Statistical Mid-Level Features for Building-Up Area Extraction from Full Polarimetric SAR Imagery
This paper addresses the problem of designing statistical features for the extraction of building-up areas (BAs) from highresolution polarimetric synthetic aperture radar (PolSAR) imagery. The idea is to represent a building-up area by the distribution of its mid-level components, called intermediates, which are statistical patterns unsupervisedly learnt from PolSAR images. More precisely, by analyzing the structural properties and the polarimetric characteristics exhibited in various terrain types, we propose two kinds of midlevel features for small regions: the cluster based statistical feature (CSF) and the scattering mechanism based statistical feature (SMSF). In detail, for the CSF, the intermediates are the K-mean clusters with Wishart distance of the PolSAR images; for the SMSF, the intermediates are the scattering mechanism categories obtained by relying on a four-component decomposition with deorientation of the PolSAR images. In contrast with existing features for describing BAs, the proposed features, i.e., CSF and SMSF, capture more complex context information of BAs. We compare the proposed features with those based on the Gaussian Markov random field (GMRF) models, which have been proven to be suitable for BAs mapping. Experimental results on RADARSAT-2 datasets demonstrate the effectiveness of the proposed features.