See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/280918524

Sentinel-2 Images and Finnish Corine Land Cover Classification

Conference Paper · April 2012

CITATIONS READS

2 304

4 authors, including:

Markus Törmä Pekka Härmä

Finnish Environment Institute Finnish Environment Institute
53 PUBLICATIONS   182 CITATIONS    27 PUBLICATIONS   279 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Markus Törmä on 12 August 2015.

The user has requested enhancement of the downloaded file.

 Sentinel-2 Images and Finnish Corine Land Cover Classification
Markus Törmä, Suvi Hatunen, Pekka Härmä, Elise Järvenpää

Finnish Environment Institute SYKE, PO.Box 140, 00251 Helsinki, Finland, markus.torma@ymparisto.fi

ABSTRACT As compared to the earlier version of CLC database, the

updated version was more time-consistent, more
This paper presents the main principles of Finnish accurate and costs were lower [3]. The main outputs of
Corine Land Cover classification process, the the IMAGE2000 and CLC2000 projects at European
characteristics of the used images, experiences level were [4]:
concerning image processing and expectations
concerning future Sentinel-2 MultiSpectral Instrument. 1. national and European wide satellite image mosaic
It is expected that Sentinel-2 MSI will provide images for the year 2000 (IMAGE2000),
with better quality and temporal resolution, and which 2. an updated national and European CORINE land
will make the processing of future IMAGE-mosaics cover classification for 2000 (CLC2000), and
easier and provide better Corine Land Cover 3. database of land cover changes between 1990 and
classifications due to better spatial, spectral and 2000 at national and European levels.
radiometric resolutions.
EEA endorsed a proposal in summer 2005 to update
1. INTRODUCTION CLC data together with high resolution land cover data
as part of the implementation of the GMES fast track
The European Commission introduced the Corine service on land monitoring. This service would provide
Programme in 1985 in order to gather information about on a regular basis core land cover and use change data
the environment of the European Union. In order to that can be used by a wide range of downstream
determine and assess the effects of Community’s services at European, national, regional and local level
environment policy, it is needed to have a proper [5]. This GMES land monitoring core service delivered
understanding concerning the different features of the the following products:
environment [1]. Corine Land Cover (CLC)
classification is produced using satellite images. The 1. Orthorectified satellite images for the reference
mapping scale is 1:100 000 and mapping accuracy is at year 2006 (+/1 year),
least 100 m. The minimum mapping unit is 25 hectares 2. European mosaic based on satellite imagery called
and minimum width of units is 100 m. Only area IMAGE2006,
elements are classified. The classification nomenclature 3. Corine land cover changes 2000 – 2006,
is hierarchical and contains five classes at the first level, 4. Corine land cover classification 2006,
15 classes at the second level and 44 classes at the third 5. High resolution core land cover data for built-up
level [1,2]. There can also be national level-4 classes. areas including degree of soil sealing for year 2006,
Originally, the CLC classification was performed as and
visual interpretation of hardcopy printout of satellite 6. High resolution core land cover data for forest areas
images by overlaying a transparency on the printout, for year 2006.
drawing polygons to transparency and digitizing drawn
polygons [2]. So far, the Corine Land Cover classifications have been
project based work organized by European Environment
In order to update the CLC data European Environment Agency EEA. Each participating country has been
Agency (EEA) and Joint Research Centre (JRC) responsible for data production in it own territory. CLC
launched the IMAGE2000 and CLC2000 projects. The updates are part of European land monitoring onwards
CLC2000 database was based on visual interpretation of 2011 which is organized within Global Monitoring of
Landsat ETM-images and ancillary data like existing Environment and Security (GMES) programme of the
maps. Interpretation was made using GIS-software [2]. European Commission [6]. Land monitoring service
includes the continuity of Corine Land Cover with a
new update for the reference year 2012, the production
of 5 additional pan-European High Resolution Layers
(HRL) and support to harmonization efforts of countries
in order to improve synergies between pan-European
and national land cover activities. All produced data sets
should be available mid-2014 [6].

