Automating length and maturity stage measurements of Antarctic Krill (Euphausia superba) from high resolution image pairs

Deep learning architecture was used to train a model to classify the maturity stage of Antarctic krill and measure their total length. These parameters are key variables in the methods for managing the Antarctic krill fishery employed by the Commission for the Conservation of Antarctic Marine Living Resources (CCAMLR). Images of Antarctic krill in both lateral and dorsal views were collected alongside manual measurements of krill length and sexual maturity stage to build a dataset of 457 photographs each comprising of at least 25 organisms. Deep learning techniques were used to develop tools to detect individual krill and pair them with labels of sexual maturity and length. The resultant dataset of 5095 unique labelled krill were used to train a maturity classifier and length regressor. A multi-task model using the DenseNet121 deep learning architecture on high resolution concatenated images provided the most accurate estimates of Antarctic krill length (root mean square error 1.76 mm) and Antarctic krill maturity stage (categorical cross entropy 90.52%). We propose that image collection, under specific guidelines, could be used to replace current fishery observer measurement effort to support CCAMLR objectives for collecting krill length and maturity stage from trawls. A stream-lined automated process could provide more consistent (between vessel and observer) and time-saving practises, permitting fisheries observers more effort to divert to other tasks.