Regional forest stock volume mapping using GEDI-based interpolation, multi-source remote sensing, and a multi-level stacking ensemble model in complex terrain

Introduction Forest Stock Volume (FSV) is a key indicator of forest ecosystem functions, including productivity and carbon storage. Although spaceborne LiDAR such as GEDI provides high-resolution information on forest vertical structure, its discrete sampling pattern poses challenges for generating continuous spatial representations. Methods This study developed an integrated framework for FSV estimation in Pinus kesiya var. langbianensis forests in Jingdong County, Yunnan, China, by combining GEDI LiDAR metrics, Landsat-8 imagery, topographic variables, and field measurements from 143 sample plots. Six spatial interpolation methods were compared to transform discrete GEDI observations into continuous predictor layers. A multi-level stacking ensemble model (MLSEM) integrating six base machine-learning algorithms was then constructed for FSV estimation. Results Among the tested interpolation methods, Sequential Gaussian Conditional Simulation (SGCS) performed best, and the SGCS-derived variables generally achieved R 2 values greater than 0.50. The final stacked model outperformed all base learners and single-level meta-learners, achieving an R 2 of 0.93 and an RMSE of 9.50 m 3 /ha. The total standing stock in 2019 was estimated at approximately 8.66 × 10 7 m 3 , with a mean volume of 43.89 m 3 /ha. Discussion These results demonstrate the value of combining interpolated GEDI structural information with multi-source remote-sensing predictors for regional FSV mapping in complex mountainous terrain. The resulting spatially explicit FSV map can provide useful support for forest inventory updating, productivity monitoring, and management prioritization.