SA-ESegNet: A Shadow Attention-Driven Framework for Accurate Arbitrary-Shaped Water Region Segmentation in Complex Aerial Imagery

Accurate water body segmentation in aerial imagery is a fundamental component of intelligent systems for flood monitoring, urban planning, and environmental management. However, the presence of shadows and surface variations often leads to significant misclassifications. This paper proposes SA-ESegNet, an expert segmentation framework that introduces a novel Shadow Attention Module (SAM) to enhance detection accuracy under challenging illumination conditions. By integrating SAM into the SegNet architecture, the system dynamically prioritizes shadow-affected regions to mitigate false negatives. The model’s performance was rigorously evaluated using a dual-validation approach involving locally acquired high-resolution imagery and independent external aerial datasets. Benchmark comparisons against standard SegNet and Reflection Attention Unit (RAU) variants demonstrate that SA-ESegNet achieves a superior ROC AUC of 0.9637 and 0.9375 on the locally collected aerial imagery dataset and the independently curated public external aerial dataset, respectively. Furthermore, ablation studies and noise-robustness tests confirm that the SAM is critical for maintaining stable performance in diverse and “out-of-domain” environments. The results indicate that SA-ESegNet offers a highly generalized and reliable solution for precise water extraction, making it suitable for deployment in automated, real-world expert systems.