AlphaMap™
Species-resolved aboveground forest biomass, structure, and height at 10 m annually updated from 2017 onward.
Species-resolved aboveground forest biomass, structure, and height at 10 m annually updated from 2017 onward.
Web App link - CONUS * in development
Web App link - Hurricane Michael case study
Web App link - Alachua County case study
Web App link - Hurricane Michael case study
Web App link - Alachua County case study
Overview: AlphaMap maps how much aboveground biomass is standing, and which tree species it belongs to, at 10 m resolution, for any year from 2017 onward with annual updates into the future. The first application is the Hurricane Michael impact area, covering 36 counties across the Florida Panhandle and adjoining Georgia and Alabama, 47,361 km². Within that area we model 35,636 km² (8,805,853 acres) as forest, carrying a mean 38.87 tons/acre of aboveground biomass in 2017. The design is built to then scale up to CONUS level in the near future. Additional ongoing improvements include GeoTiff download capability and AOI statistics export, among various ideas.
Methods: AlphaMap uses Google AlphaEarth annual satellite embeddings as the predictors: a 64-number summary of each 10 m pixel for each year, learned from Sentinel-1 radar, Sentinel-2 and Landsat optical, elevation, climate reanalysis, and GEDI spaceborne lidar, among others. One embedding exists per year from 2017 to 2025. The model is calibrated against three products - two from the USDA Forest Service FIA program, the BIGMAP and TREEMAP datasets, and one from ESA, the ESA CCI Biomass v6. The study area is masked to Google Dynamic World areas with tree canopy present, which is recomputed annually so that the mask tracks succession and clearing, with an override retaining any pixel the inventory products report as carrying biomass. A single seeded random forest is trained once and applied unchanged to all nine annual embeddings (and into the future). The series is therefore one model viewed through nine years of imagery rather than nine separately fitted models, which is what makes year-to-year comparison feasible.
Validation: Spatial transfer. Under 25 km spatial-block cross-validation - holding out whole blocks of landscape rather than scattered points - skill is r = 0.85 (slash pine), 0.79 (sweetgum), 0.78 (loblolly), 0.75 (longleaf), and 0.71 (total). Temporal transfer. We measured this independently on two products. Applying a model out of its calibration year costs about +4% RMSE (CCI, six epochs) and −4% (TreeMap, two epochs). Thus out-of-year application is close to free. Model drift over nine years. Undamaged forest, run through the frozen 2017 model on nine successive embeddings, reads 42.65 tons/acre in 2017 and 42.82 in 2025, ranging between 42.63 and 43.51 over the nine years, a 2.1% spread with no trend across roughly 173 million pixels. The model does not drift with embedding vintage.
Case Study: Hurricane Michael: - preliminary results
Pre-storm. Pines dominated, with loblolly at 20.1% of modelled biomass and slash at 18.3%, with longleaf a further 6.8%. These three pines thus hold 45.2% of standing biomass.
The storm. In the most severely damaged class of the Storm Cloud RAPID damage classification, 13.7% of standing biomass was removed. Losses fell on the pines: slash −33.1%, longleaf −33.6%, and loblolly −22.0%. Bottomland hardwoods held, with swamp tupelo at −1.0% and laurel oak at −4.0%.
Recovery. By 2025 the worst-hit class has closed only about 32% of the loss and still sits 6.9% below comparable undamaged forest. Recovery is a change in composition rather than a return: sweetgum (+23.0%), water oak, yellow poplar and loblolly pine are gaining, while slash pine is flat and longleaf shows no recovery signal.
References
Brown, C.F. et al. (2025) AlphaEarth Foundations. arXiv:2507.22291.
Brown, C.F. et al. (2022) Dynamic World, near real-time global 10 m land use land cover mapping. Scientific Data 9: 251. doi:10.1038/s41597-022-01307-4
Species biomass from the USDA Forest Service BIGMAP 2018 per-species aboveground biomass product;
Stand structure from USFS TreeMap (RDS-2025-0032);
Independent total biomass from ESA CCI Biomass v6;
Forest mask from Google Dynamic World (10 m).
Acknowledgements: AlphaMap is a project of the Spatial Ecology and Conservation (SPEC) Lab, School of Forest, Fisheries and Geomatics Sciences, at the University of Florida (www.speclab.org), maintained by (PI) Eben Broadbent. Please email [email protected] with any questions, requests, or issues. Funding support gratefully acknowledged from the USDA McIntire-Stennis, UF IFAS, FIA USFS, FAMU CSER, and the SPEC Lab. Special thanks to Angelica Almeyda Zambrano, Gabriel Prata, Todd Schroeder, Jason Drake, and Paul Medley.
