Herbariograph: a deep-learning tool to classify specimen images.

Area of Science:

• Botany
• Computer Science
• Data Science

Background:

• Herbarium specimens represent global plant diversity and are vital research resources.
• While millions of herbarium specimens exist, only about 12.8% have associated digital images, limiting research potential.
• Automated sorting of specimen images is essential for unlocking the wealth of information they contain.

Purpose of the Study:

• To develop an automated method for recognizing and categorizing herbarium specimen image types.
• To introduce Herbariograph, an open dataset and deep-learning model for image classification in herbaria.

Main Methods:

• Creation of the Herbariograph dataset, comprising 17 categories with 12,288 images each, sourced from 43 institutions.
• Training a Convolutional Neural Network (CNN) model on the Herbariograph dataset for image recognition.
• Evaluation of the model's performance using a test macro F1 score.

Main Results:

• The Herbariograph dataset contains a diverse range of 17 image categories.
• The trained CNN model achieved a high test macro F1 score of 0.9611.
• The model demonstrates strong capability in automatically recognizing various herbarium image types.

Conclusions:

• Herbariograph provides a valuable resource for automating the processing of herbarium specimen images.
• This automation facilitates faster, more efficient, and accessible creation of targeted datasets for plant science research.
• The developed deep-learning model enhances the utility of digital herbarium collections for scientific discovery.