Dense GNSS networks can do more than measure Earth deformation: they also sense atmospheric water vapor. The MEaSUREs ESESES project has developed an observation-driven system that uses GNSS tropospheric delays and machine learning to track extreme weather and identify flash-flood warning conditions in near real time.
Preprint: Real-Time Flash-Flood Warning Prediction Using GNSS Water Vapor Observations and Deep Learning
The Challenge
Flash floods can develop rapidly, especially in complex terrain, burn scars, and regions affected by intense
atmospheric rivers, monsoonal convection, or tropical cyclone remnants. Operational warnings rely on numerical
weather prediction, radar rainfall estimates, satellite products, gauge networks, and forecaster expertise.
GNSS provides an additional, independent observation of atmospheric moisture evolution over land.
Multi-task LSTM machine learning system for GNSS-based extreme weather event tracking and assess flash flood warning conditions in an operational environment. Five-minute tropospheric-delay observations (zenith wet delay and horizontal gradients) from continuously operating GNSS stations are synchronized with atmospheric-river catalogs, precipitation data, and historical flash flood warnings to form the Integrated Extreme Weather Dataset (IEWD). Model outputs are streamed to the TACLS machine learning framework and visualized via MGViz to support operational situational awareness and early warning.
GNSS Observations
5-minute tropospheric delay measurements from dense regional GNSS networks.
Atmospheric Features
Zenith wet delay and horizontal gradients track moisture surges and spatial asymmetry.
Machine Learning
A multi-task LSTM learns storm evolution and flash-flood-warning conditions.
Key Results
- Detected approximately 93% of issued flash-flood warnings in independent testing.
- Generalized across atmospheric rivers, monsoonal convection, and Tropical Storm Hilary.
- Provided near-real-time situational awareness using GNSS data available with 15-30 minutes latency.
- Early detections occurred in a subset of events, indicating that GNSS moisture signals can precede warning issuance.
Why GNSS?
| Radar / NWP / Satellite | GNSS Atmospheric Sensing |
|---|---|
| Tracks precipitation and forecast fields | Directly senses atmospheric moisture over land |
| Can be limited by terrain, latency, or model uncertainty | Operates continuously in all weather |
| Provides essential operational guidance | Provides independent observational evidence |
Broader Vision
This work demonstrates that dense GNSS networks can function as a ground-based atmospheric moisture observing
system. Combined with machine learning, GNSS atmospheric sensing provides a scalable, observation-driven pathway for near-real-time hazard monitoring that complements radar, satellite precipitation products, and numerical
weather prediction.