A comprehensive benchmark for real-world Sentinel-2 imagery super-resolution

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

Documentation: https://esaopensr.github.io/opensr-test

PyPI: https://pypi.org/project/opensr-test/

Paper: https://www.techrxiv.org/users/760184/articles/735467-a-comprehensive-benchmark-for-optical-remote-sensing-image-super-resolution

Overview

Super-Resolution (SR) aims to improve satellite imagery ground sampling distance. However, two problems are common in the literature. First, most models are tested on synthetic data, raising doubts about their real-world applicability and performance. Second, traditional evaluation metrics such as PSNR, LPIPS, and SSIM are not designed to assess SR performance. These metrics fall short, especially in conditions involving changes in luminance or spatial misalignments - scenarios frequently encountered in real world.

To address these challenges, 'opensr-test' provides a fair approach for SR benchmark. We provide three datasets carefully crafted to minimize spatial and spectral misalignment. Besides, 'opensr-test' precisely assesses SR algorithm performance across three independent metrics groups that measure consistency, synthesis, and correctness.

How to use

The example below shows how to use opensr-test to benchmark your SR model.

import torch
import opensr_test

lr = torch.rand(4, 64, 64)
hr = torch.rand(4, 256, 256)
sr = torch.rand(4, 256, 256)

metrics = opensr_test.Metrics()
metrics.compute(lr=lr, sr=sr, hr=hr)

Benchmark

Benchmark comparison of SR models. Downward arrows (↓) denote metrics in which lower values are preferable, and upward arrows (↑) indicate metrics in which higher values reflect better performance.

Installation

Install the latest version from PyPI:

pip install opensr-test

Upgrade opensr-test by running:

pip install -U opensr-test

Install the latest dev version from GitHub by running:

pip install git+https://github.com/ESAOpenSR/opensr-test

Examples

The following examples show how to use opensr-test to benchmark your SR model.

• Use opensr-test with TensorFlow model

• Use opensr-test with PyTorch model

• Use opensr-test with a diffuser model

Visualizations

The opensr-test package provides a set of visualizations to help you understand the performance of your SR model.

import torch
import opensr_test
import matplotlib.pyplot as plt

from super_image import HanModel

# Define the SR model
srmodel = HanModel.from_pretrained('eugenesiow/han', scale=4)

# Load the data
lr, hr, landuse, parameters = opensr_test.load("spot").values()

# Define the benchmark experiment
metrics = opensr_test.Metrics()

# Define the image to be tested
idx = 0
lr_img = torch.from_numpy(lr[idx, 0:3])
hr_img = torch.from_numpy(hr[idx, 0:3])
sr_img = srmodel(lr_img[None]).squeeze().detach()

# Compute the metrics
metrics.compute(
    lr=lr_img, sr=sr_img, hr=hr_img,
    stability_threshold = parameters.stability_threshold[idx],
    im_score = parameters.correctness_params[0],
    om_score = parameters.correctness_params[1],
    ha_score = parameters.correctness_params[2]
)

Now, we can visualize the results using the opensr_test.visualize module. fDisplay the triplets LR, SR and HR images:

metrics.plot_triplets()

Display the quadruplets LR, SR, HR and landuse images:

metrics.plot_quadruplets()

Display the matching points between the LR and SR images:

metrics.plot_spatial_matches()

Display a summary of all the metrics:

metrics.plot_summary()

Display the correctness of the SR image:

metrics.plot_tc()

Deeper understanding

Explore the API section for more details about personalizing your benchmark experiments.

Citation

If you use opensr-test in your research, please cite our paper:

@article{aybar2024comprehensive,
  title={A Comprehensive Benchmark for Optical Remote Sensing Image Super-Resolution},
  author={Aybar, Cesar and Montero, David and Donike, Simon and Kalaitzis, Freddie and G{\'o}mez-Chova, Luis},
  journal={Authorea Preprints},
  year={2024},
  publisher={Authorea}
}

Acknowledgements

This work was make with the support of the European Space Agency (ESA) under the project “Explainable AI: application to trustworthy super-resolution (OpenSR)”. Cesar Aybar acknowledges support by the National Council of Science, Technology, and Technological Innovation (CONCYTEC, Peru) through the “PROYECTOS DE INVESTIGACIÓN BÁSICA – 2023-01” program with contract number PE501083135-2023-PROCIENCIA. Luis Gómez-Chova acknowledges support from the Spanish Ministry of Science and Innovation (project PID2019-109026RB-I00 funded by MCIN/AEI/10.13039/501100011033).