Hyperspectral anomaly detection via low-rank and sparse decomposition with cluster subspace accumulation

Area of Science:

• Remote Sensing
• Signal Processing
• Computer Vision

Background:

• Anomaly detection (AD) is crucial in hyperspectral imagery (HSI) processing.
• Traditional methods like low-rank and sparse matrix decomposition (LRaSMD) face challenges with background interference, noise, and complex targets.
• Existing algorithms often fail to fully leverage both spatial and spectral information inherent in HSI data.

Purpose of the Study:

• To develop an advanced AD method for HSI that overcomes limitations of traditional techniques.
• To enhance the distinction between background and anomalous targets while suppressing noise.
• To improve the stability and discriminative power of HSI analysis for more robust anomaly identification.

Main Methods:

• Segmentation of HSI into subspaces using k-means to reduce band redundancy.
• Application of the fractional Fourier transform (FrFT) within subspaces to enhance anomaly-target separation and noise reduction.
• Integration of low-rank and sparse matrix decomposition (LRaSMD) for improved HSI stability and discriminative power.
• Utilization of a modified Reed-Xiaoli (RX) detector for anomaly identification within subspaces, followed by aggregation of results.

Main Results:

• The proposed method achieved a high average area under the curve (AUC) of 0.9761 across five real HSI datasets.
• A low standard deviation of 0.0156 indicates consistent performance and robustness of the algorithm.
• Experimental results demonstrate superior performance compared to existing anomaly detection techniques in HSI.

Conclusions:

• The novel HSI anomaly detection method effectively integrates spatial and spectral features for superior performance.
• The combination of k-means, FrFT, LRaSMD, and modified RX detector offers a powerful approach to HSI anomaly detection.
• The proposed algorithm is highly competitive and provides a significant advancement in the field of hyperspectral anomaly detection.