PSRN: Polarimetric Space Reconstruction Network for PolSAR Image Semantic Segmentation

To accurately extract various ground objects from polarimetric synthetic aperture radar (PolSAR) images is a challenging research topic. The deep convolutional neural networks are widely used in SAR segmentation due to their remarkable performance in optical remote sensing images. However, they are still limited by the geometric deformation of the objects, the strong scattering interference among adjacent objects, and the difficulty in distinguishing similar things. A large part of the reason lies in the insufficient utilization and the damage of feature mining for PolSAR data. In this article, focusing on the scattering matrix and the polarimetric coherency matrix, a polarimetric space reconstruction network is proposed. First, to maintain the relatively initial spatial constraints and the complete polarimetric information, the inputs are arranged by a spatial amplification coding method for the polarimetric coherency and scattering matrices. Second, a statistics enhancement module based on scattering characteristics is proposed to mine the differential expression among multiple scattering and polarimetric components, which supplements the local feature representation of convolutional operators. Third, the designed dual self-attention mechanism can capture the amplitude and phase relations of matrix element context adaptively. The proposed method keeps the spatiality of each scattering vector and fully explores the complementary information involved in the fully polarimetric data. Moreover, it accomplishes the accurate land cover classification for PolSAR images, especially the traditional confusing categories, such as water and roads. The experimental results on the E-SAR, AIRSAR, Gaofen-3, and RADARSAT-2 datasets show that the proposed method is effective and promising for PolSAR image segmentation.