93% of historical flash-flood warnings were correctly identified by a machine learning system that processes GNSS satellite signal delays in under 15 minutes. That's the headline from the Transient Artifact and Continuous Learning System (TACLS), built by NASA JPL, UCSD Scripps, and NOAA's National Weather Service.
What TACLS Actually Does
TACLS doesn't look at clouds or radar returns. It watches water vapor in the troposphere by measuring how much GNSS satellite signals slow down on their way to Earth. More water vapor means higher signal delay. The ML model, trained on over 30 years of GNSS data, acts as an anomaly detector. It flags any unusual increase in atmospheric moisture and decides whether that spike is a sensor artifact or a real transient event like heavy precipitation.
If it's a transient, the software passes the data to a visualization tool called MGViz, originally adapted from JPL's Mars mission GIS software. Human meteorologists then examine the flagged areas and decide whether to issue a flash flood warning or advisory. The whole loop from satellite data to analyst visualization runs in near real-time, producing forecasts in as little as 15 minutes.
Proven Against Real Events
The team tested TACLS against severe weather data from 2017 to 2023, covering atmospheric rivers, monsoonal convection, and tropical cyclone remnants. It successfully captured 93% of the flash-flood warnings the National Weather Service actually issued during that period. Not a contrived benchmark, but a direct comparison against operational decisions.
Credit for the lead concept goes to Yehuda Bock at UCSD Scripps, who wanted to give meteorologists a tool that cuts the time to decide on warnings. The back-end software pulls from JPL's Domain-agnostic Outlier Ranking Algorithms and Time-series Forecasting, Evaluation, and Deployment programs. No black box, no vendor lock-in: both the TACLS software and the training data will be open source.
What Changes Next
National Weather Service meteorologists are now working to integrate TACLS into existing forecasting systems for Southern California. That's a narrow first deployment, but the architecture is portable. Any region with GNSS coverage and a desire to catch flash floods faster can retrain the model on local data. Open source means you can fork it, adapt it, or build your own from scratch. The question is how quickly other weather services will pick up a system that already demonstrates 93% recall on real operational warnings.
Source: NASA Uses Machine Learning to Enhance Flash Flood Warnings
Domain: science.nasa.gov
Comments load interactively on the live page.