Improving Forest Carbon Stock Estimation: Integrating ICESat-2 LiDAR Data with Forest Classification and Hyperparameter Optimization

Accurate monitoring of forest carbon stocks represents a critical prerequisite for achieving carbon neutrality. However, conventional remote sensing‑based estimation methods frequently overlook forest heterogeneity, causing systematic overestimation or underestimation. To address this gap, we propose a novel forest carbon stock esti-mation framework that integrates two complementary strategies: (1) forest type‑specific modeling to account for forest heterogeneity, and (2) hyperparameter op-timization to enhance Random Forest model performance. Using ground‑measured carbon stocks and a CCDC‑derived forest vegetation classification map for Hangzhou City, China, we built forest‑type‑specific Random Forest models based on ICESat‑2 canopy height metrics and optimized each model via hyperparameter tuning. The re-sults show that the 70th-90th percentiles and the mean canopy height are relatively highly correlated with carbon stock. Forest‑type‑specific modeling improves estima-tion accuracy, yielding R² gains of 0.10–0.17 and reduced RMSE by 2.28–7.43 Mg C/ha over the non‑stratified model. Integrating forest classification and hyperparameter op-timization strategies improved model R² by 0.16–0.23 and lowered RMSE by 3.05–8.20 Mg C/ha. Overall, this study demonstrates that accounting for forest heterogeneity and applying hyperparameter optimization can significantly enhance the accuracy of for-est carbon stock estimation.