Two-Orientations Finite Markov Real-Time Local Anomaly Detection via Pixel-by-Pixel Processing for Hyperspectral Imagery

Real-time local anomaly detection needs to be performed simultaneously with hyperspectral image acquisition. However, the discussion on the scope of using existing data, mainly focuses on line-by-line data acquisition methods and does not address pixel-by-pixel acquisition methods. In addition, localized information from a single line has a limited impact on the pixels captured by the whisk broom hyperspectral imaging sensor. Therefore, reasonable utilization of local information from different directions around the sample data to be measured is key to improving the detection performance of pixel-by-pixel processing. Based on the whisk broom hyperspectral imaging sensor data acquisition, the new pixel-by-pixel local correlation matrix R-based anomaly detection real-time processing method is proposed in this article. A two-orientations finite Markov model is established, which uses bidirectional local information to judge the anomaly of current pixels. Based on the evaluation of the current pixels, two distinct local sample correlation matrix R are designed, and two different recursive equations for correlation matrix R anomaly detection are derived. This approach enables real-time pixel-by-pixel processing, reduces the need for large amounts of data, and enhances detection efficiency. Neither version of real-time RAD has been studied in the past. According to the experimental findings, the proposed algorithm outperforms others in real-time anomaly detection across various real hyperspectral image datasets.