A Novel Deep Learning Framework for Water Body Segmentation from Satellite Images

Deep learning techniques became crucial in analyzing satellite images for various remote sensing applications such as water body detection. Water body segmentation helps identify and analyze the statistics of various water bodies such as rivers, lakes, and reservoirs. Remote sensing-based real-time water body detection aids in providing a proper response during crises such as floods and course changes in rivers. However, the need for high-resolution multichannel satellite images is the main challenge in achieving a highly accurate water body segmentation. Most water body extraction methods described in the literature use multi-band satellite data that gather extra information from additional bands. However, the lack of such a dataset poses a significant challenge to the analysis. As a result, the research in this field is considerably weaker compared with the other related disciplines. The current study focuses on a research problem for segmenting water body regions from relatively low- to moderate-resolution RGB images without using additional multispectral channels. The overall segmentation pipeline uses a customized encoder–decoder-based convolutional neural network (CNN) model enhanced with attention gates. The architecture uses a customized deep supervised U-Net architecture with residual connections. The network is also integrated with convolution block attention module (CBAM) gates to focus the target region on improving segmentation performance. The main contributions of the proposed model include the use of satellite image sample selection with a customized pre-processing pipeline use of a novel CNN segmentation architecture by integrating various boosting modules such as attention units, deep supervision, and residual blocks. The overall performance is tested with a public satellite dataset, and the results are evaluated on standard benchmark metrics. The segmentation results show improved performance compared with state-of-the-art water body approaches for low-resolution satellite images. The proposed model outperformed the best state-of-the-art model in precision, sensitivity, Dice score, specificity, and accuracy by 3%, 3%, 4%, 5%, and 5% respectively.