developments in the standardisation and methodological

refinement era in the context of accuracy assessment.

c) What are the major steps in accuracy assessment?

16.4 CONSIDERATIONS FOR ACCURACY

ASSESSMENT
You know that there are several factors that affect accuracy assessment in
remote sensing data analysis. Understanding these factors is crucial for
improving classification quality and ensuring that the results accurately
represent real-world conditions.
There are two things we will discuss: first we will learn about the sources of
errors that affect classification outputs and then about the factors that affect
assessment of accuracy.
16.4.1 Sources of Errors in Classification Outputs
Errors in classification outputs of remote sensing data may occur due to various
factors affecting the quality of the input data, processing steps, and the
classification algorithms used. These errors can significantly impact the
accuracy of classification outputs, leading to incorrect thematic maps and
misinterpretations. Some of the major sources of errors in classification outputs
in remote sensing are shown in Table 16.1.
Table 16.1: Major sources of errors in remote sensing data based
classification outputs.
Source of Error Description Impact Possible Solution
Sensor Differences in sensor Sensor noise, Use well-calibrated sensors
characteristics types, calibration, calibration errors, and with appropriate spectral,
and limitations radiometric resolution, limited spectral bands spatial, and radiometric

84
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
Source of Error Description Impact Possible Solution
and noise levels can can introduce resolution, and apply noise
affect the data quality inaccuracies in reduction techniques, if
classification needed
Use additional spectral
bands, indices (e.g., NDVI),
Different land cover Results in or apply advanced
Spectral classes may have misclassification algorithms like machine
overlap and overlapping or similar thereby reducing learning models that can
similarity spectral signatures e.g. classification accuracy better handle overlapping
between water and shadow, as similar classes are classes or incorporate
classes leading to not distinguished additional data (e.g.,
misclassification properly texture, elevation) and
incorporate contextual
information in classification
Use sub-pixel classification
Causes
techniques, spectral
misclassification,
Pixels containing more unmixing techniques, fuzzy
especially in
than one land cover classification approach to
heterogeneous or
Mixed pixels types, especially common handle mixed pixels better
transitional zones and
and boundary at class boundaries due or higher spatial resolution
overestimation or
effects to the spatial resolution, data; Use boundary
underestimation of
leading to ambiguous smoothing techniques,
class boundaries,
classification post-classification filtering,
affecting area
or segmentation-based
calculations
classification
Radiometric noise,
Degrades the quality of Apply preprocessing steps
Noise and atmospheric effects,
spectral information, like radiometric correction,
distortions in sensor errors and
leading to incorrect atmospheric correction,
data preprocessing artifacts
classification and image normalisation
may affect data quality
Differences in sunlight Results in inconsistent
due to time of day, reflectance values,
season, or topography, affects classification Apply topographic
Shadow and
shadows from terrain, performance in shaded normalisation, shadow
illumination
buildings, or vegetation areas, especially in detection and removal, or
effects
obscure surface features, urban and forested use of multi-temporal data
altering their spectral regions or other low-
properties reflectance classes
Atmospheric effects like Causes spectral Apply atmospheric
haze, clouds, and aerosol distortion, reducing the correction like Dark Object
scattering affect sensor ability to differentiate Subtraction (DOS) or
measurements and alter between classes, atmospheric radiative
Atmospheric
the signal received by results in gaps or transfer models, use cloud
conditions
sensors, presence of incorrect classifications masking, gap-filling
clouds and their shadows in cloud-covered areas, algorithms, or multi-
obscure the land surface, especially in optical temporal compositing to
leading to missing data data reduce cloud impacts
Alters spectral
Terrain-induced signatures, causes
Topographic variations in reflectance misclassification in Apply topographic
effects due to slope, aspect, and mountainous regions or correction techniques
elevation differences areas with varied
topography
Changes in land cover Leads to outdated or
between image misleading Use recent data,
acquisition dates and the classification outputs, incorporate multi-temporal
Temporal
ground truth data especially in dynamic datasets, or apply change
changes in
collection affect landscapes; Causes detection techniques; use
landscape/
classification consistency; temporal mismatches multi-seasonal data,
seasonal
differences in vegetation and classification phenological adjustment, or
variations
phenology and land cover errors, especially in seasonal compositing
conditions across agriculture and forest techniques
seasons monitoring
Errors due to sensor Results in spatial Apply geometric correction,
Geometric
geometry, earth misalignment and orthorectification, and
distortion/
curvature, or relief incorrect positioning of accurate georeferencing
misregistration
displacement affecting features and affecting using precise ground
85
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
Source of Error Description Impact Possible Solution
pixel locations the classification control points
Differences in data scales Can lead to bias and Normalise or standardise
Scaling issues between training and poor classification data to ensure consistent
classification datasets performance scaling across datasets
Improve training data
Insufficient or through field validation,
Inaccurate, insufficient, or
unrepresentative expert annotation, and
Inadequate non-representative
training samples can sufficient number of high-
training sample training data can
lead to biased or quality, well-distributed
size and quality misguide the
inaccurate classification training samples ensuring
classification model
results balanced class
representation
Standardise procedures,
Errors introduced during use automated approaches
Leads to
Human manual classification or when possible, and
inconsistencies and
interpretation training data collection incorporate multiple
inaccuracies in
errors due to subjective interpreters for removal of
classification outputs
judgment subjectivity in interpretation
and to minimise errors
Use techniques like
Uneven distribution of Dominant classes are oversampling,
Sample design/
classes in training data over-represented, while undersampling, or
class
can bias the model minority classes are Synthetic Minority Over-
imbalance
toward dominant classes under-classified sampling Technique
(SMOTE)
Use of irrelevant or
Increases the risk of
Suboptimal redundant features that Use feature selection
overfitting, leading to
feature do not contribute methods like Principal
decreased performance
selection significantly to Component Analysis (PCA)
on unseen data
classification accuracy
Choosing an unsuitable Results in overly Use appropriate
Inappropriate classification scheme that simplistic or overly classification scheme
classification does not capture the complex classifications, suitable to resolution, scale
scheme variability of land cover leading to high error and level of classes being
classes rates mapped
Results in reduced
Inherent weaknesses in classification accuracy Use ensemble methods,
classification algorithms due to inability to hybrid approaches, or
Algorithm in handling complex or correctly model data advanced machine learning
limitations overlapping classes and complexities especially algorithms such as
non-linear class in complex or Random Forests or Neural
boundaries heterogeneous Networks
landscapes
Mistakes in collecting or Use careful field validation,
Leads to biased training
Ground truth labelling ground truth automated data collection
data and unreliable
data quality data used for training and tools, and quality control
accuracy assessments
validation measures

Errors in classification outputs in remote sensing can stem from various factors
including data quality, algorithm limitations, environmental effects, and human
errors. Each error source can significantly impact the overall classification
accuracy, leading to incorrect land cover maps and unreliable results for
decision-making. Understanding and addressing these sources of errors is
crucial to improving the accuracy of classification outputs in remote sensing,
leading to more reliable and actionable thematic maps. Use of appropriate
preprocessing techniques for correcting geometric and radiometric distortions,
atmospheric corrections, and shadow removal techniques can help reduce
errors related to data quality. Judicious selection of suitable classification
scheme and algorithm, choice of advanced classification techniques that better
handle complex data patterns and non-linear relationships among classes and
use of indices (e.g. NDVI), incorporation of contextual information such as
86
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
texture, etc. and techniques to enhance the separability of classes can
significantly improve classification outputs. Further, ensuring high-quality,
representative, and adequately balanced training data can improve the
robustness of the classifier. Application of post-classification processing
techniques such as smoothing filtering, or object-based post-processing
techniques can help reduce classification noise and improve the spatial
coherence of the output. Lastly, careful validation can greatly improve the
reliability of classification outputs in remote sensing applications.
After learning about the sources of errors in classification outputs of remote
sensing data, let us now understand the factors that affect accuracy
assessment.
16.4.2 Factors Affecting Accuracy Assessment
There are several factors that may affect assessment of accuracy. Some of the
key factors influencing accuracy assessment are given in Table 16.2.
Table 16.2: Factors affecting assessment of accuracy in remote sensing
data analysis.

