TNU-Net: Water Body Segmentation based on the Transformer Network with U-Net Model

Satellite imagery is being increasingly examined using deep learning techniques for the purpose of remote sensing applications, particularly in the realm of water body detection. The identification and analysis of rivers, lakes, and reservoirs through segmentation has become possible, thereby enabling the examination of their statistical data. In times of crises such as floods and alterations in river courses, the use of real-time water body detection through remote sensing proves to be immensely beneficial. However, achieving a highly accurate segmentation of water bodies poses a significant challenge, primarily due to the requirement of high-resolution multichannel satellite images. Most of the existing methods in literature rely on satellite data obtained from multi-band satellites to extract water bodies. In contrast, the present study focuses on the segmentation of water body regions using relatively low-resolution RGB images without the utilization of additional multi-spectral channels. To address this issue, a novel approach is proposed, wherein a customized U-Net model equipped with a sequence of transformer networks is employed for segmentation. The proposed model not only incorporates a selection of satellite imagery, but also utilizes a customized preprocessing pipeline and a novel CNN segmentation architecture that integrates various boosting modules, including attention units, deep supervision, and residual blocks. In order to evaluate the overall performance of the proposed model, it is tested against a publicly available satellite dataset, and the results are assessed against standard benchmark metrics. It achieved 93% precision, 95% sensitivity, 92% Dice score, 95% specificity and 96% accuracy. The proposed architecture outperforms even the most high-performing baseline.The segmentation results obtained demonstrate a notable enhancement in performance when compared to the state-of-the-art approaches employed for the segmentation of water bodies from low-resolution satellite imagery.