Deep learning based methods for water body extraction and flooding evolution analysis based on sentinel-1 images

Abstract

Water body extraction technique has played an important role in water source management and monitoring. In recent years, Threshold based methods, such as Bimodal threshold segmentation (BTS) and maximum between-class variance (OTSU), have widely applied in water body extraction. However, these methods only consider pixel intensity and ignore the spatial correlation among neighboring pixels, resulting in misclassified results. To address this issue, we exploit deep learning based models for water body extraction, which both considers the pixel intensity and spatial correlation among neighboring pixels. Several deep learning based methods, especially Unet, outperform threshold based methods on our hand-crafted dataset acquired from sentinel-l images. The Unet is finally applied in flooding evolution analysis of Xinxiang, Henan province in the summer of 2021, effectively showing the flooding evolution trend.