Machine learning approach for quantifying methane emissions from a flooded rice paddy in Chungcheong province

Machine learning (ML)-based estimation of methane (CH 4 ) emissions is essential to effectively assess the impacts of rice production and respond to climate change.This approach not only requires less labor and costs than field-based study, but also entails high reproducibility and applicability.However, optimal ML algorithms and combination of environmental covariates for forecasting CH 4 emissions are still lacking.Hence, the goals of this study were (1) to evaluate the regression performance of three ML algorithms (i.e., multiple linear regression, ridge regression, and random-forest) and (2) to assess the relative influence of each environmental factor within the predictive model for CH 4 emissions.The environmental covariates, which obtained from rice paddy field at Daejeon, South Korea in 2024, classified as followed: climatic factors (i.e., atmospheric, soil, and water temperatures and precipitation), fertilization method (i.e., quantity of nitrogen fertilization and fertilizer deep-placement application), and soil properties (i.e., oxidation-reduction potential and soil drainage).Our predictive model based on the random-forest algorithm outperformed than the linear-based ML algorithms.The random-forest algorithm achieved the highest reliability of 86.1%, and the mean value of the K-Fold cross-validation was 83.0%.The relative influence of each parameter was exhibited to be high in the order of atmospheric temperature (0.49), soil temperature (0.27), quantity of nitrogen fertilization (0.20), fertilizer deep-placement (0.17), and soil drainage (0.17).These findings revealed the requirement of explanation capacity for non-linear relationship in forecasting the CH 4 emissions, and need for collecting various environmental factors related to rice production.