Remote sensing water body extraction method based on weak information dynamic perception

As a core task in remote sensing image analysis, water body extraction is vital for ecological monitoring, flood early warning, and water resource management. Recent AI-based remote sensing methods for water body characterization have advanced significantly but struggle with weak-information scenarios (multi-scale small water bodies, indistinct boundaries, limited representation). This paper proposes a Data-driven Weak Information Dynamic Perception Network (DWIDP-Net) for high-precision multi-scale water body extraction. It adopts a data-driven, memory-coupled multi-scale framework, integrating Transformer’s global semantic modelling and MobileNet-v2’s lightweight local extraction to adapt to complex water morphologies. Two key modules are designed: Multi-Modal Feature Alignment Module (MAM, optimizing cross-scale consistency via mask-guided residual reconstruction) and Weak Information-Guided Feature Enhancement Method (WIGFEM, addressing feature degradation with gradient-sensitive masks and cascaded fusion). Experiments on WaterFormer and LoveDA_Water datasets show DWIDP-Net achieves IoU (96.16%/78.27%) and Recall (97.45%/86.23%), outperforming mainstream models by 1.06%–20.43% (Recall) and 1.12%–15.81% (IoU), providing technical support for flood monitoring and water management.