RESULTS The Mamdani fuzzy logic sys

RESULTS
The Mamdani fuzzy logic system was applied to classify the image into 5 land classes and the accuracy was determined. The type and position of output membership functions were changed to analyse changes in the result. First the input membership function of water body was changed while keeping all the other input membership function same and the result were studied.
Total number of pixels in image 781 × 671 = 524052
Table 1 shows the loss of pixels that is misclassification when membership functions were changed. The loss (misclassification) is very high in case of trapezoidal membership function.Losses were measured taking number of pixels for waterbody in case of output for gaussmf case as the base.
From the output obtained as shown in Figure 6, it was clear that in other cases original water body pixels were wrongly classified as other land classes. Although membership functions like triangular and trapezoidal gave sharper edges but the loss of pixel that is classification error was clearly visible in case of trapezoidal and triangular membership functions. Since only input membership function for waterbody was changed, the effect on the other classes was minimal. On modifying the arrangement of membership functions and keeping the rules same, different results were obtained. In first attempt of classification, the forest and the pasture land classes, which had input membership functions having values which were very close to each other, in one or the two bands, membership function in the output for these land classes were not placed adjacent to one another.
Figure 7 shows rule editor of first case in this arrangement mf4 which represent urban land class was placed between forest (mf3) and pasture (mf5) land class which have very similar input values in one or two bands.
Figure 8 shows the output for this (first) arrangement. The misclassification in the case of urban land class was clearly visible. Many of the pixels which were pasture or forest were misclassified under urban land class.
Table 2 shows the colour taxonomy, the yellow colour was used to depict pasture similarly green for forest, blue for water body, white for clouds and red for urban. In the second case two closely related membership functions were placed adjacent to each other and the unrelated membership function was not placed inbetween them.
Figure 9 shows forest (mf3) and pasture (mf4) land class placed adjacent to each other and the urban (mf5 here) membership function, whose input range was not overlapping in any of the bands of input mf of forest and pasture, was not placed in-between them.
Figure 10 shows the result of the second arrangement. The improvement in classification was clearly visible. Original pasture and forest class pixels were not misclassified as urban. The change in arrangement did not affect the output for two other land classes- clouds and waterbody. Idea for accuracy assessment methods of classification results comes from the selecting random sample with known classes and then let methods „say‟ what these samples are. With 100 random selected samples, Table 3 shows the comparison of two arrangements.
100 samples from the output of two arrangements were taken and they were verified with the original image. The accuracy obtained in the first arrangement was 43% whereas the same for second arrangement was 87%.