A New Method for Counting Reproductive Structures in Digitized Herbarium Specimens Using Mask R-CNN

Area of Science:

• Botany and Ecology
• Computational Biology
• Climate Change Science

Background:

• Phenology, the timing of life-history events, is crucial for understanding organismal responses to climate change.
• Herbarium specimens offer a vast historical record of plant life, but manual data extraction is a bottleneck.
• Existing machine learning efforts often simplify phenological data to binary presence/absence.

Purpose of the Study:

• To develop and validate a machine learning pipeline for automated segmentation and counting of plant phenological features from herbarium specimens.
• To assess the accuracy of machine learning in quantifying reproductive stages (buds, flowers, fruits) across multiple species.
• To compare machine learning performance against crowd-sourced and expert data.

Main Methods:

• Utilized crowd-sourced phenological data from over 3,000 herbarium specimens of six wildflower species.
• Trained Mask R-CNN models to segment and count phenological features (buds, flowers, fruits).
• Evaluated model accuracy for binary coding, relative abundance estimation, and precise feature counting.

Main Results:

• A single global Mask R-CNN model achieved >87% accuracy for binary coding of reproductive stages.
• Relative abundance estimation reached ≥90% accuracy.
• Model performance varied by phenological stage and species, with flower counting being less accurate than buds or fruits.

Conclusions:

• The developed crowd-sourcing and machine learning pipeline effectively automates the extraction of high-quality phenological data from herbarium specimens.
• This approach significantly enhances the capacity to study plant responses to climate change.
• The model demonstrated transferability to different species, showing promise for broader applications in botanical research.