Forest vertical structural diversity reflects ecosystem complexity and biodiversity potential. Foliage height diversity (FHD), an indicator measured by the Global Ecosystem Dynamics Investigation (GEDI), describes how foliage is distributed across canopy height layers. However, GEDI LiDAR samples forests within discontinuous footprints rather than continuous coverage. To produce a wall-to-wall, 10 m-resolution FHD map of Hokkaido, Japan, we propose a LightGBM regression framework trained on approximately 150,000 GEDI footprints collected from May to October 2023. Predictors include conventional remote sensing and environmental layers (Sentinel-1 C-band SAR, Sentinel-2 optical imagery, AW3D30 elevation, and CGLS forest type) alongside a foundation-model-derived feature set: Satellite Embedding V1, a 64-dimensional representation from AlphaEarth Foundations. We compared three models: (1) conventional predictors only, (2) embeddings only, and (3) combined predictors and embeddings. The embedding-only approach clearly outperformed the conventional baseline, while the combined model achieved the best overall performance (RMSE = 0.303; <em>R</em><sup>2</sup> = 0.507). These results suggest that the embeddings capture information relevant to forest vertical structure, providing complementary value beyond standard predictors. The final 10 m FHD map revealed spatially coherent patterns across Hokkaido, suggesting its potential utility for regional assessments of ecosystem complexity and biodiversity. Our findings indicate that foundation-model-derived satellite embeddings can substantially improve GEDI-based forest structure estimation when integrated with open geospatial datasets in a reproducible mapping workflow.