摘要
大斑病在全球各大玉米产区都有出现,降低了玉米的品质和产量。该病害多在病斑明显时识别,难以及时防治。本文提出一维卷积神经网络(1DCNN)高光谱模型,实现早期识别。以玉米大斑病为研究对象,手动接种大斑病后,选取吐丝期的玉米叶片进行试验,此时期刚显现病斑特征,但无法通过视觉属性观察看出是何种病害。首先采用SOC710E光谱仪采集高光谱图像,通过选取感兴趣区域获得玉米叶片的健康和大斑病两种光谱数据。使用SG卷积平滑、多元散射校正(MSC)、标准正态变换(SNV)和去趋势算法(DT)等四种光谱预处理方法,以去除光谱数据中的噪声。分别使用随机森林(RF)和K最近邻(KNN)两种监督学习算法,以准确率作为评价指标,对高光谱图像进行识别。结果表明,MSC为优选的预处理方法,两种模型预测准确率分别为88.13%和86.26%。然后采用竞争性自适应重加权算法对玉米叶片光谱数据进行特征波长提取,从原始的260个波长中优选出48个特征波长。最后建立一维卷积深度学习模型进行分类,识别准确率达到99.61%,相较于传统分类模型KNN、RF、偏最小二乘判别分析(PLS-DA)、反向传播神经网络(BP)、支持向量机(SVM),提出的模型识别准确率分别提高了5.94%、6.88%、6.48%、8.27%、12.12%。高光谱技术结合深度学习模型可以更有效识别玉米大斑病,为实现玉米病害早期识别提供了一种新的思路和方法。
Northern corn leaf blight(NCLB)occurs in major maize-producing regions globally,leading to a reduction in both maize quality and yield.Disease identification typically occurs when lesions are more obvious,making it challenging to prevent and control the disease promptly.This study proposes a one-dimensional convolutional neural network(1DCNN)model for early disease detection using hyperspectral imaging.In this research,NCLB was selected as the target disease.After manual inoculation,maize leaves at the silking stage were used for experiments,when lesions had just begun to appear,but the disease could not yet be visually identified.First,hyperspectral images were captured using the SOC710E spectrometer,and spectral data of both healthy and NCLB-infected maize leaves were obtained by selecting regions of interest.Four spectral preprocessing methods Savitzky-Golay smoothing(SG),multiplicative scatter correction(MSC),standard normal variate transformation(SNV),and detrending(DT)were applied to remove noise from the spectral data.Supervised learning algorithms,random forest(RF)and K-nearest neighbors(KNN),were employed for hyperspectral image classification,with accuracy as the evaluation metric.The results indicated that MSC was the optimal preprocessing method,achieving prediction accuracies of 88.13%and 86.26%for the RF and KNN models,respectively.Next,a competitive adaptive reweighted sampling(CARS)algorithm was applied to extract characteristic wavenumbers from the maize leaf spectral data,reducing the original 260 wavenumbers to 48 selected features.Finally,a 1DCNN deep learning model was developed for classification,achieving an accuracy of 99.61%.Compared with traditional classification models such as KNN,RF,partial least squares discriminant analysis(PLS-DA),backpropagation neural network(BP),and support vector machines(SVM),the proposed model improved recognition accuracy by 5.94%,6.88%,6.48%,8.27%,12.12%,respectively.These findings demonstrate that combining hyperspectral technology with deep learning models provides a new approach and method for early detection of maize diseases,enhancing the accuracy and timeliness of disease recognition.
作者
路阳
顾福谦
谷英楠
许思源
王鹏
LU Yang;GU Fu-qian;GU Ying-nan;XU Si-yuan;WANG Peng(College of Information and Electrical Engineering,Heilongjiang Bayi Agricultural University,Daqing 163319,China;Institute of Agricultural Remote Sensing and Information,Heilongjiang Academy of Agricultural Sciences/Postdoctoral Research Workstation of Heilongjiang Academy of Agricultural Sciences,Harbin 150086,China;Heilongjiang Provincial Key Laboratory of Networking and Intelligent Control,Northeast Petroleum University,Daqing 163318,China;Sanya Research Institute of Offshore Oil and Gas,Northeast Petroleum University,Sanya 572024,China)
出处
《光谱学与光谱分析》
北大核心
2025年第8期2302-2310,共9页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金项目(U21A2019,62476081,32401703)
中国博士后科学基金地区专项支持计划项目(2023MD744179)
黑龙江省重点研发计划项目(2023ZX01A06)
黑龙江省自然科学基金联合引导项目(LH2024F048)
黑龙江省青年科技人才托举工程项目(2022QNTJ012)
海南省科技专项(ZDYF2022SHFZ105)资助。
关键词
一维卷积神经网络
高光谱图像
玉米
大斑病
One dimensional convolutional neural network
Hyperspectral image
Maize
Northern leaf blight
作者简介
通讯作者:路阳,1976年生,黑龙江八一农垦大学信息与电气工程学院教授,e-mail:luyanga@sina.com;通讯作者:谷英楠,e-mail:guyingnan0453@163.com。
