Earth Observation · isp-uv-es
opensr-test
A benchmark dataset and evaluation suite for testing real-world Sentinel-2 satellite imagery super-resolution.
What opensr-test is
opensr-test is an Earth observation benchmark dataset designed to evaluate super-resolution algorithms on real-world satellite imagery. Developed by the Image and Signal Processing group (isp-uv-es) with support from the European Space Agency, it provides paired low-resolution Sentinel-2 scenes and high-resolution aerial and satellite captures across varied terrestrial environments.
The benchmark addresses limitations in standard super-resolution research, where models are frequently assessed on artificially degraded imagery using image quality metrics such as PSNR, SSIM, and LPIPS. These traditional evaluation methods often fail in real-world remote sensing conditions involving sensor misalignment or differences in atmospheric luminance. opensr-test provides verified alignment data and targeted evaluation criteria covering consistency, synthesis, and correctness.
How it was made
The benchmark contains five distinct sub-datasets: NAIP, SPOT, Venµs, SPAIN CROPS, and SPAIN URBAN. Each sample pairs 12-band Sentinel-2 Level-1C top-of-atmosphere or Level-2A bottom-of-atmosphere data with high-resolution visible and near-infrared (RGBNIR) imagery at x2 or x4 magnification scales. High-resolution patch sizes range from 256x256 to 512x512 pixels, matching spatial resolutions of 2.5 metres or 5 metres.
Each entry is organised as a dictionary containing raw high-resolution images, high-resolution images harmonised to the spectral response of Sentinel-2 Level-2A bands, and full Level-1C and Level-2A spectral stacks. Metadata tables provide spatial misalignment scores measured in 10-metre pixels, spectral angle differences, L1 reflectance variation scores, coordinate reference systems, and affine transformation matrices.
Key facts
- Target satellite
- Sentinel-2 (L1C and L2A products)
- Scale factors
- x2 and x4
- Sub-datasets
- 5 (NAIP, SPOT, Venµs, SPAIN CROPS, SPAIN URBAN)
- Reference ground resolutions
- 2.5 m and 5 m
- Reference spectral coverage
- RGBNIR (Visible and Near-Infrared)
- Reference patch dimensions
- 256x256, 484x484, and 512x512 pixels
What you can use it for
- Benchmarking real-world satellite super-resolution
- Assessing how well super-resolution architectures upscale real Sentinel-2 Level-1C and Level-2A imagery against genuine high-resolution aerial and satellite targets.
- Spectral and radiometric consistency verification
- Evaluating whether algorithmically generated fine details maintain reflectance accuracy and spectral signatures relative to lower-resolution satellite bands.
- Land cover super-resolution analysis
- Testing algorithmic reconstruction quality across varied target environments, including agricultural fields, forests, bare soil, urban areas, and road networks.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Pairs actual low-resolution satellite passes with genuine high-resolution captures rather than synthetic bicubic downsamplings.
- Supplies both raw reference imagery and harmonised versions adjusted to Sentinel-2 Level-2A spectral reflectance.
- Provides sample-level metadata quantifying spatial misalignment, spectral angle distance, and radiometric reflectance deviation.
- Spans multiple geographic sources and land-cover categories across Europe and North America.
Watch out for
- Reference target data is restricted to four RGBNIR spectral bands, leaving eight Sentinel-2 bands without high-resolution ground truth.
- High-resolution reference imagery is constrained to small patch extents between 256x256 and 512x512 pixels.
Questions
Can I look inside opensr-test without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is opensr-test under?
Its authors declare mit. Read the terms on the original page before reusing the data.
Which Sentinel-2 spectral bands are included in the benchmark?
Both Level-1C and Level-2A data include 12 bands: coastal aerosol (B01), blue (B02), green (B03), red (B04), red edge (B05, B06, B07), near-infrared (B08, B8A), water vapour (B09), cirrus (B10, L1C only), and short-wave infrared (B11, B12).
What is the distinction between HR and HRharm in the dataset?
HR denotes the raw high-resolution reference image in visible and near-infrared bands, whereas HRharm provides that same imagery radiometrically harmonised to match Sentinel-2 Level-2A surface reflectance.
How are spatial misalignments recorded?
The dataset records spatial misalignment values within the metadata table for each patch, specified in terms of low-resolution 10-metre Sentinel-2 pixels, alongside coordinate reference systems and affine transformations.
How to cite it
Aybar, Cesar and Montero, David and Donike, Simon and Kalaitzis, Freddie and Gómez-Chova, Luis. A Comprehensive Benchmark for Optical Remote Sensing Image Super-Resolution. Authorea Preprints, 2024.
Written from opensr-test's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.