SEN2NEON: Enabling quantitative benchmarking of Sentinel-2 super-resolution for all multispectral bands

We’re pleased to share that our work, “SEN2NEON: Enabling Quantitative Benchmarking of Sentinel-2 Super-Resolution for all Multi-Spectral bands,” has been published open access by IEEE.

Sentinel-2 super-resolution has become an important research direction for Earth observation, but the field has long faced a fundamental limitation: for the 20 m and 60 m bands, there is no straightforward high-resolution ground truth. As a result, many studies rely on synthetic degradation protocols, where low-resolution inputs are generated by downsampling higher-resolution data. While useful, these protocols do not fully capture the radiometric and spatial characteristics of real Sentinel-2 observations.

SEN2NEON addresses this gap by introducing a physically grounded benchmark dataset for Sentinel-2 super-resolution. It pairs real Sentinel-2 Level-2A reflectances with high-resolution reference data derived from AVIRIS-NG hyperspectral imagery, harmonized through sensor-aware spectral response convolution and radiometric correction.

What we built

• SEN2NEON benchmark dataset: A cross-sensor dataset designed for quantitative Sentinel-2 super-resolution evaluation across all multispectral bands, excluding the cirrus band.
• Real Sentinel-2 LR inputs: Instead of generating low-resolution inputs synthetically, SEN2NEON preserves the original Sentinel-2 observations as the LR data.
• High-resolution AVIRIS-NG references: Hyperspectral AVIRIS-NG imagery is transformed into Sentinel-2-like reflectances at 2.5 m reference resolution.
• Sensor-aware harmonization: We align AVIRIS-NG and Sentinel-2 through spectral response convolution and radiometric correction, producing physically meaningful LR–HR pairs.
• Cross-dataset consistency analysis: We quantify harmonization quality using LR–HR resampling consistency metrics, enabling a more rigorous assessment of alignment between the real Sentinel-2 measurements and the high-resolution reference.

Results at a glance

SEN2NEON provides physically aligned low- and high-resolution pairs for Sentinel-2 bands at their native S2 resolution and at 2.5 m reference resolution.

Compared to existing cross-sensor datasets, it offers:

• Broader spectral coverage across Sentinel-2 multispectral bands.
• A finer effective reference resolution.
• Stronger consistency between LR Sentinel-2 observations and HR reference data.
• A reproducible benchmark for evaluating super-resolution models under realistic sensing conditions.

By avoiding purely synthetic degradation, SEN2NEON enables a more meaningful evaluation of whether super-resolution models recover physically plausible high-resolution reflectance patterns rather than merely optimizing for artificial downsampling assumptions.

Why it matters

Super-resolution for Sentinel-2 is widely used in land-cover mapping, vegetation monitoring, water analysis, urban studies, and environmental change detection. However, without reliable high-resolution references, it is difficult to compare methods fairly or assess whether improvements are physically meaningful.

SEN2NEON helps address this by providing a benchmark that is:

• Quantitative: Models can be evaluated against high-resolution references rather than only visually inspected or tested under synthetic degradation.
• Physically grounded: The LR input is real Sentinel-2 data, while the HR reference is derived through sensor-aware harmonization.
• Multispectral: The dataset supports benchmarking beyond RGB and 10 m bands, extending evaluation to the full Sentinel-2 multispectral setting, except for the cirrus band.
• Transferable: The harmonization strategy behind SEN2NEON can be adapted to other multispectral satellite sensors, supporting broader development of physically consistent super-resolution benchmarks.

Toward better Sentinel-2 super-resolution evaluation

By combining real Sentinel-2 observations with harmonized high-resolution hyperspectral references, SEN2NEON provides a foundation for developing and comparing models that are not only visually convincing, but also radiometrically and spatially consistent.

🔗 Read the full article: SEN2NEON: Enabling Quantitative Benchmarking of Sentinel-2 Super-Resolution for all Multi-Spectral bands — IEEE Open Access

🔗 Access the Dataset: Github.com/SEN2NEON