Spatially Complete Monthly XCH4 Mapping over China from 2019 to 2024: A Multi-Model Analysis Using Machine Learning

Abstract

Satellite retrievals of the dry-air column-averaged mole fraction of methane (XCH4) contain substantial spatial gaps, limiting monthly comparisons over a fixed domain. We developed CatBoost (CATB), random forest (RF), and multilayer perceptron (MLP) models to reconstruct spatially complete monthly XCH4 fields over China at 0.1° resolution from 2019 to 2024. CATB achieved the best overall performance and showed high consistency with Total Carbon Column Observing Network (TCCON) column observations, with correlation coefficients of 0.94 and 0.96 at Xianghe and Hefei, respectively. On commonly observed grid cells, CATB spatial out-of-fold predictions reproduced pure monthly averaged TROPOspheric Monitoring Instrument (TROPOMI) observations, with a correlation of 0.998 for monthly national means and a mean monthly spatial correlation of 0.845. However, valid TROPOMI retrievals covered only 36.16% of possible China grid-months. The trained models were therefore applied to every China land-grid-month to provide model-predicted fields over a fixed domain. The three models produced similar spatial patterns, with relatively high XCH4 in Shanxi, the Sichuan Basin, and the Yangtze River Delta. Predictor-contribution and ablation analyses identified NO2 as informative, although its covariation with XCH4 may partly reflect shared TROPOMI retrieval and sampling sensitivities. The dataset supports analyses of regional monthly XCH4 variations.