PlantHealthNet: Transformer-Enhanced Hybrid Models for Disease Diagnosis and Severity Estimation in Agriculture

Global plant diseases represent a major threat to agriculture and represent significant economic losses constituting an important threat to food security. This study introduces a transformative hybrid framework for plant disease diagnosis and severity estimation, combining the strengths of advanced deep learning architectures: Detection Transformer (DETR), Swin Transformer and SAM2-UNet. The proposed methodology incorporates classification, detection, segmentation, and severity estimation in the form of a unified pipeline to address the complexities of real-world agricultural settings, including variable lighting, background variety and large-scale fields. With a hierarchical attention mechanism for efficient feature extraction, the Swin Transformer achieves a top 1 accuracy of 85.6%, and a top 5 accuracy of 96.2%. Utilizing its encoder-decoder attention technique, Detection Transformer (DETR) effectively detects diseases by employing an encoder to forecast bounding boxes and class labels for affected areas. Segmentation performance is enhanced with SAM2-UNet, achieving a pixel-level segmentation with a Dice Similarity Coefficient (DSC) of 94.7%, cleanly delineating disease-affected areas. Finally, disease impact is quantified in the final severity prediction stage to allow for targeted interventions based on disease intensity. The framework has been extensively validated through experiments, and it performs significantly better on all the evaluation metrics: Average ROC AUC of 99.97%, Average Precision-Recall AUC of 98.47%, and F1-score of 94.46%. The suggested system improves precision agriculture by making disease detection and management reliable, scalable, and efficient. This work applies cutting-edge deep learning to sustainable agriculture to reduce crop losses, optimise resource use, and provide farmers with proactive disease management knowledge.