A group of researchers with strong ties to data science at the University of Minnesota won the Boggess Award for their paper published in 2023 in the Journal of the American Water Resources Association (JAWRA). The team of nine co-authors includes Vipin Kumar, Director of the CSE Data Science Initiative and Regents Professor in Computer Science & Engineering, and Xiaowei Jia, Assistant Professor of Computer Science at the University of Pittsburgh and a CS&E alumnus. At the time of publishing, the other seven authors were at various USGS (United States Geological Survey) offices in Pennsylvania, Wisconsin, and California.

The award-winning paper, “Near-term forecasts of stream temperature using deep learning and data assimilation in support of management decisions” addressed how USGS can use Artificial Intelligence to help water reservoir managers in the Delaware River Basin protect New York City’s drinking water while also protecting trout in the river by releasing reservoir water to cool stream temperatures if they are becoming too warm for trout and other cold-water species. 

This work builds on a revolutionary framework known as knowledge guided machine learning (KGML), that has been pioneered by Vipin Kumar’s research group at the University of Minnesota.   KGML is fundamentally more powerful than pure ML approaches and  traditional process-based models, as it allows integration of scientific knowledge in standard black-box machine learning (ML) methods to produce solutions that are scientifically grounded and likely to generalize on out-of-distribution samples even with limited training data.

According to lead author Jacob Zwart (Senior Data Scientist with Water Resources Mission Area, USGS) this work “introduced a novel approach to help integrate deep learning, process-based modeling, and data assimilation to allow forecasting of stream temperature up to seven days in advance.” They used public data sets to teach deep learning how to accurately predict stream temperature dynamics. Adding data assimilation improved forecasts compared to using KGML and process-based models alone. 

Their uncertainty characterization was also well calibrated. This allowed reservoir managers to anticipate the probability of exceeding ecologically relevant temperature thresholds, which helped them make decisions about releasing reservoir water downstream. 

A critical part of the development of the model was engaging with NYC reservoir managers throughout the process. With their input, the team was able to develop a tool that NYC reservoir managers find valuable in their decision making because of the enhanced forecasting skill and confidence, improved planning for staffing coverage on weekends and holidays, and additional benefits for anglers and ecotourists.

Zwart also notes that their research was conducted with a focus on rapid learning from real-world forecasting efforts by analyzing the forecasts produced and delivered in 2021, all while navigating real-world deadlines and resource constraints. 

“Given the flexibility of deep learning models, we anticipate that future extensions and enhancements to this framework could be applicable to a wide range of environmental forecasting applications. Once trained, deep learning models incur minimal computational costs, enabling forecasts to be generated almost as quickly as the incoming forcing data, which facilitates faster management decisions.”

The research has been recognized on several occasions including a presentation at a USGS Science Town Hall with an introduction by Secretary of the Interior Deb Haaland and highlighted as a USGS Open Science Success Story in 2023. The paper also received the Ecological Forecasting Outstanding Publication Award for 2023 from the Ecological Society of America and is in the top 10% most downloaded papers in the Journal of the American Water Resources Association for 2023. 

The Boggess Award, in honor of William “Randy” Boggess, one of the founders of the American Water Resources Association, is given each year to recognize the best paper published in the previous year.