【作者】 杨宝华；

【导师】 朱艳；

【作者基本信息】 南京农业大学 ， 作物信息学， 2020， 博士

【摘要】 小麦作为中国重要的粮食作物,在农业生产和战略性粮食储备中占有重要地位。氮素不仅是小麦生育期内重要的营养元素,而且是决定小麦品质、产量和生产力的重要基础。定量监测氮素已成为当前农业遥感领域的一个重要研究方向,是智慧农业发展中作物生长监测、精确农作管理及精准施肥实施的关键。近年来,基于不同遥感平台的作物氮素营养监测取得较为成功的应用,其特征提取已逐渐成为农作物氮素营养无损监测与诊断的关键环节,极大地扩展了作物冠层的特征表达能力。然而,传统的特征提取通常将遥感数据作为同质数据考虑,往往忽略了不同遥感数据的类型、结构及维度上差异性,这不仅影响了估测模型的普适性和鲁棒性,更无法准确获取不同遥感数据的全面图谱特征。因此,明确不同遥感数据的图谱特征变化规律及其对氮素估计模型的影响,对优化作物栽培、减施增效、促进农业可持续发展具有重要意义。为此,本研究开展了多栽培因子互作的多年小麦田间试验,运用地物光谱仪、成像光谱仪和携带消费级数码相机的无人机,获取冠层反射光谱、高光谱影像、无人机RGB图像及小麦叶层氮含量。针对不同遥感监测数据的特点,利用不同的方法系统分析不同生育时期小麦冠层的传统图谱特征变化规律,探明了不同图谱特征对小麦氮素营养的敏感响应参数,明确基于不同遥感数据的图谱特征与小麦叶层氮含量的定量关系,进而构建不同遥感平台图谱特征融合的小麦叶层氮含量估算模型,为小麦生长状况监测提供理论基础和技术支撑。首先,利用不同施氮水平、不同种植密度、不同品种的小麦田间试验资料,系统分析了不同遥感监测数据提取的传统图谱特征在小麦不同生育时期的变化规律,探明了消除数据冗余和自相关后特征的响应表现,及其在不同生育时期与小麦叶层氮含量的定量关系。1)对于高光谱数据,基于反射率构建的植被指数与小麦叶层氮含量在不同生育时期均呈现强相关,且植被指数之间表现为共线性,进而通过随机森林算法优选植被指数,结果表明 DCNI#、NDRE、NDVI Ⅱ RVI Ⅳ、VOG1、VOG2、VOG3、MSR具有较高的相对重要性。与此同时,利用光谱统去除方法处理反射光谱,从而提取光谱反射位置(500-673nm、745-980nm、980-1200nm、1200-1359nm、1453-1799nm及 2000-2400nm)、光谱吸收位置(555-745nm、883-1078nm、1078-1274nm)的深度、面积、归一化深度等特征参数。利用随机森林算法优选了具有相对重要性较高的光谱位置和形状特征,包括:A_Depth1、A_Area1、A_ND1、A_Depth2、R_Depth1、R_Areal、R_ND1、R_Depth3、Dr、SDr、Rg、Ro,其中 A_ND1、R_ND1、Rg、Ro 与小麦叶层氮含量表现为负相关,其他特征表现为正相关。2)对于高光谱影像数据,在不同生育时期的植被指数与小麦叶层氮含量都表现为强相关。其中,NDVIg-b#、VOG2、VOG3、NPCI、SIPI、PSRI与小麦叶层氮含量负相关,其他植被指数都表现为正相关。通过随机森林算法优选,结果表明 NDVIg-b#、SIPI、NPCI、VOG3、VOG2、RVI Ⅰ、SAVI Ⅱ、MTVI2具有较高的相对重要性。并且利用光谱统去除方法获取小麦冠层高光谱的两个吸收位置(557-754nm,900-1030nm)和两个反射位置(500-675nm,754-960nm),进一步提取深度、面积、归一化深度等光谱吸收位置和形状特征。通过分析特征值的分布情况,发现在小麦不同生育时期具有明显差异,进而利用全生育时期的特征进行相关分析,从而提取与小麦叶层氮含量相关性较高的特征包括Rg、R_Depth1、R_Aear 1、R_ND1、A_Depth1、A_Aear1、A_ND1。3)对于小麦冠层无人机RGB图像,基于R、G和B通道构建的可见光植被指数与小麦叶层氮含量的相关性分析,确定VARI、MGRVI、GRVI、ExR、CIVE为优选的可见光植被指数。另外,经过两层离散小波分解获取垂直、水平和对角线方向的6个高频子图,分别从高频子图中提取能量(E)、熵(En)、均值(Mean)和标准差(S)特征。通过统计分析发现,小波纹理特征间存在严重共线性,且与小麦叶层氮含量相关性较弱。因此,利用主成分分析获得8个主成分小波纹理特征。其次,利用深度学习框架下的卷积神经网络模型提取深层特征。将深度学习引入到小麦氮素营养监测中,构建了基于PyTorch框架的卷积神经网络模型,其结构包括五个卷积层、三个池化层和两个全连接层。通过迁移学习、参数微调、调整输入层数据格式以适应不同遥感数据分析,实现端到端的自动提取基于不同遥感监测平台的小麦冠层反射光谱、高光谱影像及无人机RGB图像的抽象而复杂的深层特征,从而提升了特征语义表达能力。1)对于高光谱数据,去除噪声光谱后保留1811维的光谱数据,通过连续投影算法提取64维敏感波长,按照64×64的光谱矩阵图作为卷积神经网络的输入,通过卷积层、池化层和全连接层后提取256维深层特征;2)对于小麦冠层的高光谱影像和无人机RGB图像,将图像尺寸切割成大小227pixel×227pixel×3后作为卷积神经网络的输入,分别提取256维深层特征;并通过特征可视化展示了不同卷积层对小麦冠层高光谱影像和无人机RGB图像的深层特征。结果表明自动提取的不同的深层特征具有明显的鲁棒性,体现卷积神经网络在特征提取局部感受野(local field)和权重共享的优势。研究结果为提高特征的鲁棒性提供技术支撑。进一步,融合了不同维度的图谱特征。针对不同遥感平台获取遥感数据特点,利用光谱分析、连续统去除、小波变换等方法从光谱维、空间维提取反射光谱(植被指数、光谱位置和形状特征)、高光谱影像(植被指数、光谱位置和形状特征)和无人机RGB图像(可见光植被指数、小波纹理特征)的传统图谱特征。同时,通过卷积神经网络模型将低层属性转换成更健壮的抽象特征,分别提取了高光谱的深层特征(光谱维)、高光谱影像和无人机RGB图像的深层特征(空间维),通过分析传统图谱特征和深层特征与小麦叶层氮含量的相关关系,明确消除数据冗余和特征自相关后的特征响应规律,探明特征在不同生育时期与小麦叶层氮含量的影响,提升了对小麦叶层氮含量的估测精度。另外,构建了基于不同遥感监测平台的小麦叶层氮含量估测模型。探讨了基于粒子群优化支持向量回归算法,对SVR模型的惩罚系数C与核函数参数g等两个重要参数进行优化,解决了惩罚因子、核函数、灵敏度系数等参数难以选择的问题。通过建立PLSR、SVR和PSO-SVR模型,并基于独立年份的数据进行验证,进而比较所有模型的性能,详细分析预测模型的均方根误差、决定系数及剩余预测偏差,从而确定了基于融合特征(传统图谱特征和深层特征)的PSO-SVR模型的小麦叶层氮含量估测效果最优。1)对于小麦冠层高光谱(反射光谱),基于融合特征的PSO-SVR模型精度最高,校正集R~2达到0.923,比PLS、SVR模型分别提高了 8.06%、2.8%。验证集R~2达到0.855,比PLS、SVR模型分别提高了 9.25%、3.31%。2)对于小麦冠层高光谱影像,基于融合特征的PSO-SVR模型精度最高,校正集R~2达到0.9251,比PLS、SVR模型分别提高了 5.04%、1.12%。验证集R~2达到0.8663,比PLS、SVR模型分别提高了 4.85%、0.48%。3)对于小麦冠层无人机RGB图像,基于融合特征的PSO-SVR模型精度最高,校正集R~2达到0.9172,比PLS、SVR模型分别提高了 9.71%、2.92%。验证集R~2达到0.8562,比PLS、SVR模型分别提高了 11.19%、6.16%。更多还原

