Enhancing Water Bodies Detection in the Highland and Coastal Zones Through Multisensor Spectral Data Fusion and Deep Learning

Accurate mapping of inland and coastal water bodies is crucial for monitoring environmental changes, managing hydrological resources, and assessing the impacts of climatic variability. This study presents a deep learning-based semantic segmentation framework that leverages multi-band Sentinel-2 imagery for delineating glaciers and coastal lakes. The dataset comprises 400×400-pixel image patches, constructed using false-colour composites of Sentinel-2 Bands 8 (NIR), 4 (Red), and 3 (Green), which enhance spectral separation between water and non-water surfaces. These bands were strategically selected to improve water body contrast and boundary definition through multi-sensor data fusion, enabling more precise lake border extraction. Each image patch is paired with hand-labeled binary lake masks to serve as ground truth. We developed and trained a Simple U-Net in PyTorch and a shallow Convolutional Neural Network (CNN) in TensorFlow to evaluate model performance and architectural efficiency using the same dataset. Both models were assessed using standard performance metrics, including Precision, Recall, F1 Score, and Intersection over Union (IoU). Results show high segmentation accuracy across both platforms (F1 > 0.92, IoU > 0.86). The TensorFlow-based model exhibited faster training and inference, while the PyTorch U-Net provided more consistent and accurate border delineation. This work demonstrates the synergistic power of multi-band spectral fusion and deep learning for environmental feature extraction in remote sensing. The proposed models and methods are scalable and adaptable for broader applications in coastal monitoring, inland water mapping, and climate-related hydrological assessments, offering a valuable contribution to automated Earth observation workflows under changing climatic conditions.