An explainable multimodal framework for early plant disease detection through deep learning and molecular data integration

Abstract

Early and accurate plant disease diagnosis is essential for sustainable crop production and effective disease management. This study presents an Explainable Gradient-Based Convolutional Neural Network (EG-CNN) that integrates leaf image data with gene expression and metabolomics information to enhance disease classification while providing transparent and interpretable predictions. The proposed framework was evaluated on four major plant diseases, powdery mildew, blight, anthracnose, and leaf spot, using a multimodal dataset comprising 10,000 images and associated omics data. Comparative analysis against Traditional CNN, ResNet-50, Vision Transformer (ViT), and CNN-LSTM models demonstrated the superior performance of EG-CNN. The model achieved 97.4% accuracy, 97.1% precision, 96.8% recall, and a 96.9% F1-score, outperforming all benchmark approaches. Paired t-test results revealed statistically significant improvements (p < 0.05) over competing models. Confusion matrix analysis indicated minimal misclassification across disease classes, while ROC analysis produced near-perfect AUC values, confirming excellent class separability, sensitivity, and specificity. Five-fold cross-validation further demonstrated robust generalization, with accuracy ranging from 97.1% to 97.6%. Grad-CAM visualizations successfully identified disease-relevant regions, including lesions, discoloration, and necrotic tissues, thereby enhancing model transparency and user trust. The integration of omics data also improved biological interpretability by linking predictions to underlying molecular responses. In conclusion, the proposed EG-CNN framework provides a highly accurate, robust, and explainable solution for plant disease diagnosis. Its multimodal architecture and strong generalization capability make it a promising tool for precision agriculture and real-time disease monitoring applications.