Improved Point-Cloud Segmentation for Plant Phenotyping Through Class-Dependent Sampling of Training Data to Battle...

Area of Science:

• Computer Vision
• Plant Science
• Computational Biology

Background:

• Accurate plant phenotyping is crucial for plant breeding and scientific research.
• Manual measurement of complex traits like plant architecture is labor-intensive and often infeasible for large populations.
• 3D point cloud data offers a promising avenue for automated plant phenotyping, but segmentation challenges remain.

Purpose of the Study:

• To develop an improved 3D point cloud segmentation method for plant phenotyping.
• To address the class-imbalance problem in point cloud segmentation, where minority classes are underrepresented.
• To enhance the accuracy of segmenting plant organs from 3D point cloud data.

Main Methods:

• Utilized 3D point cloud data for plant organ segmentation.
• Implemented a class-dependent sampling strategy for training data, contrasting with a class-independent baseline.
• Investigated the impact of neighborhood size around anchor points on segmentation performance.
• Analyzed segmentation quality using mean intersection-over-union (IoU).

Main Results:

• The class-dependent sampling strategy improved overall segmentation quality, increasing mean IoU from 0.94 to 0.96.
• Significant improvements were observed for minority classes, with 'node' segmentation accuracy increasing by 46.0 percentage points.
• Optimal neighborhood size for segmentation varied across different plant organ classes.
• Higher levels of class balance did not universally guarantee better segmentation performance.

Conclusions:

• The proposed class-dependent sampling strategy effectively addresses class imbalance in 3D point cloud segmentation for plant phenotyping.
• This method enhances the accurate identification of plant organs, contributing to more efficient and precise plant trait analysis.
• Further research into optimal neighborhood selection per class is warranted for maximizing segmentation accuracy.