MINNEAPOLIS / ST. PAUL (04/28/2022)—A team of researchers led by the University of Minnesota has significantly improved the performance of numerical predictions for agricultural nitrous oxide emissions. The first-of-its-kind knowledge-guided machine learning model is 1,000 times faster than current systems and could significantly reduce greenhouse gas emissions from agriculture.
The research was recently published in Geoscientific Model Development, a not-for-profit international scientific journal focused on numerical models of the Earth. Researchers involved were from the University of Minnesota, the University of Illinois at Urbana-Champaign, Lawrence Berkeley National Laboratory, and the University of Pittsburgh.
Compared to greenhouse gases such as carbon dioxide and methane, nitrous oxide is not as well-known. In reality, nitrous oxide is about 300 times more powerful than carbon dioxide in trapping heat in the atmosphere. Human-induced nitrous oxide emissions (mainly from agricultural synthetic fertilizer and cattle manure) have also grown by at least 30 percent over the past four decades.
“There’s a pressing need to shut off the valve as quickly as possible, but you can’t manage what you can’t measure,” said Licheng Liu, the lead author of the study and research scientist from the University of Minnesota’s Digital Agriculture Group in the Department of Bioproducts and Biosystems Engineering.
Estimating nitrous oxide from cropland is an extremely difficult task because the related biogeochemical reactions involve complex interactions with soil, climate, crop, and human management practices—all of which are hard to quantify. Although scientists have come up with different ways to estimate nitrous oxide emission from cropland, most existing solutions are either too inaccurate when using complex computational models with physical, chemical, and biological rules or too expensive when deploying sophisticated instruments in the fields.
In this new study, researchers developed a first-of-its-kind knowledge-guided machine learning model for agroecosystem, called KGML-ag. Machine learning is a type of artificial intelligence that allows software applications to become more accurate at predicting outcomes without being explicitly programmed to do so. Previous machine learning models have been criticized, however, for being a “black-box” where scientists can’t explain what happened between inputs and outputs. Now, scientists have developed a new generation of methods that integrates scientific knowledge into machine learning to unpack the “black-box.”
KGML-ag was constructed by a special procedure that incorporates the knowledge learned from an advanced agroecosystem computational model, called ecosys, to design and train a machine learning model. In small, real-world observations, the KGML-ag turns out to be much more accurate than either ecosys or pure machine learning models and is 1,000 times faster than previously used computational models.
“This is the first-of-its-kind journey with ups and downs because there’s almost no literature to tell us how to develop a knowledge-guided machine learning model that can handle the many interactive processes in the soil, and we’re so glad things worked out,” Liu said
One unique feature of KGML-ag is that it goes beyond most machine learning methods by explicitly representing many less obvious variables related to nitrous oxide production and emission. It also captures the complex causal relationship among inputs, outputs, and other complex intermediate variables.
“Knowing these intermediate variables, such as soil water content, oxygen level, and soil nitrate content, are very important because they inform drivers of nitrous oxide emissions, and give us possibilities to reduce nitrous oxide,” said the corresponding author, Zhenong Jin, a University of Minnesota assistant professor in the Department of Bioproducts and Biosystems Engineering who also leads the Digital Agriculture Group.
The development of the KGML-ag was inspired in part by pioneering research on knowledge-guided machine learning in environmental systems led by Vipin Kumar, a University of Minnesota Regents Professor in the Department of Computer Science and Engineering and the William Norris Chair. This research includes studies for lake temperature predictions and streamflow predictions.
“This is another success story of computer scientists working closely with experts in agriculture and the environment to better protect our Earth,” Kumar said. “This new effort will further enhance existing knowledge-based machine learning activities that the University of Minnesota is currently leading nationally.”
In the future, the team will expand KGML-ag for predicting the carbon emissions from the soil using a variety of factors, including high resolution satellite imagery.
“This is revolutionary work that brings together the best of observational data, process-based models, and machine learning by integrating them together,” said Kaiyu Guan, a coauthor of the study and an associate professor at the University of Illinois at Urbana-Champaign.
Guan is also the lead researcher of the Department of Energy’s Advanced Research Projects Agency-Energy (ARPA-E) Systems for Monitoring and Analytics for Renewable Transportation Fuels from Agricultural Resources and Management (SMARTFARM) project that funds this study.
“We are really excited to continue this collaboration with the University of Minnesota team led by Zhenong Jin to explore and realize the full potentials of KGML,” Guan added.
Accurate, scalable, and cost-effective monitoring and reporting of greenhouse gas emissions are needed to verify what are called “carbon credits” or permits that offset greenhouse gas emissions. Farmers can be reimbursed for practices that reduce greenhouse gas emissions. The KGML-ag framework opens tremendous opportunities for quantifying the agricultural nitrous oxide, carbon dioxide, and methane emissions, helping to verify carbon credits and optimize farming management practices and policy making.
“There is a lot of excitement around the potential for agriculture to contribute to carbon drawdown, but unless we have accurate and cost-effective measurement tools to assess what is happening both above- and below-ground, we won’t see the market incentives we know are necessary to facilitate a transition to net-negative agriculture,” said David Babson, a program director with the U.S. Department of Energy’s ARPA-E.
“The teams working together from Minnesota, Illinois, California and Pennsylvania understand this,” Babson added. “I’m looking forward to the teams further expanding this research.”
Geoscientific Model Development
Method of Research
KGML-ag: a modeling framework of knowledge-guided machine learning to simulate agroecosystems: a case study of estimating N2O emission using data from mesocosm experiments
Article Publication Date