A flood predictability study for Hurricane Harvey with the CREST-iMAP model using high-resolution quantitative precipitation forecasts and U-Net deep learning precipitation nowcasts

Abstract

• CREST-iMAP is tested for flood predictability giving 1-hour lead-time. • NWP QPFs and U-Net AI nowcasts are used to predict for Hurricane Harvey flood. • QPFs have displacement errors but still yielded good flood peak prediction. • QPFs predict better maximum flood extent maps than AI nowcast. A flood is one of the most hazardous natural disasters, and it commonly causes fatalities and socioeconomic damages. The advances of modeling techniques and observation data in flood prediction have found success in field operations. This paper presents a comprehensive flood prediction of Hurricane Harvey in 1-hour lead-time that is not limited to 1D streamflow forecast but also 2D flood extent and 3D inundation depth. It uses high-resolution quantitative precipitation forecasts (QPFs, from operational Rapid Refresh-RAP, and High Resolution Rapid Refresh-HRRR models) and deep learning nowcasts (AI nowcasts). The results show that the QPFs have a well-known displacement issue and the AI nowcast cannot predict the precipitation intensity, and an attempt to combine the two methods (AI hybrid) failed to improve the overall accuracy. However, the 2D flood extent predictions with the HRRR and AI hybrid forcings can provide information indicating the future flooded area with about 50% accuracy (hit rate) and stream flow prediction showed that the HRRR QPF can provide relatively accurate flood peak prediction (-8%). In contrast, the AI nowcast reveals minimal displacement errors but underpredicts precipitation intensity. The deep learning method also indicates that the binary tests with low threshold, which are commonly employed in the deep learning field, neglect the importance of precipitation intensity errors for extreme event studies.