Anomaly Detection for Hyperspectral Imagery Based on Active Learning with Support Vector Data Description

The Support Vector Data Description (SVDD) method for anomaly detection in hyperspectral imagery solved the problem of large numbers of false alarm in general detection methods based on statistical theory due to the Gaussian and homogeneous assumptions of background, but the background samples are selected randomly in SVDD. The active learning provides an effective sample selection method, therefore this paper presents Active Learning Support Vector Data Description (ALSVDD) method which is used to detect anomalies in hyperspectral imagery combing with neighboring clustering segmentation. ALSVDD method uses optimized minimal hypersphere to express the background tightly and distinguish function to detect anomalous pixels, which takes full advantage of the spatial and spectral information of the hyperspectral imagery. The method reduces the number of samples that is needed in the process of the algorithm and avoids the interference of possible anomalies in background. The experiments on the simulation data and AVIRIS data show the validity, efficiency and practicability of the proposed method which greatly reduces the computation complexity and false alarm rate in detecting anomalies in hyperspectral imagery.