Automatic Fungi Recognition: Deep Learning Meets Mycology.

Area of Science:

• Mycology
• Artificial Intelligence
• Citizen Science

Background:

• Citizen science initiatives can significantly enhance ecological data collection.
• Accurate species identification is crucial for biodiversity monitoring and research.
• AI-powered tools offer potential solutions for real-time species recognition.

Purpose of the Study:

• To develop and deploy an AI-based system (FungiVision) for real-time fungi species recognition.
• To create a novel, fine-grained fungi classification dataset (Danish Fungi 2020) with rich metadata.
• To improve fungi recognition accuracy using AI and citizen-collected data.

Main Methods:

• Development of FungiVision, a mobile application with a human-in-the-loop AI for real-time fungi identification.
• Creation of the Danish Fungi 2020 dataset, featuring species-level labels and observation metadata.
• Implementation of a Vision Transformer architecture for fungi recognition, leveraging metadata.

Main Results:

• FungiVision increased citizen participation in data collection by 400%.
• The system achieved nearly 93% accuracy with human-in-the-loop validation.
• The novel AI model, trained on the DF20 dataset, reduced recognition error by 46.75% compared to the previous system.

Conclusions:

• AI-driven tools like FungiVision can effectively engage the public in citizen science.
• The integration of metadata significantly enhances the accuracy of AI-based species classification.
• A collaborative approach, with continuous data flow, fosters a virtuous cycle of improvement for both AI systems and scientific understanding.