BioCongo: A Multimodal Dataset for Studying 3D Forest Structure in the Congo Basin

The Congo Basin holds the second-largest contiguous tropical forest on Earth and the largest tropical peatland complex, yet it remains one of the world's most sparsely sampled forest regions. Recent works have begun releasing forest structure beyond scalar maps, one profile family at a time. Machine-learning ready datasets still largely omit Central Africa, and existing products ship single scalar maps rather than predictors paired with full vertical profiles. We present BioCongo, a 10 m multimodal dataset of 355 MGRS tiles centred on the basin's moist forest ecotone. It pairs six predictor modalities (Sentinel-2, Sentinel-1, ALOS-2 PALSAR-2, AlphaEarth embeddings, Copernicus GLO-30 and Dynamic World) with thirteen structural targets from GEDI and ICESat-2, including five co-registered vertical profiles and four scalar targets per GEDI shot, alongside six canopy height and biomass products resampled onto the same grid. Predictors are aligned to the 01/04/2019 - 30/03/2023 GEDI mission window that underpins the reference products. A second lidar epoch, 01/04/2024 - 31/07/2025, allows published products and user-derived maps to be tested against acquisitions that postdate them. BioCongo imposes no fixed benchmark protocol: users choose inputs, targets, and splits. Beyond scalar prediction, this supports a representation-learning question: to what extent multimodal EO observations encode lidar-observed vertical forest structure, how representations learned from one structural target transfer to others, and whether those relationships remain stable across time.