Deep Generative Endmember Modeling: An Application to Unsupervised Spectral Unmixing

This package contains the authors' implementation of the paper [1].

In order to address endmember variability in spectral unmixing, the endmember spectra are modelled using a deep generative model (a VAE), which is learned from the observed hyperspectral image. This allows us to parametrize the variable endmembers using points in the low-dimensional latent space of the generative models, which can then be simultaneously optimized with the abundances to solve the unmixing problem.

The code is implemented in MATLAB and includes:

• example1.m - a demo script comparing the algorithms (DC1)
• example2.m - a demo script comparing the algorithms (DC2)
• example3.m - a demo script comparing the algorithms (DC3)
• example4.m - a demo script comparing the algorithms (DC4)
• example_real1.m - a demo script comparing the algorithms (Houston)
• example_real2.m - a demo script comparing the algorithms (Samson)
• example_real3.m - a demo script comparing the algorithms (Jasper Ridge)
• ./DeepGUn/ - contains the MATLAB files associated with the DeepGUn algorithm
• ./python/ - contains the Python files associated with the DeepGUn algorithm
• ./other_methods/ - contains the ELMM, GLMM and PLMM methods
• ./utils/ - useful functions
• ./data/ - images used in the examples
• README - this file

UPDATE:

Changed network training and matlab loading, now compatible with Tensorflow 2 and Keras 2.4.

IMPORTANT:

If you use this software please cite the following in any resulting publication:

[1] Deep Generative Endmember Modeling: An Application to Unsupervised Spectral Unmixing
    R.A. Borsoi, T. Imbiriba, J.C.M. Bermudez.
    IEEE Transactions on Computational Imaging, 2019.

INSTALLING & RUNNING:

Just start MATLAB and run one of the demo scripts (e.g. example1.m, example2.m, etc).

Requirements

The variational autoencoders are trained in Python using the Keras package.

NOTES:

1. The ELMM algorithm was provided by Lucas Drumetz.
   Drumetz, L., Veganzones, M.-A., Henrot, S., Phlypo, R., Chanussot, J., & Jutten, C. Blind hyperspectral unmixing using an extended linearmixing model to address spectral variability. IEEE Transactions on Image Processing, 2016.

2. The GLMM algorithm was provided by Tales Imbiriba at https://github.com/talesimbiriba/GLMM.
   Imbiriba, T., Borsoi, R.A. & Bermudez J.C.M. Generalized linear mixing model accounting for endmember variability. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018.

3. The PLMM algorithm was provided by Pierre-Antoine Thouvenin.
   Thouvenin, P.-A., Dobigeon, N., & Tourneret, J.-Y. Hyperspectral unmixing with spectral variability using a perturbed linear mixing model. IEEE Transactions on Signal Processing, 2016.

4. The minFunc optimization toolbox was provided by Mark Schmidt and can be found at
   http://www.cs.ubc.ca/~schmidtm/Software/minFunc.html