SVDD Algorithm Based on Noise Cost for Anomaly Detection in Hyperspectral Imagery

Abstract

Anomaly detection algorithm based on Support Vector Data Description (SVDD) brings low detection rates due to background training samples being contaminated by anomalous data. To solve the problem, a new method based on SVDD with Noise Cost is proposed by introducing unbalanced data mining cost sensitive mind. This algorithm gives a different noise cost value to each background training samples through the neighbourhood clustering and then introduces the noise cost into SVDD to construct the SVDD hypersphere, thus making the classification interface more compact and improving the description ability of the anomaly and background value. At the same time, the sensitivity to the abnormal algorithm and the detection probability of the algorithm are greatly improved. Experimental results based on simulation data show that: compared to SVDD, this algorithm greatly reduces the false alarm rate, and improves the detection precision.