opensr-test 1.2.0

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

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://ieeexplore.ieee.org/document/10530998

---

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)
>>> {'reflectance': 0.253, 'spectral': 26.967, 'spatial': 1.0, 'synthesis': 0.2870, 'ha_percent': 0.892, 'om_percent': 0.0613, 'im_percent': 0.04625}

This model returns:

• reflectance: How SR affects the reflectance norm of the LR image. By default, it uses the L1 distance. The lower the value, the better the reflectance consistency.

• spectral: How SR affects the spectral signature of the LR image. By default, it uses the spectral angle distance (SAD). The lower the value, the better the spectral consistency. The angles are in degrees.

• spatial: The spatial alignment between the SR and LR images. By default, it uses Phase Correlation Coefficient (PCC). Some SR models introduce spatial shift, which can be detected by this metric.

• synthesis: The high-frequency details introduced by the SR model. By default, it uses the L1 distance. The higher the value, the better the synthesis quality.

• ha_metric: The amount of hallucinations in the SR image. A hallucination is a detail (high-gradient) in the SR image that is not present in the HR image. The lower the value, the better the correctness of the SR image.

• om_metric: The amount of omissions in the SR image. An omission is a detail in the HR image that is not present in the SR image. The lower the value, the better the correctness of the SR image.

• im_metric: The amount of improvements in the SR image. An improvement is a detail in the SR image that is present in the HR image and not in the LR image. The higher the value, the better the correctness of the SR image.

Benchmark

Benchmark comparison of SR models. Downward arrows (↓) denote metrics in which lower values are preferable, and upward arrows (↑) indicate that higher values reflect better performance. This table is an improved version of the one presented in the opensr-test paper, considering the latest version of the datasets and other distance metrics. To get the datasets used in the original paper, set the dataset parameter to version=020 in the opensr_test.load function.

CLIP

We use the RemoteCLIP model to measure the distance between the LR, SR, and HR images. The parameters of the experiments are: {"device": "cuda", "agg_method": "patch", "patch_size": 16, "correctness_distance": "clip"}. This distance metric allows us to quantify the amount of semantic information introduced by the SR model. By comparing the three categories, we can assess the accuracy of the semantic information introduced.

model	reflectance ↓	spectral ↓	spatial ↓	synthesis ↑	ha_metric ↓	om_metric ↓	im_metric ↑
ldm_baseline	0.1239 ± 0.0405	12.8441 ± 2.7508	0.0717 ± 0.0683	0.0409 ± 0.0290	0.5963 ± 0.3055	0.2327 ± 0.2238	0.1710 ± 0.1435
opensrmodel	0.0076 ± 0.0046	1.9739 ± 1.0507	0.0118 ± 0.0108	0.0194 ± 0.0120	0.1052 ± 0.0590	0.6927 ± 0.1343	0.2021 ± 0.0854
satlas	0.1197 ± 0.0233	15.1521 ± 2.9876	0.2766 ± 0.0741	0.0648 ± 0.0302	0.6996 ± 0.2058	0.0872 ± 0.0947	0.2132 ± 0.1393
sr4rs	0.0979 ± 0.0509	22.4905 ± 2.1168	1.0099 ± 0.0439	0.0509 ± 0.0237	0.3099 ± 0.1704	0.3486 ± 0.1753	0.3415 ± 0.1042
superimage	0.0068 ± 0.0016	1.8977 ± 1.1053	0.0004 ± 0.0032	0.0130 ± 0.0073	0.0610 ± 0.0305	0.8524 ± 0.0586	0.0866 ± 0.0395

LPIPS

We use the LPIPS model to measure the distance between the LR, SR, and HR images. The parameters of the experiments are: {"device": "cuda", "agg_method": "patch", "patch_size": 16, "correctness_distance": "lpips"}. This distance metric allows us to quantify the amount of perceptual information introduced by the SR model. By comparing the three categories, we can assess the accuracy of the perceptual information introduced.