【Abstract】 As an important crop in China,wheat occupies an important position in agricultural production and strategic food reserves.Nitrogen is an important nutrient element during the growth period of wheat,and also an important basis for determining wheat quality,yield and productivity.Quantitative monitoring of nitrogen concentration has become an important research direction in the field of agricultural remote sensing,and it is the key to the implementation of crop growth monitoring,precision farming management and precision fertilization in the development of smart agriculture.In recent years,crop nitrogen nutrition monitoring based on different remote sensing platforms has been successfully applied,and feature extraction has gradually become a key technology in non-destructive monitoring and diagnosis of crop nitrogen nutrition,which greatly expands the feature expression ability of crop canopy.However,traditional feature extraction usually considers remote sensing data as homogeneous data,and the differences in the types,structures and dimensions of different remote sensing data are often ignored,which not only affects the universality and robustness of the estimation model,but also makes it impossible to accurately obtain the comprehensive features of image and spectrum of different remote sensing data.Therefore,it is necessary to clarify the change regular of the features of image and spectrum of different remote sensing data and the influence on the nitrogen estimation model.It is of great significance to optimize crop cultivation,reduce fertilization and increase efficiency,and promote sustainable agricultural development.Therefore,multi-year wheat field trials with interaction of multiple cultivation factors were carried out in this study.Canopy reflectance spectra,hyperspectral images,RGB images of unmanned aerial vehicle(UAV),and nitrogen concentration of leaf layer in wheat were obtained by using ground feature spectrometers,imaging spectrometers and UAV carrying consumer-grade digital cameras.For the features of different remote sensing monitoring data,the traditional features of image and spectrum of wheat canopy at different growth periods were systematically analyzed,and the sensitive response parameters of different features of image and spectrum to wheat nitrogen nutrition were explored,and the features of image and spectrum and nitrogen concentration of leaf layer in wheat based on different remote sensing data were clarified.Furthermore,the models for estimating the nitrogen concentration of leaf layer in wheat based on the fusion of features of image and spectrum of different remote sensing platforms were constructed,which provided theoretical basis and technical support for monitoring of wheat growth status.First of all,the experimental data of different nitrogen fertilizer levels,different planting densities,and different varieties of wheat fields were used,and the changes regular of traditional features of image and spectrum extracted from different remote sensing monitoring data in different growth periods of wheat were systematically analyzed.The response perfornance of the features after eliminating data redundancy and autocorrelation was ascertained,and its quantitative relationship with nitrogen concentration of leaf layer in wheat at different growth stages.1)For hyperspectral data,the vegetation index constructed based on reflectance and the nitrogen concentration of leaf layer in wheat showed a strong correlation at different growth periods,and the vegetation index showed collinearity.Then the vegetation index was selected based on the random forest algorithm.The results showed that DCNI#,NDRE,NDVI Ⅱ,RVI Ⅳ,VOG1,VOG2,VOG3,MSR have high relative importance.At the same time,the reflection spectrum is processed by the spectral system removal method to extract depth,area,normalized depth parameters of the spectral reflection position(500-673nm,745-980nm,980-1200nm,1200-1359nm,1453-1799nm and 20002400nm)and the spectral absorption position(555-745nm,883-1078nm,1078-1274nm).The random forest algorithm is used to optimize the spectral