Detection of wheat panicle diseases with synthetic oversampling and quantum kernel-enhanced ResNet50-SVM

Abstract Early detection of wheat panicle diseases plays a vital role in protecting crop health and maintaining food security. This study proposes a hybrid deep learning and machine learning framework for the robust classification of four major wheat panicle diseases: This study presents a hybrid framework that accurately classifies four significant wheat panicle diseases which include Blast, Fusarium Head Blight, Loose Smut and Smut. Using ResNet50 through transfer learning, deep features were extracted from preprocessed and augmented image datasets. The deep feature space receives the Synthetic Minority Oversampling Technique (SMOTE) application to resolve class imbalance and improve model generalization. To achieve better nonlinear separability and decision boundaries, oversampled features are used to train a quantum kernel-enhanced Support Vector Machine (SVM) classifier. The framework achieved 97.58% validation accuracy and demonstrated strong performance across several evaluation measures, with a sensitivity of 90.42%, specificity of 98.57%, precision of 96.74%, F1-score of 93.07%, and MCC of 0.9213. The model demonstrated reliable performance through its low false-positive rate of 1.43% and high Cohen’s Kappa score of 0.9125. The classification accuracy of the SVM classifier was significantly improved through quantum kernels when managing the overlapping feature distributions. This study introduces an automated detection technique for wheat panicle infections that can be adapted for use in precision agriculture.