Multi-scale Adaptive Semantic Segmentation Network for Remote Sensing Images based on Attention Gating Mechanism

In this paper, a multi-scale adaptive semantic segmentation model for remote sensing images—AG-MSASNet—is proposed to address the imprecise utilization of feature information in multi-scale feature extraction. The model introduces a novel attention-gated multi-scale feature fusion mechanism, which selectively integrates multi-scale features to enhance useful information while suppressing noise, thereby improving segmentation accuracy. To further enhance semantic understanding in complex remote sensing scenes, a Global Context Block (GC Block) is designed to capture long-range dependencies and aggregate global context. Additionally, a joint loss optimization strategy is proposed to balance learning across different scales and classes, contributing to more precise and robust segmentation. Experimental results on the LoveDA dataset demonstrate that the proposed model outperforms the baseline, achieving a $\mathbf{2 . 2 \%}$ improvement in mean intersection over union (mIoU) and a $4.1 \%$ increase in overall accuracy (OA).