as post- 
xels were 
(1) 
between 
oÜnging to 
yntextual 
nizes the 
s of the 
ghboring 
ld (MRF) 
a'priori 
ilt of the 
ld (MRF) 
ities are 
classified 
s (Hord, 
'omputed 
jr matrix 
esent the 
or class 1 
ata class 
sses. The 
correctly 
presents 
a classes 
cy values 
id, 1994): 
‘the total 
the total 
iding the 
ach class 
] for that 
robability 
reference 
ling pixel 
ding the 
ach class 
issified to 
robability 
epresents 
sification 
ween the 
data and 
greement 
random 
classification. If KHAT is close to one it indicates 
that the classification result is better than a 
random classification. 
The error matrix of LVQ-classification is represented in 
Table 3. The overall accuracy was 86.4% and kappa 0.83. 
The producer’s and user’s accuracies are represented in 
Table 5. Contextual classification increased the overall 
accuracy a bit, it was 88.596 (kappa 0.85) for CI- 
classifier and 87.3% (kappa 0.84) for M1-classifier. The 
best classifier was M2 when transitional probabilities 
were estimated using 7 by 7 pixel window. In that case 
the overall accuracy was 89.8% and kappa 0.87. 
Corresponding error matrix is represented in Table 4 
and the producer’s and user’s accuracies in Table 5. 
Water and agricultural field were classified very well, 
producer’s and user’s accuracies were over 99% and 94%, 
respectively. Some pixels belonging to agricultural field 
were classified as forest with low stem volume. Urban 
area and mire were classified moderately well, 
producer’s and user’s accuracies were over 77% and 78%, 
respectively. These classes were mixed mainly with 
forest classes. Forest classes were classified from very 
poor to moderate, they were mainly mixed with each 
other or mire and urban area. Very poor accuracy of 
forest, stem volume 100 - 200 m?/ha was probably due to 
small sample size used in the training of the classifier. 
Contextual classification increased the classification 
accuracies except for forest, stem volume 100 - 200 
m/ha. 
8. CONCLUSIONS 
Generally the mean backscatter of classes is higher 
during summer than winter and the temporal variations 
of the mean backscatter are higher during summer and 
autumn than during winter. Backscatter from forest 
classes is lowest when ground is frozen and covered with 
thin layer of snow. Accumulation of new snow increases 
backscatter and aging of snow decreases it. Variations of 
forest classes increase according to the increase in stem 
volume. Also differences in backscatter between summer 
and winter increase with the stem volume. Differences 
in backscatter of mire and forest classes are largest 
during summer, otherwise backscatter behaves quite 
similar. The backscatter from urban area does not vary 
much, it is almost undependent from season. 
The mean class separabilities are largest for Jers, 
summer and late autumn ERS-1, snow-free and unfrozen 
Radarsat and frozen Radarsat images. The separabilities 
are smallest for winter and very rainy ERS-1 images, so 
ground freezing and heavy rain decrease the suitability 
of ERS-1 images for land-use classification. Median 
filtering increases the class separability. 
The results of classifications were satisfactory. The 
overall accuracy was 86.496 for LVQ-classification and 
89.8% for the best contextual classification. Water and 
agricultural field were classified very well and urban 
area and mire morerately well. Forest classes are the 
most problematic, they mix with each other, mire and in 
some cases with urban area. 
ACKNOWLEDGMENTS 
This study is part of European Union funded EUFORA- 
project and conducted under the supervision of Helsinki 
University of Technology, Laboratory of Space 
Technology. Authors wish to thank Dr.Tech. Jouni 
Pulliainen for help and comments. 
REFERENCES 
Devivjer P., Kittler J., 1982. Pattern Recognition - A 
Statistical Approach. Prentice-Hall, 1982. 
Hord M., 1986. Remote Sensing - Methods and 
Applications. John Wiley & Sons, 1986. 
Jain A., Dubes R., 1988. Algorithms for Clustering Data. 
Prentice-Hall, 1988. 
Kohonen T., 1990. The Self-Organizing Map. Proceedings 
of the IEEE, Vol. 78, no. 9, September 1990, pp. 1464- 
1480. 
Lillesand T., Kiefer R., 1994. Remote Sensing and Image 
Interpretation. John Wiley & Sons 1994. 
Pulliainen J., Mikkel4 P., Hallikainen M., Ikonen J-P., 
1996. Seasonal Dynamics of C-Band Backscatter of 
Boreal Forests with Applications to Biomass and Soil 
Moisture Estimation. IEEE Transactions on Geoscience 
and Remote Sensing, Vol. 34, no. 3, May 1996, pp. 758- 
769. 
Richards J., 1993. Remote Sensing Digital Image 
Analysis. 2nd ed. Springer-Verlag, 1993. 
Rauste Y., 1989. Methods for Analyzing SAR images. 
Tech. Res. Ctr. Finland, Lab Instrument Technol. 
Espoo, Finland, Rep. 612, 1989. 
Tomppo E., 1997. Finnish National Forest Inventory. 
Proceedings of the Finnish - Russian seminar on remote 
sensing in Helsinki, 29. Aug. - 1. Sep. 1994. Reports of 
the Finnish Geodetic Institute, 97:2. 
Tórmà M, 1997. Land-Use Classification Using SAR- 
Images. Photogrammetric Journal of Finland, Vol. 15, 
no. 2, 1997, pp. 31-47. 
Vuorela A., 1997. Satellite Image Based Land Cover and 
Forest Classification of Finland. Proceedings of the 
Finnish - Russian seminar on remote sensing in 
Helsinki, 29. Aug. - 1. Sep. 1994. Reports of the Finnish 
Geodetic Institute, 97:2. 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 571