position and shape features with relatively high importance,including:A_Depthl,A_Areal,A_ND1,A_Depth2,R_Depth1,R_Areal,R_ND1,R_Depth3,Dr,SDr,Rg,Ro.Among them,A_ND1,R_ND1 Rg,Ro and nitrogen concentration of leaf layer in wheat showed negative correlation,and other features showed positive correlation.2)For the hyperspectral image data,the vegetation index and nitrogen concentration of leaf layer in wheat at different growth stages show a strong correlation.Among them,NDVIg-b#,VOG2,VOG3,NPCI,SIPI,PSRI were negatively correlated with nitrogen concentration of leaf layer in wheat,and other vegetation indices were positively correlated.Feature selection was performed based on random forest algorithm,and the results show that NDVIg-b#,SIPI,NPCI,VOG3,VOG2,RVI Ⅰ,SAVI Ⅱ,MTVI2 have high relative importance.And the two absorption positions(557-754nm,900-1030nm)and two reflection positions(500-675nm,754-960nm)of the wheat canopy hyperspectrum were obtained using the spectral system removal method,and then the depth,area,and normalization were extracted.By analyzing the distribution of eigenvalues,it is found that there were obvious differences in different growth periods of wheat,and then the features of the whole growth period were used for correlation analysis,so that the features that have a higher correlation with the nitrogen concentration of the wheat leaf layer were extracted,including Rg,R_Depth1,R_Aear1,R_ND1,A_Depth1,A_Aear1,A_ND 1.3)For the RGB image of the wheat canopy,the correlation analysis between the visible light vegetation index constructed based on the R,G and B channels and the nitrogen concentration of leaf layer in wheat.In addition,six high-frequency sub-images in vertical,horizontal and diagonal directions were obtained through two-layer discrete wavelet decomposition,and energy(E),entropy(En),mean(Mean)and standard deviation(S)were extracted from the high-frequency sub-images.After statistical analysis,it was found that there was serious collinearity among wavelet texture features,and the correlation with the nitrogen concentration of wheat leaf layer was weak.Therefore,eight principal component wavelet texture features were obtained using principal component analysis.Then,a convolutional neural network model under the deep learning framework was used to extract deep features.Deep learning was introduced into the monitoring of wheat nitrogen nutrition,and a convolutional neural network model based on the PyTorch framework was constructed which includes five convolutional layers,three pooling layers and two fully connected layers.Through migration learning,fine-tuning of parameters,and adjustment of the input layer data format to adapt to different remote sensing data analysis.The abstract and complex deep features of the image were extracted based on the end-toend automatic extraction of wheat canopy reflectance spectra,hyperspectral images and UAV RGB images from different remote sensing monitoring platforms,which enhanced the ability of feature semantic expression.1)For hyperspectral data,after removing the noise spectrum,the 1811-dimensional spectral data was retained,and the 64-dimensional sensitive wavelength was extracted based on the successive projections algorithm(SPA),and the 64×64 spectral matrix was used as the input of the convolutional neural network,the 256-dimensional deep features were extracted through convolutional layer,pooling layer and fully connected layer;2)For the hyperspectral image and the RGB image of the wheat canopy,the image size was adjusted to 227 pixel×227 pixel×3 as the input of the convolutional neural network and extracted separately 256-dimensional deep features;the deep features of different