Automated Real-Time Identification of Medicinal Plants Species in Natural Environment Using Deep Learning Models-A...

Area of Science:

• Botany
• Computer Science
• Biodiversity Studies

Background:

• Manual plant identification is crucial for biodiversity management but is labor-intensive.
• Automated methods using digital image processing and pattern recognition are needed for efficient species identification.
• Existing automated systems face challenges in achieving high accuracy for plant species identification.

Purpose of the Study:

• To design and develop an automated real-time plant species identification system for medicinal plants in Borneo.
• To leverage computer vision and deep learning for accurate and efficient plant identification.
• To create a user-friendly mobile application for plant identification and data collection.

Main Methods:

• Developed a system comprising a computer vision module, a dynamic knowledge base, and a mobile application.
• Utilized an EfficientNet-B1 deep learning model trained on a combined dataset of public and private plant images.
• Integrated crowdsourcing feedback and geo-mapping features within the mobile application.

Main Results:

• The EfficientNet-B1 model achieved 87% (private) and 84% (public) Top-1 accuracies on test sets, surpassing baseline models by over 10%.
• Real-time mobile application testing yielded 78.5% (Top-1) and 82.6% (Top-5) accuracies.
• The system demonstrated a promising direction for automated plant species identification.

Conclusions:

• The developed automated system offers a significant improvement over manual plant identification methods.
• The mobile application facilitates real-time identification, data collection, and community feedback for Borneo's medicinal plants.
• Further refinement may address accuracy variations observed during real-time application testing.