4x Satellite Image Super-Resolution
This dataset contains 720 pairs of satellite images aligned at high (0.5 m/pixel) and low (2 m/pixel) resolutions, organized into GeoTIFF files. It is optimized for training 4X super-resolution models in remote sensing, mapping, or geospatial imaging contexts. Each pair is geographically synchronized, allowing for accurate learning.
Description
4x Satellite Image Super-Resolution is a data set dedicated to spatial super-resolution tasks. It offers pairs of high-resolution (0.5 m/pixel) and low-resolution (2 m/pixel) satellite images, rigorously aligned. The images cover varied areas (urban, rural, natural) with an organization that allows efficient processing.
What is this dataset for?
- Training super-resolution models to improve the quality of satellite images
- Test CNN, GAN, or Transformers architectures in 4x upscaling
- Develop tools for mapping or improved geospatial analysis
Can it be enriched or improved?
Yes. It is possible to add other geographic regions, to vary the seasons or to label the types of terrain. Additional annotations (e.g. soil type, building) would also make it possible to use the data for complex supervised tasks.
🔎 In summary
🧠 Recommended for
- Geospatial engineers
- Imaging researchers
- Satellite AR developers
🔧 Compatible tools
- Rasterio
- PyTorch
- TensorFlow
- DeepSR
- GDAL
💡 Tip
To avoid overfitting, integrate a data augmentation based on the change in brightness and random rotation.
Frequently Asked Questions
Does this dataset contain terrain or class annotations?
No, only pairs of high and low resolution images are provided, without segmentation labels.
Can this dataset be used in commercial or open-source projects?
Yes, the Apache 2.0 license allows free use, including in commercial projects.
What image format is used and what are the resolutions?
The images are in GeoTIFF format, with resolutions of 2 m/pixel for LR and 0.5 m/pixel for HR.
