Mapping of cotton bolls and branches with high-granularity through point cloud segmentation

Area of Science:

• Agricultural Engineering
• Plant Science
• Computer Vision

Background:

• Accurate mapping of cotton boll spatial distribution is crucial for understanding yield and fiber quality correlations.
• Previous methods lacked the granularity to analyze boll positioning on plant structures.

Purpose of the Study:

• To develop and validate a 3D point cloud segmentation workflow for high-granularity spatial mapping of cotton bolls and branches across diverse genotypes.
• To assess the efficacy of deep learning (PointNet++) and clustering algorithms for instance segmentation of cotton bolls.

Main Methods:

• Utilized high-resolution 3D point clouds of 18 cotton genotypes.
• Developed a two-approach workflow for vertical and horizontal boll distribution mapping using PointNet++ and Euclidean clustering.
• Employed TreeQSM for plant structure segmentation and Dijkstra's algorithm for branch classification.

Main Results:

• Achieved high accuracy (0.954) and mean intersection over union (mIoU) (0.896) for 2-class segmentation using PointNet++.
• Demonstrated excellent boll counting performance with R²=0.99 and RMSE=5.4.
• Successfully mapped the high-granularity spatial distribution of cotton bolls and branches for the first time.

Conclusions:

• The developed 3D point cloud segmentation workflow provides a powerful tool for precise cotton plant phenotyping.
• This methodology has the potential to accelerate cotton breeding programs and plant physiological research.
• Direct prediction of fiber quality from 3D point clouds remains an area for future investigation.