Deep Supervised U-Net++ for Semantic Segmentation of Water Bodies in Satellite Imagery

Accurate extraction of water bodies from satellite imagery is necessary for hydrological evaluation, environmental monitoring, and sustainable resource management. In this paper, we propose a deep learning method for semantic segmentation based on a U-Net++ architecture and deep supervision to detect waterbodies from RGB images. Our pipeline includes data preprocessing, augmentation and training on a well-tailored dataset of RGB images and the corresponding binary masks of underwater scenes. Images and masks were downscaled to 128×128 pixels and normalized. The model uses a multilayer U-Net with dense skip connections and multiple supervised outputs, allowing for solid training and better boundary localization. A hybrid loss function that included the Dice Coefficient and Binary Cross-Entropy (BCE) loss function was used to compute and equalizes region-wise and pixel-wise learning. The Adam optimizer was used for training along with callback functions to ensure convergence and avoid overfitting. Across a 20-epoch experimental data, Dice Coefficients increased from 0.49 to 0.53 and trended normal or stable for validations, while Mean Intersection over Union (IoU) converged time and finally reached around 0.38. Despite the moderate IoU, the model was able to segment out aquatic zones with class consistency for the training and validation stages.