Refining CNN-Based Models for Multi-Class Corn Leaf Disease Classification

Corn leaf disease significantly impacts agricultural productivity and national food security, particularly in regions with high dependence on maize as a staple commodity. Manual disease identification remains challenging due to the need for expert agronomists, inconsistent environmental conditions, and visual similarities among disease patterns, often resulting in delayed decision-making and inaccurate control measures. Deep learning-based image classification has emerged as an effective solution for plant disease identification; however, existing models often face limitations regarding overfitting, poor generalization, and insufficient performance when applied to multi-class agricultural image datasets. Therefore, this research aims to develop an Improved EfficientNetB0 model for the multi-class classification of maize leaf diseases comprising Healthy, Leaf Blight, Leaf Rust, and Leaf Spot categories. A dataset of 4,000 images was used and processed through resizing, normalization, and augmentation techniques. Five CNN backbones; EfficientNetB0, MobileNetV2, ResNet50, DenseNet121, and InceptionV3—were initially evaluated, and EfficientNetB0 demonstrated the highest baseline performance. The model was subsequently enhanced through fine-tuning, regularization (dropout and batch normalization), and cosine learning rate scheduling. Experimental results show that the Improved EfficientNetB0 achieved superior performance with an accuracy of 0.9671, macro precision of 0.9665, macro recall of 0.9666, and macro F1-score of 0.9661, exceeding all baseline models. These findings demonstrate that the proposed framework effectively improves maize disease classification accuracy and contributes a robust solution for smart agriculture applications. Future work may integrate real-time deployment and mobile-based decision support for field-level monitoring.