Related Experiment Video
Updated: Jun 23, 2026
05:03
Direct Observation and Automated Measurement of Stomatal Responses to Pseudomonas syringae pv. tomato DC3000 in Arabidopsis thaliana
Published on: February 9, 2024
Automated stomatal traits measurement in melon (Cucumis melo L.) based on vision transformers with dynamically
Yao Huang1, Lelong Yan1, Jiaxi Yang1
1State Key Laboratory of Crop Genetics and Germplasm Enhancement and Utilization, College of Horticulture, Nanjing Agricultural University, Nanjing, 210095, China.
Plant Methods
|June 20, 2026
Summary
This study introduces an advanced deep learning model for analyzing stomatal traits in dicotyledonous crops like melon. The improved Mask R-CNN framework with Vision Transformer enhances stomatal segmentation and trait quantification, aiding crop optimization.
Area of Science:
- Plant Science
- Computer Vision
- Agricultural Technology
Background:
- Stomatal trait analysis is crucial for crop photosynthesis and transpiration.
- Deep learning studies have predominantly focused on monocotyledons, neglecting dicotyledonous crops like melon.
- Accurate stomatal trait quantification is vital for breeding and optimizing crop yields.
Purpose of the Study:
- To develop and validate a deep learning model for stomatal instance segmentation and trait quantification in melon (Cucumis melo L.).
- To improve upon existing Mask R-CNN frameworks by integrating Vision Transformer (ViT) with novel attention and feature extraction modules.
- To establish a comprehensive stomatal dataset for melon to facilitate research in understudied dicotyledonous crops.
Main Methods:
- Creation of a dedicated melon stomatal dataset with 5,708 training, 1,631 validation, and 815 test images.
- Development of an improved Mask R-CNN framework utilizing a Vision Transformer (ViT) backbone.
- Integration of Dynamically Composable Multi-Head Attention (DCMHA) and a modified effective Squeeze-and-Excitation (eSE) module within the Feature Pyramid Network (FPN).
- Automated quantification of stomatal traits using ellipse fitting, validated against manual measurements.
Main Results:
- The proposed model achieved a mean average precision (mAP) of 72.40% on the melon dataset.
- The effective Squeeze-and-Excitation (eSE) module significantly improved detection metrics, while Dynamically Composable Multi-Head Attention (DCMHA) offered moderate gains.
- Automated trait quantification demonstrated strong agreement with manual measurements (Pearson r = 0.978).
- Preliminary transferability to other cucurbits (cucumber, watermelon, pumpkin, loofah) showed an average R² of 0.86.
Conclusions:
- The developed deep learning model effectively performs stomatal instance segmentation and trait quantification in melon.
- The novel architectural improvements enhance model performance and provide a valuable tool for plant science research.
- The model's adaptability suggests potential for broader application in analyzing stomatal traits across various dicotyledonous species.
