Comparing Forecast Techniques for Overlapping Tornado and Flash Flood Events

Abstract

Tornadoes and flash floods are considered two of the most hazardous weather phenomena in the United States. On rare occasions, these hazards can overlap in space and time, an event known as a TORFF (TORnado-Flash-Flood). These overlaps are especially dangerous because the rarity of combined events makes them difficult to forecast, and the protective guidance of overlapping warnings often contradict each other. Tornado warnings will advise people to seek shelter in a basement or low-level structure, contrasting flash flood warnings which often advise moving to higher ground. This divergence in guidance has proved deadly in multiple instances, highlighting the importance of ample warning time and specific protective guidance for such events. Numerous studies have shed light on the meteorological characteristics favorable for TORFF events, but little research has demonstrated their forecasting potential or if current operational forecast systems are capable of highlighting TORFF risk with sufficient calibration or spatial discrimination. The purpose of this study, therefore, is to develop a probabilistic TORFF forecasting paradigm that accurately highlights areas of dual-hazard risk to inform forecasters, emergency managers, and the public. Forecasters could use such a product to shift focus on areas where both hazards are possible to improve warning coordination, and emergency managers could use it to prepare additional resources for high-risk areas that may experience both hazards. In order to forecast TORFFs within a probabilistic framework, the statistical relationship between the individual events must be studied to discern how to best model the joint probability. Very little is known about the statistical relationship between tornadoes and flash floods, especially whether they are to be treated as independent or dependent events. To address this knowledge gap, a linear regression method is utilized to model the correlation between tornado and flash flood events across the United States (US). Results do not indicate a robust correlation across the US, but a notable region of positive correlation was present from northern Mississippi into central Iowa, motivating the use of both independent and conditional probability to derive joint probabilities. As a result, multiple forecast methods are evaluated including independent joint probabilities derived from operational Storm Prediction Center (SPC) tornado outlooks and Weather Prediction Center (WPC) excessive rainfall outlooks, conditional probability methods that account for statistical relationships between tornadoes and flash floods, and finally a random forest machine learning model trained on Global Ensemble Forecast System reforecast data. No significant difference in forecast skill is evident between regions with and without a significant tornado-flash flood correlation. Forecast methods which utilize both SPC and WPC outlooks provide the best spatial discrimination for TORFF events, but probabilities are not well calibrated. Methods that utilize conditional probability in conjunction with the SPC tornado outlook are much better calibrated, yet have worse spatial discrimination. The machine learning model is also not well calibrated but demonstrates enhanced spatial discrimination in test cases over other forecast methods, suggesting future work with machine learning is a promising avenue for forecasting TORFF events.