Machine learning satellite system could give flash flood warnings precious minutes sooner

Updated September 18, 2026 2:25 pm

Researchers working with the National Weather Service have built a satellite-and-machine-learning system that could help forecasters spot flash flood threats earlier, giving communities more time to evacuate and responders more time to act. The tool, called TACLS, is already in use at weather offices in Southern California and is scheduled to roll out more broadly across the agency in October.

The technology matters because flash floods can turn deadly in minutes, often before traditional alerts arrive. In a warming climate, where extreme rainfall events are becoming more frequent and more intense, the ability to detect flood risk earlier could save lives.

Why flash floods remain so hard to warn about

Flash floods are among the most dangerous weather events in the world because they can develop with very little lead time. In the United States, they are the second-deadliest weather-related hazard. Globally, they are the deadliest. A few inches of fast-moving water can topple a person; deeper water can sweep away cars and larger vehicles.

That danger is not hypothetical. In southern Indiana, Laura Lin learned how quickly a normal morning can become an emergency. She was working from home in Lanesville, near the Kentucky border, when intense rain began filling her yard and sending debris past her barn. Within hours, her community had taken on more water than local terrain and infrastructure could handle.

Lin and her family escaped safely, but she says the warnings came after the danger had already arrived. By the time emergency alerts urged people to get to higher ground, she says many residents were already trapped by rising water. Her account illustrates the core problem meteorologists are trying to solve: how to identify flash flooding before roads disappear and escape routes vanish.

“It would have been helpful to know [water was coming] before I was locked in,” Lin said, describing the speed of the flooding in her town. “But it came in so fast.”

What is TACLS and how does it work?

TACLS, short for the Transient Artifact and Continuous Learning System, is a forecasting aid that combines satellite data, ground sensors and machine learning to help meteorologists see flood risk sooner. Rather than replacing human forecasters, it is designed to add another layer of evidence when weather offices decide whether to issue a flash flood warning.

The system was developed through collaboration among scientists at the University of California, San Diego, the National Weather Service and NASA. Funding came from NASA’s Earth Science Technology Office through its Advanced Information Systems Technology program.

At its core, TACLS looks for changes in atmospheric moisture, or precipitable water, using data from the Global Navigation Satellite System, a network better known for applications such as earthquake monitoring. In this case, the system measures how signals between satellites and ground stations are delayed by moisture in the air. The more water vapor present, the longer the delay, giving forecasters a way to assess how much atmospheric fuel is available for heavy rainfall.

The software then compares those observations with weather forecasts and historical storm patterns. If the live moisture signal suggests a storm is moving differently than predicted, or if conditions appear favorable for rapid flooding, the tool highlights the area for further scrutiny.

How do forecasters issue flood warnings today?

Forecasters do not rely on one signal alone. They use rain gauges, stream gauges, radar, satellite imagery, flood guidance thresholds and local expertise to decide whether an area is in danger. That process is already detailed and highly technical, but it has limits.

Some parts of the country, especially deserts and remote regions, simply do not have dense enough ground monitoring networks. Even where sensors exist, rainfall measurements arrive only after precipitation has started, which means they may confirm a storm too late to extend precious lead time. Satellites also tend to provide better coverage over oceans than over land, limiting their usefulness for inland flood prediction.

Jayme Laber, a senior service hydrologist with the National Weather Service in Oxnard, California, said weather offices use local flood guidance tailored to soil type, terrain and rainfall duration to decide if an alert is needed. That guidance is based on how much rain a region can absorb before runoff becomes dangerous. All of it is weighed by trained staff before an alert is issued.

That process results in three main alert levels:

• Watch — conditions suggest flooding is possible, often 12 to 48 hours ahead.
• Advisory — flooding may occur, but it is usually considered a lower-level nuisance event.
• Warning — flooding is imminent or already underway, and life-threatening impacts are possible.

Flash flooding is distinguished from ordinary flooding by speed. If water rises within six hours of the rainfall event, it is classified as a flash flood.

Why the new system could improve warning times

TACLS matters because it gives forecasters a real-time view of what is actually happening in the atmosphere before rainfall fully develops on the ground. Instead of depending only on the forecast, meteorologists can compare predicted storm movement with live moisture measurements and decide whether the storm is arriving faster, slower or more intensely than expected.

That can be especially important in regions where storms evolve rapidly or where terrain channels water into neighborhoods, roads and low-lying areas. In those settings, even a modest delay in warning issuance can make the difference between a manageable response and a life-threatening rescue.

Ivory Small, science and operations officer at the National Weather Service in San Diego, said the advantage is straightforward: better tools can buy enough time for people to move out of harm’s way before conditions become fatal.

Small and Laber both worked closely on TACLS and say it is meant to support, not replace, the judgment of meteorologists who know their local terrain and weather patterns. The software is a decision aid, not an automated warning machine.

