SAR Image Urban Scene Classification based on an Optimized Conditional Generative Adversarial Network

Classification of urban scenes in SAR images has been challenging in the complex and various behaviors of urban areas in SAR images. Generative Adversarial networks (GANs), which can learn distribution characteristics of image scenes in an alternative training style, have attracted attention. However, these networks all suffer from difficulties in network training and stability. To overcome this issue and extract proper features, this paper introduces two effective solutions. Firstly, employing the residual structure and an extra classifier into the traditional conditional adversarial network to achieve scenes classification. Then gradient penalty is used to optimize the loss convergence during training stage. Last, we select GF-3 and Sentinel-1 SAR images to test the network. The experiment results show the usefulness of our proposed optimization.