Release correction — 20 August 2026: A previous dataset upload mistakenly provided a linearized 10 m NEON product as lr. This was corrected on 20 August 2026. The canonical LR files are now in s2_l2a_10m/ and contain the original Sentinel-2 observations exported through Google Earth Engine.

SEN2NEON - Dataset for Multispectral Sentinel-2 Super-Resolution

SEN2NEON is a PyTorch-based dataset and evaluation suite for the SEN2NEON multispectral super-resolution benchmark. The project aligns Sentinel-2 Level-2A reflectances with co-registered 1 m AVIRIS-NG hyperspectral imagery, creating the first large-scale reference set for the Sentinel-2 20 m bands under real reflectance conditions.

HR resolution: The 2.5 m HR product is the canonical SEN2NEON product described and evaluated in the paper. We also provide spatially aligned 1 m HR tiles because they are produced as an intermediate output of our processing workflow and may be useful to others. The 1 m product is supplementary; it does not redefine the paper benchmark or its reported results.

Highlights

• Harmonized dataset with 2,269 spatially aligned tiles (1024×1024 pixels at 2.5 m) spanning 12 Sentinel-2-equivalent bands (B1–B9, B8A, B11–B12).
• Land-cover aware metadata (centroids, fine/coarse classes) for stratified analyses.
• Ready-to-use PyTorch Dataset/DataModule that pairs observed Sentinel-2 LR (all bands on an aligned 10 m grid) with canonical 2.5 m or supplementary 1 m NEON-derived HR GeoTIFF tiles.
• Validation pipeline with visualization, histogram matching, and metric logging that integrates OpenSR-Test semantics.
• Baseline model adapters for SEN2SR variants, LDSR-S2, and SRGAN to reproduce the benchmarks from the accompanying paper.

---

Dataset at a glance

• LR product: Original Sentinel-2 Level-2A observations exported from GEE COPERNICUS/S2_SR_HARMONIZED. The 12 bands are stored as a common 10 m, 256×256 aligned stack; this convenience product does not preserve separate native 10/20/60 m grids.
• Canonical HR product (paper): AVIRIS-NG hyperspectral imagery harmonized to Sentinel-2A/B spectral response functions and provided at 2.5 m (1024×1024), excluding B10. This is the default throughout the code and the target used for the published benchmark and metrics.
• Supplementary HR product: The same aligned, harmonized data at 1 m (2560×2560). We include these tiles only because our production workflow creates them and they may support other research; they are not the HR definition evaluated in the SEN2NEON paper.
• Coverage: 2,269 tiles from 119 NEON acquisition assets (110 calendar dates, 2018–2024), spanning forests, grasslands, shrublands, croplands, water, barren land, and built-up areas.
• Metadata: Per-tile CSV/JSON metadata listing relative paths, projected and geographic centroids, and both detailed and superclass land-cover labels for stratified evaluation.

These characteristics make SEN2NEON suitable for benchmarking super-resolution models that target the Sentinel-2 20 m bands without relying on synthetic degradations.

---

1. Environment setup

Create a Python environment (≥3.10 recommended) and install the core dependencies:

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121  # pick the wheel for your system
pip install pandas numpy rasterio matplotlib tqdm huggingface_hub datasets opensr_test

Recent huggingface_hub releases use hf-xet for large-file transfers. Install GPU-specific PyTorch wheels as needed.

---

2. Download the dataset from Hugging Face

The dataset is hosted on the Hugging Face Hub at isp-uv-es/SEN2NEON. Use the helper script to keep the on-hub directory layout:

python data/download_SEN2NEON.py \
  --repo-id isp-uv-es/SEN2NEON \
  --out-dir ./data/sen2neon \
  --hr-resolution 2.5 \
  --high-performance     # optional, enables high-performance hf-xet transfers

The command above downloads the canonical 2.5 m HR product. To select the supplementary 1 m product instead, pass --hr-resolution 1; use --hr-resolution both to download both. Metadata indexes are downloaded in every case.

