The TM data were taken during one Landsat pass in July 1992. 
A 229x319 portion of the six nonthermal bands was used for the 
classification The maps, aerial photographs and ground truth 
were used to prepare a thematic map of this scene to assess the 
classification accuracy. By this way, 4839 (6.62% of the area) 
pixels were selected. Overall, 1196 pixels were used for training 
and 3643 pixels for testing the classifiers. Six TM channels in 
the visible and in the infrared spectrum were selected to form 
input vector for each pixel. Table 2. shows the mean and 
standard deviation of each class. 
E 
Figure 2. Channel 5 of the Landsat TM image. 
  
4. EXPERIMENT RESULTS 
Before performing classification using maximum likelihood 
classifier and neural network classifier, we have calculated the 
separability between each pair of classes to determine whether to 
modify some classes, create a new set of classes or merge those 
pairs of classes whose separability is too low. Bhatacharrya 
Distance was used as a class separability measure (Richards 
1986). This measure is monotonically related to classification 
accuracies. The larger the separability values are, the better the 
final classification results will be. Table 3. shows the separability 
matrix. 
Table 3. Separability matrix. 
350 9 6970435 13045149009 951 203) To 8957 D 
+ 
511.72 
70.00 2.00 
9/1.99 2.00 1.99 
11/1.99 2.00 2.00 1.99 
132.00 2.00 2.00 2.00 2.00 
152.00 2.00 2.00 2.00 1.99 1.99 
171.99 1.99 2.00 1.98 1.99 2.00 2.00 
1912.00 2.00 1.99 2.00 2.00 2.00 2.00 2.00 
2112.00 2.00 1.98 1.53 2.00 2.00 2.00 1.87 1.99 
230.00 2.00 1.99 2.00 2.00 2.00 2.00 2.00 1.98 2.00 
2512.00 2.00 1.93 1.98 2.00 2.00 2.00 1.99 1.74 1.94 1.99 
2711.83 1.59 2.00 1.99 1.99 2.00 2.00 1.82 2.00 1.99 2.00 1.99 
2911.99 1.99 2.00 1.99 1.96 2.00 1.99 1.99 2.00 1.99 2.00 2.00 1.99 
31/2.00 2.00 2.00 1.99 1.99 1.98 1.99 2.00 2.00 1.99 2.00 1.99 2.00 1.98 
  
As shown above, the average separability is 1.97952, which 
indicates that the overall separability among the fifteen classes is 
relatively good. Another evalution of spectral separability is 
provided by a confusion matrix (Lillesand et al. 1987) as shown 
in Table 4 and Table 6. These tables are made by classifing the 
testing set pixels. The confusion matrix gives information on 
how much of each original testing area was actually classified as 
being in the class that testing was meant to represent. 
Table 2. Means and standard deviations of the classes for training set of data. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Channel 1 Channel 2 Channel 3 Channel 4 Channel 5 Channel 7 
Class # Mean. Stan Mean Stan Mean. Stan Mean Stan Mean. Stan Mean. Stan 
value dev. value. dev. value dev. value dev. value dev. value dev. 
1 641 13 25.7 07 216 09 82.3 3.9 61.4 2.9 175 15 
2 63.8 11 249 0.5 210 03 741 2.6 552 17 15.7 1.0 
3 785 12 36.9 0.7 44.6 1.1 50.4 1.1 739 31 383 19 
4 72,5. 13 332 0.6 352 10 731 1.5 893 28 354 1.6 
5 679 12 304 0.8 269 09 88.6 2.8 65.8 1.8 20.9 0.9 
6 772 15 392 12 354 14 99.0 3.9 68.7 1.7 21 19 
7 70.1 19 302 13 268 15 105.1 4.6 584 2.0 170 0.8 
8 684 1.0 29.1 0.7 294. 12 646 53 78.5 4.5 28.5 ‚1.5 
9 877 23 434 1.5 $29. 27 $52 22 77.0 5.4 45.1 3.8 
10 732 17 326 09 371 18 559 12 862 40 398 22 
11 73.66 2.6 322.26 363 3.8 40.1 3.7 519 7.6 271 48 
12 985 34 $524 1.8 ] 6042 32 726 20 105.4 3.0 59.3 27 
13 640 13 263 07 23.1 1.1 69.2 3.9 60.5 4.7 18.7 1.8 
14 70.8 1.0 31.1 0.8 29.7 0.8 712 16 63.0 1.4 236 1.2 
15 775-17 356 1.1 368 19 772 34 73.0 3.8 288 22 
  
  
  
  
  
The average accuracy is the average of the accuracies for each 
class and overall accuracy is similar average with the accuracy of 
each class weighted by the proportion of test samples for that 
class in the total testing set. The average accuracy and the overall 
accuracy of the train pixels for the maximum likelihood 
algorithm were 93.7496 and 93.6596 respectively. The average 
accuracy and the overall accuracy of the test pixels for the 
maximum likelihood classification were 92.28% and 91.96% 
respectively. Looking at the classification of the testing samples 
(Table 4.), we see that “Arid Soil “ suffered from the worst 
classification confusion, with only 86.9% of the testing area 
classified as “Arid Soil “. We can also see that out of 344 pixels, 
7.8% were classified as “Barley I”, 0.3% were classified as 
“OilRape P’, while 4.9% were not classified at all. The training 
378 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998