Detection of Marine Mammal Vocalizations in Low SNR Environments with Ideal Binary Mask

Due to the complexity of the underwater environment, detecting underwater target signals, especially in low SNR conditions, remains challenging. In the context of marine mammal vocalization detection, several techniques have been used by human experts to distinguish and annotate vocalizations from background noise. However, manual analysis by human experts, where critical ratios (CRs) are calculated can be labor-intensive and subject to interpretation differences. These discrepancies may lead to false detections of vocalizations. To mitigate this limitation, this study proposes an ideal binary mask (IBM) method for separating marine mammal vocalizations from background noise. The effectiveness of this method was evaluated using a similarity coefficient and preserved signal ratio (PSR), which suggests that the IBM method holds potential for effectively distinguishing marine mammal vocalizations below the noise floor.