Key options:

• --hr-resolution {2.5,1,both}: choose the HR files to download; the default is canonical 2.5.
• --all: download every file in the repository instead of the curated subset.
• --high-performance: enable HF_XET_HIGH_PERFORMANCE for faster transfers.
• Set HF_TOKEN if you need to authenticate to private mirrors.

After the command finishes you should see the following structure:

./data/sen2neon
├── metadata.parquet            # default Hugging Face dataset index
├── metadata.jsonl              # portable equivalent index
├── metadata.csv                # CSV index used by the PyTorch loader
├── s2_l2a_10m/                 # canonical observed S2 LR, 12×256×256
├── s2_l2a_10m.sha256           # checksums for all canonical LR TIFFs
├── neon_2.5m_linearized/       # canonical paper HR, 12×1024×1024
├── neon_1m_linearized/         # supplementary workflow HR, 12×2560×2560
└── DATASET_RELEASE_NOTES.md

Only the HR directory or directories selected with --hr-resolution will be present.

The CSV, JSONL, and Parquet indexes contain the same one-row-per-tile metadata. hr remains a backward-compatible alias for the canonical 2.5 m path. Explicit fields include hr_2_5m_path, hr_1m_path, hr_available_resolutions_m, hr_canonical_resolution_m, and per-product width, height, pixel size, nodata value, transform, and status fields (hr_2_5m_* and hr_1m_*). The indexes also include exact S2 and NEON asset provenance, acquisition times, LR source/grid fields, land-cover metadata, and projected/geographic centroids. The historical neon_10m_linearized product is NEON-derived and is not the canonical LR.

Load the lightweight metadata index without downloading all GeoTIFFs:

from datasets import load_dataset

index = load_dataset("isp-uv-es/SEN2NEON", split="validation")
print(index[0]["lr"])
print(index[0]["hr"])             # canonical 2.5 m alias
print(index[0]["hr_1m_path"])     # supplementary 1 m path

The path fields remain relative paths because PIL cannot faithfully decode these 12-band GeoTIFFs. Use rasterio through the PyTorch loader below, or download individual paths with huggingface_hub.hf_hub_download.

---

3. Working with the Dataset API

Basic Dataset

from data.dataset import SEN2NEON
import torch

DATA_ROOT = "./data/sen2neon"
CSV_PATH = f"{DATA_ROOT}/metadata.csv"

ds = SEN2NEON(
    csv_path=CSV_PATH,
    root_dir=DATA_ROOT,
    hr_resolution=2.5,       # canonical paper product (default); use 1 for supplementary HR
    crop_size_lr=128,        # None keeps full tiles; value in LR pixels
    dtype=torch.float32,
    allow_nan=False,
)

sample = ds[0]
print(sample["lr"].shape, sample["hr"].shape)
print(sample["meta"])       # dict with name, lon/lat, land-cover labels

To use the supplementary product, download it and set hr_resolution=1. Full-tile shapes are [12, 256, 256] for LR and [12, 2560, 2560] for 1 m HR. With the default hr_resolution=2.5, the HR shape is [12, 1024, 1024].

• Scaling: GeoTIFFs store Sentinel-2 reflectance scaled by 10 000 (uint16). Convert to [0,1] by dividing by 10 000 before feeding a model.
• Cropping: crop_size_lr performs aligned random crops, automatically scaling the HR crop to match the LR patch.
• Metadata: Land-cover detail/superclass IDs and human-readable labels are included for stratified evaluation.

