SEN2NAIP v2.0 Released — A Major Boost for Sentinel-2 Super-Resolution

We’re excited to share the release of SEN2NAIP v2.0, a powerful and extensive dataset developed to support super-resolution (SR) research for Sentinel-2 imagery. Designed by our team and now hosted on Hugging Face, SEN2NAIP provides the foundation for both reference-based SR and synthetic training setups, enabling robust benchmarking and model development across varied spatial and spectral domains.

What’s Inside SEN2NAIP v2.0?

The dataset consists of two core components:

  1. Cross-Sensor Dataset

    • 2,851 image pairs pairing Sentinel-2 L2A imagery (low-resolution) with high-resolution NAIP orthophotos.

    • Each pair covers ~1.46 km², offering rich variability across landscape types.

    • Ideal for training and validating SR models in real-world remote sensing tasks.

  2. Synthetic Dataset

    • 17,657+ NAIP-S2like pairs, created by applying a trained degradation model that simulates Sentinel-2 characteristics from NAIP imagery.

    • Multiple variants using VAE and gamma-based histogram matching allow for comprehensive SR evaluation under different degradation assumptions.

The dataset also includes temporal support, cloud-optimized access, and updated harmonization models. You can explore and download it here on Hugging Face or use the ready-to-run Colab examples.

Why It Matters

SEN2NAIP addresses a critical bottleneck in SR research: the lack of realistic, high-quality, and diverse training data for satellite image enhancement. By bridging real and synthetic data pipelines with a unified framework, the dataset enables the development of robust SR models that generalize well across geographies, seasons, and sensors.

Whether you’re developing GANs, diffusion-based SR pipelines, or transformer-based upscaling networks, SEN2NAIP offers the benchmark data you need.

Dataset Usage Example

				
					import matplotlib.pyplot as plt
import numpy as np
import tacoreader

# Load the dataset
dataset = tacoreader.load("tacofoundation:sen2naipv2-unet")

# Filter the dataset for a subset of classes in "rai:admin2"
top_classes = dataset["rai:admin1"].value_counts().index[:10]  # Select top 5 most frequent classes
filtered_dataset = dataset[dataset["rai:admin1"].isin(top_classes)]

# Create subplots
fig, axes = plt.subplots(2, 2, figsize=(16, 12))

# Plot "rai:ele" (elevation)
dataset.plot("rai:ele", cmap="viridis", legend=True, ax=axes[0, 0])
axes[0, 0].set_title("Elevation (rai:ele)")

# Plot "rai:cisi" (log-scaled) on the fly
dataset.assign(rai_cisi_log=lambda df: np.log1p(df["rai:cisi"])).plot(
    "rai_cisi_log", cmap="viridis", legend=True, ax=axes[0, 1]
)
axes[0, 1].set_title("CISI (Log-Scaled, rai:cisi)")

# Plot "rai:admin0" (categorical) with a categorical colormap
dataset.plot("rai:admin0", cmap="tab20", legend=True, ax=axes[1, 0])
axes[1, 0].set_title("Administrative Boundaries (Level 0) (rai:admin0)")

# Plot filtered "rai:admin2" (categorical) with a limited number of classes
filtered_dataset.plot("rai:admin1", cmap="tab20", legend=True, ax=axes[1, 1])
axes[1, 1].set_title("Administrative Boundaries (Level 1, Filtered) (rai:admin1)")

# Adjust layout for better readability
plt.tight_layout()
plt.show()
				
			

Recent Posts

No-Code SR Demo is now live!

This demo, aimed at non-technical users, allows you to enter your coordinates and create a super-resolution product on your custom Sentinel-2 acquisition. Immediately judge wether SR can be useful for you application!

OpenSR Team @Living Planet Symposium

The OpenSR team joined ESA’s Living Planet Symposium 2025 to present our latest advances in Sentinel-2 super-resolution, dataset standards, and workflows. From latent diffusion models to FAIR-compliant data access with TACO, our tools aim to make high-resolution Earth observation more accessible and actionable.

New Release: OpenSR-UseCases Package

A lightweight validation toolkit to benchmark segmentation performance across low-, super-, and high-resolution imagery. Quantifies how well super-resolution models improve object detection and segmentation accuracy in real-world tasks. Ideal for researchers who want to go beyond visual inspection and measure actual downstream performance gains.

New Preprint: A Radiometrically and Spatially Consistent Super-Resolution Framework for Sentinel-2

We’ve published a new preprint presenting SEN2SR, a deep learning framework for super-resolving Sentinel-2 imagery with radiometric and spatial fidelity. The model leverages harmonized synthetic data, hard constraints, and xAI tools to achieve artifact-free enhancements at 2.5 m resolution.

RGB-NIR Latent Diffusion Super-Resolution Model Released!

Our Latent diffusion model, including weights, for the RGB-NIR bands of Sentinel-2 has been released.

New Publication: LDSR-S2 Model Paper

Our diffusion-based super-resolution model for Sentinel-2 imagery has been published in IEEE JSTARS! The open-access paper introduces a latent diffusion approach with pixelwise uncertainty maps—pushing the boundaries of trustworthy generative modeling in Earth observation.

SEN2NAIP v2.0 Released — A Major Boost for Sentinel-2 Super-Resolution

We’ve released SEN2NAIP v2.0, a large-scale dataset designed for training and validating super-resolution models on Sentinel-2 imagery. The dataset includes thousands of real and synthetic HR-LR image pairs, making it a cornerstone for future SR research in Earth Observation.

New Publication: SEN2NAIP published in ‘Scientific Data’

The dataset paper has been published in 'Scientific Data'.

The OpenSR team contributes to Flood Mapping for the Valencian Flash Floods

Our team at the University of Valencia has released an interactive satellite flood map of the recent Valencia flash floods, using Landsat-8 and Sentinel-2 imagery combined with a machine learning segmentation model. Leveraging super-resolution techniques, we enhanced Sentinel-2 data to 2.5m resolution, enabling more precise flood extent mapping for post-disaster analysis.

OpenSR-Utils Preview Released: A package to handle patching, tiling and overlapping for SR Products

We’ve released a preview of OpenSR-Utils, a Python package to apply super-resolution models on raw Sentinel-2 imagery. With multi-GPU support, georeferenced output, and automatic patching, it’s a practical toolkit for real-world remote sensing pipelines.

SUPERIX: Intercomparison Excercise

Presenting SUPERIX: a community-driven benchmark to rigorously compare super-resolution models for Sentinel-2 data. Using real-world datasets and tailored metrics, SUPERIX aims to uncover the true impact of SR techniques on remote sensing accuracy.

Team attends ESA SUREDOS Workshop in Frascati

Our team attended the ESA SUREDOS Workshop to discuss the role of super-resolution in enhancing Earth Observation data. The event explored cutting-edge deep learning techniques and the importance of reliable, domain-specific validation for scientific and operational EO applications.

New Publication: OpenSR-Test theoretical framework has been published

OpenSR-test is now published in IEEE GRSL! Our new paper introduces a rigorous benchmark for evaluating super-resolution in remote sensing with real-world datasets and meaningful metrics.

OpenSR-Test Framework and Datasets Released

Our framework to validate supre-resolution results is now published. It can take any SR model and create sophisticated validation metrics over mutliple datasets, enhancing the comparability of methodologies.

OpenSR-Degradation Released: Package to create Syntehtic Training Data

We’ve released OpenSR-Degradation, a toolkit to generate synthetic Sentinel-2-like imagery from NAIP using statistical, deterministic, and variational models. This open-source pipeline enables large-scale training and benchmarking for cross-sensor super-resolution.