Mapping agroforestry Above-Ground Biomass using GEDI-derived structural traits and multisensor fusion of Sentinel-1/2 and ALOS-2 PALSAR-2

Agroforestry systems are highly heterogeneous and often underrepresented in global biomass products, creating challenges for carbon stock estimation. Accurate mapping of Above-Ground Biomass (AGB) and canopy structure is therefore essential for improving monitoring and reporting. In this study, we evaluate the use of GEDI-derived canopy structural variables, namely canopy cover (CC), 95th percentile height (RH95), and foliage height diversity (FHD) as intermediate predictors for AGB estimation across smallholder agroforestry landscapes. Using Sentinel-1, Sentinel-2, ALOS-2 PALSAR-2 and ancillary data inputs, we compared early, intermediate, and late multimodal fusion strategies within a U-Net architecture. Intermediate fusion achieved the best performance for structure retrieval, particularly for RH95 (r = 0.86 against airborne LiDAR). Incorporating FHD into an allometry-informed AGB model increased accuracy from an R 2 value of 0.64 to 0.74, reducing the RMSE from 8.23 to 7.07 ton/ha, and decreased the bias from −1.27 to −0.56 ton/ha. These results demonstrate that structural diversity captured by FHD provides additional information beyond height and cover, and that fusion design influenced retrieval performance. The framework uses GEDI-derived structural traits as intermediate supervision to improve AGB estimation in data-scarce agroforestry systems with an allometry-informed and interpretable AGB model. The approach offers a scalable way for integrating LiDAR missions with multisensor satellite data for AGB monitoring and agroforestry planning.