複数の植物病の有効な診断のためのアンサンブル異種トランスフォーマーモデル

Plant diseases are a significant challenge to sustainable farming resulting in drastic losses of crop quality and quantity. Conventional diagnostic procedures like manual examination and single-model deep learning-based methods tend to be ineffective in identifying overlapping appearances, detailed textures of leaves, and environmental changes, which results in inconsistent performance. In order to address these issues, this paper presents an ensemble transformer framework that incorporates the segmentation, classification and optimization to identify multi-diseases in plants accurately. The framework has a two phase design. At the initial stage, U-Net and Swin Transformer V2 detect the disease-affected leaf areas with high accuracy, and the important features are correctly captured. In the second stage, classification is carried out using CoAtNet and its enhanced variant, which combine convolutional feature extraction with transformer-based global context learning. To further improve decision-making, a meta-heuristic fusion strategy based on the Levy Flight Honey Badger Algorithm dynamically weights classifier outputs, enhancing robustness and reducing misclassifications. Model interpretability is enhanced through GRAD-CAM visualizations, providing clear insights into the regions influencing disease classification. The framework was extensively evaluated on the PlantVillage dataset containing 54,305 images across 38 classes. Results demonstrate outstanding performance, with 99.31% accuracy, 99.32% precision, 99.31% recall, 99.32% specificity, and 99.31% F1-score. The ensemble segmentation approach exhibits a statistically significant 7.34% improvement over single-method implementations. Moreover, the heterogeneous ensemble model achieves 8.43% and 14.59% superiority over homogeneous ensembles and individual models, respectively. The integration of segmentation, hybrid transformer architectures, and meta-heuristic decision fusion delivers a powerful, interpretable, and highly reliable solution for early plant disease detection, offering strong potential for real-world agricultural deployment.