Conversion of surface CH4 concentrations from GOSAT satellite observations using XGBoost algorithm

• Published in: Atmospheric environment (1994) Vol. 301; p. 119694
• Main Authors: Wan, Yong, Chen, Fangfang, Fan, Lu, Sun, Dong, He, Hu, Dai, Yongshou, Li, Ligang, Chen, Yuyu
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.05.2023
• Subjects: Methane; Near-surface CH4; Satellite remote sensing; XGBoost; Methane; Satellite remote sensing; XGBoost; Near-surface CH4
• ISSN: 1352-2310; 1873-2844
• DOI: 10.1016/j.atmosenv.2023.119694

Methane (CH4), the second most significant greenhouse gas after carbon dioxide, contributes significantly to global warming. Owing to its wide monitoring range and long observation time, satellite remote sensing has emerged as a popular method for monitoring CH4. Although the existing algorithm for satellite retrievals of CH4 column amount is mature, it has some drawbacks: the retrieval calculation process is complicated, and there is still a discrepancy between the satellite-retried column amount of CH4 and surface CH4 concentrations. To obtain more accurate near-surface CH4 concentrations from satellite observations, this paper proposed a conversion method based on the Extreme Gradient Boosting (XGBoost) algorithm, taking column amount retrieved by the Greenhouse Gas Observation SATellite (GOSAT) satellite, meteorological factors and near-surface methane concentrations as predictor variables. Using this method, we analyzed the importance of the characteristic factors and predicted methane concentrations at four ground monitoring sites in World Center for Greenhouse Gases (WDCGG), and reached the following conclusions: (1) The model performed well on the test set, with a sample-based cross-validation coefficient of determination (R2) of 0.79, a root mean square error (RMSE) of 0.0251 ppm and a mean absolute percentage error (MAPE) of 0.88%. (2) Of the five meteorological input features, surface net solar radiation had a greater impact on the construction of the model than the other four. (3) The CH4 monthly average concentrations predicted by this model are generally consistent with the trend of the CH4 concentrations of ground monitoring stations. •A machine learning method is used to obtain surface CH4 concentrations reliably.•The R2 between the model estimated value and surface CH4 concentrations is 0.79.•The importance of features in the model is analyzed.