Abstract—Satellite imagery is increasingly being scrutinized through deep learning methodologies for remote sensing applications, particularly focusing on the detection of water bodies. The ability to identify and analyze rivers, lakes, and reservoirs through segmentation has now become feasible, enabling the exploration of their statistical information. During crises such as floods and changes in river pathways, real-time detection of water bodies via remote sensing proves to be highly advantageous. Nevertheless, achieving a precise segmentation of water bodies presents a notable challenge, mainly due to the necessity of high-resolution multi-channel satellite images. The existing literature predominantly relies on satellite data from multi-band satellites for water body extraction. Conversely, this current research emphasizes the segmentation of water body regions using relatively lower-resolution RGB images without the incorporation of extra multi-spectral channels. To tackle this challenge, a unique methodology is suggested, involving a customized U-Net model integrated with a Time-Frequency Attention network (TFAU-Net) for segmentation. To assess the comprehensive performance of the proposed model, it is evaluated against a publicly available Sentinel-2 satellite dataset, and the outcomes are compared against standard benchmark metrics. The model achieved a precision of 94%, sensitivity of 96%, Dice score of 93%, Mean IoU of 85%, and accuracy of 97%. The proposed architecture surpasses even the most high-performing baseline. The segmentation results obtained exhibit a significant improvement in performance compared to the cutting-edge methods used for water body segmentation from low-resolution satellite images.