Utilizing Machine Learning and Probabilistic High Resolution Model Guidance to Improve Flash Flood Forecasts Across Southern Utah.

The Southern Utah backcountry continues to draw an increasing number of outdoor recreationalists. For example, visitation to Zion National Park has nearly doubled in the last 10 years from 2.8 million to about 5 million. This results in increased exposure to flash flood risk in remote areas where there is little to no means for receiving Flash Flood Warnings. To address this problem, the National Weather Service (NWS) Weather Forecast Office (WFO) in Salt Lake City (SLC) has worked closely with land management agencies, including the National Park Service, to develop a Flash Flood Potential Rating which extends out to 2 days. This product is designed to highlight days of higher flash flood potential to those with plans to recreate in the backcountry . Providing this risk information at longer lead times than Flash Flood Warnings gives visitors and our partners time to better prepare or alter plans ahead of time. Partners at our National Parks and recreation areas increasingly rely on this forecast to inform and educate visitors of this risk, which includes prominent signage featuring this forecast at the entrance to popular hikes, printable forecasts at visitor centers, and included on all backcountry permits issued by park rangers, which they review with each permit holder. Given the important role of this forecast, it is vital to ensure we are providing the most reliable and scientifically robust information possible, which is why WFO SLC is pursuing two approaches to enhance the forecast starting point.

Historically, the Flash Flood Potential Rating focused on synoptic scale environments favorable for flash flooding utilizing precipitable water and storm motion. In recent years, local research focused on developing a random forest machine learning algorithm utilizing thermodynamic and kinematic parameters from point-based (BUFR) data from the North American Mesoscale (NAM) model to identify favorable days for flash flooding within larger scale basins across southern Utah. Initial versions of this algorithm showed some skill in identifying flash flood days across southern Utah, however a rather short period of record resulted in a limited training dataset resulting in inconsistent results. Robust monsoon seasons in 2021 through 2023 expanded and improved the training dataset and considerably improved the skill of the model, which is being incorporated by default into the first-guess forecast presented to WFO SLC meteorologists for the 2024 convective season.

Also in 2023, a local study focused on leveraging high resolution output from Convective Allowing Models in an attempt to produce a higher resolution flash flood potential focused on smaller scale drainages. This effort utilized QPF neighborhood probability of exceedance values for the 2023 monsoon season and produced mixed results, likely due in part to a small sample size. In order to expand this dataset, QPF neighborhood probability of exceedance values from both the HREF and SREF are being archived for the 2024 Monsoon Season, and will be added to the initial 2023 dataset. The expanded dataset will be analyzed to determine whether these forecasts are skillful in identifying potential flash flood days across small flash flood prone basins across southern Utah. Additionally, this dataset will be combined with the Machine Learning output to determine if a blended approach provides additional skill. The end goal is to develop a robust methodology utilizing high resolution model output to forecast flash flood potential within flash flood prone small scale basins across southern Utah.