RFPN-BOM: a reinforced FPN architecture with boundary optimization for small water body extraction

This paper addresses the need for accurate extraction of small water bodies from remote sensing imagery, a critical prerequisite for water resources analysis that underpins resource planning, sustainable management, and environmental protection. However, existing extraction models often suffer from detail loss, misclassification, and coarse boundaries. To overcome these limitations, this paper proposes a novel architecture that combines a Feature Pyramid Network (FPN) with a Boundary Optimization Model (BOM) to improve prediction accuracy and preserve the continuity of boundary pixels. The proposed method first constructs a reinforced FPN; adopts the Deformable Attention Transformer Encoder, path aggregation network and Feature Fusion Module to enhance the feature representation ability of small water bodies from global features, low-dimensional features and fusion features. Subsequently, the method adopts a BOM that combines morphological corrosion and expansion to optimize the edges of small water bodies, achieving precise boundary segmentation and enhancing the extraction accuracy of the features of small water bodies. Experiments on the GF-2 dataset show that the model can achieve more complete and accurate extraction of small water bodies, with the precision rate, F1 score and Intersection over Union reach 95.8%, 92.7% and 91.5% respectively. It has been proved that the method proposed in this paper provides a reliable solution for environmental monitoring and water resource management. The experimental results prove that the proposed method can offer a reliable solution for environmental monitoring and water resource management.