Towards Reliable and Interpretable Wheat Disease Diagnosis: Unified Framework Incorporating Vision Transformer and Advanced Object Detection

Wheat diseases threaten global food security, leading to yield losses exceeding 20% annually. This study presents a comparative evaluation of advanced deep learning architectures for automated wheat disease detection. We assessed five transformer-based classification models (MaxViT, Swin, MViTv2, DAvit, RDNet) and five object detection models (YOLOv7, YOLOv10, YOLOv12, RT-DETR, RT-DETRv3) using a large dataset of 14,155 images spanning 15 disease categories. Among classification models, MaxViT achieved the highest accuracy at 97.83%, while YOLOv12 demonstrated the best detection performance ($94.4 \% \text{mAP} {@} 0.5$) alongside superior computational efficiency. The results show that the unified framework, combining YOLOv12 and MaxViT achieved an overall accuracy of 97.62%. Gradient-weighted Class Activation Mapping confirmed that the models focused on biologically relevant features, reinforcing their diagnostic reliability. Our findings highlight that state-of-the-art architectures can be effectively leveraged for agro-diagnostic applications, helping mitigate crop losses and strengthen food security. This work contributes to precision agriculture by providing guidance on selecting practical deep learning models tailored to specific constraints and operational needs.