Implementation of YOLOv5 for Fingerling Detection under Turbid Conditions in an Ex-Tin Mining Pond

The utilization of ex-tin mining ponds for freshwater aquaculture in Bangka Belitung requires reliable monitoring technology, particularly for fingerling counting under variable water clarity conditions. Manual counting is time-consuming, labor-intensive, and may increase fish stress due to repeated handling. This study developed a portable real-time detection system for Nile tilapia fingerlings by integrating YOLOv5-based image processing with turbidity sensing. The system used a camera for image acquisition, an embedded computing unit for object detection, and turbidity sensors to measure water clarity during testing. A dataset of Nile tilapia fingerling images was collected, annotated, and augmented to improve model generalization under different orientations, densities, and visual conditions. Experimental results showed that the YOLOv5 model achieved 98% accuracy, 100% precision, 96% recall, 100% specificity, and an F1-score of 0.98. Detection performance remained stable at low fingerling densities, while under-counting occurred under high-density conditions when turbidity exceeded approximately 85 NTU. To address this limitation, a turbidity-based correction factor was introduced to compensate for detection loss caused by reduced contrast, scattering, and object overlap. The system also maintained stable inference performance during continuous operation, indicating its feasibility for portable aquaculture monitoring. These findings demonstrate that combining computer vision with turbidity-aware calibration can improve the reliability of fingerling counting in challenging post-mining aquatic environments.