Recording and identification of marine mammal vocalizations on the scotian shelf and slope

In October 2012, Fisheries and Oceans Canada began collection of two years of near-continuous autonomous acoustic recordings in the Gully Marine Protected Area and adjacent slope areas between the Gully and Shortland canyon, and between Shortland and Haldimand canyons south of Nova Scotia, Canada. Data were sampled at 16 ksps for 13 min alternating with either 128 ksps or 375 ksps for 2 min to overlap the full bandwidth of marine mammal vocalizations likely to be present in the area. Each recorder produced ~3.5TB (~286 days) of acoustic data per year, for a full data set of ~20TB. To efficiently analyze this data, species-specific automated detectors were developed. We created a set of `ground truth´ data by manually annotating an average of 1000 calls for each species of interest, which included calls from 1 to 6 geographic environments (recording locations). For each species, the entire dataset was split in three equal independent subsets. Two of these subsets were used at a time to determine the optimal parameters of the detectors by quantifying their performance using the precision and recall metric. Thus far, six species detectors have been developed and evaluated for the detection of marine mammals on the Scotian shelf and slope: northern bottlenose, right, blue, fin, sei, and humpback whales. We present an overview of the detectors, the methods used to test and refine the detectors, the precision and recall of each detector as a function of signal-to-noise ratio, and an overview of cetacean vocalization presence in 2013 on the Shortland-Haldimand station.