A hybrid method combining rule-based filter and machine learning to detect porpoise and vessel sounds from a pulse event recorder.

Passive acoustic monitoring is essential for assessing the impact of anthropogenic noise on marine ecosystems and detecting vocalizing marine life. While acoustic event recorders are widely used to record odontocete echolocation due to their low power and memory demands, conventional detection algorithms are often unsuitable for analyzing datasets composed of complex pulse events. Here, we developed a hybrid analytical framework combining a rule-based filter with a random forest model to efficiently detect narrow-ridged finless porpoise (Neophocaena asiaeorientalis) click trains and vessel noise events using data from the pulse event recorder. The rule-based filter effectively reduced noise from raw data, achieving detection accuracy of almost 100% for click trains and 94% for vessel noise. However, among the events detected by this filter, 45% and 81% were actually false positives. The machine learning model improved classification accuracy to 97% and 99%, respectively. This model reduced the high false positive rates to 2.8% and 0.1%. This combined method offers a robust and efficient approach to processing pulse event recorder data, specifically for A-tag. It reduces manual workload, improves detection accuracy, and facilitates rapid assessment of vessel noise impacts, thereby supporting long-term ecological monitoring of small cetacean populations in diverse and noisy marine environments.