Plant-GeoAT: Geometry-Aware Organ Identity Parsing of 3D Plant Point Clouds for Organ-Level Phenotyping

Abstract

Three-dimensional plant point clouds retain crop architecture, but organ-level phenotyping requires reliable semantic separation of leaves and stems. Plant-GeoAT is a geometry-aware parsing network that encodes local relative-XYZ neighbourhoods before RGB fusion and couples spatial neighbourhoods with feature–space relations for dense point prediction. We evaluated the model separately within the native protocol of a self-built structure-from-motion rapeseed dataset, an image-based soybean dataset, and laser-scanned Pheno4D maize and tomato datasets; these are within-dataset train/test experiments, not cross-dataset transfer or domain-generalisation tests. Across five seeds, mIoU was 92.26 ± 0.28%, 82.50 ± 0.24%, 99.74 ± 0.05%, and 94.75 ± 0.15%, respectively. Maize Stem IoU reached 99.57 ± 0.09%. Adding LLGE increased the four-dataset average mIoU from 66.38% to 85.84%, and the complete LLGE + SSCA model reached 92.31%. On the fixed six-sample test set, exploratory semantic-guided clustering achieved 86.67 ± 7.45% Count Accuracy; structural correctness ranged from 3/6 to 4/6 samples across seeds. Plant-height consistency was assessed independently on all 60 reconstructed rapeseed samples. The results indicate that geometry-to-context encoding produces organ-level semantic units for subsequent phenotyping, while independent instance-labelled datasets and additional growth stages are still needed for trait-level validation.

Code and data availability

The paper's Plant-GeoAT research-preview code (model implementation, training and evaluation scripts) is publicly available at an immutable GitHub commit under the authors' kimevans111 account. The self-built rapeseed point-cloud dataset and annotations are not public (available only on request), and third-party public