GESI™

​Global Ecosystem Structure Index (GESI)

Modeled 3D structure of forests on a 10 m grid, 2017 to 2025.

GESI estimates how forests are built from the ground up: how tall the canopy is, how leaves and branches are spread across heights, and how much aboveground biomass the forest holds. It learns from NASA's GEDI space lidar, which measures forest height profiles in scattered circles about 25 m across. It fills the gaps between them with Google AlphaEarth satellite embeddings, a yearly summary of satellite imagery for every 10 m pixel. It then calculates and provides downloads for a variety of rasters in GeoTiff format, suitable for direct use in R or any remote sensing or GIS software, and statistics in database formats.

Launch the GESI web app

Best in a desktop browser. On a phone, the profile viewer starts collapsed.
App screenclip (above). GESI of an AOI in the Ordway Swisher Biological Station in Melrose, Florida, USA: canopy layers in color (red = upper canopy, green = middle, blue = understory). Mixed pine savannas, dense Oak hammocks and wetlands are abundant.
App screenclip (above). GESI of Kanapaha Forest in Gainesville, Florida, USA: canopy layers in color (red = upper canopy, green = middle, blue = understory). Pine plantations, urban forested neighborhoods and dense Oak hammocks and wetlands are abundant.
App screenclip (above). GESI of the Osa Conservation property and surrounding area in the Osa Peninsula, Costa Rica: canopy layers in color (red = upper canopy, green = middle, blue = understory). Old-growth rainforest, cattle pastures, and secondary forests dominant the landscape.
App screenclip (above). GESI of the Los Amigos Biological Station and surrounding areas in the Peruvian Amazon: canopy layers in color (red = upper canopy, green = middle, blue = understory). Old-growth rainforest, gold mining operations (exposed soil), and secondary forests dominant the landscape.
App screenclip (above). GESI of the Cocha Cashu Biological Station and surrounding areas in the Peruvian Amazon: canopy layers in color (red = upper canopy, green = middle, blue = understory). Old-growth rainforest dominant the landscape.
What GESI maps
​For any single selected area up to about 20 x 20 km (new versions lifting the processing window, and a global asset release is planned in 2026-27) between 52° S and 52° N, and for each year from 2017 to 2025:
  • Canopy height (GEDI RH98, near the top of the canopy) and aboveground biomass (Mg/ha, metric tons per hectare), each with a 90% range: the span that should hold the GEDI value 9 times in 10.
  • Plant area (leaves and branches) in twelve 5 m height layers (the top one holds everything above 55 m), and its total, the plant area index.
  • Biomass by height: each pixel's biomass split across the same layers in proportion to its plant area.
  • Foliage height diversity: how evenly plant area is spread over the layers (a Shannon index; MacArthur & MacArthur 1961).
  • The modeled lidar height profile of every 10 m pixel (GEDI relative heights RH0 to RH100), also drawn in 1 m steps for display.
​
How it works
When you press Build, the App gathers up to 4,000 quality-filtered GEDI measurements (2019-2025) from your area and 10 km around it, widened to 20 or 40 km where measurements are few. It pairs each one with the AlphaEarth embedding of the same year. The embeddings are produced by Google and Google DeepMind with their AlphaEarth Foundations AI model, and they are the only inputs to GESI's models. Random forests, a machine-learning method, then learn canopy height, the shape of the height profile, total plant area and biomass, and map every 10 m pixel of every year.

GEDI measurements in some 2 km blocks are kept out of training. Part of them sets the 90% ranges and the rest tests the map. After each build, the App reports how well the map matches them in your area, or says when too few were available to test. At NEON field sites in the United States (the National Ecological Observatory Network), an optional button also compares the canopy height map with NEON airborne lidar, which GESI does not use for training.

Using the App
  1. Open the App. The demo area (Ordway-Swisher Biological Station, Florida) is already selected. On a wide screen, the profile viewer shows a gray EXAMPLE profile.
  2. Choose an area: keep the demo, or press "Pick a 20 km square" or "Draw an area".
  3. Build: press "Build the 3D structure". It usually takes 3 to 5 minutes. Keep the browser tab open and in view, because a hidden or minimized tab runs much slower. Then read the accuracy summary for your area.
  4. Map: pick a year and a map layer.
  5. Vertical profile: click the map inside the white outline for the profile of one 10 m pixel, or press "Draw a polygon" for the average profile of a patch.
  6. Download: choose a product and press "Get GeoTIFF links for the selected year".

