Forest carbon stock estimation in a mountainous region using locally adaptive machine learning approaches

Accurate estimation of forest carbon stocks is essential for climate change mitigation, particularly in mountainous regions where strong spatial heterogeneity constrains the performance of conventional remote sensing models. This study investigates whether incorporating local sample weighting and robust parameter optimization can enhance the reliability of forest carbon stock estimation under spatial heterogeneity. A locally weighted regression framework based on Support Vector Regression was developed by integrating neighborhood-based sample weighting with a global hyperparameter optimization strategy and applied to Huoshan County, Anhui Province, China. The analysis was based on Landsat optical data and DEM-derived topographic variables, combined with field-based forest inventory data. Results indicate that the proposed approach achieved higher predictive performance than five commonly used regression models, with a test-set coefficient of determination (R²) of 0.763 and a root mean square error (RMSE) of 6.72 t·ha^-1, and demonstrated improved robustness in spatially heterogeneous forest areas. The findings suggest that locally weighted machine learning methods provide an effective and reliable tool for regional forest carbon stock mapping in complex mountainous environments.