Optimization of deep learning models and applications of explainable artificial intelligence classification in some Tricholoma species

Abstract

Macrofungi are vital to ecosystem stability and biodiversity conservation, yet traditional species identification methods often require expert input and are prone to error. This study proposes a deep learning-based classification framework for six macrofungi species using advanced convolutional neural network (CNN) architectures. Among the tested models, DenseNet121 achieved a 90% F1-score, while EfficientNetV2-M reached 92% accuracy and 92% precision. AUC scores of 90% or higher were observed for DenseNet, EfficientNet-B4, EfficientNetV2-M, and ShuffleNet, demonstrating robust classification capabilities. Importantly, this study pioneers the integration of explainable artificial intelligence (XAI) in macrofungi classification by employing Grad-CAM. These visualizations revealed that top-performing models focused on biologically relevant features such as the cap, stem, and spore surfaces, whereas less accurate models fixated on background areas, leading to misclassification. The results highlight the importance of dataset size and variability in model performance and underscore the broader applicability of deep learning models beyond macrofungi, including spore-level analysis, plant disease diagnostics, and agricultural classification tasks. By enhancing interpretability and accuracy, this research provides a novel and effective approach for automated fungal identification, contributing significantly to biodiversity monitoring and ecological research.