Unsupervised classification methods based on polarization features are considered for the problem of associating frequency-hopped pulses according to their transmission source. The problem is considered in frequency-selective propagation channels, where discrimination cannot be obtained simply through signal amplitude information, and where angle-of-arrival methods are challenged by multipath. Source discrimination is based on polarization-frequency behavior, leveraging polarization mode dispersion (PMD) associated with each propagation channel. The passive receiver employs a dual-polarized antenna to collect orthogonally-polarized complex baseband signals, measure PMD responses, and compare the responses with a dynamic library of responses to identify signals that likely originated from the same source. The PMD responses are updated upon reception of each pulse and the number of unique sources may also be estimated.