Application of a hybrid deep learning architecture for analysing wheat diseases

This paper presents a comparative analysis of deep learning architectures for the automated classification of wheat leaf diseases using RGB image data. The study introduces a hybrid neural network model that integrates a convolutional backbone (ResNet50) with a Vision Transformer (ViT-B16) encoder to combine local feature extraction with global context modeling. A dataset of 5987 annotated images was compiled from public sources and augmented to address class imbalance. The proposed hybrid model was trained using fine-tuning techniques with mixed precision and evaluated alongside state-of-the-art CNNs (EfficientNet-B0, VGG16, MobileNetV3), Transformer-based models (ViT-B16, Swin Transformer), and traditional classifiers (SVM, Random Forest). Experimental results show that the hybrid architecture achieved the highest performance with 93.8% test accuracy and 93.7% F1-score, outperforming all baselines. Interpretability was ensured via Grad-CAM and UMAP projections, while model checkpoints and metrics were tracked using Weights & Biases. The study highlights the advantages of combining convolutional and attention-based mechanisms for robust plant disease recognition. Future work will include additional disease categories, the use of field-collected images, and integration with IoT and environmental data sources to support real-time and context-aware diagnostics.