Kernel-Band-Projection Algorithm for Anomaly Detection in Hyperspectral Imagery

• Published in: International Conference on Signal Processing (Print) pp. 300 - 303
• Main Authors: Yao, Xifeng, Zhao, Chunhui
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.08.2018
• Subjects: Anomaly detection; Covariance matrices; Detectors; fast algorithms; hyperspectral image; Hyperspectral imaging; Kernel; kernel band projection; Real-time systems
• ISSN: 2164-5221
• DOI: 10.1109/ICSP.2018.8652278

The widely-used RX detector (RXD) that was proposed by Reed and Yu has been considered as a benchmark in hyperspectral anomaly detection. Kernel RX detector (KRXD), the nonlinear version of RXD, improves the detection accuracy by mapping the inputting pixel vectors into kernel space. However, this processing style with pixels being the mapping subject brings three limitations. 1) More computing time is required for KRXD due to the thorough data kernelization; 2) the sizes of kernelized processing terms are related to the number of background pixels, which limits global processing and real-time design of KRXD; 3) the inverse of kernelized background covariance matrix is usually singular. Therefore, this paper proposes an anomaly detector with bands being the mapping subject. By mapping bands into the kernel space to construct projection matrix in which the Euclidean distance is implemented, comparable detection accuracy with KRXD can be achieved. The inverse operation of kernel projection matrix can be also avoided when the Gaussian kernel function is used. Therefore, the above limitations can be addressed. Experimental results demonstrate the effectiveness of the proposed algorithm.