Accuracy and limits
GESI is a modeled product, not a measurement, and its accuracy changes from place to place. The accuracy summary for the area you build is the one to use.

In the v52 release tests, the maps were compared with GEDI measurements set aside for testing. Biomass was compared with GEDI's own footprint biomass estimates, not with field plots. R² is the share of the variation explained: 1 is perfect, and 0 is no better than the area's average.
  • Florida pine forest (Ordway-Swisher): canopy height R² 0.51, biomass R² 0.35. The modeled height profile was off by 2.4 m on average.
  • Tall tropical forest near Manaus, Brazil: canopy height R² 0.26, biomass R² 0.22. The height profile was off by 3.8 m on average. Total plant area (R² 0.10) and plant area by height (R² 0.09) were flagged as weak.
  • 90% ranges: they held the GEDI value 89-92% of the time.

These scores are likely optimistic, because AlphaEarth was trained in part to reconstruct GEDI heights.

Against NEON airborne lidar at Ordway-Swisher (2023 flight, 253 spots), canopy height had R² 0.54 and a typical error of 5.8 m. On average the map was 0.1 m lower than the lidar, but it read about 6 m too high where the canopy was under 13 m and 2.6-4.0 m too low where it was taller. Its 90% range held the lidar height at 81% of spots. The ranges are set with GEDI, so differences between GEDI and airborne lidar count as error here.

Limits:
  • Resolution: each estimate describes a circle about 25 m across (the size of a GEDI measurement), drawn on a 10 m grid. Single trees are not resolved.
  • The 1 m profile is a display and download grid, not 1 m resolution. Detail finer than about 5 m is not resolved, so compare polygons, not single meters.
  • Coverage and years: about 52° S to 52° N (GEDI's range). GEDI has a gap in its measurements in 2023-24. The years 2017 and 2018 come before GEDI and are extrapolated.
  • Change over time: accuracy is tested across space, not across years, so differences between years have not been validated as real change.
  • Tall tropical forest: skill is weakest here due to dense tall leaf area limiting GEDI and Alpha Embeddings signal within and below the canopy.
  • Biomass by height is an estimated split, not a measurement. Trunk wood low in the canopy is under-represented.
  • 90% ranges describe the value GEDI would report for a ~25 m footprint. They leave out the error of GEDI's own biomass models.
  • Intended use: GESI is a beta research product for preliminary and informational use.

Downloads and license
The App provides zipped GeoTIFF files with 10 m pixels, in the area's UTM map projection, for the selected year: canopy height and biomass with their 90% ranges; plant area index, canopy cover and foliage height diversity; the RH0-RH100 height profile; and plant area, biomass by height and lidar return energy, each in 5 m layers or 1 m steps. A metadata file (CSV), listed with the download links, explains each band, the method and the accuracy measured for that area.

GESI outputs are released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0, creativecommons.org/licenses/by-nc/4.0/): free to share and adapt, with attribution, for non-commercial use. Commercial use requires permission from the authors ([email protected]). © 2026 University of Florida. Input data keep their own terms (see Credits).

How to cite
Broadbent EN, Almeyda Zambrano AM (2026). GESI - Global Ecosystem Structure Index, version # (month-day-year). Spatial Ecology and Conservation (SPEC) Lab, University of Florida. https://www.speclab.org/gesi.html. Accessed [date].
Cite the version you used; its number and release date are at the top of the App's sidebar.

Detailed methods

This section describes GESI version 52 in more detail. The App's "About, methods and citations" panel carries the same information, together with the full list of data credits.

