A Deep Reinforcement Learning-Based Framework for PolSAR Imagery Classification

The deep convolutional neural network (CNN) has been extensively applied to polarimetric synthetic radar (PolSAR) imagery classification. However, its success is greatly dependent on numerous labeled samples for revealing and modeling the characteristics of different targets, thus remaining a challenge in maintaining high accuracy in limited sample cases. To address this issue, a deep reinforcement learning (RL)-based PolSAR image classification framework, named deep Q-fully CNN (DQFCN), is proposed in this article. In this framework, two ways are adopted to boost the classification performance while reducing training samples. On the information utilization hand, the spatial neighboring information and polarimetric decomposition information of PolSAR data are both extracted to enrich the feature representation of the sample. Meanwhile, the 3-D CNN architecture is adopted to learn the spatial-polarimetric jointed characteristics simultaneously. On the model learning hand, two RL learning strategies are employed to promote classification performance. The first one is learning from scratch, which does not use any label information as prior knowledge but learns from its self-generated experience. Learning from pretraining is the second strategy in which the networks are sequentially trained from labeled samples and experience data to reduce the time cost. As far as we know, it is the first time that an RL-based fully CNN has been proposed for PolSAR image classification. Experiments on three benchmark datasets prove the effectiveness of the proposed framework, suggesting that the two adopted strategies achieve boosted performance in all experiments, particularly in a limited sample size.