Once widely abundant throughout the North Pacific ocean, the eastern North Pacific right whale (NPRW) population (Eubalaena japonica) is now one of the rarest populations of marine mammals. Classified as critically endangered under the International Union for the Conservation of Nature, right whales have been illegally hunted for decades and are now nearing extinction. The most recent abundance estimate for this population estimated a mere 31 individuals remaining in the Bering Sea (95% confidence estimate: 23 to 54). However, this 12-year-old estimate is considered outdated and unreliable for use in conservation and marine mammal stock management, a task which our client, the National Oceanic and Atmospheric Administration (NOAA), is responsible for. Updating this population’s abundance estimate with traditional methods such as mark-recapture using photographs is particularly difficult due to the population’s sparsity and traditionally far-reaching habitat. Through this capstone project, we were tasked with automating the range estimation process and classifying ‘gunshot‘ call type (dispersed versus non-dispersed) for this endangered population through passive acoustic data modeling.
In our project developed a pipeline for processing and labelling data and a machine learning modeling using a multi-output Convolutional Neural Network (CNN) to identify (classification) and range (regression) dispersed gunshots of North Pacific right whales from passive acoustic recordings in the Bering Sea. We also present preliminary findings for use of this CNN model and how this pipeline can be used by NOAA to build a probability detection curve to be used in density estimation.