Inputs.
  • Labels: NASA GEDI footprint products from 2019 to 2025 (Version 2):
    • relative heights RH0 to RH100 in 22 percentiles (L2A);
    • the plant-area profile in 5 m layers and the total plant area index (L2B);
    • aboveground biomass density (L4A, version 2.1).
    GEDI Version 3 footprint positions are used where they exist.
  • Predictors: the 64 bands of the AlphaEarth Foundations annual Satellite Embedding (Google and Google DeepMind) for 2017 to 2025. These are the only inputs to the models.
  • Screening only: terrain slope from the Copernicus DEM, and leaf-on timing from the MODIS land cover dynamics product.

Selecting GEDI measurements.
  • Where from: footprints from the chosen area and 10 km around it, widened to 20 or 40 km when fewer than 1,200 pass the quality checks.
  • Joining: the three GEDI products are joined shot by shot.
  • Quality checks:
    • GEDI's quality and degrade flags;
    • beam sensitivity of at least 0.95, and at least 0.98 for evergreen broadleaf forest in GEDI's Africa, South Asia and South America regions (the GEDI Version 3 rule);
    • leaf-on timing for deciduous forest;
    • terrain slope of 30° or less;
    • GEDI's rule on the highest return relative to the terrain model;
    • physical limits of 90 m canopy height and 1,500 Mg/ha biomass.
  • Order and cap: footprints are ordered and shuffled with a fixed seed, so a rebuild gives the same sample. At most 4,000 are used; above that, they are thinned to one per 200 m cell per year.

Training and testing.
Each 2 km block goes wholly into training (70%), calibration of the 90% ranges (15%) or testing (15%). The App reports skill only when at least 60 test measurements are available.
Label adjustments. Before training, three steps adjust the biomass, canopy height and total plant area labels:
  • known offsets between GEDI beam types and times of day are removed;
  • small per-orbit offsets are removed;
  • orbits with large, consistent errors are left out.

All three are estimated from out-of-sample errors, and their benefit is still being evaluated.

Models.

Random forests (Breiman 2001) predict:
  • canopy height (RH98);
  • the shape of the height profile above the ground end, as three principal components;
  • the lowest part of the profile (RH0 to RH20) separately;
  • the total plant area index;
  • aboveground biomass.

The predicted heights are sorted into increasing order. The shape of GEDI's plant-area profile in twelve 5 m layers is modeled directly, again as three principal components. It is cut above the predicted canopy height and scaled by the predicted total plant area. Where an area has fewer than 50 GEDI plant-area profiles, the shape comes instead from the predicted height profile, through the gap correction GEDI itself uses (MacArthur & Horn 1969).

Derived layers.
  • Biomass by height: each pixel's biomass split in proportion to its plant area.
  • Canopy cover: 1 − exp(−0.5 × plant area index).
  • Foliage height diversity: the Shannon index of the 5 m plant-area shares (MacArthur & MacArthur 1961).
  • 1 m steps: the lidar return energy per meter comes from the modeled height profile, with the energy below 2 m (ground and low vegetation) given as a separate number. Plant area and biomass are spread from their 5 m layers onto 1 m steps with a smooth curve that keeps every layer's total. The 1 m steps are for display and download; the model does not resolve detail finer than about 5 m.

​Uncertainty.
The 90% ranges for canopy height and biomass come from split-conformal prediction on the calibration blocks (Lei et al. 2018; Vovk et al. 2005). They are scaled by how much the forest's trees disagree, with separate multipliers for low, middle and high values where enough data exist. They describe the value GEDI would report for a footprint about 25 m across.

Accuracy checks.
  • In every build: the held-out test blocks give the accuracy summary shown after each build. It is likely optimistic, because AlphaEarth was trained in part on GEDI heights.
  • At NEON sites (optional): the map can be compared with NEON airborne lidar. The App uses the most recent released flight and compares the 98th percentile of NEON's 1 m canopy height model, in 25 m circles, with GESI's map for the year of that flight.

