Hybrid CNN–Transformer Model for Maize Leaf Blight Classification using Adaptive Genetic Optimization

Abstract

Abstract Maize leaf blight is a disastrous foliar disease in the world production of maize that causes significant losses in terms of yield annually. The classical machine learning and convolutional neural network (CNN) models are prone to poor generalization across the different conditions in the field because of differences in lighting, background, and leaf morphology. To counter these difficulties, this research suggests the use of Hybrid CNN-Transformer architecture that is trained using the Adaptive Genetic Optimization (AGO) to provide accurate classification of maize leaf blight. The hybrid model will utilize the spatial feature extraction and Global attention mechanism of Vision Transformer (ViT) to extract local and contextual patterns of diseases. AGO algorithm is dynamically adjusted with essential hyperparameters, such as the learning rate, filter size, and embedding, and adjusted according to the population diversity and fitness assessment, thus enhancing the convergence speed and classification accuracy. The experimental analysis of an augmented dataset of maize leaf disease proves that the developed model has a higher classification accuracy of 98.1, compared to the base models, including ResNet50 (96.7) and MobileNetV3 (97.2). The hybrid AGOCNN Transformer version demonstrates better solutions to the intelligent system on agricultural disease management in the real-field conditions.

Plant phenotyping relevance

トウモロコシ葉の病害状態を画像から分類するCNN・Transformer手法の開発と比較評価が研究の中心であり、植物病害フェノタイプの取得・推定に該当する。

Code and data availability

The paper uses a combined maize leaf image dataset (PlantVillage, Kaggle Maize Leaf Dataset, and field-acquired samples) but provides no public dataset link, no code repository, and no trained model release. The Data Availability statement explicitly says 'Not Applicable,' and no author URL or deposit is given for any,

No evidence-backed public reproduction asset is currently recorded.