Estimation of Global Near-Surface Methane Mixing Ratios Using Machine Learning and Tropospheric Monitoring Instrument (TROPOMI) Observations

Abstract

Methane (CH₄) is a potent greenhouse gas with a global warming potential approximately 84 times greater than carbon dioxide (CO₂) over a 20-year time horizon, accounting for at least 25% of current climate forcing. Since the Industrial Revolution, atmospheric CH₄ mixing ratios have more than doubled due to diverse natural and anthropogenic sources, including wetlands, fossil fuel extraction, agriculture, and landfills. While ground-based in situ observations offer high temporal resolution, their limited spatial coverage hampers global assessments. Satellite-based column-averaged CH₄ (XCH₄) retrievals enable large-scale monitoring but present challenges in directly estimating near-surface CH₄ mixing ratios due to vertical profile uncertainty (Hasekamp et al., 2019). This study aims to estimate global near-surface CH₄ mixing ratios by integrating Tropospheric Monitoring Instrument (TROPOMI) XCH₄ with meteorological and land cover variables using machine learning. Among four models (RF, XGBoost, LGBM, MLR), Random Forest (RF) demonstrated the best performance, with R = 0.74, RMSE = 0.033 ppmv, and MAPE = 1.23% for the validation dataset. Variable importance was assessed using both mean decrease in impurity and permutation methods, identifying TROPOMI XCH₄, tropospheric NO₂, and day of year as key predictors. Despite multicollinearity among certain variables (e.g., surface temperature, vegetation indices), stable and interpretable results were achieved. The RF model successfully estimated near-surface CH₄ mixing ratios over four continents from 2018 to 2024, capturing typical seasonal patterns in the Northern Hemisphere with winter maxima (~2.004 ppmv) and summer minima (~1.980 ppmv). Six-year means in key regions such as Western Europe and Northeast Asia ranged from 1.96 to 2.06 ppmv, with model–observation differences within ±0.005 ppmv, even in areas lacking ground-based stations. Temporal analysis at representative sites revealed that the model reproduced seasonal cycles with high fidelity at background stations (e.g., KIT, MAPE = 0.003%), but showed larger errors at sites influenced by local emissions or satellite retrieval biases (e.g., STE, MAPE = 1.11%).