Efficient and explainable deep learning for fine-grained classification of morphologically overlapping species of the truffle genus Elaphomyces

Accurate discrimination of morphologically overlapping macrofungal species remains a persistent challenge in fungal taxonomy, particularly when species-level differences are restricted to subtle ultrastructural characters. This study presents an explainable deep learning-based framework for the fine-grained classification of four closely related Elaphomyces species using high-resolution scanning electron microscopy images of ascospore surfaces. Rather than focusing solely on model accuracy, the study evaluates how different architectural designs and hybrid decision strategies capture biologically meaningful micromorphological patterns. The results showed that lightweight and medium-scale models generally produced more stable and reliable classifications than several deeper architectures, indicating that network depth alone does not guarantee improved performance in limited biological image datasets. The best hybrid configuration achieved 98.13% accuracy and provided interpretable visual evidence associated with taxonomically informative spore surface traits. Under the predefined independent hold-out design, the findings indicate that model efficiency, classifier compatibility, and biological interpretability may be more influential than architectural complexity alone for the classification of the examined fungal images. Overall, the proposed framework offers an objective and interpretable preliminary framework for supporting fungal taxonomy, biodiversity assessment, and other biological classification tasks involving subtle phenotypic variation.