RootNet: A deep learning framework for automated tomato radicle segmentation and length measurement

Abstract

Automated phenotyping of crops is a vital component of precision agriculture. As a representative economic crop, the radicle length of tomato seeds is a key phenotypic indicator for assessing seed vigor and seedling health. To reduce labor costs and improve measurement efficiency, we developed an integrated seed germination phenotype acquisition system that combines cultivation and imaging, enabling continuous image acquisition throughout the tomato seed germination process. To achieve automatic measurement of radicle length, we propose a deep learning-based segmentation framework, termed RootNet. The framework incorporates a custom-designed module, MambaNextBlock (MNB), within skip connections to enhance long-range feature modelling, and integrates an Atrous Spatial Pyramid Pooling (ASPP) module at the bottleneck to capture multi-scale contextual information. Post-segmentation, Canny edge detection is employed to extract radicle contours, from which actual lengths are computed based on contour arc length. Experimental results show that RootNet achieved 77.75% Intersection over Union (IOU), 86.03% Precision, 88.72% Recall, and 87.46% F1-Score on the root class. Compared with manual measurements conducted using ImageJ, our method showed high agreement across 1170 radicle measurements, with an R² of 0.9785, an MAE of 0.274 mm, an RMSE of 0.344 mm, and a Bias of −0.008 mm. Bland–Altman analysis further confirmed the absence of systematic bias, with 95% limits of agreement ranging from −0.682 mm to +0.666 mm. Meanwhile, measurement efficiency was improved by approximately 680-fold. Furthermore, the method was applied to evaluate the effects of drought, salinity stress, and different concentrations of Streptomyces albidoflavus (HL4) and Streptomyces virginiae (GZ2) on radicle growth. The results indicated that drought stress, salinity stress, and undiluted HL4 inhibited radicle elongation, whereas diluted HL4, as well as both undiluted and diluted GZ2, significantly promoted radicle growth. This study provides an efficient and cost-effective solution for non-destructive crop phenotyping and intelligent agricultural management in precision farming.

Plant phenotyping relevance

トマト幼根長を画像から自動抽出・測定する深層学習フレームワークを開発し、手動測定との定量的検証も行っており、植物フェノタイピング手法が研究の中心である。

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.