Spectral Clustering of Polarimetric Sar Data with Wishart-derived Distance Measures

This paper presents a new spectral clustering algorithm, which is specially tailored for segmentation of polarimetric SAR images. This is accomplished by use of certain pairwise distance measures between pixels. The measures are derived from the complex Wishart distribution, and capture the statistical information contained in the coherency matrix. We demonstrate how the pairwise distances are transformed into an affinity matrix, whose eigendecomposition determines the optimal partitioning of pixels. We further show that the obtained clustering provides an improved initialization of the classical unsupervised Wishart classifier, and that the entire classification can also be performed in a kernel induced feature space. The algorithms are tested on crop classification with promising results.