Taking shallow-water ocean-bottom seismometer (OBS) stations near the US West Coast as an example, using manually identified frequency-band results as a benchmark, this study compares the earthquake arrival-time picking performance of the Short-Term Average/Long-Term Average (STA/LTA) method and two machine learning models developed for OBS data (Blue-PhaseNet and Blue-EQTransformer). The results show that 1) Both Blue-PhaseNet and Blue-EQTransformer exhibit poor detection capability, with accuracies of 33.33% and 14.81%, respectively, which are 51.85% lower in accuracy than the traditional STA/LTA method. This is likely because the models have not been optimized for local datasets, resulting in insufficient generalization. Furthermore, this study uses large-magnitude earthquake data with high noise in shallow-water environments, whereas the training datasets consist predominantly of microseismic events from deep-water areas, leading to significant differences in data characteristics. 2) After denoising using hydrophone pressure records, the arrival time picking accuracy of the STA/LTA method improves significantly to 96.30%, and more seismic events are identified (29 events, compared with 27 in the raw data). Therefore, Blue-PhaseNet and Blue-EQTransformer cannot be directly applied to shallow-water observations. Denoising to improve the signal-to-noise ratio followed by STA/LTA processing represents a feasible automated approach for OBS earthquake arrival-time identification.