Polarimetric synthetic aperture radar (PolSAR) image classification is a challenging task due to lack of effective feature representation approaches. However, when it comes to deep feature representation, there is little research that pays enough attention to make full use of existing expert knowledge. We propose a model for deep polarimetric feature extraction, and a superpixel map is used to integrate contextual information. The proposed model uses multiple polarimetric algebra operations, polarimetric target decomposition methods, and convolutional neural network (CNN) to extract deep polarimetric features. The core idea is to utilize expert knowledge of the target scattering mechanism interpretation to assist the CNN classifier in feature extraction and employ superpixel algorithm based on Wishart distribution to improve the final classification performance. The proposed method is able to get abstract and discriminative representation from initial features, achieving robust and improved performance for PolSAR images. Compared with other state-of-the-art methods, experiments on the classification of three real PolSAR datasets are performed to demonstrate the superiority of the proposed approach.

Superpixel-level image classification methods take advantage of contextual information of pixels and reduce the time cost in training and test processes. However, extracting a superpixel-level feature is a challenging task because each superpixel has irregular size and shape. A superpixel-level polarimetric feature extraction (SLPFE) based on circular loop spatial pyramid pooling (CLSPP) (SLPFE_CLSPP) is proposed for polarimetric synthetic aperture radar (PolSAR) image classification. The main idea is that the CLSPP layer divides the feature maps of each superpixel into the same number of circular loops and produces consistent feature size for each superpixel, which then feeds into the next processing step, such as classification. SLPFE_CLSPP not only makes it possible to generate the same size feature but also to utilize the spatial information of pixels and reduce the running time during SLPFE process. Compared with the existing methods, the proposed method makes a balance between overall accuracy and running time. Experimental results on real PolSAR dataset demonstrated the superiority of the proposed approach.