marklpd/PSRGAN-Sentinel-2

PSRGAN Sentinel-2 Super-Resolution

Six test patches: degraded 64×64 input (top, shown enlarged) and Sentinel-2 target (bottom)

Overview

Sentinel-2 images every land surface on Earth every five days, but at 10 m per pixel most fields, roads and buildings are only a few pixels wide. Super-resolution models can sharpen these images. But these models also invent details that never existed, and they do not tell you where.

This dataset is built to train and evaluate super-resolution models that know when they are guessing. It contains 12,295 pairs of 64×64 low-resolution inputs and 256×256 high-resolution targets (4× upscaling) from 771 cloud-filtered Sentinel-2 L2A tiles. All of them lie in central Europe (Switzerland, Austria, northern Italy, south-eastern France, Germany and Slovenia), and eight test-only tiles come from Herzogenaurach, Wolfsburg and München. Every patch also carries its NDVI, the scene classification (SCL), the raw reflectance and its footprint, so errors and uncertainty can be analysed by land cover.

It is the dataset behind PSRGAN, a Bayesian Real-ESRGAN that outputs a super-resolved image together with a per-pixel uncertainty map.

Randomized degradation

The inputs are not a plain bicubic downscale of the targets. Each HR patch is first blurred with a random kernel (Gaussian with σ between 0.5 and 1.5, horizontal or vertical motion, or defocus, 3 to 9 px wide) and then reduced 4× with a random interpolation (bicubic, bilinear or area). So a model has to learn super-resolution instead of inverting one fixed downscaling operation. The blur and interpolation of every patch are stored in its row. The same pipeline is available as degradation.py in the model repo, to make LR inputs from your own HR images.

One HR patch through the pipeline with each blur type. The LR images are shown enlarged, the bottom row zooms into the white box.

Splits without leakage

Splits are assigned per tile, before any patch is cut. Overlapping tiles are merged into one group, every group belongs to the 0.5°×0.5° block that contains its centre, and whole blocks go to train, validation or test (about 80/10/10, seed 42). Blocks with city tiles always go to test. No tile, patch ID or patch footprint appears in two splits, which the pipeline asserts during creation. Augmentation (flips, 90° rotations) is meant for train at training time only, so it cannot leak test content into training.

split	patches	tiles	share of patches
train	9,575	601	77.9%
validation	1,392	87	11.3%
test	1,328	83	10.8%

The test split includes 128 patches from the three cities. The tile-to-split table is in metadata/tiles.csv, together with the source of every tile and the copies that are left out.

What is NOT in the dataset

• There is no high-resolution ground truth from another sensor. The targets are Sentinel-2 itself and the inputs are degraded from them, so models trained here learn 4× super-resolution of degraded Sentinel-2 imagery, not 2.5 m imagery from 10 m.
• Only RGB (B04, B03, B02) is included as image pairs. NIR (B08) only enters through the NDVI.
• Clouds and open water are under-represented because of the filters below.
• Pixels are not square. Tiles were requested as 512×512 rasters over square WGS84 boxes of 0.0721° (8 km / 111 km per degree), so one pixel covers about 10 m east-west and 15.6 m north-south.

Dataset Details

Source Data

• Copernicus Sentinel-2 Level-2A (bands B02, B03, B04, B08 and SCL) via the Sentinel Hub Process API of the Copernicus Data Space Ecosystem, as a mosaic per tile from 2025-09-20 to 2025-10-07. The newest scene comes first and older scenes fill its gaps. The notebook requested the least-cloudy order, but in a field the API ignores, so it used its default most-recent order. The city tiles use the most recent scenes with at most 40% cloud cover from the first of six two-week windows between April and October 2025 that passed the cloud filter.
• The Copernicus quota ran out during version 1.1, so 148 of its 257 new tiles come from the same Level-2A product on the Microsoft Planetary Computer. They use the same grid, scene order, band offset and nearest-neighbour resampling. On 12 existing tiles this reproduced 92.5% of the reflectance values exactly (mean difference 0.0009) and 99.7% of the SCL classes.
• Random tiles were drawn uniformly from 4.0 to 19.5°E and 45.5 to 54.5°N (the first 64 from 7.5 to 13.5°E and 46.5 to 51.5°N), without overlapping existing tiles.
• Tiles with more than 25% cloud, cloud shadow, cirrus or unclassified pixels were rejected, and so were new tiles with more than 50% water or more than 1% no-data. In this autumn window that removed almost every lowland candidate, so 98.2% of the accepted tiles lie near the Alpine region.
• Up to 16 random 256×256 crops per tile, each with at most 15% cloud pixels and normalised with a 1 to 99% percentile stretch.

In the training split, vegetation covers 70% of all pixels, bare ground 15%, snow and ice 5% and water 2%. Per-split SCL shares are in metadata/stats.json.

