Variability in the performance of the spectrogram correlation detector for Northeast Pacific blue whale calls

Spectrogram correlation has been used successfully for automatic detection of baleen whale calls. However, applying this method consistently to long time series can be challenging. To illustrate the potential challenges of the automatic detection process, recordings collected in the Southern California Bight between 2007 and 2012 were used for detection of North-east Pacific blue whale (Balaenoptera musculus) B calls. The effects of the following factors were investigated: blue whale B call frequency shift and appropriate kernel modification, seasonal variability in call abundance, analyst variability and noise. Due to intra- and inter-annual changes in the call frequency of blue whale B calls, seasonal and annual adjustments to the call detection kernel were needed. To account for seasonal variability in call production, evaluation of the detector against ground truth data was performed at multiple times during the year. Analyst variability did not affect overall long-term trends in detection, but it had an impact on the total number of detections, as well as call rate estimation. Noise, particularly from shipping, was negatively correlated with detections at hourly time scales. A detailed analysis of variability in the performance of spectrogram correlation detectors should be performed when applying this method to long-term acoustic data-sets.