基于双输出回归卷积神经网络的燕麦株高估测研究  

Oat Plant Height Estimation Based on a Dual Output Regression Convolutional Neural Network

作　　者：张建龙 邢文文 叶绍波 张超[1,2] 郑德聪 ZHANG JianLong;XING WenWen;YE ShaoBo;ZHANG Chao;ZHENG DeCong(College of Agricultural Engineering,Shanxi Agricultural University,Taigu 030801,Shanxi;Dryland Farm Machinery Key Technology and Equipment Key Laboratory of Shanxi Province,Taigu 030801,Shanxi)

机构地区：[1]山西农业大学农业工程学院,山西太谷030801 [2]旱作农业机械关键技术与装备山西省重点实验室,山西太谷030801

出　　处：《中国农业科学》2024年第20期3974-3985,共12页Scientia Agricultura Sinica

基　　金：中央引导地方科技发展资金项目(YDZJSX20231C009);山西农业大学博士科研启动项目(2021BQ85);山西农业大学学术恢复项目(2023XSHF2);山西省博士毕业生,博士后研究人员来晋工作奖励资金科研项目(SXBYKY2022019)。

摘　　要：【目的】株高影响燕麦的单株生产力,并与种植密度共同作用影响单位面积产量。探索大田环境下燕麦株高参数的自动、实时、精准获取方法,以期为燕麦田间自动化管理提供技术参考。【方法】首先基于Intel RealSense D435型深度相机和LabVIEW软件开发平台搭建燕麦深度图像采集系统,以‘品燕4号’燕麦为研究对象,获取生长全程26376组建模数据和2205组测试数据,每幅深度图像中燕麦所对应的平均株高和最高株高使用量尺测得。建模数据和测试数据在燕麦各株高区间内的数量相对均衡,并对图像进行高度还原、灰度化和缩放的简单预处理,随之给每张图像打2张标签,分别为图像中燕麦的平均株高和最高株高。基于8种经典卷积神经网络模型,将各网络模型的最后一层(分类层)去除,添加2个单节点且没有激活函数的全连接层后,分别构建双输出回归卷积神经网络估测模型,模型使用均方差函数(mean square error,MSE)评价各模型估测燕麦株高时的准确率。最终基于TensorFlow深度学习平台,采用建模数据经5折交叉验证选取Modified EfficientNet V2 L为估测模型。【结果】采用未参与模型训练的测试数据考察了Modified EfficientNet V2 L模型估测燕麦株高的泛化性能,该模型估测燕麦平均株高时平均绝对误差(mean absolute error,MAE)、均方根差(root mean square error,RMSE)和平均相对误差(mean relative error,MRE)分别为2.30 cm、2.90 cm和4.4%,估测最高株高时分别为2.24 cm、2.82 cm和4.1%,模型平均估测时间为52.14 ms。使用该方法估测作物株高时的精度与已有方法相近,平均估测时间可以满足作物株高获取的实时性要求。燕麦平均株高和最高株高估测时的相对误差随着作物株高的增加呈总体下降趋势,可能是由于作物株高较低时,估测结果受土壤起伏度影响较大。模型特征图可视化的结果表明,模型根据深度图像中�【Objective】Oat plant height affects the productivity per plant and the yield per unit area together with planting density.This study explores automatic,real-time,and precise methods for acquiring oat plant height in a field environment,aiming to provide technical references for the automated field management of oat.【Method】Firstly,an oat depth image acquisition system was built based on Intel RealSense D435 depth camera and LabVIEW software development platform.Taking Oat‘Pinyan No.4’as the research object,26376 modeling data and 2205 test data were obtained during the whole oat growth process.The average and highest plant height of oats in each depth image were measured with a scale.The quantity of modeling data and test data in each height range of oat plant was relatively balanced.The images were preprocessed by high restoration,grayscale and scaling.Each image was tagged with two labels,one for the average and one for the highest plant height of the oats in the image.Then,based on 8 classical convolutional neural network models,the last layer(classification layer)of each network model was removed,and two fully connected layers with single nodes and no activation function were added to construct the double output regression convolutional neural network estimation model.Mean square error(MSE)was used to evaluate the accuracy of each model in estimating oat plant height.Finally,based on the TensorFlow deep learning platform,Modified EfficientNet V2L was selected as the estimation model by 5-fold cross-validation using the modeling data.【Result】The generalization performance of Modified EfficientNet V2L model to estimate oat plant height was investigated using test data not involved in model training.The mean absolute error(MAE),root mean square error(RMSE)and mean relative error(MRE)to estimate oat average plant height were 2.30 cm,2.90 cm and 4.4%,respectively.Meanwhile,the MAE,RMSE and MRE to estimate highest plant height was 2.24 cm,2.82 cm and 4.1%,respectively.The average estimated time of t

关 键 词：立体图像 深度学习 株高 LABVIEW 燕麦 

分 类 号：S512.6[农业科学—作物学] TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]