Deep Learning-based Multi-scale Water Body Segmentation from UAV Imagery

Accurate detection of water bodies across varying spatial scales is essential for hydrological analysis, environmental monitoring, and disease control. Most existing approaches focus on either large- or small-scale water bodies, often failing to achieve consistent performance across scales. In this study, we propose a unified deep learning framework for detecting both large and small water bodies from UAV imagery. The approach enhances the U-Net architecture with residual connections and channel–spatial attention units for improved multi-scale feature extraction and employs a two-stage learning strategy: initial training on large-scale water datasets followed by domain adaptation to small-scale datasets. This enables robust generalization under diverse environmental and spatial conditions. Experimental results show that the proposed model outperforms the baseline UNet by approximately 22% and 14% in segmentation accuracy on large- and small-scale water datasets, respectively, and surpasses DeepLabv3+ by approximately 13% and 9%, respectively. The framework provides a simple yet effective solution for UAV-based multi-scale water body detection, supporting comprehensive water resource management and environmental assessment without the need for additional data or auxiliary tools.