SEN2NEON

LR/HR Multispectral Super-Resolution Dataset for benchmarking and research.

---

Dataset Summary

SEN2NEON provides paired low-resolution (LR, 10 m) and high-resolution (HR) GeoTIFF tiles for validating super-resolution (SR) models in remote sensing. The LR product contains observed Sentinel-2 Level-2A reflectances; the HR references are derived from NEON AVIRIS-NG hyperspectral acquisitions. Each product shares the same spatial footprint and is pixel-aligned. Companion Parquet/JSONL/CSV indexes provide paths for both HR resolutions plus per-tile provenance, acquisition timing, geometry, cloud/nodata statistics, and land-cover labels for stratified evaluation (e.g., Forest vs Built-up).

The code, examples, and validation workflows can be found at ESAOpenSR/SEN2NEON.

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. The incorrect product is excluded from this release. The quantitative LR/HR consistency values reported in the paper were calculated using the real Sentinel-2 values.

HR resolution: The 2.5 m HR product is the canonical SEN2NEON reference described and evaluated in the paper. It remains the default in the metadata and companion code. We additionally provide aligned 1 m HR tiles because they are produced by our processing workflow and may be useful to others. The 1 m product is supplementary and does not redefine the published benchmark or its reported results.

• Modality: 12-band multispectral GeoTIFFs
• Tasks: Super-resolution (image-to-image), benchmarking, and analysis
• Scale: 2,269 aligned tile IDs with one LR and two HR products
• Alignment: Canonical 2.5 m HR is 4× LR; supplementary 1 m HR is 10× LR; both are integer, isotropic, and pixel-aligned
• Geo: Projected UTM zones per tile; WGS84 centroids are included in the metadata
• Split: One validation split; this is a benchmark, not a globally representative training corpus

---

Repository Layout

.
├── README.md
├── DATASET_RELEASE_NOTES.md
├── assets/
│   └── sen2neon_banner.png
├── metadata.parquet             # default HF Dataset/Viewer index
├── metadata.jsonl               # equivalent line-delimited JSON index
├── metadata.csv                 # equivalent tabular index
├── s2_l2a_10m.sha256            # SHA-256 manifest for LR tiles
├── s2_l2a_10m/                  # observed Sentinel-2 LR, 12×256×256
├── neon_2.5m_linearized/        # canonical paper HR, 12×1024×1024
└── neon_1m_linearized/          # supplementary workflow HR, 12×2560×2560

Relative paths in all metadata indexes match the on-Hub layout. The Parquet index is the default source for the Hugging Face Dataset Viewer and load_dataset; JSONL and CSV are retained as portable equivalents.

---

Record Schema

Each record describes one aligned tile and its LR/canonical/supplementary HR paths. Core fields are:

{
  "id": "<stem>",
  "name": "<filename.tif>",
  "split": "validation",
  "lr": "s2_l2a_10m/<file>.tif",
  "hr": "neon_2.5m_linearized/<file>.tif",
  "hr_2_5m_path": "neon_2.5m_linearized/<file>.tif",
  "hr_1m_path": "neon_1m_linearized/<file>.tif",
  "hr_available_resolutions_m": [2.5, 1.0],
  "hr_canonical_resolution_m": 2.5,
  "bands": ["B1", "B2", "B3", "B4", "B5", "B6", "B7", "B8", "B8A", "B9", "B11", "B12"],
  "lon": 0.0,
  "lat": 0.0,
  "LC_detail_id": 0,
  "LC_detail_text": "<class>",
  "LC_superclass_id": 0,
  "LC_superclass_text": "<class>"
}

Additional fields record the exact Sentinel-2 and NEON asset identifiers, acquisition times, temporal separation, cloud score, CRS, reflectance scaling, and the native Sentinel-2 resolution associated with each band. Per-product hr_2_5m_* and hr_1m_* fields provide raster dimensions, pixel size, nodata value, affine transform, and canonical/supplementary status. The legacy hr and generic hr_* fields remain aliases for the canonical 2.5 m product for backward compatibility.

Notes:

• split is set to "validation" for this release.
• lon/lat are WGS84 centroids. The enriched index also provides the explicit aliases centroid_lon/centroid_lat.
• Land-cover fields come from the categorical land-cover raster used during preprocessing.
• All raster products contain reflectance stored as uint16 and scaled by 10,000.

---

Quick Download

Using the Hub snapshot cache (resumable, selective patterns):

pip install -U huggingface_hub
python - <<'PY'
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="isp-uv-es/SEN2NEON",
    repo_type="dataset",
    local_dir="./data/sen2neon",
    allow_patterns=[
        "README.md",
        "DATASET_RELEASE_NOTES.md",
        "metadata.*",
        "s2_l2a_10m.sha256",
        "s2_l2a_10m/**",
        "neon_2.5m_linearized/**",
    ],
)
PY

This example downloads the canonical 2.5 m product. Replace neon_2.5m_linearized/** with neon_1m_linearized/** for only the supplementary 1 m product, or include both patterns to download both.

