Fungi or Fatal: Ensemble Learning for Mushroom Edibility Classification in the Wild

Mushroom edibility classification is crucial for understanding biodiversity and ensuring public health. However, as many edible and poisonous mushrooms visually resemble each other, traditional expert-based classification methods are prone to errors. This study proposes a deep learning-based approach that automates mushroom classification using computer vision techniques. Experimental results indicate that pre-trained CNN models are negatively affected by background noise. To mitigate this issue, we incorporated YOLOv8-based instance segmentation to achieve more precise mushroom isolation. The existing dataset was re-annotated to support instance segmentation. The proposed approach improves accuracy by 8.80% over the baseline, achieving an overall accuracy of 87.13%. Building on this, we introduced an ensemble learning strategy, achieved an accuracy of 88.71% marking a total improvement of 10.76% over the baseline. These findings demonstrate that combining instance segmentation with ensemble deep learning significantly enhances the reliability of automated mushroom classification and lays the groundwork for more robust biodiversity analysis systems.