Open-Earth-Monitor Global Workshop 2026

Forest structural diversity - the spatial heterogeneity of canopy architecture across vertical and horizontal dimensions - is a fundamental component of ecosystem functioning. Yet its continuous global mapping remains constrained by the sparse orbital sampling of spaceborne LiDAR missions such as the Global Ecosystem Dynamics Investigation (GEDI).
Here, we integrated globally distributed GEDI-derived structural diversity metrics with dense-vector representations from a geospatial vision foundation model pretrained on multi-source satellite imagery. Specifically, we used the Google Satellite Embedding dataset, derived from the AlphaEarth Foundations model, which provides globally consistent 64-dimensional embeddings at 10 m resolution from multi-source satellite imagery. Our analysis spans GEDI's full tropical-to-temperate sampling domain (52°N–52°S), encompassing 14 major biomes from temperate conifer to tropical moist broadleaf forests.
Random forest regression models were fitted within a spatially blocked cross-validation framework stratified by biogeographic region. Cross-validated R² was consistently high across structural diversity dimensions, with low inter-fold variance indicating robust transferability across held-out biogeographic regions. Predicted structural diversity revealed strong but metric-dependent spatial gradients, reflecting the distinct axes of canopy architecture — from height and complexity to vertical profile shape - captured across the global sampling domain.
Our results demonstrate that geospatial foundation-model embeddings capture information across both vertical and horizontal dimensions of forest canopy architecture, thus providing a scalable pathway for wall-to-wall inference of forest structural diversity from existing spaceborne observations