Development of a calibration to predict maize seed composition using single kernel near infrared spectroscopy

The relative composition of protein, oil, and starch in the maize kernel has a large genetic component. Predictions of kernel composition based on single-kernel near infrared spectroscopy would enable rapid selection of individual seed with desired traits. To determine if single-kernel near infrared spectroscopy can be used to accurately predict internal kernel composition, near infrared reflectance (NIR) and near infrared transmittance (NIT) spectra were collected from 2160 maize kernels of different genotypes grown in several environments. A validation set of an additional 480 kernels was analyzed in parallel. Constituents were determined analytically by pooling kernels of the same genotype grown in the same environment. The NIT spectra had high levels of noise and were not suitable for predicting kernel composition. Partial least squares regression was used to develop predictive models from the NIR spectra for the composition results. Calibrations developed from the absolute amount of each constituent on a per kernel basis gave good predictive power, while models based on the percent composition of constituents in the meal gave poor predictions. These data suggest that single kernel NIR spectra are reporting an absolute amount of each component in the kernel.     

Investigation of fungal development in maize kernels using NIR hyperspectral imaging and multivariate data analysis

Near infrared (NIR) hyperspectral imaging and hyperspectral image analysis were evaluated for their potential to track changes in fungal contamination on and fungal activity immediately under the surface of whole maize kernels (Zea mays L.) infected with Fusarium verticillioides. Hyperspectral images of clean and infected kernels were acquired using a SisuChema hyperspectral pushbroom imaging system with a spectral range of 1000–2498 nm at predetermined time intervals after infection. Background, bad pixels and shading of acquired absorbance images were removed using exploratory principal component analysis (PCA). When plotting PC4 against PC5, with percentage sum of squares (%SS) 0.49% and 0.34%, three distinct clusters were apparent in the score plot and this was associated with degree of infection. Loading line plots, with prominent peaks at 1900 nm and 2136 nm, confirmed that the source of variation was due to changes in starch and protein. Partial least squares (PLS) regression models, with time as the Y variable, were calculated and also indicated that changes over time were apparent. Variable importance plots (VIP) confirmed the peaks observed in the PCA loading line plots. More systematic future experiments are needed to confirm this, but it can already be concluded that early detection of fungal contamination and activity is possible.     

Quantification of fructan concentration in grasses using NIR spectroscopy and PLSR

Near-infrared reflectance (NIR) spectroscopy combined with chemometrics was used to quantify fructan concentration in samples from seven grass species. Savitzky–Golay first derivative with filter width 7 and polynomial order 2 with mean centering was applied as a spectral pre-treatment method to remove unimportant baseline signals. In order to model the NIR spectroscopy data the partial least squares regression (PLSR) approach was used on the full spectra. Variable selection based on PLSR by jack-knifing within a cross-model validation (CMV) framework was applied in order to remove non-relevant spectral regions. PLSR was also used to model fructan concentrations from an augmented matrix [X|G], where X is spectra and G is correlation matrix of band specific information and X, in order to integrate the chemical band information in regression models. The present analysis showed that rapid quantification of fructans by NIR spectroscopy is possible and that jack-knifing PLSR within a CMV framework is an effective way to eliminate the wavelengths of no interest. Jack-knifing PLSR did not improve the predictive ability because the root mean square error of prediction (RMSEP) increased (1.37) compared to the full model (1.26). This was possibly due to signals from carbohydrates, which could act as cofactor in the prediction of fructans. However, jack-knifing PLSR within a CMV framework simplified the interpretation of the regression model with r² = 0.90 and RMSEP = 1.37.     

Evaluation of malting barley quality using exploratory data analysis. II. The use of kernel hardness and image analysis as screening methods

This paper presents an exploratory investigation of the use of image analysis and hardness analysis of barley kernels for characterisation and prediction of malting quality. A sample set of fifty barley samples representing 15 spring barley and 10 winter barley varieties grown at two locations in Denmark was used. The samples were micro-malted and mashed and analysed for 13 quality parameters according to the official methods of the European Brewery Convention. A sub-sample of the barley samples was analysed on two different single kernel instruments: (1) Foss Tecator GrainCheck was applied for non-destructive recording of single kernel size and shape (width, length, roundness, area, volume and total light reflectance) and (2) Perten Single Kernel Characterization System 4100 was applied for single kernel hardness and weight determinations. The eight variables from these single seed analyses have been used in two different ways, either as means and standard deviations, or as appended histogram spectra representing 250 kernels from each bulk sample. By the two methods, it has been possible to obtain reasonable Partial Least Squares Regression (PLSR) models for the structural and physical part of the malting quality complex associated to malt modification, but it was as expected impossible to predict the biochemical parameters associated with nitrogen chemistry and enzymatic power. The best model was achieved for (1→3, 1→4)-β-D-glucan in barley. The hardness of the barley kernels is by far the most important variable for describing malting performance. The additional use of the morphological data as acquired by fast non-destructive image analysis, however, also reveals some malting quality information by improving the calibration models based on hardness alone. The brightness of the kernels is by far the most important GrainCheck variable but also kernel size and shape is associated to malting performance. In general, the utilisation of the single kernel readings (used as histogram spectra), compared to sample mean and standard deviation, did not provide additional information for an improved prediction of the malting quality parameters.     

