1 Research Background
As both the intensity and frequency of storm events are projected to increase due to climate change, local agencies are in urgent need of accurate flood forecasting so that they can better protect people’s lives and infrastructures in urban areas. With access to better quality hydrologic, topographic, and meteorological data which have much finer spatial and temporal resolutions than ever, we are able to improve flood forecasting with more accurate estimations of the peak flood quantity and the time to peak flood from the initiation of a storm event. The increasing coverage of 1-meter Lidar digital elevation model (DEM) in the United States enables detailed delineation of stream geomorphology; Hourly hydrologic observations (e.g., stream discharge and stage) at United States Geological Survey (USGS) gages provide sufficient data to analyze the patterns of flood events at multiple locations; Various satellite-based data sets from the National Aeronautics and Space Administration (NASA) and the European Space Agency (ESA) can provide additional hydrometeorological data (e.g., areal average soil moisture, evapotranspiration, vegetation conditions, etc.) that can potentially promote flood forecasting in the future. Recently, a new framework, GeoFlood, was proposed1 to map inundated areas by constructing synthetic rating curves from high-resolution DEM. Coupling the GeoFlood with the National Water Model (NWM) which forecasts discharge quantity, we can estimate the water level of a stream from its rating curve and map inundation areas based on the DEM data. However, two major limitations in the GeoFlood may undermine the accuracy of flood forecasting and inundation mapping. The first issue arises from the process of constructing synthetic rating curves. An empirical equation called Manning’s equation is widely used to estimate discharge, and one of the most important constants in this equation is Manning’s roughness coefficient, n. Currently, the n value is assumed to be 0.5 for all streams in the GeoFlood, but the n value varies in magnitude among streams based on channel morphology, bedforms, vegetation and other factors. At streams without observations, we do not know the correct value of n and cannot construct an accurate rating curve (i.e., the relationship between the discharge and water stages) to forecast floods and inundation areas. The second issue is the quality of discharge forecasting from the NWM. The forecasting of discharge at many streams from the NWM is not calibrated, and the accuracy of discharge is unguaranteed. Eventually, any overestimation or underestimation of discharge can disrupt the accuracy of inundation mapping.
2 Research Objectives
Here, we propose a machine learning framework that is designed to overcome the uncertainties in flood forecasting processes framed in the GeoFlood. Machine learning is one of the most revolutionary developments in data science, and it provides a new framework for solving real-world problems using data effectively. Rather than split the flood forecasting into two steps (i.e., constructing rating curves and acquiring discharge forecasting from the NWM) to estimate water levels of flood events, we can now directly predict water levels using machine learning with near-real-time soil moisture and precipitation observations from space, high-resolution land use and land cover, and stream morphology data. We will start from data reaches from densely populated regions of the Chesapeake Bay Watershed. As our satellite data have greater coverage than the Chesapeake Bay, we can extend our research to the continental United States and other places in the world. In summary, this project has two objectives:
- Establish a machine learning frame to predict two key variables of floods: 1) the peak in-stream water levels and 2) the time to peak floods from the beginning of storms.
- With the predicted water level, map the worst-case scenario for inundation areas from topographic data.