Comparison of time-frequency distributions for target classification.

Abstract

Time frequency representations characterize signals in the time-frequency plane and can aid in the physical interpretation of backscattered chirps. However, these 2-D representations force trade-offs in resolution, computational efficiency, noise reduction, and cross-term generation. We are mainly concerned with the ability of the transform to reject noise, but we are also concerned with the effect of the cross-term artifacts that do not represent real phenomena on the ability to classify targets using machine learning approaches. First, strong backscatter returns available for a particular target were identified. In the cluttered field under study, this was accomplished from a synthetic aperture image of the field. The processing reported herein used these returns from each studied target to produce a partition of the time frequency plane that contained disjoint sets of time-frequency regions. Each target was represented by a subset of these regions. To test the separability of the targets based on these subsets, random pings were selected from each target’s high backscatter region. These pings were then compared against the disjoint sets. This procedure was carried out with several different time-frequency methods in order to determine the importance of the method’s resolution, noise reduction quality, and cross-term artifacts, in producing robust time-frequency plane partitions that may be used to classify targets. [This work was sponsored by the SERDP.]