PyTorch Lightning-style DataModule

from data.datamodule import SEN2NEONDataModule

datamodule = SEN2NEONDataModule(
    csv_path=CSV_PATH,
    root_dir=DATA_ROOT,
    hr_resolution=2.5,       # canonical default; 1 selects supplementary HR
    batch_size=4,
    num_workers=8,
    crop_size_lr=128,
)

datamodule.setup(stage="predict")
predict_loader = datamodule.predict_dataloader()
for batch in predict_loader:
    lr = batch["lr"].float()
    hr = batch["hr"].float()
    break

Visualization helpers

ds.save_example(out_path="example.png", seed=123)

This saves a 2×2 panel containing RGB and random-band comparisons between LR and HR tiles.

---

4. Validation pipeline

validate.py demonstrates how to reproduce the benchmarking experiments. Reproduction of the paper results uses hr_resolution=2.5; the supplied baseline models and reported metrics target that canonical grid, not the supplementary 1 m product.

1. Load data: instantiate SEN2NEONDataModule, call setup(), and iterate over the prediction loader.
2. Select models: models/model_selector.py exposes adapters for srgan, sen2sr, lite_sen2sr, and ldsrs2, defining their input bands and prediction calls.
3. Run inference: normalize inputs to [0,1], slice the requested bands, and call the adapter’s predict function in torch.no_grad() mode.
4. Post-process: select the SR band subset, apply histogram matching to the LR reference, and ensure SR/HR tensors align.
5. Score results: metrics.validator.SREvaluator computes PSNR, SSIM, SAM, and OpenSR-Test metrics; MetricsSink logs CSV summaries while optional quick-look figures are saved to disk.

To evaluate your own model, add a new entry to models_configs describing the required input bands, the index positions of the 20 m outputs, and a callable that performs inference. The rest of the loop remains unchanged.

---

5. Benchmark performance

The paper reports the following validation metrics on the SEN2NEON evaluation split (mean ± std):

Model	PSNR ↑	SSIM ↑	SAM ↓	Refl. Consistency ↓	Spectral Consistency ↓	Spatial Consistency ↓
S2SR-LDSR-S2	31.98 ± 4.21	0.78 ± 0.12	0.05 ± 0.03	0.0087 ± 0.000	2.20 ± 1.33	0.019 ± 0.05
S2SR-Lite	32.56 ± 3.28	0.80 ± 0.10	0.04 ± 0.02	0.0054 ± 0.000	1.52 ± 0.91	0.004 ± 0.02
S2SR-Mamba	32.44 ± 4.18	0.81 ± 0.10	0.05 ± 0.03	0.0081 ± 0.000	2.20 ± 1.25	0.006 ± 0.04
SRGAN	31.40 ± 3.85	0.73 ± 0.13	0.05 ± 0.03	0.0067 ± 0.000	1.21 ± 0.79	0.020 ± 0.03

• Frequency-transfer approaches (SEN2SR variants) deliver the best balance between sharpness and spectral fidelity.
• Diffusion-based S2SR-LD introduces additional detail but at the cost of hallucinations, while SRGAN favors perceptual sharpness with weaker physical consistency.

---

6. Land-cover-aware analysis

Each metadata row carries detailed (LC_detail_*) and superclass (LC_superclass_*) labels derived from a categorical land-cover raster. Use them to filter tiles, compute stratified metrics, or balance batches across ecosystems. Missing values are stored as empty strings/NaNs and can be dropped or imputed downstream.

---

7. Roadmap

• Release dataset & dataloaders
• Integrate land-cover annotations
• Provide visualization utilities
• Add SEN2SR / LDSR-S2 / SRGAN adapters
• Publish benchmarking pipeline and metrics integration

---

Citation

If you use SEN2NEON in your research, please cite:

@article{article,
author = {Donike, Simon and Aybar, Cesar and Contreras, Julio and Gómez-Chova, Luis},
year = {2026},
month = {01},
pages = {6013905-6013905},
title = {SEN2NEON: Enabling Quantitative Benchmarking of Sentinel-2 Superresolution for All Multispectral Bands},
volume = {23},
journal = {IEEE Geoscience and Remote Sensing Letters},
doi = {10.1109/LGRS.2026.3703947}
}```