The traditional plant part segmentation methods rely on empirical selection of threshold parameters,while the current shallow deep learning framework may lead to the loss of important geometric features of the plant cloud,and it is difficult to effectively integrate the local and global features of the plant.Therefore,a plant part segmentation network was proposed on 3D Point Cloud(LGF-SegNet),which was more suitable for expressing geometric features in plant point-cloud data by introducing double-weighted attention mechanism module and location coding.A feature aggregation module was introduced into the decoding layer of the proposed framework to fuse the local feature and global feature of the plant point cloud,so that the framework could focus on the overall feature outline of the plant while preserving the detailed plant textures(such as stems and leaves).The experimental results show that the average of intersection ratio, precision and Fl score of semantic segmentation reach 85.76%,93.18%and 91.08%,respectively.The mean precision,mean coverage and mean weighted coverage of instance segmentation reach 85.27%,78.46%and 79.63%, the proposed architecture is better than the current deep learning network architecture used in the current plant point cloud segmentation task,and is suitable for the dual tasks of plant semantic segmentation and instance segmentation. This lay a foundation for the subsequent research on plant growth prediction.