Remote sensing-based carbon stock estimation with interpretability analysis in boreal forests

China's state-owned boreal forests have shifted from timber production to ecological restoration, making carbon stock estimation vital for tracking degradation and recovery. However, common methods are labor-intensive and lack interpretability. We integrate remote sensing data with Recursive Feature Elimination with Cross-Validation to identify the optimal feature set from optical, radar, and environmental data. Then, two methods including Bayesian optimization with SHAP-based interpretability and TPOTRegressor with permutation feature importance, are compared in terms of automation and interpretability. Results show that feature optimization can reduce features and achieve higher accuracy. LightGBM model embeded in Bayesian optimization is the best compared with other machine learning models. The model has an R² of 0.59, close to that from TPOTRegressor (R² = 0.61), but requires less running time. Forest structure, soil moisture, and canopy dynamics are the key factors affecting carbon stock. This study offers a new perspective to improve carbon stock estimation accuracy and efficiency.