CFANS research offers significant improvements in flood forecasting

A new flood forecasting model from the College of Food, Agricultural, and Natural Resource Sciences (CFANS) and the College of Science and Engineering (CSE) significantly improves the accuracy of flood prediction, offering potentially life-saving solutions for communities.

The model, which combines elements of traditional physics-based forecasting with newer machine-learning techniques, can predict streamflow and flood levels with greater accuracy than current methods used across the United States.

“We have already seen increasing floods within the last few decades in many parts of Minnesota, including several flood records set within the last couple of years,” says Zac McEachran, a research hydrologist with the University of Minnesota Climate Adaptation Partnership. “It’s vital that we improve our ability to predict these events so we can protect lives and infrastructure.”

In a typical flood event, forecasters with the National Weather Service modify physics-based models in real-time with observations from the field, attempting to improve available computer-generated predictions with real-world intelligence. But this labor-intensive modification process is hard to scale up. This presents a challenge as climate change is increasing the risk of sudden and severe flooding, and the corresponding demand for forecasts.

In recent years, researchers have been attempting to apply machine-learning techniques to the flood-prediction process. While we’ve seen great strides in this approach, human forecasters using the traditional physics-based tools are still out performing the best machine-learning models.

McEachran’s research offers a third avenue forward, improving machine-learning techniques with a “knowledge-guided” approach that offers greater overall accuracy without time-intensive manual adjustment.

“We know that flood prediction experts are critical now and into the future in providing our communities with life-saving forecasts,” says Heidi Roop, director of the Climate Adaptation Partnership. “But we also know that advanced water-prediction tools like McEachran’s can add capacity and be efficiently scaled at a time when the risks of catastrophic flooding have never been greater.”

The model is particularly successful against existing forecasting methods in relatively arid areas of the United States—a wide swath from North Dakota and western Minnesota south through Arizona and New Mexico.    

McEachran’s research was published in November in the journal Water Resources Research. A companion paper on the machine-learning aspects of the work will appear in the forthcoming proceedings of the Institute of Electrical and Electronics Engineers International Conference on Data Mining.