LFA target characterization with hybrid classifiers

Low-frequency active (LFA) target echo characterization is of considerable interest from the perspective of target physics and acoustic signal processing because of the highly variable echo structure as a function of time, aspect, and target type. We explore a combination of low dimensional features and four classifier architectures that exploit temporal variabilities in the echo structure. We extract features from four projection spaces: (1) compressed phase map, (2) image compression of the reduced interference distribution (RID), (3) speech-related features from the output of the principal component inversion (PCI), and (4) segmented matched filter output shape and amplitude statistics. We demonstrate the importance of matching the classifier architecture to the underlying low dimensional feature distribution. We link good features to target physics and explore an appropriate classifier architecture that maximizes class separability for robust echo characterization