We will build a deep learning classification technique on a temporal sequence of geo-registered remote sensing images to extract the dynamics of flooding extent during extreme weather events.
JGJonathan L GoodallSchool of Engineering and Applied Science
GHGuoping HuangSchool of Architecture
DHDevin K. HarrisSchool of Engineering and Applied Science
The goal of this research is to estimate the geographic extent of flooding over large regions using remote sensing imagery and deep learning algorithms capable of change detection. Building upon prior work that utilized shallow neural networks for change detection and a multi-temporal operator to create a change mask, we will use deep learning classification techniques created for infrastructure monitoring applications on a sequence of geo-registered temporal satellite images to build a change detection system for flood mapping. The results of the technique will then be used to estimate a water depth map through the fusion of topography data. This result can be used to calibrate and evaluate state-of-the-art flood models to provide more accurate flood warnings and flood risk assessments. This will be one of the first applications of deep learning incorporating spatial, spectral as well as temporal dimensions as a remote sensing application for flood estimation and monitoring.
The research will result in a prototype system for classification of flooding extent from remote sensing imagery. The system will be tested using image data resulting from recent hurricanes in the United States, likely from Hurricane Harvey in particular. The classified imagery can be used to better assess flood dynamics and, when combined with detailed topographic data within a geographic information system, give important volumetric water data useful in hydrologic model calibration and evaluation. The use of drones for post-disaster data and imagery collection opens the door to more extensive flood extent mapping using techniques like the one that will be created through this research. The flooding extent data can also be combined with infrastructure layers to identify flood impacts across large geographic regions. The preliminary results generated through this research will form at least one proposal to an external funding agency. Potential agencies that we believe will be interested in funding this research include the Virginia Transporation Research Council, the National Science Foundation, or potentially the Federal Emergency Management Agency.
At least 50% of the funding, and we expect a much higher percentage, will go toward funding graduate and undergraduate students to work on the project as research assistants. We have students in our research labs now that are well positioned to work on the project but will also recruit other students to be a part of the team.