🌲 AGBD-Lite: A Representative Subset of the AGBD Biomass Dataset 🌳

AGBD-Lite is a compact, distribution-matched subset (≈595k samples, ~5%) of the full AGBD dataset (≈16M samples), a machine-learning-ready benchmark for global Above-Ground Biomass (AGB) estimation from satellite imagery.

It is designed as a screening tool: a small, cheap-to-download stand-in for AGBD that lets you benchmark and iterate on models before committing to a full-scale run on the complete dataset. It was introduced to make the evaluation of Geospatial Foundation Models (GFMs) tractable.

🧪 How it was built

AGBD-Lite subsamples ≈5% of GEDI footprints per Sentinel-2 tile, choosing the footprints that minimise the Wasserstein distance between the biome distributions of the subsampled and the full dataset, for each region. As a result:

• Model rankings are preserved exactly relative to the full dataset.
• For our best configuration, training on AGBD-Lite comes within 3.5% of full-data RMSE (52.72 vs. 50.92 Mg/ha) at ≈5% of the cost.
• Results on AGBD-Lite are therefore a slightly pessimistic estimate of full-AGBD performance, while relative comparisons transfer reliably.

The code that generates AGBD-Lite (footprint selection, feature subsetting, and .h5 writing) lives in the AGBD-GFMs repository under data/agbd_lite/.

📦 Contents

File	Split	Samples	Size
AGBD-Lite-train.h5	train	371,849	5.4 GB
AGBD-Lite-val.h5	val	103,550	1.5 GB
AGBD-Lite-test.h5	test	119,991	1.7 GB
AGBD-Lite-statistics.pkl	—	—	(per-feature mean/std on train, for normalisation)
	total	595,390

Each sample is a 25×25 pixel patch (250 m × 250 m at 10 m resolution) centred on a GEDI footprint, with the footprint's AGB label at the centre.

🗂️ Schema

Every .h5 file has the same layout. Datasets share a common first axis N (number of samples).

Dataset	Shape	dtype	Description
S2_bands	(N, 25, 25, 12)	uint16	Sentinel-2 L2A reflectance, 12 bands. Band order in .attrs['order']: B01,B02,B03,B04,B05,B06,B07,B08,B8A,B09,B11,B12 (SCL excluded).
ALOS_bands	(N, 25, 25, 2)	uint16	ALOS PALSAR-2 yearly mosaic, .attrs['order'] = ['HH','HV'].
DEM	(N, 25, 25)	int16	Copernicus DEM elevation (m).
LC	(N, 25, 25, 2)	uint8	Land cover (2 channels).
GEDI/agbd	(N,)	float32	Target: above-ground biomass density (Mg/ha).
GEDI/rh98	(N,)	float32	GEDI relative height at 98% (canopy height proxy, m).
GEDI/lat_decimal	(N,)	float32	Signed fractional part of the footprint latitude (see decoding below).
GEDI/lat_offset	(N,)	uint8	Unsigned integer part of the footprint latitude.
GEDI/lon_decimal	(N,)	float32	Signed fractional part of the footprint longitude.
GEDI/lon_offset	(N,)	uint8	Unsigned integer part of the footprint longitude.
GEDI/pft_class	(N,)	uint8	GEDI plant functional type class.
GEDI/region_cla	(N,)	uint8	GEDI region class.
Sentinel_metadata/S2_date	(N,)	int16	Sentinel-2 acquisition date (days offset).
Sentinel_metadata/S2_boa_offset	(N,)	uint8	Sentinel-2 BOA reflectance offset (processing-baseline correction).
region	(N,)	uint8	Region index (regions numbered alphabetically).

Decoding coordinates. Latitude and longitude are split into a signed decimal part and an unsigned integer offset. Reconstruct with: lat = sign(lat_decimal) * (abs(lat_decimal) + lat_offset) and likewise for longitude.

For the full definition of every feature (units, encodings, ancillary variables), see the AGBD dataset card.

🚀 Quickstart

AGBD-Lite ships as raw .h5 files (no datasets loader / no viewer). Download and open with h5py:

from huggingface_hub import hf_hub_download
import h5py, numpy as np

path = hf_hub_download(
    repo_id="prs-eth/AGBD-Lite",
    filename="AGBD-Lite-train.h5",
    repo_type="dataset",
)

with h5py.File(path, "r") as f:
    agbd = f["GEDI"]["agbd"][:]          # (N,)  target, Mg/ha
    s2   = f["S2_bands"]                 # (N, 25, 25, 12), order in .attrs["order"]
    print(f["S2_bands"].attrs["order"])
    print("N =", agbd.shape[0], "| mean AGB =", np.nanmean(agbd))

To grab the whole dataset at once:

from huggingface_hub import snapshot_download
local_dir = snapshot_download(repo_id="prs-eth/AGBD-Lite", repo_type="dataset")

🔗 Related resources

• Full dataset: prs-eth/AGBD
• Generation code: AGBD-GFMs → data/agbd_lite/
• Trained supervised weights & test predictions: available on Zenodo (see the generation repo README).

📄 License

Released under CC BY-NC 4.0, following the parent AGBD dataset.

🤝 Citing

AGBD-Lite is a subset of AGBD. If you use it, please cite the AGBD dataset:

@article{Sialelli2025,
  title = {AGBD: A Global-scale Biomass Dataset},
  volume = {X-G-2025},
  ISSN = {2194-9050},
  url = {http://dx.doi.org/10.5194/isprs-annals-X-G-2025-829-2025},
  DOI = {10.5194/isprs-annals-x-g-2025-829-2025},
  journal = {ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
  publisher = {Copernicus GmbH},
  author = {Sialelli, Ghjulia and Peters, Torben and Wegner, Jan D. and Schindler, Konrad},
  year = {2025},
  month = jul,
  pages = {829--838}
}