A standard for curating, sharing and combining large-scale EO datasets.
Started at ESA Φ-lab; a community project that Asterisk Labs supports and co-develops.
AI models like square chips of rasters. Mostly. We define a grid, starting from the equator and dividing our planet into strips of 10 km. In each row (one strip), we divide the circumference of that circle of latitude into 10 km segments, to create approximately square cells. There are 4,008 of them at the equator, but only 2,472 of them at 51.9°N. In total, Earth is covered by 5,113,745 cells of Major TOM at 10 km spacing.
We index cells by indicating how many rows up or down (U/D) and how far to the left or right (L/R) a given segment is. For example, you can find the Bonnie Banks o' Loch Lomond if you go 624 rows up from the equator, and move 29 cells to the left of the Greenwich meridian in that row. This leaves us with a code of 624U_29L. You can see that on top of that, we define a 10.68 km footprint for a chip in order to cover the cell in a product with UTM projection. Newer Major TOM products often use 10.56 km chips, centred on the cell, for optimised chunking. The underlying 10.00 km grid remains the same in both cases.
This is how a larger (about 110 km on each side) Sentinel-2 product gets cropped into a Major TOM sample. We can do this to any source, in any projection, although it is best when it is in UTM. See 4 examples of Major TOM products over the same cell. Most of them can be easily built using majortom.build functionality in our code.
Major TOM is a standard for developing AI-ready datasets with environmental raster data. By defining a common grid of points, independent developers of datasets can produce samples that natively overlap over the same cells.
Apart from a framework for defining these samples, the initial release produced Major TOM Core datasets, which at the time, were the largest (nearly 70 TB of data altogether) open and free pre-training datasets with Copernicus (Sentinel-2, Sentinel-1) data.
Since the release, these datasets have been used on many occasions to train the next generation of AI models for Earth observation. These models can also be used to encode data, an example of which are the Major TOM embedding datasets, offering a consistent experimentation sandbox by deriving embeddings from the same predefined (and accessible) set of samples.
The initial Core release is a set of near-global, open datasets built on one shared grid:
| Dataset | Source | Patches | Size |
|---|---|---|---|
| Core-S2L2A | Sentinel-2 L2A | 2.25 M | ~23 TB |
| Core-S2L1C | Sentinel-2 L1C | 2.25 M | ~23 TB |
| Core-S1RTC | Sentinel-1 RTC (SAR) | 1.47 M | ~16 TB |
| Core-DEM | Copernicus DEM 30 | 1.84 M | ~1 TB |
| Core-VIIRS-Nighttime-Light | VIIRS night lights | 2.11 M | ~400 GB |
On the same grid, ten embedding datasets encode these patches with published models — SSL4EO (Sentinel-2 and Sentinel-1), DINOv2, SigLIP, DeCUR, MMEarth, UniverSat, and the AlphaEarth satellite embeddings — so embeddings from different models line up patch-for-patch.
Browse the Core datasets in your browser:
Asterisk Labs used Major TOM to train BetaEarth and will use it for many other experimentation scenarios. At this time, we are working hard to improve the AI-ready aspect of Major TOM by rereleasing some of the data using our latest developments like TACO, rumi and GeoZL.
asterisk labs