Improving seabird species identification in electronic monitoring applications using machine learning systems
Electronic Monitoring (EM) systems have been used in many fisheries for a variety of purposes, such as compliance monitoring and characterizing catch or bycatch. Accurate and precise enumeration of catch and bycatch using video imaging has been a challenge whether it is for commercial fishery management or research fishery cruises. Scientists in the U.S. Pacific Northwest, including National Marine Fisheries Service (NOAA Fisheries) staff in collaboration with University of Washington have been exploring technology that addresses this challenge. The success with accurate fish species identification led to this proof of concept research applying this technology and methodology to seabird species identification. In a laboratory setting, a multi-spectral camera chute was set up and birds collected for necropsy were presented to the imaging cameras. Training images (1,837) of a variety of species were used to support feature extraction by the camera systems. Test images (213) of 16 species or species groups were then examined. Overall accuracy was 93%, with some species (Black-footed and Laysan Albatross) at 100% accuracy. With the favourable results of the proof of concept, further research, development, and testing will be conducted.