Joint Prediction of U.S. Rice Yields and Methane Emissions: A Machine Learning Approach

Despite the United States’s major role as a rice exporter with significant economic and environmental impacts, most remote sensing and machine learning research on rice yield prediction has focused on Asian production regions. Moreover, while rice cultivation generates substantial methane emissions, few studies have explored the trade-offs between productivity and environmental impact. This study leverages remote sensing and machine learning techniques to predict U.S. county-level rice yields and methane emissions from 2008 to 2022 across 67 counties in six major rice-producing states. We use eight different machine learning models for predictions. XGBoost and EBM emerge as top performers, accurately predicting yields and emissions, individually, without overfitting. A key finding reveals that these models excel at out-of-season forecasts, accurately predicting yields as early as April-June of the growing season. Feature importance analysis highlights soil properties, particularly pH and texture at various depths, as critical predictors for both yield and emissions. Most importantly, this study advances an integrated economic-environmental modeling in agriculture by analyzing yield-emissions trade-offs through the Non-dominated Sorting Genetic Algorithm II (NSGA-II), revealing an unexpected synergy where practices that improve economic productivity also reduce environmental impact, as higher yields correlate with lower methane emissions.