基于水稻冠层多信息融合的监测褐飞虱种群大小的BP神经网络方法  

BP neural network method for monitoring the population size of Nilaparvata lugens(Hemiptera:Delphacidae)based on multi-source data collected from rice canopy

作　　者：熊志强 王嘉汉 刘向东[1] XIONG Zhi-Qiang;WANG Jia-Han;LIU Xiang-Dong(Department of Entomology,College of Plant Protection,Nanjing Agricultural University,Nanjing 210095,China)

机构地区：[1]南京农业大学植物保护学院昆虫学系,南京210095

出　　处：《昆虫学报》2024年第4期572-581,共10页Acta Entomologica Sinica

基　　金：国家重点研发计划项目(2021YFD1401102);江苏省现代农业-重点及面上项目(BE2022387)。

摘　　要：【目的】褐飞虱Nilaparvata lugens种群监测的自动化和智能化尚未实现。本研究旨在探究水稻受褐飞虱危害后冠层光谱和温度以及叶片叶绿素相对含量与为害虫量的关系,建立基于高光谱、热成像和叶绿素等多信息融合的误差反向传播(back propagation,BP)神经网络监测褐飞虱的方法,为田间褐飞虱种群监测向自动化与智能化方向发展提供方法支持。【方法】在可控条件下利用方形塑料框培育水稻,并在分蘖期接入雌雄1∶1配对的不同对数(0,1,2,3,4,5,6和7对)的褐飞虱雌、雄成虫,然后连续多次(接虫后16,27,32,44和60 d时)调查接虫区水稻上褐飞虱虫量(每4穴稻上个体数),并采用高光谱仪和热成像仪分别测定水稻冠层光谱反射率和冠层温度,利用土壤和植物分析仪器开发(soil and plant analyzer development,SPAD)叶绿素仪测定叶片叶绿素的相对含量(SPAD值);采用Pearson相关法分析各测量指标与褐飞虱虫量的相关性;采用多元散射校正对光谱反射率数据进行降噪处理;采用连续投影算法对高光谱反射率数据进行降维和敏感波段筛选;分别以光谱反射率单一信息及其与冠层温度和SPAD值融合后的多源信息为输入量,采用普通和加入粒子群算法优化的BP神经网络建模,构建褐飞虱为害不同时段后种群大小的神经网络监测模型。【结果】褐飞虱为害后水稻冠层光谱在近红外的730-930 nm波段反射率、水稻冠层温度与气温的差值(冠气温差)和叶片的SPAD值均与褐飞虱虫量呈显著负相关。利用连续投影算法筛选出的冠层光谱敏感波段处反射率并降噪后建立的BP神经网络监测5个危害时段褐飞虱虫量的预测集决定系数R^(2)在0.504~0.892之间;融合冠层光谱、冠气温差和叶片SPAD值等多信息建立的BP神经网络监测褐飞虱虫量的预测集R^(2)提升到0.640~0.975;在多源信息基础上再选用粒子群优化(particle swarm optimization,P【Aim】The automation and intelligence of population monitoring of the brown planthopper,Nilaparvata lugens,have not been resolved now.The aim of this study is to explore the relationship of canopy spectrum and temperature,and leaf chlorophyll content with the number of N.lugens on rice plants,and to establish a back propagation(BP)neural network to monitor the population size of N.lugens based on multi-source information fusion of hyperspectral,thermal imaging,and chlorophyll,so as to provide a new method for the development of automation and intelligence in monitoring N.lugens populations.【Methods】Under controlled conditions,rice was cultivated using the square plastic box,and different pairs(0,1,2,3,4,5,6 and 7 pairs)of female and male adults of N.lugens(female to male ratio=1∶1)were released onto rice plants at the tillering stage.Then,the number of N.lugens on rice(number of individuals per 4 hills of rice)was investigated on day 16,27,32,44 and 60 post original infestation,and the spectral reflectance and temperature of rice canopy were measured using a hyperspectral spectrometer and a thermal imager,respectively.The relative content of chlorophyll in leaves was measured using a chlorophyll meter(soil and plant analyzer development,SPAD)as SPAD readings.The Pearson correlation method was used to analyze the correlations between these measured indexes and the number of N.lugens.The multivariate scattering correction was used to process the spectral reflectance data to reduce noise.The successive projection algorithm was adopted for dimensionality reduction and screening the sensitive band of hyperspectral reflectance.Using single source of spectral reflectance information and its multi-source information fusion with canopy temperature and SPAD readings as inputs,the modeling methods,the general(BPNN)and optimized BP neural networks by particle swarm optimization(PSO-BPNN)were used to establish the neural network models to monitor the population sizes of N.lugens damaging different periods.【Results�

关 键 词：褐飞虱 冠层光谱反射率 冠层温度 叶绿素SPAD BP神经网络 粒子群优化 

分 类 号：Q968[生物学—昆虫学]