Fully Polarized SAR imagery Classification Based on Deep Reinforcement Learning Method Using Multiple Polarimetric Features

Most traditional supervised classification methods for polarimetric synthetic aperture radar (PolSAR) imagery require abundant manually selected samples, and the classification results are affected by the size and quality of the samples. In this paper, we propose an improved deep Q-network (DQN) method for PolSAR image classification, which can generate amounts of valid data by interacting with the agent using the ε-greedy strategy. The PolSAR data are first preprocessed to reduce the influence of speckle noise and extract the multi-dimensional features. The multi-dimensional feature image and corresponding training image are then fed into a deep reinforcement learning model tailored for PolSAR image classification. After many epochs of training, the method was applied to identify different land cover types in two PolSAR images acquired by different sensors. The experimental results demonstrate that the proposed method has a better classification performance compared with traditional supervised classification methods, such as convolutional neural network (CNN), random forest (RF), and L2-loss linear support vector machine (L2-SVM), and also has a good performance compared with the deep learning method CNN-SVM, which integrates the synergy of the SVM and CNN methods, especially in small sample sizes. This study also provides a toolset for the DQN (kiwi.server) on the GitHub development platform for training and visualization.