Field-Oriented Potato Disease Diagnosis: An Optimized EfficientNet-B0 Framework for Real-World Agricultural Deployment

Abstract

Potato crops face devastating 30-50% annual yield losses from diseases, threatening global food security. While deep learning shows promise, most systems fail in real-world conditions due to laboratory-centric development. This study develops an optimized EfficientNet-B0 framework achieving 83.19% accuracy across 7 disease categories in uncontrolled field environments the broadest coverage validated in real agricultural settings. Trained on 3,076 field images from Indonesian farms, our approach incorporates strategic data augmentation and class-aware optimization to handle environmental variability and severe class imbalance. The model achieves 95.35% accuracy on bacterial infections and 90% nematode detection despite minimal training samples. Grad-CAM interpretability confirms pathologically relevant feature localization with 85% alignment to disease-specific pathology. Unlike previous field studies achieving 70-85% accuracy on 3-5 classes, this work establishes the first pathologically-verified potato disease classifier for practical farmer deployment.