Machine learning approaches, XGBoost (XGB) and Random Forest (RF), were applied to estimate methane (CH₄) emissions from paddy fields in Gimje, South Korea, using three years of chamber-based observations. Both models achieved Nash–Sutcliffe efficiency (NSE) values above 0.5, indicating acceptatble predictive performance. XGB better captured extreme values and temporal variability, whereas RF reproduced mean emission trends more reliably, highlighting complementary strengths. Regional simulations produced mean emission factors of 2.64 and 2.35 kg CH₄ ha⁻1 d⁻1 for XGB and RF, respectively, comparable to or slightly higher than the Tier 2 country-specific factor (2.32 kg CH₄ ha⁻1 d⁻1) currently used in Korea’s national greenhouse gas inventory. Spatially, XGB identified high-emission hotspots with greater sensitivity, while RF yielded smoother patterns with lower uncertainty. Model-based regional estimates (6.68–7.51 Gg CH₄ yr⁻1) showed uncertainty ranges of ± 12–13%, substantially lower than the ± 50% default suggested by IPCC guidelines, underscoring the models’ robustness. These results demonstrate the feasibility of applying machine learning to CH₄ estimation at regional scales, thereby addressing the limitations of uniform Tier 2 emission factors. Incorporating region-specific, data-driven approaches can improve inventory accuracy and provide a stronger scientific basis for mitigation strategies. This study highlights the potential of machine learning models to complement existing methods and contribute to the advancement toward a Tier 3 inventory framework for agricultural greenhouse gas reporting.
Lee et al. (Mon,) studied this question.