Fusion Schemes for Ensembles of Hyperspectral Anomaly Detection Algorithms

Hyperspectral imaging is playing an ever increasing role in our military's remote sensing operations. The exponential increase in collection operations generates more data than can be evaluated by analysts unassisted. Anomaly detectors attempt to reduce this load on analysts by identifying potential target pixels which appear anomalous when compared to what are determined to be background, or non-target, pixels. However, there is no one individual algorithm that is best suited for all situations and it can be difficult to choose the best algorithm for each individual task. Fusion techniques have been shown to reduce errors and increase generalization, eliminating the need to always find the best algorithm for a given scenario. The utility of decision level fusion methods is examined, utilizing combinations of the emerging Autonomous Global Anomaly Detector and the Support Vector Data Description anomaly detection algorithms, along with the well-established Reed-Xiaoli detector. The fusion techniques investigated include algebraic combiners and voting methods. This research demonstrates that, with a modest amount of diversity among a minimal number of individual ensemble members, fusion offers reduced error rates and good generalization characteristics.