Datasets:
SatClean-Bench v1 – 600 Sentinel-2 patches, 3 degradation levels
Browse files
README.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
pretty_name: SatClean-Bench
|
| 4 |
+
task_categories:
|
| 5 |
+
- image-to-image
|
| 6 |
+
tags:
|
| 7 |
+
- remote-sensing
|
| 8 |
+
- satellite
|
| 9 |
+
- super-resolution
|
| 10 |
+
- denoising
|
| 11 |
+
- sentinel-2
|
| 12 |
+
- benchmark
|
| 13 |
+
- earth-observation
|
| 14 |
+
size_categories:
|
| 15 |
+
- n<1K
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
<div align="center">
|
| 19 |
+
|
| 20 |
+
# SatClean-Bench
|
| 21 |
+
|
| 22 |
+
**Fixed evaluation benchmark for joint denoising + 3× super-resolution of Sentinel-2 imagery**
|
| 23 |
+
|
| 24 |
+
[](https://github.com/Nora-Research-Lab)
|
| 25 |
+
[](https://huggingface.co/NoraResearchLab)
|
| 26 |
+
[](https://www.linkedin.com/company/nora-research-lab)
|
| 27 |
+
[](https://x.com/noraresearchlab)
|
| 28 |
+
[](https://noraresearchlab.site)
|
| 29 |
+
|
| 30 |
+
</div>
|
| 31 |
+
|
| 32 |
+
---
|
| 33 |
+
|
| 34 |
+
## Task
|
| 35 |
+
|
| 36 |
+
Given a **noisy 30 m RGB** image, recover a **clean 10 m RGB** image.
|
| 37 |
+
|
| 38 |
+
The benchmark supplies three fixed, reproducible degradation levels that cover (and exceed) typical atmospheric and sensor artefacts found in real Sentinel-2 data:
|
| 39 |
+
|
| 40 |
+
| Level | Noise σ (DN of 255) | Striping | Haze (airlight blend) | Description |
|
| 41 |
+
|-------|---------------------|----------|-----------------------|--------------------------|
|
| 42 |
+
| 1 | 15 | none | none | mild noise |
|
| 43 |
+
| 2 | 25 | 8 DN | none | noise + striping |
|
| 44 |
+
| 3 | 50 | none | 0.35 | heavy noise + haze |
|
| 45 |
+
|
| 46 |
+
**Degradation pipeline** (locked in `bench_spec.json`):
|
| 47 |
+
|
| 48 |
+
1. 3× box downsample (10 m → 30 m)
|
| 49 |
+
2. Mild Gaussian PSF blur (σ = 0.5 px)
|
| 50 |
+
3. Optional haze (smooth airlight field)
|
| 51 |
+
4. Additive Gaussian noise
|
| 52 |
+
5. Optional row/column striping
|
| 53 |
+
6. Clamp → 8-bit quantisation
|
| 54 |
+
|
| 55 |
+
## Files
|
| 56 |
+
|
| 57 |
+
| File | Shape | Role |
|
| 58 |
+
|-------------------|---------------------------|-------------------------------|
|
| 59 |
+
| `test_hr.npy` | (600, 384, 384, 3) uint8 | Clean high-resolution targets |
|
| 60 |
+
| `level1_lr.npy` | (600, 3, 128, 128) uint8 | Degraded inputs – Level 1 |
|
| 61 |
+
| `level2_lr.npy` | (600, 3, 128, 128) uint8 | Degraded inputs – Level 2 |
|
| 62 |
+
| `level3_lr.npy` | (600, 3, 128, 128) uint8 | Degraded inputs – Level 3 |
|
| 63 |
+
| `bench_spec.json` | — | Exact parameters, seed, metrics definition |
|
| 64 |
+
|
| 65 |
+
## Metrics
|
| 66 |
+
|
| 67 |
+
All metrics are computed on the **uint8-quantised** RGB output:
|
| 68 |
+
|
| 69 |
+
- **PSNR** / **SSIM**
|
| 70 |
+
- **LPIPS** (AlexNet backbone)
|
| 71 |
+
- Optional efficiency: FLOPs on a 256×256 LR tile and tiled fp16 latency on 1024×1024 LR input
|
| 72 |
+
|
| 73 |
+
## Source data
|
| 74 |
+
|
| 75 |
+
High-resolution patches are extracted from the public
|
| 76 |
+
[Major-TOM / Core-S2L2A](https://huggingface.co/datasets/Major-TOM/Core-S2L2A) dataset
|
| 77 |
+
(Sentinel-2 Level-2A, 10 m).
|
| 78 |
+
|
| 79 |
+
Tone curve applied: `clip(DN / 3000) ** 0.7`.
|
| 80 |
+
|
| 81 |
+
## Citation
|
| 82 |
+
|
| 83 |
+
If you use SatClean-Bench please cite both the original Major-TOM paper and this benchmark:
|
| 84 |
+
|
| 85 |
+
```bibtex
|
| 86 |
+
@inproceedings{Major_TOM,
|
| 87 |
+
title = {Major TOM: Expandable Datasets for Earth Observation},
|
| 88 |
+
author = {Alistair Francis and Mikolaj Czerkawski},
|
| 89 |
+
year = {2024},
|
| 90 |
+
booktitle = {IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium},
|
| 91 |
+
eprint = {2402.12095},
|
| 92 |
+
archivePrefix = {arXiv},
|
| 93 |
+
primaryClass = {cs.CV}
|
| 94 |
+
}
|
| 95 |
+
```
|
| 96 |
+
|
| 97 |
+
## Maintainer
|
| 98 |
+
|
| 99 |
+
**NORA Research Lab**
|
| 100 |
+
|
| 101 |
+
[](https://github.com/Nora-Research-Lab)
|
| 102 |
+
[](https://huggingface.co/NoraResearchLab)
|
| 103 |
+
[](https://www.linkedin.com/company/nora-research-lab)
|
| 104 |
+
[](https://x.com/noraresearchlab)
|
| 105 |
+
|
| 106 |
+
Quick links:
|
| 107 |
+
[Website](https://noraresearchlab.site) ·
|
| 108 |
+
[GitHub](https://github.com/Nora-Research-Lab) ·
|
| 109 |
+
[Hugging Face](https://huggingface.co/NoraResearchLab) ·
|
| 110 |
+
[LinkedIn](https://www.linkedin.com/company/nora-research-lab) ·
|
| 111 |
+
[X](https://x.com/noraresearchlab)
|