[Paper Review] Open High-Resolution Satellite Imagery: The WorldStrat Dataset -- With Application to Super-Resolution

Julien Cornebise, Ivan Oršolić+1 more|arXiv (Cornell University)|Jul 13, 2022

Advanced Image Fusion Techniques17 citations

TL;DR

This paper introduces the WorldStrat dataset, a large-scale, open-access collection of high-resolution (1.5 m/pixel) Airbus SPOT 6/7 satellite imagery paired with multi-temporal, lower-resolution (10 m/pixel) Sentinel-2 images, stratified across diverse global land uses including underrepresented humanitarian sites. The dataset enables training of super-resolution models to derive high-fidelity imagery from free low-resolution data, with open-source tools and code for reproducibility and community extension.

ABSTRACT

Analyzing the planet at scale with satellite imagery and machine learning is a dream that has been constantly hindered by the cost of difficult-to-access highly-representative high-resolution imagery. To remediate this, we introduce here the WorldStrat dataset. The largest and most varied such publicly available dataset, at Airbus SPOT 6/7 satellites' high resolution of up to 1.5 m/pixel, empowered by European Space Agency's Phi-Lab as part of the ESA-funded QueryPlanet project, we curate nearly 10,000 sqkm of unique locations to ensure stratified representation of all types of land-use across the world: from agriculture to ice caps, from forests to multiple urbanization densities. We also enrich those with locations typically under-represented in ML datasets: sites of humanitarian interest, illegal mining sites, and settlements of persons at risk. We temporally-match each high-resolution image with multiple low-resolution images from the freely accessible lower-resolution Sentinel-2 satellites at 10 m/pixel. We accompany this dataset with an open-source Python package to: rebuild or extend the WorldStrat dataset, train and infer baseline algorithms, and learn with abundant tutorials, all compatible with the popular EO-learn toolbox. We hereby hope to foster broad-spectrum applications of ML to satellite imagery, and possibly develop from free public low-resolution Sentinel2 imagery the same power of analysis allowed by costly private high-resolution imagery. We illustrate this specific point by training and releasing several highly compute-efficient baselines on the task of Multi-Frame Super-Resolution. High-resolution Airbus imagery is CC BY-NC, while the labels and Sentinel2 imagery are CC BY, and the source code and pre-trained models under BSD. The dataset is available at https://zenodo.org/record/6810791 and the software package at https://github.com/worldstrat/worldstrat .

Motivation & Objective

• To address the scarcity of open, high-resolution, and geographically stratified satellite imagery for machine learning research.
• To enable super-resolution tasks by pairing high-resolution SPOT 6/7 imagery with multiple low-resolution Sentinel-2 revisits.
• To support broad-spectrum applications in remote sensing by providing a publicly accessible, diverse, and well-structured dataset.
• To lower barriers to entry in satellite image analysis by offering open-source tools and pre-trained models compatible with EO-learn.
• To foster community-driven data enrichment and reproducible research through versioned, open licenses and transparent data curation.

Proposed method

• Curated nearly 10,000 km² of unique high-resolution SPOT 6/7 imagery across 23 LCCS land use classes and 6 IPCC categories, ensuring global geographic and land-use diversity.
• Stratified sampling based on land use and urban density (using GHSL-SMOD), with inclusion of underrepresented sites such as conflict zones and informal settlements.
• Paired each high-resolution SPOT image with multiple temporally aligned Sentinel-2 images (10 m/pixel) from the same time window to enable multi-frame super-resolution.
• Developed an open-source Python package with EO-learn integration to rebuild, extend, and train on the dataset, including tutorials and baseline models.
• Implemented and released multiple compute-efficient super-resolution baselines (e.g., Multi-Frame SRCNN, HighResNet) trained on the dataset.
• Applied versioned DOI hosting on Zenodo to ensure persistent, traceable, and updatable dataset releases with community contribution pathways.

Experimental results

Research questions

• RQ1 Can a large-scale, open, and geographically stratified high-resolution satellite imagery dataset be created to support diverse machine learning applications?
• RQ2 How effective is multi-temporal, low-resolution Sentinel-2 data in enabling super-resolution when paired with high-resolution SPOT 6/7 imagery?
• RQ3 To what extent can open-source tools and community contributions enhance the scalability and utility of satellite image datasets?
• RQ4 What are the performance characteristics of super-resolution models trained on this paired, stratified dataset compared to standard benchmarks?
• RQ5 How can open licensing and versioned data distribution promote reproducibility and long-term sustainability in Earth observation research?

Key findings

• The WorldStrat dataset comprises nearly 10,000 km² of high-resolution satellite imagery, covering 23 LCCS land use classes and 6 IPCC categories with stratified global representation.
• The dataset includes 1,150 unique high-resolution SPOT 6/7 scenes paired with multiple Sentinel-2 revisits, enabling multi-frame super-resolution training.
• The authors trained and released multiple super-resolution baselines (e.g., Multi-Frame SRCNN, HighResNet) that achieve competitive PSNR and SSIM scores on the validation set.
• Statistical analysis revealed high variability in model performance across the validation set, indicating that mean-based metrics alone are insufficient for benchmarking.
• The dataset is hosted on Zenodo with versioned DOIs, ensuring long-term persistence and traceability, with plans to support community contributions and corrections.
• Licensing model enables broad reuse: high-res imagery under CC-BY-NC, Sentinel-2 and labels under CC-BY, and code/models under BSD, maximizing accessibility and integration.