Improving physical correction methods based on topographic information for inundation area mapping using deep learning

Abstract

In this paper, we investigated the refinement of physical correction methods based on topographic information with the aim of improving the accuracy of flood area mapping from satellite images using deep learning. Specifically, we applied corrections based on multiple conditions that account for water flow along the terrain to the results of flood area segmentation from SAR satellite images obtained via deep learning, using high-resolution DEMs of Japan. As a result, using a high-resolution DEM, such as those available in Japan, for correction not only corrected flooded areas that could not be detected by the deep learning method but also reduced false positives, yielding a maximum improvement in IoU of 0.2520. Additionally, the number of missed flooded areas was significantly reduced. On the other hand, using a higher-resolution DEM does not necessarily have a positive effect on the correction; in fact, we also observed that correction accuracy actually decreased when using a 1-meter mesh DEM.