土壤有机质含量可见-近红外光谱反演模型校正集优选方法

土壤有机质含量可见-近红外光谱反演过程中校正集的构建策略对模型的预测精度有重要影响。以江汉平原洪湖地区水稻土为研究对象,采用Kennard-Stone(KS)法,Rank-KS(RKS)和Sample set Partitioning based on joint X-Y distance(SPXY)法,构建样本数占总校正集不同比例的子校正集,通过偏最小二乘回归,建立土壤有机质含量的可见—近红外光谱反演模型。结果表明:KS法无法提高模型预测精度,但可以在保证标准差与预测均方根误差比(ratio of performance to standard deviation,RPD)2.0的前提下减少30%的校正样本;基于SPXY法的模型,当子校正集样本比例为总校正集的50%时达到最佳的模型预测精度,RPD为2.557;RKS法能够在保证预测精度的情况下(RPD2.0),最多减少总校正集70%的样本,对应模型RPD为2.212。当校正集与验证集的有机质含量分布相近时,能够以较少的建模样本达到与总校正集相近甚至更高的模型预测精度,提升土壤有机质光谱反演模型的实用性。     

水稻遥感识别偏差修正的地统计学方法

为了进一步提高农作物遥感识别精度,充分利用高分辨率遥感影像上不同地物之间的邻域空间关系,提出农作物遥感识别偏差修正的地统计学方法。该方法综合考虑目标地物的光谱特征与空间信息,以类别隶属度偏差为研究对象,首先利用类别指示向量和类别后验概率向量之间的差异实现目标地物的类别隶属度偏差量化,然后对训练样本的类别隶属度偏差进行变异函数建模,并采用带局部均值的简单克里格插值方法预测总体类别隶属度偏差,之后用总体偏差的预测值对光谱分类所得的类别后验概率进行修正,重新确定识别结果,实现农作物遥感识别的偏差修正。以安徽省南部的一景 SPOT-5影像覆盖范围为研究区,选择2块典型区域分别作为试验区和验证区,以一季稻和晚稻为目标农作物,以支持向量机作为光谱分类的分类器,建立了水稻遥感识别的偏差修正流程;采用地面实测数据对修正效果进行评估,并与最大似然分类、模糊分类和支持向量机分类的结果进行比较。试验结果表明,该方法的总体分类精度能够达到90%以上,与传统分类方法相比,总体精度提高了近14%;且该方法能够大幅提高一季稻和晚稻的生产者精度和用户精度,有效改善了研究区的水稻识别结果,可以为中国南方复杂种植条件下的水稻识别提供参考。     

Prediction of Starch Amylose Content Versus Total Grain Amylose Content in Corn by Near-Infrared Transmittance Spectroscopy

A study was conducted to investigate methods of improving a near-infrared transmittance spectroscopy (NITS) amylose calibration that could serve as a rapid, nondestructive alternative to traditional methods for determining amylose content in corn. Calibrations were developed using a set of genotypes possessing endosperm mutations in single- and double-mutant combinations ranging in starch-amylose content (SAC) from -8.5 to 76%, relative to a standard curve. The influence of three factors were examined including comparing calibrations made against SAC versus grain amylose content (GAC), developing calibrations using partial least squares (PLS) analysis versus artificial neural networking (ANN), and using all samples in the calibrations set versus using progressively narrower ranges of SAC or GAC in the calibration set. Grain samples were divided into calibration and validation sets for PLS analysis while samples used in ANN were assigned to a training set, test set, and validation set. Performance statistics of the validation sets that were considered were the coefficient of determination (R), the standard error of prediction (SEP), and the ratio of the standard deviation of amylose values to the SEP (RPD), which were used to compare all NITS models. The study revealed an NITS prediction model for SAC (R = 0.96, SEP = 5.1%, RDP = 3.8) of similar precision to the best GAC model (R = 0.96, SEP = 2.7%, RPD = 3.5). Narrowing the amylose range of the calibration set generally did not improve performance statistics except for PLS models for SAC in which a decrease in SEP values was observed. In one model, the SEP improved while R and RPD remained constant (R = 0.94, SEP = 4.2%, RPD = 2.8) when samples with SAC values <20% were removed from the calibration set. Although the NITS amylose calibrations in this study are of limited precision, they may be useful when a rough screening method is needed for SAC. For example, NITS may be useful to detect severe contamination during transport and storage of specialty grains or to aid breeders when selecting for amylose content from large numbers of grain samples.     

