Deep Learning Insights into Seamless Reconstruction of XCO2 in China: Spatiotemporal Patterns and Driving Mechanisms

Accurate quantification of atmospheric column-averaged dry-air CO2 mole fractions (XCO2) is pivotal for quantifying carbon sources and supporting China’s dual carbon goals. However, existing satellite observations are limited by spatiotemporal gaps due to orbital constraints and atmospheric conditions. To bridge these gaps, we utilized a deep learning framework featuring a dual self-attention mechanism, Air-Transformer, to capture complex long-range spatiotemporal dependencies and non-linear interactions among variables. Utilizing OCO-2 retrievals and multi-source data, this approach generated a spatiotemporally consistent, daily 0.1° XCO2 dataset over China during 2015–2020. Cross-validation demonstrates superior accuracy (R2 = 0.98), with robust performance confirmed by spatial and temporal validation and ground-based TCCON benchmarks. The estimated full-coverage outputs reveal a national mean annual increase of 2.68 ppm, characterized by a distinct east-high/west-low pattern. Interpretable analysis based on Shapley Additive Explanations (SHAP) elucidates the non-linear interactions between XCO2 and environmental drivers and exhibits significant regional heterogeneity. This spatiotemporally consistent and interpretable XCO2 dataset offers vital data support for regional carbon monitoring and differentiated policy-making.