Amino acid profile of the quinoa (Chenopodium quinoa Willd.) using near infrared spectroscopy and chemometric techniques

The high content of amino acids of the quinoa, especially essential amino acids (higher than other cereals) makes a food increasingly demanded by consumers. A total of twelve amino acids (arginine, cystine, isoleucine, leucine, lysine, phenylalanine, proline, serine, threonine, tryptophan, tyrosine and valine) were analyzed in quinoa samples from Chile by near infrared spectroscopy (NIR) with direct application to the samples of a remote fiber-optic reflectance probe. The calibration results using modified partial least squares (MPLS) regression satisfactorily allowed the determination of the concentrations of this amino acid group with high multiple correlation coefficients (RSQ = 0.97–0.71) and low standard prediction errors (SEPC = 0.07–0.20). The prediction capacity (RPD) for the arginine, the cystine, the isoleucine, the lysine, the serine, the threonine, the tryptophan, the tyrosine and the valine ranged between 2.6 and 5.2, for the rest of amino acids were higher to 1.8, indicating that the NIRS equations obtained were applicable to unknown samples. It has confirmed that NIRS technology is a method that may be useful to replace the traditional methods for routine analysis of some amino acids.     

Multivariate analysis of nitrogen content for rice at the heading stage using reflectance of airborne hyperspectral remote sensing

Airborne hyperspectral remote sensing was adapted to establish a general-purpose model for quantifying nitrogen content of rice plants at the heading stage using three years of data. There was a difference in dry mass and nitrogen concentration due to the difference in the accumulated daily radiation (ADR) and effective cumulative temperature (ECT). Because of these environmental differences, there was also a significant difference in nitrogen content among the three years. In the multiple linear regression (MLR) analysis, the accuracy (coefficient of determination: R², root mean square of error: RMSE and relative error: RE) of two-year models was better than that of single-year models as shown by R² ≥ 0.693, RMSE ≤ 1.405 g m⁻² and RE ≤ 9.136%. The accuracy of the three-year model was R² = 0.893, RMSE = 1.092 g m⁻² and RE = 8.550% with eight variables. When each model was verified using the other data, the range of RE for two-year models was similar or increased compared with that for single-year models. In the partial least square regression (PLSR) model for the validation, the accuracy of two-year models was also better than that of single-year models as R² ≥ 0.699, RMSE ≤ 1.611 g m⁻² and RE ≤ 13.36%. The accuracy of the three-year model was R² = 0.837, RMSE = 1.401 g m⁻² and RE = 11.23% with four latent variables. When each model was verified, the range of RE for two-year models was similar or decreased compared with that for single-year models. The similarities and differences of loading weights for each latent variable depending on hyperspectral reflectance might have affected the regression coefficients and the accuracy of each prediction model. The accuracy of the single-year MLR models was better than that of the single-year PLSR models. However, accuracy of the multi-year PLSR models was better than that of the multi-year MLR models. Therefore, PLSR model might be more suitable than MLR model to predict the nitrogen contents at the heading stage using the hyperspectral reflectance because PLSR models have more sensitive than MLR models for the inhomogeneous results. Although there were differences in the environmental variables (ADR and ECT), it is possible to establish a general-purpose prediction model for nitrogen content at the heading stage using airborne hyperspectral remote sensing.     

利用近红外技术预测向日葵籽仁品质性状

为实现向日葵育种材料的品质性状快速预测,选取154份向日葵籽仁样品,结合化学测定值和近红外光谱,利用化学计量学手段建立向日葵籽仁品质指标的近红外模型,评估其在向日葵籽仁粗蛋白、粗脂肪、油酸、亚油酸等重要品质性状测定中的可行性。结果表明,改进偏最小二乘法建模效果最佳,其粗脂肪、粗蛋白、油酸、亚油酸、饱和脂肪酸及不饱和脂肪酸的定标相关系数分别为0.975、0.950、0.973、0.951和0.913,交叉验证相关系数分别 为0.969、0.939、0.915、0.927和0.711。用检验集对模型进行验证,粗脂肪、蛋白质、油酸、亚油酸、饱和及不饱和脂肪酸的外部检验相关系数（R2）分别为0.959、0.950、0.937、0.906和0.930。本研究建立的模型质量较高,能够满足向日葵籽仁品质成分的快速测定,可为向日葵品质育种前期大量、快速的筛选提供技术支持。     

