A Multi-Temporal Machine Learning Framework for National-Scale Mapping of Methane Emissions from Rice Paddies

Rice paddies are a dominant source of methane emissions in the global agricultural sector, accounting for approximately 40% of total anthropogenic methane emissions. However, methane fluxes vary strongly with crop growth stages and environmental conditions, which introduces substantial uncertainty in quantitative assessments at the national scale. Although chamber measurements and eddy covariance observations can directly capture these emission processes, their spatial representativeness is limited. To address this limitation, approaches that integrate remote sensing data with machine learning have been proposed as effective alternatives. Nevertheless, the selection of temporal resolution in model development requires a careful balance between analytical objectives and computational efficiency. In this study, we developed separate daily and hourly machine-learning models to estimate methane emissions from rice paddies in South Korea and systematically compared their predictive performance and practical applicability across temporal scales.

Methane flux observations from four rice paddy sites in the FLUXNET-CH₄ network (IT-Cas, JP-Mse, KR-CRK, and US-HRC) were combined with MODIS-derived vegetation indices, ERA5-Land reanalysis meteorological and soil variables, and HWSD soil information to train FLAML AutoML regression models. The daily model showed strong predictive performance, achieving a correlation coefficient of 0.897 in 5-fold cross-validation and 0.819 in Leave-One-Year-Out validation. SHAP-based interpretation identified soil temperature as the dominant driver, explaining approximately half of the total variance. Under low-emission conditions, temperature effects were most pronounced, whereas during peak emission periods the relative importance of vegetation indices (NDVI and GNDWI) increased. The hourly model exhibited stronger interactions among NDVI, soil moisture, and evapotranspiration, revealing a more complex multivariate structure governing methane emissions. Although both models showed some nonlinear bias, these biases were largely offset at monthly and seasonal aggregation scales, remaining within ±1% and thus having negligible influence on long-term emission estimates.

These results indicate that temporal resolution should be selected strategically based on the intended application. The daily model is well suited for national greenhouse gas inventory development and long-term emission assessments, while the hourly model is more effective for detecting high-emission events and monitoring short-term variability. The modeling framework presented here provides a quantitative basis for mapping methane emission patterns in South Korean rice paddies and offers a methodological foundation for extension to agricultural systems across East Asia.