BIGGER PICTURE Disaster preparedness
Which storm on the screen is about to become dangerous?
ProbSevere combines radar, satellite, lightning and weather-model clues into a probability attached to each storm.

- Horizon
- The next hour
- Hazards
- Hail, damaging wind and tornadoes
- Version 3
- Operational at NOAA since August 2025
How the idea works
Track a storm
Radar and satellite data identify a moving storm object.
Combine hazard clues
Lightning, structure and environment inform four models.
Support warning decisions
Forecasters inspect probabilities and their trends.
One storm, several kinds of evidence
A forecaster watching developing thunderstorms has more than one screen’s worth of evidence. Radar sees precipitation and storm structure; satellites show cloud evolution; lightning and short-range weather models add other clues. ProbSevere follows individual storms and combines these signals into next-hour probabilities, helping a forecaster decide which developing cell deserves closer attention.
The probability travels with the storm
Version 3 uses four machine-learning models: one each for severe hail, strong wind, tornadoes and any of these hazards. The operational display outlines storms with probability-coded contours; a mouse hover reveals probabilities and underlying predictor values. That makes the estimate inspectable in the same working environment where the meteorologist already follows the storm.
Improvement was not equal for every hazard
The 2024 study evaluated data from 2018–2023 and found improved performance over version 2 across much of the United States and across seasons. Gains were strongest for hail, wind and the combined severe-weather model; tornado improvements were modest. Forecasters in simulated real-time tests generally preferred the new guidance. NOAA made version 3 operational in August 2025.
A low number does not end the investigation
The operational quick guide lists concrete blind spots: shallow storms and storms with little lightning can receive low probabilities, while tracking problems in squall lines can distort trends. A storm must first be identified. These limits explain why the system supports a warning decision rather than making one alone; no specific saved-life count is established by this evaluation.
The original sources
- Weather and Forecasting / NOAA repositoryProbSevere Version 3: Improved Exploitation of Data Fusion and Machine Learning for Nowcasting Severe Weather
- University of Wisconsin–Madison CIMSSProbSevere v 3
- NOAA / CIMSSProbSevere Quick Guide