The aim of this article is to describe the Finnish Corine

Land Cover classification process and discuss the
consequences using Sentinel-2 images in this
interpretation work by giving examples about the
difficulties experienced with previous images and
Figure 2. Production flow of CLC2000 in Finland.
anticipated advantages of Sentinel-2.

An automated generalization procedures were

2. FINNISH CORINE LAND COVER
developed in order to produce CLC database with 25 ha
CLASSIFICATION
MMU and changes with 5 ha MMU according to
specifications of EEA.
The production of Finnish CLC database is based on
automated interpretation of satellite images and data
The same approach which was applied in CLC2000
integration with existing digital map data, as illustrated
project was repeated in CLC2006 project [8].
in Fig. 1 and 2 [7]. Map data provides information
Automated interpretation of satellite images and data
describing land use and soils, and satellite images
integration with existing digital map data was applied.
provide information about land cover and are used to
Additionally, specific classes were interpreted manually
update map data. Continuous land cover variables are
with the aid of IMAGE2006 and ancillary data. In order
transformed into discrete CLC classes using thresholds
to detect changes between 2000 and 2006 two
according to the class descriptions of CLC
approaches were combined (fig. 3): 1. the differences
nomenclature. Land cover data and changes were
between high resolution land cover data sets 2000 and
produced in the original spatial resolution of satellite
2006 were evaluated together with the 2. changes
data. This enables the production of more detailed
detected using satellite data, i.e. IMAGE2000 and
information for national use. This was necessary in
IMAGE2006.
order to get national funding for the work and access to
the national land cover databases.

Figure 3. Production of Finnish CORINE land cover

Figure 1. The main data sources for main CLC- 2006 and changes 2000 – 2006.
classes in case of Finnish CLC2000.
 Landsat-7 ETM Spot-4 HRVIR IRS P6 LISS III Sentinel-2 MSI
Channels (µm) B1: 0.433 – 0.453
ETM1: 0.45 – 0.52 B2: 0.458 – 0.523
ETM2: 0.53 – 0.61 XI1: 0.50 – 0.59 MS1: 0.52 – 0.59 B3: 0.543 – 0.578
ETM3: 0.63 – 0.69 XI2: 0.61 – 0.68 MS2: 0.62 – 0.68 B4: 0.650 – 0.680
B5: 0.698 – 0.713
B6: 0.733 – 0.748
B7: 0.773 – 0.793
ETM4: 0.75 – 0.90 XI3: 0.78 – 0.89 MS3: 0.77 – 0.86 B8: 0.785 – 0.900
B8a: 0.855 – 0.875
B9: 0.935 – 0.955
B10: 1.360 – 1.390
ETM5: 1.55 – 1.75 XI4: 1.58 – 1.75 MS4: 1.55 – 1.70 B11: 1.565 – 1.655
ETM7: 2.09 – 2.35 B12: 2.100 – 2.280
ETM6: 10.4 – 12.5
PAN: 0.52 – 0.90 PAN: 0.61 – 0.68
Pixel size (m) 30 multispectral 20 multispectral 23.5 10: B2, B3, B4, B8
15 panchromatic 10 panchromatic 20: B5, B6, B7, B8a,
B11, B12
60: B1, B9, B10
Quantization (bits) 8 8 7 12
Swath width (km) 180 60-80 141 290
Table l. Characteristics of Landsat-7, Spot-4 and IRS P-6, as well as the Sentinel-2 MultiSpectral Instrument.

Figure 4. IMAGE2000 on the left, IMAGE2006 in the middle and Sentinel-2 image (simulated using RapidEye-image)
on the right. Red: red channel, Green: NIR channel and Blue: green channel.

