Datasets:

ArXiv:
File size: 3,261 Bytes
1901721
 
 
 
37170fd
 
1901721
37170fd
 
1901721
37170fd
 
 
1901721
37170fd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e6c070d
37170fd
 
 
 
 
 
 
 
e6c070d
37170fd
605b0c0
37170fd
e6c070d
 
 
 
37170fd
 
e6c070d
b15b7aa
 
 
 
 
 
 
 
 
 
 
 
e6c070d
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
---
task_categories:
- image-segmentation
---
# NLCD-L 
This dataset incorporates both SSL4EO-L Benchmark dataset and the NLCD-L dataset which is derived from the original SSL4EO-L Benchmark dataset by combining optical data from Landsat-7 and Landsat 8-9 with NLCD ground-truth labels, originally proposed in SSL4EO-L. The dataset contains 20 MSI bands, deliberately exceeding Sentinel-2’s channel count. It comprises 17,500 training samples, 3,750 validation samples, and 3,750 test samples.

Please refer to the original SSL4EO-L paper for more detailed information about the original SSL4EO-L Benchmark dataset:
- Paper: https://arxiv.org/abs/2306.09424

## How to Use This Dataset
```python
from datasets import load_dataset

# To access NLCD-L, set name to etm_oli_toa_nlcd in load_dataset function
dataset = load_dataset("GFM-Bench/SSL4EO-L-Benchmark", name="etm_oli_toa_nlcd")
```

Also, please see our [GFM-Bench](https://github.com/uiuctml/GFM-Bench) repository for more information about how to use the dataset! 🤗

## Dataset Metadata

The following metadata provides details about the Landsat imagery used in the dataset:
| Configuration Name       | Number of Bands | Number of Label Classes | Spatial Resolution |
|:---------------:|:------------:|:------------:|:------------:|
| etm_sr_cdl | 6 | 134 | 30 |
| etm_sr_nlcd | 6 | 21 | 30 |
| etm_toa_cdl | 9 | 134 | 30 |
| etm_toa_nlcd | 9 | 21 | 30 |
| oli_sr_nlcd | 7 | 134 | 30 |
| oli_sr_nlcd | 7 | 21 | 30 |
| oli_tirs_toa_cdl | 11 | 134 | 30 |
| oli_tirs_toa_nlcd | 11 | 21 | 30 |
| **etm_oli_toa_cdl** | 20 | 134 | 30 |
| **etm_oli_toa_nlcd** | 20 | 21 | 30 |

## Dataset Splits
The **NLCD-L** and SSL4EO-L Benchmark dataset consist following splits:
- **train**: 17,500 samples
- **val**: 3,750 samples
- **test**: 3,750 samples

## Dataset Features:
The **NLCD-L** and SSL4EO-L dataset consist of following features:
<!--- **radar**: the Sentinel-1 image.-->
- **optical**: the Landsat image.
- **label**: the segmentation labels.
<!--- **radar_channel_wv**: the central wavelength of each Sentinel-1 bands.-->
- **optical_channel_wv**: the central wavelength of each Landsat bands.
- **spatial_resolution**: the spatial resolution of images.
## Citation
If you use either the NLCD-L dataset or the original SSL4EO-L Benchmark dataset in your work, please cite the original paper:
```
@article{stewart2023ssl4eo,
  title={Ssl4eo-l: Datasets and foundation models for landsat imagery},
  author={Stewart, Adam and Lehmann, Nils and Corley, Isaac and Wang, Yi and Chang, Yi-Chia and Ait Ali Braham, Nassim Ait and Sehgal, Shradha and Robinson, Caleb and Banerjee, Arindam},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  pages={59787--59807},
  year={2023}
}
```
and if you also find our benchmark useful, please consider citing our paper:
```
@misc{si2025scalablefoundationmodelmultimodal,
      title={Towards Scalable Foundation Model for Multi-modal and Hyperspectral Geospatial Data}, 
      author={Haozhe Si and Yuxuan Wan and Minh Do and Deepak Vasisht and Han Zhao and Hendrik F. Hamann},
      year={2025},
      eprint={2503.12843},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.12843}, 
}
```