Diagnosing the spatiotemporal evolution of anthropogenic CO2 emission drivers in China under systemic disruptions using interpretable machine learning.

Understanding how anthropogenic CO2 emissions (ACE) respond to large-scale systemic disruptions is essential for climate mitigation and environmental management. Here, we develop an interpretable machine learning framework that integrates an XGBoost model with SHAP (SHapley Additive exPlanations) to diagnose how the key drivers of the ACE inventory's spatiotemporal variability evolved in mainland China, using the COVID-19 period as a natural experiment. The framework reliably reproduces spatial and temporal patterns of the inventory and enables systematic diagnostic analysis of predictor importance and interactions. Socio-geographic variables, including longitude, population, latitude and elevation, form a stable core associated with persistent spatial structure, whereas meteorological, biospheric and socioeconomic predictors contribute through dynamic and interactive influences. Temporal variables exhibit increasing importance over time, indicating growing temporal heterogeneity and regime shifts not fully captured by physical predictors alone. Interaction analysis further reveals evolving relationships among environmental and socioeconomic variables, including shifts in the coupling between vegetation activity and air-quality indicators and contrasting regional responses across western and eastern China. For example, the analysis identifies a differential decline of extreme pollution events between high- and low-GOSIF (Global Sun-Induced Chlorophyll Fluorescence) regions, as well as a coincidence of declining ACE with intensified PM2.5 pollution in western China. This study demonstrates how interpretable machine learning can serve as a diagnostic tool for examining inventory-consistent emission variability and generating hypotheses about how systemic disruptions reshape the predictor structure underlying emission patterns.