Factor Description and impact

The categories or classes defined for classification should be consistent with reference
Classification scheme to the data being used for mapping; inconsistent or incompatible schemes can cause
errors in accuracy assessment

Inaccurate, outdated, or imprecise ground truth data used for validation can affect
Quality of ground truth
data
validation accuracy e.g. inaccurate location data can lead to misleading accuracy
metrics and poor validation of classification results

The choice of method used and poor sampling design to collect reference data, (e.g.,
Sampling design random, systematic, stratified) can influence the representativeness of validation data,
bias results and under-represent classes

Inaccessibility to Inaccessibility to the sites being used for validation and over-sampling in easily
sampling location on accessible areas would lead to bias for those classes
ground

Characteristics of the The level of detail in terms of spatial and spectral resolutions of the remote sensing data
reference remote being used as reference data; can impact the ability to distinguish between different
sensing data being features; frequency of data acquisition and its relevance to the study period can affects
used for validation the relevance of data to the study period leading to incorrect reference data

Choice of suitable Maps generated through different classification method may require suitable accuracy
accuracy assessment assessment approach, e.g. hard and soft classification, per-pixel, object based, object
approach detection outputs may require separate approaches for accuracy assessment

Site specific Unique features of the study area that may affect data interpretation can lead to
characteristics misclassification, if not accounted for

Difference in the geographic coordinates in the data from which mapping has been
Positional accuracy done and on the ground can lead to wrong interpretation and affect quality of the data
being collected on ground

Differences in the time of data acquisition between the remote sensing data used for
Temporal mismatch classification and reference data being collected (e.g., seasonal variations) can lead to
inaccurate comparisons

Mistakes made during data collection, interpretation, or recording by field surveyors

Human error in ground
truthing
(e.g. misidentifying plant species) during field surveys can introduce inaccuracies into
the reference data, affecting overall assessment quality
87
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
The factors listed in the Table 16.2 highlights the complexity of accuracy
assessment in remote sensing. Properly addressing them can significantly
improve the quality and reliability of classification results, ensuring more
effective use of remote sensing data for various applications.
You have read about recoding as one of the post classification steps in MGY-
102. You have learnt that after image classification, it is necessary to recode the
classified image and merge or give same code or information class label to a
specific class. Recoding operation allows us to assign a new class value or to
merge multiple classes of the same land cover type. Further, recoding helps to
remove duplication of thematic classes and edit some observable errors in the
classes. It also enables us to recode the wrong class to the right class based on
their context (known as contextual editing) (Deshmukh et al., 2005) and to place
the number of output classes to the predefined value.

Fig. 16.2: Kadmat atoll in Lakshadweep archipelago, shown in standard false

colour composite in the left panel and eco-morphological zonation map
of the coral reef is shown in the central panel. Coral atolls have their
characteristic central lagoon. Seagrass beds are found here in the
shallow lagoon but its spectral signature is similar to the bottom
materials found in depth in the coralline shelf as marked in the FCC. You
may note that due to that spectral similarity the seagrass bed has been
misclassified as coralline shelf. The classified image was then subjected
to contextual editing to improve accuracy. Contextual editing is the
application of decision rules where merging of the classes takes place.
The context applied here is that the seagrass beds do not occur in
coralline shelf region, and the coralline shelf region does not occur in
the lagoonal area. The right panel shows the classified image after the
contextual correction and therefore having improved map accuracy.
(Source: modified from SAC, 2003)

16.4.3 Sampling Design Consideration

You know that accuracy assessment is done based on the reference data
primarily derived from the ground. Reference data may be taken from either
88
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
ground-based measurements or an existing map or even from some type of
remote sensing data having higher spatial resolution. Reference data is
collected from the study area at some locations. You have learnt earlier that
poor sampling design can bias results and under-represent classes. Hence,
consideration of sampling design is an important aspect. Determining the
appropriate sampling pattern for accuracy assessment in remote sensing is
crucial for generating reliable and unbiased results. The choice of sampling
pattern can significantly affect the quality of the accuracy assessment,
depending on the type of data, classification outputs, and the landscape
characteristics. Proper sampling ensures that the assessment results are
reliable and representative of the study area. However, following factors
influence our choice of sampling design:
 Objectives of the study
 Class distribution and variability
 Spatial autocorrelation
 Accessibility of the study area and available resources
Objective of the study such as to estimate detailed class accuracy or overall
accuracy can dictate the sampling approach. Class distribution and variability
and also accessibility to the study area can influence the choice of design. If
spatial autocorrelation is high, systematic or spatially stratified designs can be
more effective.
You have already read about the sampling scheme in MGY-102 but Table 16.3
outlines different methods used for determining sampling design in accuracy
assessment.
Table 16.3: Comparison of sampling design methods used for accuracy
assessment.
Method Description Suitability Advantage Limitation
Homogeneous Can be inefficient, if
Samples are
Simple landscapes e.g. Easy to implement; the study area is
chosen randomly
random forests minimises bias, if very large; may
over the entire
sampling well-randomised under-sample some
study area
classes
Regular or evenly
Samples are taken
spaced areas e.g. Simple to implement Can miss patterns,
Systematic at regular intervals
monitoring and uniform area if there is a regular
sampling (e.g., every nth
agriculture in regular coverage structure in the data
pixel or grid cell)
fields
Study area is Heterogeneous area
Ensures
Stratified divided into strata with known classes Requires prior
representation of all
random (i.e. classes), and e.g. urban-rural knowledge of strata;
classes; increases
sampling samples are taken classification can be complex
precision
from each stratum
Systematic Mixed
Stratified Reduces systematic
sampling within regular/irregular Complex to
Systematic bias; well-
strata with landscapes like implement
Unaligned represented
unaligned grids urban-rural areas
The area is divided Large areas or when Can be less
Cost-effective for
Cluster into clusters (e.g., individual sampling is precise, if clusters
large areas; easy to
sampling grid cells or impractical e.g. vast are not
manage
patches), and entire forested or homogeneous
89
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
Method Description Suitability Advantage Limitation
clusters are mountainous areas
sampled
Sampling strategy Areas with high Can improve
More complex to
adapts based on variability e.g. efficiency and
Adaptive design and
observations, often wetland with sharp accuracy; focuses
sampling implement; can
focusing on areas class transitions resources, where
introduce bias
with more variability needed
Equalised Equal samples from
Imbalanced class Improves small class Over-sampling of
stratified each stratum
areas e.g. wetlands representation. rare classes.
random regardless of size.
Samples Classes with varying Rare classes may
Proportional Reflects class
proportional to areas, e.g. forest and be
stratified distribution efficiently
stratum size non-forest underrepresented
Samples are taken For assessing
Useful for assessing Can miss variability
Transect along predefined gradients or linear
spatial gradients; perpendicular to
sampling lines or transects features
easy to implement transects
across the area
Combines different For complex surveys
More complex to
sampling methods needing detailed Flexible and can
Multi-stage design and analyse;
(e.g., stratified information provide detailed
sampling requires multiple
followed by random information
stages
sampling)

These methods help in visualising how sampling designs are implemented and
the trade-offs between them. Each method has its unique strengths and
weaknesses, which influence the choice of method depending on the specific
goals and constraints of the remote sensing project.
16.4.4 Determining Sample Size
You have learnt earlier that sample size is an important consideration while
assessing the accuracy of remotely sensed data. Proportion of correctly
identified locations will represent the accuracy of the map. Determination of
sample size depends on number of map classes and level of detail/rigor
required. Collection of sample size requires time and resources. Therefore, it
must be kept to a minimum. Yet it is critical to maintain a large enough sample
size so that analysis performed is statistically valid. In remote sensing literature,
many researchers have published equations and guidelines for choosing the
appropriate sample size. Majority of them have used an equation based on the
binomial distribution or the normal approximation to the binomial probability
distribution to compute the required sample size. A generally accepted rule of
thumb is to use a minimum of 50 samples for each class in the error matrix but
also to adjust based on class variability and landscape complexity. This rule
also tends to agree with the results of computing sample size using the
multinomial distribution. Given an error matrix with n land cover classes, for a
given class there is 1 (one) correct answer and n–1 incorrect answers. Sufficient
samples must be acquired to be able to adequately represent this confusion.
Therefore, use of the multinomial probability distribution is recommended.
Determining the appropriate sample size for accuracy assessment in remote
sensing is crucial to ensure reliable and statistically valid results. The sample
size influences the accuracy, precision, and confidence of the assessment.
There are different methods (as given in Table 16.4) used to calculate sample

