Multi-Modal Foundation · isp-uv-es
SuperIX
Enhances Sentinel-2 satellite imagery to finer resolutions and evaluates super-resolution algorithms for researchers.
What SuperIX is
SuperIX is an explainable AI super-resolution initiative and framework developed by isp-uv-es for the European Space Agency Sentinel-2 mission. Super-resolution methods are increasingly applied to 10-metre Sentinel-2 satellite data to improve spatial clarity for downstream Earth science and climate applications, such as crop delineation and road detection. However, machine learning super-resolution can alter surface reflectance or invent ungrounded visual artefacts, raising questions regarding whether enhancements offer analytical value or merely aesthetic appeal.
SuperIX provides an intercomparison framework to benchmark how conservatively super-resolution algorithms preserve original radiometric and spectral integrity while generating finer spatial details. Accompanied by interpretable uncertainty assessments, the project benchmarks model outputs across standardised open-source reference datasets to determine whether reconstructed high-frequency elements reflect true surface structures or unphysical distortions.
How it works
The evaluation framework operates on x2 and x4 spatial scaling factors using Sentinel-2 Level-1C and Level-2A imagery paired with high-resolution reference datasets. To evaluate radiometric consistency, the super-resolved output is downsampled back to low resolution using a triangular anti-aliasing filter and bilinear interpolation. It is then measured against the original input using Mean Absolute Error (MAE) for surface reflectance, Spectral Angle Distance (SAM) for spectral curve fidelity, and Phase Correlation Coefficient (PCC) for spatial shift detection.
Reconstructed high-frequency details are evaluated and decomposed into three normalised scores that sum to one: improvements (similarity to the high-resolution reference), omissions (similarity to the low-resolution baseline), and hallucinations (deviation from both reference sources). These scores are calculated using MAE on small pixel patches (2x2 for x2 scale, 4x4 for x4 scale) to evaluate intensity differences, or Learned Perceptual Image Patch Similarity (LPIPS) across 32x32 patches to assess structural differences.
Key facts
- Target imagery
- ESA Sentinel-2 (Level-1C and Level-2A)
- Scale factors
- x2 and x4
- Supported formats
- ONNX, GeoTIFF, tf-keras
- Benchmark dataset
- isp-uv-es/opensr-test
- Primary metrics
- MAE (reflectance), SAM (spectral), PCC (spatial), LPIPS
What you can use it for
- Sentinel-2 Spatial Resolution Enhancement
- Upscaling 10-metre multispectral Sentinel-2 imagery to higher spatial resolutions to aid agricultural field boundary mapping, road extraction, and object recognition.
- Algorithm Intercomparison and Benchmarking
- Systematically comparing the behaviour, accuracy, and detail synthesis of open-source and closed-source satellite super-resolution models on standardised test scenes.
- Reflectance and Spectral Integrity Verification
- Assessing whether super-resolution processing introduces radiometric bias or shifts multispectral signatures away from ground observations.
- Hallucination and Artefact Auditing
- Quantifying the proportion of high-frequency spatial features introduced by a model that do not exist in high-resolution ground-truth imagery.
Run it here
We list this model but cannot run it yet: no loader for library: tf-keras.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Separates evaluation of high-frequency synthesis into explicit, complementary metrics for improvements, omissions, and hallucinations.
- Combines intensity-sensitive MAE metrics with structural LPIPS perceptual assessments.
- Measures physical preservation using Spectral Angle Distance and downsampled Mean Absolute Error rather than relying purely on visual inspection.
- Validates performance across diverse cross-sensor reference data covering urban, agricultural, and forest landscapes.
Watch out for
- Evaluation outcomes vary across reference datasets due to cross-sensor differences in radiometry, spectral response, and spatial alignment.
- Models designed specifically for x4 scaling must be evaluated at x2 magnification via artificial downsampling by a factor of two.
- Evaluation is restricted to specific patch dimensions, namely 32x32 patches for LPIPS and 2x2 or 4x4 patches for MAE.
Inputs and outputs
You provide
Image
PNG, JPEG or TIFF. Sent to the model as-is.
You get back
The returned image.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use SuperIX commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What reference datasets are used to validate model outputs?
Validation uses the OpenSR-test suite, which includes NAIP aerial orthophotos of US agriculture and forests, SPOT images from Worldstrat, Spanish urban and agricultural orthophotos, and SEN2VENµS imagery.
How are hallucinations defined and calculated?
Hallucinations are quantified using a dedicated score (ha_score) that measures when high-frequency details generated in the super-resolved output deviate significantly from both the low-resolution input and the high-resolution reference.
What formats can be submitted for closed-source models?
Closed-source model participants submit their super-resolved results in GeoTIFF format matching the exact spatial resolution of the high-resolution reference image, accompanied by a metadata JSON file.
How does SuperIX verify that low-resolution reflectance was preserved?
The super-resolved image is downsampled to native resolution using a triangular anti-aliasing filter and bilinear interpolation, followed by Mean Absolute Error measurement against the original input.
Written from SuperIX's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.