Marine Mammal

2.2 | Marine mammal detection and classification Alternating files from the first 4 days of each month of 2014, approximately 4% of the full data set, were manually analyzed through audio-visual analysis of 30 s, full bandwidth, multipart spectrograms in PAMlab (JASCO Applied Sciences, 2015; spectrogram parameters in Table 1). Spectrograms were displayed on a logarithmic frequency axis to enable annotation of multiple species in a single pass by improving visibility of low frequency signals, toggling to linear display when necessary to better visualize high frequency signals. One call per species per file was annotated. The first call encountered was usually selected, though in some cases a different or additional call was annotated to vary the call types and noise conditions in the data set or choose a clearer call that could be more confidently attributed to a species when the first call was unclear. These annotated data and others from five other sites provided by six analysts at the Canadian Department of Fisheries and Oceans and JASCO Applied Sciences were used by JASCO Applied Sciences to develop feature extraction and classification packages. The resultant detection and feature extraction algorithms were used to extract 94 time domain, frequency domain, and time-frequency features from each signal detected in the 20 Hz to 8 kHz band. This limited analysis to PWSD pulsed calls, as echolocation signals are outside of this band (Figure 3). Given that the frequency band of PWSD echolocation far exceeds the recording bandwidth (20 kHz to 100 kHz and 1 Hz to 32 kHz, respectively), and isolated PWSD clicks sometimes appear similar to those of other dolphin species within this bandwidth (Soldevilla et al., 2008), echolocation clicks were only used to validate species classifications for pulsed call events and were not otherwise included in the data set. A random forest classifier with 100 trees, minimum leaf size of 1, and confidence threshold of 0.2 was trained using the treebagger algorithm in MATLAB R2016a (Breiman 1984, 2001; Loh 2002; Loh & Shih 1997; Meinshausen 2006; The Mathworks Inc., 2016). The training data included 1,808 marine mammal annotations from Barkley Canyon and 500 randomly selected detections from files not containing marine mammal sounds. The classifier was trained to classify signals into one of five roughly balanced classes representing each species identified in the annotated data set: humpback whale (Megaptera novaeangliae), killer whale (Orcinus orca), sperm whale (Physeter macrocephalus), PWSD, and “other,” where other included all detected nonmammalian sounds. The classifier was validated with 100 repetitions of 10-fold cross-validation (Kim, 2009), yielding precision, recall, and F -score of 0.7903, 0.7986, and 0.7906, respectively, for the PWSD class (Davis & Goadrich, 2006). To avoid performance overestimation resulting from distributing nonindependent data across validation folds, training data were divided as evenly as possible across folds such that no two folds contained data from the same file or the same encounter (Bianco et al., 2019; Roberts et al., 2016). An encounter was defined as a bout of vocalizations from a single species with silences between vocalizations lasting <15 min. This classifier was used to classify detections for the full May 2013 to January 2015 data set produced by the same detection and feature extraction algorithms used for the training data. A binary presence/absence data set was generated from the classification results, and each file containing at least one sound classified as PWSD was manually reviewed through audio-visual analysis in PAMLab (JASCO Applied Sciences, 2015) to remove false positives from the data set. Detections of Risso's dolphins, who's echolocation clicks are similar to those of PWSDs within the recording bandwidth but who's pulsed calls are different (Corkeron & Parijs, 2001; Henderson et al., 2011; Neves, 2013; Soldevilla et al., 2008), were removed at this stage. Long-term spectral averages (LTSA) of each detected PWSD calling event produced using the MATLABbased Triton software (Scripps Whale Acoustics Lab, 2016; 8 s time resolution, 100 Hz frequency resolution) were assessed for presence of the banded echolocation clicks characteristic of PWSDs (Figure 3). Events not containing these diagnostic clicks were removed from the data set to exclude northern right whale dolphins (Lissodelphis borealis), whose pulsed calls are indistinguishable from those of PWSDs (Henderson et al., 2011; Rankin et al., 2007; Soldevilla et al., 2008, 2010). While some ambiguous events that could have been PWSDs were removed due to an absence of any clicking behavior, these events were not considered to represent a significant portion of the data set as they were not numerous and tended to be quite short, often lasting <5 min. The files confirmed to contain PWSD signals totaled 6.42% of the data.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.