This study integrated airborne LiDAR point cloud features with a hyperparameter optimization algorithm to develop a precise carbon stock estimation model for Pinus elliottii plantations in wetlands. The aim is to provide a scientific basis for large-scale carbon sink monitoring, thereby supporting forest management decisions that align with the "dual carbon" goals.

Taking Pinus elliottii plantation in Guangdong Province as the research object, 101 point cloud feature variables were extracted based on airborne LiDAR data. The recursive feature elimination combined with cross-validation (RFE-CV) was used to screen the optimal feature combination, and the Optuna hyperparameter optimization framework based on the TPE algorithm (with RMSE as the objective function) was introduced to construct random forest (RF), k-nearest neighbor (KNN) and extreme gradient boosting (XGBoost) models. The model performance was evaluated by R², RMSE and MAE, and the SHAP value was used to analyze the contribution mechanism of features to the prediction of carbon storage of Pinus elliottii.

1) Five optimal feature variables (A_IH40, A_IH80, H₁₀, H₂₅, I_max) closely related to the estimation of Pinus elliottii carbon storage were selected by 5-fold RFE-CV. Compared with the model using all variables, feature selection significantly improved the prediction accuracy of each model. Among them, the XGBoost model has the most obvious effect (R² increased from 0.78 to 0.89, RMSE decreased from 0.63 to 0.46 t, MAE decreased from 0.48 to 0.41 t); 2) The Optuna hyperparameter optimization framework based on the TPE algorithm is significantly better than random search: the RMSE value of the XGBoost model is reduced by 2.31%, and the error adjustment percentage (13.13%) is 3.57 times that of random search. The convergence iterations of the RF and KNN models are reduced by 33.33% and 49.18%, respectively; 3) After optimization by Optuna-TPE, the accuracy of each model has been significantly improved: the XGBoost model achieves the highest accuracy, with its R² significantly increasing from 0.89 to 0.93, thus becoming the optimal model; 4) SHAP value analysis revealed the difference of feature dependence in different slash pine models: A_IH40 dominated XGBoost prediction (SHAP = 0.64), while H₂₅ was more critical to the RF model (SHAP = 0.28).

This study showed that feature selection and Optuna hyperparameter tuning based on the TPE algorithm are crucial to improve the performance of the Pinus elliottii carbon storage estimation model. The optimized XGBoost model combined with the optimal feature combination realized the high-precision estimation of the carbon storage of Pinus elliottii plantations.