【作者】 黄建荣；

【导师】 刘向东；

【作者基本信息】 南京农业大学 ， 农业昆虫与害虫防治， 2013， 博士

【摘要】 本文采用便携式光谱仪在叶片、小区和大田水平上测定了不同生育期(分蘖期、孕穗期和扬花期)水稻受稻纵卷叶螟和褐飞虱为害后的光谱反射率,建立了光谱反射率与虫害程度间的线性回归模型,以及利用径向基函数神经网络方法监测卷叶率和褐飞虱虫量的方法,获得了以下主要结果：(1)稻纵卷叶螟为害后水稻的叶片光谱反射率表现为,随着稻叶受害程度的升高,分蘖期在400-420nm及610-700nm波段内升高,在530-570nn和700-1050nm波段内降低；孕穗期在659-678nm波段内升高,在426-475nm、496-630nnm和690-1000nm波段内降低；扬花期在632-691nm波段内升高,510-616nm和697-1000nm波段内降低。利用健康叶和受害卷叶平铺组成的不同卷叶率组合叶片,其光谱反射率随卷叶率的升高,分蘖期和孕穗期均在450-500nm及610-700nm波段内升高,在530-570nm和700-1000nm波段内降低。小区水稻受稻纵卷叶螟为害后的冠层光谱反射率表现为,随着卷叶率级别的升高,孕穗期水稻在717-1000nm波段内降低,扬花期在530-600nn1和717-1000nm波段内降低。大田水稻受稻纵卷叶螟为害后的冠层光谱反射率表现为,随着卷叶率的升高,分蘖期在515-596nm和698-1000nm波段内降低；孕穗期在664-684nm波段内升高,在714-1000nm波段内降低；扬花期在623-692nm波段内升高,在725-1000nm波段内降低。(2)建立了基于叶片光谱指数的叶片受害程度的一元线性回归监测模型,分蘖期以741nn1处的反射率监测效果较好,进行100次监测的准确率为90%；孕穗期以NDVI的监测效果较好,进行86次监测的准确率为76%；扬花期以Rgreen/Bred的监测效果较好,进行80次监测的准确率为69%。建立了基于光谱指数对平铺组合叶片卷叶率监测的一元线性回归模型,分蘖期以Syellow的监测效果较好,进行14次监测的准确率为86%；孕穗期以Sred的效果较好,进行14次监测的准确率为79%。建立了基于小区水稻冠层光谱指数的卷叶率级别(Y,0-5级)的逐步回归监测模型,孕穗期以Y=14.47-0.60DVI542-500-142.98RVSI-6254.44FD540-16.76(Sred-Sblue)效果较好,模型的诊断误差RMSE为0.216；扬花期以Y=8.7356-44.7232DVI730-670+40.3448TCARI+539.1316FD715+44.3701SD695的效果较好,模型诊断误差为0.3035。对400-1000nm内大田水稻受稻纵卷叶螟为害后的光谱反射率采用逐步回归、因子分析和偏最小二乘法进行降维,并以逐步回归筛选出的各反射率、因子分析得到的两个主要因子的得分值、偏最小二乘法得到的各因子的得分值、以及筛选出的18个光谱指数作为神经网络的输入层因子,卷叶率作为输出层,进行大田卷叶率的径向基函数神经网络的学习和监测,结果表明,对分蘖期、孕穗期和扬花期水稻的31、23和27个卷叶率进行监测时,逐步回归方法所得因子为输入时最大的监测准确率分别为61%、65%和63%；因子分析所得的两个因子值为输入时最大的监测准确率分别为68%、69%和63%；偏最小二乘法得到的最少因子值为输入时最大的监测准确率分别为74%、78%和70%；18个光谱指数为输入因子时最大的监测准确率分别为81%、83%和78%。(3)水稻受褐飞虱为害后的光谱特征表明,随着虫量的增加,孕穗期水稻叶片的光谱反射率在570-690nm内上升,而在700-1000nm波段内下降,扬花期在710-1000nm波段内下降。水稻冠层光谱反射率随虫量增加,拔节-孕穗期在484-515nm和570-707nm波段内升高,在720-1000nm波段内下降；孕穗期在525-565nm和700-1000nm波段内下降；扬花期在725-1000nm波段内下降。建立了基于叶片光谱指数的褐飞虱虫量(Y)的一元线性回归监测模型,孕穗期(褐飞虱0-106.25头)以Chl Index和NPQI组建的模型监测效果较好,误差分别为17.3和17.33头；扬花期(褐飞虱0-17.5头)以DMRnir-green构建的模型效果较好,误差为3.37头。对25个叶片光谱指数与褐飞虱虫量进行逐步回归建模,孕穗期和扬花期水稻上褐飞虱虫量的监测分别以Y=174.831-79.361Chllndex-1655.217NPQI和Y=-39.903+491.616DMnir-green-791.617DMRnir-red+120.714G(Lich)模型的效果较好,监测误差分别为13.71和2.87头。建立了水稻冠层光谱指数的褐飞虱虫量(Y)监测的一元线性回归模型,拔节-孕穗期(褐飞虱虫量0-112.5)和孕穗期(褐飞虱虫量0-106.25)水稻褐飞虱虫量的监测均以指数MRnir构建的模型效果较好,模型的误差分别为19.28和18.22头；扬花期(褐飞虱0-17.5)以DVI935-670构建的模型效果较好,误差为3.81头。利用25个冠层光谱指数构建褐飞虱虫量的逐步回归模型,拔节-孕穗期、孕穗期和扬花期监测褐飞虱虫量较好的模型分别为：Y=179.85-1481.037MRnir+2341.975DVI730-542,Y=353.754-472.028MRnir-108.122WI+88.052DVI542-500,Y=158.46-1.251WI-65.499DFI935-670-133.89AI,模型的误差分别为15.06、10.6和3.29头。对400-1000nm波段内的水稻冠层光谱反射率,采用逐步回归筛选出主要的反射率、因子分析得到两个主要因子的得分、偏最小二乘方法得到各因子的得分、以及筛选的25个冠层光谱指数,作为径向基函数神经网络的输入层因子,褐飞虱虫量作为输出层,进行褐飞虱数量的神经网络学习和监测,结果表明,对拔节-孕穗期、孕穗期和扬花期水稻的6、7和7个褐飞虱虫量进行监测,逐步回归方法所得因子最大的监测准确率分别为67%、57%和57%；因子分析所得两个因子的最大监测准确率分别为50%、57%和57%；偏最小二乘法得到的最少因子监测的最大准确率分别为83%、71%和71%；25个光谱指数的监测的最大准确率分别为83%、71%和71%。(4)不同氮肥施用量水稻被褐飞虱为害后的光谱特征表现为,随着氮肥用量的升高,水稻冠层光谱反射率在790-1000nm内上升,650-690nm内下降。水稻叶片SPAD值随氮肥用量的增加显著升高。随着褐飞虱虫量的增加,低氮肥用量下的水稻在褐飞虱为害14天后,冠层光谱反射率在近红外区域显著降低,但在高氮肥用量水稻上,褐飞虱为害35天后在近红外区域的冠层光谱反射率才显著降低。水稻倒四叶的SPAD值与褐飞虱的数量相关较好,而倒二叶与倒四叶SPAD的差值与施氮水平相对不敏感。水稻冠层光谱指数和SPAD指数的结合可对不同施N水平下水稻中的褐飞虱数量进行监测。更多还原

【Abstract】 The spectral reflectance from the leaves and canopy of rice damaged by rice leaf folder (RLF) Cnaphalocrocis medinalis (Guenee) and brown planthopper (BPH) Nilaparvata lugens (Stal) were measured in plot and field at different growth stages (tillering, booting and flowering) using ASD Hand-held Spectroradiometer. The linear models for diagnosing leaf-roll rate and the number of BPH based on spectral reflectance and radial basis function (RBF) neural network method were established to supply some effective method for monitoring RLF and BPH in rice field. The main results were list as followings.