Size-specific escapement and a lack of fine-scale spatial information on sampled species can affect standard trawl sampling and its effectiveness for ground truthing acoustic data, undermining the accuracy of data collected on scientific fisheries surveys. By having a camera “eye in the trawl” it is possible to gather continuous records of the organisms caught in the net, including ones too small to be effectively retained. This thesis aims to evaluate the potential of the Deep Vision in-trawl camera system for annotating acoustic recordings by comparing the data collected from three pelagic stations in the Norwegian Sea to acoustic registrations and catch results. Manual analyses are also compared with results from automated machine learning techniques. As expected, total counts and length frequency distributions of <10 cm long mesopelagic fish differed significantly between catch and image-data. Apart from a few exceptions, patterns of acoustic NASC values and Deep Vision counts matched vertically and horizontally for the two 20-30 cm long pelagic species encountered in the survey. Species and swimming orientation were observed to be two important factors to consider for scaling counts by species obtained while approximating the true fish count when using machine learning algorithms to better approximate the true fish count. These results suggest that if the accuracy of the automated analysis is increased, in-trawl cameras have the potential to provide a more accurate understanding of the biological components in the water column.
Managing Thesis and Internship Exchanges
