Solar Vision: Harnessing Automated Detection and

Quantification of Solar Panels in Urban Areas promoting

Renewable Energy Adoption achieving Sustainable

Development Goals

Vishu Kalier

Virat Srivastava

Durgesh Kumar Singh

National Remote Sensing Center, Hyderabad

Dr. Jaya Saxena, Head of Student Project Interface Division,

National Remote Sensing Center, Hyderabad

Acknowledgement

We would like to extend my heartfelt gratitude to Dr. Jaya Saxena, Head of the Student Interface Division,
National Remote Sensing Center, for her unwavering support and invaluable guidance throughout the course of
this project, Solar Vision : Harnessing Automated Detection and Quantification of Solar Panels in Urban Areas
promoting Renewable Energy Adoption achieving Sustainable Development Program. Her expertise and
leadership have greatly contributed to the success of this project. Her insightful feedback and encouragement
have consistently driven me to push beyond my limits, ensuring the highest level of quality in my work.
Thank you, Dr. Jaya Saxena, for your dedication and for fostering an environment of growth and
learning. Your contributions have been instrumental to my journey.

Sincerely,
Durgesh Kumar Singh
Virat Srivastava
Vishu Kalier
Solar Vision - NRSC ISRO, Hyderabad

Abstract (for professional papers)

Solar Energy production is one of the most important Sustainable Development Goals launched by the
Government of India, following United Nations Sustainable goal development 7 - Ensure access to affordable,
reliable, sustainable and modern energy for all. This initiative aims to improve and expand the solar energy
production throughout the country - urban power generation. This project aims to detect solar panel arrays and
the number of solar panels and the solar energy produced by the panels [METHOD OF POWER
CALCULATION TO BE ADDED]. The project is partitioned into three major parts. The first task uses the
extracted data from Google Earth Engine containing images of solar panels of varying sizes and orientations.
Second phase determines and detects the solar panels and then evaluates the number of solar panels contained in
the processed image. Thirdly, the number of solar panels counted are used to calculate the solar energy
produced by the solar panels in the corresponding image.

Keywords: Solar Vision, Urban Areas, Solar Panel Detection, Sustainable Development Goals
Solar Vision - NRSC ISRO, Hyderabad

1. Introduction

Solar Energy is one of the most usable and readily available renewable energy and currently pan India is
developing methodologies to extract the maximum solar energy and is drifting gradually from non-renewable to
renewable solar energy. For the above-mentioned reason, India has developed solar parks across multiple rural
areas across metropolitan cities like Hyderabad, Delhi, and Pune. Nowadays, the generation of electricity from
solar energy has reached urban areas as well, for which many houses have attached solar panels of varying
shapes and sizes on their rooftops. Thus, it is essential and crucial at the same time to evaluate the amount of
solar energy produced from rooftops since they are entirely distributed, non-homogeneous in density and vary
in great aspects from city to city and state to state.

The project uses deep learning and neural network architectures to detect solar panels in the images. The
images chosen are of wide variety and contain both individual solar panels and arrays of solar panels, in which
the solar panels are classified against the rooftops, the blue sheds and the shadows present in the images of the
urban areas. It is a fast-forward approach which reduces the manual labor of processing data.

(a)
(b) (c)

(d)
(e) (f)

Figure 1: Dataset diversity for solar panels- (a) clustered solar panels (b) scattered solar panels
(c) null (no solar panels) (d) arrays of solar panels (e) tilted solar panels (f) single grid of solar panels
Solar Vision - NRSC ISRO, Hyderabad

2. Objective

The project performs operations towards completion of multiple goals aligned towards Sustainable
Development Goals. The project uses Convolutional Neural Networks for detection and evaluation of solar
panels. The objectives encapsulated under the project are defined below-

i. Accurate Detection of Solar Panels – The main or primary objective of the project is to extract the
regions where the solar panels or arrays of solar panels are present in the image, if there are any.

ii. Quantification of Solar Panels present in the Image – The other parameter of the project which it
targets aligned with the Sustainable Development Goal is to provide the exact number of solar panels
present within the image.

iii. Evaluation of Output Solar Energy – The other objective of heightened significance is to evaluate the
total energy produced by the solar panels provided in the image collectively.
Solar Vision - NRSC ISRO, Hyderabad

3. Related Work

There are many models and Machine Learning Architectures defined to detect and identify solar panels in
rural areas or areas with sparse populations like farmlands, parks and small vegetation. These works are mostly
inclined towards agricultural land and rural areas. In India, the estimation of generation of solar energy by solar
parks is done on a very large scale, yet, it has not reached the urban areas and the metropolitan cities. Most of
the solar parks are situated on the outskirts of the cities and are comparatively easier to detect and define than
the vaguely oriented solar panels on the rooftops of houses. This project revolves around the similar technique
of estimation of solar panels and arrays of solar panels and thus, extends itself to the rural areas and
metropolitan cities.

