Anomaly Detection Through Adaptive Background Class Extraction From Dynamic Hyperspectral Data

We propose a computationally efficient methodology to determine the background classes and anomalies in a moving sequence of hyperspectral data. A target material may be considered as an anomaly in an image that has a spectral signature different from the spectral signatures of the background objects. In order to detect such anomalies in an image, the classes associated with the background have to be known. The method is based on the assumption that anomaly pixels are relatively rare in comparison with the abundance of the background class pixels. It consists of robust clustering of subsets of the image pixels. The resulting clusters may be considered as representatives of the background classes in the image. The clusters are obtained using a self-organising map (SOM) clustered using the local minima of the U-matrix (distance matrix). In a first stage the classes are built using a randomly chosen small percentage of the image pixels. Subsequently, the background classes are updated in a process where existing classes are reinforced or deleted from the feature space, new classes are formed and anomalies are detected. Experiments are conducted using realistic hyperspectral data where the use of the method for real-time on-fly target detection is demonstrated