M3LEO

community
Activity Feed

AI & ML interests

Earth Observation Datasets

Recent Activity

mataln  updated a Space about 2 months ago
M3LEO/README
mataln  updated a dataset about 2 months ago
M3LEO/southamerica
mataln  updated a dataset about 2 months ago
M3LEO/china
View all activity

image/png

M3LEO: A Multi-Modal Multi-Label Earth Observation Dataset

This repository contains information about the multi-modal multi-label, wide area Earth Observation (EO) datasets collated during the 2023 Frontier Development Lab. It contains around 40 TB of co-aligned machine learning ready data tiles, spanning 9 EO datasets and 6 geographic regions. For ease of access, the dataset has been compressed as parquet files. For a smaller (uncompressed) version of our dataset, check out the M3LEO miniset.

PAPER @ NeurIPS 2024 D&B track

Decompression

If you need to decompress the files, please see the main README at the github repo.
If you want to use them directly from the parquet files, the original .tif/.nc files were read into the rows as binary file data sources

Tile Definitions

Each data tile covers an area of 4480m x 4480m (448x448 pixels at 10m/pixel) and is labelled with a unique identifier based on location.

image/png

Areas of Interest

Our areas of interest (AOIs) span China, Conus, Europe, the Middle East, Pakin, and South America. Each AOI has a '.geojson' file associated with the geometries and identifiers of each data tile.

image/png

Train-Test-Validation Splits

For each geopgraphic area, we provide '.csv' files with predefined train, test and validation splits that can be used for repeatability and comparability of experiments. 60% of tiles are allocated for training, 20% for validation, and 20% for testing.

Temporal Coverage

As of now, M3LEO contains data from 2018 - 2020 for SAR amplitude and multi-spectral Sentinel-2 imagery. Other datasets are provided for 2020 only. Future iterations might extend the dataset to other years.

Datasets

The M3LEO dataset spans 9 diverse EO data types, covering input EO imagery and associated labels.

image/png

Synthetic Aperture Radar Datasets

  • s1grd-[2018-2020]: Sentinel-1 C-band SAR Ground Range Detected (GRD) amplitude data. Contains VV and VH polarization channels at 10m resolution, with seasonal median composites (4 seasons/year) for ascending and descending orbits.

  • gssic: Global Seasonal Sentinel-1 Interferometric Coherence and Backscatter dataset at 90m resolution.

  • gunw-dateinit_dateend: ARIA Sentinel-1 Geocoded Unwrapped Interferograms at 90m resolution. Date selection prioritizes maximum interferometric pair availability within the specified period [dateinit, dateend].

Optical Imagery

  • s2srm-[2018-2020]: Harmonized Sentinel-2 Level 2A surface reflectance data at 10m resolution. Includes 10 spectral bands (B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12) as cloudless monthly medians for March, June, September, and December.

Labeled Datasets

  • biomass-2020: ESA Climate Change Initiative (CCI) Above Ground Biomass annual maps at 90m resolution.

  • esaworldcover-2020: ESA World Cover land use/land cover maps at 10m resolution.

  • modis44b006veg: MODIS Vegetation Continuous Fields (VCF) annual maps at 250m resolution.

  • ghsbuilts-2020: European Commission JRC Global Human Settlement Layer (GHSL) Built-up Surface dataset at 100m resolution.

Digital Elevation Model

  • srtmdem: NASA Shuttle Radar Topography Mission (SRTM) digital elevation model at 30m resolution.

Acknowledgements

This work has been enabled by Frontier Development Lab Europe a public / private partnership between the European Space Agency (ESA), Trillium Technologies, the University of Oxford and leaders in commercial AI supported by Google Cloud and NVIDIA, developing open science for all Humankind.