Construction of a 0.01° Monthly Seamless XCO₂ Dataset over China: Based on a Temporally Adaptive Forest Model

Abstract High-resolution column-averaged carbon dioxide (XCO₂) data are of significant importance for accurately characterizing regional carbon emission patterns and constraining carbon flux processes, particularly with urgent demand in urban and regional carbon management practices. However, most existing XCO₂ products are based on static modeling assumptions, making it difficult to overcome concept drift in long-term time series data. This results in insufficient reliability at fine regional scales and an inability to effectively capture the non-stationary characteristics of the time series. To address this, this study integrates OCO-2/3 observations with multi-source environmental data and proposes a Temporal Adaptive Forest model. This model employs time decomposition based on the piecewise stationary approximation theory and adopts a feature-space adaptive alignment strategy. Through annual training and a dynamic one-hot encoding mechanism, it effectively mitigates the modeling challenges posed by temporal non-stationarity and enables dynamic diagnosis of driving factors. Based on this model, a seamless monthly XCO₂ dataset covering China with a spatial resolution of 0.01° was constructed. Independent validation results from TCCON stations indicate that this dataset can provide crucial data support for precise emission source identification, separation of carbon sources and sinks, and fine-scale characterization of regional carbon emission patterns.