Application of FT-MIR spectroscopy for fast control of red grape phenolic ripening

The content of phenolic compounds determines the state of phenolic ripening of red grapes and is a key criterion in setting the harvest date to produce quality red wines. In this study, the feasibility of Fourier transform mid-infrared (FT-MIR) spectroscopy combined with partial least-squares (PLS) regression to quantify phenolic compounds is reported. The reference methods used for quantifying these compounds (which were evaluated as total phenolic compounds, total anthocyanins, and condensed tannins) were the usual ones used in cellars that employed UV-vis spectroscopy. To take into account the high natural variability of grapes when building the calibration models, fresh grapes from six varieties, at different phenolic ripening states were harvested during three vintages. Destemmed and crushed grapes were subjected to an accelerated extraction process and used as calibration standards. A total of 192 extracts (objects) were obtained, and these were divided into a training set (106 objects) and a test set (86 objects) to evaluate the predictive ability of the models. Among the different MIR regions of the extract raw spectra, those that provided the highest variability on the absorption were selected. The results showed that the best PLS regression model was the one obtained when working in the region of 1168-1457 cm(-1) because it gave the most accurate and robust prediction for total phenolic compounds (RMSEP%=4.3 and RPD=4.5), total anthocyanins (RMSEP%=5.9 and RPD=3.5), and condensed tannins (RMSEP%=5.8 and RPD=3.8). Therefore, it can be concluded that FT-MIR spectroscopy can be a fast and reliable technique for monitoring the phenolic ripening in red grapes during the harvest period.     

基于EPO算法去除水分影响的土壤有机质高光谱估算

野外进行土壤有机质的光谱快速预测时需考虑土壤含水量的影响。在室内设计人工加湿实验分别获取9个土壤含水量梯度(0~32%,间隔4%)的土壤光谱数据,分析土壤含水量变化对光谱的影响,再利用外部参数正交化法(external parameter orthogonalization,EPO)进行湿土光谱校正,并结合偏最小二乘回归和支持向量机回归分别建立土壤有机质预测模型。结果表明,土壤光谱反射率随着土壤含水量的增加呈非线性降低趋势,偏最小二乘回归模型的预测偏差比为1.16,模型不可用;经EPO算法校正后,各土壤含水量梯度之间的光谱差异性降低,能实现土壤有机质在不同土壤含水量梯度的有效估算,偏最小二乘回归和支持向量机回归模型的预测偏差比分别提高至1.76和2.15。研究结果可为田间快速预测土壤有机质提供必要参考。     

Simultaneous assessment of key properties of arid soil by combined PXRF and Vis–NIR data

Arid soil is common worldwide and has unique properties that often limit agronomic productivity, specifically, salinity expressed as soluble salts and large amounts of calcium carbonate and gypsum. Currently, the most common methods for evaluating these properties in soil are laboratory‐based techniques such as titration, gasometry and electrical conductivity. In this research, we used two proximal sensors (portable X‐ray fluorescence (PXRF) and visible near‐infrared diffuse reflectance spectroscopy (Vis–NIR DRS)) to scan a diverse set (n = 116) of samples from arid soil in Spain. Then, samples were processed by standard laboratory procedures and the two datasets were compared with advanced statistical techniques. The latter included penalized spline regression (PSR), support vector regression (SVR) and random forest (RF) analysis, which were applied to Vis–NIR DRS data, PXRF data and PXRF and Vis–NIR DRS data, respectively. Independent validation (30% of the data) of the calibration equations showed that PSR + RF predicted gypsum with a ratio of performance to interquartile distance (RPIQ) of 5.90 and residual prediction deviation (RPD) of 4.60, electrical conductivity (1:5 soil : water) with RPIQ of 3.14 and RPD of 2.10, Ca content with RPIQ of 2.92 and RPD of 2.07 and calcium carbonate equivalent with RPIQ of 2.13 and RPD of 1.74. The combined PXRF and Vis–NIR DRS approach was superior to those that use data from a single proximal sensor, enabling the prediction of several properties from two simple, rapid, non‐destructive scans.     