Key references
  • Dubayah, R., Hofton, M., Blair, J., Armston, J., Tang, H. & Luthcke, S. (2021). GEDI L2A Elevation and Height Metrics Data Global Footprint Level V002. NASA EOSDIS Land Processes DAAC. doi:10.5067/GEDI/GEDI02_A.002
  • Dubayah, R., Tang, H., Armston, J., Luthcke, S., Hofton, M. & Blair, J. (2021). GEDI L2B Canopy Cover and Vertical Profile Metrics Data Global Footprint Level V002. NASA EOSDIS Land Processes DAAC. doi:10.5067/GEDI/GEDI02_B.002
  • Dubayah, R.O., Armston, J., Kellner, J.R., et al. (2022). GEDI L4A Footprint Level Aboveground Biomass Density, Version 2.1. ORNL DAAC. doi:10.3334/ORNLDAAC/2056
  • Brown, C.F., Kazmierski, M.R., Pasquarella, V.J., et al. (2025). AlphaEarth Foundations: an embedding field model for accurate and efficient global mapping from sparse label data. arXiv:2507.22291. doi:10.48550/arXiv.2507.22291
  • NEON (National Ecological Observatory Network) (2026). Ecosystem structure (DP3.30015.001), RELEASE-2026. doi:10.48443/8qst-0w84
  • MacArthur, R.H. & MacArthur, J.W. (1961). On bird species diversity. Ecology 42, 594-598. doi:10.2307/1932254
  • MacArthur, R.H. & Horn, H.S. (1969). Foliage profile by vertical measurements. Ecology 50, 802-804. doi:10.2307/1933693
  • Breiman, L. (2001). Random forests. Machine Learning 45, 5-32. doi:10.1023/A:1010933404324
  • Lei, J., G'Sell, M., Rinaldo, A., Tibshirani, R.J. & Wasserman, L. (2018). Distribution-free predictive inference for regression. Journal of the American Statistical Association 113, 1094-1111. doi:10.1080/01621459.2017.1307116
  • Vovk, V., Gammerman, A. & Shafer, G. (2005). Algorithmic Learning in a Random World. Springer, New York. doi:10.1007/b106715
The full list of data credits and methods references is in the App's "About, methods and citations" panel.

Versions
  • v52, released 09/27/2026 (current). The public release: an example profile on first load, plain-language guidance, simpler and more standard methods, more repeatable builds, an optional check against NEON airborne lidar, and a CC BY-NC 4.0 license.
  • v51, released 09/25/2026. Added the vertical profile in 1 m steps (lidar return energy, plant area and biomass per meter, with the ground as a separate number) on the map, in the profile viewer and in downloads.
  • v48 to v50 (September 23-25, 2026): early builds of the current App.
  • v7, released 12/15/2021 (prototype). One 30 m map, representing 2019-20, of about 50 x 50 km around Los Amigos Biological Station in the Peruvian Amazon. Random forests predicted about ten GEDI relative heights per pixel from ALOS PALSAR radar, SRTM elevation and a GEDI-Landsat canopy height map. Its profiles showed lidar return energy in 0.5 m steps, not plant area. Superseded and not comparable with current versions.

Credits
GESI is a project of the Spatial Ecology and Conservation (SPEC) Lab, School of Forest, Fisheries and Geomatics Sciences, University of Florida (www.speclab.org), maintained by (PI) Eben Broadbent. It uses NASA GEDI and MODIS data, the AlphaEarth Foundations Satellite Embedding dataset (Google and Google DeepMind, CC BY 4.0), the Copernicus DEM and NEON airborne lidar, and runs on Google Earth Engine. Each data source keeps its own terms; full credits are in the App and in each metadata file. Portions of the software development and text revisions were completed with assistance from AI tools. All final work was reviewed by the project team, which takes responsibility for it. Questions, bug reports and commercial-use requests: Eben Broadbent, [email protected]. For a bug report, the App's "Technical log (for bug reports)" button, under About, collects the details needed.

Disclaimer
This is a beta system. The results provided herein are intended for preliminary and informational purposes only. These are modeled layers - NASA GEDI lidar footprint measurements extended to a 10 m grid with Google AlphaEarth satellite embeddings - and not field measurements. GESI is an independent research product; it is not produced, reviewed or endorsed by NASA, Google, Google DeepMind, ESA or any other data provider. The University of Florida (UF), the Spatial Ecology and Conservation Lab (SPEC Lab), 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. The organizations in charge of the Copernicus program by law or by delegation do not incur any liability for any use of the Copernicus WorldDEM-30. This system runs on Google Earth Engine.