NASA-funded TACLS uses GNSS satellites and machine learning to speed up flash flood warnings

A flood that outran the warnings

On the morning of June 9th, Laura Lin was on a video call at her home in Lanesville, a small Indiana town near the Kentucky border, when she noticed pieces of her barn drifting across the yard. She woke her children and left immediately, and the family sheltered with a neighbor. As she told The Verge, the official messages urging residents to climb onto their roofs arrived only after much of the town had already flooded. Lanesville had absorbed more than 8 inches of rain in just a few hours, far beyond what counts as heavy rainfall.

Her experience frames a report from The Verge on a new piece of software built to close that timing gap.

What TACLS does

The system is called the Transient Artifact and Continuous Learning System, or TACLS. It was developed by scientists from UC San Diego, the National Weather Service and NASA, with funding from NASA's Earth Science Technology Office through its Advanced Information Systems Technology program. It combines satellite data with machine learning to flag places at risk of tipping from heavy rain into a genuinely dangerous flash flood, and to help NWS forecasters decide when to issue alerts. According to The Verge, TACLS currently covers only California, but it is set to become available to all 122 NWS weather forecast offices across the United States and its territories.

How flood warnings work today

Forecast offices already lean on a mix of instruments: rain and stream gauges, weather radar and satellites, plus flash flood guidance thresholds derived from local soil types and the steepness of slopes. Jayme Laber, senior service hydrologist at the NWS office in Oxnard, California, told The Verge that this gauge network is uneven, with deserts and sparsely populated areas covered far less densely than elsewhere.

Alerts come in three levels. A watch, issued 12 to 48 hours in advance, signals that a flash flood is possible and gives emergency teams time to stage sandbags, set barriers and review evacuation plans. An advisory covers nuisance-level flooding that threatens neither life nor property. A warning is the gravest tier, describing situations where someone could be swept off their feet. The NWS classifies a flood as a flash flood when it develops in under six hours.

The limitations are structural, as Yehuda Bock, the TACLS project lead and a research geodesist at the Scripps Institution of Oceanography, explained to The Verge. The NWS satellites resolve more detail over oceans than over land, and monitoring rainfall as it happens provides almost no lead time, because once precipitation is falling, the event has already begun.

Measuring moisture through signal delays

TACLS draws on the Global Navigation Satellite System, or GNSS, a network of satellites and ground sensors normally associated with earthquake prediction. Bock notes that water vapor slows the signals traveling between GNSS satellites and ground stations, so the size of the delay reveals how much moisture the atmosphere holds. That measurement of precipitable water gives forecasters a real-time view of what the sky is doing as a storm builds, which they can check against what weather models predicted. Laber said this comparison helps staff make better calls in the warning decision process.

The machine learning layer

Raw moisture data still needs interpretation. Bhavik Chandna, a UC San Diego graduate student, spent about a year building the model using long short-term memory architecture, an approach well suited to weather patterns that evolve over time, such as developing storms. The model was trained on years of GNSS atmospheric measurements together with records of atmospheric rivers, precipitation and past flash flood warnings, so it learns both how moisture shifts as a storm develops and whether conditions justify issuing a warning.

Outside researchers counsel realism. Joel Johnson, an associate professor in the department of earth and planetary sciences at the University of Texas at Austin, who was not involved in the project, told The Verge that machine learning handles some tasks poorly and others very well.

Why it matters

According to The Verge, floods are the deadliest weather event worldwide and the second-deadliest in the United States. Just 6 inches of fast-flowing water can knock an adult down, a foot of water can lift a car, and two feet can move trucks and SUVs. A warming climate is driving more extreme rainfall in the US, multiplying the chances of flash floods, as seen in the deadly flooding in Nepal in August after a glacier collapse. Ivory Small, science and operations officer at the NWS San Diego office, summed up the stakes: without TACLS, a storm moving into an area could kill some folks, whereas with it, you can put out the warning and save some folks.