基于相似光谱匹配预测土壤有机质和阳离子交换量

土壤可见光-近红外波段光谱(350~2 500 nm)包含了大量的土壤属性信息,相同类型的土壤具有相似的光谱曲线特征,但相似光谱曲线是否具有相似的属性含量?探讨此问题可为土壤光谱库的应用提供依据,从而最终服务于快速获取土壤信息技术体系的构建。该研究以安徽宣城为研究区,根据母质、地形特征和土地利用等信息,采集91个典型土壤剖面,共含400个土壤发生层样品,测定了有机质(soil organic matter,SOM)和阳离子交换量(cation exchange capacity,CEC)含量,同时采用VARIAN公司的Cary 5000分光光度计测定了土壤光谱,并将光谱数据变换为反射率(R)、反射率一阶导数(FDR)和吸收度(Log(1/R))3种形式。该文采用光谱角(spectral angle mapper,SAM)、偏最小二乘回归(partial least square regression,PLSR)和SAM-PLSR(spectral angle mapper-partial least square regression,SAM-PLSR)3种方法预测土壤SOM和CEC。SAM方法是通过对测试集104个光谱曲线与参考集的296个光谱曲线进行相似性计算,并以此实现土壤SOM和CEC含量的预测。SAM-PLSR方法以SAM算法下的匹配结果作为建模样本建立PLSR模型和进行预测分析。结果表明,具有相似光谱曲线的土壤具有相似的SOM和CEC含量,SAM算法下相似光谱匹配可直接预测SOM(R2=0.78,RPD=2.17)和CEC(R2=0.82,RPD=2.41)。PLSR方法可很好地预测SOM(R2=0.87,RPD=2.77)和CEC(R2=0.87,RPD=2.59);相较之下,SAM-PLSR方法不仅可以更加准确预测SOM(R2=0.89,RPD=3.00)和CEC(R2=0.91,RPD=3.06),而且大大减少了建模样本的数量。该研究使可见光-近红外光谱可更加高效地用于土壤属性分析,并为土壤光谱数据库的建设及应用提供技术参考。     

Principal component analysis applied to Fourier transform infrared spectroscopy for the design of calibration sets for glycerol prediction models in wine and for the detection and classification of outlier samples

Principal component analysis (PCA) was used to identify the main sources of variation in the Fourier transform infrared (FT-IR) spectra of 329 wines of various styles. The FT-IR spectra were gathered using a specialized WineScan instrument. The main sources of variation included the reducing sugar and alcohol content of the samples, as well as the stage of fermentation and the maturation period of the wines. The implications of the variation between the different wine styles for the design of calibration models with accurate predictive abilities were investigated using glycerol calibration in wine as a model system. PCA enabled the identification and interpretation of samples that were poorly predicted by the calibration models, as well as the detection of individual samples in the sample set that had atypical spectra (i.e., outlier samples). The Soft Independent Modeling of Class Analogy (SIMCA) approach was used to establish a model for the classification of the outlier samples. A glycerol calibration for wine was developed (reducing sugar content < 30 g/L, alcohol > 8% v/v) with satisfactory predictive ability (SEP = 0.40 g/L). The RPD value (ratio of the standard deviation of the data to the standard error of prediction) was 5.6, indicating that the calibration is suitable for quantification purposes. A calibration for glycerol in special late harvest and noble late harvest wines (RS 31-147 g/L, alcohol > 11.6% v/v) with a prediction error SECV = 0.65 g/L, was also established. This study yielded an analytical strategy that combined the careful design of calibration sets with measures that facilitated the early detection and interpretation of poorly predicted samples and outlier samples in a sample set. The strategy provided a powerful means of quality control, which is necessary for the generation of accurate prediction data and therefore for the successful implementation of FT-IR in the routine analytical laboratory.     

