Fishing for data: AI approaches to advance recreational fisheries monitoring

Abstract

Recreational fishing is an important social and economic activity on a global scale. In Australia, the quantification of the recreational fishing catch has relied on voluntary surveys, which require significant human resourcing for data collection and also limit the sampling and the precision of the estimates. Artificial intelligence (AI) incorporated into fish cleaning tables offers the opportunity to collect data in an almost continuous manner with little human effort. To date, the feasibility of such an approach has not been tested nor optimised. Here, we present a rudimentary image classification model that we developed with a dataset of only 2000 images, and test its accuracy and the deployment variables that may affect its performance. We assessed the influence of camera height, image resolution, species, fish orientation and position on identification and measurement accuracy. Lowering camera height and photographing fish in their dorsal orientation increased identification accuracy, while poor image resolution decreased measurement accuracy. These results informed subsequent development of an object detection model with increased performance (80% true positives compared to 30% for the image classification model). This assessment indicates the substantial promise of AI to sample recreational fisheries catches, with the potential to dramatically increase data collection for recreational fisheries management.

Bibliographic Information