Experimental feature-based SAR ATR performance evaluation under different operational conditions

Automatic target recognition (ATR) system performance over various operating conditions is of great interest in military applications. The performance of ATR system depends on many factors, such as the characteristics of input data, feature extraction methods, and classification algorithms. Generally speaking, ATR performance evaluation can be performed either theoretically or empirically. The theoretical evaluation method requires reasonably accurate underlying models for characterizing target/clutter data, which in many cases is unavailable. The empirical (experimental) evaluation method, on the other hand, needs a fairly large data set in order to conduct meaningful experimental tests. In this paper, we present experimental performance evaluation of ATR algorithms using the Moving and Stationary Target Acquisition and Recognition (MSTAR) data set. We conduct a comprehensive analysis of the ATR performance under different operating conditions. In the experimental tests, different feature extraction techniques, Principle Component Analysis (PCA), Linear Discriminant Analysis (LDA) and kernel PCA, are employed on target SAR imagery to reduce the feature dimension. A number of classification approaches, Nearest Neighbor, Naive Bayes, Support Vector Machine are tested and compared for their classification accuracy under different conditions such as various feature dimensions, target classes, feature selection methods and input data quality. Our experimental results provide a guideline for selecting features and classifiers in ATR system using synthetic aperture radar (SAR) imagery.