International Journal of Science, Engineering and TechnologyAuthors: Dr. Pankaj Malik, Rahul Verma, Ashish Chourey, Sameer Tandon, Riddhi Yadav
Abstract: Litchi (Litchi chinensis Sonn.) is a high-value tropical fruit crop that is highly susceptible to various diseases, including anthracnose, fruit rot, pericarp browning, sooty mold, and algal spot, which significantly reduce fruit quality, market value, and yield. Accurate and timely disease diagnosis is essential for effective disease management and sustainable litchi production. Traditional visual inspection methods are labor-intensive, subjective, and often incapable of detecting diseases at early stages. Recent advances in deep learning have demonstrated promising results for automated plant disease detection; however, conventional Convolutional Neural Networks (CNNs) primarily focus on local features and often fail to capture global contextual information. Furthermore, the lack of model interpretability limits their practical adoption in smart agriculture applications. To address these challenges, this study proposes a novel Hybrid CNN–Vision Transformer (HCVT) framework for multi-class litchi fruit disease detection and severity assessment. The proposed model integrates EfficientNet-B3 and Vision Transformer (ViT) architectures through an attention-based feature fusion mechanism, enabling simultaneous extraction of local lesion characteristics and global contextual representations. The framework incorporates comprehensive image preprocessing, data augmentation, explainable artificial intelligence (XAI) techniques using Grad-CAM and SHAP, and a disease severity estimation module based on lesion segmentation. Experimental evaluation was conducted on a multi-class litchi disease dataset comprising healthy and diseased fruit images. The proposed HCVT model achieved an overall classification accuracy of 97.63%, precision of 97.31%, recall of 97.05%, F1-score of 97.18%, and AUC of 0.983, outperforming standalone CNN, Vision Transformer, and conventional hybrid models. Five-fold cross-validation confirmed the robustness and generalization capability of the framework. Grad-CAM visualizations successfully localized disease-affected regions, while SHAP analysis identified lesion texture, color variation, and infected area characteristics as the most influential features. Additionally, the proposed severity estimation module effectively quantified disease progression using lesion-area analysis. The results demonstrate that the proposed HCVT framework provides a highly accurate, interpretable, and reliable solution for automated litchi disease diagnosis. The integration of deep learning, transformer-based representation learning, explainable AI, and severity estimation offers significant potential for deployment in smart agriculture, precision farming, and future UAV-based crop health monitoring systems.
