【作者】 王丽爱；

【导师】 郭文善；

【作者基本信息】 扬州大学 ， 作物栽培学与耕作学， 2016， 博士

【摘要】 小麦是我国也是江苏省的主要粮食作物之一。在目前小麦栽培与管理的实际过程中,相对于采用技术人员定点调查等传统的小麦生长生产监测方法,遥感技术能及时、快捷并无破坏地监测预报大面积的小麦生长状态,但是,遥感监测预报精度还不能完全满足实际生产管理需要,为此,前人已进行了相关的研究,其中利用遥感数据构建经验统计模型定量反演能够反映小麦生长状态的一些苗情关键参数的研究得到了广泛开展,在此方向的研究显示,构建统计模型所需的算法对遥感监测精度有很重要的影响。然而,现阶段与建模算法相关的国内外文献数量不多,且其中多数文献局限于研究使用相关算法遥感监测某一个农学参数,或只是针对小麦某一个生育期展开使用相关算法遥感监测其农学参数的研究,而针对小麦主要生育期的多个苗情诊断关键参数,系统分析、评价并比较基于不同算法构建遥感监测模型的实用性和预测性研究却少有涉及。鉴于如上所述,为提高小麦生长状态的遥感监测精度,进而指导大面积小麦田间生产,最终实现高产、高效、安全、低耗、优质生产,本研究以江苏省的泰兴、姜堰、兴化和仪征地区2010（-2）013年稻茬小麦为研究对象,围绕小麦拔节、孕穗和开花期的叶面积指数(LAI)、生物量、叶片氮含量(LNC)和叶片SPAD值等苗情诊断关键参数,以各生育期相应的环境卫星HJ-CCD遥感信息为支撑,比较基于不同多元回归算法构建的遥感监测模型的可行性及预测性,旨在为提高大田尺度下的小麦苗情诊断关键参数遥感监测预报精度提供技术和方法。本研究的主要内容和结果如下：(1)分析主要生育期小麦苗情诊断关键参数与遥感变量间的相关性,以0.01水平下显著相关为依据,结果表明在拔节和开花期分别选用NRI、RVI、NDVI、GNDVI、SIPI、SAVI、 OSAVI和PSRI共8种植被指数,孕穗期选用NDVI、GNDVI、SIPI、SAVI、OSAVI和IPSRI这6种植被指数,监测小麦叶面积指数是可行的；遥感监测小麦拔节、孕穗和开花期的地上干生物量时,各期均可选用NDVI、SAVI、OSAVI、NRI、GNDVI、SIPI、PSRI、RVI、 CRI、EVI、MSR、NLI、RDVI、TVI、和MTVI2这15个植被指数作为敏感遥感变量；遥感监测小麦叶片氮含量时,拔节期选用NDVI、GNDVI、SIPI、RVI、SAVI、OSAVI、MSAVI和EVI这8种植被指数,孕穗期选用NRI、PSRI、NDVI、GNDVI、SIPI、RVI、SAVI、OSAVI、 MSAVI和EVI共10种指数,开花期选用NDVI、NRI和PSRI这3种植被指数,分别作为各生育期的敏感遥感变量；遥感监测小麦叶片SPAD值时,拔节期选用NRI、RVI、 NDVI、GNDVI、 SIPI、SAVI、OSAVI和PSRI共8种植被指数,孕穗期选用NDVI、NRI、RVI、SAVI和OSAVI这5种植被指数,开花期选用RVI、NDVI、GNDVI、SIPI、SAVI、OSAVI和PSRI共7种植被指数,分别作为各生育期的敏感遥感变量。(2)使用多种回归算法构建主要生育期小麦各苗情诊断关键参数的遥感监测模型。分别针对每个生育期(拔节、孕穗和开花期)的LAI、生物量、LNC和SPAD各长势参数,以2010（-2）012年小麦数据为训练集,将筛选的敏感遥感变量作为输入变量,以长势参数为输出变量,分别采用传统多元线性(ML, Multiple Linear)、偏最小二乘(PLS, Partial Least Squares)、人工神经网络(ANN, Artificial Neural Networks)、单核支持向量回归(SK-SVR, Single-Kernel Support Vector Regression)、双核支持向量回归(DK-SVR, Double-Kernel Support Vector Regression)和随机森林回归(RF, Random Forest)共6种回归算法建立每个生育期各长势参数的遥感反演多因子模型。利用独立于建模集的2013年小麦实测数据,结合决定系数R~2和均方根误差RMSE这两个评价指标及实测值与模型预测值之间的1：1关系图,系统性地评价并比较了所建模型的实用性及预测性,进而通过对各生育期模型性能比较确定最佳模型。(3)明确了遥感监测小麦主要生育期叶面积指数(LAI)的各模型性能。研究结果表明,监测拔节、孕穗和开花期小麦LAI的最佳模型均为DK-SVR模型,3个生育期模型估测值与实测值之间的决定系数R~2和均方根误差RMSE依次为：0.76和0.29,0.80和0.47,0.67和0.55,且各生育期的模型预测值与实测值拟合良好,同时,基于ML和ANN的模型监测这3个生育期LAI的效果最差；拔节期其余3个模型性能由高到低依次为SK-SVR模型(R~2=0.71,RMSE=0.43)、PLS模型(R~2=0.65, RMSE=0.40)和RF模型(R~2=0.49,RMSE=1.41);孕穗期其余3个模型性能由高到低依次为SK-SVR模型(R~2=0.78,RMSE=0.58)、PLS模型(R~2=0.75, RMSE=0.74)和RF模型(R~2=0.32, RMSE=1.13);开花期其余3个模型性能由高到低依次为RF模型(R~2=0.52, RMSE=0.57). PLS模型(R~2=0.45, RMSE=0.64)和SK-SVR模型(R~2=0.33, RMSE=0.84)。(4)明确了遥感监测小麦主要生育期地上干生物量的各模型性能。研究结果显示,在小麦拔节、孕穗和开花各期的6个模型中,RF模型均表现出最佳的预测能力,3个生育期RF预测值与实测值间的决定系数R~2和均方根误差RMSE依次为：0.53和477 kg.hm’2,0.72和1126 kg.hm（-2）,0.79和1808 kg.hm（-2）。拔节期次优模型为DK-SVR模型(R~2=0.50,RMSE=505.5 kg.hm（-2）)和SK-SVR模型(R~2=0.47, RMSE=509.5kg.hm"2),而基于ML. PLS和ANN的模型监测该生育期生物量的效果最差；在孕穗期和开花期,基于ANN的模型监测小麦生物量的效果最差,其余4个模型性能由高到低依次为DK-SVR模型(孕穗期R~2=0.67,RMSE=1389.2 kg.hm（-2）;开花期R~2=0.65, RMSE=2058.1 kg.hm（-2）)、SK-SVR模型(孕穗期R~2=0.51, RMSE=1422.3 kg.hm（-2）;开花期R~2=0.62, RMSE=2174.2kg.hm（-2）)、ML模型(孕穗期R~2=0.53, RMSE=1461.5 kg.hm（-2）;开花期R~2=0.49, RMSE=2454.4kg.hm（-2）)和PLS模型(孕穗期R~2=0.48, RMSE=1521.7kg.hm（-2）;开花期R~2=0.49, RMSE=2803.6kg.hm（-2）)。