Events - Event View

Rapid, regional scale riverine and pluvial flood risk assessment and forecasting is complicated by hydrology and hydraulic complexity. Complexity is not fully captured by the most detailed of hydrology models—causing model results to deviate from observations over long time scales. Moreover, as the spatial extents (i.e., scale) and resolution of the study area increase, the associated traditional hydraulic models become computationally expensive. Accordingly, traditional hydrology and hydraulic modeling seemingly is at odds with the efficiency needed for regional flood analysis. Here, we show that an effective, rapid prediction of pluvial flood inundation is achieved through a direct statistical simulation of hydrologic averages coupled to the insight of a machine learning model that extends high-resolution detailed HEC-RAS 2D model results over regional areas. We found that the direct statistical simulation of the watershed water balance provides a reliable baseline of the watershed moisture status (in comparison to satellite data) and so provides for a reasonable estimation of runoff. This runoff estimation is then mapped to inundation extents based on the machine learning process. While the machine learning accuracy is around 80-percent in comparison to the detailed HEC-RAS 2D models (based on high-resolution digital elevation models), we show the results are more reasonable and detailed when compared to previous mapping efforts. For two case study areas, our results demonstrate how machine learning can provide insights from data whether it be remote sensing or model derived. We anticipate that similar machine learning approaches will start to complement traditional hydraulic modeling efforts in rapidly extending results to large regions. As machine learning approaches evolve, we foresee data driven machine learning approaches capturing more of the complex functional dynamics of flood modeling without the parameter uncertainty of traditional modeling approaches.