A data-driven approach for clustering scatter values in hybrid-polarimetric SAR images

In this article, we discuss clustering of hybrid-polarimetric SAR images using finite mixture model framework. We use, particularly, Gaussian mixture model to fit the distribution of data. As features of clustering technique, we use three decomposition techniques of hybrid-polarimetric synthetic aperture radar (SAR) data, namely, , , and decompositions. The three pseudo-power components that are derived from these decompositions are clustered using data-driven multivariate Gaussian modelling. These clustered images would give an idea of land use in terms of three broad classes, viz., surface, volume, and dihedral, which are defined with reference to SAR scattering characteristics. We show the efficiency of this technique by successfully clustering the data where dominant scatter-based clustering could not properly distinguish the scattering phenomenon. We also discuss an incremental directional smoothing for speckle suppression of polarimetric SAR images. For experiments, we use hybrid-polarimetric C-band SAR data from Indian Space Research Organisation's Radar Imaging Satellite (RISAT-1). We also show the results on a simulated hybrid-polarimetric SAR image from Advanced Land Observing Satellite's (ALOS-1) full-polarimetric SAR data. From our experiments, it is seen that the proposed technique performs consistently and satisfactorily in clustering hybrid-polarimetric SAR data on different decomposition techniques.