(5)研究表明了遥感监测小麦主要生育期叶片SPAD值的各模型性能。综合考虑模型预测值与实测值之间的一致性,并且同时考虑二者之间的决定系数R~2和均方根误差RMSE,结果显示,基于ML的模型监测小麦拔节、孕穗和开花期的叶片SPAD值时效果最差；在小麦拔节期,确定DK-SVR模型为最佳模型,预测值与实测值间的R~2和IRMSE为：0.65和1.58,其它4个模型按性能由高到低依次为：RF模型(R~2=0.55, RMSE=2.11)、SK-SVR模型(R~2=0.57,RMSE=2.31)、ANN模型(R~2=0.43, RMSE=3.06)和PLS模型(R~2=0.40,RMSE=3.42);在孕穗期,确定RF为最优模型(R~2=0.72, RMSE=2.2),其余4个性能由高到低的模型为：DK-SVR模型(R~2=0.57, RMSE=2.10)、SK-SVR(R~2=0.52, RMSE=2.30)、 PLS模型(R~2=0.47, RMSE=5.76)和ANN模型(R~2=0.43, RMSE=2.80);在开花期,确定RF为最优模型(R~2=0.60, RMSE=3.16),同时结果显示,基于PLS的模型预测效果不好,其余3个模型为：DK-SVR模型(R~2=0.52, RMSE=3.03)、SK-SVR (R~2=0.48, RMSE=3.07)和ANN模型(R~2=0.46, RMSE=3.20).(6)阐明了遥感监测小麦主要生育期叶片氮含量(LNC)的各模型性能。结果表明,基于DK-SVR算法构建的模型是分别遥感监测小麦拔节、孕穗和开花各期LNC的最佳模型,各生育期的模型预测值与实测值之间显示出良好的一致性,二者间的决定系数R~2和均方根误差RMSE分别为：拔节期的R~2=0.73, RMSE=0.13,孕穗期的R~2=0.82, RMSE=0.21,开花期的R~2=0.75, RMSE=0.20; SK-SVR模型和PLS模型均为各生育期的次优模型,3个生育期SK-SVR模型的R~2和RMSE依次为：0.61和0.16,0.77和0.29,0.72和0.21,同时,PLS模型的R~2和RMSE依次为：0.59和0.23,0.77和0.31,0.52和0.26；结果还表明,基于ML、ANN和RF的模型监测这3个生育期LNC的效果最差。更多还原

【Abstract】 Wheat is one of the main crops in China and Jiangsu province. Compared with the traditional method of monitoring the growth and production of wheat by relying on technical person to implement fixed-point investigation, the remote sensing technique is used for monitoring wheat growth at large scale because of its real-time and non-destructive monitoring in the process of wheat cultivation and management. However, its estimation accuracy should be further improved for guiding on-farm crop management. Therefore, predecessors have carried out relevant research, among which, constructing statistical models based on remote sensing data to monitor growth diagnosis parameters which can reflect the growth status of wheat at main growth stages of wheat has been widely carried out. The research about this direction has showed that algorithms used to construct statistical models are very important for improve monitoring accuracy. However, to our knowledge, only a few studies have related to the algorithms of modeling both at home and abroad recently, and among which, most of them limited on remotely monitoring one growth diagnosis parameter using related algorithms, or only aimed at one growth stage. But few studies have done to monitor multiple growth diagnosis parameters of different main growth stages of wheat based on different algorithms, and furthermore, to systemly analysis, evaluate and compare corresponding model practicability and prediction to determine the best model.In view of the above mentioned, to improve remote sensing monitoring accuracy of wheat growth to guide the production of large area wheat field for