model	reflectance ↓	spectral ↓	spatial ↓	synthesis ↑	ha_metric ↓	om_metric ↓	im_metric ↑
ldm_baseline	0.1239 ± 0.0405	12.8441 ± 2.7508	0.0717 ± 0.0683	0.0409 ± 0.0290	0.4558 ± 0.2932	0.3558 ± 0.2518	0.1884 ± 0.1232
opensrmodel	0.0076 ± 0.0046	1.9739 ± 1.0507	0.0118 ± 0.0108	0.0194 ± 0.0120	0.0642 ± 0.0271	0.6690 ± 0.1291	0.2668 ± 0.1071
satlas	0.1197 ± 0.0233	15.1521 ± 2.9876	0.2766 ± 0.0741	0.0648 ± 0.0302	0.5999 ± 0.2182	0.0588 ± 0.0552	0.3413 ± 0.1858
sr4rs	0.0979 ± 0.0509	22.4905 ± 2.1168	1.0099 ± 0.0439	0.0509 ± 0.0237	0.3417 ± 0.1833	0.1924 ± 0.1402	0.4659 ± 0.1448
superimage	0.0068 ± 0.0016	1.8977 ± 1.1053	0.0004 ± 0.0032	0.0130 ± 0.0073	0.0357 ± 0.0200	0.8844 ± 0.0391	0.0800 ± 0.0301

Normalized Difference (ND)

We use the normalized difference (ND) distance to measure the distance between the LR, SR, and HR images. In contrast to L1 distance, ND is less sensitive to the magnitude of the reflectance values, providing a more fair detection of the hallucinations and omissions when assessing at the pixel level. The parameters of the experiments are: {"device": "cuda", "agg_method": "patch", "patch_size": 1, "correctness_distance": "nd"}. This distance metric allows us to quantify high-frequency details introduced by the SR model at pixel level. However, at pixel level, small changes in the reflectance values could introduce noise in the hallucination, omission, and improvement estimation.

model	reflectance ↓	spectral ↓	spatial ↓	synthesis ↑	ha_metric ↓	om_metric ↓	im_metric ↑
ldm_baseline	0.0505 ± 0.0161	9.6923 ± 2.1742	0.0715 ± 0.0679	0.0285 ± 0.0307	0.6067 ± 0.2172	0.3088 ± 0.1786	0.0845 ± 0.0428
opensrmodel	0.0031 ± 0.0018	1.2632 ± 0.5878	0.0114 ± 0.0111	0.0068 ± 0.0044	0.3431 ± 0.0738	0.4593 ± 0.0781	0.1976 ± 0.0328
satlas	0.0489 ± 0.0086	12.1231 ± 3.1529	0.2742 ± 0.0748	0.0227 ± 0.0107	0.8004 ± 0.0641	0.1073 ± 0.0393	0.0923 ± 0.0266
sr4rs	0.0396 ± 0.0198	3.4044 ± 1.6882	1.0037 ± 0.1520	0.0177 ± 0.0083	0.7274 ± 0.0840	0.1637 ± 0.0572	0.1089 ± 0.0292
superimage	0.0029 ± 0.0009	1.5672 ± 1.0692	0.0132 ± 0.1131	0.0046 ± 0.0027	0.2026 ± 0.0692	0.6288 ± 0.0754	0.1686 ± 0.0302

To reproduce the results, check this Colab notebook.

Installation

Install the latest version from PyPI:

pip install opensr-test
pip install opensr-test[perceptual] # Install to test perceptual metrics

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

Datasets

The opensr-test package provides five datasets for benchmarking SR models. These datasets are carefully crafted to minimize spatial and spectral misalignment. See our Hugging Face repository for more details about the datasets. https://huggingface.co/datasets/isp-uv-es/opensr-test

Dataset	Scale factor	Number of images	HR patch size
NAIP	x4	62	484x484
SPOT	x4	9	512x512
Venµs	x2	59	256x256
SPAIN CROPS	x4	28	512x512
SPAIN URBAN	x4	20	512x512

NAIP (X4 scale factor)

The National Agriculture Imagery Program (NAIP) dataset is a high-resolution aerial imagery dataset that covers the continental United States. The dataset consists of 2.5m NAIP imagery captured in the visible and near-infrared spectrum (RGBNIR) and all Sentinel-2 L1C and L2A bands. The dataset focus in crop fields, forests, and bare soil areas.

import opensr_test

naip = opensr_test.load("naip")

SPOT (X4 scale factor)

