Combining Model and Observational Data Using Machine Learning for Short-Term Severe Weather Hazard Prediction

Loken, Eric D.; Clark, Adam J.; Calhoun, Kristin M.; Sandmael, Thea N.; Skinner, Patrick; Burke, Patrick C.; Heinselman, Pamela; Cintineo, John L.; Reinhart, Anthony

doi:10.1175/AIES-D-24-0102.1

i

Combining Model and Observational Data Using Machine Learning for Short-Term Severe Weather Hazard Prediction

2025
By Loken, Eric D. ; Clark, Adam J. ; Calhoun, Kristin M. ; ...

Select the Download button to view the document

Please click the download button to view the document.

Details

Journal Title:

Artificial Intelligence for the Earth Systems
Personal Author:

Loken, Eric D. ; Clark, Adam J. ; Calhoun, Kristin M. ; Sandmael, Thea N. ; Skinner, Patrick ; Burke, Patrick C. ; Heinselman, Pamela ; Cintineo, John L. ; Reinhart, Anthony
NOAA Program & Office:

OAR (Oceanic and Atmospheric Research) ; NSSL (National Severe Storms Laboratory) ; WPO (Weather Program Office) ; CIWRO (Cooperative Institute for Severe and High-Impact Weather Research and Operations)
Description:

The Warn-on-Forecast System (WoFS) is a convection-allowing ensemble that rapidly assimilates high-resolution radar, satellite, and other observational data to enhance forecasting capabilities at 0–6-h lead times. However, data latency and model spinup can reduce WoFS’ utility, especially within the first hour after initialization. On the other hand, Probability of Severe (ProbSevere), version 2 (PS2), is a set of statistical models that considers radar, satellite, and numerical weather prediction environment data to provide skillful severe hail, wind, and tornado probabilities at 0–1-h lead times, updated every 2 min. To enable seamless, probabilistic severe weather hazard forecasts leveraging the strengths of both systems, a random forest (RF) algorithm is developed that considers predictors from both WoFS and PS2. Experiments are conducted using different combinations of predictors at various spatial radii and lead times, and predictor importance is assessed. RFs configured using predictors from both WoFS and PS2 are found to outperform RFs configured from only one system, with the greatest gains in forecast skill from the all-predictor RFs coming at lead times less than 90 min. PS2 (WoFS) predictors are identified as more important at earlier (later) lead times. Severe hail forecasts were the most skillful of the three hazards, followed by severe wind and tornadoes. This new algorithm shows how combining complementary datasets using machine learning can improve short-term severe hazard guidance. Upcoming work will report on the results from testing during real-time forecasting experiments.
Keywords:

Artificial Intelligence Decision Trees Machine Learning Artificial Intelligence Decision Trees Machine Learning Model Interpretation And Visualization Nowcasting Short-range Prediction
Source:

Artif. Intell. Earth Syst., 4, e240102
DOI:

https://doi.org/10.1175/AIES-D-24-0102.1
Format:

PDF
Document Type:

Journal Article
Funding:

Grant no. NA21OAR4320204 ; Grant no. NA22OAR4590530
License:

https://creativecommons.org/licenses/by/4.0/
Rights Information:

CC BY
Compliance:

Submitted
Main Document Checksum:

urn:sha-512:4cc917c69bd3d7dd0f2730354654581aa7d5c9ba262af15c20a96520c50460e31360f5b394872385e59da2513d9782e0816d5c0b0ebd9c12e9236d1df9241bc8
Download URL:

https://repository.library.noaa.gov/view/noaa/71974/noaa_71974_DS1.pdf
File Type:

[PDF - 25.71 MB ]

ON THIS PAGE

Details

The NOAA IR serves as an archival repository of NOAA-published products including scientific findings, journal articles, guidelines, recommendations, or other information authored or co-authored by NOAA or funded partners. As a repository, the NOAA IR retains documents in their original published format to ensure public access to scientific information.