EFMSANet: an edge-fused multi-scale attention network for building extraction from remote sensing images

As the basic component of urban space, the accurate extraction of buildings is the key basis for the application of geographic information. In complex urban scenarios, extracting buildings from high-resolution remote sensing images often faces issues such as blurred boundaries, severe background interference, and insufficient multi-scale feature fusion. To solve these problems, an Edge-Fused Multi-Scale Attention Network (EFMSANet) is proposed in this paper. The network achieves this through a synergistic integration of explicit edge guidance, multi-scale feature fusion, and attention-enhanced feature encoding. An edge detection branch based on dilated convolutions and residual mapping extracts boundary probability maps from the input image, which are then adaptively integrated into decoder features through a spatial-channel dual-attention fusion module that selectively enhances boundary regions while suppressing background interference. A Dual-Branch Multi-Scale Feature Fusion Module (DB-MSFF) captures both global context and local detail in parallel to accommodate buildings of varying scales. In building extraction experiments on Satellite Dataset II, the Aerial Imagery Dataset, and the Massachusetts Buildings Dataset, IoU reached 64.1662%, 88.1049%, and 70.2682%, and F1 scores were 78.1722%, 93.6763%, and 82.5382%, respectively. In the boundary extraction experiments, compared with existing methods, the proposed EFMSANet achieves IoU values of 50.9732%, 83.9798%, and 72.8380% on the three datasets, respectively, which outperforms the suboptimal methods by 1.76, 2.51, and 0.61% points and obtains the optimal results. Experimental results show that the proposed method can achieve high-precision and robust building segmentation in complex urban scenarios.