eventually realizing high yield, high efficiency, safety, low cost and high quality production of wheat, we carried out experiments in four counties (YiZheng, JiangYan, Xing Hua, and TaiXing) of Jiangsu province, China during the winter wheat growing seasons of 2010 through 2012 including the jointing stage, the booting stage and the anthesis stage respectively. Around wheat growth diagnosis parameters such as leaf area index, biomass, leaf nitrogen content and leaf SPAD at main growth stages, and supporting by synchronous China’s domestic HJ-CCD multi-spectral data, we studied the feasibility and prediction accuracy of different multiple regression models for remotely monitoring each parameter of wheat. The objective of this research is to provide a useful exploratory and predictive tool for improving prediction accuracy of monitoring growth diagnosis parameters at main growth stages of wheat by remote sensing technology. The main research contents and results of this study are as follows:(1) The study on analyzing the correlation between growth diagnosis parameters of wheat at different growth stages and remote sensing variables. The results showed that, based on the significant correlation at 0.01 level, when monitoring wheat leaf area index (LAI), the vegetation indices including NRI、RVI、NDVI, GNDVI, SIPI, SAVI, OSAVI and PSRI could be selected as sensitive remote-sensing variables in the jointing and anthesis stage respectively, and the vegetation indices including NDVI, GNDVI, SIPI, SAVI, OSAVI and PSRI were sensitive to LAI in the booting stage. The vegetation indices such as NDVI、SAVI、OSAVI、NRI、GNDVI、 SIPI、PSRI、RVI、CRI、EVI、MSR、 NLI、RDVI、TVI and MTVI2 could be used to monitoring above ground dry biomass of wheat of the jointing, booting and anthesis respectively. When monitoring nitrogen content in wheat leaves (LNC), the vegetation indices such as NDVI, GNDVI, SIPI, RVI,SAVI, OSAVI, MSAVI and EVI were sensitive remote-sensing variables in the jointing stage, and the vegetation indices including NRI、PSRI、NDVI,GNDVI, SIPI, RVI,SAVI,OSAVI, MSAVI and EVI were sensitive remote-sensing variables in the booting stage, and the vegetation indices such as NDVI, NRI and PSRI could be used. When monitoring wheat leaf SPAD, the sensitive remote-sensing variables respectively were NRI、RVI、NDVI, GNDVI, SIPI, SAVI, OSAVI and PSRI in the jointing stage, and NDVI、NRI、RVI、SAVI and OSAVI in the booting stage, and RVI、NDVI、GNDVI, SIPI，SAVI, OSAVI and PSRI in the anthesis.