BECSL: boundary-enhanced consistency semi-supervised learning model for water extraction from remote sensing images

Accurate water body identification and extraction presents a critical challenge in geographic information systems, particularly for boundary-sensitive applications. While deep learning offers promising solutions for automated geospatial analysis, most semi-supervised methods inadequately model edge information. This study introduces a Boundary-Enhanced Consistency Semi-supervised Learning (BECSL) framework to address this gap. Our approach integrates Segment Anything Model (SAM) with OpenStreetMap (OSM) data to create multi-modal training datasets. The framework employs a dual-decoder architecture combining reverse attention and boundary enhancement mechanisms. This design generates refined pseudo-labels for unlabeled data supervision. We further develop a self-contrast strategy targeting regions with boundary prediction inconsistencies. Comprehensive evaluation on 2024EarthVQA and custom datasets demonstrates our method's effectiveness. The framework achieves superior performance using merely 10% labeled data while maintaining precise boundary delineation. This work provides both theoretical and practical advances in resource-efficient water extraction from remote sensing images.