This study classified red-tilapia weight ranges from UAV imagery using CNN-XGBoost and EfficientNet-B0-XGBoost hybrids. On 5×5 m images preserving cage context, CNN-XGBoost reached 98.8% mean accuracy and was far faster than EfficientNet. With context-limited 2×2 m images, EfficientNet-XGBoost performed better at 90.0%. Cross-farm, seasonal and water-condition validation is still needed.
Key findings
- For 5×5 m images, CNN-XGBoost achieved 0.988 ± 0.008 accuracy at 20 tuning units versus 0.977 ± 0.021 for EfficientNet-XGBoost at 40 units. Mean workflow time was 0.038 ± 0.001 seconds versus 65.007 ± 6.141 seconds per image. At 2×2 m, EfficientNet-XGBoost reached 0.900 ± 0.010 at 30 units, outperforming CNN-XGBoost at 0.850 ± 0.022.
Why this matters globally
Non-contact size assessment could reduce labour and fish stress while improving feeding and harvest planning. Fast wide-view inference is operationally attractive, but it must be linked to weight error, economic value and farm decisions.
Thai researcher contribution
Kasetsart University fisheries researchers linked cage-aquaculture expertise, UAV imaging and machine learning around a red-tilapia production problem.
Limitations to consider
The abstract does not detail image counts, farms, split strategy or weight-range labelling, so leakage and test independence cannot be assessed. Classification accuracy does not report gram-level error. Lighting, turbidity, waves, stocking density and cage design may reduce transfer performance.