Graph-based feature integration for Mango leaf image reconstruction and classification using hybrid GCNN-vision transformer

Abstract

This manuscript presents a comprehensive study on automated systems for identifying and classifying mango leaf diseases, crucial for timely agricultural interventions. We employed advanced machine learning and deep learning techniques, including convolutional neural networks (CNNs) and pre-trained networks such as VGG16, AlexNet, DenseNet, and Vision Transformer (ViT), to enhance classification accuracy through feature integration and reconstruction. The proposed graph-based model, Graph Convolutional Neural Networks (GCNN), integrates CNNs’ hierarchical feature learning with ViT's global context understanding, resulting in a novel approach, GCNN-ViT, for effective mango leaf analysis. Our methodology comprises three phases: (1) Convolutional layer and graph node transformation to capture spatial dependencies; (2) Graph convolution and deconvolution to enhance complex relationship detection and accurate representation; and (3) Classification using ViT for precise disease pattern identification. Comparative analysis of GCNN-ViT with CNN, VGG16, AlexNet, DenseNet, and ViT demonstrates superior performance metrics. GCNN-ViT achieved perfect scores with accuracy, specificity, sensitivity, F1-Score, and AUC, featuring 251 true positives, no false negatives or positives, and 161 true negatives. In contrast, ViT also performed exceptionally well, while traditional CNN architectures showed improvements over the baseline model. GCNN-ViT's flawless AUC score of 1.00 and a mean accuracy of 0.998 across k-fold cross-validation highlight its robustness and effectiveness. A Chi-square value of 8.7 further supports the model's fit. These results underscore the GCNN-ViT model's significant advancements in mango leaf disease classification, showcasing its potential for real-world applications in agriculture and plant pathology, where enhanced accuracy and efficiency are critical.