Enhanced U-Net with Boundary Aware Modules for Waterbody Area Segmentation of Satellite Images

Satellite imaging is essential for environmental monitoring, resource management, and disaster response, especially in the identification and analysis of waterbody regions. However, challenges such as uneven waterbody configurations, low contrast with adjacent terrain, and fluctuating spectral characteristics due to environmental influences hinder accurate segmentation. This study presents an enhanced U-Net architecture, attaining notable performance enhancements with average metrics of 96.1% recall, 97.2% F1 score, 94.3% Jaccard index, 97.7% accuracy, 99.0% specificity, and 97.2% Dice coefficient. The proposed method integrates preprocessing, and sophisticated edge-detection techniques to enhance segmentation precision. Boundary refinement blocks and dense connections augment the network's capacity to identify and outline complex waterbody edges. Experiments using publicly accessible datasets illustrate the model's resilience and versatility across many environmental situations, affirming its reliability as a tool for water resource management and conservation.