Multifractal analysis of SAR images for unsupervised classification

In present paper an attempt has been made for unsupervised classification of SAR images based on the surface roughness using multifractal technique. Surface roughness is measured with the help of fractal dimension (D), which lies in the range 2.0 and 3.0. Based on roughness values, i.e., D, various land classes are grouped in different classes. The D values are estimated for a number of local window sizes and thus the window size is very important for classification. The window size is optimized for best classification and in present case it is 9times9. The K-means classifier has been used for this procedure which clusters various land classes according to D values. Although fractal dimension is able to provide the roughness values for various land classes, it can not uniquely identify all classes. In order to remove this discrepancy, the multifractal analysis has been performed. The multifractal dimension has been estimated as 5 generalized dimensions providing 5 multifractal images and then these images are classified. The overall classification accuracy using fractal dimension alone comes to be nearly 60% while it increases to 67% with multifractal images.