Machine Learning Transformation Could Boost National Water Model's Forecasting in Mountainous Areas

Big DataBig Data

Posted: February 29, 2024

Machine Learning Transformation Could Boost National Water Model's Forecasting in Mountainous Areas

"The team is encouraged by the promise that successful application of machine learning in the National Water Model could be a significant leap forward in hydrological forecasting."

A team of University of Vermont researchers have found that the integration of machine learning into the National Water Model (NWM) could significantly enhance its streamflow forecasting abilities, particularly in the challenging montane regions of the United States, as seen in Vermont.

Forecasting such as this is crucial for effective water resource management, flood prevention, and environmental protection in areas where water flow prediction has traditionally been challenging due to complex terrain and diverse climates. Especially as a warming climate brings more moisture into these areas.

Machine learning (ML), a branch of artificial intelligence, enables the NWM to process vast datasets, identify patterns, and make more accurate predictions. This approach is crucial in areas with sparse data collection points and diverse environmental features. Research indicates that ML could improve the NWM performance in key efficiency metrics, creating a more reliable forecasting tool for hydrologists and planners.

The team was interested in changes to the Nash-Sutcliffe Efficiency (NSE) metric, used to assess the predictive power of hydrological models, which showed a significant increase, indicating that the ML-enhanced NWM's forecasts align more closely with observed data. Similarly, the Kling-Gupta Efficiency (KGE), which combines correlation, bias, and variability into a single score, demonstrates improved alignment between predicted and observed streamflow.

Another metric, the Pearson correlation coefficient, measuring the linear correlation between predicted and observed values, also showed a noteworthy rise, suggesting a stronger linear relationship due to the ML integration.

Moreover, the percent bias, indicating the average tendency of the model predictions to be larger or smaller than their corresponding observed values, has been substantially reduced. This reduction in bias means using these ML tools the NWM could offer a more balanced and accurate representation of streamflow, especially in smaller montane headwater streams.

The team is encouraged by the promise that successful application of machine learning in the NWM could be a significant leap forward in hydrological forecasting. It would not only enhance the model's performance in Vermont but also sets a precedent for improving streamflow predictions in similar mountainous regions across the country.

Assessing and Improving National Water Model Performance in Montane Headwater Catchments” was presented as part of “Machine Learning for Hydrologic Modeling VI” on December 14, 2023 during the American Geophysical Union Annual Fall Meeting

  • Mirce Ivon Morales Velazquez, University of Vermont (First Author, Presenting Author)
  • Andrew Schroth, University of Vermont
  • Beverley Wemple, University of Vermont
  • James Shanley, United States Geological Survey
  • Donna Rizzo, University of Vermont
  • John Kemper, University of Vermont
  • Kristen Underwood, University of Vermont