SSL4EO-L: Landsat Self-Supervised Pretraining Dataset (TM, ETM+, OLI/TIRS, 5M patches)
- Provider
- Stewart, Lehmann, Corley, Wang et al. (TorchGeo; UIUC, TUM, Microsoft AI for Good)
- License
- CC0-1.0 (source)
- Size
- 47 files
- Format
- CSV (in sampled files)
- Stored at
- Hugging Face (https://huggingface.co/datasets/torchgeo/ssl4eo_l)
- Access
- Free and public. No account with the data provider is needed.
- DOI
- 10.48550/arXiv.2306.09424
- Category
- Earth Observation & Geospatial
- Source
- https://torchgeo.readthedocs.io/
About this dataset
SSL4EO-L is the first large self-supervised pretraining corpus for the Landsat family, released with the NeurIPS 2023 Datasets and Benchmarks paper by Stewart et al. Patches are sampled around the 10,000 most populous cities with a 50 km Gaussian, non-overlapping, at 250,000 locations, each imaged in four seasons, giving 1 million unlabeled 264x264 pixel patches (7.92 km at 30 m) per sensor and product and 5 million in total. Thermal and panchromatic bands are resampled to 30 m so every patch is one multi-band file. Five products are provided as separate top-level folders: tm_toa (Landsat 4/5 TM top of atmosphere), etm_toa and etm_sr (Landsat 7 ETM+ TOA and surface reflectance), oli_tirs_toa and oli_sr (Landsat 8/9 OLI/TIRS), plus sampled_locations.csv with the patch centres. Data were exported from Google Earth Engine; about 1.7 TB. Load with torchgeo.datasets.SSL4EOL. The paper states that all datasets it creates are released under CC0 1.0 and that the source USGS Landsat data are public domain; cite Stewart et al. 2023.
How to load this dataset in Python
import pandas as pd
df = pd.read_csv("hf://datasets/torchgeo/ssl4eo_l/sampled_locations.csv")
df.head()Data structure
File types in sampled files: CSV
Files
Example file: sampled_locations.csv
Analyze this dataset without downloading it
On Scigantic, SSL4EO-L: Landsat Self-Supervised Pretraining Dataset (TM, ETM+, OLI/TIRS, 5M patches) opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.
This page is also available as Markdown and as JSON.
Related Earth Observation & Geospatial datasets
- Kartverket Sentinel-2 Cloud-free Mosaics of Norway, 10 m RGB, 2015-2017, 2018 and 2021-2025 (Norwegian Mapping Authority, Source Cooperative)
- PoroTomo
- USGS Topographic Map Indices: 7.5-Minute, 15-Minute, 30x60-Minute and 1x2-Degree Quadrangle Grids
- Digital Earth Africa Fractional Cover
- EarthDEM
- Overture Maps - Places Dataset - Riyadh, Saudi Arabia
- Cloud Native MGRS
- OPERA Radiometric Terrain Corrected SAR Backscatter from Sentinel-1 validated product (Version 1)