Created with support from NASA’s Earth Science and Technology Office (ESTO), TACLS uses machine learning to automatically identify evidence of impending flash flooding (an abnormal increase in atmospheric moisture) that meteorologists may miss when analyzing large amounts of data. TACLS flags that evidence, shows where flash flooding is likely to occur, and displays that information in user-friendly visualizations for human analysts to interpret. These analysts can then decide whether to issue a flash flood warning or weather warning.
This new framework for tracking extreme weather events and predicting impending flash floods operates in near real-time and generates predictions in just 15 minutes.
“This is what we really wanted to do: give meteorologists a tool to help them make decisions about flash flood warnings,” said Yehuda Bock, a fellow at UCSD Scripps Institution of Oceanography and principal investigator of TACLS.
In simulation tests, TACLS successfully captured 93% of flash flood warnings issued between 2017 and 2023 using data from a variety of severe weather phenomena, including atmospheric rivers, monsoon convection, and remnants of tropical cyclones. National Weather Service meteorologists are currently working to incorporate TACLS into existing systems for predicting flash floods in Southern California.
This learning system has two main components. First, the analytics backend software suite uses machine learning algorithms to process satellite data to identify areas at risk of flooding. Second, user-friendly visualization software highlights those areas for further human analysis.
TACLS backend software analyzes data from satellites of the Global Navigation Satellite System (GNSS), a constellation of satellite networks that power navigation services around the world. Water vapor in the troposphere delays the signals of these satellites as they reach Earth. This signal delay can be analyzed to calculate the amount of water vapor in the atmosphere at a particular location on Earth.
The TACLS analytics backend software suite features machine learning models trained using over 30 years of historical GNSS data. This model is an anomaly detector that tracks abnormal increases in atmospheric moisture. The model then carefully examines that atmospheric moisture data to determine whether it is an artifact (mischaracterized or distorted data) or a transient event (time-sensitive physical event such as heavy rain) that requires interpretation by a human analyst.
If TACLS determines that the data is temporary, such as extreme weather events that warrant a flash flood warning, the data is forwarded to TACLS visualization software (MGViz) for further human evaluation. Analysts use their judgment and experience to interpret these events, determine whether flagged data indicates the possibility of flash flooding, and issue flash flood warnings if necessary.
Several past innovations developed at JPL are leveraged by TACLS to process GNSS data and display results. The analytical backend software suite incorporates elements of JPL’s domain-independent outlier ranking algorithm program and time series forecasting, evaluation, and deployment program. The TACLS visualizer is based on the Multimission Geographic Information System originally developed at JPL for NASA’s Mars mission.
TACLS software combines all these components into a new system that enhances existing methods and reduces the time it takes for human analysts to determine whether to issue a flash flood warning.
Both the TACLS software and the data used to train it will be open source, allowing scientists to adapt the model to their own research needs or create their own models from scratch.
For more information, see the NASA TechPort entry for this project.
Project leader: Dr. Yehuda Bock of the University of California, San Diego;
Sponsoring organization: NASA Earth Science and Technology Office Advanced Information Systems Technology Program. JPL; NOAA; National Weather Service.