convolutional layers on wheat canopy hyperspectral images and UAV RGB images were displayed with feature visualization.The results show that the different deep features automatically extracted have obvious robustness,which reflects the advantages of the convolutional neural network in feature extraction local field and weight sharing.The research results provide technical support for improving the robustness of features.At the same time,the features of image and spectrum of different dimensions were integrated.According to the features of remote sensing data acquired by different remote sensing platforms,the traditional features of image and spectrum based on successive projections algorithm and wavelet transform were extracted from reflectance spectra(vegetation index,spectral position and shape features),hyperspectral images(vegetation index,spectral position and shape features),and UAV RGB images(visible light vegetation index,wavelet texture features).At the same time,the low-level attributes were converted into more robust and abstract features through the convolutional neural network model,and the deep features of the hyperspectral(spectral dimension),the deep features of the hyperspectral image and the UAV RGB image(spatial dimension)were extracted respectively.The correlation between traditional features of image and spectrum and deep features and nitrogen concentration of leaf layer in wheat was analyzed,which help to clarify the feature response regular after eliminating data redundancy and feature autocorrelation.The results of the study improved the estimation accuracy of the nitrogen concentration of leaf layer in wheat.In addition,an estimation model of nitrogen concentration of leaf layer in wheat based on different remote sensing monitoring platforms was constructed.The support vector regression algorithm based on particle swarm optimization optimized two important parameters such as the penalty coefficient C and kernel function parameters g of the S VR model,and solved the problem that the parameters such as penalty factor,kernel function,and sensitivity coefficient were difficult to choose.Through the establishment of PLSR,SVR and PSO-SVR models,and verification based on independent year data,the performance of all models was compared with the root mean square error(RMSE),coefficient of determination(R~2)and residual predictive deviation(RPD)of the prediction model were analyzed in detail.Therefore,the PSOSVR model estimated nitrogen concentration of leaf layer in wheat based on fusion features(traditional features of image and spectrum and deep features)was determined to be the most effective.1)For the wheat canopy hyperspectrum(reflectance spectrum),the PSO-SVR model based on the fusion features has the highest accuracy,and the calibration set R~2 reaches 0.923,which was 8.06%and 2.8%higher than the PLS and SVR models respectively.The validation set R~2 reaches 0.855,which was 9.25%and 3.31%higher than the PLS and SVR models respectively.2)For wheat canopy hyperspectral images,the PSOSVR model based on the fusion features has the highest accuracy,and the calibration set R~2 reaches 0.9251,which was 5.04%and 1.12%higher than that of the PLS and SVR models respectively.The validation set R~2 reaches 0.8663,which was 4.85%and 0.48%higher than tthat of he PLS and SVR models respectively.3)For the wheat canopy UAV RGB image,the PSO-SVR model based on fusion features has the highest accuracy,and the calibration set R~2 reaches 0.9172,which was 9.71%and 2.92%higher than the PLS and SVR models respectively.The validation set R~2 reaches 0.8562,which was 11.19%and 6.16%higher than that of the PLS and SVR models respectively.更多还原