Water Body Extraction from Sentinel-2 Imagery with Deep Convolutional Networks and Pixelwise Category Transplantation

A common task in land-cover classification is water body extraction, wherein each pixel an image labelled as either or background. Water detection integral to the field of urban hydrology, with applications ranging from early flood warning resource management. Although traditional index-based methods such Normalized Difference Index (NDWI) and Modified (MNDWI) have been used detect bodies for decades, deep convolutional neural networks (DCNNs) recently demonstrated promising results. However, training these requires access large quantities high-quality accurately data, which often lacking remotely sensed imagery. Another challenge stems fact that category interest typically occupies only a small portion thus grossly underrepresented data. We propose novel approach data augmentation—pixelwise transplantation (PCT)—as potential solution both problems. Experimental results demonstrate PCT’s ability improve performance on variety models datasets, achieving average improvement 0.749 mean intersection over union (mIoU). Moreover, PCT enables us outperform previous high score achieved same dataset without introducing new model architecture. also explore suitability several state-of-the-art segmentation loss functions extraction. Finally, we address shortcomings works by assessing RGB, NIR, multispectral features ascertain relative advantages approach. In particular, find significant benefit inclusion bands, outperforming visible-spectrum 4.193 mIoU.