Schema

field	type	description
patch_id	string	<tile_id>_pNN, unique
tile_id	string	source tile (aoi_XXXX random, custom_XXXX city)
source	string	random or city
aoi_name	string	city name for city tiles, empty otherwise
lr	image (PNG, 64×64 RGB)	degraded input
hr	image (PNG, 256×256 RGB)	target, B04/B03/B02 after a 1 to 99% percentile stretch
ndvi	float32 [256, 256]	(B08 − B04) / (B08 + B04), clipped to [−1, 1]
scl	uint8 [256, 256]	Sentinel-2 scene classification (0 to 11)
reflectance	uint16 [3, 256, 256]	B04, B03, B02 surface reflectance × 10,000 (unstretched)
lon, lat	float64	patch centre (WGS84)
lon_min, lat_min, lon_max, lat_max	float64	patch footprint (WGS84)
crop_y, crop_x	int32	top-left pixel of the patch in its 512×512 tile
cloud_fraction, water_fraction	float32	share of SCL classes 3, 7, 8, 9, 10 and of class 6
stretch_low, stretch_high	float32	1st and 99th percentile of the patch reflectance (0 to 1 scale)
blur, blur_kernel, blur_sigma	string, int32, float32	blur type (gaussian, motion_h, motion_v, defocus), kernel size in px, Gaussian σ (0 otherwise)
interpolation	string	bicubic, bilinear or area
time_from, time_to	string	acquisition window of the mosaic

hr is clip((reflectance / 10000 − stretch_low) / (stretch_high − stretch_low), 0, 1) quantised to 8 bit. lr is hr (in float) blurred with the stored kernel, downsampled 4× with the stored interpolation and quantised to 8 bit.

Evaluation Note

The HR images are nearest-neighbour resamples of the 10 m Sentinel-2 grid onto a latitude-longitude grid (for the random tiles 0.072° per 512 px, so 15.6 m north-south and 9 to 11 m east-west depending on latitude). Edges can therefore jump by one pixel in an irregular pattern that the LR images do not contain, which lowers PSNR and SSIM significantly for any method evaluated on this data. The blur filters only change the LR images, the HR images stay as downloaded.

How to use it

from datasets import load_dataset

ds = load_dataset("marklpd/PSRGAN-Sentinel-2")
sample = ds["train"][0]
sample["lr"], sample["hr"]          # PIL images, 64×64 and 256×256

# Only the image pairs (much faster, skips NDVI, SCL and reflectance)
pairs = load_dataset("marklpd/PSRGAN-Sentinel-2", split="test", columns=["patch_id", "lr", "hr"])

PyTorch tensors in [0, 1]:

import numpy as np, torch

def to_tensor(batch):
    batch["lr"] = [torch.from_numpy(np.asarray(im, np.float32) / 255).permute(2, 0, 1) for im in batch["lr"]]
    batch["hr"] = [torch.from_numpy(np.asarray(im, np.float32) / 255).permute(2, 0, 1) for im in batch["hr"]]
    return batch

train = load_dataset("marklpd/PSRGAN-Sentinel-2", split="train", columns=["lr", "hr"]).with_transform(to_tensor)

Intended Use

Research on super-resolution and uncertainty estimation in Earth observation. Use validation for model selection and keep test for final numbers. Super-resolved Sentinel-2 images are estimates, not measurements, so details a model adds beyond the input should not be used for decisions about individual parcels, buildings or people. Models trained here should also be expected to perform worse outside central Europe, in other seasons and in other climates.

Data Developer

Mark-Daniel Leupold.

Release Date and Version

Version 1.1 (12,295 patches, 18.0 GB of parquet), released on Hugging Face on October 20, 2026. Version 1.0 contained 128 tiles that were pixel-identical copies of 64 other tiles (2,048 patches) and 3 patches that repeated a crop position within their tile. Version 1.1 removes them and adds 4,111 patches from 257 new tiles. All other patches and their splits are unchanged. A new tile in an existing 0.5° block joins that block's split, and new blocks fill the splits toward 80/10/10. PSRGAN is trained on version 1.1.

Version 1.0 replaced an earlier internal version with 12,253 patches.

License/Terms of Use

Creative Commons Attribution 4.0 International (CC BY 4.0). The dataset contains modified Copernicus Sentinel data (2025), processed via the Copernicus Data Space Ecosystem, and use of the Sentinel data is governed by the Legal Notice on the use of Copernicus Sentinel Data and Service Information.

Citation

@misc{leupold2026psrgans2,
  author = {Leupold, Mark-Daniel},
  title = {PSRGAN Sentinel-2 Super-Resolution: Spatially Split LR/HR Pairs with Land-Cover Context},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/marklpd/PSRGAN-Sentinel-2}
}