Contents

Details

Analytics

Home

All Products

Docs

Source Cooperative is a Radiant Earth project

A vertical vegetation structure model of the Earth · GeoAI@UCPH · Source Cooperative | Source Cooperative

A vertical vegetation structure model of the Earth

Name: A vertical vegetation structure model of the Earth
Creator: GeoAI@UCPH

To better support carbon and biodiversity monitoring and to support the SDGs, EUDR, and 30x30, we have developed a global vertical vegetation structure model (VSM) to describe the full 3D-structure of vegetation at high spatial resolution.

Contents

root

Details

Visibility: Public
Owner: GeoAI@UCPH
Created: 5 May 2026
Last Updated: 12 Oct 2026

Analytics

28 days

Downloads

97,330

Data served

21.5 GB

Countries

28-day downloads

18 Sep 2026

15 Oct 2026

Analytics

README

VSM: A Vertical Vegetation Structure Model of the Earth

The VSM brings a vertical dimension to global vegetation mapping, extending beyond canopy height to characterzing vegetation structure at 10 m resolution. Trained with 178 million sparse GEDI lidar observations, a deep learning model translates Sentinel-2 imagery into continuous maps of 101 relative height metrics (RH0–RH100), each with predictive quantiles that describe its uncertainty. By capturing vegetation structure beyond canopy height, VSM supports applications in forest monitoring, biodiversity assessment, and ecosystem research.

Explore

📄 Preprint: eartharxiv.org/repository/view/15380
🗺️ Interactive map: huizhml.github.io/map-explorer
💻 Code: github.com/huizhml/global-vegetation-structure
📦 Storage: s3://geoai-ucph/gvsm/ on the endpoint https://data.source.coop

Key facts


Variable	Relative height profile (RH0 - RH100) , in decimeters
Uncertainty	Median (Q1) available; 5th percentile (Q2) and 95th percentile (Q0) coming soon
Resolution	10 m (main product), 1 km (global mosaics)
Coverage	Global land
Years	2020 (available), 2024 (in preparation)
Input	Sentinel-2 L2A
Reference data	GEDI L2A, 178 million height profiles
Format	Cloud Optimized GeoTIFF, Int16, lossless LERC+ZSTD (tiles) / ZSTD (mosaics)
Volume	≈ 360 TB per year (full RH profile × 3 quantiles)
License	CC BY 4.0

What is an RH metric?

GEDI records laser energy reflected by vegetation and the ground within each footprint. RHn is the height above ground below which n % of that returned energy has accumulated.

RH98–RH100: approximate canopy top height, comparable to conventional canopy height products.
RH50: the height below which half of the total returned energy has accumulated.
Low RHs (RH0–RH25): describe the lower portion of the cumulative return, influenced by both vegetation and ground returns.

Taken together, RH0 … RH100 for a pixel make a cumulative vertical profile. You can use it to study canopy layering, understory density and structural diversity, not only the top height.

Global map of VSM vertical vegetation structure, with RH layers stacked as a 3-D datacube (height in metres, 0–50)

Data organization

File names follow RH{level}_Q{quantile}.tif, with level from 0 to 100 and quantile from 0 to 2.

Path	Content	Grid
`<year>/<tile>/`	Full-resolution predictions. One folder per Sentinel-2 MGRS tile, one file per RH level and quantile.	Native UTM zone of the tile, 10 m, 10 980 × 10 980 px
`mosaics/<year>/`	Global overview mosaics (10 m tiles averaged to 0.01°), for browsing and continental- to global-scale analysis	EPSG:4326, 0.01°, 36 000 × 13 889 px

To find the tile that covers your area, use the Sentinel-2 MGRS tiling grid. Neighbouring tiles overlap by about 10 km.

Quantiles: Q0, Q1, Q2

The model is trained with quantile regression, so each RH level is predicted at three quantiles:

Suffix	Quantile	Use
`Q1`	50th percentile (median)	The main prediction. Use this by default.
`Q2`	5th percentile	Lower bound of the prediction interval
`Q0`	95th percentile	Upper bound of the prediction interval

Q0 and Q2 together give a 90 % prediction interval for each pixel and RH level. The interval is calibrated with conformal prediction, so it reaches the target coverage against held-out GEDI data. The interval width (Q0 − Q2) is a simple per-pixel uncertainty measure.

Note: Only Q1 (the median) is published at the moment. Q0 and Q2 will be uploaded to the same folders soon.

Pixel values

Data type Int16. Values are in decimetres: divide by 10 to get metres.
NoData = 32767. This marks pixels without a valid prediction, such as water, snow and ice, persistent cloud.
Low RH levels can be negative. This follows the GEDI convention: the lowest part of the return waveform lies below the detected ground elevation.

Quick start

GDAL: inspect a file over HTTP without downloading it:

Python: read the vertical profile at a point:

AWS CLI: list or download data:

The COGs also open directly in QGIS (Layer → Add Raster Layer → Protocol: HTTP(S)) and in rioxarray/xarray.

