NDR U-Net: Segmentation of Water Bodies in Remote Sensing Satellite Images Using Nested Dense Residual U-Net

The utilization of remote sensing satellite imagery is crucial for advancing scientific knowledge in identifying water bodies.This technology has diverse applications, including natural disaster forecasting, as well as the detection of droughts and floods.Currently, there are few programs available specifically designed to locate naturally disappearing water bodies.Accurately delineating aquatic features from Very high-resolution (VHR) remotely sensed and moderately high-resolution images poses a significant challenge in remote sensing.The intricate spectrum combinations resulting from aquatic vegetation, various lake/river colours, mud along the sand, and shadows from surrounding plants make it difficult to determine water body boundaries.To enhance water body extraction from VHR and moderately high-resolution remote sensing images, it is imperative to increase feature variety and semantic data.In this study, we employed the nested dense residual network, known as D3net, to detect water bodies in satellite images.The Adam optimizer was utilized for training the satellite images, minimizing losses in the process.The optimizer also identified the optimal values for the activation function and the number of nodes in each layer.For segmentation, 5682 Sentinel-2 satellite images (comprising 2841 satellite images and their respective masks) were used, focusing on the Europe geo location.These masks were generated using the Normalized Water Difference Index (NWDI).The proposed model demonstrated a performance of 95.36% Intersection over Union (IOU) and 96.99% accuracy, making it suitable for edge detection, blurry image recognition, and lowresolution image detection.While the model is reliable and accurate in its predictions, it requires more memory due to the utilization of VHR and moderately high-resolution images for segmentation.