Tomato seedling stem and leaf segmentation method based on an improved ResNet architecture

Lina Zhang, Minghao Yin, Zhiyin Yang, Bo Yang, Shengpeng Yu, Shuai Zhao, Ziyi Huang, Xingrui Zhang, Yang Han, Yi-Xing Lin, Helong Yu, Manqi YangFrontiers in Plant Science 2025 · DOI

Abstract

The proposed X-ResNet model exhibits superior segmentation performance, demonstrating high accuracy in phenotypic trait extraction. The strong correlations and low errors between extracted and manually measured data validate the feasibility of 3D point cloud technology for tomato phenotyping. This study provides a valuable benchmark for plant phenotyping research, with significant practical and theoretical implications.

Introduction

In the burgeoning field of smart agriculture, automated phenotyping—the measurement of physical plant traits—is essential for optimizing crop yields and monitoring growth. While 3D point cloud technology offers rich spatial data, processing these complex datasets remains a significant computational challenge. Traditional deep learning models often struggle with high complexity, overfitting, and the difficulty of distinguishing fine structures like thin stems from leaf petioles. This research introduces X-ResNet, a lightweight deep learning architecture specifically designed to segment 3D point clouds of tomato seedlings with high precision and efficiency.

Methodology

The X-ResNet architecture improves upon the standard ResNet18 framework through several key innovations designed for biological data processing.

Model Architecture

To reduce computational overhead, X-ResNet implements bottleneck modules using 1x1 convolutions to compress data dimensionality before feature extraction. This modification reduces the parameter count by nearly 60%. Furthermore, the model integrates geometric features, such as surface curvature, directly into its custom convolutional layers. This allows the network to "sense" the sharp transitions where leaves meet stems, improving segmentation accuracy in complex organic shapes. The architecture follows a U-shaped encoder-decoder flow, utilizing Farthest Point Sampling (FPS) to manage point cloud density without losing structural integrity.

Data Acquisition and Annotation

The study utilized a dataset of 480 tomato plants. Data was captured using a high-precision 3DScanner-630w, which provides sub-millimeter accuracy.

Experimental setup for 3D point cloud data acquisition of tomato seedlings.

The resulting point clouds were manually labeled into "stem" and "leaf" categories using CloudCompare software to create a ground-truth baseline for training the model.

Key Results

X-ResNet outperformed several industry-standard models, including PointNet++, U-Net, and the original ResNet18, across all primary metrics.

Segmentation and Efficiency

The model achieved a segmentation accuracy of 95.11%, a 3.26% improvement over the baseline ResNet18. Despite this performance boost, the model is significantly more efficient: the parameter count dropped from 11.2M to 4.52M, and the testing time was reduced to 4.02 seconds, making it 25% faster than its predecessor.

Segmentation result of a tomato seedling point cloud, with leaves in red and the stem in blue.

Phenotypic Precision

Automated measurements derived from the X-ResNet segmentation showed exceptionally high correlations with manual "gold standard" measurements:

  • Plant Height: $R^2 = 0.941$
  • Leaf Area: $R^2 = 0.945$
  • Leaf Inclination: $R^2 = 0.943$

The Absolute Percentage Error (APE) for all extracted traits remained below 6%, proving the model's reliability for real-world agricultural research.

Discussion and Conclusion

The success of X-ResNet suggests that for specific biological tasks, specialized lightweight models can outperform heavier, general-purpose architectures by reducing noise and focusing on relevant geometric features. While the current study focused on the vegetative growth stage, future work will involve adapting X-ResNet to more complex growth stages and different crop types, such as peppers and cucumbers. Ultimately, this lightweight architecture is well-suited for deployment on low-power edge computing devices for real-time monitoring in commercial greenhouses.

Summary updated

Abstract

The proposed X-ResNet model exhibits superior segmentation performance, demonstrating high accuracy in phenotypic trait extraction. The strong correlations and low errors between extracted and manually measured data validate the feasibility of 3D point cloud technology for tomato phenotyping. This study provides a valuable benchmark for plant phenotyping research, with significant practical and theoretical implications.
BibTeX
@inproceedings{zhang2025tomato,
  author = {Lina Zhang and Minghao Yin and Zhiyin Yang and Bo Yang and Shengpeng Yu and Shuai Zhao and Ziyi Huang and Xingrui Zhang and Yang Han and Yi-Xing Lin and Helong Yu and Manqi Yang},
  title = {{Tomato seedling stem and leaf segmentation method based on an improved ResNet architecture}},
  booktitle = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1571445},
}
Figure