HagPF: Hierarchical-annotation-guided phenotypic framework for stem instance segmentation and length measurement in plant point clouds

Abstract

As a critical structural component that connects almost all other types of plant organs, the stem system not only supports the weight of the total plant, but also serves as a vital channel for nutriment transportation. Accurate phenotypic measurement of stem instances is of practical significance for assessing crop growth dynamics and predicting yield. To address current 3D phenotyping challenges of crops such as the difficulty in separating stem segments from the stem system and the low accuracy in stem length measurement, we propose a Hierarchical-annotation-guided Phenotypic Framework (HagPF) for Stem Instance Segmentation and Length Measurement in plant point clouds. Specifically, a hierarchical leaf-stem organ instance annotation strategy is devised to effectively train a PSegNet network for leaf and stem instance segmentation. The segmentation is then followed by a shape-adaptive measurement algorithm to automatically measure the length of stem segments that are morphologically diverse in space. On a 3D dataset comprising four crop species, the proposed framework achieved an Intersection over Union (IoU) of 95.45% for organ semantic segmentation and a Mean Weighted Coverage (mWCov) of 87.87% for instance segmentation (both stem and leaf). Regarding to the stem length measurement, the method obtained an average Root Mean Square Error (RMSE) of 1.044 cm and a relative error of 11.907%, outperforming 7 mainstream methods. The relevant dataset and source code can be found at: https://github.com/Jinx00/stem-length-measurement .

Plant phenotyping relevance

植物点群から茎のインスタンスを分割し、茎長を自動測定する3D表現型解析フレームワークの開発・比較検証が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.