90
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
size based on statistical principles, data characteristics, and the purpose of the
study.
Table 16.4: Commonly used methods for determining sample size.
Method Description Suitability Advantage Limitation
A simple heuristic
often used to
for preliminary
determine sample size
estimates or when Easy to apply; quick May not be statistically
Rule of thumb based on general
detailed information is estimation rigorous; lacks precision
recommendations or
unavailable
past experience for
quick estimation
for stratified
Sample size is
populations to ensure
Proportional allocated to different Ensures representation Requires accurate
each group is
allocation strata or groups in proportional to class size knowledge of strata sizes
proportionally
proportion to their size
represented
Can over- or under-sample
Equal samples for for situations where
Simple to implement; easy certain classes as does
Equal allocation each class regardless strata are considered
to balance across strata not account for differences
of their size equally important
in stratum sizes
Divides the population for heterogeneous
Increases precision by
Stratified into distinct strata and populations where Requires detailed
accounting for population
random samples randomly each stratum must be information on strata
heterogeneity
within each stratum represented
Calculates sample Assumes normal
for estimating
size for a specified Provides precise sample distribution; complex for
Cochran’s proportions with
level of precision and size for given confidence large populations; requires
formula known or estimated
confidence and precision knowledge of population
proportions
considering variability size
Used for binary
outcomes (e.g.,
Limited to binary
presence/absence) to binary classifications Simple to apply;
Binomial outcomes; does not
estimate the sample with known or statistically robust sample
distribution handle multiclass
size needed for a assumed proportions size
scenarios
specific level of
accuracy
Extends binomial to Requires knowledge of
Handles multiple classes;
Multinomial multiple categories or suitable for problems proportions for all classes;
more flexible than binomial
method classes for multiclass with multiple classes more complex
classification
Estimates sample size
required to detect a for hypothesis testing
Accounts for effect size, Requires prior knowledge
specific effect size and experiments
Power analysis power, and significance of effect size and variance;
with a given level of requiring power
level; flexible complex calculations
statistical power and estimation
significance
Dynamically adjusts
Suitable for Efficient for Requires ongoing
the sample size based
Adaptive heterogeneous or heterogeneous monitoring and
on observed data
sampling spatially variable populations; can reduce adjustments; can be
patterns or conditions
populations costs complex
during the study
Determines sample Suitable when
size based on expert statistical methods Provides practical Highly subjective; may
Expert
opinion and are impractical or recommendations based lack consistency and
judgement
experience rather than when expert insights on experience. scientific rigor
statistical formulas are crucial

It is required to have a balance between representation and resources. Some

methods, like power analysis and binomial distribution, require estimates of
expected accuracy, variability, or effect sizes. Although, methods based on
statistical distributions (e.g. binomial, Cochran’s formula) provide more reliable
sample size estimates but require careful parameter selection. Understanding
these methods and their implications helps ensure that the chosen sample size
91
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
for accuracy assessment is both statistically valid and practically feasible,
leading to more reliable remote sensing classification outcomes.
To learn to choose a sample selection method and determine sample size, and
allocate the sample to strata based on estimation objectives follow the link
https://openmrv.org/web/guest/w/modules/mrv/modules_3/sampling-design-for-
estimation-of-area-and-map-accuracy.
The important factors governing choice of the method are: study’s goals,
available resources, and the level of statistical rigor required. The determination
of sampling points and design in remote sensing accuracy assessment requires
a balance between statistical rigour, class representation, and practical
considerations. By carefully selecting and employing an appropriate sampling
design, researchers can ensure that the accuracy assessment is both reliable
and meaningful, reflecting the true quality of the classification results.
Let us spend 5 minutes to check your progress.

SAQ II
a) List the sources of errors in image classification.

b) How does sampling pattern affect accuracy assessment?

c) List the methods of determining sample size in the context of accuracy

assessment.

16.5 APPROACHES TO ACCURACY ASSESSMENT

You have read about the procedure of accuracy assessment in Unit 17 of the
course MGY-102. In the Units 13 and 14 of this course i.e. MGY-005, you have
been introduced to various techniques of classification of remote sensing
images. In this section, you will learn about accuracy assessment approaches to
various kinds of classification outputs such as per-pixel, object based, object
detection outputs, hard and soft classification, etc.
Let us first learn about the approaches for Per-pixel and object based
classifications.
16.5.1 For Per-pixel and Object Based Classification
Outputs
As you have learnt, per-pixel classification assigns each pixel in an image to a
specific class based on its spectral properties whereas object-based
classification groups pixels into meaningful objects based on both spectral and
spatial information. Hence there could be different approaches to accuracy
assessment of these two kinds of outputs.
Accuracy assessment methods for per-pixel classification include use of
confusion matrix, the most common method, which compares the classified
image with a reference (ground truth) data to calculate metrics like overall
accuracy, producer’s accuracy, user’s accuracy, and the Kappa coefficient.
Object based accuracy assessment evaluates classification accuracy by
comparing classified objects (segments) to reference data on an object level. At
the initial level, it involves visual inspection of the classified objects and
92
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
comparing them with reference data to assess accuracy. However, it may be
subjective and may require multiple interpreters to minimise bias. It requires
accurate delineation of objects and also considers scale and segmentation
accuracy. In the case of per-pixel classification the focus is generally on the
spectral errors whereas in case of object based classification it includes both
spectral and spatial errors.
Accuracy assessment of object-based classification includes use of object
confusion matrix, which is similar to the pixel-based confusion matrix but the
major difference is that it is applied to objects instead of individual pixels. It
means that it operates at the object level, comparing classified objects to
reference objects. It accounts for variability in object size and shape and also
boundaries of objects. For object based classification, it is important to evaluate
i) quality of the segmentation process (with metrics such as under-segmentation
and over-segmentation indices), ii) area of correctly classified objects (area
based metrics), and also iii) shape accuracy of the classified objects (shape
based metrics). Shape and size based metrics evaluate accuracy based on the
shape and size of classified objects compared to reference data. However,
shape irregularities and size discrepancies can impact accuracy. There are also
fuzzy accuracy measures that are applied to assess accuracy where some
degree of uncertainty or overlap exists. It is useful for handling ambiguity and
mixed pixels in object-based classifications.
For some applications, we may not be interested in preparing thematic maps
having multiples classes rather we may be interested in knowing locations of a
certain object. In such scenario, object detection is used which is the technique
for identifying and classifying objects of certain defined classes in images.
Output of object detection is an image with bounding boxes and labels on
detected objects instead of having boundaries of the objects marked as we see
in the classification outputs. In case of object detection, we would like to know
how accurately the model has predicted the class and also how close is the
bounding box to the ground truth. Hence, accuracy assessment of object
detection results would require a different set of accuracy metrics that help
assess how accurately an object detection model identifies, localises, and
classifies objects within images.
Intersection over Union (IoU) ratio, confusion matrix (Precision, Recall, F1-
score), Mean Average Precision (mAP), Area-based accuracy (ROC curve),
feature similarity, distance difference are used for the purpose.
IoU ratio is ratio of intersection of the two bounding boxes to the union of the
two bounding boxes. It quantifies object localisation accuracy and measures the
overlap between the predicted bounding box and the ground truth bounding box;
its values ranges between 0 (no overlap) and 1(perfect overlap) is useful to
measure the quality of a predicted box against the ground truth; hence, it can
serve as a threshold to discard or accept predictions. Precision is ratio of true
positive detections to all positive predictions (true positives + false positives),
Recall is the ratio of true positive detections to all actual positives (true positives
+ false negatives); and F1-score is the harmonic mean of precision and recall.
Here, the terms True Positive (TP) refers to correct classification of the object,
False Positive (FP)-incorrect classification of the object, True Negative(TN) -
correct classification of the object as not being that class, and False Negative
93
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
(FN) - incorrect classification of the object as not being that class. These terms
are borrowed from the binary classification tasks.
Average Precision (AP) is weighted mean of precisions at each threshold; mAP
is the average of AP of each class. ROC curve is created by plotting the true
positive rate (TPR) against the false positive rate (FPR) at various threshold
settings; (where, TPR is the proportion of positive data points correctly
considered as positive, with respect to all positive data points; and FPR is the
proportion of negative data points that are mistakenly considered as positive,
with respect to all negative data points).
To learn accuracy assessment of a two class problem with a sample confusion
matrix follow the steps given in the article by Nicolau et al 2024, wherein they
have shown how to calculate producer’s, user’s, overall and kappa coefficient
through TP, TN, FP and FN.
16.5.2 For Hard and Soft Classification Outputs
You have read earlier that hard classification assigns each pixel to a single
class whereas soft or probabilistic classification assigns each pixel a probability
distribution over multiple classes, rather than a single, discrete class. So, it
allows pixels to belong to multiple classes with varying degrees of membership.
Soft classification approach is useful for dealing with mixed pixels where
multiple classes may be present. Since the two types of classification
approaches are different, the accuracy assessment measures of these two
types of classified outputs should be different. Gu et al 2015 discuss the impact
of positional errors on soft classification accuracy assessment.
Table 16.5: Difference between accuracy assessment of hard and soft
classification outputs.
Aspect Hard Classification Soft Classification
Probability distribution over multiple classes
Output Discrete class labels for each pixel
for each pixel
Error matrix (confusion matrix), overall Fuzzy confusion matrix, fuzzy accuracy
accuracy, Kappa coefficient measures, probabilistic metrics
Overall accuracy: proportion of correctly Fuzzy overall accuracy considering partial
classified pixels memberships
Producer’s accuracy: probability that a
Adjusted to account for partial memberships
reference pixel is correctly classified
Accuracy
metrics User’s accuracy: probability that a classified
Adjusted to account for partial memberships
pixel represents the true class
Kappa coefficient: measures agreement Fuzzy kappa coefficient adjusted for partial
beyond chance memberships
Used to assess uncertainty and robustness
Entropy-based measures: not typically used and also in understanding the distribution of
class memberships
Sampling Similar sampling methods, but with
Random or stratified sampling
design consideration for fuzzy memberships
Consideration Does not account for mixed pixels or Accounts for uncertainty and partial class
of uncertainty uncertainty memberships
Relatively simple and straightforward to More complex, requires specialized methods
Complexity
compute and interpret for computation and interpretation
Spatial Limited spatial context, focuses on individual Can include spatial accuracy measures,
context pixel accuracy considering spatial coherence

