Precision-oriented two-stage YOLOv8–Faster R-CNN framework for intelligent fish monitoring in sustainable aquaculture systems

Sustainable aquaponics requires accurate and stable real-time fish monitoring under visually degraded conditions such as turbidity, glare, reflections, occlusion, and dynamic illumination. Although YOLOv8 enables efficient real-time detection, its performance in low-clarity aquatic environments is often limited by localization instability and background-induced false positives. To address this issue, this study proposes a precision-oriented two-stage YOLOv8–Faster R-CNN framework, where YOLOv8 performs initial fish localization and region-of-interest (ROI) cropping, followed by Faster R-CNN refinement with a ResNet-50–FPN backbone to enhance spatial consistency and background suppression. Data were collected from continuous monitoring in a solar-powered aquaponics greenhouse, capturing real operational disturbances affecting visual clarity. A staged evaluation was conducted, including controlled aquarium validation and greenhouse deployment. In controlled conditions, the model achieved mAP@0.5 ≈ 0.98 and mAP@0.5:0.95 ≈ 0.55 with stable convergence. Under greenhouse conditions, the proposed framework improved strict localization (mAP@0.5:0.95 ≈ 0.27 vs. 0.23 for YOLOv8), increased precision (~ 0.56 vs. ~ 0.53), reduced false positives, and produced more temporally stable bounding boxes, although with slightly lower recall for extreme fish sizes. These results indicate that the contribution lies in robustness and detection stability rather than raw accuracy gains. Limitations include single-site data, mild overfitting, and higher computational overhead. Future work will focus on multi-site datasets, lightweight edge optimization, and temporal modeling to support reliable, energy-efficient fish monitoring in PV-powered sustainable aquaponics systems.