FusionViT CNN adaptive framework for insect associated plant leaf disease classification

The timely and proper diagnosis of the leaf diseases caused by insects is important in reducing the losses of crops and also maintaining food security. Nonetheless, traditional Convolutional Neural Network (CNNs) can tend to be more susceptible to false alarms due to a lack of features in minority disease classes. To solve this problem, we present FusionViT-CNN, a high-precision hybrid stacking ensemble which combines the global context-awareness of Vision Transformers with the local feature-extraction ability of CNNs. Through the stacking meta-learner ensemble strategy, this architecture has been able to suppress background noise and overcome the visual ambiguities that single models have difficulty solving. The proposed ensemble was tested on a mixture of 31,098 images and reached a state-of-the-art of 99.65 percent, surpassing conventional baselines such as ResNet50 (93.28 percent) and MobileNetV2 (96.30 percent). More importantly, our error analysis showed that the ensemble minimized diagnostic errors by 94.7% and the overall count of misclassified samples 209 (ResNet50) to only 11. In particular, it removed the issue of hallucination observed in minority classes with 100% Precision on Apple Cedar Apple Rust, where baseline models performed considerably poorly. We also checked the credibility of the model with the help of Explainable AI (XAI) methods; Grad-CAM and LIME visualizations have proven that the decision-making patterns of the model are biologically motivated, and it concentrates solely on necrotic lesions and rust pustules instead of environmental artefacts. These findings indicate that the FusionViT-CNN framework is a very reliable, accuracy based, solution to automated agricultural diagnostics.