A feature selection method for multimodal multispectral sensing

This repository is related to the paper "A feature selection method for multimodal multispectral sensing", which includes:

The dataset collected from 30 material specimens in 4 material classes by using our multimodal multispectral (MM) LiDAR prototype
The proposed multiclass group feature selection algorithm based on an all-in-one support vector machine (MGSVM FS)
The 10-fold cross-validation as the evaluation framework and three evaluation metrics (i.e., F1 score, stability index, and selection probability) for decision-making
The ablation study comparing MGSVM FS with three other feature selection algorithms (i.e., random FS, RF-MDPA FS, and MRMR FS)

* Environment setup

The list of required packages:

    cvxpy==1.4.3 
    matplotlib==3.6.2
    mrmr-selection==0.2.8
    pandas==2.0.3
    scikit-learn==1.1.3
    scipy==1.9.3
    seaborn==0.12.2
    numpy==1.23.3

1. MM feature dataset

The dataset of the MM features collected from the material collection can be found in the folder Dataset_MM_features.

To plot the MM features (as shown in Fig. 7 of the paper), please implement the following command.

    python MM_features_plotting.py

2. Feature selection algorithm: MGSVM FS

Multiclass group feature selection algorithm based on an all-in-one support vector machine

Embedded feature selection
Group feature selection / Group LASSO for structured data / Structural sparsity
Multiclass classification / All-in-one SVM

Example command -- Plot the output of selected features from MGSVM FS

Please refer to the subfolder FS_algorithm\MGSVM FS.

3. Evaluation framework and metrics

The evaluation framework

    10-fold cross-validation

Three evaluation metrics

    1. F1 score
    2. Kuncheva's stability index
    3. Selection probability

Options of modality combinations

    1. DoLP
    2. R
    3. d
    4. R+d
    5. R+DoLP
    6. d+DoLP
    7. R+d+DoLP
    
    R: Reflectance
    d: Distance
    DoLP: Degree of linear polarization

Target number of selected spectral channels (feature groups)

     1 --> 28

Example command -- Evaluate the feature selection algorithms for arbitrary modality combinations and save the evaluation metrics

Different modality combinations can be set in the file Evaluation_feature_selection_methods.py.

The corresponding evaluation metrics can be output by implementing the following command to run the evaluation framework.

    python Evaluation_feature_selection_methods.py

4. Ablation study

Four feature selection algorithms

    1. Random feature selection 
    2. RF-MDPA feature selection
    3. MRMR feature selection
    4. MGSVM feature selection

Example command -- Plot the final evaluation metrics

All evaluation metrics for all modality combinations can be found in the subfolder Results_evaluation_metrics.

To plot the evaluation metrics (as shown in Figs. 8-10 of the paper), please implement the following command.

    python Evaluation_metrics_plotting.py

* Citation

If you find our dataset and/or code useful, we appreciate your consideration in citing our publication.

    @article{HAN202442,
    title = {A feature selection method for multimodal multispectral LiDAR sensing},
    journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
    volume = {212},
    pages = {42-57},
    year = {2024},
    issn = {0924-2716},
    doi = {https://doi.org/10.1016/j.isprsjprs.2024.04.022},
    author = {Yu Han and David Salido-Monzú and Jemil Avers Butt and Sebastian Schweizer and Andreas Wieser}
    }

yuhan-yhyh/Dataset_Code_MGSVM-FS-MM-LiDAR