Few-Shot Learning Enables Population-Scale Analysis of Leaf Traits in Populus trichocarpa

Area of Science:

• Plant Science
• Computational Biology
• Genetics

Background:

• Plant phenotyping is crucial for understanding plant adaptation and genetic architecture but is often time-consuming and expensive.
• Current methods face bottlenecks in processing large-scale population data efficiently.

Purpose of the Study:

• To develop a rapid and accurate image-based plant phenotyping method using few-shot learning.
• To address the challenges of data preprocessing and minimal training sample requirements in current phenotyping techniques.
• To provide a new, large-scale dataset and open-source tools for the plant science and machine learning communities.

Main Methods:

• Leveraged few-shot learning with convolutional neural networks to segment leaf body and venation from 2,906 field-collected Populus trichocarpa images.
• Utilized raw, full-resolution RGB images without requiring experimental or image preprocessing.
• Extracted leaf morphology and vein topology traits using open-source image-processing tools and validated them with physical measurements.

Main Results:

• Successfully segmented leaf structures with minimal training data (e.g., 8 images for vein segmentation).
• Extracted validated traits for leaf morphology and vein topology.
• Conducted a genome-wide association study to identify genes controlling these traits, utilizing a dataset of 68 distinct leaf phenotypes.

Conclusions:

• The developed few-shot learning approach offers a fast, accurate, and data-efficient solution for plant phenotyping.
• The study provides valuable open-source tools and a comprehensive dataset to advance plant science and machine learning research.
• This methodology accelerates the understanding of genetic architecture underlying complex plant traits at population scale.