IndoHerb: Indonesia medicinal plants recognition using transfer learning and deep learning.

Area of Science:

• Ethnobotany
• Computer Vision
• Machine Learning

Background:

• Indonesia possesses rich herbal plant diversity crucial for traditional healing and ethnobotany.
• Modernization threatens the recognition and preservation of valuable Indonesian herbal plant heritage.
• Accurate plant identification is vital for maintaining traditional practices and utilizing nutritional benefits.

Purpose of the Study:

• To develop an efficient method for identifying Indonesian herbal plants using computer vision.
• To classify Indonesian herbal plants through transfer learning of Convolutional Neural Networks (CNNs).

Main Methods:

• Curated an extensive dataset of Indonesian herbal plant images.
• Applied rigorous data preprocessing techniques.
• Utilized transfer learning with five CNN models (ResNet, DenseNet, VGG, ConvNeXt, Swin Transformer) for classification.

Main Results:

• ConvNeXt model achieved the highest classification accuracy of 92.5%.
• A model trained from scratch yielded an accuracy of 53.9%.
• Key hyperparameters included ExponentialLR scheduler, Adam optimizer, Cross-Entropy Loss, and 50 training epochs.

Conclusions:

• Automated identification of Indonesian medicinal plants is feasible and effective using CNN transfer learning.
• This technology can aid in preserving ethnobotanical knowledge.
• Enhances agricultural practices through better cultivation of valuable medicinal resources.