Implementation of a corn leaf disease classification system using EfficientNetV2-S and transfer learning

Manual detection of corn leaf diseases makes it difficult for farmers to recognize early symptoms, resulting in delayed treatment. Meanwhile, deep learning methods generally require high computing power that is not compatible with simple devices in the field. To overcome this, this study developed a corn leaf disease classification system using the lightweight and efficient EfficientNetV2-S architecture. The research data came from a public dataset with four classes, namely Healthy, Leaf Blight, Common Rust, and Leaf Spot. The model was trained with transfer learning from ImageNet for 30 epochs (±22 minutes on GPU). The evaluation showed an accuracy of 84% with an average precision of 83%, recall of 84%, and F1-score of 83%. Per class, the best performance was obtained for Common Rust (F1 = 0.98) and Healthy (F1 = 0.88), while the lowest was for Leaf Spot (F1 = 0.64). The best results were obtained in the scenario without data augmentation. The system was implemented as a Flask-based web application optimized for mobile devices. The results of this study show that EfficientNetV2-S is efficient and performs well and has the potential to be a practical solution for corn disease detection at the farmer level.