[Paper Review] StoManager1: Automated, High-throughput Tool to Measure Leaf Stomata Using Convolutional Neural Networks

Jiaxin Wang, Heidi J. Renninger+1 more|arXiv (Cornell University)|Nov 17, 2023

Plant Water Relations and Carbon Dynamics4 citations

TL;DR

StoManager1 is a high-throughput, automated tool using convolutional neural networks and empirical/theoretical algorithms to detect, count, and measure over 30 stomatal and guard cell metrics with mAP@0.5 > 0.993. It enables precise, scalable analysis of stomatal traits linked to productivity and intrinsic water use efficiency in hardwoods, enhancing plant physiological modeling and climate adaptation research.

ABSTRACT

The characteristics of stomata on leaves are crucial for the performance of plants and their impact on global water and carbon cycling. However, manually counting stomata can be time-consuming, prone to bias, and limited to small scales and sample sizes. We have created StoManager1, a high-throughput tool that automates detecting, counting, and measuring stomata to address this issue. StoManager1 uses convolutional neural networks to estimate parameters such as stomatal density, area, orientation, and variance. Our results show that StoManager1 is highly precise and has an excellent recall for the stomatal characterizing leaves from various species. This tool can automate measuring leaf stomata, making it easier to explore how leaf stomata control and regulate plant growth and adaptation to environmental stress and climate change. An online demonstration of StoManager1 is available on GitHub at https://github.com/JiaxinWang123/StoManager.git. We have also developed a standalone, user-friendly Windows application for StoManager1 that does not require any programming or coding experience.

Motivation & Objective

• Address the lack of automated, high-throughput tools for comprehensive stomatal metric measurement in plant physiology and ecology.
• Overcome limitations of manual and semi-automated methods that are time-consuming, biased, and limited in scale.
• Develop a robust, accurate, and scalable solution for measuring diverse stomatal and guard cell morphological traits across species.
• Enable large-scale integration of stomatal metrics into global carbon and water cycle models and ecological functional trait studies.
• Facilitate genetic and physiological research on stomatal adaptation to environmental stress and climate change.

Proposed method

• Employ a dual deep learning approach combining bounding box detection and instance segmentation via convolutional neural networks (CNNs).
• Integrate empirical and theoretical algorithms to compute complex stomatal metrics such as area, length, width, orientation, and aggregation index.
• Use the Shoelace formula for precise area calculation of irregular stomatal and guard cell shapes from segmented outputs.
• Train and validate the model on 17 hardwood species with diverse stomatal morphologies to ensure generalizability.
• Implement a multi-stage pipeline: image preprocessing, stomatal detection, instance segmentation, metric extraction, and validation against ground-truth data.
• Provide both a command-line Python package and a standalone Windows application for broad accessibility and usability.

Experimental results

Research questions

• RQ1 Can a deep learning-based system achieve high-precision, high-recall detection and measurement of over 30 stomatal and guard cell metrics across diverse hardwood species?
• RQ2 How do newly derived stomatal metrics—such as the ratio of stomatal area to guard cell area—relate to leaf physiological performance and tree growth?
• RQ3 To what extent do stomatal traits, when combined with leaf functional traits, explain variance in productivity and intrinsic water use efficiency (iWUE)?
• RQ4 Can automated, high-throughput stomatal phenotyping improve the accuracy and scalability of ecological and physiological modeling?
• RQ5 How does StoManager1’s performance compare to manual and semi-automated methods in terms of precision, recall, and time efficiency?

Key findings

• StoManager1 achieved exceptional precision and recall with a mean average precision at 0.5 IOU (mAP@0.5) exceeding 0.993 across diverse hardwood species.
• The integration of stomatal and leaf functional traits explained 78% of productivity variance and 93% of intrinsic water use efficiency (iWUE) variance in hardwoods.
• Newly derived metrics such as the ratio of stomatal area to guard cell area showed significant correlations with leaf physiology and tree growth.
• The tool successfully measured over 30 stomatal and guard cell metrics, including area, length, width, orientation, evenness, divergence, and aggregation index.
• StoManager1 demonstrated high sensitivity and robustness in capturing morphological diversity across species, enabling reliable large-scale phenotyping.
• The source code and a user-friendly Windows application are publicly available on GitHub and Zenodo, supporting widespread adoption and reproducibility.