94
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
The assessment of classification accuracy for hard and soft classification
outputs involves different approaches and considerations due to the nature of
each classification type. In case of hard classification focuses on discrete class
assignments and uses traditional accuracy metrics whereas in case of soft
classification, it accounts for partial memberships and uses extended metrics to
capture the nuances of class distributions. Important aspects of the accuracy
assessment of both hard and soft classification outputs are given in Table 16.5.
In this section, you have learnt about accuracy assessment approaches
applicable to various kinds of classification outputs and object detection results.
However, there are other types of data that do not fall under these categories. In
the next section, you will be introduced to approaches of accuracy assessment
for other kinds of data or outputs.
Let us spend 5 minutes to check your progress.

SAQ III
a) Differentiate between the accuracy assessment of hard and soft
classification outputs?

b) List the approaches used for accuracy assessment of object based

classification outputs.

c) How is accuracy assessment of object detection outputs carried out?

16.6 APPROACHES TO ACCURACY ASSESSMENT

FOR OTHER DATA/ OUTPUTS
There are some other types of data that do not fall under the categories that
have not been covered in the previous section. In the next section, you will be
introduced to approaches of accuracy assessment for other kinds of data or
outputs such as change detection maps, elevation data, modelled outputs and
vector data.
Let us learn about these data and the applicable approaches under different
subsections.
16.6.1 Change Detection Outputs
You have learnt about change detection techniques in Unit 15 of this course.
Change detection involves identifying and analysing the changes that have
taken place in the landscape between the acquisitions of two or more temporal
images. You have seen that the outputs can be either raster or vector data.
Accuracy assessment of the change detection outputs may require a different
approach. In this case, assessment focuses on the accuracy of change
detection rather than individual classifications and it requires comparison of
change maps and validation of detected changes using ground truth or
reference data from different times. For the purpose, change matrix is a
common approach, which is similar to confusion matrix but focuses on detecting
and classifying changes over time. Some other quantitative metrics are used
that measure accuracy of detected changes, such as the number of correctly
identified change pixels, etc.
95
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
Direct comparison, post-classification comparison, temporal trajectory analysis,
multi-temporal composite analysis, object-based change detection are some of
the approaches used for the purpose.
Direct comparison directly compares pixel values from multi-temporal images or
change in vegetation cover using NDVI. However, radiometric normalisation is
required to ensure consistency. Post-classification comparison compares
classified images from different times to detect changes. Temporal trajectory
analysis analyses the temporal sequence of images to detect changes over time
using time-series analysis techniques e.g. Monitoring deforestation trends over
several years. Multi-temporal composite analysis combines multiple images into
a single composite to detect changes, e.g. detecting urban expansion using
multi-temporal composites. Object-based change detection uses objects rather
than pixels for change detection object change accuracy.
Let us now learn about the approaches used for elevation data.
16.6.2 Digital Elevation Models
There are various types of digital elevation models available. Accuracy
assessment of elevation data, such as the DEMs is essential for evaluating the
quality and reliability of the elevation information derived from various sources
like LiDAR, satellite imagery, and photogrammetry. Accuracy of a DEM is
generally assessed by comparing the elevation values in the DEM with
reference data, often derived from ground surveys or higher-accuracy data
sources. Visual inspection, point-to-point comparison, profile method, surface
comparison (area-based assessment), cross-validation, slope and aspect
comparison and contour line comparison are some of the methods used for the
purpose.
In visual inspection, DEM outputs are visually compared with high-resolution
imagery or reference maps, often involving expert judgement. Point-to-point
comparison method involves comparing the elevation values of the DEM with
precise ground control points (GCPs) collected from field surveys or high-
precision GPS measurements. In profile method, elevation values are compared
along linear transects or profiles (e.g., roads, rivers) that have been surveyed
using high-precision instruments. Surface comparison (area-based assessment)
compares entire DEM surfaces between the DEM and a reference DEM of
higher accuracy. In cross-validation, the DEM is divided into training and
validation sets, where portions of the data are used to assess the accuracy of
interpolated or modeled elevations. Slope and aspect comparison approach
compares derived slope/aspect maps with reference data to assess the
accuracy of terrain representation. And, contour line comparison method
compares DEM-generated contours with reference maps such as topographic
maps or survey data.
There are various accuracy assessment methods used for DEMs but the
choice of method depends on the type of terrain, data availability, resources,
and the intended application of the DEM. Among the metric used, RMSE is
the most commonly used, which provides a direct measure of elevation
accuracy. It measures the average distance between the predicted and
observed values. It tells you how concentrated the data is around the line of

