AUTHOR=Zhang Lina , Li Xinying , Yang Zhiyin , Yang Bo , Yu Shengpeng , Zhao Shuai , Huang Ziyi , Zhang Xingrui , Yang Han , Lin Yixing , Yu Helong , Yang Minglai 

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

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1571445

DOI=10.3389/fpls.2025.1571445

ISSN=1664-462X

ABSTRACT=IntroductionThe phenotypic traits of tomato plants reflect their growth status, and investigating these characteristics can improve tomato production. Traditional deep learning models face challenges such as excessive parameters, high complexity, and susceptibility to overfitting in point cloud segmentation tasks. To address these limitations, this paper proposes a lightweight improved model based on the ResNet architecture.MethodsThe proposed network optimizes the traditional residual block by integrating bottleneck modules and downsampling techniques. Additionally, by combining curvature features and geometric characteristics, we custom-designed specialized convolutional layers to enhance segmentation accuracy for tomato stem and leaf point clouds. The model further employs adaptive average pooling to improve generalization and robustness.ResultsExperimental validation demonstrated that the optimized model achieved a training accuracy of 95.11%, a 3.26% improvement over the traditional ResNet18 model. Testing time was reduced to 4.02 seconds (25% faster than ResNet18’s 5.37 seconds). Phenotypic parameter extraction yielded high correlation with manual measurements, with coefficients of determination (R²) of 0.941 (plant height), 0.752 (stem diameter), 0.945 (leaf area), and 0.943 (leaf inclination angle). The root mean square errors (RMSE) were 0.506, 0.129, 0.980, and 3.619, respectively, while absolute percentage errors (APE) remained below 6% (1.965%–5.526%).DiscussionThe 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.