EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

LBD-PointNet++: a point cloud segmentation network for phenotypic trait extraction of broccoli seedlings

Haoqi Wang, Baoqiang Yin, Siying Liu, Yuanyu Xia, Ziyi Zhang, Hongbin Wu, Xiuqing Fu

Frontiers in Plant Science · 31.07.2026 · 17

DOI: 10.3389/fpls.2026.1835285 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Frontiers in Plant Science, online erschienen am 31.07.2026. Verfasst von Haoqi Wang und 6 weiteren Autor:innen. Beteiligte Einrichtung: College of Engineering, Nanjing Agricultural University, Nanjing, China. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Broccoli is a globally significant vegetable, but climate change and soil salinization increasingly threaten its productivity. Precise seedling phenotyping is essential for selecting salt-tolerant germplasm, yet traditional manual methods are labor-intensive and error-prone. This study develops LBD-PointNet++, an optimized 3D point cloud semantic segmentation model for automated phenotypic parameter extraction of broccoli seedlings at the germination and early developmental phase under salt stress. High-fidelity 3D point clouds were reconstructed from a precision three-view imaging system using Structure from Motion (SfM) algorithms. LBD-PointNet++ introduces three core optimizations: (1) a Large Kernel Attention (LKA) mechanism using 3D sparse decomposition to capture long-range global dependencies; (2) a Dual Uncertainty and Shape-Adaptive Sampling (DUSAS) mechanism to preserve high-frequency features of fragile stems and margins; and (3) a joint Boundary-Aware Nested Contrastive and Adaptive Varifocal Joint Loss (BNCV-Loss) to effectively isolate overlapping leaves. Experimental results demonstrate superior performance, achieving an overall mean Intersection over Union (mIoU) of 88.07% across all three categories (Leaf, Stem, and Pot) and a Mean F1-score of 93.48%. Compared to state-of-the-art Transformer architectures like PTv3, LBD-PointNet++ achieves higher accuracy with less than 6% of the parameter volume and over twofold faster inference speed. Furthermore, dynamic monitoring across NaCl gradients (0–250 mmol/L) revealed a potential non-linear threshold effect, identifying 100 mmol/L as a preliminary phenotypic threshold under these conditions. Beyond this threshold, growth inhibition intensified rapidly; At 250 mmol/L, plant height decreased by 54.43% and the 3D entity volume shrank to approximately one-fifth of the control group. In summary, LBD-PointNet++ provides a high-efficiency solution for phenotypic identification and digital breeding of salt-tolerant Brassicaceae crops.

Schlagwörter

Stichwörter der Autor:innen

3D point cloudLBD-PointNet++broccoli seedlingslarge kernel attentionphenotypingsalt stresssemantic segmentation

Autor:innen und Institutionen

  • College of Engineering, Nanjing Agricultural University, Nanjing, China.

Zitieren

APA

Wang, H., Yin, B., Liu, S., Xia, Y., Zhang, Z., Wu, H., & Fu, X. (2026). LBD-PointNet++: a point cloud segmentation network for phenotypic trait extraction of broccoli seedlings. Frontiers in Plant Science, 17. https://doi.org/10.3389/fpls.2026.1835285

Vancouver

Wang H, Yin B, Liu S, Xia Y, Zhang Z, Wu H, et al. LBD-PointNet++: a point cloud segmentation network for phenotypic trait extraction of broccoli seedlings. Frontiers in Plant Science. 2026;17. doi:10.3389/fpls.2026.1835285

BibTeX
@article{Wang2026,
  author = {Wang, Haoqi and Yin, Baoqiang and Liu, Siying and Xia, Yuanyu and Zhang, Ziyi and Wu, Hongbin and Fu, Xiuqing},
  title = {LBD-PointNet++: a point cloud segmentation network for phenotypic trait extraction of broccoli seedlings},
  journal = {Frontiers in Plant Science},
  year = {2026},
  volume = {17},
  doi = {10.3389/fpls.2026.1835285}
}
Kostenlos starten →