Deep Learning Techniques for Identifying Fungi in Vine Wood Micrographs

Fungal infections represent a major challenge in agriculture, negatively impacting crop health, reducing yields, and leading to substantial financial losses. Early and accurate identification of these infections is vital for timely intervention and effective disease control. Traditional diagnostic techniques, including manual examination and laboratory analysis, are often slow, labor-intensive, and reliant on specialized expertise. To address these limitations, this study introduces an automated deep learning framework designed for fungal disease classification and segmentation. The proposed system integrates EfficientNetV2 and Vision Transformers (ViTs) for classification tasks and employs a CNN-Transformer hybrid model for detailed segmentation. The classification pipeline is trained on a dataset comprising highresolution TIFF images, divided into two categories “With Fungi” and “Without Fungi.” This classification model delivers a high performance, achieving a 97% accuracy, with precision, recall, and F1-scores all exceeding 95%, ensuring reliable differentiation between healthy and infected samples. The segmentation challenge detects fungal diseases in vine wood using a hybrid deep learning model that combines convolutional neural networks (CNNs) and transformers. The architecture incorporates a Feature Pyramid Network (FPN) for multi-scale feature processing, as well as a Convolutional Block Attention Module (CBAM) that focuses on crucial image regions to improve detection accuracy. Post-processing with Conditional Random Fields (CRFs) improves segmentation boundaries and reduces mistakes. The model performs well, with high Dice and IoU scores, indicating accurate fungal area mapping.