Cucumber Seedling Segmentation Network Based on a Multiview Geometric Graph Encoder from 3D Point Clouds.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Biology

Background:

• Plant phenotyping is crucial for understanding plant growth and development.
• Accurate plant organ segmentation from 3D point clouds is essential but challenging using traditional geometric features.
• Existing methods struggle with precise segmentation and measurement of plant structures.

Purpose of the Study:

• To develop an advanced segmentation network for accurate plant organ identification from 3D point clouds.
• To enhance plant phenotyping by improving the precision of geometric feature extraction.
• To propose a novel network, SN-MGGE, leveraging multiview geometric graphs for improved segmentation.

Main Methods:

• A point cloud acquisition platform was used to create a cucumber seedling dataset.
• The proposed SN-MGGE network utilizes a geometric graph encoder (GGE) to extract features in Euclidean and hyperbolic spaces.
• Semantic segmentation is achieved through downsampling and multilayer perceptron, followed by K-means clustering for parameter extraction.

Main Results:

• The SN-MGGE network achieved high segmentation accuracy, with mIoU of 94.90% and OA of 97.43%, outperforming existing methods.
• Extracted phenotypic parameters (plant height, leaf length, width, area) showed high correlation with ground truth (R² values up to 0.98).
• Ablation studies and generalization experiments confirmed the robustness and broad applicability of the SN-MGGE network.

Conclusions:

• The SN-MGGE network offers a significant advancement in plant organ segmentation from 3D point clouds.
• This method provides a robust and accurate approach for plant phenotyping and trait measurement.
• The findings have implications for accelerating crop research and breeding programs through automated analysis.