CHIME: Canopy Height Inference from Multispectral Earth Observation

Spaceborne lidar captures forest vertical structure along sparse orbital tracks, whereas Sentinel-2 provides repeated wall-to-wall imagery. We present CHIME, a 16.6-million-parameter sparse-to-dense model that predicts 10 m grids of eight ordered GEDI relative-height percentiles and aleatoric scale estimates from native-resolution Sentinel-2 imagery. A footprint-aware operator connects dense predictions to sparse GEDI supervision. We use CHIME to examine whether geographically broader training data improve transfer. Holding architecture, seed, and optimization budget fixed, expanding training from 225 to 1,025 MGRS tiles reduces RH95 MAE by 13.9% on a fixed development benchmark. On a separately selected, label-blind cohort of 689,562 GEDI footprints in 106 tiles across six continents, the same expansion reduces MAE by 11.8%, lowers height-balanced and biome-balanced error, and reduces MAE in every represented continent. For system-level context, expanded CHIME also has lower MAE than a SenFus-inspired CNN on development and sampled GLAD 2019 on its covered transfer footprints. These results support CHIME as a scalable foundation for forest-structure screening and for targeting field and airborne-lidar surveys.