Dual-Path Deep Learning and Hierarchical Spectral Clustering: A Versatile Framework for Semantic and Instance Segmentation of Plant Point Clouds

Accurate plant organ segmentation is essential for high-throughput phenotyping and ideotype selection. However, current methods struggle with plants of complex morphology, particularly small organ categories with sparse point distributions. In addition, severe leaf adhesion in dense canopies often hinders reliable leaf instance segmentation using conventional clustering methods. To address these challenges, we propose a dual-path fusion network (DPFuseNet) for semantic segmentation and a hierarchical multi-scale spectral clustering algorithm (HMSC) for instance segmentation of plant point clouds. DPFuseNet introduces three innovations: a high-frequency information embedding strategy, a dual-path feature extraction module integrating CNN and Transformer branches, and a cross-attention–based dual-granularity feature fusion block. Evaluated on tomato, cabbage, and soybean datasets, DPFuseNet achieved superior performance over state-of-the-art baselines such as Stratified Transformer and Point Transformer v3, reaching average precision, recall, F1-score, and IoU of 96.51%, 96.27%, 96.38%, and 93.32%, respectively. Compared with the current leading single-branch model Point Transformer v3, DPFuseNet improves these metrics by 1.19%, 1.20%, 1.21%, and 2.05%, and by 0.89%, 1.14%, 1.02%, and 1.70% over the dual-branch model PVDST. For instance segmentation, the proposed HMSC algorithm, combined with region growing, achieved mPrec 89.65%, mRec 78.70%, mCov 76.88%, and mWCov 85.11% on multi-stage tomato, cabbage, and soybean datasets, consistently outperforming conventional spectral clustering. Overall, the proposed framework demonstrates robustness and efficiency in both semantic and instance segmentation, offering a novel pathway for advancing plant point cloud analysis and smart agriculture.