Hybrid AI Tool Boosts National Flood Forecast Accuracy

Publication

The study was published in the journal AGU Advances, which publishes high-impact, open-access research and commentary across the Earth and space sciences.

How the Model Works

This AI is a neural network, or a deep learning model, and is trained to find errors. The network was trained on historical observational and National Water Model-simulated data for rainfall and flooding. Researchers created a hybrid program by combining the AI system with the National Water Model, which forecasts streamflow for the entire United States.

The national model utilizes physics to determine the landscape, water levels, and other factors to predict flooding. When only AI is used to analyze flood predictions, such as Google's AI-only system, more complex factors, like topography or land use, may be overlooked.

Overview of medium-range (10-day) streamflow forecasting framework over the CONUS. (a) Integration of NWM with ECN. ECNs are trained using NWM hindcasts, climate forcing, and USGS streamflow observations to estimate an ensemble of errors, which combine with NWM forecasts for probabilistic forecasts. (b) Forecast framework showing data sets used for 1 to 10-day error predictions: observed (gray) and forecasted (colored) meteorological forcings, observed flows, and NWM forecasts. “+” indicates data combination sequence; N represents lookback window for observed data.

Limitations of Pure AI Models

"So especially for floods, the performance of the pure AI model is quite poor," said Vinh Ngoc Tran, hydrologist at the University of Michigan and head researcher on the study. "The advantage of the AI model is that they are very simple. You only need to use the data to train the model and provide the forecast, but the most important thing we need to be concerned about is ensuring prediction accuracy for flood events that can cause significant damage."

Role of NOAA Data

Across the United States, NOAA has nearly 11,000 operational water gauges that collect data on previous floods and water levels. However, NOAA also tracks data beyond water levels. The agency gathers detailed information on variables such as vegetation, urbanization, and drainage networks for each gauge.

The sheer amount of available information is helpful, but it also makes it harder to narrow down the research to identify where things went wrong or to account for everything when creating a flood model. This will cause errors in the forecasting system, and that is where AI can step in.

Errorcastnet: The Hybrid Approach

Aptly named Errorcastnet, the AI-based system looks for errors in the national model. It examined old flood records and compared them with the model's flood forecasts. For times when NOAA's model did not correctly forecast the flood, the AI would categorize the errors into two groups: those that could be reduced and those that could not be fixed.

The AI learns the problems within the model and works to correct them. Errors that cannot be fixed, such as limitations within the model itself or incomplete data, are still essential to track. It helps the AI continue to train and improve the forecasting by focusing only on the errors that it can fix.

Physics Still Matters

"You can't throw away physics," said Valeriy Ivanov, physical hydrologist at the University of Michigan and an author of the study. "It's just by definition you can't. You have to understand that systems are different. The landscapes are different. You have to account for dominant physical processes in your predictive model."

Researchers found that when using Google's AI flood forecasting program, which relies on historical data to make predictions but doesn't consider details such as elevation, vegetation, and reservoirs that the National Water Model incorporates, the model generally underpredicts flood flows.

"We understand the power of AI," Ivanov said. "No one denies it. It's definitely there. But it should not negate decades of research. It should not negate the understanding of physics and understanding of complexity of physical processes in watersheds."

Future Implications

By improving NOAA's forecasting model, the researchers believe it could also enhance the potential economic impacts of floods. More accurate flood predictions could enable businesses to better prepare for the upcoming floods. Tran, Ivanov, and their team hope that as the program grows, potential floods can be predicted in detail with several days or more of advance notice.