Within a Classification System (CS), prior to classification, feature extraction techniques are used to reduce the
dimensionality of features. A standard unsupervised technique is the Principal Component Analysis (PCA). In
this paper we apply a supervised method for feature extraction, the Non-Parametric eigenvalue-based Feature
Extraction (NPFE), to CS data sets and compare the performance of the two different feature extraction
schemes based on classification accuracies obtained with the LVQ cluster algorithm. Furthermore, to reduce
computational complexity of NFPE, we introduce an approximate NFPE and show that it provides significantly
reduced computation time with almost identical performance as in the full NPFE algorithm.
Missile Warning Systems (MWS) have the task to identify missile threats to support timely counter measures. Key difficulty is that anything which can be detected must be considered a potential alarm at the output of the classifier. Hence, MWS must be optimized at a certain threshold on the receiver operating characteristic to trade probability of declaration against false alarm rate. To identify actual threats, two neural-network based discrimination algorithms are presented. In the first approach, measured object features from each spectral band over time are used to derive temporal features that model the temporal object behavior. These temporal features are fed into a static neural network. In the second approach, the measured object features are fed directly into a dynamic neural network which has a context layer. We present performance results of the two approaches based on simulated missile data overlaid with recorded background data.