AUTHOR=Avsar Ercan , Feekings Jordan P. , Krag Ludvig Ahm TITLE=Estimating catch rates in real time: Development of a deep learning based Nephrops (Nephrops norvegicus) counter for demersal trawl fisheries JOURNAL=Frontiers in Marine Science VOLUME=10 YEAR=2023 URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2023.1129852 DOI=10.3389/fmars.2023.1129852 ISSN=2296-7745 ABSTRACT=

Demersal trawling is largely a blind process where information on catch rates and compositions is only available once the catch is taken onboard the vessel. Obtaining quantitative information on catch rates of target species while fishing can improve a fisheries economic and environmental performance as fishers would be able to use this information to make informed decisions during fishing. Despite there are real-time underwater monitoring systems developed for this purpose, the video data produced by these systems is not analyzed in near real-time. In other words, the user is expected to watch the video feed continuously to evaluate catch rates and composition. This is obviously a demanding process in which quantification of the fish counts will be of a qualitative nature. In this study, underwater footages collected using an in-trawl video recording system were processed to detect, track, and count the number of individuals of the target species, Nephrops norvegicus, entering the trawl in real-time. The detection was accomplished using a You Only Look Once v4 (YOLOv4) algorithm. Two other variants of the YOLOv4 algorithm (tiny and scaled) were included in the study to compare their effects on the accuracy of the subsequent steps and overall speed of the processing. SORT algorithm was used as the tracker and any Nephrops that cross the horizontal level at 4/5 of the frame height were counted as catch. The detection performance of the YOLOv4 model provided a mean average precision (mAP@50) value of 97.82%, which is higher than the other two variants. However, the average processing speed of the tiny model is the highest with 253.51 frames per second. A correct count rate of 80.73% was achieved by YOLOv4 when the total number of Nephrops are considered in all the test videos. In conclusion, this approach was successful in processing underwater images in real time to determine the catch rates of the target species. The approach has great potential to process multiple species simultaneously in order to provide quantitative information not only on the target species but also bycatch and unwanted species to provide a comprehensive picture of the catch composition.