CEA-SAM: Context-Enhanced Adaptation of Segment Anything Model for Robust Water Body Extraction from UAV Imagery

Extracting water bodies from Unmanned Aerial Vehicle (UAV) imagery is challenging due to complex boundaries, reflections, and occlusions. While Convolutional Neural Networks (CNNs) suffer from limited receptive fields yielding discontinuous results, large foundation models like the Segment Anything Model (SAM) lack fine-grained perception and are costly to adapt. To overcome these issues, we propose CEA-SAM, a context-enhanced and adapted framework. CEA-SAM first leverages a Spatial Feature Enhancement Module (SFEM) to amplify boundary details, compensating for the SAM encoder's limitations. It then employs efficient Weight-Decomposed Low-Rank Adaptation (DoRA) to specialize the powerful SAM encoder for the water extraction task. Finally, a decoder-encoder fusion strategy integrates high-level semantics with spatial details to refine segmentation masks. Experiments on the Varied Drone Dataset (VDD) validate our approach, achieving state-of-the-art performance with a 94.24 % Intersection over Union (IoU) and $\mathbf{9 7. 0 3 \%}$ F1-Score. Notably, our model attains the highest Recall of $\mathbf{9 8. 1 9 \%}$, demonstrating superior robustness in preventing missed detections. Our work presents an effective paradigm for adapting large vision models to specialized remote sensing applications.