JulioContrerasH/Major-TOM-TACO

This dataset follows the TACO specification.

Major TOM Core-Combo (TACO)

A TACO-formatted multimodal dataset built on Major TOM. Each sample aligns Sentinel-2 L2A, Sentinel-1 RTC, Copernicus DEM 30, and optional Major TOM embeddings under a single grid so patches are spatially matched and ready for training and evaluation.

Description

Dataset

This dataset packages co-registered patches drawn from the Major TOM core datasets (S2 L2A / S2 L1C / S1 RTC / DEM) and the official Major TOM embeddings (SSL4EO, DINOv2, SigLIP, DeCUR, MMEarth, AlphaEarth). Major TOM defines a geographical indexing grid and a metadata structure to merge multi-source EO data—ideal for large-scale pretraining, representation learning, and multimodal fusion.

What each sample contains:

• S2 L2A (10 m) — 13 MSI bands (B1–B12 incl. B10), with native 20 m/60 m bands resampled to 10 m for a consistent stack.
• S1 RTC (10 m) — backscatter (VV/VH) and optional geometry/angle layers, co-registered to the S2 grid.
• DEM (30 m → 10 m) — Copernicus DEM 30 resampled to 10 m, with optional derived slope/aspect.
• Embeddings (optional) — one or more per-patch vectors from Major TOM families (e.g., SSL4EO, DINOv2, SigLIP, DeCUR, MMEarth, AlphaEarth).
• Metadata — acquisition dates, orbit/pass, QA (e.g., S2 cloud metrics when available), CRS and affine transform, plus upstream lineage.

The dataset inherits global land coverage from Major TOM Core and is extensible (you can enable/disable modalities and embeddings per tortilla).

Sensors used

• Sentinel-2 MSI (L2A/L1C) — optical multispectral, 13 bands (443–2190 nm) at 10/20/60 m; all represented on a unified 10 m grid.
• Sentinel-1 RTC — SAR backscatter (VV/VH) in analysis-ready RTC format at ~10 m.
• Copernicus DEM 30 — global 30 m elevation resampled to 10 m for alignment.
• Embeddings — model-derived features aligned to the same grid (families: SSL4EO, DINOv2, SigLIP, DeCUR, MMEarth; optional AlphaEarth subset).

Creators

• ESA Φ-lab & collaborators

Original datasets

See the official Major TOM organization for coverage, counts and updates:

• Major TOM organization overview (Hugging Face)
• Core datasets: Core-S2L2A, Core-S2L1C, Core-S1RTC, Core-DEM
• Embedding releases: Core-S2L1C-SSL4EO, Core-S1RTC-SSL4EO, Core-S2RGB-DINOv2, Core-S2RGB-SigLIP, Core-S2L1C-DeCUR, Core-S1RTC-DeCUR, Core-S2L2A-MMEarth, Core-AlphaEarth-Embeddings
• Spec & paper: “Major TOM: Expandable Datasets for Earth Observation” (arXiv) and the IGARSS 2024 citation (see Publications)

Note: sizes/coverage and licenses are inherited from upstream. Always consult the upstream dataset cards.

TACO dataset

Construction

• Grid & tiling: adopt Major TOM’s global grid; fixed-size 512 × 512 px @ 10 m (~5.12 km) windows keyed by grid cell.
• Temporal pairing: select closest-in-time S1/S2 acquisitions per cell (configurable), preferring clear-sky S2 for optical stacks.
• Embeddings: if enabled, fetch and store per-patch embedding vectors (Parquet/NPY sidecar) indexed by the tortilla.
• Lineage metadata: upstream identifiers/commits, acquisition timestamps, processing levels, and provenance.

Default patch geometry

• Spatial extent: 5160 m × 5160 m
• S2 tensor (bands): 13 (B1–B12 incl. B10) → 512 × 512 × 13
• S1 tensor (VV/VH): 2 (plus optional angle) → 512 × 512 × 2–3
• DEM tensor: 1 (or +slope/aspect) → 512 × 512 × 1–3
• Embeddings: 1..N vectors stored as sidecars (per-patch, not per-pixel)

Spectral Bands (S2 MSI)

We expose the native Sentinel-2 MSI band set and place all on a unified 10 m grid:

idx	Band	Name	Central λ	Nominal Res.	Notes
0	B1	Coastal Aerosol	443 nm	60 m	resampled to 10 m
1	B2	Blue	492 nm	10 m
2	B3	Green	560 nm	10 m
3	B4	Red	665 nm	10 m
4	B5	Red Edge 1	704 nm	20 m	resampled to 10 m
5	B6	Red Edge 2	740 nm	20 m	resampled to 10 m
6	B7	Red Edge 3	783 nm	20 m	resampled to 10 m
7	B8	NIR (Broad)	833 nm	10 m
8	B8A	NIR (Narrow)	865 nm	20 m	resampled to 10 m
9	B9	Water Vapour	945 nm	60 m	resampled to 10 m
10	B10	Cirrus (WV 1375 nm)	1375 nm	60 m	optional for ML
11	B11	SWIR 1	1610 nm	20 m	resampled to 10 m
12	B12	SWIR 2	2200 nm	20 m	resampled to 10 m

---

🔄 Reproducible Example

import tacoreader
import rasterio as rio
import numpy as np
import matplotlib.pyplot as plt

# Load the dataset (replace with the final registry name when published)
ds = tacoreader.load("tacofoundation:majortom-core-combo")

