GMSINet for Building Extraction from Very-High-Resolution Remote Sensing Imagery

Abstract

Building extraction from very-high-resolution (VHR) remote sensing imagery remains challenging because buildings exhibit large variations in scale and are often embedded in complex textured backgrounds. These characteristics increase visual similarity and spatial confusion between buildings and their surrounding environments, making it difficult for building extraction models to achieve both multi-scale representation and detail preservation. To address these issues, we propose the gated multi-scale interaction network (GMSINet), a U-shaped encoder–decoder framework with a hierarchical shifted-window self-attention Transformer backbone. To reduce feature redundancy, background-noise propagation, and unstable multi-scale fusion in skip connections, a gated multi-scale interaction module is designed. This module adaptively filters enhanced information through SimAM-based recalibration, multi-receptive-field interaction, and channel-wise gating, thereby improving feature fusion quality in the decoding stage without changing feature resolution. In addition, a hybrid loss function is introduced to balance pixel-level supervision stability and region-level structural consistency. Experiments on the Massachusetts, INRIA, and WHU building data sets show that GMSINet achieves intersection over union scores of 74.50%, 77.35%, and 90.52%, respectively, outperforming the second-best models by 0.95%, 0.17%, and 0.25%, while maintaining stable overall accuracy and F1 score. These results demonstrate the effectiveness of GMSINet for building extraction from VHR remote sensing imagery.