ViT-U-Net fusion model for accurate localization and recognition of plant leaf diseases

Plant disease diagnosis is a critical component of precision agriculture, enabling early detection and mitigation of crop losses. With the rapid growth of deep learning, vision-based models have become an essential tool for automated plant pathology. However, many existing approaches focus solely on classification and fail to provide lesion-level localization, limiting their practical applicability in real field environments. Additionally, background noise, variable lighting, and overlapping disease regions often reduce classification reliability. This study utilizes the New Plant Diseases Dataset (Kaggle), comprising over 87,000 annotated images spanning 38 plant disease classes. The dataset offers sufficient variety for evaluating both disease recognition and lesion segmentation tasks. The proposed methodology introduces a hybrid ViT–U-Net Fusion Model, where a Vision Transformer encoder extracts global context while a U-Net decoder produces high-resolution lesion masks. A segmentation-aware classifier then uses the extracted ROI to enhance disease identification. The novelty lies in the joint segmentation–classification framework, enabling simultaneous disease localization and recognition within a unified pipeline. Performance is assessed using Accuracy, F1-score, IoU, Dice coefficient, and robustness metrics. The model achieves 99.21% classification accuracy and 93.87% IoU, outperforming baseline modules and demonstrating significant improvements validated through statistical analysis.