Abstract: Forest ecosystems serve as the largest carbon sink on land and play a critical role in the global carbon cycle and climate change mitigation. The accuracy of forest carbon stock estimation significantly affects the development of reliable carbon cycle models and the effective implementation of carbon neutrality policies. Conventional methods for dynamic carbon stock assessment are generally conducted at the stand scale, which lacks the resolution to capture carbon stock variation at the individual tree level. To address this limitation, this study investigated 12 poplar (Populus L.) plantation plots with varying planting spacings and clones. Using high-resolution terrestrial laser scanning (TLS) data acquired in 2019 and 2021, multiple fundamental tree structural metrics were extracted via a novel algorithm of the incomplete simulation of tree transmitting water and nutrients (ISTTWN), which relies on geometric characteristics for extracting tree branch skeletons. Following data extraction, correlation analysis and multicollinearity diagnostics were performed for both linear and nonlinear models to identify significant predictors. The Boruta algorithm was applied for variable selection to four machine learning models: random forest, k-nearest neighbors, support vector machine, and CatBoost. The Optuna algorithm was then used to optimize hyperparameters based on negative mean squared error during 5-fold cross-validation. The best-performing model was selected for estimating individual tree aboveground carbon stock and carbon stock growth. The study employed both direct and indirect methods to estimate aboveground carbon stock growth at the individual tree level. The direct approach modeled growth directly by selecting relevant variables and constructing a predictive model. In contrast, the indirect approach treated carbon stock as the dependent variable, estimated stocks at two time points for the same tree, and derived growth by difference. By comparing the accuracy of both methods, this study aimed to determine the optimal approach for estimating individual tree aboveground carbon stock growth and to identify the best planting configuration for poplar trees in the study area. Results showed that the Random Forest model, built with Boruta-selected variables, outperformed other models, achieving an R² of 0.944 for estimating aboveground carbon stock and an R² of 0.798 for carbon stock growth. Specifically, the direct estimation method using Random Forest produced the most accurate predictions for growth, with an R² of 0.821, a root mean square error (RMSE) of 0.920 kg, and a mean absolute error (MAE) of 0.733 kg. These results demonstrate high precision and robustness, highlighting the ability of machine learning to capture complex nonlinear relationships between tree structural attributes and carbon dynamics. Furthermore, the study found that NL-797 poplar clones planted at a spacing of 6 m × 6 m not only achieved the highest absolute carbon stock growth per tree but also exhibited a high carbon accumulation rate during the study period. This indicates that combining specific genotypes with optimal planting density can significantly enhance carbon sequestration in poplar plantations. In conclusion, this study proposes a non-destructive and highly accurate methodology for estimating aboveground carbon stock growth in individual trees using TLS and advanced machine learning models. The findings provide valuable insights for forest management strategies aimed at maximizing carbon storage and offer a scientific basis for the design and management of carbon-oriented plantation forests.