Explainable deep learning ensemble framework for accurate classification of wild poisonous mushroom species

This study introduces an explainable deep learning framework for the accurate classification of wild poisonous mushroom species, contributing to food safety. A balanced dataset of 3600 high-resolution images representing 18 species was employed, split into training, validation, and test subsets. To enhance variability and reduce overfitting, the images were expanded using advanced augmentation techniques (rotation, flipping, brightness/contrast adjustments, noise injection, etc.), increasing the training set to 7200 samples. Four pretrained CNN architectures DenseNet121, EfficientNet-B3, MobileNet-V3, and ShuffleNet-V2 were fine-tuned via transfer learning and evaluated with multiple performance metrics. Among the individual models, EfficientNet-B3 achieved the highest accuracy of 93.0%. However, ensemble strategies based on soft voting consistently outperformed single models. The four-model ensemble (DenseNet121, EfficientNet-B3, MobileNet-V3, ShuffleNet-V2) achieved the best results with 95.67% accuracy, 95.42% MCC, and a log loss of 0.175. Explainable AI methods (Grad-CAM, Grad-CAM++) revealed that classification decisions corresponded to biologically meaningful regions, thereby improving interpretability and reliability. This study holds direct life-saving potential by reducing poisoning incidents caused by misidentifications. In addition, it contributes to food safety by supporting reliable identification of toxic species within the agricultural and food supply chain. Furthermore, it pioneers the integration of AI methodologies in fungal taxonomy, providing a robust foundation for future ecological, agricultural, and biotechnological research.