Deep Learning for Inundation Extent Mapping

Abstract:

Flood occurrence is increasing due to the expansion of urbanization and extreme weather like Hurricane; hence, several types of research on this issue and methods of inundation monitoring and mapping are also increasing to reduce the severe impacts of flood disasters. Generating accurate flood maps during and after a flood event is essential to supporting emergency-response planning and providing damage assessments. Recently, deep learning has made rapid progress for remote sensing and photogrammetry classification and mapping applications. Especially in recent years, it is creating massive opportunities that were not possible before. Deep learning brings computers the ability to perform a task that typically requires some level of human intelligence. The study evaluates the potential of different deep learning models such as FCN-16, FCN-8s, SegNet, and U-Net for flood extent mapping. In this study, these deep learning models are fine-tuned to segment the inundation areas. High-resolution UAV imagery collected during Hurricane Matthew (2016) flood events and hurricane Florence event in Lumberton used to train and test these models, respectively. The results show that the FCN achieved a higher overall accuracy ( 97.72%) than other classifiers. The results imply that a deep learning model such as FCN-8s is more efficient than other classifiers for creating a flood extent map.