Growing Stock Volume Estimation in Complex Mountainous Terrain: Sentinel-2 Topographic Correction and Fusion with Fully Polarimetric SAR Data

Growing stock volume (GSV) estimation in complex mountainous forests is complicated by terrain-induced radiometric distortions and the heterogeneous characteristics of multisource remote-sensing data. This study evaluated the effects of topographic correction and optical–microwave fusion on GSV estimation. Data from 93 field plots in Beishan Forest Farm, northeastern Qinghai Province, China, were integrated with Sentinel-2 imagery, fully polarimetric SAOCOM SAR data, and a digital elevation model (DEM). Seven Sentinel-2 topographic correction schemes were compared, and SAOCOM data before and after radiometric terrain correction (RTC) were evaluated using multiple feature-selection strategies and machine-learning models. SHapley Additive exPlanations (SHAP) were further used to quantify feature contributions. The results showed that the slope-stratified VECA correction achieved the highest mean GSV estimation performance among the Sentinel-2 configurations (R2 = 0.727; RMSE = 49.736 m3·ha−1). The optical–microwave fusion model combining Boruta + RF-RFECV feature selection with random forest achieved the highest mean performance (R2 = 0.734; RMSE = 48.906 m3·ha−1), with a modest 1.67% reduction in RMSE relative to the selected optical model. Cross-validated SHAP analysis further showed that Sentinel-2 features accounted for 92.98% ± 3.54% of the total normalized absolute SHAP contribution, compared with 7.02% ± 3.54% for SAOCOM features. The B5 red-edge band and its multi-window mean-texture features showed the highest SHAP contributions among optical predictors, while the MCSM and An–Yang4 volume-scattering components were among the leading SAOCOM predictors. These findings highlight the potential value of topographic correction for GSV estimation in complex mountainous forests, with the addition of fully polarimetric SAOCOM data providing only a small numerical improvement in predictive performance.