A Novel Deep Learning Approach for Extraction of Water Bodies using Satellite Imagery

This paper presents SegxResNet, a novel deep learning framework for the precise extraction of water bodies from satellite imagery. The primary challenge in water body segmentation lies in the variability of their size, shape, and spectral properties, which often leads to information loss in conventional models due to max-pooling operations. To overcome this, SegxResNet integrates SegxNet and ResNet-50 architectures, leveraging their strengths to improve segmentation accuracy. The methodology involves an encoder-decoder structure, where SegxNet processes the input, ResNet-50 extracts deep features, and the SegxNet decoder generates the final segmented output. The accuracy of the segmentation is evaluated against ground truth images, with an inverse segmentation refinement step further enhancing precision. Experimental results demonstrate that SegxResNet significantly improves segmentation accuracy, outperforming traditional models and highlighting its potential for enhanced remote sensing applications in water body extraction.