辣椒根系图像的原位监测与图像处理  

In-situ monitoring and image processing of pepper root image

作　　者：王凤杰 鲁伟 汪小旵[1] WANG Fengjie;LU Wei;WANG Xiaochan(College of Engineering,Nanjing Agricultural University,Nanjing 210031,China;College of Agricultural Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China)

机构地区：[1]南京农业大学工学院,江苏南京210031 [2]江苏大学农业工程学院,江苏镇江212013

出　　处：《江西农业大学学报》2025年第1期214-222,共9页Acta Agriculturae Universitatis Jiangxiensis

基　　金：国家自然科学基金青年基金项目(32301697);山东省博士后创新项目(SDCX-ZG-202400103)。

摘　　要：【目的】根系的主要作用是吸收水分和吸收养分,对根系的研究是植物营养学、育种学和基因组学中十分重要的内容。根系形态的准确获取是农业生产中基因改良、提高水肥利用效率、改善作物品质和提高产量的先决条件。根系形态的原位监测与根系参数实时获取是作物营养供给、基因调控以及作物生理过程等研究的重要手段。由于根系形态获取困难、根系图像复杂以及背景对比度低等原因,根系形态原位监测与根系图像的分割较为困难。【方法】研究使用自制的根系形态采集系统获取根系原位图像,并通过在U2Net网络中构造感兴趣区域建立了一种高效、准确的复杂根毛图像分割模型。首先利用根系图像阈值分割结果作为区域基,在区域基的周围添加部分像素,来定位根系所在的区域,从而构造感兴趣区域;随后引入U2Net网络,增加迁移学习部分进行预训练,提供模型权重初始参数;并在网络中加入根系形态的先验知识对图像进行分割,并提取根系参数。【结果】该模型在不同种类根轴和根毛分割方面效果较好(F=0.922)。使用聚乙烯线制作了仿真根系与该方法对比,本试验的根系形态采集系统及其图像处理算法能较为准确的获得采集系统视野范围内的根系参数,对根轴的提取最大误差不超过6%,对根毛的提取最大误差不超过10%。对于不同的根系均能以较高的精确度获得根长、平均直径等参数,为进一步使用该系统进行原位监测提供了较为可靠的技术手段和方法。使用该方法对根系生长形态随时间的变化进行了监测,发现根系在第22天时,主根数量较少,根毛也较少;第33天时,根系发育旺盛,根轴较为丰满,有丰富的根毛;到第55天时,根轴仍然很丰满,但根毛却减少。【结论】该方法能够分割和提取根毛信息,并能够对同一个部位的形态进行长期动态监测。本研究方[Objective]The main function of root system is to absorb water and nutrients.The study of roots is very important in plant nutrition,breeding and genomics.The accurate acquisition of root morphology is a prerequisite for gene improvement,improving water and fertilizer use efficiency,improving crop quality and increasing yield in agricultural production.Root system morphology in situ monitoring and real-time acquisition of root parameters are important means for crop nutrient supply,gene regulation,and crop physiological processes research.Due to the difficulties of root system morphology acquisition,complex root system images,and low background contrast,in-situ monitoring of root morphology and segmentation of root images are difficult.[Method]In this study,a root morphological acquisition system was used to obtain root in-situ images,and an efficient and accurate image segmentation model was established by constructing regions of interest in the U2Net network.Firstly,the segmentation result of root image threshold was used as the region base,and some pixels were added around the region base to include the region where root system was located,in order to construct the region of interest.Then the U2Net network is introduced,and the transfer learning part is added for pre-training to provide the initial parameters of the model weight.The prior knowledge of root morphology is added to the network to segment the image and extract the root parameters.[Result]Results showed that the model is effective in the segmentation of different kinds of root axes and root hairs(F=0.922).The simulation root system was made by using polyethylene wire.Compared with the simulation system,the root morphology acquisition system and its image processing algorithm in this study can accurately obtain the root parameters in the field of view of the acquisition system.The maximum error of root axis extraction is less than 6%,and the maximum error of root hair extraction is less than 10%.For different roots,the parameters such as root length

关 键 词：辣椒根系 原位监测 图像分割 感兴趣区域 迁移学习 

分 类 号：S641.3[农业科学—蔬菜学] TP391.4[农业科学—园艺学]