Seabirds are often caught during commercial and recreational fishing, and for certain species, these fatalities can pose a significant threat to their population viability. The counts of seabirds around fishing vessels, conducted by government fisheries observers, may provide a greater understanding of the interactions between seabirds and fishing vessels. Moreover, the count will help manage fishing to reduce seabird bycatch. However, observers' experiences and skills vary, and collecting seabird count data is challenging. Because of this variability, there is an inherent difficulty in accurately counting birds behind fishing vessels.
This work presents an AIML approach to counting seabirds in longline fisheries in challenging environments prone to occlusions and variable lighting conditions. Data is collected by human observers, who log bird counts at specific times while simultaneously recording video footage. These videos are manually annotated with bounding boxes to train machine learning models, which are then used to compare counting accuracy against the human observer data.
We have conducted experiments to assess the accuracy of various state-of-the-art object detection models in counting seabirds and assess the effectiveness of AIML in counting seabirds analysed compared to human observers.