Can AI improve management of recreational fisheries: Assessing the effectiveness of early AI models to identify and measure fishes

<p dir="ltr">Recreational fishing is a socially and economically important pastime and industry in New South Wales (NSW), boasting nearly half-a-million licenced participants. Due to its size, the activity needs to be well-managed to avoid negative environmental impacts, such as overfishing. To do so, robust data are required on the total catch and size-frequency of the major targeted species. At present, due to costs and logistics, the industry uses telephone diary surveys, which provide limited-quality data for recreational fishing management. Here, we create and test a rudimentary Artificial Intelligence (AI) model to identify fish species, size and biomass. If effective, this approach could be implemented at boat ramp fish cleaning tables to provide industry with a more complete large-scale assessment of the recreational catch and thus improve recreational fishery management. We assessed a rudimentary AI model to identify species, measure fish length and width, and predict fish biomass, and tested the effects of a suite of variables on AI model performance. We found that the rudimentary model, currently trained using ~2000 fish images, could not adequately identify fish species (between 49 – 2 % accuracy depending on the species). Besides species, lowering camera height, and photographing fish in their dorsal orientation increased the accuracy of identification. The AI model provided accurate length estimates for fishes, with more than 95% accuracy for four species caught in NSW. The accuracy of AI measurements was negatively impacted by poor image resolution. The accuracy of biomass predictions derived by using AI measurements and established length-weight relationships varied between species, with tiger flathead (Platycephalus richardsoni) and Australasian snapper (Pagrus auratus) being estimated most and least accurately, respectively. These results were influenced by the fish cleaning process, where discarded viscera caused an overestimate of biomass. Our findings indicate that there is an exciting prospect in using AI for recreational fisheries monitoring. Our assessment of a very simple AI model found that this technology can improve length-weight relationships for endemic species and/or develop area-weight relationships using more sophisticated computer vision techniques. Additionally, we suggest further research into fish waste discard at cleaning tables to improve biomass predictions and significantly enhance the management of recreational fishing.</p>