Classification of polarimetric SAR data based on subspace projection segmentation

The fully polarimetric SAR (PolSAR) data offers four polarimetric modes (i.e., HH, HV, VH, VV) and have shown the ability to provide better interpretation than single polarization case, which led to high classification accuracy. In this study, an efficient classification method for PolSAR data based on a subspace projection segmentation (SPS) approach is proposed to improve the classification accuracy. Instead of performing the comparison of multi-dimensional (MD) polarimetric feature vectors, the SPS first transforms the MD polarization feature vectors into one-dimensional (1D) projection lengths by projecting the feature vectors onto one reference subspace which is chosen to maximize the separation of two types of data. After the transformation, any 1D thresholding technique, such as the Otsu's thresholding, can be applied to perform segmentation efficiently, which results in the reduction of computation complexity in segmentation. The proposed SPS can divide the data into proper homogeneous regions, that is, PolSAR data with similar polarization features being grouped together into regions. In the study, the polarimetric feature vectors are extracted from the coherent/covariance matrices obtained by the polarimetric scattering information. In addition, the referenced projection subspace is selected based on the coherent/covariance matrices of PolSAR data. Finally, the performance of the proposed SPS method is validated by simulations on PolSAR data obtained by NASA Airborne Synthetic Aperture Radar System (AIRSAR) during the PacRim II project and Advanced Land Observing Satellite (ALOS). Simulation results show that the proposed approach can reduce the computational complexity more effectively than other existing methods, and also achieve good classification accuracy.