Explainable deep learning for SEM-based classification of six Sclerodermataceae taxa

Accurate identification of closely related fungal species remains challenging because of substantial morphological similarities and overlapping diagnostic characters. This study evaluated explainable deep learning approaches for the image-level classification of six predefined Sclerodermataceae taxa using scanning electron microscope (SEM) images of fungal spores An original dataset consisting of 961 SEM micrographs representing Scleroderma areolatum , S. cepa , S. bovista , S. polyrhizum , S. verrucosum , and Pisolithus arhizus was constructed under specimen-preparation protocol. Image preprocessing included grayscale conversion, brightness normalization, histogram equalization, and annotation removal. The dataset was divided into training, validation, and test subsets, and data augmentation was applied to improve model generalization. Nine pretrained convolutional neural network architectures, including ResNet-50, DenseNet-121, RegNet-Y8GF, ConvNeXt-Small, SENet-154, VGG-16, Inception-V3, Xception, and EfficientNet-B3, were comparatively evaluated using identical training conditions. Performance assessment was conducted using accuracy, precision, recall, F1-score, Matthews correlation coefficient, area under the receiver operating characteristic curve, balanced accuracy, Cohen’s kappa, expected calibration error, and brier score. ResNet-50 achieved the highest classification performance with 99.48% accuracy, 99.49% precision, 99.48% recall, 99.48% F1-score, and 99.38% MCC, misclassifying only one test sample. Grad-CAM++ and occlusion sensitivity analyses suggested that model attention was predominantly associated with biologically relevant spore ornamentation regions. These findings support the applicability of SEM-based deep learning and qualitative XAI approaches for image-level classification among the six predefined Sclerodermataceae taxa examined under the present experimental conditions.