3. SATELLITE IMAGES 3.1 IMAGE2000

The production of land cover information, the update of Finnish satellite image coverage IMAGE2000 consists
map databases, the interpretation of specific land use of 36 Landsat-7 ETM-images [9]. The target year was
classes and the detection of land cover changes are year 2000 (12 images) supplemented by images taken
based on satellite images. Tab. 1 presents the 1999 (9 images), 2001 (11 images) and 2002 (4
characteristics of Landsat-7 ETM+-, Spot 4 MS and IRS images). Images were taken during mid- or late growing
P6 LISS III-images used in IMAGE2000, 2006 and season, usually in July but some images taken even in
2009, as well as the MultiSpectral Instrument of September had to be used.
Sentinel-2. Fig. 4 presents an example of IMAGE2000,
IMAGE2006 and Sentinel-2 (simulated using The orthocorrection was performed by Metria Sweden.
RapidEye-image) –images. Channels 1 – 5 and 7 were resampled to 25 m and
panchromatic channel to 12.5 m grid using cubic
convolution interpolation. Clouds and shadows were correction was done using ATCOR2 of Erdas Imagine,
interpreted visually using channels 1, 2 and 6. because the software used with IMAGE2000 did not
Atmospheric correction was made using semiempirical work with Spot- and IRS-images. The aim of
correction software developed by Technical Research atmospheric correction was to remove the effects of
Center of Finland. The correction is based on SMAC atmospheric disturbances and make the corrected
algorithm and an independent atmospheric data, but images as similar as possible with IMAGE2000
parameters like aerosol optical thickness (AOD) can be mosaics. There are two variables affecting atmospheric
estimated from the image [9]. Topographic correction correction that can be changed, visibility and aerosol
was made at Northern Finland using Ekstrand correction model. The idea was to compare dense old forests from
[10]. IMAGE2000 and 2006 images, and fine-tune visibility
and aerosol model so that the dense forest reflectance of
The mosaicking of individual satellite images was IMAGE2006 image would be the same as in
carried out in order to get stratumwise and nationwide IMAGE2000. Unfortunately, this was very difficult to
mosaics for visualization purposes. Stratumwise achieve, possibly because the aerosol models do not
mosaics were also used in the interpretation process. model the state of atmosphere very well for Finnish
conditions, or there was something wrong with
3.2 IMAGE2006 calibration coefficients of images.

Finnish IMAGE2006 consists of 80 IRS P6 LISS and 51 Topographic correction was made in Northern Finland
Spot 4/5 images [11]. Target year was 2006 (57 IRS + for strata 4c and 4d. The chosen method was statistical-
41 Spot images) but there were some images from years empirical correction [12]. The Ekstrand correction that
2005 (21+2) and 2007 (2+8). There were two image was used with IMAGE2000 was not made because it
coverages; one summer coverage (47 + 35) and another requires more work and ATCOR3 did not work due to
spring/autumn coverage (33+16). Number of images large images sizes. Stratumwise mosaics were made as
was increased because the size of Spot and IRS images in the case of IMAGE2000. Histogram matching was
is smaller than Landsat ETM+-images. Other used to fine-tune pixel values, because there were
differences were that the number of channels was considerable differences of reflectance values between
smaller and the spatial resolutions were better. overlapping images in some cases (Fig. 5). Histogram
matching was made using stable areas like large forests
The orthocorrection was performed by Metria Sweden. and other natural areas. Areas like agricultural areas,
Images were resampled to 20 meter pixel using cubic wetlands and water were omitted. Mosaics were also
convolution interpolation. Clouds and their shadows resampled to 25 m pixel size for change detection.
were interpreted visually and masked out. Atmospheric

Figure 5. An example concerning difficulties processing IMAGE2006: differences in surface reflectances between
neighboring images after atmospheric correction (on the left), and the result of histogram matching (on the right).
 borders of sand pit), automatic detection of marshes or
new classification system of wetlands.

3.4 IMAGE2012

IMAGE2012 should consist of IRS P6 LISS III-images

in summer coverage and RapidEye-images in spring
coverage. RapidEye-images should be resampled to 20
m pixel in orthocorrection [14].

