Advanced Deep Learning Approaches for Accurate Macrofungi Species Classification Using Optimized ResNet-150 and Vision Transformer

Read full text Summarise with VIRAT
Authors: Duong Thi Kim, Nghi N. Nguyen, Thanh Quang Nguyen
Cited by: 0
References: 70

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate identification of macrofungi is critical for biodiversity monitoring, ecological research, and preventing mushroom poisoning incidents. However, conventional identification based on morphology and molecular barcoding is time-consuming and requires expert knowledge. In this study, we investigate advanced deep learning approaches for fine-grained classification of 47 macrofungi species, including 25 poisonous and 22 non-poisonous classes, using a real-world image dataset of 2,820 samples collected under unconstrained conditions. We develop and evaluate two deep models: (i) an optimized ResNet-150 transfer learning pipeline and (ii) a Vision Transformer (ViT-L/16) model. The proposed ResNet-150 pipeline incorporates a lightweight task-specific classification head and an aggressive augmentation strategy tailored to small macrofungi datasets, enabling robust learning despite limited samples. Experimental results show that the improved ResNet-150 achieves a test accuracy of 93%, outperforming ViT-L/16 (91.4% accuracy) and previously published mushroom classification systems based on Swin Transformer and DenseNet-121. The ResNet-150 model also attains macro-averaged precision, recall, and F1-score of 0.95, 0.93, and 0.93, respectively, demonstrating strong balanced performance across 47 species. Beyond predictive accuracy, we analyze model behavior using attention visualization to highlight key morphological regions that drive decisions, and discuss how these attention maps can be integrated with biochemical data and ITS sequences in future multimodal frameworks. The findings indicate that optimized convolutional backbones remain highly competitive on small, fine-grained biological datasets, while transformer-based architectures open promising directions for interpretable macrofungi characterization. The proposed framework provides a practical and extensible baseline for AI-assisted mushroom identification in smart agriculture and food safety applications.

Fields of Study & Subjects

Advanced Neural Network Applications Smart Agriculture and AI Domain Adaptation and Few-Shot Learning

Publication details

Journal / source
Research Square
Publication date
31 March 2026
Language
EN
Document type
Preprint