Hyperspectral Band Selection using Self-Representation Learning with Sparse 1D-Operational Autoencoder (SRL-SOA)

This repository includes the implentation of the hyperspectral image (HSI) band selection method in SRL-SOA: Self-Representation Learning with Sparse 1D-Operational Autoencoder for Hyperspectral Image Band Selection.

Software environment:

# Python with the following version and libraries.
python == 3.7.9
tensorflow == 2.4.1
Numpy == 1.21.2
argparse == 1.1
SciPy == 1.6.2
scikit-learn == 0.23.2
scikit-feature == 1.1.0 # For ISSC method (pip install skfeature-chappers==1.1.0)
munkres == 1.1.4 # For EGCSR method (pip install munkres==1.1.4)

Creating the environment using conda:

conda create -n srl-soa python=3.7
conda activate srl-soa
conda install tensorflow-gpu=2.4.1
conda install scikit-learn=0.23.2
pip install skfeature-chappers==1.1.0
pip install munkres==1.1.4

Citation

If you use the provided method in this repository, please cite the following paper:

@INPROCEEDINGS{ahishali2022srlsoa,
  author={Ahishali, Mete and Kiranyaz, Serkan and Ahmad, Iftikhar and Gabbouj, Moncef},
  booktitle={2022 IEEE International Conference on Image Processing (ICIP)}, 
  title={SRL-SOA: Self-Representation Learning with Sparse 1D-Operational Autoencoder for Hyperspectral Image Band Selection}, 
  year={2022},
  pages={2296-2300},
  doi={10.1109/ICIP46576.2022.9897863}}

Downloading the HSI Datasets

In the experiments, the Indian Pines and Salinas-A HSI datasets are used and they are accessible from here.

Downloading the Indian Pines, Salinas-A scenes, and corresponding ground-truth data:

wget http://www.ehu.eus/ccwintco/uploads/6/67/Indian_pines_corrected.mat
wget http://www.ehu.eus/ccwintco/uploads/c/c4/Indian_pines_gt.mat
wget http://www.ehu.eus/ccwintco/uploads/1/1a/SalinasA_corrected.mat
wget http://www.ehu.eus/ccwintco/uploads/a/aa/SalinasA_gt.mat

Place them under the data folder:

mkdir data
mv *mat data/

Band Selection using SRL-SOA

The proposed SRL-SOA method can be run for the band selection as follows,

python main.py --dataset Indian_pines_corrected --method SRL-SOA --q 3 --bands 25

You can specify the dataset Indian_pines_corrected or SalinasA_corrected, polynomial order (see Operational Layers) --q, and the selected number of bands --bands. The weights of the SRL-SOA are already provided under weights/ folder to reproduce the reported results in the paper. If you desire to train them from scratch, --weights argument should be set to False:

python main.py --dataset Indian_pines_corrected --method SRL-SOA --weights False --q 3 --bands 25

Note that once the network training is finished, the same network can be tested using different number of bands (--bands) with the same q value. Hence, you only re-train the network if you change the q value.

After the band selection, the classification is performed using SVM classifier and the SVM parameters are searched via the grid search. The parameter search can be disabled by setting parameterSearch = False in main.py.

If you desire to skip the band selection procedure, set --method None. In this way, the classifier uses all the available bands:

python main.py --dataset Indian_pines_corrected --method None 

Compare with Other Band Selection Methods

For the performance comparison with SRL-SOA, the provided main.py supports the following band selection methods: Principal Component Analysis (PCA), Sparse Representation based Band Selection (SpaBS) [1], Efficient Graph Convolutional Self-Representation (EGCSR) [2], and Improved Sparse Subspace Clustering (ISSC) [3]. Similarly, they can be evaluated by setting the following arguments: --method, --dataset (Indian_pines_corrected or SalinasA_corrected), and number of bands by --bands. Using PCA method:

python main.py --method PCA --dataset Indian_pines_corrected --bands 25

On the other hand, the following repositories are needed for SpaBS, EGCSR, and ISSC:

git clone https://github.com/AngryCai/BandSelection.git
git clone https://github.com/AngryCai/GCSR-BS
mv -f GCSR-BS/ GCSR_BS/

Then, similarly, the evaluation of the methods can be performed:

python main.py --method SpaBS --dataset Indian_pines_corrected --bands 25
python main.py --method EGCSR_R --dataset Indian_pines_corrected --bands 25
python main.py --method ISSC --dataset Indian_pines_corrected --bands 25

References

[1] S. Li and H. Qi, "Sparse representation based band selection for hyperspectral images," in IEEE Int. Conf. Image Process., 2011, pp. 2693–2696.
[2] Y. Cai, Z. Zhang, X. Liu, and Z. Cai, "Efficient graph convolutional self-representation for band selection of hyperspectral image," IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 13, pp. 4869–4880, 2020.
[3] W. Sun, L. Zhang, B. Du, W. Li, and Y. M. Lai, "Band selection using improved sparse subspace clustering for hyperspectral imagery classification," IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 8, no. 6, pp. 2784–2797, 2015.