Machine Learning for Impact-Based Flash Flood Warnings: Hazard Report Operationalization for Impact Predictions

Abstract

Floods account for approximately one third of all global geophysical hazards, and flash floods allow for extremely short lead times for warnings to be emitted. Flash flood warnings are weather-related alerts which serve to inform of potential hazardous conditions which threaten life or property. The National Weather Service has transitioned to an impact-based format for flash flood warnings, which aims to provide additional valuable information about hazards, that facilitate improved public response and decision making. This work responds to the need for new decision support tools, which enable forecasters to anticipate distinct levels of impacts associated with flash flood forecasts, and provide support for issuing impact-based flash flood warnings. This dissertation proposes a foundation over which said decision support systems can be built. First and foremost, by constituting an unprecedented data set of historical flash flood reports with associated impact categories, achieved by the systematic application of a language-based impact framework (Flash Flood Severity Index) as a natural language processing task piped through a large language model (OpenAI's GPT-3.5-turbo). Secondly, through a proof-of-concept machine learning model trained to predict the severity of flash flood forecasts, based on operational flash flood forecasts, geomorphological data, and vulnerability layers.