The SPOT imagery were obtained from the worldstat dataset. The dataset consists of 2.5m SPOT imagery captured in the visible and near-infrared spectrum (RGBNIR) and all Sentinel-2 L1C and L2A bands. The dataset focus in urban areas, crop fields, and bare soil areas.

import opensr_test

spot = opensr_test.load("spot")

Venµs (X2 scale factor)

The Venµs images were obtained from the Sen2Venµs dataset. The dataset consists of 5m Venµs imagery captured in the visible and near-infrared spectrum (RGBNIR) and all Sentinel-2 L1C and L2A bands. The dataset focus in crop fields, forests, urban areas, and bare soil areas.

import opensr_test

venus = opensr_test.load("venus")

SPAIN CROPS (x4 scale factor)

The SPAIN CROPS dataset consists of 2.5m aerial imagery captured in the visible and near-infrared spectrum (RGBNIR) by the Spanish National Geographic Institute (IGN). The dataset includes all Sentinel-2 L1C and L2A bands. The dataset focus in crop fields and forests.

import opensr_test

spain_crops = opensr_test.load("spain_crops")

SPAIN URBAN (x4 scale factor)

The SPAIN URBAN dataset consists of 2.5m aerial imagery captured in the visible and near-infrared spectrum (RGBNIR) by the Spanish National Geographic Institute (IGN). The dataset includes all Sentinel-2 L1C and L2A bands. The dataset focus in urban areas.

spain_urban = opensr_test.load("spain_urban")

Examples

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

Model	Framework	Link
SR4RS	TensorFlow
SuperImage	PyTorch
LDMSuperResolutionPipeline	Diffuser
opensr-model	Pytorch
EvoLand	Pytorch
SWIN2-MOSE	Pytorch

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)
srmodel.eval()

# Load the data
dataset = opensr_test.load("spot")
lr, hr = dataset["L2A"], dataset["HRharm"]

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

# Define the image to be tested
idx = 0
lr_img = torch.from_numpy(lr[idx, 0:3]) / 10000
hr_img = torch.from_numpy(hr[idx, 0:3]) / 10000
with torch.no_grad():
    sr_img = srmodel(lr_img[None]).squeeze()

# Compute the metrics
metrics.compute(
    lr=lr_img, sr=sr_img, hr=hr_img
)

Now, we can visualize the results using the opensr_test.plot_* module. To display the triplets LR, SR and HR images:

metrics.plot_triplets()

Display the summary of the metrics. The plot shows:

• LR: Low Resolution image
• LRdown: Downsampled Low Resolution image using bilinear interpolation and triangular antialiasing filter.
• SR: Super-Resolved image.
• SRharm: Harmonized super-resolution image.
• HR: High Resolution image.
• Reflectance Consistency: Reflectance consistency between the LR and HR images.
• Spectral Consistency: Spectral consistency between the LR and HR images.
• Distance normalized to the Omissiom, Hallucination, and Improvement spaces.

metrics.plot_summary()

Display the correctness of the SR image. The blue color represents the pixels closer to the improvement space (HR), green pixels are closer to the omission space (LR), and red pixels are closer to the hallucination space (neither in HR nor LR). The distance SR-LR and SR-HR are normalized to the LR-HR distance that is independent of the SR model. Threfore a pixel or patch with a improvement distance ($d_{im}$) of 3 means that the SR is further to the HR in 3 LR-HR units. The same applies to the omission distance.

metrics.plot_tc()

Display the histogram of the distances before and after the normalization:

metrics.plot_stats()

Display a ternary plot of the metrics:

metrics.plot_ternary()

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{10530998,
  author={Aybar, Cesar and Montero, David and Donike, Simon and Kalaitzis, Freddie and Gómez-Chova, Luis},
  journal={IEEE Geoscience and Remote Sensing Letters}, 
  title={A Comprehensive Benchmark for Optical Remote Sensing Image Super-Resolution}, 
  year={2024},
  volume={21},
  number={},
  pages={1-5},
  keywords={Measurement;Remote sensing;Spatial resolution;Superresolution;Reflectivity;Protocols;Inspection;Benchmarking;datasets;deep learning;NAIP;Sentinel-2 (S2);SPOT;super-resolution (SR)},
  doi={10.1109/LGRS.2024.3401394}}

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).