Explainable convolutional neural network architectures for high-performance taxonomic classification of gasteroid macrofungi

Gasteroid fungi represent a morphologically diverse and taxonomically challenging group due to their convergent evolution and closed fruiting bodies. This study presents a novel deep learning-based framework for the classification of six macrofungi species: Battarrea phalloides, Crucibulum laeve, Cyathus olla, C. striatus, Tulostoma brumale, and T. fimbriatum. A total of 1200 high-resolution images were processed using eleven convolutional neural networks (CNNs), each pre-trained on ImageNet and fine-tuned for this classification task. DenseNet121 emerged as the best-performing model with 96.11% accuracy, 96.09% F1-score, and an AUC of 99.89%. ResNeXt and RepVGG followed closely with accuracies of 95.00% and 93.89%, respectively. Operational metrics showed that ShuffleNetV2 achieved the fastest inference time (0.80 s), while RepVGG exhibited the highest energy efficiency (16.5%). Explainable AI techniques, including Grad-CAM and Guided Backpropagation, were applied to enhance model interpretability, revealing biologically meaningful image regions. These results demonstrate that deep learning architectures can be effectively applied to fungal taxonomy with high accuracy and transparency. Furthermore, the proposed methodology is scalable and adaptable to other biological domains such as fungal spore classification, plant pollen analysis, and rare species identification in ecological monitoring. This study offers a robust, interpretable, and computationally efficient solution for automating biodiversity assessments using image-based AI techniques.