---
title: Latent_Spectral_Operators_Cave_and_Harvard
canonical_url: "https://www.modelscope.cn/datasets/WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard"
md_url: "https://www.modelscope.cn/datasets/WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard.md"
repository: WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard
last_updated: 2026-05-26
license: "MIT License"
storage_size: "22 GB"
downloads: 249
stars: 0
---

# Latent_Spectral_Operators_Cave_and_Harvard

> Latent_Spectral_Operators_Cave_and_Harvard - WeiLi419 在 ModelScope 开源的数据集。The dataset used in the paper Solving Spatial-Spectral Fusion with Latent Spectral Operators.

WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard 是 ModelScope 魔搭社区上的数据集，存储大小 22 GB，采用 MIT License 许可。

- **Repository**: WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard
- **License**: MIT License
- **Storage size**: 22 GB
- **Downloads**: 249
- **Stars**: 0
- **Last updated**: 2026-05-26

Source: https://www.modelscope.cn/datasets/WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard

---

# Latent Spectral Operators - CAVE and Harvard

This repository provides the processed CAVE and Harvard hyperspectral image (HSI) data used to evaluate Latent Spectral Operators. Each sample is simulated from a ground-truth high-resolution HSI (HR-HSI) and paired with low-resolution HSI (LR-HSI) and high-resolution MSI (HR-MSI) observations.

Code repository: [weili419/LSO](https://github.com/weili419/LSO)

The full LFS payload is approximately **22.44 GiB** (**24.10 GB**).

## Dataset Overview

We evaluate on the CAVE and Harvard datasets.

- **CAVE** contains 32 HSIs. Each image has 31 spectral bands covering 400-700 nm at a 10 nm interval. Following prior work, 21 images are used for training and the remaining 11 images are used for testing.
- **Harvard** contains 77 HSIs covering indoor and outdoor scenes. Each image has a spatial resolution of 1392 x 1040 and 31 spectral bands in the 420-720 nm range. We select 20 images, crop the upper-left 1000 x 1000 region, and split them evenly into 10 training images and 10 testing images.

![Visualization examples from the CAVE and Harvard datasets](assets/vis.png)

## Simulation Protocol

The model takes a pair of LR-HSI and HR-MSI observations, denoted as `(I_LR, I_HR)`, and predicts an HR-HSI output `O_HR` for supervision. Since real paired LR-HSI, HR-MSI, and HR-HSI measurements are unavailable, the dataset follows a standard simulation protocol.

For CAVE, the 21 training images are cropped into 3,920 overlapping ground-truth patches of size 64 x 64 x 31. LR-HSI inputs are synthesized by blurring each HR-HSI patch with a Gaussian kernel and downsampling by a factor of 4 or 8. HR-MSI inputs are generated by applying the Nikon D7003 spectral response function to the HR-HSI patches. This produces LR-HSI patches of size 16 x 16 x 31 for the x4 setting and HR-MSI patches of size 64 x 64 x 3. The simulated samples are split into 80% training and 20% validation. The Harvard data follows the same simulation procedure.

## Ground-Truth Shapes

| Dataset | Train GT shape | Val GT shape | Test GT shape |
| --- | --- | --- | --- |
| CAVE | `(3136, 31, 64, 64)` | `(784, 31, 64, 64)` | `(11, 31, 512, 512)` |
| Harvard | `(3136, 31, 64, 64)` | `(784, 31, 64, 64)` | `(10, 31, 1000, 1000)` |

## Repository Structure

```text
.
|-- assets/
|   `-- vis.png
|-- cave/
|   |-- train_cave(with_up)x4.h5
|   |-- validation_cave(with_up)x4.h5
|   `-- test_cave(with_up)x4.h5
`-- harvard/
    |-- train_harvard(with_up)x4.h5
    |-- validation_harvard(with_up)x4.h5
    `-- test_harvardv3(with_up)x4.h5
```

## Files

| Folder | Split | File | Size |
| --- | --- | --- | --- |
| `cave` | Train | `cave/train_cave(with_up)x4.h5` | 6.41 GiB |
| `cave` | Validation | `cave/validation_cave(with_up)x4.h5` | 1.60 GiB |
| `cave` | Test | `cave/test_cave(with_up)x4.h5` | 1.44 GiB |
| `harvard` | Train | `harvard/train_harvard(with_up)x4.h5` | 6.41 GiB |
| `harvard` | Validation | `harvard/validation_harvard(with_up)x4.h5` | 1.60 GiB |
| `harvard` | Test | `harvard/test_harvardv3(with_up)x4.h5` | 4.99 GiB |

## Download

You can download the dataset from ModelScope:

:modelscope-code[]{type="sdk"}
:modelscope-code[]{type="git"}

The files are stored with Git LFS. To clone the repository without immediately downloading all large files:

```bash
GIT_LFS_SKIP_SMUDGE=1 git clone https://www.modelscope.cn/datasets/WeiLi419/Latent_Spectral_Operators_Cave_and_Harvard.git
cd Latent_Spectral_Operators_Cave_and_Harvard
```

Fetch all HDF5 files:

```bash
git lfs pull
```

Or fetch a single subset, for example the CAVE files:

```bash
git lfs pull --include="cave/*.h5"
```

## Inspecting HDF5 Files

The internal HDF5 keys, shapes, and dtypes can be inspected with `h5py` after downloading the target file:

```python
import h5py

path = "cave/train_cave(with_up)x4.h5"

with h5py.File(path, "r") as f:
    def visit(name, obj):
        if isinstance(obj, h5py.Dataset):
            print(f"{name}: shape={obj.shape}, dtype={obj.dtype}")
        else:
            print(f"{name}/")

    f.visititems(visit)
```

## License

This dataset repository is released under the MIT License.

## Citation and Contact

If this repository is useful for your research, please cite:

```bibtex
@inproceedings{li2026lso,
  title     = {Solving Spatial-Spectral Fusion with Latent Spectral Operators},
  author    = {Li, Wei and Pei, Jieyuan and Xu, Junnan and Ding, Xuanfeng and Zhu, Junwei and Chen, Wanjun and Zheng, Jianwei},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  publisher = {PMLR},
  year      = {2026}
}
```

For questions, please contact the dataset maintainer through the ModelScope repository page.