Hybrid Deep Learning Framework for High-Accuracy Classification of Morphologically Similar Puffball Species Using CNN and Transformer Architectures

Puffballs, a group of macrofungi belonging to the <i>Basidiomycota</i>, pose taxonomic challenges due to their convergent morphological features, including spherical basidiocarps and similar peridial structures, which often hinder accurate species-level identification. This study proposes a deep learning-based classification framework for eight ecologically and taxonomically important puffball species: <i>Apioperdon pyriforme</i>, <i>Bovista plumbea</i>, <i>Bovistella utriformis</i>, <i>Lycoperdon echinatum</i>, <i>L. excipuliforme</i>, <i>L. molle</i>, <i>L. perlatum</i>, and <i>Mycenastrum corium</i>. A balanced dataset of 1600 images (200 per species) was used, divided into 70% training, 15% validation, and 15% testing. To enhance generalizability, images were augmented to simulate natural variability in orientation, lighting, and background. In this study, five different deep learning models (ConvNeXt-Base, Swin Transformer, ViT, MaxViT, EfficientNet-B3) were comparatively evaluated on a balanced dataset of eight puffball species. Among these, the ConvNeXt-Base model achieved the highest performance, with 95.41% accuracy, and proved especially effective in distinguishing morphologically similar species such as Mycenastrum corium and Lycoperdon excipuliforme. The findings demonstrate that deep learning models can serve as powerful tools for the accurate classification of visually similar fungal species. This technological approach shows promise for developing automated mushroom identification systems that support citizen science, amateur naturalists, and conservation professionals.