LAND-USE CLASSIFICATION USING MULTITEMPORAL 
RADARSAT, ERS-1 AND JERS SAR-IMAGES 
Markus Törmä!, Jarkko Koskinen? 
Helsinki University of Technology 
Institute of Photogrammetry and Remote Sensing, Otakaari 1, 02150 Espoo, Finland 
Tel: +358-9-451 3896, Fax: +358-9-465 077, Email: Markus.Torma@hut.fi 
?Laboratory of Space Technology, Otakaari 5 A, 02150 Espoo, Finland 
Tel: +358-9-451 2170, Fax: +358-9-451 2898, Email: jarkko@avasun.hut.fi 
Commission VII, Working Group 6 
KEY WORDS: SAR, land-use, contextual classification, Markov random field, learning vector quantization 
ABSTRACT 
Land-use classification was performed by using a set of ERS-1, JERS- and Radarsat images. Classes were water, 
forest (with three subclasses according to stem volume), agricultural field, mire and urban area. The effect of 
environmental conditions to class separability was investigated by using Bhattacharyya distance. The classes were 
most separable in JERS, summer and late autumn ERS-1 and Radarsat images with steep incidence angle. Median 
filtering was used for speckle reduction and principal component analysis for feature extraction. Spectral classification 
was performed by using self-organizing feature map and learning vector quantization. Contextual classification was 
performed as post-processing step. The overall accuracy of the spectral classification was 86.4% and the best 
contextual classification 89.8%. 
1. INTRODUCTION 
Many governmental institutions have a continuing 
requirement to form and implement laws and policies 
that involve existing or future land-use. Optical aerial 
and satellite images have been used long time to produce 
information about the current land-use. Unfortunately 
weather conditions limit the use of optical data. For 
example, here in Finland summer is usually quite 
cloudy, there are usually only few days when large area 
of Finland is cloud-free, and during winter there is dark 
also daytime. These facts have lead to investigate the 
use of microwave data to obtain spatial information. 
The aim of this study is to investigate the potential of 
ERS-1, Radarsat and JERS SAR-images to obtain 
reliable land-use classification. Classes are water, forest 
(with subclasses according to stem volume), agricultural 
field, mire and urban area. Single day images do not 
provide enough information for reliable classification, so 
several images taken in different weather conditions and 
different times are used. Factors influencing 
classification accuracy are speckle reduction and 
selecting or extracting relevant features. Median 
filtering was used for speckle reduction and principal 
component analysis for feature extraction. 
2. TEST AREA AND DATA 
The test area Porvoo is situated in southern Finland. 
The area includes several land-use classes like forests 
(mostly conifer-dominated mixed forests), mires, 
agricultural fields, lakes and urban areas. The most 
usual tree species are Norway spruce and Scots pine. 
The overall relief is quite low, elevation is well below 
100m, but not flat because of small hills (Pulliainen, 
1996). 
SAR-images used in this study were ERS-1 (frequency: 
5.3 GHz, polarization: VV, incidence angle: 23°), Jers 
(1.28 GHz, HH, 35°) and Radarsat (5.3 GHz, HH, 23° 
(S1) or 47° (S7)) images. Table 1 represents the dates 
and environmental conditions of the images. 
Temperature is the mean temperature of air during that 
day and precipitation is three-day cumulative sum prior 
image acquisition. 14 ERS-1 images (images 1 to 14 in 
the table 1) were taken during 30.6.1993 - 25.4.1994 
within ESA AO-programme. ERS-1 images have been 
topographically corrected (Rauste, 1989) and averaged to 
25 by 25 m? pixelsize. JERS-images were taken 
18.12.1994 and 30.4.1995 (images 15 and 16 in the table 
1) and Radarsat-images (images 17 to 20 in the table 1) 
were taken 19.11.1997 (incidence angle: 47°), 30.11.1997 
(23°), 18.12.1997 (47°) and 30.1.1998 (47°). JERS- and 
Radarsat-images were not topographically corrected due 
to technical difficulties with software. Also these images 
were averaged to 25 by 25 m° pixelsize. Area used in 
this study covers 12.5 by 12.5 km? so size of each image 
was 500 by 500 pixels. 
Digital land-use map made by National Land Survey 
(Vuorela, 1997) and stem volume map made by Finnish 
Forest Research Institute (Tomppo, 1997) were used as 
reference data. By combining these maps reference data 
for following classes were obtained (border pixels and 
small areas were removed): 1. mire (stem volume under 
100 m“/ha, 794 pixels for training + 3178 for testing), 2. 
566 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998