(1) At tillering stage, the reflectance of rice leaf decreased at530-570nm and700-1050nm with the increase of RLF damage, whereas it increased at400-420nm and610-700nm. At booting stage, spectral reflectance of rice leaf decreased at426-475nm,496-630nm and690-1000nm and it increased at659-678nm with the increase of RLF damage. At flowering stage, the spectral reflectance of rice leaf increased at632-691nm and decreased at510-616nm and697-1000nm.The leaf-roll rates were simulated by combining different pieces of injured and healthy leaves at booting and flowering stage of rice on a black cloth in laboratory and reflectance from the combined rice leaves was measured. The results indicated that spectral reflectance decreased at530-570nm and700-1000nm and increased at450-500nm and610-700nm with the increase of leaf-roll rates.Canopy reflectance from a plot of rice decreased significantly at717-1000nm at booting, and it decreased at530-600nm and717-1000nm at flowering stage as the infestation scales of RLF increased.In rice field, canopy reflectance of rice decreased significantly at515-596nm and698-1000nm at tillering stage as the leaf-roll rates increased. At booting stage, spectral reflectance increased at664-684nm and decreased at714-1000nm. At flowering stage, spectral reflectance increased at623-692nm and decreased at725-1000nm as the leaf-roll rates increased. (2) Linear regression models based on spectral indices were established for detecting the damage degrees of rice at different grow stages. At tillering stage, The model based on the reflectance at741nm was better, and its accuracy was90%for100times of prediction. At booting stage, the model based on NDVI was better which had76%accuracy for86times of prediction. At flowering stage, the model based on Rgreen/Rred was the better which accuracy was69%for80times of prediction.Linear regression models for diagnosing the leaf-roll rates of combined leaves of rice were built. For the tillering stage leaves, the model based on Syellow was better, and its diagnostic accuracy was86%for14times prediction. For the booting stage leaves, the model based on Sred was better, and the diagnostic accuracy was79%for14times prediction.Linear Regression models based on spectral indices were built using multiple stepwise regression (MSR) method to detect infestation scales (Y,0-5) in plot of rice. At booting stage, the model Y=14.47-0.60DVI542.500-142.98RVSI-6254.44FD540-16.76(Sred-Sblue) was better and it root mean square error (RMSE) was0.216. At flowering stage, the model Y=8.7356-44.7232DVI730-670+40.3448TCARI+539.1316FD715+44.3701SD695was better with a smaller RMSE of0.3035.In rice field, diagnostic methods for leaf-roll rates were established using the RBF neural network method based on18spectral indices and reflectance at400-1000nm after reducing dimensions by MSR, factor analysis, and partial least squares (PLS) method. Factors or spectral indices entered into the RBF neural network as the input vectors, and the leaf-roll rates entered into RBF neural network as the target vectors. After31,23,27times prediction at tillering, booting and flowering stage, the predicting correction rate by the built network based on these factors from MSR was61%,65%and63%, respectively, based on the principal factors was68%,69%and63%, based on PLS factors was74%,78%and70%, and based on18spectral indices was81%,83%and78%, respectively.