Harnessing Large-Scale Herbarium Image Datasets Through Representation Learning.

Area of Science:

• Botany
• Computer Science
• Data Science

Background:

• Herbarium digitization generates large image datasets suitable for machine learning.
• Barriers to adoption include limited computational resources and uneven digitization progress.
• Representation learning offers a potential solution for extracting valuable data from specimen images.

Purpose of the Study:

• Investigate the potential of representation learning for herbarium specimen images.
• Develop and compare deep learning models for extracting useful image representations.
• Assess the transferability of learned representations across different research applications.

Main Methods:

• Built three neural networks using a dataset of over 2 million herbarium specimen images.
• Employed representation learning techniques to extract image features.
• Compared the performance of different network architectures and their extracted representations.

Main Results:

• A triplet network demonstrated superior performance in transferring learned representations across various applications.
• The study successfully extracted herbarium specimen image representations applicable to diverse research tasks.
• Results indicate the feasibility of using deep learning for analyzing large-scale herbarium data.

Conclusions:

• Representation learning is a viable approach to unlock the information within digitized herbarium specimens.
• Triplet networks show significant potential for creating generalizable feature representations.
• Further development is needed to fully leverage representation learning for global herbarium data analysis.