Understanding forest carbon storage has been crucial for the complete and accurate quantification of carbon emissions and environmental monitoring in the context of climate change. The use of remote sensing data sources has proven to be an effective method for estimating carbon reserves at the regional scale. Taked Jiangxi Province as the research area, using the seventh National Forest Continuous Inventory (NFCI) data along with Landsat-5 TM remote sensing data. Image processing steps were performed on the Google Earth Engine (GEE) platform. Forest carbon storage was estimated by employing two feature selection methods(recursive feature elimination (RFE) and Boruta), combined with support vector regression (SVR) and three ensemble learning algorithms, including random forest (RF), extreme gradient boosting (XGBoost), and stacking integration (Stacking), to comparatively analyze the estimation accuracies of different models in detail. In addition, the Optuna hyperparameter optimization framework was used to determine the hyperparameters of each model. Inverted the forest carbon storage in Jiangxi Province based on the optimal estimation model and drew a spatial distribution map. The driving forces of carbon stocks' spatial distribution patterns were analyzed using geographic detectors. The results show that: (1) According to the feature importance ranking, RFE screened out 30 variables and Boruta screened out 11 variables. The combination of appropriate feature subsets and regression algorithms can significantly improve the accuracy of estimation. (2) Optuna-based iteratively adjusted the hyperparameters of each model. It was found that when different subsets of features were combined with machine learning algorithms, there was a significant disparity in the importance and values of hyperparameters within the model. The optimal feature subset was screened using RFE, which achieved the best predictive performance when utilized in regression simulations combined with Stacking models (R²=0.527, RMSE=15.85Mg/hm², MAE=12.31Mg/hm²). The model effectively utilized the training data and combined the advantages of multiple algorithms to reduce bias, which significantly improved the problems of underestimation of high carbon density values and overestimation of low carbon density values. (3) The optimal estimation model was inverted to obtain the average forest carbon density in Jiangxi Province in 2006 was 33.356Mg/hm² (2.585-88.943Mg/hm²), and the total forest carbon stock was 321.507Tg. (4) Among the natural environment factors, elevation and slope were the main driving factors influencing the spatial distribution pattern of carbon stocks. All factors showed nonlinear enhancement and two-factor enhancement under interaction. The spatial distribution pattern of carbon storage was the result of the synergistic effect of natural and anthropogenic factors.