96
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
best fit. However, other metrics like MAE, standard deviation, and error
distribution are also important for a holistic assessment. MAE is measured as
the average absolute difference between the predicted values and the actual
values. You must note that accuracy of the reference data (i.e. GCPs, high-
precision DEMs) significantly impacts the reliability of the assessment.
Let us now learn about the accuracy assessment approaches used for
modelled outputs.
16.6.3 Modelled Outputs
Accuracy assessment of modeled outputs in remote sensing involves evaluating
how well the predicted outputs (e.g., land cover maps, change detection results,
environmental models) match the reference or ground truth data. This process
is crucial for validating model performance and ensuring that outputs are reliable
for decision-making. Depending on the nature of the modelled data, the type of
output, and the complexity of the task various types of methods are used for the
purpose. Each method has limitations, such as sensitivity to specific data
characteristics, computational demands, or subjective biases, which must be
considered during assessment. Combining multiple methods often provides a
more comprehensive evaluation, especially when dealing with complex or
heterogeneous data.
Error matrix-based statistics, Root Mean Square Error (RMSE), Coefficient of
Determination [R-squared (R²)], Bias and Variance, Mean Absolute Error (MAE),
Correlation coefficients (e.g., Pearson’s r, Spearman’s Rank, ρ), Cohen’s Kappa
Coefficient, F1 Score and Precision-Recall analysis, NDVI metrics, Cross-
validation, Spatial Autocorrelation metrics (e.g., Moran’s I), and visual
interpretation are some of the methods used for the purpose.
Error matrix-based statistics extends confusion matrix with omission and
commission errors and is useful for thematic maps and categorical data. Root
Mean Square Error (RMSE) measures average squared difference between
predicted and observed values and is useful for continuous outputs such as
elevation models, temperature predictions, and biophysical variables.
Coefficient of Determination [R-squared (R²)] indicates the proportion of
variance in the dependent variable predictable from the independent variables
and is good for assessing the fit of linear regression models. Bias and variance
measures the error due to bias (systematic error) and variance (sensitivity to
fluctuations in the training set) and is good for evaluating trade-offs in model
complexity. Mean Absolute Error (MAE) measures average magnitude of errors
between predicted and observed values, regardless of direction and is suitable
for continuous data to see error between predicted and observed values or
pollution levels. Correlation coefficients (e.g., Pearson’s r, Spearman’s Rank, ρ)
measures the strength and direction of the linear (or rank-based) relationship
between predicted and observed values and is useful for continuous model
outputs, such as biophysical variables. Cohen’s Kappa Coefficient measures the
agreement between predicted and observed classes, adjusting for the
agreement that occurs by chance and is good for classification outputs. F1-
Score and Precision-Recall analysis combines precision (positive predictive
value) and recall (sensitivity) into a single metric for class performance and is
suitable for binary classification outputs and imbalanced classes. NDVI metrics
97
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
uses indices (like NDVI) to compare model outputs to reference data, assessing
performance in vegetation mapping and is useful for assessing models in
agriculture, forestry, and environmental monitoring.
Cross-validation splits the dataset into training and validation subsets multiple
times to assess model performance and variability and is useful for machine
learning models to ensure robust accuracy estimates. Spatial Autocorrelation
metrics (e.g., Moran’s I) measures the degree of spatial clustering of errors in
the model outputs and is ideal for spatial models, particularly when assessing
spatial bias. Visual interpretation method involves expert analysis and visual
comparison of model outputs with reference data and is useful when
quantitative assessment is challenging, such as with complex or novel outputs.
You will learn more about various types of spatial models and uncertainty
analysis in the course MGY-006.
Let us now learn about the approaches employed for accuracy assessment of
vector data.
16.5.5 Vector Data
Some of the data may not be raster data in that case some other approaches
are used. Let us now learn about the approaches of accuracy assessment used
for vector data.
Accuracy assessment of vector data involves evaluating positional, class and
temporal attributes, topological accuracies, etc. Depending on the nature of the
data and its characteristics and also the requirements of accuracy assessment
task various types of methods are used for the purpose. Commonly used
methods are presented in the Table 16.6.
Table 16.6: Comparison of the commonly used methods employed for
accuracy assessment of vector data outputs.

Method Description Metrics Used

Positional Evaluates the accuracy of the geographic coordinates of Root Mean Square Error
accuracy vector data, e.g. road networks, use of high-precision (RMSE), Mean Absolute
assessment GPS and correction techniques is recommended Error (MAE)

Measures the correctness of the attributes associated

Attribute
with vector features, e.g. validating land use Accuracy, Precision,
accuracy
classifications in a GIS database, cross-validation with Recall, F1-Score
assessment
reliable reference data is recommended

Topological Assesses the correctness of the spatial relationships

Topological errors (e.g.,
accuracy between vector features, e.g. checking for topological
overlaps, gaps)
assessment errors in cadastral maps

Thematic Confusion matrix,

Evaluates the accuracy of the classification of vector
accuracy overall accuracy,
features
assessment Kappa coefficient

Measures the accuracy of temporal attributes associated

Temporal
with vector data, e.g. validating the accuracy of time- Temporal accuracy,
accuracy
stamped events; cross-checking with reliable temporal time lag
assessment
reference data is recommended

98
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
These methods help in evaluating the accuracy of vector data by providing
various metrics and approaches to identify strengths and weaknesses.

16.7 RECENT TRENDS

You have learnt about the factors that affect accuracy assessment in the
subsection 16.4.2. Let us now recall the challenges to accuracy assessment.
16.7.1 Challenges
There are several challenges to accuracy assessment due to the complexity of
data, variations in environmental conditions, and methodological limitations.
Following are some of the key challenges:
 Availability and quality of reference data
 Mixed pixels and boundary issues
 Spatial and temporal variability
 Spectral similarity between classes
 Selection of appropriate accuracy metrics
 Handling uncertainty in soft classifications
 Automation and scalability issues
 Influence of sensor characteristics and data preprocessing
 Biases in sample design and data collection
These factors highlight the complex nature of accuracy assessment in remote
sensing.
Let us now briefly discuss the approaches used to improve accuracy.
16.7.2 Approaches to Improving Accuracy
Improving the classification accuracy of remote sensing images is crucial for
enhancing the reliability and applicability of the results in various fields, such as
environmental monitoring, land use planning, agriculture, and urban studies.
Here are some key strategies for improving the classification accuracy of remote
sensing images:
Employing suitable pre-processing techniques: use of appropriate
radiometric, geometric, and atmospheric corrections techniques reduces errors
caused by noise and inconsistencies in data and also enhances the visibility of
key features, aiding accurate classification.
Feature selection and extraction: selecting relevant features (bands, indices,
texture features, etc.) reduces data complexity and computational load and
enhances the model's ability to differentiate between classes thereby improving
the classification process by focusing on the most informative aspects of the
data.
Advanced classification algorithms: as compared to traditional methods
employing sophisticated algorithms such as support vector machine, Random
Forests, deep learning, OBIA approaches, etc. can significantly improve
classification accuracy as these are better able to handle complex, non-linear
relationships between features and classes, and adaptable to various data
types, from multispectral to hyperspectral.

99
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
Incorporating ancillary data: supplementing remote sensing images with
additional data sources, such as digital elevation models (DEMs), soil maps, or
land use data, temporal data, contextual information, etc. can improve
classification as they provide additional context that pure spectral information
may not capture, particularly in complex landscapes.
Data augmentation and synthesis: Data augmentation techniques artificially
increase the diversity of training data by creating variations thereby helping
improve robustness of classification models. It reduces overfitting by exposing
the model to a broader range of conditions and are particularly useful when
training data is limited.
Ensemble methods: combining multiple classifiers or models (ensemble
learning) can boost classification accuracy by leveraging the strengths of each
model. While Bagging (Bootstrap Aggregating) aggregates multiple models
trained on different subsets of the data to reduce variance and improve stability,
Boosting sequentially trains models to correct the errors of previous models
thereby enhancing accuracy. And, stacking combines outputs from different
models using a meta-classifier for final prediction. This approach increases
overall model robustness and reduces the impact of individual model
weaknesses.
Post-processing techniques: these refine the classified outputs to reduce
noise and improve spatial coherence. While smoothing filters reduces
classification noise and improve the visual quality of the map majority filters
corrects isolated misclassified pixels by replacing them with the majority class of
neighboring pixels. These techniques enhance the visual consistency and
accuracy of the final classified map and reduce salt-and-pepper noise often
seen in pixel-based classifications. And as you have learnt in subsection 16.4.2
with Fig. 16.1, contextual editing incorporating spatial rules (e.g., adjacency,
size constraints, location, association, etc.) can be used to refine and improve
classification results.
Active learning and interactive classification: it involves iteratively selecting
the most informative samples for labelling, which improves classification
accuracy by focusing on uncertain or misclassified areas. It involves expert input
to refine classification decisions interactively and focuses on additional sampling
efforts on areas with high classification uncertainty. It improves accuracy with
targeted data collection and also helps to fine-tune models based on expert
feedback.
Cross-validation and hyperparameter tuning: while cross-validation divides
the data into multiple subsets to ensure the model is evaluated on different
parts, and hyperparameter tuning optimises the model’s settings. For example,
K-fold cross-validation provides a more reliable accuracy estimate by using
multiple subsets of the data for training and validation and method such as grid
search/random search systematically explores different combinations of
hyperparameters to find the best model configuration. Such methods reduce the
risk of overfitting and improve model generalisation and ensures the selected
model parameters are optimal for classification.

