A Robust Multi-Scale Deep Learning Framework for Automated Waterbody Segmentation from Sentinel-2 Multispectral Imagery

Waterbody segmentation using satellite images is an essential step for flood risk assessment, irrigation planning, reservoir management, wetland preservation, and sustainable water resources management. Sentinel-2 multispectral images, which provide spectral and spatial resolution of 10 m, serve as an effective source of information for waterbody segmentation. Nevertheless, reliable segmentation of waterbodies is complicated by seasonality, water level changes, presence of cloud shadows, atmosphere, turbidity, mixture of land classes in one pixel, and spectral similarity between water, dark soils, and dense vegetation. In this study, a robust deep learning approach is proposed for automated waterbody segmentation based on a publicly available Sentinel-2 Water Bodies dataset from the Kaggle repository with Sentinel-2 satellite images and binary water body masks obtained via Normalized Difference Water Index (NDWI). The developed approach employs spectral index bootstrapping using NDWI and Modified NDWI (MNDWI) together with terrain information in order to produce augmented inputs for a multi-scale encoder-decoder neural network. The distillation mechanism is introduced to transfer semantic knowledge from the deeper layers of the network to intermediate layers. Moreover, incorporating uncertainty into the loss function will help diminish the effect of those uncertain pixels related to shadow areas caused by clouds, turbid water, and land-water boundary areas, thus improving the robustness of the algorithm in challenging environments. The techniques used in this research include the application of seasonal domain adaptation and dynamic data augmentation to enhance model generalization to geographical locations and seasons. Experimental evaluation shows that the developed method gives F1-score of approximately 95%, precision more than 95%, and Intersection over Union (IoU) around 90%, which is superior to the conventional spectral index-based models, such as NDWI, MNDWI, and Automated Water Extraction Index (AWEI) and baseline U-Net model.