Spaceborne LiDAR footprint-level AGB estimation and mapping across forest types using multisource remote sensing and U-Net

Spaceborne light detection and ranging (LiDAR), including ICESat-2 and GEDI, supports large-scale forest aboveground biomass (AGB) mapping, but accurate estimation across heterogeneous forest types remains challenging. This study developed forest-type-specific footprint-level AGB equations using ICESat-2 and GEDI observations and integrated the derived footprint AGB labels with Sentinel-1 and Sentinel-2 imagery in a U-Net framework to generate continuous AGB maps. After photon denoising and geolocation correction, GEDI RH95 showed stronger agreement with ALS-derived PH95 than ICESat-2 (R = 0.62–0.65 and 0.52–0.58, respectively), whereas ICESat-2 showed smaller bias magnitudes. At the footprint-level, the two sensors showed complementary performance. ICESat-2 performed better in broadleaf forests (R = 0.85, RMSE = 23.96 Mg/ha) and mixed forests (R = 0.89, RMSE = 20.39 Mg/ha), while GEDI performed better in coniferous forests (R = 0.85, RMSE = 22.50 Mg/ha). The U-Net model achieved improved regional AGB prediction with R = 0.91 and RMSE = 17.21 Mg/ha on the independent test set. These results demonstrate the value of combining ICESat-2, GEDI, Sentinel imagery, and deep learning for forest-type-specific AGB estimation and continuous biomass mapping.