Use of AI-mediated image analysis to investigate the influence of environmental factors and recreational scuba diving on fish activity patterns in a marine protected area

Abstract

Due to inherent constraints on dive duration and inevitable observer effects, traditional underwater visual censuses (UVC) using scuba are of limited use for long-term behavioral monitoring of fish communities and the accurate assessment of any anthropogenic impact on them. Instead, we used an unmanned underwater camera system and YOLOv5 AI technology to assess temporal changes in fish activity patterns relative to diurnal cycles and water temperature changes in a Marine Protected Area (MPA) off the coast of Chaojing Park, Keelung, Taiwan. We confirmed a clear diurnal pattern in fish activity and identified temperature as a possible driver of species-specific activity patterns and behavioral cycles. This is a popular site for recreational diving, and the presence of scuba divers was associated with a significant but temporary avoidance response in fishes, which subsided quickly after the divers had departed. Hour-by-hour data nonetheless indicated that the number of fish present at the site may be influenced by the number of visitors at any given time. An increase in visitors was found to suppress the activity of damselfishes (Pomacentridae) for a short time, but visitor frequency did not significantly alter their daily mean abundance or overall diel activity patterns. While the rapid recovery observed within minutes after diver departure is convincing, the dataset does not allow robust inferences regarding longterm or cumulative effects. Although recreational diving currently appears to induce only transient behavioral shifts in this site's fish, the potential for lasting ecological impact, particularly if visitor density increases beyond current levels, remains uncertain.