Deep learning algorithm outperforms experienced human observer at detection of blue whale D‐calls: a double‐observer analysis

Abstract

An automated algorithm for passive acoustic detection of blue whale D‐calls was developed based on established deep learning methods for image recognition via the DenseNet architecture. The detector was trained on annotated acoustic recordings from the Antarctic, and performance of the detector was assessed by calculating precision and recall using a separate independent dataset also from the Antarctic. Detections from both the human analyst and automated detector were then inspected by an independent judge to identify any calls missed by either approach and to adjudicate whether the apparent false‐positive detections from the automated approach were actually true positives. A final performance assessment was conducted using double‐observer methods (via a closed‐population Huggins mark–recapture model) to assess the probability of detection of calls by both the human analyst and automated detector, based on the assumption of false‐positive‐free adjudicated detections. According to our double‐observer analysis, the automated detector showed superior performance with higher recall and fewer false positives than the original human analyst, and with performance similar to existing top automated detectors. To understand the performance of both detectors we inspected the time‐series and signal‐to‐noise ratio (SNR) of detections for the test dataset, and found that most of the advantages from the automated detector occurred at low and medium SNR.

A New Acoustical Autonomous Method for Identifying Endangered Whale Calls: A Case Study of Blue Whale and Fin Whale

“…In fact, WST can assist us to extract significant features for LSTM in those situations when it is not possible to learn efficient features with the available training data in case of data scarcity. This makes our proposed scheme for use, such as few-shot learning [ 29 ] for identification of endangered whale calls unlike the existing recent deep-learning-based methods [ 2 , 3 , 7 , 8 ] that require large datasets for learning.…”

“…Blue whales are known to produce downswept FM (frequency-modulated) calls that are often referred as D-calls. Both male and female blue whales have been found to produce such calls [ 3 ]. On the contrary, it is observed that only males produce song, and with that, these calls are associated with the breeding season.…”

“…Studies in the Ref. [ 3 ] showed that the DenseNet-automated blue whale D-Call detector, which is based on conventional Convolutional Neural Networks (CNN) provided better results in terms of detection probability than that of human observers’ analyses, particularly at low and medium signal-to-noise ratios (SNRs). Higher detection probabilities (0.905 and 0.910 for low and medium SNRs) were obtained compared to the detection probabilities obtained by human observers’ analyses (0.699 and 0.697 for low and medium SNRs).…”

“…In the Ref. [ 3 ], a long-term recording dataset, particularly the Australian Antarctic Division’s “Casey2019” dataset was used for the results. In the Ref.…”

Ecological inferences about marine mammals from passive acoustic data

“…Traditionally, signals in passive acoustic data are classified by human analysts who generally are supported by software that provides visualisation, audition, and analysis of the recordings. Signal analysis by humans usually is treated as the standard against which automated detectors should be compared (although see Miller et al ., 2023), but can be slow and incur labour costs and delay. Differences in expertise among analysts, a given analyst's attention span on a particular day, and other inconsistencies lead to variability within and among human analysts (Širović, 2016; Leroy et al ., 2018), creating non‐systematic biases.…”