The code of the paper ‘Hyperspectral Image Super-Resolution Network Based on Cross-Scale Nonlocal Attention’.(TGRS)
2024/06/28:UPDATE: We have open source the real ziyuan hyperspectral datasets including the HSI/MSI/PAN images to contribute the hyperspectral image fusion field.
Include the hyperspectral image/multi-spectral image/pan image, which could be freely downloaded from the website:
https://aistudio.baidu.com/datasetdetail/281612
or from the baidunetdisk: link：https://pan.baidu.com/s/1kVO6KyE1HHJBtRjLqjOzTA?pwd=abcd password：abcd
or from the website: https://www.kaggle.com/datasets/shuangliangli123/real-hsimsipan-image-dataset/data
Note that this real dataset is acquired from the Ziyuan 1-02D satellite. and they have been spatially registered.
You could load these images and crop them by the code in the "save_image_ziyuan_hmp.py" as:
from save_image_ziyuan_hmp import generate_data, crop_data
hs, ms, pan, label = generate_data(ratio=12)
hs_crop, ms_crop, pan_crop, label_crop = crop_data(hs, ms, pan, label, training_size=16) # B×C×H×W
Note that the RGB image is shown by utilizing the 35th, 16th, and 7th bands of the original HSI, and the 3th, 2th, and 1th bands of the original MSI.
2023/08:UPDATE: We have open source the real ziyuan hyperspectral datasets used in this paper to contribute the hyperspectral image fusion field.
Include the hyperspectral image/pan image and hyperspectral image/multi-spectral image, which could be freely downloaded from the website:
hyperspectral image/pan image:https://aistudio.baidu.com/aistudio/datasetdetail/95831
hyperspectral image/multi-spectral image:https://aistudio.baidu.com/aistudio/datasetdetail/120965
The google drive link of datasets: https://drive.google.com/drive/folders/1oy2NaGmW-7MVZE5r42q3Zi_qUjcFHMDH?usp=drive_link
Note that these two real datasets are acquired by the Ziyuan 1-02D satellite. and they have been spatially registered, respectively.
You could load these two image datasets by the code in the "main_hmfusion.py" and "main_hpfusion.py" as:
main_hmfusion.py:
from save_image_ziyuan_reduce import generate_data, crop_data
name = 'zy'
ms, pan, label = generate_data(ratio=3)
ms_crop, pan_crop, label_crop = crop_data(ms, pan, label, training_size=16) # B×C×H×W
main_hpfusion.py
from save_image_ziyuan_hp import generate_data, crop_data
name = 'zy'
ms, pan, label = generate_data(ratio=12)
ms_crop, pan_crop, label_crop = crop_data(ms, pan, label, training_size=16) # B×C×H×W
And remember to adjust the train and test patch number for the different "training_size".
The fusion between the hyperspectral and multi-spectral images could be conducted by running the ‘main_hmfusion.py’.
The fusion between the hyperspectral and PAN images could be conducted by running the ‘main_hpfusion.py’.
Note that the main network difference between these two fusion taks is the training strategy, including optimizer, training epochs.
If you have any questions, please feel free to contact me.
Please consider cite the paper if you find it helpful.
Email:whu_lsl@whu.edu.cn