Usman Usman, Fatchul Arifin, Rustam Asnawi
Determining exactly when oil palm fruit is ripe is crucial for maximizing harvest efficiency, but current methods achieve only 65-75% accuracy, resulting in 15-20% annual losses. This research introduces a hybrid GA-PSO-SA metaheuristic optimization framework featuring self-adaptive dynamic attention for multi-objective CNN-based classification. The best CNN+GA+PSO setup used 10, 000 photos from five Indonesian plantation sites, with four ripeness classifications (Unripe, Underripe, Ripe, Overripe). It achieved 95.73% accuracy and 95.84% F1-score, indicating a better balance between bias and variance with only a 0.16% precision-recall gap. The suggested technique achieves a mean test accuracy of 96.12% ± 0.39% (max 96.86%) across 10 independent runs, with a narrow 95% confidence range [95.85%, 96.39%]. This is much better than baseline methods like MobileNetV2 (95.53%), ResNet50 (95.27%), and ViT-tiny (96.06%). The ablation study shows that dual-metaheuristic optimization is 18.23% better than single methods. However, adding more architectural complexity (SA, CBAM, SADA) worsens performance due to the curse of dimensionality. Leave-Site-Out cross-validation achieved 94.66% ± 0.54% accuracy across previously unobserved locations, thereby validating spatial generalization. The Pareto-optimal solution has a latency of 2.61 ms and a model size of 43.6 MB, which is 4.1 times quicker than ViT-tiny. This enables the deployment of precision agriculture apps on edge devices in real time. © This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License. License details: https://creativecommons.org/licenses/by-sa/4.0/
Department of Engineering Science, Faculty of Engineering, Yogyakarta State University, Yogyakarta, Indonesia; Department of Electronics and Informatics Engineering Education, Yogyakarta State University, Yogyakarta, Indonesia; Department of Electrical Engineering Education, Yogyakarta State University, Yogyakarta, Indonesia
