FSA-SARNet: A Frequency-Guided Semantic Segmentation Network for Remote Sensing SAR Images

Synthetic Aperture Radar (SAR) image semantic segmentation is crucial for all-weather, high-precision land-cover mapping. However, existing segmentation methods have limited performance due to their inability to resist well the speckle noise inherent in SAR images and neglecting the importance of frequency domain information. In this letter, we propose FSA-SARNet, a lightweight encoder-decoder network based on Frequency-Guided Attention. The main innovations of our method are as follows: (1) We design a Log-domain Guided Denoising Module (LGDM) to denoise by transforming the multiplicative speckle noise to additive noise. (2) We design a Frequency-Guided Spatial Attention Module (FGSA) to enable frequency-domain features to guide spatial-domain feature learning. (3) We introduce dynamic upsampling for the decoder to recover feature map details. Evaluated on the FUSAR-Map and WHU-OPT-SAR datasets, our method delivers superior performance and efficiency, achieving competitive MIoU scores of 42.67% and 47.00% with a compact model size of only 7.03M parameters.