ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Cotton architecture trait extraction via 3D reconstruction and instance segmentation

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601278.(doi:10.13031/aim.202601278)
Authors:   Chenjiao Tan, Lizhi Jiang, Edward Scott MacGregor, Matthew Bauer, Peng Chee, Changying Li
Keywords:   3D reconstruction, instance segmentation, architecture traits, 3DGS, phenotyping.

Abstract. Accurate characterization of cotton plant architecture is essential for yield estimation, disease analysis, and genotype evaluation. Traditional manual methods are labor-intensive and inefficient for large-scale phenotyping. While two-dimensional imaging approaches have been widely explored, they are inherently limited by occlusion and lack of depth information. To address these challenges, this study proposes a pipeline for cotton architectural trait extraction by integrating multi-view imaging, 3D Gaussian Splatting, and point cloud instance segmentation. Multi-view images of 50 cotton plants from diverse genotypes were acquired using a low-cost smartphone-based system and reconstructed into high-fidelity 3D models using 3DGS. The reconstructed models were converted into point clouds, and a transformer-based Mask3D model was trained to perform instance-level segmentation of key components, including main stems, branches, and bolls. Based on the segmentation results, structural traits such as total boll number, branch number, and boll distribution per branch were automatically extracted. Experimental results demonstrate that the proposed method achieves high reconstruction quality (PSNR = 32.88, SSIM = 0.95, LPIPS = 0.07) and strong segmentation performance (mAP = 0.933). The extracted traits show high agreement with ground truth measurements, with R² = 0.92 for boll counting and R² = 0.80 for branch counting. These results indicate that the proposed approach provides an accurate, scalable, and cost-effective solution for 3D cotton phenotyping, offering valuable insights for yield evaluation, breeding programs, and plant–environment interaction studies.

(Download PDF)    (Export to EndNotes)