demo

SpectralSpatial Out-of-Distribution-Based Unsupervised Band Selection Method for Hyperspectral Anomaly Detection

Paper

[Title] SpectralSpatial Out-of-Distribution-Based Unsupervised Band Selection Method for Hyperspectral Anomaly Detection

[Journal] IEEE Transactions on Geoscience and Remote Sensing (TGRS)

[DOI] https://doi.org/10.1109/TGRS.2024.3493879

---

Acknowledgement

This code is built upon the DP-GMM-HSI project developed by zsyOAOA.

---

Overview

This repository contains the implementation of HADUBS (Hyperspectral Anomaly Detection using Out-of-Distribution-Based Unsupervised Band Selection), an unsupervised band selection method for hyperspectral image (HSI) anomaly detection.

The method leverages:

• Hierarchical Dirichlet Process (HDP) for spectral denoising
• LOF (Local Outlier Factor) for spatial anomaly detection
• Mutual Information for optimal band ranking

---

Installation

Dependencies

• MATLAB (R2018b or later recommended)
• PROPACK Toolbox - Required for lansvd function
  • Download from: https://github.com/jedlofh/propack/releases
  • Add to MATLAB path: addpath('path/to/PROPACK');

File Structure

Abnormal_band_selection/
├── demo.m                     % Main entry point
├── set_opt_HSI.m             % HDP parameter configuration
├── hdp_denoise.m            % HDP denoising algorithm
├── inexact_alm_rpca.m        % RPCA (Robust PCA) algorithm
├── LOF.m                     % Local Outlier Factor implementation
├── Mi_my.m                   % Mutual Information computation
└── common/                   % Common utility functions
    ├── choosvd.m
    ├── PCA.m
    ├── img_quality_HSI.m
    ├── ErrRelGlobAdimSyn.m
    ├── ssim_index.m
    ├── psnr_img.m
    └── logsumexp.m

---

Usage

1. Configure data path in demo.m:

   addpath(genpath('YOUR_DATA_PATH'));
   load forest.mat  % or your HSI data

2. Run the main algorithm:

   demo

Key Parameters

Parameter	Default	Description
area_threshold	25	Connected component area threshold
quantile_upper	0.96	Upper quantile threshold for LOF
lof_k	3	K neighbors for LOF algorithm
rank_r	3	Rank of low-rank component
initT_hdp	7	Initial number of local MoG components
initK_hdp	15	Initial number of global MoG components

---

Data Format

The code expects a .mat file containing:

• data: HSI data cube of shape [Height, Width, Bands]
• map (optional): Ground truth map or mask

---

Citation

If you find this code useful in your research, please cite our paper:

@ARTICLE{HADUBS2024,
  author = {Authors},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  title = {SpectralSpatial Out-of-Distribution-Based Unsupervised Band Selection Method for Hyperspectral Anomaly Detection},
  year = {2024},
  doi = {10.1109/TGRS.2024.3493879}
}

---

License

[To be determined - please specify your license]

---

Contact

[Your contact information]