If a Flood Falls in a (Random) Forest, Does It Get Counted? Advances and Challenges in Predicting Excessive Precipitation Using Machine Learning

Since 2017, machine-learning models based on random forests have been used for probabilistic prediction of excessive rainfall, in a collaboration between Colorado State University (CSU) and NOAA's Weather Prediction Center (WPC). These forecasts can be used as a "first guess'' for WPC forecasters in generating their Excessive Rainfall Outlooks. The forecasts were initially generated for days 2-3 (lead times of 36-60 and 60-84 hours, respectively), have been tested as part of the Flash Flood and Intense Rainfall (FFaIR) experiment, and are now running operationally at WPC. The CSU-Machine Learning Probabilities (CSU-MLP) system has demonstrated skill in its probabilistic predictions at these lead times, yet some persistent biases occur, especially in portions of the western US.

This presentation will include a brief overview of the CSU-MLP system, and will focus on the causes of these biases and efforts to address them. In particular, these problems stem from the lack of accepted definitions of "excessive rainfall" and "flash flood", as well as the difficulties associated with quantitative precipitation estimation (QPE) in the complex terrain of the western US. In turn, it becomes challenging to train an algorithm to predict "excessive rainfall": the algorithm will do a good job predicting what it is trained to predict, but that may not correspond to what a forecaster is actually concerned about. Furthermore, inconsistencies in the reporting of flash flooding compounds these challenges. Nonetheless, by using multiple proxies for excessive rainfall, some of these problems can be alleviated. Results from recent updates to the CSU-MLP system, including efforts to produce day-1 forecasts will be presented, which show improvements over previous versions and yield more reliable forecasts in most regions.