Multiple-detector fusion for anomaly detection in multispectral imagery based on maximum entropy and nonparametric estimation

With the development of sensors capable of high spatial and spectral resolution, anomaly detection in Multispectral Imagery has gained more attention recently. However, using single detector meets great limitation due to that many of the conditions and parameters that govern performance are unknown or poorly characterized in an operational setting. Unlike the conventional fusion approaches simply using logical operators (e.g. AND, OR), which lead to produce highly variable performance results from one case and thus difficult to specify the "best" fusion logic in advance, a multiple-detector fusion (MDF) algorithm is proposed in this paper. Three successive procedures are included as follows: First, we use series anomaly detectors including well-known RX and its varieties to get the pilot detection results. Second, in order to estimate the pdf statistics of each individual detector's output more accurately, a nonparametric method called kernel density estimation (KDE) with bandwidth adjusted adoptively is used. The obtained probabilistic information are then fused using a modeled joint distribution by the principle of maximum entropy. Finally, the MDF approach is applied to real multispectral imagery. Experimental results and theoretical analysis demonstrate the effectiveness of proposed algorithm.