A hierarchical growth method for extracting 3D phenotypic trait of apple tree branch in edge computing

Accurately obtaining the length, quantity and distribution of fruit branches plays an important role in orchard irrigation management, disease control and improving fruits' yield and quality. Recently, edge computing has been proposed for digital orchard management as it can increase computing power for computationally intensive applications. However, due to the diversity of fruit tree morphological structures and the complexity of the planting environment, the traditional method of obtaining fruit tree phenotypes with centralized computing on cloud servers faces many challenges in terms of efficiency and accuracy. In this paper, we propose a hierarchical growing method (HG) suitable for deployment at the edge to achieve semantic segmentation and instance segmentation of fruit tree point clouds at the organ scale. Furthermore, extract fruit trees phenotypic trait at the organ scale based on the result of point clouds segmentation. Numerous experiment show that the proposed HG method can efficiently carry on instance segmentation of branches and phenotypic trait extraction by the joint analysis and processing of point cloud data. The MAE, RMSE and IOU of the primary branches reach 0.025 m, 0.113 and 0.720, respectively.

Automated summary

Accurately obtaining the length, quantity, and distribution of fruit branches is crucial for orchard management, disease control, and improving fruit yield and quality. Edge computing has been proposed to address the challenges of efficiency and accuracy that traditional centralized computing methods face due to the diversity of fruit tree morphological structures and complex planting environments. In this study, a hierarchical growing method (HG) is proposed for edge deployment to achieve semantic and instance segmentation of fruit tree point clouds and extract phenotypic traits at the organ scale. Experimental results demonstrate that the HG method efficiently performs instance segmentation and phenotypic trait extraction with high accuracy.