How machine learning makes TACLS useful

The machine learning side of TACLS was developed by Bhavik Chandna, a graduate student at UC San Diego, using long short-term memory architecture, a type of neural network that is especially useful for time-based patterns. That matters because storms are dynamic. They evolve hour by hour, sometimes minute by minute, and forecasting flood risk requires recognizing those changes as they happen.

To train the model, Chandna and colleagues used years of atmospheric measurements from the GNSS network, along with data on atmospheric rivers, rainfall and past flash flood warnings. The system learned to recognize the moisture signatures that often appear as a storm intensifies and to flag conditions that could justify a warning.

Because machine learning systems can be prone to false positives, the model was designed with safeguards. If one station reports an unusual signal but neighboring stations do not, the system can suppress that alert. If multiple nearby stations show the same pattern, confidence rises that the signal reflects a real weather event rather than a sensor glitch.

Chandna said the goal was not to chase every anomaly, but to detect patterns that are broad enough to represent a real atmospheric change rather than a local error.

That approach is important because weather data is not always clean. Sensors fail, readings drift and isolated points can mislead an algorithm. Independent researchers say that challenge is common across machine learning in the physical sciences. But the TACLS design appears to account for that by looking for spatial consistency across several stations.

Where is TACLS being used now?

TACLS is already active in weather forecast offices covering Los Angeles and San Diego, two regions where fast-moving runoff and terrain-driven flooding are common concerns. Those offices are among the places most likely to benefit from real-time moisture data because flash flooding in the West often develops quickly and can be highly localized.

In its current form, the software displays a graph of live conditions so forecasters can review signals as storms develop. A newer version, with improved maps and graphics, is nearly ready to be deployed across the National Weather Service system. Laber said the updated software is expected to reach all NWS offices in the second half of October.

The broader goal is not just prettier visualization. A clearer interface can help forecasters absorb a lot of data more quickly under pressure, especially during active weather events when every minute matters.

What makes the Western U.S. a likely early winner?

The Western United States may be the first region to get the most out of TACLS because much of the GNSS sensor network is concentrated in earthquake-prone states, where ground monitoring is already more common. That gives forecasters more data points to work with and improves the system’s ability to detect atmospheric moisture patterns.

Still, researchers believe the approach could extend beyond the West. If a region has enough GNSS stations and local weather data, the same framework could be adapted elsewhere in the U.S. and potentially in other countries.

That flexibility matters because flash flooding is not a regional problem. It affects mountain towns, suburban neighborhoods, urban watersheds and remote communities alike. In many places, the challenge is not the absence of risk but the absence of enough high-quality observations to detect that risk quickly.

Could TACLS work outside California?

Yes. The system could be adapted to other regions as long as sufficient satellite-linked sensor coverage and local weather observations are available. Researchers say the underlying method is portable, even if the exact deployment will vary from one location to another.

Why researchers say the system still needs humans

Even with machine learning and satellite sensing, forecasters remain essential. The NWS does not want an algorithm making warnings in isolation, and the TACLS team does not see the tool as a substitute for human expertise. Instead, it is one more source of evidence for weather officers who must decide whether to send out watches, advisories or warnings.

That human layer matters because every storm is different. Local geography, soil conditions, urban drainage, burn scars, storm speed and rainfall rate can all influence how water behaves once it hits the ground. No model can fully capture those variables without real-time judgment.

Small said every storm teaches meteorologists something new, and that post-event analysis remains critical for improving future warnings and understanding why some floods behave differently than expected.

In that sense, TACLS is part of a broader shift in weather forecasting: not automation replacing forecasters, but better tools extending their reach.

Key facts about TACLS and flash flood warnings

Timeline: from storm threat to warning decision

What this means for communities at risk

For people living in flood-prone areas, the value of TACLS is not abstract. More accurate lead time can mean the difference between driving away safely and being stranded by water that rises faster than expected. It can give emergency managers time to close roads, position rescue crews, prepare shelters and warn neighborhoods before access disappears.

The system will not prevent extreme rain, and it will not eliminate flash flooding. But it could narrow the gap between when danger begins and when warnings reach the public. That gap has long been one of the deadliest weaknesses in severe-weather response.

Climate change is likely to make that gap more important. As storms become more intense and rainfall rates increase, the atmosphere is presenting more opportunities for rapid runoff and sudden flooding. That means forecasting tools able to read moisture signals sooner will likely become more valuable, not less.

What comes next?

The next step is wider deployment and continued testing. As TACLS spreads from California to more weather forecast offices, forecasters will learn how well it performs in different terrains, climates and sensor networks. The technology will also continue to evolve as scientists add improved graphics, refine the machine learning model and compare outcomes with real storm events.

Researchers involved in the project say the system is still part of a bigger scientific process. Every major flood event offers a chance to improve the next forecast, sharpen the next alert and better understand how water behaves under extreme conditions.

For residents like Lin, that progress cannot come soon enough. The hope is that future warnings will arrive before water fills a yard, blocks a road or turns an ordinary morning into a rescue.

In weather forecasting, minutes can save lives. TACLS is designed to make sure more of them are available.