Classification of Mycena and Marasmius Species Using Deep Learning Models: An Ecological and Taxonomic Approach.

Area of Science:

• Mycology
• Computational Biology
• Ecosystem Science

Background:

• Fungi are vital to ecosystems, offering biodiversity and biotechnological value.
• Accurate classification of fungal species is crucial for ecological and taxonomic studies.
• Existing methods may not fully capture the complex morphological and ecological traits of fungi.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for classifying seven macrofungi species from the *Mycena* and *Marasmius* genera.
• To integrate custom Convolutional Neural Network (CNN) with Self-Organizing Map (SOM) and Kolmogorov-Arnold Network (KAN) for enhanced fungal classification.
• To assess the performance of advanced pretrained models, including MaxViT-S and ResNetV2-50, for fungal species identification.

Main Methods:

• Development of a custom CNN integrated with a supervised SOM and a KAN layer.
• Utilizing unique ecological and morphological characteristics of macrofungi for classification.
• Employing advanced pretrained models (MaxViT-S, ResNetV2-50) for comparative analysis.
• Statistical validation of classification results using the chi-square test.

Main Results:

• The CNN-SOM and CNN-KAN architectures demonstrated significant improvements in classification metrics.
• MaxViT-S achieved a high accuracy rate of 98.9% in macrofungi classification.
• Chi-square tests confirmed the statistical reliability of the evaluation metrics.
• This study marks the first application of SOM for fungal classification.

Conclusions:

• Deep learning frameworks, particularly CNN-SOM and CNN-KAN, show significant potential for advancing fungal taxonomy.
• High accuracy achieved by models like MaxViT-S underscores the efficacy of deep learning in fungal identification.
• Future research should focus on KAN optimization and dataset expansion for broader fungal class coverage.