Copyright pending © 2026, University of Florida
Methods: AlphaMap uses Google AlphaEarth annual satellite embeddings as the predictors: a 64-number summary of each 10 m pixel for each year, learned from Sentinel-1 radar, Sentinel-2 and Landsat optical, elevation, climate reanalysis, and GEDI spaceborne lidar, among others. One embedding exists per year from 2017 to 2025. The model is calibrated against three products - two from the USDA Forest Service FIA program, the BIGMAP and TREEMAP datasets, and one from ESA, the ESA CCI Biomass v6. The study area is masked to Google Dynamic World areas with tree canopy present, which is recomputed annually so that the mask tracks succession and clearing, with an override retaining any pixel the inventory products report as carrying biomass. A single seeded random forest is trained once and applied unchanged to all nine annual embeddings (and into the future). The series is therefore one model viewed through nine years of imagery rather than nine separately fitted models, which is what makes year-to-year comparison feasible.
Validation: Spatial transfer. Under 25 km spatial-block cross-validation - holding out whole blocks of landscape rather than scattered points - skill is r = 0.85 (slash pine), 0.79 (sweetgum), 0.78 (loblolly), 0.75 (longleaf), and 0.71 (total). Temporal transfer. We measured this independently on two products. Applying a model out of its calibration year costs about +4% RMSE (CCI, six epochs) and −4% (TreeMap, two epochs). Thus out-of-year application is close to free. Model drift over nine years. Undamaged forest, run through the frozen 2017 model on nine successive embeddings, reads 42.65 tons/acre in 2017 and 42.82 in 2025, ranging between 42.63 and 43.51 over the nine years, a 2.1% spread with no trend across roughly 173 million pixels. The model does not drift with embedding vintage.
Case Study: Hurricane Michael: - preliminary results
Pre-storm. Pines dominated, with loblolly at 20.1% of modelled biomass and slash at 18.3%, with longleaf a further 6.8%. These three pines thus hold 45.2% of standing biomass.
The storm. In the most severely damaged class of the Storm Cloud RAPID damage classification, 13.7% of standing biomass was removed. Losses fell on the pines: slash −33.1%, longleaf −33.6%, and loblolly −22.0%. Bottomland hardwoods held, with swamp tupelo at −1.0% and laurel oak at −4.0%.
Recovery. By 2025 the worst-hit class has closed only about 32% of the loss and still sits 6.9% below comparable undamaged forest. Recovery is a change in composition rather than a return: sweetgum (+23.0%), water oak, yellow poplar and loblolly pine are gaining, while slash pine is flat and longleaf shows no recovery signal.
References
Brown, C.F. et al. (2025) AlphaEarth Foundations. arXiv:2507.22291.
Brown, C.F. et al. (2022) Dynamic World, near real-time global 10 m land use land cover mapping. Scientific Data 9: 251. doi:10.1038/s41597-022-01307-4
Species biomass from the USDA Forest Service BIGMAP 2018 per-species aboveground biomass product;
Stand structure from USFS TreeMap (RDS-2025-0032);
Independent total biomass from ESA CCI Biomass v6;
Forest mask from Google Dynamic World (10 m).
Acknowledgements: AlphaMap is a project of the Spatial Ecology and Conservation (SPEC) Lab, School of Forest, Fisheries and Geomatics Sciences, at the University of Florida (www.speclab.org), maintained by (PI) Eben Broadbent. Please email [email protected] with any questions, requests, or issues. Funding support gratefully acknowledged from the USDA McIntire-Stennis, UF IFAS, FIA USFS, FAMU CSER, and the SPEC Lab. Special thanks to Angelica Almeyda Zambrano, Gabriel Prata, Todd Schroeder, Jason Drake, and Paul Medley.
Copyright pending © 2026, University of Florida
Disclaimer: This is a beta system. The results provided herein are intended for preliminary and informational purposes only. The University of Florida (UFL), the Spatial Ecology and Conservation Lab (SPECLab), and all collaborating institutions and individuals make no representations or warranties regarding the accuracy or completeness of these results. We expressly disclaim any liability for any direct, indirect, incidental, consequential, or special damages arising out of or in any way connected with the access to or use of these results. Users are solely responsible for the interpretation and application of the data, and any use or misuse thereof. This system is powered by Google Earth Engine, utilizing low- to zero-carbon energy sources. Portions of the software development and text revisions were completed with assistance from AI tools, under the direction of and reviewed by the authors.
