An RF-Guided Dual-Strategy Feature-Selection Framework for Multi-Source Remote-Sensing-Based Estimation of Forest Aboveground Carbon Stock in Mountainous Terrain

This study proposed an RF-guided heuristic feature-selection framework that integrates multi-source remote-sensing data for estimating Pinus densata aboveground carbon stock (AGCS) in Shangri-La, Yunnan Province, China. Compared with four baseline feature-selection methods, the Random Forest–Alpha Evolution (RFA) and Random Forest–Markov Chain Monte Carlo (RFM) algorithms generated more informative feature subsets and improved model performance, with the Optuna-optimized AdaBoost model based on RFM features achieving the highest accuracy (R2 = 0.71, RMSE = 10.53 t/ha). These results suggest that RF-guided heuristic feature selection can effectively improve AGCS estimation in complex mountainous environments. Vegetation indices and texture features were consistently prioritized across different feature-selection methods. Shapley Additive Explanations (SHAP)-based interpretation revealed that the most influential predictors were the Sentinel-2A green normalized difference vegetation index (S2_GNDVI) and precipitation of the wettest month (bio13) in the RFA Method, and the Sentinel-2A red-edge normalized difference vegetation index (S2_NDVI45) and bio13 in the RFM Method. These findings underscore the critical importance of canopy greenness, moisture availability, and structural complexity in regulating carbon accumulation in montane conifer forests. The final AGCS maps yielded total estimates of 9.83 Mt (RFA) and 10.46 Mt (RFM), and revealed a consistent spatial pattern, with moderate AGCS values dominating the landscape and a general tendency for higher values in the northwest and lower values in the southeast. In summary, the combination of RF-guided heuristic feature selection, Optuna-optimized machine learning and SHAP provides an effective and interpretable framework for AGCS estimation in mountain forests.