Investigating Accurate Water Body Extraction from Satellite Imagery Using Convolutional Neural Network with Water Indices

Water bodies like rivers and lakes are essential human, animal, and vegetation resources. Acquiring and analyzing these data are essential for better water resources and environmental management. In recent decades, many methods have been used to extract water bodies from remote sensing images. A critical way was using water indices such as normalized difference water index (NDWI) and modified normalized difference water index (MNDWI). Furthermore, extracting small water bodies (such as creeks) from satellite images is challenging and needs high spatial resolution images. However, there are several difficulties in using water indices in high spatial resolution images because of the limited spectral resolution. Most high spatial resolution satellites lack SWIR, which is necessary to calculate MNDWI. But, WorldView-3 overcomes this problem by having NIR and SWIR bands with high spatial resolution. This study used a Kaggle dataset called DSTL, which includes WorldView-3 images for the same region (but the dataset's provider obscures the location of these images) with 0.31 m, 1.24 m, and 7.5 m spatial resolution at nadir for panchromatic, multi-spectral, and SWIR, respectively. A convolutional neural network (CNN) was proposed to improve water body extraction from satellite images. In this network, three indices (normalized difference vegetation index (NDVI), NDWI, and MNDWI) were input layers to U-Net architecture. To evaluate this approach, the results were compared with the same U-Net architecture using RGB and all WorldView-3 bands as input layers, SegNet, and FCN-8. The Jaccard index was calculated as 90.7%, 90%, 89.3%, 85%, and 80% for U-Net with three indices, RGB, all WorldView-3 bands, SegNet, and FCN-8, respectively. The results show that using indices as input layers to U-Net improves the water body extraction, which is necessary to be accurate in several hydrological and water resources studies.