Facility of tomato plant organ segmentation and phenotypic trait extraction via deep learning

Abstract

In facility tomato cultivation, analyzing and managing crop phenotypic traits across different growth stages is vital for improving quality and yield. Traditional manual methods are time-consuming and labor-intensive, particularly in densely cultivated environments. This study presents an automated pipeline for extracting key phenotypic parameters, including plant height, stem diameter, and fruit clusters, through three-dimensional (3D) model construction, organ segmentation, and phenotypic analysis. Point cloud data were captured at the seedling, flowering, and fruiting stages using Kinect v2 cameras and registered with the FPFH-ICP algorithm to mitigate occlusion-induced data loss. To better extract phenotypic parameters such as stem diameter and fruit clusters, constructed four multi-feature point clouds datasets, incorporating geometric coordinates, normal vectors, and color information (XYZ, XYZ-RGB, XYZ-Normal, and XYZ-Normal-RGB). RGB data highlighted regions with distinct color contrasts, such as fruits and leaves, while normal vectors captured surface details, enhancing the description of fine structures. To improve segmentation performance, we introduced CAFPoint, a modified PointNet++ model incorporating a multibranch structure and a CrossAttentionFusion (CAF) module. This structure can extract and fuse XYZ, normal and color information in a targeted manner. When applied to XYZ-Normal-RGB data, CAFPoint achieved an accuracy of 0.959 and a mean intersection over union (mIoU) of 0.863, demonstrating its capacity to enhance boundary delineation and feature representation. Phenotypic parameters derived from the segmentation, including stem diameter, fruit cluster count, and plant height, showed strong correlations with manual measurements (R² > 0.79 for stem diameter and fruit cluster count; R² > 0.92 for plant height). The proposed method enables efficient and accurate phenotypic trait extraction for tomatoes at various growth stages in greenhouse facilities, offering a valuable reference for automated trait analysis in controlled agricultural environments.

Plant phenotyping relevance

3D点群取得、器官セグメンテーション、深層学習モデルを統合し、トマトの草丈・茎径・果房数を自動抽出して手測定と検証しており、表現型取得手法が研究の中心である。

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.