(3) When rice was infested by BPH, reflectance from leaves increased at570-690nm and decreased at700-1000nm at booting stage of rice, and at flowering stage, reflectance decreased significantly at710-1000nm regions. At canopy-level of rice, reflectance decreased at720-1000nm and increased at484-515nm and570-707nm at joint-booting stage of rice. At booting stage, reflectance decreased at525-565nm and700-1000nm, and at flowering stage, it decreased significantly at725-1000nm.Linear regression models for diagnosing the number of BPH infested rice were established based on the reflectance indices from rice leaves. At the booting stage (the number of BPH was0-106.25), the better models were Y=381.928-135.891CHl Index and Y=-71.254-2842.01NPQI with RMSE17.3and17.33, respectively. At the flowering stage (the number of BPH was0-17.5), the model based on DMRnir-green (Y=33.162-199.721DMRnir-green) was better with RMSE3.37. The multiple stepwise regression models based on all these25spectral indices were built, and the Y=174.831-79.361Chl Index-1655.217NPQI was better to detect the number of BPH at booting stage of rice, and its RMSE was13.71. The model Y=-39.903+491.616D MRnir-green-791.617DMRnir-red+120.714G(Lich) was better to detect the number of BPH at flowering stage with RMSE of2.87.At canopy-level, linear regression models for diagnosing the number of BPH (7) were built. The model based on MRnir was better to diagnose the number of BPH on joint-booting stage and booting stage of rice and their RMSE were18.22and19.28, respectively. At flowering stage of rice, the model based on DVI935-670was better with3.81RMSE. The stepwise regression models for diagnosing the number of BPH at joint-booting, booting and flowering stage of rice using all25indices were Y=179.85-1481.037MRnir+2341.975DVI730-542,Y=353.754-472.028MRnir-108.122WI+88.052DVI542-500, and Y=158.46-1.251WI-65.499DVI935-670-133.89AI, which had RMSE15.06,10.6and3.29, respectively.The diagnostic method for the number of BPH in rice based on RBF were studied using the MSR, factor analysis, and PLS to reduce the dimensions of reflectance at400-1000nm and25spectral indices. For6,7,7times prediction of BPH numbers at joint-booting, booting and flowering stage, the correction rates of RBF network built here based on factors from MSR were67%,57%and57%, respectively, and they were50%,57%and57%based on principal factors, and they were83%,71%and71%based on factors from PLS. The predicting correction rates of RBF network based on25spectral indices were83%,71%and71%for BPH numbers at joint-booting, booting and flowering stage, respectively.(4) Reflectance and SPAD readings were measured in rice using different levels of nitrogen-fertilizer and infested by different number of BPH. The results showed that the reflectance from rice canopy at790-1000nm and SPAD values from leaves increased significantly, and they decreased at650-690nm regions as N fertilizer rates increased. In the pots used low N fertilizer rate, Canopy reflectance of rice in the near-infrared region and SPAD value from leaves decreased significantly as the number of BPH increased when the BPH infested for14days, but in the high N fertilizer rate, canopy reflectance of near-infrared region decreased significantly when BPH infested for35days. SPAD value of the fourth fully expanded leaf (4LFT) had significant correlation with the number of BPH, and the relative ratio between SPAD reading of the second full expanded leaf (2LFT) and the fourth fully expanded leaf (4LFT) was relatively insensitive to N fertilizer rate but was sensitive to the number of BPH in rice used different N fertilizer rates. Spectral indices and SPAD index had great potential for detecting BPH damage in rice which was used different N fertilizer.更多还原