基于Dualex植物多酚-叶绿素仪的冬小麦叶绿素含量高光谱估算

为探讨基于Dualex植物多酚-叶绿素仪和高光谱遥感技术反演小麦叶绿素含量的可行性,利用Dualex植物多酚-叶绿素仪,测定不同生育时期冬小麦叶片叶绿素含量(Chl),同时进行叶片光谱测定,以对Chl敏感的1个一阶导数波段、3个三边参数和3个植被指数作为自变量,利用偏最小二乘法(PLS)和支持向量回归(SVR)构建估测模型,并利用验证样本对各生育时期估测模型进行精度检验,同时与传统的单因素模型进行了比较。结果表明,冬小麦反射光谱曲线在不同生育时期有所不同,且随着叶绿素含量的增加,可见光波段的光谱反射率不断降低;在以一阶导数光谱敏感波段、三边参数以及植被指数构建的冬小麦Chl单因素估算模型中,基于各生育时期显著相关的植被指数构建的模型精度最优;以7个参数作为自变量,利用偏最小二乘法(PLS)和支持向量回归(SVR)构建的模型在各生育时期均表现出较好的拟合性及预测精度,尤其利用SVR建立的模型建模决定系数在0.8以上,预测决定系数在0.7以上,是进行冬小麦叶片Chl估测的最优模型。     

基于PROSAIL模型和无人机高光谱数据的冬小麦LAI反演

为及时准确高效监测小麦叶面积指数(leaf area index, LAI),获取了冬小麦挑旗期和开花期地面实测光谱与无人机高光谱遥感影像数据,并基于查找表建立PROSAIL辐射传输模型得到冬小麦冠层模拟光谱数据,利用数学统计回归模型与偏最小二乘回归法分别构建冬小麦LAI单变量、多变量预测模型,以实测LAI数据对预测结果进行精度评价,将最佳预测模型应用于无人机高光谱影像以分析LAI空间分布情况。结果表明,冬小麦各生育时期的预测模型均具有较高的预测精度,单变量预测模型和多变量预测模型的决定系数分别为0.598～0.717和0.577～0.755,其中以基于植被指数的多变量预测模型表现最优,其在开花期的验证精度最高,RMSE和MAPE分别为0.405和12.90%。在LAI空间分布图中,开花期预测效果优于挑旗期,各试验小区的LAI分布较为均匀。     

基于主成分分析和随机森林回归的冬小麦冠层叶绿素含量估算

为提高冬小麦冠层光谱对叶绿素含量的估算精度,以陕西省乾县冬小麦为研究对象,利用SVC-1024i光谱仪和SPAD-502型叶绿素仪实测了冬小麦冠层反射率和叶绿素含量,分析了一阶导数光谱、10种特征参数和9种植被指数与叶绿素含量的相关性,并利用主成分分析(PCA)对叶绿素敏感的可见光波段(390～780 nm)一阶导数光谱进行降维,将特征值大于1的主分量结合特征参数和植被指数形成不同的输入变量,用偏最小二乘回归和随机森林回归构建冬小麦冠层叶绿素估算模型,并利用独立样本对模型进行验证。结果表明,小麦冠层叶绿素含量与一阶导数光谱在751 nm处的相关性最高(r=0.71),特征参数中红边蓝边归一化(SDr-SDb)/(SDr+SDb)与叶绿素含量的相关性最高(r=0.66),植被指数(VI)中修正归一化差异指数(mND705)相关性最高(r=0.74)。在输入变量相同的情况下,基于随机森林(RF)回归的预测模型均优于偏最小二乘回归(PLSR)模型,其中PCA-VI-RF模型的各精度指标均达到最优(r²=0.94,RMSE=1.05,RPD=3.70),是冬小麦冠层叶绿素...     

