3D Deep Learning-based Segmentation to Reveal the Spatial Distribution of Cotton Bolls

Lizhi Jiang,Changying Li,Longsheng Fu

DOI: https://doi.org/10.13031/aim.202200361

2022-01-01

Abstract:Highly accurate three-dimensional point clouds provide an opportunity to map spatial distribution of cotton bolls, which assists breeders in understanding the relationship between boll positions on each branch and yield. In this study, a high-resolution terrestrial LiDAR was used to obtain the point clouds of cotton plants, and a data processing pipeline was developed to describe precisely the position of every boll on every branch. The data processing pipeline consisted of four main steps: First, TreeQSM was used to segment main stalks and individual branches of cotton plants. Second, the 3D deep learning model PointNet++ was used to segment the cotton bolls from the branches. Then, the instance segmentation of cotton bolls was realized by Euclidean clustering, and the spatial location of every boll was determined by K-means. Finally, the distance of each boll to the main stalk was calculated to map the boll relative to the main stalk. PointNet++ was also used to segment bolls directly from plants, then boll point clouds were sliced into 1cm and the number of points in each slice were counted to map the distribution relative to plant height. Additionally, boll counts were achieved for the whole plant and for individual branches. The results demonstrate that the accuracy and mean intersection over union (mIoU) of the 2-class segmentation results based on PointNet++ on the whole plant dataset reached 0.958 and 0.902, respectively. Meanwhile, each cotton boll can be accurately mapped to the corresponding branch. This method provides a valuable tool for 3D plant mapping to advance plant breeding programs.