A Region-Based Deep Reinforcement Learning Classification Method for GF-3 PolSAR Imagery Classification

The C-band GF-3 satellite is the first fully polarimetric synthetic aperture (PolSAR) designed by China, which has multi-polarization image modes and been used in various applications. Land cover classification is an effective approach in PolSAR image interpretation and further application. However, conventional classification methods are mostly pixel-based and are easily affected by inherent speckle noise. In addition, the feature selection of PolSAR image and the amount of training samples are also critical to the performance of classifiers. To solve these problems, in this paper, we propose a region-based PolSAR image classification method, which uses reinforcement learning method altogether with statistical region merging algorithm to improve the classification performance. The contributions of our method are mainly reflected in three aspects: First, the T3 matrix is considered as the only feature set in our method, including image segmentation and classification. Second, the region is produced via statistical region merging algorithm. Finally, a deep reinforcement learning model is used to obtain PolSAR image classification result. To evaluate the performance of the proposed method, two real GF-3 images are performed in the experiments, and the experimental results illustrate that the proposed method outperforms the conventional methods (support vector machine, random forest, and convolution neural network) in terms of accuracy and achieves the state-of-art results.