Featured Products

Source Cooperative is a Radiant Earth project

3Dclouds | Source Cooperative

3Dclouds

This repository hosts the AI-ready datasets for the paper "Global 3D Reconstruction of Clouds & Tropical Cyclones." It contains paired 2D geostationary imagery (GOES, MSG, Himawari) and 3D vertical profiles (CloudSat/CALIPSO) used to train ML models for 3D cloud reconstruction, with a dedicated dataset for tropical cyclones.

Product Details

Visibility: Public
Owner: Cesar Aybar
Created: 20 Oct 2025
Last Updated: 27 Oct 2025

Product Contents

root

README

A Global, AI-Ready Dataset for 3D Cloud reconstruction

Cesar Aybar¹ • Shirin Ermis² • Lilli Freischem² • Stella Girtsou³⁴ • Kyriaki-Margarita Bintsi⁵
Emiliano Diaz Salas-Porras¹ • Michael Eisinger⁶ • William Jones² • Anna Jungbluth⁶ • Benoit Tremblay⁷

_{¹Universitat de València, ²University of Oxford, ³National Observatory of Athens, ⁴National Technical University of Athens,

⁵Harvard Medical School & Massachusetts General Hospital, ⁶European Space Agency, ⁷Environment and Climate Change Canada}

About This Dataset

This repository contains the AI-ready datasets presented in the paper "A Global, AI-Ready Dataset for 3D Cloud reconstruction". This foundational dataset provides a large-scale collection of paired 2D multispectral geostationary (GEO) satellite imagery and co-located 3D vertical cloud property profiles from the CloudSat radar. The data is curated to train deep learning models for 3D cloud structure reconstruction and is organized into three distinct components:

A Pre-training dataset for self-supervised learning.
A Clouds (Fine-tuning) dataset for supervised training on general cloud scenes.
A Tropical Cyclones (TC) benchmark dataset for model evaluation on intense storms.

1. Input Data (2D Geostationary Imagery)

The model inputs are 2D patches from three geostationary satellites, which together provide near-global coverage:

GOES-16/ABI (Americas)
Himawari-8/AHI (Asia & Oceania)
Meteosat Second Generation (MSG)/SEVIRI (Europe & Africa)

To create a unified, sensor-agnostic input for the models, the 11 closest spectral channels (a mix of reflectances and brightness temperatures) are selected from each satellite, normalized, and used as the input tensor.

2. Target Data (3D Vertical Profiles)

The ground truth data (the prediction target) consists of vertical profiles from the CloudSat CPR 2B products. These profiles are spatiotemporally aligned to the nearest GEO pixel. The primary target variables provided are:

Radar Reflectivity (Z): The primary measure of cloud structure and intensity (in dBZ).
Ice Water Content (IWC): A key microphysical property (in $\mathrm{gm^{-3}}$).
Effective Radius ($\mathrm{r_e}$): A measure of cloud droplet size (in $\mathrm{\mu m}$).

3. Dataset Splits

The splits distinguish between general cloud scenes and targeted tropical cyclone (TC) events for both pre-training and fine-tuning/evaluation.

Split	Data Type	Purpose	Number of Files	Total Size
Pre-training	General Clouds	Self-supervised pre-training	150,000	~0.6 TB
	Tropical Cyclones	Self-supervised pre-training	50,000	~0.4 TB
Fine-tuning	General Clouds	Supervised training & validation	~270,000	~273 GB
	Tropical Cyclones	Supervised evaluation/benchmark	~703	~1.7 GB

How to Access

The complete dataset is hosted on Amazon Web Services (AWS) S3 via the source.coop public data initiative. You can download the data using the tacoreader API, or any other tool compatible with AWS S3.

Usage Examples (Google Colab)

We provide Google Colab notebooks to help you get started with loading, exploring, and visualizing the data. These notebooks demonstrate how to read the Zarr files and plot the input (GEO) and target (CloudSat) data.

Purpose	Description	Colab Link
Test the dataset (no labels)	Visualize and explore the geostationary satellite inputs only, without requiring CloudSat targets. Useful for quick inspection or sensor-based analysis.
Test the dataset (with CloudSat labels)	Visualize both the geostationary inputs and the co-located CloudSat 3D profiles. Useful for evaluating input–target alignment and dataset integrity.

Data Curation Methodology

This dataset was created using a multi-step pipeline to precisely align 2D geostationary (GEO) imagery with 3D CloudSat vertical profiles. The key steps included:

Co-location: We identified all instances where a CloudSat satellite overpass occurred within the field of view of a GEO satellite (GOES, MSG, Himawari) within a 5-minute time window.
Alignment: The narrow 1D vertical profiles from CloudSat were spatially mapped to their nearest corresponding 2D GEO satellite pixel.
Patch Extraction: 256x256 pixel patches were extracted from the GEO imagery, centered around the aligned CloudSat track.
Quality Filtering: Patches were filtered to ensure data quality, including a requirement that at least 25% of the vertical columns in a patch contained clouds.
TC Dataset Creation: A dedicated benchmark dataset was created by intersecting the aligned pairs with the International Best Track Archive for Climate Stewardship (IBTrACS) database, isolating overpasses of tropical cyclones.

License

This dataset is made available under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the data for any purpose, even commercially, as long as you give appropriate credit.

Acknowledgements

This work was enabled by the Frontier Development Lab (FDL) Earth Systems Lab, a public-private partnership between the European Space Agency (ESA), Trillium Technologies, and the University of Oxford. We acknowledge the support of the CloudSat mission team at NASA JPL for providing access to the CloudSat data products. We also thank the teams behind the GOES, Himawari, and Meteosat satellites for their invaluable data. This research was supported by computational resources from Google Cloud, Scan Computers, Nvidia Corporation, and Pasteur Labs.