isp-uv-es/satbird-kenya-taco

SatBird Kenya

This is a repackaging, not a new dataset. It is SatBird Kenya by Mila / Quebec AI Institute (SatBird authors), Cornell Lab of Ornithology / eBird, ESA Copernicus / Sentinel-2, converted to TACO. Pixel values and labels are kept as released except where the description below says otherwise. All credit belongs to the original authors: if you use it, please cite them and follow their licence.

original dataset · paper · licence: CC-BY-NC-4.0

Repackaged into TACO by the Image and Signal Processing Group (ISP), Universitat de València, within the ELLIOT project.

Citation

Please cite the original work:

@inproceedings{teng2023satbird,
  title     = {SatBird: a Dataset for Bird Species Distribution Modeling using Remote Sensing and Citizen Science Data},
  author    = {Teng, M{\'e}lisande and Elmustafa, Amna and Akera, Benjamin and Bengio, Yoshua and Radi, Hager and Larochelle, Hugo and Rolnick, David},
  booktitle = {Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track},
  pages     = {75925--75950},
  year      = {2023}
}

About the data

9763 eBird hotspots, each pairing a 10 m Sentinel-2 patch with the encounter rate of every one of 1054 bird species -- the fraction of complete checklists at that hotspot reporting the species -- so the target is 1054 continuous values rather than a label or a set.

9,763 samples · splits: test 1,322 · train 6,789 · validation 1,652 · tasks: regression

Packaged as TACO v3.

Full description

Splits. The release's own: 6789 train, 1652 valid, 1322 test.

Scope. Kenya only. The release's USA summer and winter splits are not included; this is the Kenya subset.

Georeferencing. The files' georeferencing is not usable and is not carried: every patch is tagged with its Sentinel-2 tile's transform, so 9,763 rasters share just 63 distinct (CRS, origin) pairs, and taken at face value every patch of a tile would sit on the same 6 km of ground. The location of record is the release's CSV, lon/lat plus footprint_wkt. Each polygon's projected extent divided by its raster's shape gives 9.977-9.979 m by 9.994-9.997 m, which reproduces the declared 10 m to within 0.25%, so the pixel size is true even though the origin is not, and it is declared on the bands as gsd_m.

Bands. Order is B, G, R, NIR, settled against the release's own RGB rendering pixelwise (0.986, 0.986, 0.987 on the diagonal; band 3 correlates with no visual channel) rather than taken from the folder name. Reflectance is surface (BOA) x 10,000: blue is the darkest band at the 1st percentile and reaches 0, which TOA does not do. nodata is 0, contiguous off-swath edge blocks covering up to 93.5% of one patch, and zero_frac carries the fill fraction per sample because a scalar sentinel cannot state a joint across-band condition.

Bioclimatic layers. The release's (19, 6, 6) rasters are not carried: 316 of 9,763 are (19, 6, 0), i.e. empty, and 25 more are truncated, so the layer cannot satisfy a structure contract that is identical for every sample. The same 19 variables ride as bio_1..bio_19 columns, with bio_imputed True on the 206 hotspots the release's second CSV shows were filled with a single constant vector.

Checklists. num_complete_checklists is carried twice because the target JSON and the hotspot table disagree on 705 hotspots; the JSON's is the denominator that goes with the rates.

Target sparsity. A mean of 21.8 species of 1054 per hotspot, 6 hotspots all-zero, and rate sums from 0.00 to 212.00, because these are per-species fractions and not a distribution over species.

Getting started

git clone --recursive https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the reader: C++23, CMake, Ninja, pkg-config, libcurl >= 7.83, OpenSSL >= 3

Read it straight from the Hub:

from huggingface_hub import snapshot_download
from taco.ml import Dataset, plot_sample

path = snapshot_download("isp-uv-es/satbird-kenya-taco", repo_type="dataset")
ds = Dataset(path)                  # the repository is the container: parts + .tacocat
plot_sample(ds[0])

or from a local copy:

ds = Dataset("satbird-kenya")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["s2"].array.shape

Metadata without decoding anything:

import taco
taco.read("satbird-kenya/.tacocat")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	s2	raster	optical	4 band(s), unit 1, scaled
input	visual	raster	optical	3 band(s), render
target	species_encounter_rate	class_distribution		1054 classes, unit 1

Licence

CC-BY-NC-4.0

Terms of use:

• Non-commercial use only (CC-BY-NC-4.0).
• The data are provided under the same terms as the eBird data they derive from: the eBird Terms of Use (https://www.birds.cornell.edu/home/terms-of-use/) and the eBird Status and Trends Products Terms of Use (https://science.ebird.org/en/status-and-trends/products-access-terms-of-use).
• Anyone sharing data derived from this dataset must share it under the same terms.

Required credits:

• SatBird: Teng et al. (Mila)
• Bird observations: eBird, Cornell Lab of Ornithology
• Contains modified Copernicus Sentinel data
• Bioclimatic variables: WorldClim (Hijmans et al., 2005)

Providers: Mila / Quebec AI Institute (SatBird authors), Cornell Lab of Ornithology / eBird, ESA Copernicus / Sentinel-2

Acknowledgements

TACO was designed by César Aybar and is specified at https://asterisk.coop/taco/spec/.

Built by Oscar Pellicer within the ELLIOT project at the Image and Signal Processing Group (ISP), Universitat de València.