(2) The study on constructing remote sensing models based on using multiple regression algorithms for monitoring growth diagnosis parameters at main growth stages of wheat. For LAI, LNC, SPAD and biomass of wheat in the jointing, booting and anthesis respectively, taking the pooled data of 2010,2011 and 2012 as the training set, meanwhile, taking the sensitive remote-sensing variables as input variables and each growth parameter as output variables, the monitoring models of each growth parameter were established based on the traditional multiple linear (ML), Partial Least Squares (PLS), Artificial Neural Networks (ANN), Single-Kernel Support Vector Regression (SK-SVR), Double-Kernel Support Vector Regression (DK-SVR) and Random Forest (RF) regression algorithms respectively. Finally, taking the determination coefficients (R~2) and the root mean square error (RMSE) as metrics, meanwhile, combining with the 1:1 relationship between the observed values and predicted values of model, and taking data of each stage from 2013 as the testing set, the practicability and performance of all the models at each stage were not only evaluated on the testing set of corresponding stage but also compared to identify the best model for each growth stage.(3) Each model performance for monitoring wheat LAI of primary growth stages was made clear. The results showed that, the DK-SVR model had the best prediction accuracy in the jointing, booting and antehsis stage respectively. The coefficients of determination (R~2) of estimated-versus-measured LAI values respectively were 0.76,0.80 and 0.67, meanwhile the corresponding root mean square errors (RMSE) were 0.29,0.47 and 0.55, and the^predicted values of the model were in good agreement with the measured values. Models on the basis of ML and ANN were unable to monitor LAI of each stage. From high to low, the performance of the other 3 models in jointing stage were SK-SVR model (R~2=0.71, RMSE=0.43), PLS model (R~2=0.65, RMSE=0.40) and RF model (R~2=0.49, RMSE=1.41); in booting stage the order was SK-SVR model (R~2=0.78, RMSE=0.58), PLS model (R~2=0.75, RMSE=0.7) and RF model (R~2=0.32, RMSE=1.13); in anthesis the order was RF model (R~2=0.52, RMSE=0.57)、PLS model (R~2=0.45, RMSE=0.64) and SK-SVR model (R~2=0.33, RMSE=0.84).(4) Each model performance for monitoring wheat above ground dry biomass of primary growth stages was identified. The results pointed out, the model based on RF algorithm respectively showed the best predictive ability among the six models of each stage including jointing, booting and anthesis. The determination coefficients (R~2) and the root mean square error (RMSE) of estimated-versus-measured biomass values successively were 0.53 and 477 kg.hm（-2）,0.72 and 1126 kg.hm（-2）,0.79 and 1808 kg.hm（-2）. In the jointing, the DK-SVR model (R~2=0.50, RMSE=505.5 kg.hm（-2）) and SK-SVR model (R~2=0.47, RMSE=509.5 kg.hm（-2）) were the suboptimal models, and meanwhile, the results also showed that models based on ML, PLS and ANN were unable to monitor biomass of this stage. In the booting and anthesis, the ANN models were unable to monitoring