Real-time AI-augmented visual object detection in aquaculture: A deep learning and statistical framework for YOLOv5-based crayfish monitoring under variable molting and lighting Conditions

Abstract

This study investigates the application of the YOLOv5 object detection model to accurately identify and count post-larval redclaw crayfish (Cherax quadricarinatus) across diverse aquaculture environments. A 2×2 factorial experimental design was employed to examine the effects of environmental lighting (covered vs. open-air ponds) and biological molting status (molted vs. non-molted) on detection performance. A dataset of 1,200 meticulously annotated images was collected under controlled conditions and used to train and evaluate the YOLOv5s variant. Detection effectiveness was measured using precision, recall, F1-score, and mean average precision at an IoU threshold of 0.5 (mAP@0.5). Results showed that non-molted crayfish in open-air ponds achieved the highest detection accuracy with an F1-score of 0.93, whereas molted crayfish in covered ponds exhibited the lowest performance with an F1-score of 0.85. Two-way ANOVA confirmed significant main effects of lighting and molting status, as well as their interaction, on model accuracy (p < 0.05). These findings highlight the critical roles of biological pigmentation and lighting conditions in optimizing object detection accuracy for aquaculture monitoring. This study offers valuable insights toward practical applications of deep learning–based monitoring systems for sustainable aquaculture operations.