Deep learning-based semantic segmentation for surface water extraction from Sentinel-2 imagery: Case study of Kuş and Uluabat lakes, Türkiye

Abstract

This study presents a deep learning-based approach for high-precision surface water extraction from Sentinel-2 multispectral imagery. A modified U-Net architecture, trained and evaluated on two Turkish lake systems (Kuş and Uluabat Lakes), achieved superior performance compared to traditional methods. The model attained an overall accuracy of 0.9980, precision of 0.9980, recall of 0.9980, F1-score of 0.9980, and Intersection over Union (IoU) of 0.9961, outperforming both Normalized Difference Water Index (NDWI) and Modified NDWI (MNDWI). Analysis reveals that the U-Net effectively mitigates spectral confusion in heterogeneous environments, demonstrating its potential for enhanced water resource monitoring, flood mapping, and hydrological modeling applications. While NDWI and MNDWI achieved IoU scores of 0.9956 and 0.9953, respectively, the deep learning model's higher IoU signifies more accurate boundary delineation. The improved performance highlights the value of deep learning in automated surface water mapping for enhanced decision-making in water resource management. These results suggest that while traditional spectral indices are useful for preliminary analysis, deep learning approaches offer a more refined classification, particularly in complex or heterogeneous landscapes.