Multi-Date Sentinel-2 Super-Resolution

Implementation of 'Self-Supervised Multi-Date Multi-Spectral Super-Resolution for Sentinel-2 Imagery'

Graphical Abstract

Here is a graphical abstract of the proposed method from the paper:

Usage

Requirements

• Python 3.11
• GPU Nvidia
• Install pytorch and dependencies:

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Pre-trained Models

Available models in _trained_models/:

Model	Description
mdselfsen2.safetensors	Full model (10m + 20m): In paper MD-SelfS2
mdselfsen2_only10m.safetensors	10m bands only
sdselfsen2.safetensors	Single-date version (10m + 20m): In paper SD-SelfS2
sdselfsen2_only10m.safetensors	Single-date 10m only

Inference on Examples

We provide a run single- and multi date super-resolution on example ROIs _examples/ this will generate outputs in _workspace/run_examples/:

python run_examples.py

Training

• Download dateset from Zenodo (coming soon) and place in _data/ folder:

_data/
└── md_l1c_l1b_data/    (training data)
    ├── info/           (metadata for training/test sets, product info, date acquisition, orbits etc)
    ├── train.h5        (training data stored in HDF5 format)
    └── test.h5         (test data stored in HDF5 format)
└── ...

And run the script to train a new model:

python run_training.py [CONFIG]

where [CONFIG] is the name of the configuration (paper parameters) functions defined in run_training.py (e.g., mdselfsen2, sdselfsen2, mdselfsen2_only10m, sdselfsen2_only10m):

You can training by editing run_training.py with desired parameters:

• Number of epochs
• Learning rate
• Batch size
• Input/output bands
• Number of auxiliary frames

OpenSR Test

• Download dateset from Zenodo (coming soon) and place in _data/ folder:

_data/
├── ...
├── opensr_original/    (original OpenSR test data, single date)
└── opensr_md_v2/       (auxiliary frames for multi-date OpenSR data)

1. Predict and export results as georeferenced GeoTIFF:

python run_opensr_gen_geotiff.py

2. Calculate quality metrics (OpenSR Metrics, PSNR, SSIM, etc.):

python run_opensr_compute_metrics.py

License

This repository's source code is licensed under the MIT License. See the LICENSE file for details. The associated datasets are provided separately and are subject to their own licenses and usage restrictions. Please consult the official dataset pages for access conditions and licensing information.

Acknowledgments

• https://github.com/ESAOpenSR/opensr-test