Analyzing trait-climate relationships within and among taxa using machine learning and herbarium specimens

Area of Science:

• Botany
• Computational Biology
• Ecology

Background:

• Large-scale plant trait studies are crucial for understanding environmental adaptation but face data limitations.
• Digitized herbarium collections offer a vast, underutilized resource for plant trait data.
• Automating leaf trait measurement from herbarium specimens can overcome data acquisition challenges.

Purpose of the Study:

• To develop and validate a machine learning approach for automated leaf identification and measurement from digitized herbarium specimens.
• To analyze the relationship between leaf size and climate across and within species for the genera Syzygium and Ficus.
• To assess the potential of machine learning on herbarium data for expanding global plant trait datasets.

Main Methods:

• Convolutional neural network (CNN) models were employed to detect and measure complete leaves in digitized herbarium images.
• Model performance was compared using user-selected versus random image training sets.
• Validated CNN models were applied to analyze leaf area and its association with climate variables in Syzygium and Ficus specimens.

Main Results:

• User-selected training yielded more comprehensive leaf measurements with greater size variation compared to random training.
• Leaf size demonstrated a positive association with temperature and rainfall across both genera.
• Within-species correlations between leaf size and environmental variables were notably weaker than across-species correlations.

Conclusions:

• Machine learning applied to herbarium specimens provides accurate leaf trait data, significantly expanding available datasets.
• Weak within-species trait-environment relationships suggest strong influences of population history and gene flow.
• This approach offers a scalable method to enhance trait data sampling for evolutionary studies.