LMDE-UNet: local multi-directional state space modeling with dynamic edge enhancement for building extraction from remote sensing images

Abstract Accurate building extraction from remote sensing imagery requires both contextual information and local structural detail. This paper presents LMDE-UNet , an encoder–decoder network that combines window-based local multi-directional scanning (WLMS), a bottleneck feature aggregation module (BFAM), and an edge enhancement module (DEEM) in the skip connections. Each local visual selective scan block models global context before applying local scanning within non-overlapping windows. Experiments on three building-extraction datasets show that LMDE-UNet achieves higher intersection over union (IoU) and F1-scores than the evaluated baselines, with competitive precision and recall. On the Typical Chinese Urban Building Dataset, LMDE-UNet achieves an IoU of 73.47% and an F1-score of 85.16%, exceeding the corresponding SwinUNet scores by 0.99 and 0.63 percentage points, respectively. Its precision reaches 86.73%. The ablation results suggest that WLMS provides larger IoU gains when combined with BFAM and DEEM than when added alone.