Guided Ridge Regression-Based Polarimetric-Spatial Feature Extraction for Classification of Polarimetric SAR Images

Synthetic aperture radar (SAR) acquires high resolution images containing rich spatial information. The Polarimetric SAR (PolSAR) images are a good source of polarimetric and spatial information appropriate for land cover classification. Two PolSAR image classification methods are introduced in this work: ridge regression-based polarimetric-spatial (RRPS) and the guided RRPS. The RRPS feature extraction method generates polarimetric-spatial features with minimum overlapping and redundant information. To this end, each polarimetric-spatial channel of the PolSAR data is represented through a ridge regression model using the farthest neighbors of that channel. The weights of the regression model compose the projection matrix for dimensionality reduction. The guided RRPS method uses the weights of the guided filter to revise the probability maps corresponding to the initial classification map. The proposed RRPS and guided RRPS methods show high performance for PolSAR image classification in small sample size situations.