PTV2-Fr: a point cloud segmentation network for phenotypic trait extraction and gibberellin effect analysis in sorghum seedlings

TL;DR

This paper introduces PTV2-Fr, a 3D point cloud segmentation model for automatically measuring sorghum seedling traits and analyzing the effects of gibberellin on growth.

Contribution

PTV2-Fr introduces novel mechanisms (MRDCA, PG-InvFR, EL Loss) to improve segmentation accuracy and robustness for sorghum phenotyping.

Findings

PTV2-Fr outperforms PTV2 by 2.5% in accuracy with better Recall and mean F1-score.

50–100 mg/L gibberellin promotes seedling growth, while over 200 mg/L inhibits it.

MRDCA, PG-InvFR, and EL Loss modules significantly enhance model performance.

Abstract

Sorghum is a globally important crop. Under the breeding goals of high yield and stress resistance, the precise selection of elite germplasm is crucial. Phenotypic parameters such as plant height and leaf area at the seedling stage are core indicators for evaluating growth vitality. However, traditional manual measurement is inefficient and error-prone, making it difficult to meet the needs of high-throughput research. To address this, this study proposes an improved model (PTV2-Fr) based on Point Transformer V2 (PTV2), which combines 3D point cloud technology to realize the automatic extraction of sorghum seedling phenotypic parameters and explores the regulatory effects of different gibberellin (GA3) concentrations. In this study, videos of sorghum seedlings were collected using the relevant system of Nanjing Agricultural University, and reconstructed into.ply format 3D point cloud files via the open-source software Colmap. The core optimizations of the PTV2-Fr model are as follows: Firstly, it proposes a Multi-Radius Dual-Coordinate Attention (MRDCA) mechanism to address the problems of leaf overlap and uneven point cloud density, thereby enhancing feature discrimination ability; Secondly, it introduces a Point-Graph Invariant Feature Refinement (PG-InvFR) module to improve the sensitivity of the segmentation head to local geometric details; Thirdly, it constructs a composite loss function (EL Loss) combining class-weighted cross-entropy loss and Lovász loss to alleviate class imbalance and boost segmentation accuracy. We selected 50 valid datasets from 112 video groups, annotated into three categories: Stem, Leaf, and Pot. The results show that PTV2-Fr outperforms PTV2 by 2.5% in accuracy, with significant improvements in Recall and mean F1-score (mF1). Ablation experiments confirm the positive effects of MRDCA, PG-InvFR, and EL Loss. Furthermore, PTV2-Fr demonstrates good robustness in analyzing GA concentrations, revealing that 50–100 mg/L GA concentrations promote seedling growth, while concentrations exceeding 200 mg/L inhibit growth. The PTV2-Fr model provides an efficient solution for the automatic determination of sorghum seedling phenotypes, and the revealed GA3 regulatory mechanism can offer theoretical references for high-quality seedling cultivation and hormone management.

Figures