Phase-Preserving Trimodal Transformer for Tropical Forest Biomass Estimation Using Optical and PolInSAR Data

The paper introduces a physics-informed Trimodal Coherent Co-attention Transformer (TCCT) for estimating tropical-forest above-ground biomass. It combines Landsat-5 surface reflectance with complex-valued P- and L-band PolInSAR data. Complex encoders preserve spatial phase coherence, while dynamic co-attention reduces reliance on cloud-corrupted optical pixels. In experiments over the Paracou region of the Amazon, localized fine-tuning produced a 3.78 m CHM RMSE and an R² of 0.33. The resulting biomass map reportedly achieved a 4.51% relative RMSE in dense forest above 50 Mg/ha.