JLSAE: Joint Low-Rank and Sparse Autoencoder for Hyperspectral Anomaly Detection

Official implementation of JLSAE: Joint Low-Rank and Sparse Autoencoder for Hyperspectral Anomaly Detection.

JLSAE jointly models the low-rank structure of the background and the sparse characteristics of anomalies within a unified autoencoder framework. It can be viewed as a further development of Auto-AD and DeepLR.

This repository provides a simple implementation for reproducing and comparing hyperspectral anomaly detection methods.

Framework

Overall framework of JLSAE.

Requirements

A standard PyTorch environment is sufficient to run the code.

The main required packages include:

Python
PyTorch
NumPy
SciPy
Matplotlib

A CUDA-enabled PyTorch environment is recommended for GPU acceleration.

Data Preparation

Place the test hyperspectral data in the data folder.

Each data file should be stored in MATLAB .mat format and contain the following variables:

• data: hyperspectral image data
• map: ground-truth anomaly map

An example directory structure is:

JLSAE/
├── data/
│   └── example.mat
├── models/
├── utils/
├── result/
└── train.py

The main folders are organized as follows:

• data/: stores the hyperspectral data sets used for testing.
• result/: stores the generated detection maps, AUC values, or other output results.
• models/: contains the network architecture of JLSAE.
• utils/: contains commonly used utility functions, such as data loading, ROC calculation, AUC calculation, and result visualization.
• figures/: contains the images used in this README, including the framework and detection-result figures.

Before running the code, modify the data file name in train.py:

path = 'data/'
file_name = 'example.mat'

Running the Code

Run the following command:

python train.py

Results

The detection results are automatically saved in the result folder.

The saved .mat file contains:

• detection: anomaly detection map
• auc: area under the ROC curve

For example:

result/example_result.mat

Parameter Settings

The following parameters can be adjusted in train.py:

Number of Iterations

num_iter = 150

num_iter controls the number of optimization iterations. The default value for a new data set is 150. It can be adjusted according to the convergence behavior and detection performance.

Recommended candidate values include:

[100, 150, 200, 250, 300]

Penalty Parameter

pho = 1

pho is the penalty parameter used in the optimization process. If the detection result is unsatisfactory, it can be tuned using values such as:

[1, 1.25, 1.5, 1.75]

Trade-Off Parameter

lam = 0.5

lam is the trade-off parameter associated with the sparse anomaly component. It can be adjusted for different data sets to obtain better detection performance.

Reproducibility

The network parameters and the input noise tensor Z are randomly initialized for each run. Therefore, the detection results and AUC values may vary slightly across different runs.

Detection Results

Original HSI

Ground Truth

Reconstructed Background

Anomaly Component E

Detection Map

Additional Information

Please contact fanjiahui24@mails.ucas.ac.cn if you have any questions about the code or its implementation. If you find this code helpful, please cite:

@ARTICLE{fan2026jlsae,
  author={Fan, Jiahui and Sun, Xiaotong and Wang, Degang and Sun, Xu and Gao, Lianru},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={JLSAE: Joint Low-Rank and Sparse Autoencoder for Hyperspectral Anomaly Detection}, 
  year={2026},
  volume={64},
  number={},
  pages={5522017-5522017},
  keywords={Ranking (statistics);Modeling;Anomaly detection;Optimization;Signal detection;Pixel;Probability;Educational institutions;Personal digital devices;Hyperspectral imaging;Autoencoder (AE);background reconstruction;hyperspectral anomaly detection (HAD);low-rank prior;model-driven deep learning;sparse prior},
  doi={10.1109/TGRS.2026.3716370}}