Combat identification (CID) produces a mapping from data collected on detected objects to a label which identifies the object to a war fighter. Template matching classifiers perform this task through the comparison of known representations of targets of interest, called templates, with features generated from data collected on unknown objects of interest. In a forced decision scenario unknown objects are given the label of the target class associated with the template that is most similar to the unknown object. This research uses an unforced decision approach to CID. Instead of forcing a decision, objects that are unlike any used to train the classifier are labeled as out-of-library (OOL). Objects similar to more than one object represented in the library of known objects are given a non-declaration (NDEC) designation. The development of OOL and NDEC methodologies that will function well in operating conditions the classifier was not designed for, extended operating conditions, is a significant challenge and a major focus of this research.

This research develops an OOL method based on squared Mahalanobis distance scores and NDEC methodologies based on information theory. The philosophy used in this research is that OOL label assignments occur before all other labeling or non-declaration processing. Under this philosophy, the limiting effect an OOL methodology may have to classification improvements achievable with nondeclaration methods is explored.

A measure of statistical separation of in-library targets using Bhattacharyya distance is developed. This measure is used to determine the context in which a template matching classifier may experience low classification accuracy. It is found that the aspect angle the targets present to the collection platform is exploitable. Three philosophies that use knowledge of the estimated aspect angle to improve classifier performance are explored. In each philosophy classification performance improvements are achieved through modification of thresholds used by the OOL or NDEC methods to process target records collected at different aspect angles.

Experimental results demonstrate the utility of the OOL and NDEC methodologies as well as the potential gains in classifier performance achievable through context-based threshold modifications. In the experiments, a mathematical framework is used to choose the optimal classifier parameters, methodologies, and threshold adjustment methods for fixed number of observations and sensor choices while enforcing classifier performance constraints consistent with war fighter preferences.

A feature saliency method based on squared Mahalanobis distance is developed and applied to reduce the dimensionality of the input feature space. After confirmation of the feature reduction choices with the Fisher ratio, experimental results using the reduced feature space demonstrate the potential of the saliency method.

The extensibility of the developed OOL and NDEC methodologies are shown by implementing the methodologies with some modifications for use with a hidden Markov model classifier. The robustness of the methodology relative to the choice of receiver operating characteristic thresholds is demonstrated by observing the change in classification accuracy under extended operating conditions.