A method for phenotyping lettuce volume and structure from 3D images

Area of Science:

• Agricultural Engineering
• Plant Science
• Computer Vision

Background:

• Accurate plant growth monitoring is vital for crop management.
• RGBD cameras offer a non-invasive method for modeling plants like lettuce.
• Current deep learning models for fresh weight estimation lack domain adaptation and dataset availability.

Purpose of the Study:

• To develop a novel method for estimating lettuce rosette structure and volume using plant geometric features.
• To compare the proposed method with existing surface reconstruction techniques.
• To improve the accuracy of fresh weight estimation in lettuce.

Main Methods:

• A new method was developed to create a tight hull around plant point clouds, preserving rosette detail and filling surface holes.
• The proposed method was compared against Ball Pivoting and Alpha Shapes for surface reconstruction.
• Linear regression was used to estimate fresh weight based on plant volume and geometric features.
• A dataset of 402 lettuce plant point clouds was created using multi-view 3D cameras.

Main Results:

• The proposed method effectively models lettuce rosette structure and volume.
• Fresh weight estimation achieved a root mean square error (RMSE) of 18.2 g using plant volume alone.
• Including geometric features improved RMSE to 17.3 g.
• New geometric features characterizing leaf density were introduced.

Conclusions:

• Plant geometric features provide a robust method for lettuce growth modeling and fresh weight estimation.
• The developed approach overcomes limitations of existing deep learning models.
• The introduced geometric features have potential applications in plant breeding.