NDR-UNet: Segmentation of Water Bodies in Remote Sensing using Nested Dense

Abstract

Abstract The identification and accurate delineation of water bodies in remote sensing satellite images have significant implications for scientific research and various applications such as natural disaster forecasting, drought and flood detection, and monitoring disappearing water bodies. However, this task poses challenges due to complex spectral variations caused by factors like aquatic vegetation, different colors of lakes/rivers, mud along the sand, and shadows from surrounding plants. To address these challenges and improve water body extraction from high-resolution and moderate high-resolution remote sensing images, we propose a method called D3net (Nested Dense Residual Network). The Adam optimizer is employed to train the satellite images, minimizing the associated losses. The activation function and the number of nodes in each layer are optimized to achieve the best performance. To ensure data integrity and protect the identified water bodies during transmission, we utilize the Elliptic Curve Digital Signature Algorithm as a security component. This algorithm creates a digital signature for the projected area of water bodies, providing data protection. For our study, we used a dataset consisting of 5682 Sentinel-2 satellite images, including 2841 images and their corresponding masks. The masks were generated using the Normalized Water Difference Index (NWDI) for a specific geographic location in Europe. The suggested model achieves a performance IOU (Intersection over Union) of 93.27% and a recall rate of 95.60%. Additionally, the model can be applied to tasks such as edge detection, blurry image recognition, and low resolution image detection. It exhibits reliability and accuracy in its predictions, although it may require more memory due to the utilization of high-resolution and moderate high-resolution images for segmentation.