100
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
Improving classification accuracy in remote sensing involves a combination of
data pre-processing, advanced algorithms, incorporation of ancillary data,
ensemble approaches, and careful tuning of model parameters. Applying these
methods helps to create more reliable, accurate classifications that better
represent real-world conditions.
16.7.3 Recent Trends and Developments
We have discussed about various approaches to accuracy assessment for
variety of raster and vector data in the previous two sections. We have also
learnt about various challenges and ways to improve accuracy in the previous
two subsections. Let us now learn about the recent trends and developments
taking place in this field.
There are several developments taking place in the field. These trends aimed at
enhancing the precision, reliability, and applicability of remote sensing
classifications reflect advancements in terms of data availability and
methodology. Morales-Barquero et al (2019) have analysed how accuracy
assessment practices have evolved over the past two decades and provide
valuable insights into the current state and future directions of accuracy
assessment in remote sensing, and emphasise on the need for more rigorous
and standardised practices.
Some of the recent trends and developments in the field are mentioned here:
 Integration of machine learning and deep learning to automate the process,
to identify and analyse classification errors more effectively.
 Incorporation of multi-source and multi-resolution data to improve
classification accuracy and assessment, and also for cross-sensor validation.
 Enhanced spatial and temporal accuracy assessment
 Crowdsourcing and citizen science to gather reference data and validate
remote sensing classification outputs.
 Development of new metrics and methodologies such as development of
fuzzy kappa coefficient and other fuzzy metrics that account for the
uncertainty in soft classifications, new methods such as spatial kappa and
object-based accuracy measures for evaluating spatial accuracy and
coherence of classifications, and even combining multiple metrics to provide
a more comprehensive assessment
 Integration of contextual and auxiliary information including demographic or
infrastructure information, to validate and improve classifications (e.g. using
demographic data to validate classifications of social or economic features).
 Standardisation and benchmarking: to ensure consistency and comparability,
and creating benchmark datasets for evaluating and comparing classification
algorithms and accuracy assessment methods.
Google Earth Engine has emerged as a platform for thematic information
extraction including image classification. Accuracy assessment of the classified
maps generated through this are also required to be evaluated. Recent trends in
accuracy assessment for remote sensing are driven by advancements in
technology, the integration of diverse data sources, and the development of new
methodologies. These trends aim to improve the precision, reliability, and

101
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
applicability of remote sensing classifications, providing more accurate and
actionable insights for a range of applications.
Let us spend 5 minutes to check your progress.

SAQ IV
a) List the approaches used for accuracy assessment of digital elevation
models.

b) List the approaches used for accuracy assessment of modeled outputs.

c) How is accuracy assessment of vector data carried out?

d) List any five recent trends and developments in accuracy assessment.

16.8 SUMMARY
Let us now summarise what you have studied in this unit:
 Accuracy assessment is an important step in remote sensing after
preparation of thematic maps. It determines the suitability of the maps for
further decision making and planning.
 There are several metrics used for assessing accuracy. The most common
are user’s, producer’s, overall and kappa coefficient. Besides, precision,
recall, F-1 score, r, R2, mAP, MAE, etc. are some other metrics.
 Assessing accuracy for each class as well as for the whole image is essential
to compare the results of various classification techniques and quality and
reliability of the results obtained.
 Error/confusion matrix can be used for accuracy and reliability assessments.
Overall accuracy is a measure of accuracy for the whole image across all
categories. Kappa coefficient is another method for accuracy assessment
having a number of advantages over other methods.
 Accuracy assessment has evolved from initial visual comparison to now
incorporation of machine learning approaches.
 Sampling size and pattern are important consideration for accuracy
assessment and sufficient number of samples should be taken for the same.
 The approaches for accuracy assessment vary greatly depending upon the
types of image classification methods employed, types of outputs, types of
data, etc.
 There are several challenges to accuracy assessment and suitable measures
can be employed to improve accuracy.

16.9 TERMINAL QUESTIONS

1. What are the various metrics used in accuracy assessment?

2. Write about the major milestones in history of accuracy assessment in

remote sensing.

102
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
3. Discuss the approaches of accuracy assessment for various types of image
classification outputs.
4. Describe the approach of accuracy assessment of change detection outputs.

5. Discuss the various approaches to improving accuracy.

16.10 REFERENCES
 Abdullah, Q. A. (2023) Photogrammetric Engineering & Remote Sensing
Vol. 89, No. 10, October 2023, pp. 581-588.
https://www.asprs.org/wpcontent/uploads/2024/03/October2023_HLA-
Positional_Accuracy_Standards.pdf
 ASPRS (2014) New ASPRS positional accuracy standards for digital
geospatial data (2014) by Smith, D. L., Abdullah, Q. A., Maune, D. &
Heidemann, K. H.
 ASPRS (2023) ASPRS Positional Accuracy Standards for Digital Geospatial
Data, 2nd Ed., Version 1.0, August 23, 2023.
https://publicdocuments.asprs.org/PositionalAccuracyStd-Ed2-V1
 ASPRS (2024) ASPRS Positional Accuracy Standards for Digital Geospatial
Data, 2nd Ed., Version 2. https://www.asprs.org/revisions-to-the-asprs-
positional-accuracy-standards-for-geospatial-data-2023
 Chavez Jr, P. S. (1988) An improved dark-object subtraction technique for
atmospheric scattering correction of multispectral data. Remote sensing of
environment, 24(3), 459-479.
 Chavez, P. S. (1996) Image-based atmospheric corrections-revisited and
improved. Photogrammetric engineering and remote sensing, 62(9), 1025-
1035.
 Cohen, J. (1960) A coefficient of agreement for nominal scales. Educational
and Psychological Measurement, 20, pp. 37-46.
 Congalton, R. G. & Green, K. (2019) Assessing the Accuracy of Remotely
Sensed Data: Principles and Practices, 3rd Edition, CRC Press.
 Congalton, R. G. (1991) A review of assessing the accuracy of
classifications of remotely sensed data, Remote Sensing of Environment,
Vol 37, pp. 35-46.
 Congalton, R. G., & Green, K. (2008) Assessing the Accuracy of Remotely
Sensed Data: Principles and Practices. CRC Press.
 Deshmukh, B., Bahuguna, A., Nayak, S., Dhargalkar, V. K. & Jagtap, T. G.
(2005) Eco-geomorphological zonation of the Bangaram reef,
Lakshadweep. Journal of the Indian Society of Remote Sensing, 33, 99-106.
 Foody, G. M. (2002). Status of land cover classification accuracy
assessment. Remote sensing of environment, 80(1), 185-201.
 Foody, G. M. (2020) Explaining the unsuitability of the kappa coefficient in
the assessment and comparison of the accuracy of thematic maps obtained
by image classification. Remote sensing of environment, 239, 111630.
 Foody, G. M. & Cox, D. P. (1994) Sub-pixel land cover composition
estimation using a linear mixture model and fuzzy membership functions.
Remote sensing, 15(3), 619-631.
 Gu, J., Congalton, R. G. & Pan, Y. (2015) The impact of positional errors on
soft classification accuracy assessment: A simulation analysis. Remote
Sensing, 7(1), 579-599.
103
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….……………………………………….
 https://deepchecks.com/how-to-check-the-accuracy-of-your-machine-
learning-model/
 https://medium.com/@prathameshamrutkar3/the-complete-guide-to-object-
detection-evaluation-metrics-from-iou-to-map-and-more-1a23c0ea3c9d
 https://pressbooks.lib.vt.edu/remotesensing/chapter/chapter-25-accuracy-
assessment/
 https://pro.arcgis.com/en/pro-app/latest/tool-reference/image-analyst/how-
compute-accuracy-for-object-detection-works.htm
 https://storymaps.arcgis.com/stories/fddf80b8900f4168addbfefec336f4bd
 https://www.ridgerun.ai/post/mean-average-precision-map-and-other-object-
detection-metrics
 Landis, J. R. & Koch, G. G. (1977) The measurement of observer
agreement for categorical data. Biometrics, 33, 159-174.
 Liu, J. & Mason, P. J. (2009) Essential Image Processing and GIS for
Remote Sensing. John Wiley & Sons.
 Lowell, K. & Jaton, A. (1999) Spatial Accuracy Assessment: Land
Information Uncertainty in Natural Resources, Ann Arbor Press, Michigan.
 Lu, D. & Weng, Q. (2007) A survey of image classification methods and
techniques for improving classification performance. International journal of
Remote sensing, 28(5), 823-870.
 Lunetta, R. S. & Lyon, J. G. (2004) Remote sensing and GIS accuracy
assessment. CRC press.
 Lunetta, R., Congalton, R., Fenstermaker, L., Jensen, J., Mcgwire, K. &
Tinney, L. R. (1991) Remote sensing and geographic information system
data integration- Error sources and research issues. Photogrammetric
engineering and remote sensing, 57(6), 677-687.
 Mather, P. M. & Tso, B. (2009) Classification Methods for Remotely Sensed
Data. CRC Press.
 Morales-Barquero, L., Lyons, M. B., Phinn, S. R. & Roelfsema, C. M. (2019)
Trends in remote sensing accuracy assessment approaches in the context
of natural resources. Remote sensing, 11(19), 2305.
 Mustak, S. (2013) Correction of atmospheric haze in Resourcesat-1 Liss-4
Mx Data for urban analysis: an improved dark object subtraction approach.
The International Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, 40, 283-287.
 Schott, J. R., Salvaggio, C. & Volchok, W. J. (1988) Radiometric scene
normalization using pseudoinvariant features. Remote sensing of
Environment, 26(1), 1-16.
 Stehman, S. V. & Czaplewski, R. L. (1998) Design and analysis for thematic
map accuracy assessment: fundamental principles. Remote sensing of
environment, 64(3), 331-344.
 Story, M. & Congalton, R. G. (1986) Accuracy assessment: a user’s
perspective. Photogrammetric Engineering and remote sensing, 52(3), pp.
397-399.
16.11 FURTHER/SUGGESTED READINGS
 Chapter 25: Accuracy Assessment – Remote Sensing with ArcGIS Pro
(second edition) (vt.edu)

