Edge-Preserving Multi-Scale Network for Plant Point Cloud Segmentation

Abstract

Accurate plant organ segmentation is essential for enabling high-throughput extraction of plant phenotypes, as it provides the foundational data for both trait measurement and structural modeling. Although existing studies have made significant progress, plant semantic segmentation (stems and leaves) across multiple species remains underexplored. To this end, this paper introduces an edge-aware downsampling algorithm and a novel network for segmenting plant point clouds at multiple scales, named MSPlantSegNet. Experimental results on a dataset of tobacco, tomato, and sorghum demonstrate that MSPlantSegNet attained superior performance across all four key metrics-precision (97.13 %), recall (95.63 %), F1-score (96.20 %), and IoU (93.14 %). MSPlantSegNet surpasses a set of leading models, including PointNet++, PointNet, ASIS, DGCNN, PlantNet, PSegNet, and PointNeXt. This research has valuable implications for plant phenotyping, the development of smart agriculture, and ideal type selection.

Plant phenotyping relevance

植物点群から茎・葉を分割する新規ネットワークとダウンサンプリング法を開発し、複数種データで性能比較しており、表現型抽出の基盤手法が中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.