Combination Ensemble and Explainable Deep Learning Framework for High-Accuracy Classification of Wild Edible...

Area of Science:

• Mycology
• Computer Science
• Artificial Intelligence

Background:

• Accurate identification of wild edible macrofungi is vital for biodiversity, food safety, and ecological sustainability.
• Morphological similarities between edible and toxic species pose significant identification challenges.

Purpose of the Study:

• To develop and evaluate advanced machine learning models for precise wild edible macrofungi identification.
• To enhance model interpretability using explainable AI (XAI) techniques.

Main Methods:

• Analysis of a curated dataset of 24 wild edible macrofungi species.
• Benchmarking six state-of-the-art convolutional neural networks (CNNs) and four ensemble configurations.
• Integration of EfficientNetB0, ResNet50, and RegNetY in a hierarchical voting Combination Model.
• Application of XAI methods (Grad-CAM, Eigen-CAM, LIME) for model interpretability.

Main Results:

• The Combination Model achieved the highest accuracy (97.36%), AUC (0.9996), and MCC (0.9725).
• EfficientNetB0 performed best among individual CNNs (95.55% accuracy).
• XAI methods successfully highlighted biologically relevant regions, enhancing model transparency.

Conclusions:

• Engineered ensemble learning combined with XAI provides a robust and scalable solution for fine-grained fungal classification.
• This approach significantly advances mycological research and offers potential for broader applications in ecological monitoring and species recognition.