PGMMF-Net: a POI-Guided Multi-modal Mamba Fusion Network for urban impervious surface extraction from optical and SAR imagery

Accurate extraction of urban impervious surfaces (IS) is essential for urban environmental analysis. However, optical-based methods are affected by shadows and spectral confusion, while optical-SAR fusion lacks urban functional semantics. Point-of-interest (POI) data provide complementary semantic information but are difficult to integrate because of heterogeneous representations and semantic noise. To address these issues, we propose a POI-Guided Multi-modal Mamba Fusion Network (PGMMF-Net). The network employs a multi-branch encoder-decoder architecture based on a State Space Model (SSM) to capture long-range dependencies. An adaptive fusion strategy and cross-modal interaction mechanism effectively integrate optical, SAR, and POI features, while a boundary-aware enhancement mechanism improves boundary localization. Experiments on a Wuhan multi-source dataset achieve 0.8957 MIOU, 96.48% OA, and 0.9439 F1-score, outperforming nine representative methods. The results demonstrate that integrating social sensing information with multi-source remote sensing data significantly improves IS mapping accuracy.