Aim: Motivated by the need to enhance the accuracy of forest aboveground carbon storage (ACS) assessments, this study aimed to explore the effectiveness of different machine learning models in predicting ACS across various subtropical forest types in southern China. Location: The study was conducted in southern China, focusing on different types of subtropical forests. This region harbors several types of subtropical forests, which are rarely found at similar latitudes in the world. Methods: Variance inflation factor was employed to screen independent variables, resulting in the selection of 13 significant predictors. Four machine learning models—support vector machine (SVM), random forest (RF), multi-layer perceptron (MLP), and extreme gradient boosting (XGB)—were constructed to estimate carbon storage. Model performance was evaluated using root mean square error, coefficient of determination (R 2), and mean absolute error. The model with the best generalization ability was selected to calculate SHAP values for each predictor. Results: The XGB model demonstrated superior performance across all forest types, with R 2 values ranging from 0.898 to 0.974. In mountainous evergreen broad-leaved forests, the prediction accuracy followed the order of XGB > MLP > SVM > RF. In valley rainforests, MLP showed the highest R 2 value, but with higher MAE and RMSE, making it the second-best choice. The RF model performed moderately, while the SVM model showed the poorest performance. The SHAP values indicated that maximum diameter at breast height, slope, mean DBH, species evenness, altitude, and maximum tree height had significant effects on ACS. Conclusions: XGB model exhibits the best prediction performance and strongest adaptability for estimating ACS in subtropical southern China forests.Machine learning methods provide valuable references for predicting and assessing ACS in different types of zonal forests.