3.5 Sentinel-2 MSI

The characteristics of Sentinel-2 MSI [15] are presented

Figure 6. Typical example of IMAGE2009 cloud mask. in tab. 1. Compared to images used previously, there
Cloud mask does not cover all thick clouds, and thin will be more channels, their spectral resolution will be
clouds and shadows are all left out. better, spatial and radiometric resolutions will increase
and swath width will be larger. The instrument should
Because there were quite a lot gaps in IMAGE2006 be installed to two satellites; the first one should be
mosaic, SYKE received set of DMC-satellite images launched during 2014 [16].
[13], acquired by SLIM-6 instrument of BEIJING-1
mission. 13 images were taken during summer 2007. 4. EXPERIENCES CONCERNING IMAGE
Images had three channels: Green, Red and NIR PROCESSING AND INTERPRETATION
corresponding to ETM+ channels 2, 3 and 4. Good
property of these images is large image size (swath Acquisition of a cloud-free multitemporal (spring and
width about 320 km), drawbacks are small number of summer) HR satellite data coverage over Finland every
channels and worse spatial resolution (32 m at nadir, 3 years is challenging with the present instruments due
resampled to 30m in orthocorrection) that in case of to short growing season and excessive cloud cover.
ETM+, IRS or Spot. Clouds and their shadows were Acquisition of IMAGE2000 data took 4 years (1999-
removed using visual interpretation and images were 2002) and acquisition of IMAGE2006 took 3 years. The
resampled to 25 m pixel size. These images were used IMAGE2000 coverage is almost cloud-free, data is
for change detection in gaps of IMAGE2006. missing only on 0.55% (about 1800 km2) of the total
land area of Finland. There are missing data or cloudy
3.3 IMAGE2009 areas covering 4.5 % in IMAGE2006 mid-summer
coverage. In the spring coverage the corresponding
Finnish IMAGE2009 consists of Spot 4/5 MS and IRS figure is even 13.5 %. Additionally, some spring images
P6 LISS III –images with summer and spring/autumn were received in the snow-covered period.
coverage. There are smaller number of images and quite
heavy cloud cover, so the areal coverage of images is The Finnish method in CLC data production is based on
much worse than in the case of IMAGE2006. New combined usage of satellite data and existing national
feature was that cloud masks were provided for images. land cover data sets. The dates of national data and
Unfortunately these masks cover only thick clouds, not satellite images are different since national data sets are
thin clouds and shadows, and are therefore useless (Fig. extracted from databases in the nominal year of CLC
6). These images are currently used to test the new mappings and images are received within 3 years
production methods of Finnish Corine Land Cover around these nominal years. This leads to the need to
2012, like the better enlargement of point database extra modification of specific elements of national data
information (e.g. sand pit has coordinates for one point, sets using semi-automatic methods together with image
segmented satellite image is used to determine the data, which increases the production costs significantly.
There have been difficulties to make atmospheric Cover classification has been 25 meters, but this could
correction for different instruments so well that the be decreased to 10 meters in the future due to Sentinel-2
corrected images would be directly comparable with images.
each other. It was difficult to adjust the correction of
IMAGE2006 images so that the spectral reflectance of 5.3 Temporal resolution
stable targets (i.e. old coniferous forests) would
correspond to the reflectance of these targets in Temporal resolution will increase quite a lot due to
IMAGE2000 images. Also, there were surprisingly larger swath width and because the use of two satellites.
large reflectance differences between neighboring Also, the distance between neighboring paths is smaller
IMAGE2006 images in some cases. Therefore, the in Finland because it is quite far from the Equator. This
mosaicking of images were made using histogram has great significance in Finland due to dark winter,
matching between neighboring images using stable short growing season and excessive cloud cover. So far,
targets in overlapping areas in the case of IMAGE2006. it has been difficult to acquire satellite image mosaic
covering whole Finland within one growing season, the
So far, the geometric correction of images has been time-span of IMAGE2000 images was four years (1999-
quite successful. The digital elevation model used in the 2002) and IMAGE2006 three years (2005-2007).
orthocorrection should have pixel size at least the same
as in images and it would be better if the pixel size of Another consequence of increased temporal resolution
the DEM would be three times better than the size of is the possibility to use time series data (e.g. spring,
satellite image pixel, because the computation of slope summer and autumn coverages) in image interpretation.
and aspect of terrain requires 3x3 pixel window. This will increase the interpretability of certain land
Currently the best DEM available from Finnish National cover classes, like lake reed beds, coniferous vs.
Land Survey which covers whole Finland has 10 m deciduous trees or agricultural areas vs. natural
pixel size and this should be used for orthocorrection. In grasslands. Also disaster monitoring applications, e.g.
the future there will be DEM with 2 m pixel size. storm damages in forests, will benefit from increased
temporal resolution.
5. SENTINEL-2 MSI AND FINNISH CORINE
LAND COVER INTERPRETATION 5.4 Radiometric resolution

