HYBRID QUANTUM–CLASSICAL DEEP LEARNING FRAMEWORK FOR IMBALANCED MULTI-CLASS RICE LEAF DISEASE CLASSIFICATION

Abstract

Rice leaf diseases are a threat to agricultural productivity and food security, particularly in regions where access to expert diagnosis is limited. While deep learning models have shown great accuracy in disease classification, their usefulness is often hindered by extreme class imbalance, poor performance on rare diseases, and increasing computational complexity. On the other hand, hybrid quantum-classical machine learning has recently been explored as a means to enhance feature representation, although its use in large-scale agricultural image analysis is still in its infancy. This paper presents a hybrid quantum-classical neural network for multi-class classification of rice leaf diseases, integrating a pre-trained DenseNet121 feature extractor with a six-qubit variational quantum circuit. The network is optimized in an end-to-end manner that is NISQ-friendly. For handling the problem of class imbalance in the dataset, stratified data splitting and dynamic oversampling are employed during the training process. The experiments are performed on a large class-imbalanced dataset with eight classes of rice leaf diseases. The proposed framework has been able to obtain 98.0% accuracy with a macro-averaged precision, recall, and F1-score of 0.98, which signifies a balanced performance on the classes. The weighted AUC-ROC value of 0.9975 is another proof of its high discriminatory power. The ablation study shows a 5.7% improvement over the classical baselines, which signifies the effectiveness of quantum-enhanced feature transformation. These outcomes show that the hybrid quantum-classical learning approach offers a feasible solution to the existing limitations in precision agriculture.