ELGCot3D: a lightweight 3D cotton point cloud segmentation model based on EdgeConv-Local Attention-GCN and semantic feature enhancement

Efficient and non-destructive cotton organ extraction is crucial for automatic cotton phenotyping. However, limited by leaf occlusion, large model parameters, and inefficient manual observation, it fails to meet current high-throughput phenotyping demands. To address these challenges, this paper propose ELGCot3D, a lightweight 3D point cloud-based cotton organ segmentation method, enabling high-precision segmentation in resource-constrained environments. First, a new module called ELG3D replaces traditional Set Abstraction structures, enhancing local cotton data learning capability via multi-mechanism feature fusion and boosting segmentation accuracy. Second, a cotton-specific feature enhancement module is proposed to secondary optimize the features output from the Feature Propagation layer. This module significantly increases feature discriminability while substantially reducing redundant and high consumption network layers, achieving a balance between performance and efficiency. Finally, a cotton point cloud-adapted training strategy improves model training stability and prediction accuracy. Experimental results on the Crops3D dataset show ELGCot3D achieves 76.7% mIoU and 86.1% OA for cotton segmentation, Meanwhile, the number of parameters and computational complexity are reduced by 50.1% and 50.7%, respectively, demonstrating the model's lightweight characteristics. Notably, it performs well in segmenting other Crops3D crops and exhibits strong generalization on the other cotton point cloud datasets. The proposed method offers a reliable approach for cotton phenotyping and precision agriculture. Future work will extend its high-throughput extraction capability for individual plant organs in large cotton fields, providing breeders with accurate data to support efficient breeding and new variety development.