基于堆栈稀疏自编码器的小麦赤霉病高光谱遥感检测  

Hyperspectral remote sensing detection of Fusarium head blight in wheat based on the stacked sparse auto-encoder algorithm

作　　者：林芬芳 陈星宇[1] 周维勋 王倩 张东彦[2] LIN Fen-Fang;CHEN Xing-Yu;ZHOU Wei-Xun;WANG Qian;ZHANG Dong-Yan(School of Remote Sensing and Geomatics Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,Jiangsu,China;National Engineering Research Center for Agro-Ecological Big Data Analysis and Application(Anhui University),Hefei 230601,Anhui,China;Key Laboratory of Geospatial Technology for the Middle and Lower Yellow River Regions(Henan University),Kaifeng 475004,Henan,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,江苏南京210044 [2]农业生态大数据分析与应用技术国家地方联合工程研究中心(安徽大学),安徽合肥230601 [3]黄河中下游数字地理技术教育部重点实验室(河南大学),河南开封475004

出　　处：《作物学报》2023年第8期2275-2287,共13页Acta Agronomica Sinica

基　　金：国家自然科学基金项目(42271364);江苏省科技计划项目(BK20211287);黄河中下游数字地理技术教育部重点实验室(河南大学)开放课题项目(GTYR202104)资助。

摘　　要：小麦赤霉病具有发病快、周期短的特点,利用深度学习特征提取方法建立病害严重度检测模型,可为小麦赤霉病的防治提供科学指导。研究于2018—2020年间采集3个品种小麦在扬花期、灌浆期和成熟期的麦穗高光谱数据,通过形态学处理去除麦芒,提取出麦穗光谱曲线,使用多源散射校正对光谱进行去噪处理,再采用堆栈稀疏自编码器(Stacked Sparse Auto-encoder,SSAE)提取小麦赤霉病的光谱特征,利用该特征分别结合Softmax分类器和偏最小二乘回归方法构建小麦赤霉病严重度判别和预测模型。通过预训练,具有12~6个神经元的双层SSAE模型表现较好,模型均方误差更低,而且各个病害等级的特征差异明显;以训练的SSAE模型提取的深度学习特征为基础分别建立赤霉病严重度等级判别模型和严重度预测模型,在严重度等级判别的分类结果中,模型的总体精度和Kappa系数分别为88.2%和0.84,其中“淮麦35”品种的总体精度最高;在严重度预测模型中,模型对所有品种测试集的预测决定系数和均方根误差分别为0.927和0.062,对各品种的预测决定系数均在0.95左右;相比常见的几种小麦赤霉病光谱指数,基于SSAE深度学习特征的赤霉病预测模型精度更高。高光谱遥感数据量大、光谱波段多,堆栈稀疏自编码器通过在自编码器模型中加入稀疏表示的限定条件,并增加隐含层数及隐含神经元数来构建更为复杂的模型,所提取的光谱特征更能全方面地体现小麦赤霉病的光谱特征,利用该特征构建的小麦赤霉病检测模型具有更高的精度,可为精准监测小麦赤霉病提供科学依据。Fusarium head blight(FHB)has the characteristics of rapid onset and short cycle.The deep learning feature extraction method was used to establish a disease severity detection model to provide guidance for the prevention and control of FHB.The hyperspectral data of wheat ears from flowering to maturity under three varieties from 2018 to 2020 were collected.The spectral curves of wheat ears were obtained by morphological processing and multi-source scattering correction.Then spectral features of FHB were extracted by stacked sparse auto-encoder(SSAE),combined with Softmax classifier and the partial least squares regression method to detect FHB.Through pre-training,the two-layer SSAE model with 12–6 neurons performed better,the mean square error of the model was lower,and the characteristics of each disease level were significantly different.The deep learning features extracted by the trained SSAE model were the basis of the establishment of FHB disease severity level discrimination model and severity prediction model.The overall accuracy and Kappa coefficient of the model were 88.2%and 0.84,respectively,and the accuracy was the highest for the variety of‘Huaimai 35’.The prediction coefficient of determination(R2)and root mean square error(RMSE)of the model for the test set of all varieties were 0.927 and 0.062 in the severity prediction model,respectively,and R2 for each variety was around 0.95.The FHB prediction model based on SSAE deep learning features has higher accuracy than those with several common FHB spectral indices.Hyperspectral remote sensing had the characteristics of large amount of data and many spectral bands.The stack sparse auto-encoder builded a more complex model by adding the limiting conditions of sparse representation to the auto-encoder model,and increasing the number of hidden layers and hidden neurons.The extracted spectral features can better reflect the spectral characteristics of FHB in all aspects,so the detection model of FHB constructed by using these features has higher accura

关 键 词：赤霉病 堆栈稀疏自编码器 高光谱 检测 小麦 

分 类 号：S435.121.45[农业科学—农业昆虫与害虫防治] S127[农业科学—植物保护]