International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
delta rules. This algorithm operates in the batch mode and is 
invoked using the train. 
5. EXPERIMENT AND RESULTS 
In this section, the classification process stages that have 
been explained in previous sections are implemented. Figure 
2 shows the flowchart of the algorithm stages. To implement 
the algorithm, IRS-1D satellite image from a region in the 
north of Iran has been used. Bands 2, 3, and 4 of Liss-ITI 
image have been selected and fused with the PAN band to 
construct an image with 5.8 m spatial resolution. 
  
Image 
  
  
SOM 
  
  
  
Y 
Segmentation 
  
  
  
  
Feature 
extraction 
  
  
  
  
  
  
Training » MLP 
  
  
  
  
  
  
Y 
Classification 
  
  
  
  
Classified 
image 
  
  
  
Figure 2. Flowchart of land cover classification using ANN. 
The proposed segmentation utilizes a self-organizing map to 
detect the main features present in the image. The features are 
represented by their chromaticity value, which expresses 
colors hue and saturation, avoiding the luminance 
component. 
The network is composed of an orthogonal grid of cluster 
units, each associated to three orthogonal weights for the 
chromaticity data. The initial values for weights are set to 
random values before the learning phase. The produced 
segmented image includes 40 different segments. The image 
constructed from segmentation process is illustrated in figure 
4(b). One indicator pixel from each kind of segments has 
been selected. 
Then, the features have been extracted from image with a 
window centered on the indicator pixel. The sizes of the 
windows are variable. The statistical features were computed 
from a window with 3x3, 5x5, 7x7 pixels size. In order to 
extract texture features, the size of the feature extraction 
window has been variable as 9x9, 11x11 and 13x13 pixels. 
Then, multilayer perceptron network has been designed and 
trained using training data. The training diagram for network 
is shown in figure 3. 
Performance is 0.00195328, Goal is 0 
10 : ' r X 
Training-Blue 
  
  
L i i 
0 5 10 15 20 25 x 
30 Epochs 
Figure 3. Training diagram for MLC network 
Finally, the extracted statistical and texture features with the 
intensity of the central pixel of each window are arranged in a 
vector and fed to the MLP network. The MLP identifies the 
class number of each segment. The classified image has been 
shown in the figure 4(c). 
The MLP experimental results have been compared to 
Maximum Likelihood Classification (MLC) method that is 
one of the conventional classification methods based on 
statistics. The maximum likelihood decision rule is based on 
the probability that a pixel belongs to a particular class. The 
basic equation assumes that these probabilities are equal for 
all classes, and that the input bands have normal distributions. 
Figure 4(d) shows a classified image obtained using MLC. 
In order to assess the accuracy of each classification method, 
200 random points were considered for each classification 
maps. Then, they have been compared with the reference map 
to compute the overall classification accuracy. The obtained 
overall accuracies using MLC and BPNN methods are 
78.50% and 86.19% respectively. Hence, from a global 
accuracy assessment view, the NN method employed in this 
research has shown a significant improvement in 
classification of IRS images compared to MLC method. 
6. CONCLUSIONS 
In comparison of the neural network classification method 
implemented in this study to the maximum likelihood method 
on an IRS-1D image, it has been concluded that the neural 
network method is more accurate than MLC method. As 
mentioned previously, the overall accuracy in the neural 
network method has shown great improvement. Also, the 
neural network method is more sensitive to training sites than 
MLC. The sensitivity may be regarded as an improvement to 
the algorithm. 
By applying the segmentation prior to the classification, the 
classified image contains more improved edges compared to 
the MLC method. As another improvement, the MLP method 
does not classify those unnecessary pixels that is not trained 
for and avoid misclassifications. However, the MLC method 
particularly when using this image has classified all pixels 
even for those that has not been trained for. 
Internatio 
hl 
  
  
  
Using se 
classificati 
the segme 
time cons 
feature ex 
the proces: 
faster thar 
applied. It 
a modest | 
to store th 
equivalent 
Augustei 
“Performa 
cover iden 
network c 
Remote Se