Researchers have built a deep learning model that fuses radar and satellite imagery to measure Amazon forest biomass — and it meets the European Space Agency's accuracy target.
A team developed the Trimodal Coherent Co-attention Transformer (TCCT), combining optical surface reflectance data from Landsat-5 with Polarimetric SAR Interferometry (PolInSAR) data across P and L radar bands. The key engineering move is using complex-valued encoders that preserve the radar signal's phase information — something conventional fusion approaches discard. A co-attention mechanism automatically down-weights cloud-corrupted optical pixels and shifts reliance to radar when the view from space is blocked. In testing on the Paracou site in French Guiana, the model hit a relative biomass error of 4.51% in dense forest and a canopy height RMSE of 3.78 meters, outperforming Random Forest, CNN, and standard Vision Transformer baselines.
Accurately measuring how much carbon sits in tropical forests is not an academic exercise. It feeds carbon offset markets, national emissions inventories, and the climate models governments rely on. Cloud cover and radar signal saturation have long made dense tropical regions one of the harder remote sensing targets; a model that handles both constraints at once could be practical infrastructure for the ESA's BIOMASS satellite mission, which sets a benchmark of under 20% error.
The results come from a single research site, so whether this calibration holds across the broader Amazon — let alone other tropical biomes — is the obvious next question.