Anomaly Detection for Hyperspectral Imagery Based on Incremental Support Vector Data Description

Abstract

This paper presented incremental support vector data description (ISVDD) method and used it to detect anomalies in hyperspectral images. Anomaly detection is essentially a problem of one-class classification, a good solution of which is SVDD, using optimized minimal hypersphere to express tightly the background and using distinguish function to detect anomalous pixels. The method avoided the problem that general detect method based on statistical theory make large numbers of false alarms due to the assumptions that background is Gaussian and homogeneous. High dimension character of hyperspectral imagery increased the operation amount, while proposed ISVDD reduced the operation amount multiply and reduced the interference of background to decrease numbers of false alarms. The experiment on the simulation data shows the validity and practicability of the method and the performance of anomaly detection exceeded obviously SVDD method.