Scene classification from PolSAR image using medium-level features

A medium-level feature detection frame for fully polarimetric synthetic aperture radar (SAR) data is proposed. The medium-level feature is used for striding over the semantic gap between the low-level polarization scattering characteristics and the high-level semantics of land-use categories. It represents a pixel/image by a fixed-length vector of occurrence frequencies of pre-estimated elemental scattering-mechanisms. Six different state-of-art unsupervised polSAR classification methods are tried to obtain these elemental scattering-mechanisms. A supervised method combining the medium-level feature and support vector machine (SVM) is proposed for land-use classification. The test is performed on a Radarsat-2 fully polarimetric data. The results reveal its potential on classifying complex scenes such as building area and forest.