Microsoft used Convolutional neural Network for detection of change in temporal data in solar parks in
India including cities like Chennai, Pune, Delhi and Bangalore. The approach integrates the use of Neural
Networks and Artificial Intelligence to encompass the presence of solar parks in an Image. This project
stretches the domain of the above work a few steps ahead to detect, quantify and estimate the solar panels and
their energy production in urban metropolitan areas respectively.
Solar Vision - NRSC ISRO, Hyderabad

4. Approach

The Project uses multiple Convolutional Neural Network models and before using them perform a wide
variety of data cleaning and data preprocessing on the images. These processed images are then masked and
object bound to ascertain or highlight the areas which contain the arrays of solar panels.

4.1 Dataset

The dataset is extracted via a methodology of web scraping from Google Earth Engine’s application
programming interface. This interface is used to cumulatively extract multiple images on the basis of certain
parameters like location and city name. The cities from which the dataset is developed are Mumbai, Pune,
Hyderabad, Delhi, Chennai, Lucknow, Chandigarh, Ahmedabad, Rajkot, Kolkata and Los Angeles, New York,
California, Washington DC, Florida, Miami, and Denver. The data is partitioned across medium resolution of
different scale and high resolution of same scale of 640 x 640, 1280 x 1280, 1280 x 1280 (black-edge fit), 640 x
640 (resized), 1920 x 1920 (black-edge fit), enumerating to 297 pan India and 300 foreign images which are
later augmented.

Figure 2 : Dataset Classification (public, private previliged)

4.2 Data Preprocessing

The dataset is preprocessed using roboflow and certain methodologies namely annotation and bounding
box is applied to extract the regions in the image containing the arrays of solar panels and/or solar panels. This
data processing is a crucial step to enhance the model performance to a great extent. All these data pre-
processing techniques are used in the format of a data pipeline where the workflow behaves in asynchronous
Solar Vision - NRSC ISRO, Hyderabad

manner to reduce the resource utilization and time.

(a) (b)

Figure 3.a : Image contains building rooftop with solar panel array and surrounding foliage

Figure 3.b : Image contains building rooftop with solar panel array with bound box mask.(class: solar-panel)

4.3 Data Augmentation

Data Augmentation is applied on both the datasets having medium and high resolution image datasets.
Augmentation involves both clockwise and counter-clockwise. This augmented data is then fed to the Model
and the neural network for training of the data and detection of regions of solar panels and array of solar panels.

S. No. Location Data Augmented

1. India 300 900

2. Foreign 300 900

Table 1 : Table representing the total number of Images augmented, both india and foreign

4.5 Model Training

The neural network architecture is used for training of data to decipher the regions containing arrays of
solar panels. The Neural Network learns and adapts its weights as per the masking done on the images. It
quickly learns the different orientations of solar panels and then finds out the regions which are not solar panels.
There are multiple models which are used to detect and identify the solar panels which are described below.

4.5.1 You Only Look Once (YOLO) v8

This is the Convolutional Neural Network architecture which was previously used in many related works to
detect the solar parks in the outskirts of the cities or the solar panels in the farmlands or sparse population
zones. Similarly, the same model is trained to extract the regions and defined the bound box across the arrays of
solar panels in the augmented dataset.

Figure 4: You Only Look Once (YOLO) v8 Architecture Diagram

This YOLO v8 model has 24 convolution and multiple padding layers and marks or bounds the arrays of solar
panels relatively well as compared to the other object detection algorithms. This algorithm is quick and linearly
propagating due to the fact that it processed any pixel in the image only once. As compared to other models like
AlexNet, ResNet and VGG16, this model is a faster approach and is currently being used for single label Object
Detection and multiple objects Image Segmentation. The bound boxes generated are parallel to the X-axis of the
Images and are both rectangular and square in shape differing from image to image.