This chapter discusses the characteristics of Sentinel-2 Radiometric resolution will increase from 7 (IRS LISS)
MultiSpectral Instrument and their assumed effect to the or 8 (Landsat ETM+, Spot) bits to 12 bits. This means
production of Finnish Corine Land Cover classification. that it is possible to detect smaller differences in
measured radiance. As consequence, it is expected that
5.1 Image size this will help to differentiate similar classes like
different forest classes, and increase the accuracy of the
The swath width of Sentinel-2 MSI is 290 km. It is estimation of the vegetation parameters like tree height
expected that 9 or 10 images should cover whole and crown cover.
Finland. Larger image size means that the image mosaic
covering whole Finland has fewer images, depending on 5.5 Spectral resolution
the cloud cover, and therefore mosaicking process is
easier, less time-consuming and there will be fewer Spectral resolution will increase due to larger number of
borders between images. channels and decreased width of channels. There are
also new channels like red-edge (B5 and B6), channels
5.2 Spatial resolution for vegetation analysis (B7 and B8a) and channels for
atmospheric correction (B1, B9 and B10). It is expected
The increase of spatial resolution means that it is that these will increase the separability of land cover
possible to differentiate smaller spatial details (Fig. 4). classes and make atmospheric correction easier if
So far, the pixel size of Finnish version of Corine Land specific software will be developed for this purpose.
5.6 Radiometric correction because they contained only thick clouds, not thin
clouds or shadows.
Atmospheric correction of IMAGE2000 Landsat ETM+
images was successful because there was only one 6. CONCLUSIONS
instrument, easy-to-use software and the instrument had
channels to estimate atmospheric optical depth from This paper presented the main principles of Finnish
images. The correction of IMAGE2006 images was Corine Land Cover classification process, the
more difficult because there were two different characteristics of the used images, experiences
instruments, software was more difficult to use and the concerning image processing and expectations
aim was to make correction so that the images would be concerning future Sentinel-2 MultiSpectral Instrument.
radiometrically comparable to IMAGE2000 images. It
is assumed that the atmospheric correction will be So far, the main problem with IMAGE2000 and
simpler in the future, because images from one IMAGE2006 images has been the clouds. Because of
instrument type will be used, this instrument has them, the temporal resolution has been quite poor, for
specific channels for atmospheric correction and example IMAGE2000 has images from four different
instrument has channels (i.e. 2 µm SWIR and RED) to years. Another main difficulty has been the use of
reliable estimate atmospheric optical depth. several different instruments. It has been difficult to
make good atmospheric correction so that images would
It is expected that the topographic correction will be be radiometrically comparable with each other.
performed as before. The critical factor is the quality of
digital elevation model which is increasing in Finland It is expected that Sentinel-2 MSI will provide images
due to lidar campaigns. with better quality and temporal resolution, and which
will make the processing of future IMAGE-mosaics
5.7 Geometric correction easier and provide better Corine Land Cover
classifications due to better spatial, spectral and
Nowadays orthocorrection is the required way to radiometric resolutions.
perform geometric correction. Also here the critical