# Read a sample
i = 0
row = ds.read(i)
row_id = ds.iloc[i]["tortilla:id"]

s2_path = row.read("S2_L2A")      # GeoTIFF
s1_path = row.read("S1_RTC")      # GeoTIFF
dem_path = row.read("DEM")        # GeoTIFF
# emb_path = row.read("EMB")      # Optional embeddings (Parquet/NPY)

with rio.open(s2_path) as s2, rio.open(s1_path) as s1, rio.open(dem_path) as dem:
    # Simple S2 RGB (B4,B3,B2)
    rgb = np.stack([s2.read(4), s2.read(3), s2.read(2)], axis=0)
    rgb = np.transpose(rgb, (1,2,0))
    rgb_norm = np.clip(rgb / 2000.0, 0, 1)

    vv = s1.read(1)
    dem_elev = dem.read(1)

fig, ax = plt.subplots(1,3, figsize=(13,4.2))
ax[0].imshow(rgb_norm); ax[0].set_title(f"S2 RGB — {row_id}"); ax[0].axis("off")
ax[1].imshow(vv);       ax[1].set_title("S1 VV"); ax[1].axis("off")
ax[2].imshow(dem_elev); ax[2].set_title("DEM");   ax[2].axis("off")
plt.tight_layout(); plt.show()

🛰️ Sensor Information

Sources in this dataset: sentinel2msi, sentinel1-rtc, cop-dem30, and Major TOM embeddings (SSL4EO, DINOv2, SigLIP, DeCUR, MMEarth; optional AlphaEarth).

🎯 Tasks

General-purpose: self-supervised pretraining, representation learning, multimodal fusion, semantic segmentation, change detection, classification, retrieval.

📂 Original Data Repositories (Upstream)

• Major TOM organization page.

💬 Discussion

Use the Major TOM org discussions/Spaces and the satellite-image-deep-learning Discord (Major TOM channels) to coordinate contributions and combinations.

🔀 Split Strategy

All train.

📚 Scientific Publications

Publication 01

• arXiv: Major TOM: Expandable Datasets for Earth Observation (Francis & Czerkawski, 2024). BibTeX:

@article{MajorTOM2024,
author  = {Francis, Alistair and Czerkawski, Mikolaj},
title   = {Major TOM: Expandable Datasets for Earth Observation},
journal = {arXiv preprint arXiv:2402.12095},
year    = {2024},
doi     = {10.1109/IGARSS53475.2024.10640760}
}

Publication 02

• arXiv (2024): Global and Dense Embeddings of Earth: Major TOM Floating in the Latent Space.

    @article{czerkawski2024global,
    title={Global and dense embeddings of earth: Major tom floating in the latent space},
    author={Czerkawski, Mikolaj and Kluczek, Marcin and Bojanowski, J{\"A} and others},
    journal={arXiv preprint arXiv:2412.05600},
    year={2024},
    doi= {10.48550/arXiv.2412.05600}
    }
    

🤝 Data Providers

Name	Role	URL
ESA Φ-lab	producer / coordinator	https://philab.esa.int/
Major TOM (HF)	publisher (org)	https://huggingface.co/Major-TOM
CloudFerro (embeddings infra)	collaborator	https://cloudferro.com/

---

🧑‍🔬 Curators

Name	Organization	URL
Julio Contreras	Image & Signal Processing	https://juliocontrerash.github.io/
TACO Foundation	Curation (TACO format)	https://huggingface.co/datasets/tacofoundation/

---

🎞 Official Image Datasets (from Major TOM)

Dataset	Modality	Number of Patches	Sensing Type	Comments
Core-S2L2A	Sentinel-2 Level 2A	2,245,886	Multi-Spectral	Global (≈23 TB)
Core-S2L1C	Sentinel-2 Level 1C	2,245,886	Multi-Spectral	Global (≈23 TB)
Core-S1RTC	Sentinel-1 RTC	1,469,955	SAR	Global (≈16 TB)
Core-DEM	Copernicus DEM 30	1,837,843	Digital Surface Model	Global (≈1 TB)

---

📊 Official Embedding Datasets (from Major TOM)

Dataset	Modality	# Embeddings	Sensing Type	Source Dataset	Source Model	Size
Core-S2L1C-SSL4EO	Sentinel-2 Level 1C	56,147,150	Multi-Spectral	Core-S2L1C	SSL4EO-ResNet50-DINO	252.9 GB
Core-S1RTC-SSL4EO	Sentinel-1 RTC	36,748,875	SAR	Core-S1RTC	SSL4EO-ResNet50-MOCO	332.5 GB
Core-S2RGB-DINOv2	Sentinel-2 L2A (RGB)	56,147,150	True Colour	Core-S2L2A	DINOv2	223.1 GB
Core-S2RGB-SigLIP	Sentinel-2 L2A (RGB)	20,212,974	True Colour	Core-S2L2A	SigLIP-SO400M-384	41.3 GB
Core-S2L1C-DeCUR	Sentinel-2 Level 1C	56,147,150	Multi-Spectral	Core-S2L1C	SSL4EO-ResNet50-DeCUR	252.9 GB
Core-S1RTC-DeCUR	Sentinel-1 RTC	36,748,875	SAR	Core-S1RTC	SSL4EO-ResNet50-DeCUR	332.5 GB
Core-S2L2A-MMEarth	Sentinel-2 L2A (MSI)	39,727,477,454	Multi-Spectral	Core-S2L2A	MMEarth	5080 GB
Core-AlphaEarth-Embeddings	Multimodal	71,276,453,136	Multiple	AlphaEarth	AlphaEarth	6070 GB

---

Key sources for facts & tables: Major TOM org page with official tables and IGARSS citation (org card), the arXiv paper describing the framework and grid, specific upstream dataset cards (Core-S2L2A / Core-S1RTC / Core-DEM), the AlphaEarth subset card, and ESA Φ-lab’s post on embedding expansions. (huggingface.co, arxiv.org, philab.esa.int)