Efficient Ensemble of Pre-Trained CNN Models for Improving Classification of Maize Diseases

Abstract

Maize is one of the world's leading staples crops, whose yields are compromised by foliar diseases such as rust, blight, and gray leaf spot. Speedy and correct diagnosis must be employed to reduce loss and maintain food security. Deep learning or Convolutional Neural Networks (CNNs) has been promising for auto-classifying diseases, but one-model systems are not robust and generalizable. Moreover, previous research did not address issues of dataset imbalance leaf orientation variation in the out-worldly setting, and field deployability for precision farming. In this research, a feature ensemble-fusion classifier model for maize disease identification is presented. Pre-processing and augmentation were performed on a 4,188-leaf image dataset divided into four classes (Common Rust, Gray Leaf Spot, Blight, and Healthy). Six pre-trained CNNs (EfficientNetB0, DenseNet201, ResNet50V2, NasNetMobile, MobileNetV2, and VGG16) were tested in frozen and partially fine-tuned states. EfficientNetB0, DenseNet201, and ResNet50V2 were the top three models averaged using an ensemble approach. The proposed system achieved a total average of 98% accuracy across all diseases, with 97% precision, 96% recall, and 96.5% F1-score, and at least 92% precision, 96% recall, and 96% F1-score on Gray Leaf Spot disease thus performing better than single CNNs. With improved class imbalance handling and environmental robustness, the scheme offers an efficient and adaptive solution to real-world maize disease diagnosis and precision agriculture.