Recognition of targets from their electromagnetic scattered signals is a complicated problem as such data are highly aspect and polarization dependent. Most of the suggested target recognition techniques in resonance region make use of target's system poles either directly or indirectly because the complete set of system poles constitutes an aspect and polarization independent descriptor of a given scattering object. The WD-PCA based classifier and the MUSIC algorithm based classifier are recently suggested electromagnetic target classifiers demonstrated to be very successful in classifying both conducting and dielectric objects of arbitrary shapes using their late-time (natural resonance based) scattered data recorded in resonance region. In this paper, noise performances of these two target classification techniques are investigated in a comparative manner using a set of five perfectly conducting spheres with radii of 8, 9, 10, 11 and 12 cm. This target library is chosen as a worst case testing library under noisy data because the natural resonances of a conducting sphere decay, very fast in time, and hence, even a small amount of random noise can badly contaminate its late-time scattered response. It is demonstrated in this work that a slightly noisy set of reference data (instead of noise-free reference data) must be used in the classifier design phase of both techniques to obtain acceptable noise performance in classifying low-Q targets (i.e., targets having natural resonances with low quality factors) such as perfectly conducting spheres.