104
Contributor: Prof. Benidhar Deshmukh
 Unit 16 Accuracy Assessment of Thematic Maps
…………………………….……………………………………...………………………………………..
https://pressbooks.lib.vt.edu/remotesensing/chapter/chapter-25-accuracy-
assessment/#sdfootnote1anc
 Foody, G. M. (2020) Explaining the unsuitability of the kappa coefficient in
the assessment and comparison of the accuracy of thematic maps obtained
by image classification. Remote sensing of environment, 239, 111630.
 Hand, D. J. (2012) Assessing the performance of classification methods,
International Statistical Review, 80 (3), 400-414.
 https://doc.arcgis.com/en/imagery/workflows/resources/accuracy-
assessment-of-orthomosaics.htm
 https://openmrv.org/web/guest/w/modules/mrv/modules_3/sampling-design-
for-estimation-of-area-and-map-accuracy
 https://pro.arcgis.com/en/pro-app/latest/tool-reference/spatial-
analyst/generalization-of-classified-raster-imagery.htm
 Janssen, L. L. F. and van der Wel, F. J. M. (1994) Accuracy assessment of
satellite derived land-cover data: a review, Photogrammetric Engineering and
Remote Sensing, Vol 60, pp 419-426.
 Nicolau, A.P., Dyson, K., Saah, D., & Clinton, N. (2024). Accuracy
Assessment: Quantifying Classification Quality. In: Cardille, J.A., Crowley,
M.A., Saah, D., Clinton, N.E. (Eds) Cloud-Based Remote Sensing with
Google Earth Engine. Springer, Cham. https://doi.org/10.1007/978-3-031-
26588-4_7
 Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., &
Wulder, M. A. (2014) Good practices for estimating area and assessing
accuracy of land change. Remote sensing of Environment, 148, 42-57.
https://doi.org/10.1016/j.rse.2014.02.015
 Stehman, S. V. & Foody, G. M. (2019) Key issues in rigorous accuracy
assessment of land cover products. Remote Sens Environ
231:111199. https://doi.org/10.1016/j.rse.2019.05.018
 Stehman, S. V. (2009) Sampling designs for accuracy assessment of land
cover. International Journal of Remote Sensing, 30(20), 5243–5272.
https://doi.org/10.1080/01431160903131000
16.12 ANSWERS
SAQ I
a) There are two types: qualitative and quantitative accuracy. Define the two
and also the site-specific and non-site specific accuracy as mentioned in
subsection 16.2.2.
b) There have been several developments in the standardisation and
methodological refinement era. Include the milestones in your answer as
discussed in section 16.3.
c) Major steps in accuracy assessment are discussed in subsection 16.2.4.
Refer to it to write your answer.
SAQ II
a) There are several sources of errors in image classification as listed in a
table under subsection 16.4.1. Refer to it to prepare your answer.
b) Discuss the point as given under subsection 16.4.1.

105
Contributor: Prof. Benidhar Deshmukh
Block 4 Image Classification and Change Detection Techniques
…………………………………………………………………….………………………………………
c) There are several methods for determining sample size in the context of
accuracy assessment as discussed under subsection 16.4.4. Refer to it
to write your answer.
SAQ III
a) Several differences are discussed in subsection 16.5.2. Refer to it to
write your answer.
b) Several approaches are discussed in subsection 16.5.1. Refer to it to
write your answer for accuracy assessment of object based classification
outputs.
c) There is a basic difference in how object detection outputs are assessed
for their accuracy. Refer to subsection 16.5.1 for preparing your answer.
SAQ IV
a) Some approaches used for accuracy assessment of digital elevation
models are mentioned under subsection 16.6.2. Prepare your answer
from it.
b) Please refer to subsection 16.6.3.
c) Please refer to subsection 16.6.4.
d) Please refer to subsection 16.7.3.
Terminal Questions
1. Please refer to subsection 16.2.3.
2. Please refer to section 16.3.
3. Please refer to section 16.5.
4. Please refer to subsection 16.6.1.
5. Please refer to subsection 16.7.2.

106 Contributor: Prof. Benidhar Deshmukh

 GLOSSARY
The technique of correcting images collected by satellite
Atmospheric
: or airborne sensors for atmospheric effects such as
correction
scattering and absorption.
Average Precision
: It is weighted mean of precisions at each threshold.
(AP)
It is a process of identifying, describing, and quantifying
variations between images of the same area taken at
Change detection :
different times or under various circumstances through a
technique.
A change vector of a pixel is the vector difference
Change vector : between the multi-band digital vector of the pixel on two
different dates.
A classification problem's prediction outcomes are
compiled in a confusion matrix. Count values are used to
describe the number of accurate and inaccurate
Confusion matrix :
predictions for each class. This is the confusion matrix's
secret. The confusion matrix demonstrates how your
classification model might be improved.
Contextual In this type of classification, spatially neighbouring pixel
:
classification information is used in image classification.

Correlation coefficients (e.g., Pearson’s r, Spearman’s

Rank, ρ) measures the strength and direction of the
Correlation
: linear (or rank-based) relationship between predicted and
coefficient
observed values and is useful for continuous model
outputs, such as biophysical variables.

