PolSAR Target Recognition With CNNs Optimizing Discrete Polarimetric Correlation Pattern

Target recognition plays a crucial role in the intelligent interpretation of synthetic aperture radar (SAR) images. However, polarimetric information holding great potential in target recognition has not been fully studied. In this article, we propose a novel method for target recognition in polarimetric SAR (PolSAR) images by using convolutional neural networks (CNNs) to optimize discrete polarimetric correlation pattern. Discrete polarimetric correlation pattern transfers PolSAR images from image domain to rotation domain, and achieves a high-dimensional representation of the target. We then formulate an optimization problem, which is the basis for target recognition, to unfold the low-dimensional embeddings from the raw representations. The optimization problem aims to achieve intraclass compactness and interclass separation of the target embeddings. Interestingly, we employ CNN as a powerful tool to solve it. By combining the rotation domain features with the neural network, we obtain a discriminative representation that reflects the target's polarimetric scattering mechanism, and finally realize high-performance target recognition. Experiments performed on both simulated and real datasets demonstrate that the proposed method outperforms reference methods significantly in almost all metrics. It is worth mentioning that even on low-resolution images, the proposed method can still achieve high-precision recognition performance. In addition, through feature visualization, we gain deeper insights into the network behavior. Finally, feature separability issue is also discussed, further confirming that the optimized features do have the characteristics of intraclass compactness and interclass separation.