A spatio-temporal machine learning method for estimating high-resolution XCO2 in China

Column-averaged dry air mole fraction of carbon dioxide(XCO 2 ) data is of great significance for addressing global climate change, monitoring carbon emissions. Currently, satellite XCO 2 exhibit significant spatial discontinuity, which makes it difficult to meet the needs of research at small spatial scales. Although machine learning methods have been widely used to fill the gaps in satellite XCO 2 data, mainstream methods are mostly data-driven mode, which, to some extent, limits the accuracy and generalization ability of the models. Given the limitations of existing studies in mining the spatiotemporal characteristics of XCO 2 , this study innovatively proposes a new spatiotemporal XGBoost model(XGBKT) to generate high-resolution XCO 2 dataset covering the entire territory of China. This model focuses on the three major spatiotemporal characteristics of XCO 2 , namely spatial correlation, temporal heterogeneity, and temporal periodicity. Through the spatiotemporal encoding strategy, these characteristics are skillfully transformed into features that the XGBoost model can efficiently utilize, thereby enabling the model to explore the spatiotemporal distribution pattern of XCO 2 and significantly improve its estimation accuracy and reliability. The research results indicate that: The XGBKT model significantly enhances the estimation performance and generalization ability of machine learning models, demonstrating clear advantages compared to mainstream machine learning methods; The XGBKT model validates the effectiveness of spatiotemporal characteristics, thereby further strengthening the interpretability of machine learning models. Overall, XGBKT is an effective method for accurately estimating XCO 2 , providing a reliable data foundation for the fine-scale quantification of regional carbon cycling.