isp-uv-es/ssl4eo-l-benchmark-oli-cdl-taco

SSL4EO-L Benchmark (Landsat 8 OLI SR / CDL)

This is a repackaging, not a new dataset. It is SSL4EO-L Benchmark (Landsat 8 OLI SR / CDL) by SSL4EO-L authors (Stewart et al.), USGS Landsat, USDA NASS (Cropland Data Layer), 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: CC0-1.0

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{stewart2023ssl4eol,
  title     = {SSL4EO-L: Datasets and Foundation Models for Landsat Imagery},
  author    = {Stewart, Adam J. and Lehmann, Nils and Corley, Isaac A. and Wang, Yi and Chang, Yi-Chia and Ait Ali Braham, Nassim and Sehgal, Shradha and Robinson, Caleb and Banerjee, Arindam},
  booktitle = {Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track},
  year      = {2023},
  doi       = {10.52202/075280-2612}
}

About the data

25000 Landsat 8 OLI surface-reflectance chips of 264x264 at 30 m over the continental United States, each paired with the USDA Cropland Data Layer of its acquisition year as a 134-class crop-type mask.

25,000 samples · splits: test 3,750 · train 17,500 · validation 3,750 · tasks: semantic-segmentation

Packaged as TACO v3.

Full description

Radiometry. The imagery is 8-bit as published: surface reflectance 0..0.3 mapped linearly onto 0..255 and clipped, by torchgeo's compress_oli_sr.sh. Approximately SR = (DN + 0.5) x 0.3 / 255 (approx_scale_factor); a 255 means SR >= 0.3, which is 52% of NIR, 18% of SWIR1 and 7% of SWIR2 pixels.

Legend. CDL's sparse codes are remapped to a dense 0..133 legend, with the original CDL code of each class documented in the contract. The names are CDL's own, reconstructed from USDA NASS per-state metadata.

Splits. The release ships none. torchgeo's deterministic 70/15/15 shuffle (numpy seed 0 over the sorted path list) is reproduced exactly, giving 17500 train / 3750 val / 3750 test, and split_source records that it is derived rather than official. The split is chip-level and the chips come from only 380 Landsat acquisitions, so most acquisitions appear in more than one split: spatially disjoint, but not acquisition-disjoint.

Scope. Only the OLI/CDL cell of the benchmark is carried; its NLCD masks and ETM+/TM sensors are not included.

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:

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

path = hf_hub_download("isp-uv-es/ssl4eo-l-benchmark-oli-cdl-taco", "ssl4eo-l-benchmark-oli-cdl.zip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])

or from a local copy:

ds = Dataset("ssl4eo-l-benchmark-oli-cdl.zip")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["oli"].array.shape

Metadata without decoding anything:

import taco
taco.read("ssl4eo-l-benchmark-oli-cdl.zip")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	oli	raster	multispectral	7 band(s), unit 1, requantised
target	crop_type	mask	label_raster	134 classes

Licence

CC0-1.0

Providers: SSL4EO-L authors (Stewart et al.), USGS Landsat, USDA NASS (Cropland Data Layer)

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.