Anomaly detection in multi-band spectral imagery based on multiple-detector maximum entropy fusion

Anomaly detection in multi-band spectral imagery using single detector has great deficiency due to the model limitation, thus a multiple-detector fusion algorithm is proposed in this paper for this problem. Several different anomaly detectors are selected for obtaining pilot detection results, then a nonparametric method called Kernel Density Estimation (KDE) with bandwidth adjusted adaptively is used to estimate the Probability Density Function (PDF) statistics of the output of each individual detector, which preserves the long-tail behavior of multi-band spectral imagery to avoid the model error. The obtained probabilistic information are then transformed to a space with standard Gaussian marginal distribution, in which optimal probabilistic fusion of multi-detector on the decision level is accomplished utilizing a modeled joint distribution under maximum entropy principle. Target detection is finally achieved by likelihood function test in the original data space. Experimental results on EPS-A aerial multi-band spectral imagery demonstrate the effectiveness of the proposed algorithm.