Complex-Valued U-Net with Capsule Embedded for Semantic Segmentation of PolSAR Image

In recent years, semantic segmentation with pixel-level classification has become one of the types research focus in field polarimetric synthetic aperture radar (PolSAR) image interpretation. Fully convolutional network (FCN) can achieve end-to-end segmentation, which provides a basic framework for subsequent improved networks. As classic FCN-based network, U-Net been applied to remote sensing images. Although good results have obtained, scalar neurons made it difficult obtain multiple properties entities image. The vector used capsule effectively solve this problem. paper, we propose complex-valued (CV) CV embedded PolSAR structure is lightweight match small data, and designed extract more abundant features than U-Net. Furthermore, dynamic routing proposed realize connection between capsules two adjacent layers. Experiments on airborne datasets Gaofen-3 dataset show that capable distinguishing different land covers similar scattering mechanism extracting complex boundaries covers. achieves better performance other state-of-art networks, especially when training set size small.