Unsupervised Terrain Classification Based on Adaptive Decomposition of Generalized Scattering Model for Hybrid Compact Polarimetric Sar Data
Linlin Zhang,Gui Gao,Xi Zhang,Heng-Chao Li,Zenghui Zhang,Tao Liu
DOI: https://doi.org/10.2139/ssrn.4680423
2023-01-01
Abstract:Hybrid compact polarimetric (HCP) SAR is an emerging polarimetric SAR technology, which has great potential for Earth observation. However, existing target decomposition methods for HCP SAR data have problems such as target idealization, overestimation of volume scattering component, and poor model adaptability, resulting in confusion and errors in physical characteristics extraction and terrain classification. This paper proposed an unsupervised classification method based on adaptive decomposition of generalized scattering model for HCP SAR data. First, a new combination of two selected discriminators is established to identify the dominant scattering mechanism for each pixel, providing better guidance for model simplification in decomposition process. Second, by introducing depolarization for surface and dihedral scattering components, the generalized scattering model is constructed, describing different types of targets of any randomness and alleviating the overestimation of volume scattering; when solving the generalized scattering model, several strategies are utilized to deal with the under-determined problem, and the distribution of surface, dihedral or volume scattering component can be adaptively determined according to the input HCP SAR data. Third, a novel unsupervised classification based on the decomposition results is developed. Pixels without clear representative scattering mechanisms are considered as mixed pixels, and they are clustered and merged into dominant clusters based on the shortest Wishart distance and polarimetric scattering characteristics matching. Experiments using the real HCP SAR dataset RISAT-1 show that the proposed method can effectively alleviate the overestimate of volume scattering component, accurately highlight the representative scattering mechanism in different typical regions, and obtain the best comprehensive decomposition and classification results among existing methods. In addition, the convention for the sign of the fourth element of the Stokes vector is given in this paper.