A team of researchers at Michigan State University has developed a new machine learning system capable of predicting nitrous oxide emissions from U.S. croplands with significantly improved accuracy, offering potential advances for greenhouse gas accounting and mitigation strategies in agriculture.
The research was led by Prateek Sharma, a former Michigan State University graduate student, and Bruno Basso, a Hannah Distinguished Professor in the Department of Earth and Environmental Sciences and the W.K. Kellogg Biological Station. G. Philip Robertson, a University Distinguished Professor at the Kellogg Biological Station and in the Department of Plant, Soil, and Microbial Sciences, co-led the study.
Nitrous oxide is a greenhouse gas commonly produced during agricultural operations through the use of nitrogen fertilizers. Predicting emissions has historically been challenging because emissions depend on complex interactions among weather conditions, soil characteristics, and crop management practices that influence microbial activity in agricultural soils.
The researchers addressed this challenge by creating a hybrid modeling system that integrates machine learning techniques with ecosystem models to estimate daily nitrous oxide emissions. The system was trained using more than 12,000 nitrous oxide measurements collected from 17 sites across the U.S. Midwest and Great Plains, covering six cropping systems and 35 management practices.
The dataset represents one of the most comprehensive collections of agricultural nitrous oxide measurements to date. By combining machine learning with process-based ecosystem modeling, the team developed an ensemble approach that more effectively captures the interactions between environmental conditions and farming practices that drive emissions.
According to the researchers, traditional single-model approaches typically achieve about 20% accuracy in predicting nitrous oxide emissions. In contrast, the new ensemble modeling system achieved more than 80% prediction accuracy.
The study was published in the journal Proceedings of the U.S. National Academy of Sciences. The research team also included Aditya Manuraj, Neville Millar, Tommaso Tadiello, Mukta Sharma, and Mathieu Delandmeter from Michigan State University’s Department of Earth and Environmental Sciences, as well as Michael Murillo from the Department of Computational Mathematics, Science and Engineering.
The project received support from multiple organizations, including the Great Lakes Bioenergy Research Center, the U.S. Department of Energy Office of Science, the National Science Foundation Long-Term Ecological Research Program at the Kellogg Biological Station, the USDA National Institute of Food and Agriculture, the USDA Long-term Agroecosystem Research Program, the CERCA–Foundation for Food and Agriculture Research Project, Climate Trace, the Soil Inventory Project and Michigan State University AgBioResearch.
KEY QUOTES:
“One of the limiting factors of current predictive models is that they rely on outdated national greenhouse gas emission inventories and often need to be calibrated to a specific site. With this effort, we’ve moved past these limitations to provide management-specific predictions for crucial combinations of cropping systems, soils, management practices and weather conditions. We’re hopeful this approach can lead to field-specific emission mitigation strategies, as well as much-needed updates to estimates of greenhouse gas emissions from agriculture.”
Bruno Basso, Hannah Distinguished Professor, Department of Earth and Environmental Sciences, Michigan State University
