Sparse Unmixing using Archetypal Analysis (SUnAA) is an innovative semi-supervised unmixing technique that leverages the assumption that real endmembers can be represented as convex combinations of the library endmembers. Building upon the principles of archetypal analysis, we propose a new model formulation for sparse unmixing. The motivation behind this method is to address mismatches commonly encountered between endmembers from spectral libraries and endmembers from a dataset, which often manifest as scaling factors. These mismatches can arise from various sources such as noise, atmospheric effects, illumination variations, and the intrinsic variability of materials. In contrast to conventional sparse unmixing methods, SUnAA estimates both the endmembers and abundances by utilizing the endmember library as a reference.
If you use this code please cite the following paper
B. Rasti, A. Zouaoui, J. Mairal and J. Chanussot, "SUnAA: Sparse Unmixing using Archetypal Analysis," in IEEE Geoscience and Remote Sensing Letters, doi: 10.1109/LGRS.2023.3284221.
- Clone the repository and move to the repository
git clone [email protected]:BehnoodRasti/SUnAA.git
cd SUnAA- We recommend using
condato install the package, as follows:
conda create --name sunaa python=3.10- Activate your freshly created environment
conda activate sunaa- Install the required python packages:
pip install -r requirements.txt- In case you face some issues regarding the installation of
spams, we recommend installing it from source using the instructions found on the official PyPI package website. - For windows users, we suggest removing the line 4 in the requirements.txt (spams==2.6.5.4) and after installing the requirements, install spams using
pip install spams-bin.
We provide a simple demo script alongside the data to run it (DC1 in our paper).
You can run it at the desired SNR (here 20) using the following command:
python demo.py --SNR 20Take a look at the demo.py file to better understand the parameters that can be tweaked as well as changing the dataset to fit your needs.