Deep learning to capture leaf shape in plant images: Validation by geometric morphometrics.

Area of Science:

• Botany
• Computer Science
• Ecology

Background:

• Automated plant species identification relies on deep learning (DL) models analyzing leaf images.
• Reproducibly capturing leaf shape variation with DL is challenging due to the
• black box
• nature of these models.

Purpose of the Study:

• To evaluate DL's effectiveness in capturing leaf shape variation.
• To use geometric morphometrics (GM) as an eXplainable Artificial Intelligence (XAI) tool to interpret DL features.
• To assess the association between DL-extracted features and biologically relevant leaf shape variation.

Main Methods:

• Photographed Ranunculus auricomus leaves in situ and after herbarization.
• Extracted DL features from leaf images using a neural network.
• Digitized leaf shapes using GM and analyzed the association between DL features and GM shapes via dimension reduction and covariation models.

Main Results:

• DL features successfully clustered leaf images by source populations for both in situ and herbarized datasets.
• Significant associations were found between certain DL features and biological leaf shape variation inferred by GM.
• DL features enabled classification of leaves into morpho-phylogenomic groups within the R. auricomus species complex.

Conclusions:

• Simple in situ leaf imaging combined with DL reproducibly captures population-level leaf shape variation.
• Integrating DL with GM provides crucial insights into the shape information extracted by computer vision.
• This approach is essential for reliable automated plant phenotyping.