Detection and classification of protected species bycatch in Swedish small-scale fisheries

Abstract

Bycatches are an adverse side effect of fishing. Though rare, their occurrences have a serious impact on PETS(protected, endangered, and threatened species). In this paper, the feasibility of machine learning in the video feed analysis process to aid bycatch research is conducted and it is found that this niche benefits from applied machine learning. Different object detection algorithms are implemented on bycatch datasets built from scratch. The object detection models are compared on metrics such as average precision, mean average precision and recall to pick a model that is best suited for the bycatch dataset. It is also discussed how the machine learning model could benefit from diversifying the dataset while addressing key concerns of sharing data between different stakeholders. This concern is addressed by the adaptation of federated learning. A hierarchical federated machine learning framework (FEDn from Scaleout) is implemented to train YOLOv5s with Swedish and Danish clients(local models). The results obtained show that although the clients learn from each other, the rate of convergence is far slower than the locally trained models therefore requires fine-tuning and needs rethinking of global weights aggregation that determines how the clients learn from each other. Finally, it is concluded that with a good quality dataset the object detection model can be used as an aid for researchers, potentially helping them identify bycatch even when a human fails to identify them.