<![CDATA[The utilization of abandoned tin mining ponds for freshwater aquaculture in Bangka Belitung requires reliable monitoring technology, particularly for fingerling counting under variable water clarity conditions. Because in-situ underwater counting within large, deep, and unconfined pond waters is technically impractical, this study developed an on-bank portable inspection station integrating YOLOv5-based image processing on an embedded Raspberry Pi unit with multi-point turbidity sensing using water extracted directly from abandoned tin mining ponds. A dataset of 80 original images was strictly partitioned prior to augmentation into training (70%), validation (20%), and independent test (10%) sets to ensure zero data leakage, followed by training set expansion to 225 images via offline augmentations. Trained at 640×640 resolution across 50 epochs, the YOLOv5 model achieved 98% accuracy, 100% precision, 96% recall, 100% specificity, and an F1-score of 0.98 at an IoU threshold of 0.5. While vision-only detection remained stable at low densities, severe under-counting (up to 26%–30% loss) occurred under high-density conditions (40–50 fingerlings) when turbidity exceeded 85 NTU due to light scattering and boundary occlusion. To address this limitation, an Adaptive Turbidity Correction Factor (α = 1.36) was introduced, successfully restoring biomass estimation accuracy to over 96%. The embedded prototype maintained an average inference speed of 10.5 FPS with a peak CPU temperature of 67.2°C during a continuous 2-hour stress test. These findings confirm that combining edge vision with turbidity-aware calibration substantially improves fingerling counting reliability in challenging aquatic environments of abandoned tin mining ponds.]]>
Copyrights © 2026
