to the direction of the maximum variance of data 
determined by the largest eigenvalue of 
covariancematrix and its direction is determined by 
corresponding eigenvector. Second principal component 
axis is placed according to second largest variance of 
data and so that it is orthogonal to first axis and so on. 
The result of the transformation is new set of images, 
where in principle first images correspond to information 
needed in classification and latter images correspond to 
random components like speckle (Richards, 1993). 
Intensity variations between neighboring pixels were 
characterized by simple texture measure. First, the 
mean of the differencies between center pixel and 
neighboring pixels were computed for each pixel and 
then resulting image was mean filtered by 7 by 7 pixel 
window (Tôrmä, 1997). 
Steps in image processing: 
1. Images were filtered by using median filtering (size 
of window 3 by 3 pixels). 
2. PCA was performed to 14 ERS-1 images and three 
first principal component images containing 58% 
about overall variance were used in classification. 
3. PCA was performed to 4 Radarsat-images and two 
first principal component images containing 88% 
about overall variance were used in classification. 
4. PCA was performed to 2 JERS-images and first 
principal component image containing 85% about 
overall variance was used in classification. 
5. Texture features were computed from original SAR- 
images and then PCA was performed to these 
images. Two first principal component images 
containing 84% about overall variance were used in 
classification. 
6. Finally, each of these eight images was normalized 
to zero mean and unit variance (Jain, 1988). 
6. CLASSIFICATION 
Spectral classification was performed by using self- 
organizing feature map (SOM) and learning vector 
quantization (LVQ). SOM is a self-organizing neural 
network which uses competitive learning to adapt itself 
to the density function of the input patterns. Network 
consists of processing elements arranged usually to two- 
dimensional sheet and each processing element has 
weight vector. During learning the weight vectors 
converge to places where they approximate the density 
function of the input patterns (Kohonen, 1990). After 
training, weight vectors are labelled by using training 
set. LVQ methods try to find optimal places for the 
weight vectors representing different classes by using 
supervised learning. LVQ method type 3 has been used 
in this study (Kohonen, 1990). Classification is 
performed as nearest neighbor classification, when 
weight vectors are used as training data. In other words, 
input vector is classified according to the nearest weight 
vector. 
Contextual classification was performed as post- 
processing step. A’posteriori probabilities for pixels were 
estimated as follows: 
Pio s oer à 
EM MDC 
J=1 
where d,,(x,wy;) is squared Euclidean distance between 
pixel x and the nearest weight vector wy; belonging to 
class i and c is number of classes. Following contextual 
classifier were implemented: 
C1: Pixel is classified to class, which maximizes the 
product of the a'posteriori probabilities of the 
center and four neighboring pixels. Neighboring 
pixels are supposed to be independent. 
MI: Classifier is based on Markov random field (MRF) 
(Devivjer, 1982) and transitional and a’priori 
probabilities are estimated from the result of the 
spectral classification (Hord, 1986). 
M2: Classifier is based on Markov random field (MRF) 
and transitional and a’priori probabilities are 
estimated locally in the neighborhood of classified 
pixel by using a'posteriori probabilities (Hord, 
1986). 
7. RESULTS 
The error matrices and accuracy values were computed 
from classification results. The columns of error matrix 
represent the reference data and the rows represent the 
classification result. In other words, column for class 1 
represents how many pixels from reference data class 
swamp have been classified into different classes. The 
diagonal elements represent the amount of correctly 
classified pixels. Likewise, row for class 1 represents 
how many pixels from different reference data classes 
are classified as swamp. The following accuracy values 
were computed from the error matrices (Lillesand, 1994): 
x Overall accuracy: is computed by dividing the total 
number of correctly classified pixels by the total 
number of reference pixels. 
* Producer’s accuracy: is computed by dividing the 
number of correctly classified pixels in each class 
by the number of training set pixels used for that 
class (column total). It indicates the probability 
that a pixel taken randomly from the reference 
data has the same class as the corresponding pixel 
in the classification result. 
* Users accuracy: is computed by dividing the 
number of correctly classified pixels in each class 
by the total number of pixels that were classified to 
that class (row total). It indicates the probability 
that a pixel classified to a class actually represents 
that class in the reference data. 
x KHAT: is a measure of randomness of classification 
result. It measures the difference between the 
actual agreement between the reference data and 
the classification result and the change agreement 
between the reference data and a random 
570 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
ba 
for 
vol 
an 
in 
dui 
Sin 
Th 
SUI 
Rai 
are 
gro 
of 
filt 
The 
ove 
89. 
are