An eigenvalue decomposition-based Gabor network for polarimetric synthetic aperture radar image classification

Textural characteristics of the scattering components of polarimetric synthetic aperture radar (PolSAR) images are investigated for land cover classification in this work. To this end, an eigenvalue decomposition-based Gabor network (EGN) is introduced. EGN is a two-part network with four branches. In one part, three scattering components extracted by an eigenvalue decomposition method are individually inputted to a Gabor filter bank for texture feature extraction. The Gabor feature maps are analyzed by convolutional networks and fused together. In another part, the polarimetric channels are attended by weighting through the normalized Gabor features. Outputs of the two parts are further processed using a shallow convolutional network for fusion of polarimetric, scattering and contextual features. The proposed EGN shows highly accurate classification results with a less or comparable number of labeled samples for the training process when compared to several well-known PolSAR image classification methods.