Key points are not available for this paper at this time.
Three-dimensional point cloud (3DPC) data capture detailed geometric and structural plant traits beyond the capability of 2D imaging. When combined with artificial intelligence (AI), it offers a powerful, non-invasive tool for plant phenotyping, which is crucial for driving advancements in plant breeding and agriculture. However, challenges related to data complexity, limited datasets, and model generalization hinder 3DPC’s widespread adoption. To provide a comprehensive overview and guide future research in this area, we conducted a systematic literature review (SLR) following Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines by analysing 381 papers published between January 2017 and October 2025 from major databases. Our review examines the advantages, current status, limitations, and future directions of AI applications in 3DPC-based plant phenotyping. Our findings indicate a rapid increase in publications since 2022, with deep learning (DL) methods, especially pointwise MLP-based networks, driving much of this growth, with a notable recent surge in Transformer-based, Graph-based, and particularly Hybrid models that combine their strengths. Furthermore, novel methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) are emerging as powerful tools for 3D reconstruction and scene synthesis. Time-of-Flight (ToF) and Structure from Motion and Multi-View Stereo (SfM-MVS) technologies remain the predominant 3DPC data acquisition techniques. Research in this area focuses on trees/shrubs and cereals, typically involving single-species studies. Although the overall use of public datasets remains low (18.9%), their adoption has significantly increased since 2020. Key limitations identified include: (1) a lack of standardized data collection and formats, (2) insufficient model robustness and generalization, especially from lab to field, (3) high computational demands, and (4) a reliance on species-specific models. The future of AI-driven 3DPC phenotyping hinges on overcoming these bottlenecks. Priority should be given to: developing field-deployable, computationally efficient models; exploring the potential of the foundation model; establishing diverse and standardized public datasets; and strengthening the integration of 3D phenomics with genomics to bridge the genotype-to-phenotype gap. This review provides a foundational roadmap to guide research in plant phenomics, crop breeding, and plant science.
Chang et al. (Tue,) studied this question.