270
11: Teerasit Kasetkasem, Manoj K. Arora, Pramod K. Varshney
Grass
Tree
. - ~" . .,. , :
' .
. ..- ,
.' -IT
Road
I ~- r
. : ~~ t ;
' -' "
,
.
Roof 1
Shadow
· .' r
r"r
•
., n
I
I
Roof 2
. .
,
, .
.
Fig. 11.9. Resulting MRF model based sub-pixel classification for grass, roofl, tree, shadow,
road and roof2
Table 11.2. Overall accuracies of MLE and MRF derived sub-pixel classification
Data type
MLE derived sub-pixel classification
MRF derived sub-pixel classification
Overall accuracy
59.49 %
70.03 %
clearly demonstrates the excellent performance of the MRF based algorithm
to produce the sub-pixel map from multispectral data.
Further, as a by-product of the SPM, a new set of sub-pixel classifications
or fraction images has also been generated (Fig. 11.9). The improvement over
the initial fraction images (Fig. 11.8) can easily be observed in the MRF derived fraction images. Many isolated pixels particularly in areas covered with
grass and roof2 have been eliminated. Thus, the land cover class proportions
have smoother appearance, which matches well with fraction reference images
shown in Fig. 11.5. To quantitatively evaluate the performance of sub-pixel
classification (i. e. fraction images), an overall accuracy measure based on the
fuzzy error matrix (see Sect. 2.6.5 of Chap. 2) has been used. The significant
increase in the overall accuracy (Table 11.2) of the resulting MRF derived
sub-pixel classification over the initial MLC derived sub-pixel classification,
illustrates the excellent performance of the proposed algorithm.