This dissertation contains a literature review and three studies concerned with the development, assessment, and use of machine learning (ML) algorithms to explore automatically generated predictions of flash floods. The literature review explores several relevant issues: how flash floods are defined, the organization and structure of the flash flood forecasting and alerting enterprise in the U.S., proposed methods and tools for understanding and forecasting flash floods, the statistical underpinnings of ML, and how ML techniques can be applied to a wide variety of complex scientific problems, including those of a meteorological bent.

Using an archive of numerical weather predictions (NWP) from the Global Forecast System (GFS) model and a historical archive of reports of flash floods across the U.S., I develop a set of machine learning models that output forecasts of the probability of receiving a Storm Data report of a flash flood given a certain set of atmospheric and hydrologic conditions as forecast by the GFS model. I explore the statistical characteristics of these predictions, including their skill, across various regions and time periods. Then I expound upon how various atmospheric fields affect the probability of receiving a report of a flash flood and discuss different methods for interpreting the results from the proposed ML models. Finally, I explore how the mooted system could be operationalized, by delving into two case studies of past impactful flash floods in the U.S., by presenting results of National Weather Service forecasters using and interacting with the proposed tools in a research-to-operations testbed environment, and by geographically extending the predictions to cover additional parts of the world’s landmass via a set of case studies on the European continent.

One ML algorithm in particular, the random forest technique, is used throughout the vast majority of the dissertation, because it is quite successful at incorporating large amounts of information in a computationally-efficient manner and because it results in reasonably skillful predictions. The system is largely successful at identifying flash floods resulting from synoptically-forced events, but struggles with isolated flash floods that arise as a result of weather systems largely unresolvable by the coarse resolution of a global NWP system. The results from this collection of studies suggest that automatic probabilistic predictions of flash floods are a plausible way forward in operational forecasting, but that future research could focus upon applying these methods to finer-scale NWP guidance, to NWP ensembles, to new regions of the world, and to longer forecast lead times.