Semantic embedding-guided graph self-attention network for plant stem–leaf separation from 3D point clouds

Abstract

With the advancement of agricultural modernization, precise plant phenotyping—such as stem-leaf separation—has gained significant importance in the fields of intelligent plant breeding and phenotypic trait extraction. Although deep learning techniques offer superior solutions in the task of complex plant structure segmentation, challenges remain due to insufficient feature representation and low interclass separability. To address this issue, this paper proposes a semantic embedding-guided graph self-attention network for plant stem–leaf separation from 3D point clouds. Specifically, the proposed method is built on an encoder–decoder architecture. The multiscale features are first extracted as the receptive field progressively increases to learn the local geometric representation. Following this, the proposed method constructs a feature enhancement module that integrates graph convolution and self-attention mechanisms. By leveraging graph convolution and the self-attention mechanism, both local and global sets of information are aggregated across multiple scales, capturing intricate geometric and topological relationships to ensure highly descriptive and distinguishing feature representations. Afterwards, we employ a hierarchical decoding structure that combines upsampling and feature fusion to progressively reconstruct high-resolution point cloud feature representations. Finally, the integration of semantic-aware discriminative loss with cross-entropy loss is designed to increase intraclass compactness, interclass separability, and regularization, thereby further strengthening class distinction and segmentation quality. To validate the effectiveness and reliability of the proposed method, experiments were conducted on publicly available Plant-3D and Pheno4D datasets. The results demonstrate that the proposed method achieves superior performance from both quantitative and qualitative perspectives in terms of stem-leaf separation, demonstrating a trend towards outperforming existing methods on the tested datasets, with improvements of 3.97% in precision, 4.35% in recall, 4.3% in the F1 score, 5.23% in the IoU and 7.64% in the mIoU. Additionally, t-SNE visualization and qualitative comparisons further confirm the model’s superiority in feature clustering and structural boundary recognition. Our code is publicly available at https://github.com/Ahaoyang1/3D-SemSeg-RandlAnet.

Plant phenotyping relevance

植物の3D点群から茎葉を分離する深層学習手法を開発し、公開フェノタイピングデータセットで性能検証しているため、植物形態の取得・抽出法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.