ED004-0030
Evaluation of a Neural Network for Automated Classification of Beaked Whale Echolocation Clicks

Monday, 7 December 2020
Poster
Chelsea Field1, Alba Solsona Berga2, Rebecca Cohen3, Jennifer S Trickey3, Liam Mueller-Brennan4, Taylor Ackerknecht2, Kaitlin E Frasier5, Danielle Cholewiak6, Sofie M. Van Parijs7 and Simone Baumann-Pickering8, (1)Scripps Institution of Oceanography, La Jolla, CA, United States, (2)Scripps Institution of Oceanography, La Jolla, United States, (3)University of California, San Diego, Scripps Institution of Oceanography, La Jolla, CA, United States, (4)NOAA, Boulder, United States, (5)University of California San Diego, Scripps Institution of Oceanography, La Jolla, CA, United States, (6)NOAA Fisheries Woods Hole Laboratory, Protected Species Branch, Woods Hole, MA, United States, (7)Northeast Fisheries Science Center, National Marine Fisheries Service, National Oceanic and Atmospheric Administration, Woods Hole, MA, United States, (8)Scripps Institution of Oceanography, University of California, San Diego, Acoustic Ecology Laboratory, La Jolla, CA, United States
Abstract:
Passive acoustic monitoring (PAM) has become an important method in studying the ecology of visually elusive beaked whale species. The use of long-term PAM has allowed for acquisition of extensive data sets and has significantly increased our ability to assess anthropogenic influence on cetaceans. However, the large data sets generated present a challenge for species-level classification, due to the highly time-intensive nature of manual labeling by expert analysts.

The development of an autonomous network-based classifier has allowed for rapid analysis of these large data sets. We used an unsupervised clustering algorithm to identify dominant click types, then used the clustering output to train our automated classifier. The classifier then categorizes novel data based on its spectral and temporal features. Due to the wide bandwidth of data acquisition, the occurrence of other consequential anthropogenic and biological signals presents an opportunity to fine-tune the classifier to separate these noise sources from our species of interest and improve accuracy.

In order to characterize the performance of the automated classifier, an assessment of accuracy and error was needed. Manual click classifications, performed independently on the same data, served as the ground truth against which the automated classifier would be assessed. Our research aimed to quantify the discrepancies between the automated classifier and the ground truth. The performance of the automated classifier was evaluated through the ability to classify five beaked whale species (Gervais’, Cuvier’s, True’s, Blainville’s, Sowerby’s) at sites off the US Atlantic coast. Click-type profiles were based on the ability to match the spectral average and waveform for each species, regardless of noise. A comparison framework was used to quantify similarities and differences between the two methods. Discrepancies were further assessed at the individual click level.

This assessment provided the insights necessary to improve the network-based classifiers’ performance on long-term acoustic data. Optimization of the automated classifier greatly reduces acoustic data processing time, thus allowing efforts to be fostered towards follow-up questions that may expand our understanding of these cryptic whales.

<< Previous Abstract | Next Abstract >>