Complex-Valued Multi-Viewed UNet - An Application to SAR Image Segmentation

Abstract

At present, the most researches only used the intensity information to do the semantic segmentation of Synthetic Aperture Radar (SAR) images. With the improvement of the resolution of SAR image, the phase information is no longer as invalid as before, an effective use of the phase information can help us to achieve better results in SAR images segmentation. However, most of the current methods that segmented SAR images with complex-valued data are based on full-polarization data. For those dual-polarimetric data, such as Envisat ASAR AP data and Sentinel-1 data, there are no effective methods to make full use of the multi-polarization complex-valued data. Consequently, a SAR image segmentation method named Complex-Valued Multi-Viewed UNet (CVMV-Unet) is proposed, which can make comprehensive use of phase and dual-polarization information. The main contributions of this paper are as follows: first, we propose a multi-view network based on the UNet structure, which can comprehensively use the dual-polarization SAR images to improve land cover segmentation. Second, considering the importance of phase information in high-resolution SAR images, we study the performance of complex-valued neural networks in SAR image segmentation. Limited experimental results based on Sentinel-1 SAR images indicate the CVMV-UNet we proposed can achieve state-of-art classification accuracy in SAR images land cover segmentation.