Panicle-3D: Efficient Phenotyping Tool for Precise Semantic Segmentation of Rice Panicle Point Cloud

Area of Science:

• Agricultural Engineering
• Computer Vision
• Plant Science

Background:

• Automated measurement of crop phenotypic parameters is crucial for quantitative growth studies.
• Current crop point cloud segmentation faces challenges like limited data, occlusion, and lack of specialized models.
• Traditional clustering methods lack the required accuracy for plant organ segmentation.

Purpose of the Study:

• To develop an automated method for segmenting and classifying crop point clouds for phenotypic parameter measurement.
• To address the limitations of existing methods in segmenting complex crop structures like rice panicles.
• To design a targeted network model for improved point cloud segmentation accuracy.

Main Methods:

• Construction of a desktop-level point cloud scanning apparatus using structured-light projection.
• Acquisition and creation of a dedicated rice ear point cloud dataset.
• Implementation of data augmentation techniques to enhance sample utilization and training accuracy.
• Design and application of a novel 3D convolutional neural network, Panicle-3D, incorporating PointConv and skip connections.

Main Results:

• The Panicle-3D model achieved a segmentation accuracy of 93.4%, outperforming PointNet.
• The network design effectively handles multiscale features and reduces feature loss during downsampling.
• The developed scanning apparatus facilitates efficient point cloud acquisition.
• Data augmentation improved sample utilization and training accuracy.

Conclusions:

• Panicle-3D offers a significant advancement in automated crop point cloud segmentation, particularly for rice panicles.
• The model's architecture is effective in addressing challenges like feature scale variation and downsampling.
• Panicle-3D demonstrates suitability for segmenting other similar crop point cloud data, paving the way for broader applications.