A Generic Land-Cover Classification Framework for Polarimetric SAR Images Using the Optimum Touzi Decomposition Parameter Subset—An Insight on Mutual Information-Based Feature Selection Techniques

This correspondence proposes a generic framework for land-cover classification using support vector machine (SVM) classifier for polarimetric synthetic aperture radar (SAR) images considering the optimum Touzi decomposition parameters. Some new concerns have been raised recently with the Cloude-Pottier decomposition. Cloude´s α scattering type ambiguities may take place for certain scatterers, and some of the Cloude-Pottier´s parameters may not be roll-invariant for asymmetric targets. The Touzi decomposition is a relatively new roll-invariant target scattering decomposition, and it uses the target helicity, symmetric scattering type magnitude and phase. The parameters generated by the Touzi decomposition are of different physical significances, i.e., some of them are angular in nature where others are from R. Thus, classification using the Touzi parameters requires them to be normalized within the similar dynamic range preserving their physical properties. Here, a linear normalization technique has been introduced, which maps the angular parameters to R without loss of generalization. The power of mutual information (MI) has been explored hence after for selecting the optimum set of classification parameters. A third-order class-dependent MI-based method and another method based on the Eigen-space decomposition of the class conditional MI matrix have been introduced for this purpose. For SVM-based final classification, a normalized histogram intersection kernel (NIKSVM) has been proposed that boosts the generalization accuracy to a considerable extent as compared to normal histogram intersection kernel. An ALOS L-band SAR image of Mumbai area, India has been considered here to exhibit the performance of the proposed cost-effective classification framework.