Biomazon dataset unlocks ML-ready 3D forest structure modeling for Amazon

Most machine learning models for tropical forests predict single scalar targets like canopy height or biomass, ignoring the full vertical structure. The Biomazon dataset, introduced by Sayan Mandal and colleagues, fills this gap with a 20m-resolution multimodal benchmark covering the entire Amazon Basin. It combines GEDI LiDAR-derived RH percentiles (RH10–RH98) and aboveground biomass density (AGBD) as targets, paired with predictors from Sentinel-1/2, ALOS-2 PALSAR-2, Copernicus DEM, Dynamic World land use/land cover, and AlphaEarth embeddings. The dataset includes standardized spatial splits and evaluation protocols, plus a baseline encoder-decoder framework with task-specific heads for joint or separate prediction.

In extensive ablation studies, the authors explored backbone scale, modality contributions, and fusion strategies, quantifying tradeoffs between single-target and joint-target performance. They contextualized results by comparing against existing gridded products like GEDI L4D at matching temporal scales. The 32-page paper (21 figures) establishes Biomazon as a reference benchmark for physically consistent RH-profile prediction and structure-biomass modeling. This benchmark is a critical step for ML-driven carbon accounting and ecosystem monitoring in tropical forests, enabling more accurate estimation of carbon stocks and forest dynamics.