Caveats

These are model predictions, not lidar measurements. Accuracy varies with biome, terrain slope and canopy density. In dense tall forests, predicted heights tend to be biased toward the mean.
GEDI only samples between about 51.6° N and 51.6° S. Predictions further north and south rely on the model extrapolating beyond its training data, so treat them with extra caution.
RH metrics are not the same as airborne-lidar canopy height. GEDI's RH metrics come from a ~25 m footprint waveform. They are related to, but not identical to, canopy height models derived from airborne laser scanning (ALS).
Each year's map is built from Sentinel-2 imagery of that year. Differences between years mix real change with differences in the input imagery, so change analysis needs care.
The global mosaics are 10 m predictions averaged to ~1 km. They smooth out fine-scale structure, so use the 10 m tiles for quantitative work.

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Citation

Zhang, Hui, Nico Lang, Mikolaj Mazurczyk, Stefan Oehmcke, Ke Huang, Martin Brandt, Rasmus Fensholt, Ankit Kariryaa, and Christian Igel. "A vertical vegetation structure model of the Earth." EarthArXiv preprint (2026). https://doi.org/10.31223/X5M22F

Contact

Managed by GeoAI@UCPH, University of Copenhagen.

Hui Zhang (huzh@di.ku.dk) · Nico Lang (nila@di.ku.dk) · Christian Igel (igel@di.ku.dk)

README

VSM: A Vertical Vegetation Structure Model of the Earth

Explore

📄 Preprint: eartharxiv.org/repository/view/15380
🗺️ Interactive map: huizhml.github.io/map-explorer
💻 Code: github.com/huizhml/global-vegetation-structure
📦 Storage: s3://geoai-ucph/gvsm/ on the endpoint https://data.source.coop

Key facts


Variable	Relative height profile (RH0 - RH100) , in decimeters
Uncertainty	Median (Q1) available; 5th percentile (Q2) and 95th percentile (Q0) coming soon
Resolution	10 m (main product), 1 km (global mosaics)
Coverage	Global land
Years	2020 (available), 2024 (in preparation)
Input	Sentinel-2 L2A
Reference data	GEDI L2A, 178 million height profiles
Format	Cloud Optimized GeoTIFF, Int16, lossless LERC+ZSTD (tiles) / ZSTD (mosaics)
Volume	≈ 360 TB per year (full RH profile × 3 quantiles)
License	CC BY 4.0

What is an RH metric?

GEDI records laser energy reflected by vegetation and the ground within each footprint. RHn is the height above ground below which n % of that returned energy has accumulated.

RH98–RH100: approximate canopy top height, comparable to conventional canopy height products.
RH50: the height below which half of the total returned energy has accumulated.
Low RHs (RH0–RH25): describe the lower portion of the cumulative return, influenced by both vegetation and ground returns.

Taken together, RH0 … RH100 for a pixel make a cumulative vertical profile. You can use it to study canopy layering, understory density and structural diversity, not only the top height.

Data organization

File names follow RH{level}_Q{quantile}.tif, with level from 0 to 100 and quantile from 0 to 2.

Path	Content	Grid
`<year>/<tile>/`	Full-resolution predictions. One folder per Sentinel-2 MGRS tile, one file per RH level and quantile.	Native UTM zone of the tile, 10 m, 10 980 × 10 980 px
`mosaics/<year>/`	Global overview mosaics (10 m tiles averaged to 0.01°), for browsing and continental- to global-scale analysis	EPSG:4326, 0.01°, 36 000 × 13 889 px

To find the tile that covers your area, use the Sentinel-2 MGRS tiling grid. Neighbouring tiles overlap by about 10 km.

Quantiles: Q0, Q1, Q2

The model is trained with quantile regression, so each RH level is predicted at three quantiles:

Suffix	Quantile	Use
`Q1`	50th percentile (median)	The main prediction. Use this by default.
`Q2`	5th percentile	Lower bound of the prediction interval
`Q0`	95th percentile	Upper bound of the prediction interval

Note: Only Q1 (the median) is published at the moment. Q0 and Q2 will be uploaded to the same folders soon.

Pixel values

Data type Int16. Values are in decimetres: divide by 10 to get metres.
NoData = 32767. This marks pixels without a valid prediction, such as water, snow and ice, persistent cloud.
Low RH levels can be negative. This follows the GEDI convention: the lowest part of the return waveform lies below the detected ground elevation.

Quick start

GDAL: inspect a file over HTTP without downloading it:

Python: read the vertical profile at a point:

AWS CLI: list or download data:

The COGs also open directly in QGIS (Layer → Add Raster Layer → Protocol: HTTP(S)) and in rioxarray/xarray.

