Development of a Lightweight SAR Semantic Segmentation Framework with ASFEM for Monitoring Pastoral Ecosystems

Abstract Synthetic Aperture Radar (SAR) provides reliable land-cover observations under varying illumination and weather conditions, but semantic segmentation remains challenging because of speckle noise and heterogeneous backscatter. This study adapts the lightweight PP-LiteSeg architecture to Sentinel-1 SAR semantic segmentation and introduces an Adaptive Spatial Feature Enhancement Module (ASFEM) to improve spatial feature representation. ASFEM combines convolutional processing, Coordinate Attention, depthwise convolution, and residual learning with limited additional computational cost. Experiments on the TurkeySAR dataset show that the adapted PP-LiteSeg baseline achieves 69.45% mIoU, while PP-LiteSeg-ASFEM improves mIoU to 70.70%, with an overall accuracy of 89.67%. The proposed module increases the parameter count by only 0.17% and the computational cost by approximately 0.06%, while maintaining an inference speed of 187.81 FPS. These results demonstrate that PP-LiteSeg can be effectively adapted to SAR imagery and that ASFEM improves its segmentation capability while largely preserving the efficiency of the original lightweight architecture. Keywords Synthetic Aperture Radar (SAR); Sentinel-1; semantic segmentation; land-cover mapping; PP-LiteSeg; ASFEM; Coordinate Attention; lightweight deep learning; remote sensing