Cross-validation splits the dataset into training and

validation subsets multiple times to assess model
Cross-validation :
performance and variability and is useful for machine
learning models to ensure robust accuracy estimates.

It is a subset of machine learning methods based on

Deep learning :
neural networks with representation learning.

It is the harmonic mean of precision and recall. Here,

True Positive (TP) refers to correct classification of the
object, False Positive (FP)-incorrect classification of the
F1-score : object, True Negative(TN) - correct classification of the
object as not being that class, and False Negative (FN) -
incorrect classification of the object as not being that
class.

107
 It is a type of clustering algorithm widely used for soft
classification. It allows each data point to belong to
multiple clusters with varying degrees of membership.
This approach is particularly useful where land cover
Fuzzy C-Means types often have mixed characteristics, making it difficult
:
(FCM) clustering to classify them into discrete categories. It assigns a
membership value to each data point for all clusters.
These membership values range between 0 and 1, with
the sum of membership values for each data point equal
to 1.
Hard / crisp An image classification process in which each pixel or
: object is assigned to a single, discrete class.
classification
It is a commonly used technique for unsupervised
classification of remote sensing images. It groups pixels
into clusters based on their attributes (i.e. spectral
Hierarchical
: similarity) without prior knowledge of the class labels.
clustering
Hierarchical clustering organises data into a tree-like
structure called a dendrogram, which represents the
nested grouping of pixels or objects and their similarities.
It is a process through which pixels in the image is
grouped into various informational classes/ objects based
Image classification : on their spectral signatures or reflectance properties.
There are broadly two approaches of image classification
i.e. unsupervised and supervised.
ISODATA (Iterative Self-Organising Data Analysis
Technique) algorithm is an unsupervised clustering
method designed to partition a dataset into a predefined
number of clusters based on the data's inherent
ISODATA clustering : characteristics. Developed by Stuart Lloyd in 1982,
ISODATA draws inspiration from the K-means clustering
algorithm but introduces adaptive mechanisms to handle
varying cluster shapes and sizes during the clustering
process.

It is a discrete multivariate technique used to assess

classification accuracy from an error matrix. Kappa
analysis generates a kappa coefficient or Khat statistics,
the value of which ranges between 0 and 1. Through
Kappa statistics : Kappa statistics we know how well the classification
performed in comparison to randomly assigning the pixels
to a specific class. Kappa coefficient (Khat) is a measure
of the agreement between two maps taking into account
all elements of error matrix. It is a statistical measure that
compares the observed accuracy with an expected

108
 accuracy (random chance). In this off-diagonal elements
of an error matrix are incorporated as a product of the row
and column marginal totals.

It is one of the most widely used unsupervised

classification algorithms. It partitions the input image into
k-means clustering :
`k` clusters based on the nearest mean value of pixel
intensities in a multi-dimensional spectral space.
It is a field of artificial intelligence that develops and
Machine learning : studies algorithms that can learn from data and generalise
to unseen data.
One of the commonly used techniques for supervised
image classification is MXL. It estimates the chance that a
Maximum likelihood
: given pixel belongs to a certain class based on the
Classification (MXL)
assumption that the statistics for each class in each band
are normally distributed.

It measures average magnitude of errors between

Mean Absolute predicted and observed values, regardless of direction
:
Error (MAE) and is suitable for continuous data to see error between
predicted and observed values or pollution levels.

Mean Average
: mAP is the average of AP of each class.
Precision (mAP)
Mixel pixel : In a remote sensing image, in transition zones or at the
edges of different land cover classes you see a pixel
covering more than one land cover class because spatial
resolution of the image i.e. the area covered by a pixel is
larger than the individual objects on the ground. Such
pixels are called mixed pixels and its signature is a
combination of the land cover classes covered in it. Mixed
pixels pose challenges in the analysis of remote sensing
data because its spectral signature does not correspond
to any single land cover type hence resulting in
misclassification.

A principal component analysis where two or more

images from the same location taken at different times are
Multi-date Principal
superimposed and analysed as a single image. The local
Component :
alterations are revealed by the minor component images,
Analysis
while the major component images display the albedo
(reflectance) and radiometric variances (minor changes).
Multitemporal Images captured on different time periods.
:
image
Non-parametric The type of classifier does not assume any specific
:
classifier statistical distribution of data.
OBIA is a technique for analysing remote sensing data
that segments an image into meaningful objects and
Object-based image
: classifies them based on spectral, shape, and contextual
analysis (OBIA)
data. In this classification is conducted based on the
objects, instead of an individual pixel
It is the proportion of correctly classified pixels (both true
positives and true negatives) out of the total number of
pixels. It is the commonly cited measure of mapping
accuracy which is the number of correctly classified pixels
Overall Accuracy : (sum of major diagonal cells in the error matrix) divided by
total number of pixels checked. Though, overall accuracy
is a measure of accuracy for the entire image across all
classes, it ignores off-diagonal elements (i.e. errors of
omission and commission).
The type of classifier that assumes that input data follows
Parametric a known statistical distribution (e.g., Gaussian) and uses
:
classifier parameters (e.g. mean vector and covariance matrix)
generated from training samples
An image classification process in which each pixel is
Per-Pixel
: classified individually based on its spectral properties
classification
without considering spatial context.

It is the ratio of true positive detections to all positive

predictions (true positives + false positives). True Positive
Precision : (TP) refers to correct classification of the object, and
False Positive (FP) refers to incorrect classification of the
object.

It measures how well the classification process identifies

pixels of a particular class. It is a probability that a
Producer’s reference pixel has been correctly classified and shows
accuracy (PA) or : what percentage of a particular ground class was correctly
Omission error classified. It is calculated by dividing the number of
correctly classified pixels of a class by the total number of
reference pixels of that class.

It is the ratio of true positive detections to all actual

positives (true positives + false negatives). True Positive
Recall : (TP) refers to correct classification of the object, and
False Negative (FN) refers to incorrect classification of
the object as not being that class.

ROC curve : ROC curve is created by plotting the true positive rate

110
 (TPR) against the false positive rate (FPR) at various
threshold settings.
(Where, TPR is the proportion of positive data points
correctly considered as positive, with respect to all
positive data points; and FPR is the proportion of negative
data points that are mistakenly considered as positive,
with respect to all negative data points).

The process of checking of spectral signatures for their

Signature representativeness of the class they attempt to describe
:
Evaluation and also to ensure as small spectral overlap between
signatures of different classes as possible.
An image classification process which provides
Soft / fuzzy
: probabilistic or fuzzy outputs, assigning a pixel multiple
classification
classes with varying probabilities
Spatial Autocorrelation metrics (e.g., Moran’s I) measures
Spatial
the degree of spatial clustering of errors in the model
Autocorrelation :
outputs and is ideal for spatial models, particularly when
metrics
assessing spatial bias.

In this type of classification, pure spectral information is

used in image classification. A ‘noisy’ classification result
Spectral
: is often produced due to the high variation in the spatial
classification
distribution of the same class e.g. Maximum likelihood,
minimum distance, artificial neural network
In this type of classification, spectral and spatial
Spectral
information is used in classification; parametric or non-
contextual
: parametric classifiers are used to generate initial
(/spectral-spatial)
classification images and then contextual classifiers are
classification
implemented in the classified images.
An image classification process in which spectral value of
each pixel is assumed to be a linear or non-linear
Sub-pixel
: combination of defined pure materials (or endmembers),
classification
providing proportional membership of each pixel to each
endmember.
It is the process of identification of classes within a
Supervised
: remotely sensed image with inputs from and as directed
Classification
by the user in the form of training data.
It is the process of automatic identification of natural
Unsupervised groups or structures within a remotely sensed image. In
:
classification this user input is minimum and the process is guided by
the spectral similarity of the objects present in the image.
User’s or It measures the reliability of a classification result,
:
Consumer’s indicating the likelihood that a pixel classified into a given
accuracy (CA) or class actually represents that class on the ground. It is the
Commission error probability that the class of a pixel actually represents that
same class on the ground. It is a measure of the reliability
of a map generated from a classification.

112