Machine Learning for Multi‑Scale Carbon Emission Prediction and Environmental Risk Assessment

Accurate prediction of carbon emissions is crucial for addressing climate change and protecting ecological safety, but many existing models lack rigorous validation on independent data. This study develops and compares two machine learning models—random forest (RF) and convolutional neural network (CNN)—for predicting CO₂ emissions at both international (163 countries) and sub-national (284 Chinese cities) scales. Using two distinct datasets, models were trained and optimized via grid search and 10-fold cross-validation, then evaluated on fully independent test sets. For the global dataset, RF achieved superior performance (test set R² = 0.984, RMSE = 0.427), while for the more complex urban dataset, CNN demonstrated better generalization (test set R² = 0.890, RMSE = 0.199). SHAP analysis revealed key drivers: trade openness and GDP were dominant at the country level, whereas economic structure, carbon sequestration, and urban form factors played significant roles at the city level. The study highlights the importance of external validation and shows that model performance depends on data scale and feature complexity, providing robust tools for emission forecasting, environmental risk assessment, and policy support.