Development of NIRS equations for food grain quality traits through exploitation of a core collection of cultivated sorghum

A sorghum core collection representing a wide range of genetic diversity and used in the framework of a sorghum breeding and genetics program was evaluated by near-infrared reflectance spectroscopy (NIRS) to predict food grain quality traits: amylose content (AM), protein content (PR), lipid content (LI), endosperm texture (ET), and hardness (HD). A total of 278 sorghum samples were scanned as whole and ground grain to develop calibration equations. Laboratory analyses were performed on NIRS sample subsets that preserved the core collection racial distribution. Principal component analysis performed on NIRS spectra evidenced a level of structure following known sorghum races, which underlined the importance of using a wide range of genetic diversity. Performances of calibration equations were evaluated by the coefficient of determination, bias, standard error of laboratory (SEL), and ratio of performance deviation (RPD). Ground grain spectra gave better calibration equations than whole grain. PR equation (RPD of 5.7) can be used for quality control. ET, LI, and HD equations (RPD of 2.9, 2.6, and 2.6, respectively) can be used for screening steps. Even with a small SEL in whole sample analysis, a RPD of 1.8 for AM confirmed that this variable is not easy to predict with NIRS.     

Use of NIRS for quantification of mangiferin and hesperidin contents of dried green honeybush (Cyclopia genistoides) plant material

Cyclopia genistoides, normally used for the preparation of an herbal tea, honeybush, is a good source of the bio-active compounds mangiferin and hesperidin and is in demand for the preparation of xanthone-enriched extracts. Near-infrared spectroscopy (NIRS) was used to develop calibration models to predict the mangiferin and hesperidin contents of the dried green plant material. NIRS measurements of plant material and pure compounds were performed in diffuse reflectance mode. The calibration sets for mangiferin and hesperidin contents ranged from 0.7 to 7.21 and 0.64-4.80 g/100 g, respectively. Using independent validation, it was shown that the NIRS calibration models for the prediction of mangiferin (SEP=0.46 g/100 g; R2=0.74; and RPD=1.96) and hesperidin (SEP=0.38 g/100 g; R2=0.72; and RDP=1.90) contents of the dried plant material are adequate for screening purposes, based on RPD values.     

基于可见/近红外光谱技术的土壤总磷分析研究

Overabundance of phosphorus (P) in soils and water is of great concern and has received much attention in Florida, USA. Therefore, it is essential to analyze and predict the distribution of P in soils across large areas. This study was undertaken to model the variation of soil total phosphorus (TP) in Florida. A total of 448 soil samples were collected from different soil types. Soil samples were analyzed by chemical reference method and scanned in the visible/near-infrared (VNIR) region of 350--2 500 nm. Partial least squares regression (PLSR) calibration model was developed between chemical reference values and VNIR values. The coefficient of determination (R²) and the root mean squares error (RMSE) of calibration and validation sets, and the residual prediction deviation (RPD) were used to evaluate the models. The R² in calibration and validation for log-transformed TP (log TP) were 0.69 and 0.65, respectively, indicating that VNIR calibration obtained in this study accounted for at least 65% of the variance in log TP using only VNIR spectra, and the high RPD of 2.82 obtained suggested that the spectral model derived in this study was suitable and robust to predict TP in a wide range of soil types, being representative of Florida soil conditions.     

