Unsupervised Classification of Polarimetric SAR Images by Gamma-Correction of Features using Self Organizing Map

Unsupervised classification algorithms have tendency to form more clusters in the upper range of tin- features used for classification. In case of polarimetric SAR.. this means that, different areas having low polarimetric response will be merged into a single cluster and an area with large polarimetric response will be segmented into many clusters. This, for example, occurs in the areas like water, grass, roads and ot her flat surfaces which, owing to t heir low polarimetric response, are often assigned to the same cluster. Urban areas and other man-made objects, on the other hand, generally have large polarimetric response, and have tendency to form more clusters than required. To avoid this, we scale the features by applying gamma-correct ion or a power law transformation to nonlinearly reduce the dynamic range of features. A gamma-corrected feature has a more evenly distributed histogram as compared to the untransformed feature. In the transformed feature space thus obtained, the areas with low polarimetric response are mapped to a larger range of the total range as compared to the input feature space. This transformation of the feature space may provide a chance to the clustering algorithm to separate the low polarimetric response areas from each other. A neural network based on self organizing map (SOM) is used to classify I lie polarimetric SAR image into different clusters. A variety of features have been utilized to form the feature vectors. Elements of the polarimetric coherency matrix, derived features like Freeman decompositions. Alpha, Entropy. Anisotropy, and Eigen values of the coherency matrix may be used. The results show the comparison of the images classified by using both transformed and untransformed features.