Scale-Aware Fusion and Spatial-Frequency Collaborative Network for Remote Sensing Imagery Semantic Segmentation

Abstract

Semantic segmentation of remote sensing images is crucial for serial earth observation tasks. However, significant scale variations in remote sensing scenes and insufficient exploitation of frequency-domain information often cause small-scale objects to be overwhelmed by large backgrounds under imbalanced multiscale features, while missing high-frequency boundary details lead to boundary–background confusion, ultimately degrading segmentation accuracy. To address these issues, this article proposes a scale-aware fusion and apatial-frequency collaborative network (SFSNet). Specifically, a scale-aware fusion module (SAFM) is designed to enhance multiscale feature representation through cross-level feature interaction, multireceptive-field calibration, and cross-stage channel fusion, thereby alleviating the suppression of small-scale objects by background regions as well as the imbalance among multiscale features, and ultimately improving segmentation accuracy. In parallel, a spatial-frequency collaborative enhanced attention module (SCEAM) is developed to strengthen high-frequency components in the frequency domain for sharper boundary details, while introducing global constraints in the spatial domain to improve regional consistency, thus enhancing segmentation quality for complex boundaries. Extensive experiments on the Vaihingen, Potsdam, and LoveDA datasets demonstrate that the proposed SFSNet achieves excellent performance in remote sensing image segmentation tasks.