Deep Unsupervised Blind Hyperspectral and Multispectral Data Fusion, GRSL. (PyTorch)

Jiaxin Li 李嘉鑫, Ke Zheng 郑珂, Jing Yao 姚靖, Lianru Gao 高连如, and Danfeng Hong 洪丹枫, IEEE Geoscience and Remote Sensing Letter (GRSL).

文章可在这里下载🖼️PDF，The final version can be downloaded in 🖼️PDF

我的英文版本个人简历可在隔壁仓库下载，如您需要此简历模板可以通过微信联系我。 My english CV can be downloaded in this repository .

2025.09——, 就职于重庆邮电大学 计算机科学与技术学院 文峰副教授 $\color{red}{博后导师：韩军伟教授}$ 【官网，谷歌学术主页】

2020.09-2025.07 就读于中国科学院 空天信息创新研究院 五年制直博生 $\color{red}{导师：高连如研究员}$ 【导师空天院官网，谷歌学术主页】

2016.09-2020.7 就读于重庆大学 土木工程学院 测绘工程专业

From 2025.09, I work at the School of Computer Science and Technology (National Exemplary Software School), Chongqing University of Posts and Telecommunications, as a Wenfeng associate professor. My postdoctoral supervisor is Junwei Han.

From 2020.09 to 2025.07, I am a PhD candidate at the Key Laboratory of Computational Optical Imaging Technology, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. My supervisor is Lianru Gao.

From 2016.0 to 2020.7, I studied in the school of civil engineering at Chongqing University, Chongqing, China, for a Bachelor of Engineering.

这是我的谷歌学术和ResearchGate，More information can be found in my Google Scholar Citations and my ResearchGate

🖼️遇到任何问题，包括但不限于代码调试、数据仿真、运行结果等，随时添加 $\color{red}{我的微信(WeChat): BatAug，欢迎交流与合作}$

Fig.1. Architecture of the proposed unsupervised degradations adaptive learning network, abbreviated as UDALN, for the task of HSI-MSI fusion.

Fig.2. Directory structure. There are three folders and six .py files in UDALN_GRSL-master.

This folder is used to store the training results and a folder named houston18_5_S1=0.001_20000_10000_S2=0.001_30000_20000_S3=6e-05_15000_5000 is given as a example.

• convolution_hr2msi.pth is the trained result of SpeDnet, PSF.pth is the trained result of SpaDnet, and spectral_upsample.pth is the trained result of SpeUnet.

• opt.txt is used to store the training configuration.

• precision.txt is used to store the training precision.

• My_Out.mat is the final reconstructed HHSI.

This folder is used to store the ground true HHSI and corresponding spectral response of multispectral imager. The HSI data used in 2018 IEEE GRSS Data Fusion Contest and spectral response of WorldView 2 multispectral imager are given as a example here.

This folder consists four .py files, including spatial_downsample.py(SpaDnet), spectral_downsample.py(SpeDnet), spectral_upsample.py(SpeUnet), and __init__.py.

• config.py: all the parameters in our methed.

• Data_loader.py: generate the simulated low HSI and high MSI.

• evaluation.py: compute five metrics, which will be stored in precision.txt.

• func.py: some functions used in train_all_special.py.

• train_all_special.py: main.py

• Requirements: codes of networks were tested using PyTorch 1.9.0 version (CUDA 11.4) in Python 3.8.10 on Windows system. For the required packages, please refer to detailed .py files.

• Parameters: all the parameters need fine-tunning can be found in config.py, including the learning rate decay strategy of three training stages.

• Data: put your HSI data and MSI spectral reponse in ./data/data_name and ./data/spectral_response, respectively.The HSI data used in 2018 IEEE GRSS Data Fusion Contest and spectral response of WorldView 2 multispectral imager are given as a example here.

• Run: just simply run train_all_special.py after adjust the parameters in config.py.

• Results: one folder named dataname_SF_S1=x1_y1_z1_S2=x2_y2_z2_S3=x3_y3_z3 will be generated once train_all_special.py is run and all the results will be stored in the new folder. A folder named houston18_5_S1=0.001_20000_10000_S2=0.001_30000_20000_S3=6e-05_15000_5000 is given as a example here.

Our work is inspired by the following paper

[1] Zheng, Ke, et al. "Coupled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super-resolution." IEEE Transactions on Geoscience and Remote Sensing (2020), DOI: 10.1109/TGRS.2020.3006534.

[2] Yao, Jing, et al. "Cross-attention in coupled unmixing nets for unsupervised hyperspectral super-resolution." In Proceedings of the European Conference on Computer Vision (ECCV) (2020), pp. 208-224.

[3] Han, Xiaolin, et al. "Hyperspectral and Multispectral Image Fusion Using Cluster-Based Multi-Branch BP Neural Networks" Remote Sensing (2019), DOI: 10.3390/rs11101173.

If you encounter any bugs while using this code, please do not hesitate to contact us.

🖼️遇到任何问题，包括但不限于代码调试、数据仿真、运行结果等，随时添加 $\color{red}{我的微信(WeChat): BatAug，欢迎交流与合作}$