Wenquan Dong

Accurate forest aboveground biomass (AGB) estimation across heterogeneous subtropical regions is essential for carbon accounting and climate change mitigation. We developed XGBoost and random forest models using GEDI L4A Lidar samples and multi-source remote sensing features (Sentinel-1/2, topography) to predict AGB in Xijiang Forest Farm (Guangdong) and transferred them to Simao District (Yunnan). XGBoost demonstrated superior performance and transferability, with parameter fine-tuning effectively adapting the source-domain model to the target region (R² = 0.48) using only 20% of target samples, while full retraining achieved R² = 0.52. SHAP analysis identified spectral indices (SIPI and SAVI) and SAR backscatter (VH) as key predictors. Monte Carlo uncertainty decomposition revealed GEDI measurement error accounts for 36%–39% of total predictive uncertainty. This transfer learning framework enables cost-effective AGB mapping in data-limited regions, supporting regional carbon monitoring and forest management.