HairNet2: deep learning to quantify cotton leaf hairiness, a complex genetic and environmental trait

Area of Science:

• Agricultural Science
• Plant Biology
• Computational Biology

Background:

• Cotton is a major natural fiber source, with leaf hairiness influencing key agricultural traits.
• Current methods for assessing leaf hairiness are qualitative, relying on visual scoring or the HairNet deep-learning model (GHS).
• A need exists for a more precise and quantitative method to evaluate cotton leaf hairiness.

Purpose of the Study:

• To introduce HairNet2, a novel quantitative deep-learning model for measuring cotton leaf hairiness.
• To develop a model that detects leaf hairs (trichomes) and provides a segmentation mask and a quantitative Leaf Trichome Score (LTS).

Main Methods:

• Annotation of 1250 trichome images (AnnCoT).
• Testing of six Feature Extractor and five Segmentation modules with various loss functions and data augmentation.
• Validation of HairNet2 on multiple datasets (CotLeaf-1, CotLeaf-2, CotLeaf-X).

Main Results:

• Leaf number, environment, and image position did not significantly impact results.
• While Genotype Hairiness Score (GHS) and Leaf Trichome Score (LTS) generally correlated, LTS revealed significant heterogeneity within GHS classes at the genotype level.
• LTS demonstrated a strong correlation with expert image scoring.

Conclusions:

• HairNet2 is the first quantitative and scalable deep-learning model for measuring leaf hairiness.
• The model aligns with qualitative scores at the extremes but suggests reordering for intermediate phenotypes (GHS 3-4+).
• HairNet2 assists in fine-ranking phenotypes and enables selection for specific hairiness traits linked to beneficial characteristics.