Figure 5: Solar Panels Detection and formation of bounding box using YOLO v8
Solar Vision - NRSC ISRO, Hyderabad

4.5.2 YOLO v8 Oriented Bound Box (OBB)

This is a slight modification of the original version of YOLO v8, where the bounding boxes were parallel to the
X-axis. This model improves the efficiency of original YOLO v8 by integrating 344 layers as the bounding
boxes are formed for the tilted orientation of the arrays of solar panels. It uses Artificial Intelligence and
Machine Learning to detect, learn and estimate the orientation of the solar panels and then slowly craves out the
bounding box.

Figure 6: Solar Panel Detection and formation of bounding box using YOLO v8 OBB

4.5.3 COCO CNN Models

The COCO Models are an application of Convolutional Neural Networks on the COCO Dataset. They are
Recurrent Convolutional Neural Networks (RCNN), You Only Look Once (YOLO) and Mask Recurrent
Convolutional Neural Network (Mask R-CNN) that are trained and evaluated on the COCO Dataset. These
models are widely used in Object Detection and Image Segmentation and Panoptic Segmentation. This model is
used to classify and detect the domains in the image which contains the arrays of solar panels.

Figure 7 : Solar Panel detection using COCO CNN base model

4.5.4 Image Segmentation

Image Segmentation is used to emphasize the areas containing the solar panels. This technique is fruitful
in highlighting the bound boxes containing the solar panels allowing reduction of noise from the original data as
well as the augmented data. This leads to finer classification by the model as there are only areas containing the
solar panels and other parts are eradicated from the Image leading to Image segmentation.

Figure 8 : Black masked image formed by Image Segmentation of Solar Panels

4.5.5 Edge Detection

There are multiple edge detection techniques like Canny edge detection, Harris Corner and Edge
Detection, but among them the most suitable for the project was Holistically Nested Edge Detection Algorithm
(HED). This algorithm defined the outlines in the image in such a way that it gives the outlines a varying
thickness on the basis of the depth in the image due to which it was preferred over Canny Edge Detection. This
algorithm performs well in identifying the solar panels as the number of closed boundary objects in the masked
black image.

Figure : to be added
Solar Vision - NRSC ISRO, Hyderabad

5. Results

This project validates the proof of concept of detection of solar panels or the arrays of solar panels. For this, a
bound box model was trained initially on roboflow to check the feasibility and manageability of the Object
Detection and Image Segmentation. {} [MODEL 2 - BOUND BOX DETECTION - 1280]

5.1 Proof of Concept

The COCO Model was used as a proof of concept to detect and decipher the arrays of solar panels or solar
panels. The model is a test on the proof of concept to check for the validation.

Layers 50

Parameters 10,72,346

Gradients 10,72,568

GPU T4

Epochs 300

Time 14 minutes

Table 1 : Parameter Summary Table for Proof of Concept

(a) (b)

Figure 9 : Statistics for the Proof of Concept a) Class loss plot b) Object loss plot
Solar Vision - NRSC ISRO, Hyderabad

5.2 You Only Look Once (YOLO) v8 OBB Model - Size Medium (M)

This model was used to detect and segment out the solar panels or the arrays of solar panels which were present
in different orientations in the augmented image. The summary of the model is attached herewith the output
bounded box image encapsulating the solar panels.

Layers 344

Parameters 23,782,294

Gradients 23,782,278

GPU 4th Iteration on T4

Epochs 100

Time 2.657 hours

Table 2 : Parameter Summary Table for YOLO v8 OBB - Size M

Figure 10 : Statistics for YOLO v8 OBB Model giving the evaluation metrics

5.3 You Only Look Once (YOLO) v8 OBB Model - Size (N)

This model was used to detect and segment out the solar panels or the arrays of solar panels which were
present in different orientations in the augmented image. The summary of the model is attached herewith the
output bounded box image encapsulating the solar panels.
Solar Vision - NRSC ISRO, Hyderabad

Layers 128

Parameters 2,000,000

Gradients 0

GPU 2th Iteration on T4

Epochs 100

Time 30 minutes

Table 3 : Parameters Summary Table for YOLO v8 Model- Size N

Figure 11 : Statistics during Model Training of YOLO v8 OBB Model of size N

6. Future Scope

This project aligns with the sustainable development goals and is a step to detect solar panels across
PAN India in urban areas. It is a step forward to achieve sustainable development goals as ascertained by the
government of India. This project can be used to gather the information about the total energy produced by
urban areas which can significantly optimize the energy production by the electricity department of individual
cities.
Solar Vision - NRSC ISRO, Hyderabad

7. References