Fusion-Based Explainable Deep Learning Framework for Species Classification of Jelly Fungi from Field Images

Abstract Jelly fungi are a morphologically diverse and taxonomically complex group within Basidiomycota, often exhibiting visually similar yet distinct traits. This study introduces a novel deep learning framework combining convolutional neural networks (CNNs), model fusion, and explainable AI (XAI) to automate and interpret the classification of jelly fungi. A balanced dataset of 1.800 high-resolution images representing nine species was developed under natural conditions. Ten CNN architectures were evaluated, with EdgeNeXT achieving 96.45% accuracy and 0.9990 AUC. The best performance came from a fusion model (EdgeNeXT + RepVGG), reaching 96.81% accuracy, 96.86% F1-score, 99.86% AUC, and 0.9641 MCC, while reducing misclassification between similar species from 20 to 3.1%. Grad-CAM and Integrated Gradients provided visual explanations aligned with relevant fungal structures, enhancing model interpretability. This is among the first studies applying fusion-based XAI to jelly fungi classification. The proposed method offers broad potential in biological applications such as spore and pollen identification, contributing to biodiversity monitoring and advancing AI-driven ecological informatics.