Generating segmentation masks of herbarium specimens and a data set for training segmentation models using deep...

Area of Science:

• Botany
• Computer Science
• Digital Imaging

Background:

• Digitized herbarium specimens contain visual noise and bias from digitization and mounting processes.
• Accurate segmentation of plant tissues is crucial for deriving reliable biological insights from herbarium data.
• Existing methods may not adequately address the challenges of noise and bias in herbarium specimen imagery.

Purpose of the Study:

• To develop a deep learning workflow for segmenting plant tissues in herbarium specimen images.
• To create a dataset of high-resolution image masks for training deep learning models.
• To systematically remove noise and minimize bias in digitized herbarium specimens.

Main Methods:

• Generated 400 high-resolution image masks of ferns using automated and manual techniques.
• Trained a U-Net-style deep learning model for image segmentation using the curated masks.
• Achieved a Sørensen-Dice coefficient of 0.96 for accurate segmentation.

Main Results:

• Developed a deep learning model capable of efficient and accurate segmentation of herbarium specimen images.
• The model effectively segments plant tissues, particularly for ferns, removing background pixels.
• Demonstrated the potential for automated processing of large-scale digitized herbarium collections.

Conclusions:

• Deep learning in herbarium science necessitates transparent protocols for training data generation.
• Shared segmentation ground-truth masks and the trained model to facilitate broader herbarium applications.
• Encourages further development and transfer learning opportunities for herbarium data analysis.