Abstract: Using field-measured data collected in 2019 and 2021, this study focused on Pinus densata forests in Shangri-La City. Sentinel–1A, Sentinel–2A and Landsat 8 OLI optical remote sensing data, environmental variables, and DEM-derived topographic factors were integrated to estimate aboveground carbon stock (AGC). The Boruta algorithm was employed for feature selection, and four machine-learning models, including Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), Random Forest Regression (RFR), and eXtreme Gradient Boosting (XGBoost), were developed. Furthermore, the Simulated Annealing (SA) algorithm was applied to optimize the hyperparameters of each model. The optimal model was subsequently used to map the spatial distribution of aboveground carbon stock in 2021.The results indicated that the integration of multi-source remote sensing data achieved higher estimation accuracy than any single remote sensing data source. Hyperparameter optimization using the SA algorithm effectively improved model performance across all machine-learning models. Among them, the SA-optimized XGBoost model exhibited the best performance, with an R² of 0.64, an RMSE of 11.75 t·hm⁻², an MAE of 8.87 t·hm⁻², and an rRMSE of 28.33%. The estimated aboveground carbon stock of Pinus densata forests in 2021 was 9.92 Mt. Spatially, carbon stock exhibited a distinct pattern of being higher in the northwest and lower in the southeast. Areas with high carbon stock were primarily concentrated in the central and northwestern regions, whereas low-carbon-stock areas were mainly distributed in the eastern and southeastern regions.These findings demonstrate that the integration of multi-source remote sensing data with simulated annealing-based hyperparameter optimization can effectively improve the accuracy of forest carbon stock estimation and provide a reliable approach for regional forest carbon assessment.