VGDS-PointNet++ for organ segmentation and phenotypic trait estimation in greenhouse tomato seedlings

Area of Science:

• Agricultural Engineering
• Computer Vision
• Plant Science

Background:

• Accurate phenotypic analysis is crucial for greenhouse tomato breeding and variety selection.
• Traditional methods for measuring plant traits can be time-consuming and labor-intensive.

Purpose of the Study:

• To develop an automated system for segmenting tomato plant point clouds and calculating key phenotypic traits.
• To enable efficient and accurate data acquisition for plant breeding using depth cameras.

Main Methods:

• A novel model combining voxel grid downsampling (VGDS) with PointNet++ was proposed for point cloud segmentation.
• A tomato point cloud dataset (TPCD) was created and augmented, with VGDS replacing the original sampling strategy.
• Leaf length and stem diameter were automatically calculated after segmentation using surface fitting techniques.

Main Results:

• The VGDS-PointNet++ model achieved high accuracy (96.80%) and mean Intersection over Union (mIoU) (88.95%) in segmentation.
• The model demonstrated a significant runtime improvement with an average VGDS of 0.132 s.
• High determination coefficients (R ² = 0.93 for leaf length, R ² = 0.87 for stem diameter) were observed between automatic and manual measurements.

Conclusions:

• The proposed VGDS-PointNet++ model offers an efficient and accurate solution for automated phenotypic trait extraction in tomatoes.
• This technology can significantly aid in the phenotypic analysis and variety selection of greenhouse tomato plants.
• Depth camera-based point cloud analysis presents a promising avenue for precision agriculture applications.