Cloud 3D - GOES Pretraining Dataset

Description

GOES (Geostationary Operational Environmental Satellite) imagery subset from the Global 3D Cloud Reconstruction Dataset. Contains multispectral geostationary satellite imagery from GOES-16/ABI for 3D cloud structure reconstruction. Each sample contains 20 bands: 16 spectral channels plus satellite and solar angles. 512x512 pixel patches in Cloud-Optimized GeoTIFF format.

Dataset Overview

105 partitions 2018 - 2024 temporal coverage

Spatial Coverage

Click on any region to view partition details

Keywords

cloud microphysics 3d reconstruction geostationary satellites GOES-16 remote sensing tropical cyclones deep learning

ML Tasks

regression foundation-model

TACO Structure (Root-Sibling Uniform Tree)

Hierarchical structure showing representative samples across levels. The "..." notation indicates additional samples following the same pattern. All samples at the same level share identical structure (RSUT constraint).

Hierarchy Details

Level	Types	Total Samples	Sample IDs (preview)
Level 0	`All FILE`	91,423	Root level samples

Metadata Fields by Level

These fields are available for querying with SQL when using TacoReader.

LEVEL0 (20 fields)

Field Name	Type	Description
id	string	Unique sample identifier within parent scope. Must be unique among siblings.
type	string	Sample type discriminator (FILE or FOLDER).
taco:header	binary	Binary TACOTIFF header (35 bytes + tile counts) for fast reading without IFD parsing
stac:crs	string	Coordinate reference system (WKT2, EPSG, or PROJ string)
stac:tensor_shape	list<item: int64>	Raster dimensions e.g. [bands, height, width]
stac:geotransform	list<item: double>	GDAL affine transform [origin_x, pixel_width, rot_x, origin_y, rot_y, pixel_height]
stac:time_start	timestamp[us]	Acquisition start timestamp (microseconds since Unix epoch, UTC)
stac:centroid	binary	Raster center point in EPSG:4326 (WKB binary format)
stac:time_end	timestamp[us]	Acquisition end timestamp (microseconds since Unix epoch, UTC)
stac:time_middle	timestamp[us]	Midpoint between start and end timestamps (microseconds since Unix epoch, UTC)
geotiff:stats	list<item: list<item: float>>	Per-band statistics (List[List[Float32]]): categorical mode returns class probabilities, continuous mode returns [min, max, mean, std, valid%, p25, p50, p75, p95]
cloud3d:satellite	string	Geostationary satellite platform (GOES, HIMAWARI, or MSG)
cloud3d:cyclone	bool	Whether the sample contains tropical cyclone imagery
majortom:code	string	MajorTOM spherical grid cell identifier (e.g., 0100km_0003U_0005R) with ~dist_km spacing
geoenrich:elevation	float	Mean elevation in meters (GLO-30 DEM)
geoenrich:precipitation	float	Mean annual precipitation in mm estimated from GPM data
geoenrich:temperature	float	Mean annual temperature in °C estimated from MODIS LST data
geoenrich:admin_countries	string	Country name at centroid location
internal:current_id	int64	Current sample position at this level (0-indexed). Enables O(1) random access and relational JOINs (ZIP, FOLDER, TACOCAT).
internal:parent_id	int64	Foreign key referencing parent sample position in previous level (ZIP, FOLDER, TACOCAT).

Loading the Dataset

# pip install tacoreader
import tacoreader

# Load dataset
ds = tacoreader.load("https://data.source.coop/taco/3dclouds/pretraining/goes/")

# Basic info
print(f"ID: {ds.id}")
print(f"Version: {ds.version}")
print(f"Samples: {len(ds.data)}")

# Coming soon: R support is planned but not yet available
# install.packages("tacoreader")
library(tacoreader)

# Load dataset
ds <- load_taco("https://data.source.coop/taco/3dclouds/pretraining/goes/")

# Basic info
cat(sprintf("ID: %s\n", ds$id))
cat(sprintf("Version: %s\n", ds$version))
cat(sprintf("Samples: %d\n", nrow(ds$data)))

# Coming soon: Julia support is planned but not yet available
# using Pkg; Pkg.add("TacoReader")
using TacoReader

# Load dataset
ds = load_taco("https://data.source.coop/taco/3dclouds/pretraining/goes/")

# Basic info
println("ID: ", ds.id)
println("Version: ", ds.version)
println("Samples: ", size(ds.data, 1))

Providers & Curators

Data Providers

NOAA producerhttps://www.noaa.gov

European Space Agency (ESA) licensorhttps://www.esa.int

source.coop hosthttps://source.coop

Dataset Curators

Name	Organization	Email
Cesar Aybar	Universitat de València	cesar.aybar@uv.es
Shirin Ermis	University of Oxford	—
Lilli Freischem	University of Oxford	—
Stella Girtsou	National Observatory of Athens	—
Kyriaki-Margarita Bintsi	Harvard Medical School	—
Emiliano Diaz Salas-Porras	Universitat de València	—
Michael Eisinger	European Space Agency	—
William Jones	University of Oxford	—
Anna Jungbluth	European Space Agency	—
Benoit Tremblay	Environment and Climate Change Canada	—

Publications & Citations

How to Cite This Dataset

If you use this dataset in your research, please cite:

10.48550/arXiv.2511.04773

Ermis, S., Aybar, C., Freischem, L., Girtsou, S., Bintsi, K.-M., Diaz Salas-Porras, E., Eisinger, M., Jones, W., Jungbluth, A., & Tremblay, B. (2025). Global 3D Reconstruction of Clouds & Tropical Cyclones. Tackling Climate Change with Machine Learning Workshop at NeurIPS 2025.

BibTeX

@dataset{cloud3d-pretraining-goes0,
  title = {Cloud 3D - GOES Pretraining Dataset},
  author = {Cesar Aybar and Shirin Ermis and Lilli Freischem and Stella Girtsou and Kyriaki-Margarita Bintsi and Emiliano Diaz Salas-Porras and Michael Eisinger and William Jones and Anna Jungbluth and Benoit Tremblay},
  year = {2018},
  version = {0.1.0},
  publisher = {Universitat de València}
}