Automated detection of sperm whale clicks under interference from noise transient
Sperm whales produce powerful and broadband echolocation clicks that serve as their primary means of communication and prey localization in deep ocean environments. Passive Acoustic Monitoring (PAM) systems provide a non-invasive approach to study these animals by detecting and analyzing their clicks over long temporal and spatial scales. However, accurate click detection is often hindered by both natural and anthropogenic noise transient sources that overlap with sperm whale clicks in the time and frequency domains. Such overlap leads to high false detection rates and reduced system reliability. In this study, we propose a novel PAM-based detection method that leverages the impulsive characteristics of sperm whale clicks to separate them from noise transients. It begins with the estimation of a click model derived from 3000 manually verified clicks. Then, we design a custom wavelet specifically matching the temporal and spectral structure of the estimated click model. Acoustic recordings are then decomposed using this tailored wavelet and then reconstructed under optimized wavelet settings to enhance impulsive components. This reconstruction efficiently preserves click energy while attenuating Gaussian background noise and suppressing strong noise transient events. Performance evaluation is conducted using Receiver Operating Characteristic (ROC) analysis to assess the trade-off between true and false positive detection probabilities under varying signal-to-noise ratios (SNRs). The evaluation utilizes several datasets: 25,000 manually annotated clicks from Dominica Island recordings, 3.5 hours of Dominica noise-only data, a 3.6-hour subset of seven months of Mediterranean Sea recordings contains manually verified noise transients, and synthetic datasets as additive Gaussian noise and noise transients. Compared with the Teager–Kaiser Energy Operator (TKEO) benchmark, our method demonstrates superior robustness, achieving a more favorable balance between probability of true positive and false positive detections across all SNR levels and consistent behavior across geographically distinct noise environments in terms of false alarm rate.