Real-Time Fish Detection with YOLOv8 on a Semi-Autonomous Underwater Vehicle for Aquaculture Monitoring

Aquaculture is a vital sector for supporting food security, yet fish monitoring in floating net cages is still dominated by manual inspection, which is inefficient and prone to error. Advances in computer vision have enabled more reliable automated underwater observation, particularly with modern deep learning detectors. This paper presents a real-time fish detection system that integrates the YOLOv8s model with a semi-autonomous underwater vehicle (AUV) for monitoring aquaculture environments. The AUV captures underwater video using a Logitech C525 camera, while all inference is computed on a local GPU laptop, achieving an effective processing rate of 10–15 FPS. The model was trained on a custom dataset of 3,477 annotated underwater images containing Nile tilapia (Oreochromis niloticus), and evaluated using precision, recall, and mean Average Precision (mAP). The trained YOLOv8s model achieved a mAP@0.5 of 0.900, precision of 0.869, and recall of 0.804, outperforming a YOLOv5 baseline by 2.38%. Results show that detection performance remains stable under challenging conditions such as turbidity, occlusion, and low illumination. These findings demonstrate the feasibility of combining YOLO-based perception with AUV platforms to enhance the efficiency of aquaculture monitoring and support future extensions toward fish length estimation and large-scale deployment.