Proximal sensing of the seasonal variability of pasture nutritive value using multispectral radiometry

The nutritive value of pasture is an important determinant of the performance of grazing livestock. Proximal sensing of in situ pasture is a potential technique for rapid prediction of nutritive value. In this study, multispectral radiometry was used to obtain pasture spectral reflectance during different seasons (autumn, spring and summer) in 2009–2010 from commercial farms throughout New Zealand. The analytical data set (n = 420) was analysed to develop season‐specific and combined models for predicting pasture nutritive‐value parameters. The predicted parameters included crude protein (CP), acid detergent fibre (ADF), neutral detergent fibre (NDF), ash, lignin, lipid, metabolizable energy (ME) and organic matter digestibility (OMD) using a partial least squares regression analysis. The calibration models were tested by internal and external validation. The results suggested that the global models can predict the pasture nutritive value parameters (CP, ADF, NDF, lignin, ME and OMD) with moderate accuracy (0·64 ≤ r² ≤ 0·70) while ash and lipid are poorly predicted (0·33 ≤ r² ≤ 0·40). However, the season‐specific models improved the prediction accuracy, in autumn (0·73 ≤ r² ≤ 0·83) for CP, ADF, NDF and lignin; in spring (0·61 ≤ r² ≤ 0·78) for CP and ash; in summer (0·77 ≤ r² ≤ 0·80) for CP and ash, indicating a seasonal impact on spectral response.     

基于机器学习算法的冬小麦不同生育时期生物量高光谱估算

为探索适用于冬小麦不同生育时期的高光谱估算方法,基于4年大田试验,以江苏省主要冬小麦品种为材料,以8种对常用生物量敏感的高光谱指数为基础,分别采用偏最小二乘算法、支持向量回归算法、随机森林算法在冬小麦4个主要生育时期(抽穗期前、抽穗期、开花期和灌浆期)进行了高光谱生物量估算和预测能力比较。结果表明,在冬小麦不同生育时期,高光谱估算生物量精度差异显著;利用随机森林构建的生物量估算模型在4个生育时期均表现出很好的效果,决定系数(r^2)和均方根误差(RMSE)在抽穗期前分别为0.79和44.82 g·m^-2,在抽穗期分别为0.71和62.07 g·m^-2,在开花期分别为0.70和97.63 g·m^-2,在灌浆期分别为0.71和106.98 g·m^-2;随机森林模型在4个生育时期的预测能力都高于或接近于支持向量回归模型,高于偏最小二乘回归模型,r^2和RMSE在抽穗期前分别为0.60和72.54 g·m^-2,在抽穗期分别为0.60和75.07 g·m^-2,在开花期分别为0.68和109.9 g·m^-2,在灌浆期分别为0.61和127.93 g·m^-2。这说明随机森林算法在冬小麦不同生育时期生物量高光谱遥感估算方面具有较高的精度和稳定性。     

Analysis of hyperspectral reflectance data for monitoring growth and development of lesquerella

Seed oil from lesquerella (Lesquerella fendleri (Gray) Wats.) is currently being developed as a biorenewable petroleum substitute, but several issues related to crop management and breeding must be resolved before the crop will be commercially viable. Due particularly to the prominent yellow flowers exhibited by lesquerella canopies, remote sensing may be a useful tool for monitoring and managing the crop. In this study, we used a hand-held spectroradiometer to measure spectral reflectance over lesquerella canopies in 512 narrow wavebands from 268 to 1095 nm over two growing seasons at Maricopa, Arizona. Biomass samples were also regularly collected and processed to obtain aboveground dry weight, flower counts, and silique counts. Partial least squares regression was used to develop predictive models for estimating the three lesquerella biophysical variables from canopy spectral reflectance. For model fitting and model testing, the root mean squared prediction errors between measured and modeled aboveground dry weight, flower counts, and silique counts were 2.1 and 2.3 Mg ha⁻¹, 251 and 304 flowers, and 1018 and 1019 siliques, respectively. Analysis of partial least squares regression coefficients and loadings highlighted the most sensitive spectral wavebands for estimating each biophysical variable. For example, the flower count model heavily emphasized the reflectance of yellow light at 583 nm, and contrasted that with reflectance in the blue (483 nm) and at the red edge (721 nm). Because of the indeterminate nature of lesquerella flowering patterns, remote sensing methods that monitor flowering progression may aid management decisions related to the timing of irrigations, desiccant application, and crop harvest.     

Trait dissection of maize kernel weight: Towards integrating hierarchical scales using a plant growth approach

Maize (Zea mays L.) yield is a function of the number harvested kernels per unit land area and the individual kernel weight (KW). Kernel weight and its development show a wide variability due to the genotype, the environment, the crop management, and all possible interactions. Commercial maize hybrids differ markedly in the patterns (rate and duration of kernel growth) behind differences in final KW. The same can be observed when public or elite proprietary maize inbred lines are analyzed. To progress in our understanding of KW variability, we reviewed and discussed current knowledge for analyzing kernel growth as an integrated system, modulated by processes linking different levels of organization (the different kernel tissues, the whole kernel, the plant, the canopy). Ideas on how to integrate this knowledge towards the development of a multi-hierarchical scale framework for predicting KW under different growth environments are currently needed, as they have high relevance for dissecting the genetic basis of kernel growth and maize yield definition at the canopy level.