Entropic Gradient Archetypal Analysis for Blind Hyperspectral Unmixing

Official PyTorch implementation of the paper Entropic Gradient Archetypal Analysis for Blind Hyperspectral Unmixing

Introduction

This repository mainly implements the entropic gradient archetypal analysis method for blind hyperspectral unmixing.

In addition, we include various unmixing methods (supervised and blind) on several standard real datasets.

Citing EDAA

The authors kindly ask users to cite EDAA in their publications if they use it in their research:

A. Zouaoui, G. Muhawenayo, B. Rasti, J. Chanussot, J. Mairal, "Entropic descent archetypal analysis for blind hyperspectral unmixing", in IEEE Transactions on Image Processing, doi: 10.1109/TIP.2023.3301769

Installation

This repository was developed using Ubuntu 20.04 LTS, Python 3.8.8 and MATLAB 2020b.

We recommend using conda to handle the Python distribution and pip to install the Python packages.

conda create --name edaa --file conda.txt

Activate the new conda environment to install the Python packages and run the code:

conda activate edaa 

git clone https://github.com/inria-thoth/EDAA
cd EDAA && pip install -r requirements.txt

---

Run matlab code in Python

It is possible to run the NMF-QMV code inside Python to make sure that the reported results are consistent across methods.

To install matlab.engine, first visit the official webpage.

1. Locate matlabroot (e.g. ~/softwares/matlab-2020b/)
2. Create a matlab build directory (e.g. ~/matlab/)
3. Locate your current conda environment (e.g. ~/conda/envs/hsu)

cd $matlabroot/extern/engines/python
conda activate edaa
python setup.py build -b ~/matlab install --prefix ~/conda/envs/edaa

---

Data download

The data can be downloaded from http://pascal.inrialpes.fr/data2/azouaoui/data.zip.

To extract it in the appropriate ./data directory, simply run:

python -m utils.prepare_datasets

If for whatever reasons you were not able to use the script above:

1. Download the file directly from the link.
2. Create a new folder ./data in the root folder of this repository.
3. Extract data.zip into ./data.
4. You should have all 7 datasets directly available in the ./data folder.

---

Getting started

This repository uses hydra to seamlessly manage different configurations over the command line.

To run the EDAA model on the Samson dataset, use the following command:

python main.py dataset=Samson mode=blind model=BlindEDAA

The different datasets available are listed under ./hsi_unmixing/config/dataset:

• Samson
• JasperRidge
• Urban{4,5,6}
• TinyAPEX
• WDC

To run the EDAA model on all datasets, use the following command:

python main.py dataset=Samson,JasperRidge,Urban4,Urban6,TinyAPEX,WDC mode=blind model=BlindEDAA --multirun

---

Other methods

This repository implements other supervised and blind hyperspectral unmixing techniques:

• Supervised methods
  • FCLSU using VCA (Python implementation) and DecompSimplex (DS) from SPAMS .
  • FCLSU using VCA (Python implementation) and FCLS from pysptools (slower).

Command to run FCLSU using VCA+DS on Samson:

python main.py dataset=Samson mode=supervised torch=False model=DS initializer=VCA

• Blind methods
  • NMF-QMV using a Python wrapper on the Matlab implementation.
  • Archetypal Analysis (AA) from SPAMS.
  • Robust AA (RAA) from SPAMS.

Command to run AA on Samson:

python main.py dataset=Samson mode=blind torch=False model=AA