A scalable and efficient UAV-based pipeline and deep learning framework for phenotyping sorghum panicle morphology...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Breeding

Background:

• Sorghum canopy architecture, influenced by traits like panicle morphology, is crucial for light capture and yield.
• Panicle morphology shows significant variation due to genetics, environment, and management.
• Accurate phenotyping of sorghum panicles is essential for crop improvement.

Purpose of the Study:

• To develop a scalable framework for 3D reconstruction and phenotyping of sorghum canopies and panicles.
• To create annotated datasets for training deep learning models for panicle detection.
• To introduce SegVoteNet, a novel deep learning model for 3D semantic segmentation and detection of sorghum panicles.

Main Methods:

• A Unmanned Aerial Vehicle (UAV)-based protocol using videos and Neural Radiance Fields (NeRFs) for 3D sorghum canopy reconstruction.
• Development of a 3D simulation model to generate annotated datasets for deep learning.
• Implementation of SegVoteNet, a multi-task deep learning model integrating VoteNet and PointNet++ for 3D point cloud analysis.

Main Results:

• High-quality 3D point cloud reconstructions of sorghum canopies were generated.
• SegVoteNet achieved high accuracy in sorghum panicle detection: 0.986 mAP on synthetic data and 0.850 mAP on real-world data.
• The proposed pipeline offers a robust and scalable method for field-based sorghum panicle phenotyping.

Conclusions:

• The developed framework enables efficient and accurate 3D phenotyping of sorghum panicles.
• SegVoteNet demonstrates strong performance in detecting sorghum panicles from 3D point cloud data.
• This approach provides valuable tools for sorghum breeding programs and commercial applications.