Flash floods are the second-deadliest weather event in the United States and the deadliest worldwide, according to the National Weather Service. They can turn a dry creek bed into a killing channel in minutes. Six inches of fast-moving water can knock an adult off their feet. A foot of water can float a car. Two feet can carry away trucks and SUVs. And because the most intense rainfall often falls in places where monitoring equipment is sparse, forecasters frequently have to issue warnings with incomplete information.
A new software tool called TACLS — short for Transient Artifact and Continuous Learning System — is designed to close some of that gap. Developed by scientists at the University of California San Diego, the National Weather Service, and NASA, and funded by NASA’s Earth Science Technology Office, TACLS combines satellite data with machine learning to help forecasters identify flash flood risk areas sooner. It is currently operating only in California, but a new version with improved graphics is being loaded and will be available to all 122 NWS forecast offices in the second half of October.
How TACLS Works
At the heart of TACLS is the Global Navigation Satellite System, or GNSS — the same constellation of satellites that underpins GPS. GNSS signals travel from satellites to ground sensors, and as they pass through the atmosphere, they slow down slightly depending on how much moisture is present. That delay can be translated into a measurement called precipitable water, which tells forecasters how much water vapor is available to fall as rain.
“Satellites are more detailed over oceans than land,” said Yehuda Bock, the TACLS project lead at Scripps Institution of Oceanography at UCSD. He noted that tracking rainfall only as it happens does not give forecasters enough lead time to warn people ahead of a flash flood.
The machine learning side of TACLS was built by Bhavik Chandna, a UCSD graduate student who spent about a year developing it. The model uses a long short-term memory architecture, a type of machine learning well suited to sequences of data that change over time. Chandna said LSTM works well for shifting weather patterns. The model was trained on GNSS atmospheric measurements paired with historical flood data.
In practice, TACLS provides real-time precipitable water data, displayed as a graph with live data points. Forecasters can compare that real-time information against what their models predicted, helping them judge whether a storm is moving at the expected speed and whether it is likely to stall or intensify over a particular area. The new version being rolled out adds a map with more detailed graphics.
Assisting, Not Replacing, Human Forecasters
Chandna emphasized that TACLS is not intended to replace human forecasters. It is a decision-support tool. Ivory Small, science and operations officer at the NWS San Diego Weather Forecast Office, said TACLS will help save lives and that without it, storms could kill people.
Jayme Laber, a senior service hydrologist at the NWS Oxnard office, which oversees the Los Angeles area, said TACLS helps forecasters make better warning decisions. He also noted that the Western United States deals with quicker flash flooding than other regions, making fast detection especially valuable there.
The NWS issues three levels of flood alerts. A flood watch means there is a chance of flash flooding, usually issued 12 to 48 hours in advance. During a watch, local emergency teams review evacuation plans and set up sandbags and road barriers. A flood advisory covers nuisance-level flooding, while a flood warning signals life-threatening flooding. A flash flood is defined as flooding that occurs within six hours of the rainfall that causes it.
Current monitoring relies on rain gauges, stream gauges, satellites, and radar. Flash flood guidance is based on soil and slope conditions. But deserts and less populated areas have fewer rain gauges, leaving gaps in coverage. GNSS sensors can help fill those gaps, though the network was primarily built for earthquake prediction, and the majority of its sensors are located in earthquake-prone areas. That means TACLS is likely to be used most in the Western United States, though it could potentially work elsewhere if GNSS sensors and local weather data are available.
Limitations and False Positives
Machine learning systems are not flawless. Joel Johnson, an associate professor at the University of Texas at Austin, said machine learning can produce artifacts and false positives. He added that TACLS seems well suited for flash flood prediction, but the possibility of false alarms remains a consideration.
One way the system builds confidence is by looking at multiple nearby GNSS stations. Chandna said that when several stations show similar signals, confidence in a detection increases. That redundancy helps reduce the chance that a single anomalous reading triggers an unnecessary warning.
Why Earlier Warnings Matter
The need for faster alerts is not abstract. On June 9, a flood struck Lanesville, Indiana, a small community about 15 miles from the Kentucky border. More than 8 inches of rain fell within hours. Laura Lin, a resident, described flooding at her home and said the alerts she received came too late. She and her family evacuated safely, but she said an earlier alert would have been helpful.
Flash flooding is also a global threat. In August, a glacier collapse in Nepal caused deadly flash flooding. A warming climate is increasing extreme rainfall events in the United States, according to the brief, adding pressure to improve warning systems.
Small, the NWS San Diego science and operations officer, said research after storms is imperative — studying what happened and why can improve future forecasts. TACLS is one part of that effort. It is already in use at the Los Angeles and San Diego forecast offices, and its expansion to all NWS offices in October will put the tool in front of more forecasters who need to make split-second decisions.
For now, TACLS remains a helper, not a replacement. It gives forecasters another stream of data — atmospheric moisture measured from space — and a machine learning model that can flag when conditions look dangerous. Combined with human judgment, that could mean the difference between a warning that arrives in time and one that arrives too late.
