Fully PolSAR image reconstruction for enhanced land cover mapping

Polarimetric Synthetic Aperture Radar (PolSAR) images provide richer ground object information than optical images due to their full-polarization capabilities. However, annotating PolSAR images is time-consuming, limiting the available research data and increasing the risk of overfitting. Addressing the scarcity of labeled data is crucial for land cover mapping. For this, we propose a PolSAR image reconstruction method called PolSAR to RGB (P2R), which expands the dataset by generating reconstructed RGB images from original PolSAR images. While the reconstructed images augment the dataset, they still contain lots of speckle noise, leading to lower land cover mapping performance. To alleviate this defect, we also design a Gate Denoised Module (GDM) consisting of two components: Image Denoising guided by Original Polarization (IDOP) and Feature Generation with Gated Unit (FGGU) module. Specifically, the IDOP module reduces the impact of noise on the feature map by learning similar characteristics between the original polarization and reconstructed RGB images. The FGGU module adaptively fuses the low-level feature map with the denoised feature map by a weighted gate, which helps remove the noise in images without affecting their details. Experiments on several publicly available PolSAR datasets show the effectiveness of our land cover mapping method, reaching 53.94% mIoU in AIR-PolSAR-Seg, 76.71% mIoU in PolSF-RS2, and 90.34% mIoU in PolSF-GF3.