Miguel Andrei P. Algarme, Sealtiel Rae P. Villegas
Profiling mushrooms in the Philippines is crucial for their conservation, contributions to different fields of study, and raising health awareness regarding edible and non-edible species. The identification of mushrooms faces several challenges, including environmental climate conditions, the introduction of non-native species, and the variability in mushroom appearances across the country. To address these challenges, this study introduces an ensemble transfer learning model utilizing ResNet50, InceptionV3, and EfficientNet-B5 models. These models, pre-trained on ImageNet, have shown significant accuracy in image classification while maintaining a balance of efficiency and cost. The models are trained using 8,000 samples from the Mushroom Observer dataset as well as images of mushrooms collected from the Philippines. A stacked ensemble technique is employed, and XGBoost is used as the meta-model to combine the base models' predictions. The testing results achieved an accuracy of 94.25%, with precision at 92.08%, recall at 96.83%, and an F1-score of 94.39%. These results show the strength of the ensemble model in identifying the edibility of various mushroom species. This study contributes to advancements in image classification models and machine learning techniques, with a focus on promoting informed mushroom consumption and enhancing awareness across fields of study. © 2025 IEEE.
University of San Carlos, Department of Computer, Information Sciences, and Mathematics, Cebu City, Philippines
