Classifying fungi biodiversity using hybrid transformer models.

Area of Science:

• Ecology and Biodiversity
• Computational Biology
• Mycology

Background:

• Fungi are crucial for ecosystems, impacting nutrient cycling, agriculture, and medicine.
• Accurate fungal species classification is vital for understanding biodiversity and harnessing ecological benefits.
• Existing classification methods can be improved with advanced computational techniques.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model for multiclass fungal classification.
• To merge Vision Transformer and Swin Transformer models with transfer learning frameworks.
• To enhance the accuracy and efficiency of fungal species identification.

Main Methods:

• Utilized a dataset of 9115 fungal images across five species.
• Employed data augmentation to address class imbalance.
• Integrated Vision Transformer (ViT) and Swin Transformer with MobileNetV2, DenseNet121, and EfficientNetB0.
• Applied five-fold cross-validation and Grad-CAM for model interpretability.

Main Results:

• The Swin Transformer combined with DenseNet121 achieved the highest accuracy (96.96% training, 95.97% validation, 95.57% testing).
• Hybrid models demonstrated effective generalization and minimized misclassifications.
• Grad-CAM visualizations confirmed model focus on biologically relevant fungal structures.

Conclusions:

• Hybrid deep learning models offer a scalable and efficient approach for complex fungal classification tasks.
• These models balance local feature modeling with global context capture.
• The findings support improved fungal management, biodiversity conservation, sustainable agriculture, and medical diagnostics.