Caveats

These are model predictions, not lidar measurements. Accuracy varies with biome, terrain slope and canopy density. In dense tall forests, predicted heights tend to be biased toward the mean.
GEDI only samples between about 51.6° N and 51.6° S. Predictions further north and south rely on the model extrapolating beyond its training data, so treat them with extra caution.
RH metrics are not the same as airborne-lidar canopy height. GEDI's RH metrics come from a ~25 m footprint waveform. They are related to, but not identical to, canopy height models derived from airborne laser scanning (ALS).
Each year's map is built from Sentinel-2 imagery of that year. Differences between years mix real change with differences in the input imagery, so change analysis needs care.
The global mosaics are 10 m predictions averaged to ~1 km. They smooth out fine-scale structure, so use the 10 m tiles for quantitative work.

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Citation

Contact

Managed by GeoAI@UCPH, University of Copenhagen.

Hui Zhang (huzh@di.ku.dk) · Nico Lang (nila@di.ku.dk) · Christian Igel (igel@di.ku.dk)

1gvsm/
2├── <year>/                      # acquisition year, e.g. 2020
3│   └── <MGRS tile>/             # Sentinel-2 MGRS tile ID, e.g. 32UNG
4│       ├── RH0_Q1.tif
5│       ├── RH1_Q1.tif
6│       ├── …
7│       └── RH100_Q1.tif
8└── mosaics/
9    └── <year>/
10        ├── RH0_Q1.tif
11        ├── …
12        └── RH100_Q1.tif

1gvsm/
2├── <year>/                      # acquisition year, e.g. 2020
3│   └── <MGRS tile>/             # Sentinel-2 MGRS tile ID, e.g. 32UNG
4│       ├── RH0_Q1.tif
5│       ├── RH1_Q1.tif
6│       ├── …
7│       └── RH100_Q1.tif
8└── mosaics/
9    └── <year>/
10        ├── RH0_Q1.tif
11        ├── …
12        └── RH100_Q1.tif

1gdalinfo /vsicurl/https://data.source.coop/geoai-ucph/gvsm/2020/32UNG/RH98_Q1.tif

1gdalinfo /vsicurl/https://data.source.coop/geoai-ucph/gvsm/2020/32UNG/RH98_Q1.tif

1import numpy as np
2import rasterio
3from pyproj import Transformer
4
5base = "https://data.source.coop/geoai-ucph/gvsm/2020/32UNG"
6lon, lat = 9.7915, 55.4934
7
8profile = []
9for rh in range(101):
10    with rasterio.open(f"{base}/RH{rh}_Q1.tif") as src:
11        x, y = Transformer.from_crs("EPSG:4326", src.crs, always_xy=True).transform(lon, lat)
12        v = next(src.sample([(x, y)]))[0]
13        profile.append(np.nan if v == src.nodata else v / 10)  # decimetres -> metres

1import numpy as np
2import rasterio
3from pyproj import Transformer
4
5base = "https://data.source.coop/geoai-ucph/gvsm/2020/32UNG"
6lon, lat = 9.7915, 55.4934
7
8profile = []
9for rh in range(101):
10    with rasterio.open(f"{base}/RH{rh}_Q1.tif") as src:
11        x, y = Transformer.from_crs("EPSG:4326", src.crs, always_xy=True).transform(lon, lat)
12        v = next(src.sample([(x, y)]))[0]
13        profile.append(np.nan if v == src.nodata else v / 10)  # decimetres -> metres

1aws s3 ls --no-sign-request --endpoint-url https://data.source.coop s3://geoai-ucph/gvsm/2020/32UNG/
2aws s3 sync --no-sign-request --endpoint-url https://data.source.coop s3://geoai-ucph/gvsm/mosaics/2020/ ./gvsm_mosaics_2020/

1aws s3 ls --no-sign-request --endpoint-url https://data.source.coop s3://geoai-ucph/gvsm/2020/32UNG/
2aws s3 sync --no-sign-request --endpoint-url https://data.source.coop s3://geoai-ucph/gvsm/mosaics/2020/ ./gvsm_mosaics_2020/

1@misc{zhang2026vertical,
2  title     = {A vertical vegetation structure model of the {Earth}},
3  author    = {Zhang, Hui and Lang, Nico and Mazurczyk, Mikolaj and Oehmcke, Stefan and Huang, Ke and Brandt, Martin and Fensholt, Rasmus and Kariryaa, Ankit and Igel, Christian},
4  year      = {2026},
5  publisher = {EarthArXiv},
6  note      = {Preprint},
7  doi       = {10.31223/X5M22F},
8  url       = {https://doi.org/10.31223/X5M22F}
9}

1@misc{zhang2026vertical,
2  title     = {A vertical vegetation structure model of the {Earth}},
3  author    = {Zhang, Hui and Lang, Nico and Mazurczyk, Mikolaj and Oehmcke, Stefan and Huang, Ke and Brandt, Martin and Fensholt, Rasmus and Kariryaa, Ankit and Igel, Christian},
4  year      = {2026},
5  publisher = {EarthArXiv},
6  note      = {Preprint},
7  doi       = {10.31223/X5M22F},
8  url       = {https://doi.org/10.31223/X5M22F}
9}