LAPFuse: Language-Anchored and Physics-Informed Fusion of Sentinel and GEDI Data for Forest Aboveground Biomass Estimation

Accurate forest aboveground biomass (AGB) estimates are needed for forest inventories and carbon assessment, yet optical, synthetic aperture radar (SAR), and spaceborne lidar products describe forest properties at different spatial supports. We present LAPFuse, a multimodal framework that combines dense Sentinel-2 and Sentinel-1 observations with sparse GEDI Level 4A footprint-level aboveground biomass density records. Modality-specific encoders preserve the sensing characteristics of each input. Fixed language anchors regularize their projected representations, while distance-weighted sparse–dense attention relates GEDI footprint tokens to spatially distributed Sentinel features. A physics-informed neural process then produces a predictive mean and interval using an allometry-based soft constraint. We evaluate LAPFuse with 1847 permanent plots from the 2022–2024 growing seasons in three forest regions of China. On the held-out set of 369 plots, LAPFuse achieves R2=0.882, RMSE =15.34 Mg ha−1, and MAE =9.87 Mg ha−1. Its 95% prediction intervals attain a coverage probability of 0.949 with a mean width of 67.2 Mg ha−1. Sequential ablation shows progressive improvements as the alignment, attention, probabilistic, and allometric components are introduced. Zero-shot cross-region R2 ranges from 0.762 to 0.798, and adaptation with 100 target-region plots reduces the performance gap to 0.1–1.5% relative to in-domain performance. These results demonstrate the value of support-aware multimodal fusion for plot-level AGB estimation under this sensor configuration and these regional conditions.