Abstract
The northern shrimp (Pandalus borealis) and the Atlantic cod (Gadus morhua) fisheries are prone to bycatch of polar cod (Boreogadus saida), a key Arctic forage fish species. Discrimination between the acoustic signals from these coinciding species could provide information on the risk of bycatch in addition to improving the accuracy of non-lethal scientific stock assessment surveys. As a step towards automatic in situ classification, we conducted a series of single-species mesocosm experiments for broadband target strength spectra measurements of Atlantic cod, polar cod and northern shrimp. Mesocosm experiments were completed with a Wideband Autonomous Transceiver (WBAT) and collected individual target strength spectra, TS(f), between 90–170 kHz and 185–255 kHz. Hundreds of TS(f) were extracted for each species and used to train machine-learning classification algorithms (i.e. classifiers). We found that two supervised learning classifiers, LightGBM and support vector machine, were able to achieve high classification performance (89%) on target spectra shape with a single 200 kHz transducer operating in broadband mode. This is promising for acoustic classification from autonomous platforms with limited payload. We explore the utilization of single transducer target spectra shape variability and provide recommendations to overcome challenges associated with scaling the method successfully for in situ marine species classification not only in the Arctic, but globally.