基于3D Gaussian Splatting的小麦植株三维表型构建分析OA

To address inefficiencies and limited automation in traditional methods for acquiring mature wheat three-dimensional pheno-typing,which struggle to balance efficiency and precision,a comprehensive phenotypic workflow based on 3D Gaussian Splatting(3DGS)was established.This approach integrated three technical modules:high-fidelity 3D reconstruction from multi-view images using 3DGS,extraction and accuracy validation of phenotypic traits represented by plant height,and organ segmentation(leaf,stem,spike)via point cloud analysis using PointNet++model.Experimental results showed that 3DGS could efficiently reconstruct detailed three-dimensional wheat plant models,achieving peak signal-to-noise ratios,structural similarity indices,and learned perceptual image block similarity of 36.959 4 dB,0.974 6,and 0.114 6,respectively.Plant height measurements showed high consistency with manual data(determination coefficient R2=0.971 3,root mean square error was 1.565 cm).PointNet++model achieved best organ segmenta-tion accuracy and average intersection-over-union ratios of 0.780 69 and 0.639 54 under optimized parameters(10 000 sampling center poi-nts).On test set,ear segmentation accuracy was the highest,with precision rate of 0.860 4 and intersection over union of 0.754 7.Three-dimensional models of wheat generated using this method exhibited high-quality reconstruction and precision,confirming its ad-vantages in efficiency and precision for three-dimensional phenotyping analysis and demonstrating strong application potential.