Fast Real-Time Causal Linewise Progressive Hyperspectral Anomaly Detection via Cholesky Decomposition

Real-time processing of anomaly detection has become one of the most important issues in hyperspectral remote sensing. Due to the fact that most widely used hyperspectral imaging spectrometers work in a pushbroom fashion, it is necessary to process the incoming data line in a causal linewise progressive manner with no future data involved. In this study, we proposed several processes to well improve the computational performance of real-time causal linewise progressive anomaly detection (RCLPAD). At first, Cholesky decomposition along with linear system solving (CDLSS) was used since the background statistical matrix are symmetric positive definite. The computational performance as well as the numerical stabilities is well improved. In order to show the computational advantage of the proposed method, we did a comprehensive comparative analysis regarding the computational complexity of different linewise processing techniques, in terms of the theoretical floating point operations (flops) and the real computer processing time. Moreover, the symmetric property of some intermediate resultingmatrices in the process is considered for further computational optimization. Finally, from an onboard detection point of view, we defined the line-varying global background (i.e., an area covered by recently acquired data lines) to improve the detection power. To substantiate the performance of the CDLSS-based RCLP-AD regarding the accuracy and efficiency, two hyperspectral datasets were used in our experiments.