Rethinking the Optimal Strategy of Deep Learning for Sar Image Semantic Segmentation

Deep learning framework based on transformer have achieved excellent performance in semantic segmentation tasks. However, due to the different properties of SAR image, it is difficult to ensure the rationality and effectiveness if the optical training pipeline is directly transferred to the SAR semantic segmentation task. In this paper, the optimal strategies get rethink when it comes to SAR image semantic segmentation. First, for the imaging process of SAR images, unreasonable data augmentation methods are removed. Second, the implementation and feasibility verification of other optimization strategies, such as Online Hard Example Mining (OHEM), Mosaic and self-training process, are carried out. Finally, comparison experiments demonstrate the effectiveness of the entire system on a large open dataset of built-up areas in SAR Images.