We’re excited to announce the release of OpenSR-Degradation, our open-source toolset for creating synthetic Sentinel-2-like datasets from high-resolution NAIP imagery! Designed to support robust super-resolution research and training, this framework enables users to transform NAIP images into S2-like (Sentinel-2 style) data using three complementary degradation models.

Why Synthetic S2-like Data?

Access to high-resolution and temporally consistent satellite data is limited. By generating synthetic Sentinel-2 imagery from publicly available high-res NAIP data, we can simulate real-world sensor characteristics and enable cross-sensor super-resolution training at scale—without relying solely on sparse real HR-LR image pairs.

What’s Under the Hood?

OpenSR-Degradation offers three unique degradation pipelines:

• Statistical Model
  Learns gamma correction curves per band using multivariate Gaussian distributions—ideal for statistical harmonization across spectral channels.

• Deterministic Model
  Combines bilinear downsampling with a U-Net (EfficientNet-B0 backbone) and histogram correction to simulate Sentinel-2 reflectance accurately.

• Variational Model
  A VAE-based histogram matcher that transforms NAIP spectral histograms into Sentinel-2-like distributions using learned priors.

Each method outputs harmonized S2-like images (NAIPhat), making them suitable for training and benchmarking super-resolution models with consistent spectral properties.

Ready to Try?

🔗 GitHub: github.com/ESAOpenSR/opensr-degradation
📦 PyPI: opensr-degradation
📖 Docs: esaopensr.github.io/opensr-degradation