Abstract
Abstract
Increasing anthropogenic pressures such as climate change, overfishing, pollution, and marine energy extraction, reinforce the need for enhanced marine monitoring to support ecosystem management. Underwater video monitoring is increasingly used for this purpose but has been constrained by the labour-intensive nature of manual video analysis. Here, we developed a YOLO-based object detection pipeline to automatically enumerate two commercially important fish taxa (
Gadidae
;
mAP
= 0.896 ± 0.009,
Pleuronectiformes
;
mAP
= 0.814 ± 0.009) in Baited Remote Underwater Video (BRUV) data collected at offshore wind farms in the northern North Sea. We introduce a custom metric, Mean Peak Success (
MPS
), to evaluate the model’s ability to estimate the maximum number of individuals observed (
MaxN
) in BRUV deployments using an abundance-appropriate error tolerance. Across two survey years within a single monitoring programme, the automated pipeline partly replaced manual annotation. Applying the model to stereo-BRUV data, we found that combining both camera viewpoints (
max-MaxN
) can increase the sensitivity of
MaxN
estimates to experimental variables relative to a single viewpoint. Tailored object detection can therefore extend the efficiency and analytical scope of video-based marine monitoring within a defined survey system.