Deep Unsupervised Blind Hyperspectral and Multispectral Data Fusion

WeChat: BatAug

Jiaxin Li, Ke Zheng, Jing Yao, Lianru Gao, and Danfeng Hong

Our paper is accpeted by IEEE Geoscience and Remote Sensing Letters (GRSL).

The early access version can be downloaded in my researchgate.

More information can be found in my Google Scholar Citations.

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

Directory structure

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

checkpoints

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.

data

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.

model

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

other five .py files

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

How to run our code

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.

References

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.

Contact

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

Jiaxin Li (:incoming_envelope: lijiaxin203@mails.ucas.ac.cn) is currently pursuing the Ph.D. degree in cartography and geographic information system with the Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China.

JiaxinLiCAS / UDALN_GRSL