A Comparative Analysis of CNN Architectures, Fusion Strategies, and Explainable AI for Fine-Grained Macrofungi Classification

Abstract

This study was motivated by the persistent difficulty of accurately identifying morphologically similar macrofungi species, which remains a significant challenge in fungal taxonomy and biodiversity monitoring. This study presents a deep learning framework for the automated classification of seven morphologically similar coprinoid macrofungi species. A curated dataset of 1692 high-resolution images was used to evaluate ten state-of-the-art convolutional neural networks (CNNs) and three novel fusion models. The Dual Path Network (DPN) achieved the highest performance as a single model with 89.35% accuracy, a 0.8764 Matthews Correlation Coefficient (MCC), and a 0.9886 Area Under the Curve (AUC). The feature-level fusion of Xception and DPN yielded competitive results, reaching 88.89% accuracy and 0.8803 MCC, demonstrating the synergistic potential of combining architectures. In contrast, lighter models like LCNet and MixNet showed lower performance, achieving only 72.05% accuracy. Explainable AI (XAI) techniques, including Grad-CAM and Integrated Gradients, confirmed that high-performing models focused accurately on discriminative morphological structures such as caps and gills. The results underscore the efficacy of deep learning, particularly deeper architectures and strategic fusion models, in overcoming the challenges of fine-grained visual classification in mycology. This work provides a robust, interpretable computational tool for automated fungal identification, with significant implications for biodiversity research and taxonomic studies.