factor is the quality of digital elevation model which REFERENCES
should be as good and spatially accurate as possible.
The search of Ground Control Points should be based 1. Heymann, Y. Steenmans, Ch., Croissille, G. &
on image matching between reference image mosaic Bossard, M. (1994). Corine Land Cover –
and image to georeferenced in order to provide time- Technical Guide, Office for Official Publications of
series of images which match geometrically well to each the European Communities, Luxembourg.
other. This should make change detection more easy 2. Bossard, M., Feranec, J. & Otahel, J. (2000).
and error-free. CORINE Land Cover Technical Guide – Addendum
2000, European Environment Agency Technical
5.8 Cloud cover Report 40.
3. Büttner, G., Feranec, J. & Jaffrain, G. (2002).
Corine Land Cover Update 2000 - Technical
One important factor which decreases temporal
guidelines, European Environment Agency
resolution in Finland is cloud cover, which can be quite
Technical Report 89.
excessive during summer. Therefore it has been difficult
4. Krynitz, M. (2001). Land Cover Annual Topic
to acquire cloud-free mosaics covering Finland within
Update 2000, European Environment Agency
one summer. So far, the cloud masking has been done
Topic report 4.
using manual interpretation, because although it is easy
5. EEA, (2006). EEA Project Implementation Plan
to recognize thick clouds, thin clouds and especially GMES Land FTS 2006 – 2008, European
their shadows are much more difficult to recognize Environment Agency AMP 2006 – 1.2.4 AST /
automatically. IMAGE2009 images were provided with AMP 2007 – 1.2.6 ASO draft 16.11.2006.
cloud masks (Fig. 6); unfortunately they were useless
 6. EEA, (2011). GMES Initial Operations 2011-2013
Land monitoring Services: Annex I - Tender
Specifications, European Environment Agency
http://www.eea.europa.eu/about-us/tenders/eea-ses-
11-004-framework/tender-specifications.
7. Härmä, P., Teiniranta, R., Törmä, M., Repo, R.,
Järvenpää, E. & Kallio, E. (2004). The production
of Finnish Corine land cover 2000 classification. In
International Archives of Photogrammetry and
Remote Sensing volume XXXV Part B4, 1330-
1335.
8. Haakana, M., Hatunen, S., Härmä, P., Kallio, M.,
Katila, M., Kiiski, T., Mäkisara, K., Peräsaari, J.,
Piepponen, H., Repo, R., Teiniranta, R., Tomppo,
E. & Törmä, M. (2008). Finnish CORINE 2006
project: determining changes in land cover in
Finland between 2000 and 2006. In Remote
Sensing for Environmental Monitoring GIS
Applications and Geology VIII SPIE Proc. 7110.
9. SYKE (2005). CLC2000-Finland. Final Report,
Finnish Environment Institute (SYKE),
Geoinformatics and Land Use Division (GEO).
10. Ekstrand, S. (1996). Landsat TM Based Forest
Damage Assessment: Correction for Topographic
Effects. Photogrammetric Engineering and Remote
Sensing, vol. 62, no. 2, pp. 151-161.
11. SYKE (2009). CLC2006 Finland. Final Technical
Report, Finnish Environment Institute (SYKE),
Geoinformatics and Land Use Division (GEO).
12. Itten, K.& Meyer, P. (1993). Geometric and
Radiometric Correction of TM Data of
Mountainous Forested Areas. IEEE Transactions
on Geoscience and Remote Sensing, vol. 31, no. 4,
pp. 764-770.
13. Crowley, G. (2010). DMC Data Product Manual.
DMC International Imaging Ltd., 127.
14. ESA (2011). GMES Space Component Data Access
Portfolio: Data Warehouse 2011-2014. European
Space Agency GMES-PMAN-EOPG-TN-11-0006.
15. ESA (2010). GMES Sentinel-2 Mission
Requirements Document. European Space Agency
EOP-SM/1163/MR-dr.
16. ESA (2012). Sentinel 2. European Space Agency
www-page https://earth.esa.int/web/guest/missions/
esa-future-missions/sentinel-2, accessed 14.4.2012.

View publication stats