PERFORMANCE ANALYSIS AND OPTIMIZATION OF YOLOV8 FOR UNDERWATER FISH DETECTION AND COUNTING

Abstract

Manual monitoring of fish populations in floating net cages is time-consuming and prone to producing inconsistent data. This study evaluated the performance of YOLOv8n, a lightweight variant of YOLOv8, for automated fish detection and counting in floating net cages in Ambon, Indonesia. A total of 780 underwater images, obtained from publicly available datasets on Kaggle and field data collection, were annotated and preprocessed through resizing, contrast enhancement using contrast-limited adaptive histogram equalization (CLAHE), and grayscale conversion. Two models were trained and compared: Model 1 without data augmentation and Model 2 with horizontal flipping and rotation. Model 1 achieved a more balanced performance, with an mAP@0.5:0.95 of 0.603, an mAP@0.5 of 0.762, a precision of 0.737, a recall of 0.711, and an F1-score of 0.724. Model 2 achieved a higher precision of 0.861 but a lower recall. Field testing integrating YOLOv8n with DeepSORT demonstrated reliable fish detection and counting, although accuracy declined in turbid water and densely packed fish schools. These findings suggest that a lightweight YOLOv8 model without architectural modifications can serve as a viable starting point for developing automated fish monitoring systems for precision aquaculture in coastal regions such as Ambon.