Underwater detection of the presence of a ship is usually achieved by intercepting the noise radiated by its propeller. In the far-field and in a noisy marine environment, this task is hindered by the rapid decline of the SNR in the frequency band containing the harmonics of the propeller rotation. The traditional approach, largely grounded on empirical observations, consists in demodulating the broadband noise generated by propeller cavitation in the high-frequency end of the spectrum where the SNR is somewhat better preserved. This communication gives a theoretical justification to this methodology by recognizing that high-frequency propeller noise is essentially second-order cyclostationary. Using cyclostationary signal processing offers advantages over existing techniques, in that no user interaction is required to design pass band filters, and superior "cyclic frequency" resolution can be obtained thereby affording more accurate shaft speed estimates and the ability to discriminate multiple propeller shafts operating at slightly different speeds. This paper presents further development of the cyclostationary detection technique, focusing on the definition of statistical thresholds to allow automatic detection of propeller craft based on superior detection metrics. The performance of these thresholds and improved metrics is demonstrated using simulation and measured signals.