---

Load with 🤗 Datasets

The default configuration loads the metadata index without downloading all imagery:

from datasets import load_dataset

ds = load_dataset("isp-uv-es/SEN2NEON", split="validation")
row = ds[0]
print(row["id"], row["lr"])
print(row["hr"])             # canonical 2.5 m alias
print(row["hr_2_5m_path"])   # explicit canonical path
print(row["hr_1m_path"])     # supplementary path

The LR and HR path columns intentionally remain relative path strings. Hugging Face's standard Image decoder relies on PIL, which cannot faithfully decode these 12-band GeoTIFFs. Download individual pairs lazily and read them with rasterio:

from huggingface_hub import hf_hub_download
import rasterio

lr_path = hf_hub_download(
    repo_id="isp-uv-es/SEN2NEON",
    filename=row["lr"],
    repo_type="dataset",
)
hr_key = "hr"  # use "hr_1m_path" for the supplementary 1 m product
hr_path = hf_hub_download(
    repo_id="isp-uv-es/SEN2NEON",
    filename=row[hr_key],
    repo_type="dataset",
)

with rasterio.open(lr_path) as src:
    lr = src.read()  # (12, 256, 256)
with rasterio.open(hr_path) as src:
    hr = src.read()  # (12, 1024, 1024) by default; 2560×2560 for 1 m

---

Load with PyTorch (Local Files)

After downloading into ./data/sen2neon/, use the CSV-driven loader from the companion code repository:

from data.dataset import SEN2NEON
from torch.utils.data import DataLoader

root = "./data/sen2neon"
csv_path = f"{root}/metadata.csv"

ds = SEN2NEON(
    csv_path=csv_path,
    root_dir=root,
    hr_resolution=2.5,  # canonical default; use 1 for supplementary HR
    crop_size_lr=None,
)
loader = DataLoader(ds, batch_size=2, shuffle=True, num_workers=4, pin_memory=True)

batch = next(iter(loader))
lr, hr, meta = batch["lr"], batch["hr"], batch["meta"]
print(lr.shape, hr.shape)

---

Land-cover Integration

Each sample is joined with land-cover information derived from an external categorical land-cover raster covering the study area. The LR tile footprint is reprojected to the land-cover CRS; the mode value within that window is taken as the label.

• LC_detail_id / LC_detail_text: fine-grained class (for example, 41 → “Deciduous”).
• LC_superclass_id / LC_superclass_text: coarser super-group (for example, 40 → “Forest”).
• The enriched index also provides normalized land_cover_detail* and land_cover_superclass* aliases.
• These fields enable stratified metrics such as PSNR, SSIM, and SAM by environment type.
• Missing values may occur outside coverage or over NoData areas.

---

Geospatial & Data Notes

• CRS: Tiles use projected UTM zones; the exact CRS and affine transforms are included per record. Centroids are provided in WGS84.
• Alignment: Canonical 2.5 m HR is an exact 4× scaling of LR; supplementary 1 m HR is an exact 10× scaling. Both are pixel-grid aligned.
• Nodata: LR uses 65535 and HR uses 0 as the GeoTIFF nodata value.
• Bands: All raster products contain B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B11, and B12. B10 is excluded.
• LR grid: The LR files preserve observed Sentinel-2 radiometry, but all 12 bands are stored on one common 10 m tile grid. Bands with native 20 or 60 m resolution were sampled onto that grid by the GEE export; separate native-resolution grids are not included.
• HR availability: The 2.5 m HR reference is the canonical paper product and default. The aligned 1 m workflow product is provided as a supplementary resource for other potential uses.

---

Intended Uses

• Benchmarking SR models (classical, CNN, diffusion, and GAN approaches).
• Stratified evaluation by land-cover class.
• Qualitative visualization and error analysis.
• Testing spectral and radiometric consistency across all released Sentinel-2 bands.
• Exploring applications that benefit from the supplementary aligned 1 m workflow product.

Limitations

• Land-cover labels are window-mode summaries, not per-pixel annotations.
• Some tiles may lack land-cover labels or cross class boundaries.
• Cross-sensor pairs may retain differences caused by temporal separation, atmosphere, illumination, sensor point-spread functions, and residual coregistration.
• Coverage is limited to North American NEON sites and is imbalanced toward natural land-cover classes.
• The published SEN2NEON benchmark and reported metrics use the canonical 2.5 m HR product, not the supplementary 1 m product.

---

Integrity

s2_l2a_10m.sha256 records the SHA-256 checksum of every canonical LR file. The corrected files were verified byte-for-byte against the archived original GEE exports before this release was staged.

---

License

SEN2NEON is distributed under CC BY 4.0. Please retain attribution when redistributing or deriving work from the dataset.

---

Citation

If you use SEN2NEON in your research, please cite:

@article{donike2026sen2neon,
  author  = {Donike, Simon and Aybar, Cesar and Contreras, Julio and G{\'o}mez-Chova, Luis},
  title   = {SEN2NEON: Enabling Quantitative Benchmarking of Sentinel-2 Superresolution for All Multispectral Bands},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  volume  = {23},
  pages   = {6013905--6013905},
  year    = {2026},
  doi     = {10.1109/LGRS.2026.3703947}
}

---

Contact

• Maintainer: Image Processing Laboratory, University of Valencia, Spain
• Issues and questions: open a discussion at ESAOpenSR/SEN2NEON.