biomass, and from high to low, the performance of the other 4 models of these two stages were DK-SVR model (R~2=0.67 and RMSE=1389.2 kg.hm"2 in the booting; R~2=0.65 and RMSE=2058.1 kg.hm（-2） in the anthesis), SK-SVR model (R~2=0.51 and RMSE=1422.3 kg.hm"2 in the booting; R~2=0.62 and RMSE=2174.2 kg.hm（-2） in the anthesis), ML model (R~2=0.53 and RMSE=1461.5 kg.hm"2 in the booting; R~2=0.49 and RMSE=2454.4 kg.hm"2 in the anthesis) and PLS model (R~2=0.48 and RMSE=1521.7 kg.hm（-2） in the booting; R~2=0.49 and RMSE=2803.6 kg.hm（-2） in the anthesis).(5) Each model performance for monitoring wheat leaf SPAD of primary growth stages was identified. Considering the determination coefficients (R~2) and the root mean square error (RMSE) between measured values and predicted values of the model, meanwhile considering the consistency between the estimated-versus-measured values, models based on ML was unable to monitor leaf SPAD of jointing, booting and anthesis stage respectively. In the jointing, the DK-SVR model was the best model, and its R~2 and RMSE were 0.65 and 1.58 respectively, and meanwhile, from high to low, the performance of the other 4 models were RF model (R~2=0.55 and RMSE=2.11), SK-SVR model (R~2=0.57 and RMSE=2.31), ANN model (R~2=0.43 and RMSE=3.06) and PLS model (R~2=0.40 and RMSE=3.42). In the booting, the best model was the RF model (R~2=0.72 and RMSE=2.2), the other 4 models successively were DK-SVR model (R~2=0.57 and RMSE=2.10), SK-SVR model (R~2=0.52 and RMSE=2.30), PLS moel (R~2=0.47 and RMSE=5.76) and ANN model (R~2=0.43 and RMSE=2.80). In the anthesis, RF model (R~2=0.60 and RMSE=3.16) showed the best performance, while PLS model was unable to monitoring leaf SPAD values, and the other 3 models successively were DK-SVR model (R~2=0.52 and RMSE=3.03), SK-SVR model (R~2=0.48 and RMSE=3.07) and ANN model (R~2=0.46 and RMSE=3.20).(6) Each model performance for monitoring wheat LNC of primary growth stages was made clear. The results indicated that, the DK-SVR model showed the best prediction accuracy in the jointing, booting and antehsis stage respectively. Good agreement was observed between the estimated values and the measured values in each stage, and the determination coefficients (R~2) and the root mean square error (RMSE) of estimated-versus-measured LNC values were 0.73 and 0.13 in the jointing,0.82 and 0.21 in the booting, and 0.75 and 0.20 in the anthesis. Both SK-SVR model and PLS model were the suboptimal model of each stage, and the R~2 and RMSE of SK-SVR model successively were 0.61 and 0.16,0.77 and 0.29,0.72 and 0.21, meanwhile, the R~2 and RMSE of PLS model successively were 0.59 and 0.23,0.77 and 0.31,0.52 and 0.26. The results also showed that models on the basis of ML, ANN and RF were unable to monitor LNC of each stage.更多还原