andersoncotrim/md_l1c_l1b_data

Multi-Date Sentinel-2 L1C/L1B Dataset

Dataset accompanying the paper "Self-Supervised Multi-Date Multi-Spectral Super-Resolution for Sentinel-2 Imagery".

Each sample contains paired Sentinel-2 L1C and L1B images, spatial correspondence grids, validity masks, and auxiliary L1C acquisitions from multiple dates. Image data are stored in HDF5 files, while Parquet tables provide acquisition metadata, valid auxiliary frame indices, and dataset splits.

Code and training: GitHub repository.

Directory structure

README.md
train.h5
test.h5
info/
├── ref-train.parquet
├── ref-val.parquet
├── ref-train-val.parquet
├── ref-test.parquet
├── aux-random-train-val.parquet
└── aux-random-test.parquet

Training and validation samples share train.h5 (31,910 samples), whereas test samples are stored in test.h5 (6,440 samples). The corresponding partitions and auxiliary frame indices are defined by the Parquet tables in info/.

HDF5 structure

Each HDF5 file contains 12 root-level datasets compressed with GZIP level 1:

• Reference images: l1c and l1c_20m contain the reference L1C acquisition.
• Supervision images: l1b and l1b_20m contain the corresponding L1B acquisition in the L1B geometry.
• Auxiliary images: pastl1c* and futl1c* contain L1C acquisitions before and after the reference date, respectively.
• Spatial correspondence: grid and grid_20m provide per-band correspondence grids with the spatial dimensions of the L1B images.
• Validity masks: mask and mask_20m provide single-channel boolean masks in the L1B geometry.

The _20m suffix identifies the six bands at 20 m nominal resolution. Datasets without this suffix contain the four 10 m bands, in the following order:

Resolution	Band order
10 m	B04, B03, B02, B08
20 m	B05, B06, B07, B8A, B11, B12

Image values are stored as uint16. Reference and supervision images follow (N, C, H, W), while auxiliary sequences follow (N, T, C, H, W), where N denotes samples, T temporal slots, C spectral bands, and H, W spatial dimensions.

Dataset shapes

Dataset	Type	train.h5 shape	test.h5 shape
l1c	uint16	(31910, 4, 206, 130)	(6440, 4, 220, 134)
l1c_20m	uint16	(31910, 6, 103, 65)	(6440, 6, 110, 67)
l1b	uint16	(31910, 4, 200, 200)	(6440, 4, 200, 200)
l1b_20m	uint16	(31910, 6, 100, 100)	(6440, 6, 100, 100)
pastl1c	uint16	(31910, 30, 4, 206, 130)	(6440, 30, 4, 220, 134)
pastl1c_20m	uint16	(31910, 30, 6, 103, 65)	(6440, 30, 6, 110, 67)
futl1c	uint16	(31910, 16, 4, 206, 130)	(6440, 16, 4, 220, 134)
futl1c_20m	uint16	(31910, 16, 6, 103, 65)	(6440, 16, 6, 110, 67)
grid	float32	(31910, 4, 200, 200, 2)	(6440, 4, 200, 200, 2)
grid_20m	float32	(31910, 6, 100, 100, 2)	(6440, 6, 100, 100, 2)
mask	bool	(31910, 1, 200, 200)	(6440, 1, 200, 200)
mask_20m	bool	(31910, 1, 100, 100)	(6440, 1, 100, 100)

Each sample provides 30 past and 16 future temporal slots. Unused slots may contain zeros; valid frames must be identified through the auxiliary metadata tables described below.

Metadata

Reference acquisitions (ref-*.parquet)

These tables identify reference acquisitions and define the dataset partitions.

File	Rows	HDF5 file	Split
ref-train.parquet	27,007	train.h5	Training
ref-val.parquet	4,903	train.h5	Validation
ref-train-val.parquet	31,910	train.h5	Training + validation
ref-test.parquet	6,440	test.h5	Test

Training and validation have disjoint sample IDs, and their union covers train.h5. Sample IDs in test.h5 are indexed independently.

Column	Type	Description
sample_id	int64	Zero-based index in the associated HDF5 file
product_id	string	Reference Sentinel-2 L1C product identifier
date	UTC timestamp (µs)	Acquisition date and time
orbit	int64	Relative orbit number

Auxiliary acquisitions (aux-random-*.parquet)

These tables list valid auxiliary acquisitions and their temporal indices.

File	Rows	Samples	Frames per sample
aux-random-train-val.parquet	675,542	31,910	4–45
aux-random-test.parquet	121,078	6,440	12–42

Each auxiliary frame is uniquely identified within its table by (sample_id, source, date_index):

• source = 'p': selects pastl1c*, ordered from the most recent to the oldest acquisition.
• source = 'f': selects futl1c*, ordered from the earliest to the latest acquisition.
• date_index: zero-based index within the selected temporal sequence.

Column	Type	Description
sample_id	int64	Zero-based index in the associated HDF5 file
source	string	Temporal source (p or f)
date_index	int64	Temporal slot index
product_id	string	Auxiliary Sentinel-2 L1C product identifier
date	UTC timestamp (ns)	Acquisition date and time
orbit	int64	Relative orbit number
cloud_cover	float64	Product-level cloud coverage
dy	float64	Vertical alignment shift (10 m pixels)
dx	float64	Horizontal alignment shift (10 m pixels)
score	float64	Alignment score

Auxiliary L1C images were aligned with the reference using Fourier-based phase correlation, with subpixel refinement through sinc fitting. The resulting translations (dy, dx) are expressed in 10m pixels.

Reading a sample

The following example loads a reference L1C image, its L1B supervision image, spatial correspondence data, and one valid auxiliary acquisition from the test set. It requires h5py, pandas, and pyarrow.

from pathlib import Path

import h5py
import pandas as pd

root = Path(".")  # Dataset root directory

ref = pd.read_parquet(root / "info/ref-test.parquet")
aux = pd.read_parquet(root / "info/aux-random-test.parquet")

sample_id = int(ref.iloc[0]["sample_id"])
selected = aux.loc[aux["sample_id"] == sample_id].iloc[0]

prefix = {"p": "pastl1c", "f": "futl1c"}[selected["source"]]
date_index = int(selected["date_index"])

with h5py.File(root / "test.h5", "r") as f:
    reference = f["l1c"][sample_id]    # (4, 220, 134)
    target = f["l1b"][sample_id]       # (4, 200, 200)
    grid = f["grid"][sample_id]        # (4, 200, 200, 2)
    mask = f["mask"][sample_id]        # (1, 200, 200)

    auxiliary = f[prefix][sample_id, date_index]
    auxiliary_20m = f[prefix + "_20m"][sample_id, date_index]

Training and reproducibility

The training implementation, pretrained models, and evaluation scripts are available in the official GitHub repository.

To train a model, place the dataset in the following directory within the cloned repository:

_data/
└── md_l1c_l1b_data/
    ├── info/
    ├── train.h5
    └── test.h5

Citation

If you use this dataset in your research, please cite our paper:

@article{cotrim2026selfsupervised,
  title   = {Self-Supervised Multi-Date Multi-Spectral Super-Resolution for Sentinel-2 Imagery},
  author  = {Cotrim, A. Nogueira and Anger, J. and Facciolo, G. and Santos, C. A. Nogueira and Pedrini, H.},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year    = {2026},
  note    = {Accepted for publication}
}

Acknowledgments

We thank the Sentinel-2 team at ESA for enabling access to the L1B products. This paper contains modified Copernicus Sentinel data 2024.