Flood Extent Mapping in 3D Using Deep Learning from High-Resolution Remote Sensing Data

• Main Author: Gebrehiwot, Asmamaw Amare
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2021
• Subjects: Artificial intelligence; Geographic information science; Remote sensing
• ISBN: 9798516942440

Flooding is one of the greatest threats of natural disasters to human life and property, especially in densely populated urban areas. Real-time and precise floodwater extent and depth are vital to supporting emergency response planning and providing damage assessment in spatial and temporal measurements. This study investigates the potential of deep learning algorithms such as Convolutional Neural Network (CNN) for automatically generating flood extent maps using high-resolution optical imagery. Pre-trained CNN based models were fine-tuned using Unmanned Aerial Vehicle (UAV) datasets that were collected during Hurricane Matthew and Florence flooding events and manually annotated. Although the dataset contained only one hundred training samples, the classification results were very promising in the open area visible of the imagery. Detecting floods underneath vegetation canopies is challenging to implement with optical images because the view is obscured by the forest. This research attempts to address this problem by proposing an integrated CNN and Region Growing (RG) method to map both open and underneath vegetation canopies flooded areas, which is essential to supporting effective flood emergency response and recovery activities. The research also proposed two methods for inundation depth estimation by integrating deep learning-based flood extent and Digital Elevation Model (DEM). The methods were compared and validated using the United States Geology Survey (USGS) gauge water level data acquired during the flood event. Experimental results showed that the deep learning method could precisely extract inundation areas from the remote sensing images in comparison with the conventional algorithms. In addition, the proposed approaches achieved significant results in estimating floodwater depth.