Complex-valued U-Net for PolSAR Image Semantic Segmentation

As an image semantic segmentation network, U-Net has the advantage of simple structure which is suitable for semantic segmentation of PolSAR images with small datasets. However, the original U-Net is a real-valued (RV) network, whose input must be RV. If it is directly used in the segmentation of PolSAR image, the complex-valued (CV) input must be converted into RV, which results in the loss of information. In this paper, a CV U-Net, which is mathematically strict, is proposed for semantic segmentation of PolSAR images. Considering that the PolSAR dataset is small, the structure and parameters of CV U-Net are furtherly simplified based on the original U-Net to prevent overfitting. Experimental results on the Flevoland dataset show that the proposed CV U-Net has better segmentation performance than the original RV U-Net and some other semantic segmentation networks.