基于变量优选与机器学习的干旱区湿地土壤盐渍化数字制图

土壤盐渍化是导致土壤退化和生态系统恶化的主要原因之一,对干旱区的可持续发展构成主要威胁。为了尽可能精确地监测土壤盐渍化的空间变异性,该研究收集新疆艾比湖湿地78个典型样点,其中选取54个样本作为训练集,24个样本作为独立验证集。基于Sientinel-2多光谱传感器(Multi-Spectral Instrument,MSI)、数字高程模型(Digital Elevation Model,DEM)数据提取3类指数(红边光谱指数、植被指数和地形指数),经过极端梯度提升(Extreme Gradient Boosting,XGBoost)算法筛选有效特征变量,构建了关于土壤电导率(Electrical Conductivity,EC)的随机森林(Random Forest,RF)、极限学习机(Extra Learning Machine,ELM)和偏最小二乘回归(Partial Least Squares Regression,PLSR)预测模型,并选择最优模型绘制了艾比湖湿地盐渍化分布图。结果表明:优选的红边光谱指数基本能够预测EC的空间变化;红边光谱指数与植被指数组合建模效果总体上优于...     

Use of visible and near infrared spectroscopy and least squares-support vector machine to determine soluble solids content and pH of cola beverage

Visible and near infrared (VIS/NIR) transmission spectroscopy and chemometric methods were utilized for the fast determination of soluble solids content (SSC) and pH of cola beverage. A total of 180 samples were used for the calibration set, whereas 60 samples were used for the validation set. Some preprocessing methods were applied before developing the calibration models. Several PLS factors, extracted by partial least squares (PLS) analysis, were used as the inputs of least squares-support vector machine (LS-SVM) model according to their accumulative reliabilities. The correlation coefficient (r), root mean square error of prediction (rmsEP), bias, and RPD were 0.959, 1.136, -0.185, and 3.5 for SSC, whereas 0.973, 0.053, 0.017, and 4.1 for pH, respectively. An excellent prediction precision was achieved by LS-SVM compared with PLS. The results indicated that VIS/NIR spectroscopy combined with LS-SVM could be applied as a rapid and alternative way for the fast determination of SSC and pH of cola beverage.     

采用SEPLS_ELM模型估算夏玉米地上部生物量和叶面积指数

利用高光谱数据进行作物生长状况监测具有无损和高效的特点,是现代精准农业发展的必要手段。该研究以连续3 a（2018-2020年）不同水氮供应下夏玉米营养生长期采集的212份植物样品（地上部生物量和叶面积指数）和高光谱实测数据为数据源,分别采用偏最小二乘回归（Partial Least Squares Regression,PLS）、极限学习机（Extreme Learning Machine,ELM）、随机森林（Random Forest,RF）和基于PLS叠加策略的叠加极限学习机算法（Stacked Ensemble Extreme Learning Machine based on the PLS,SEPLS_ELM）构建了夏玉米营养生长期地上部生物量和叶面积指数估算模型。结果表明：基于PLS和ELM构建的夏玉米地上部生物量和叶面积指数估算模型的精度均较低,前者验证集R2低于0.85、均方根误差高于550 kg/hm2,后者R2低于0.90、均方根误差高于0.40 cm2/cm2。相比之下,基于RF和SEPLS_ELM构建的夏玉米营养生长期地上部生物量和叶面积指数估算模型均有着较高的估算精度,SEPLS_ELM模型表现尤为突出,其地上部生物量和叶面积指数估算模型验证集的R2分别为0.955和0.969,均方根误差分别为307.3 kg/hm2和0.24 cm2/cm2,表明叠加集成模型能够充分利用高光谱数据并提高作物地上部生物量和叶面积指数估算精度。     

Quality Prediction of Rice Flour by Multiple Regression Model with Instrumental Texture Parameters of Single Cooked Milled Rice Grains

The purpose of this study was to develop highly accurate regression models with texture parameters of cooked milled rice grains for predicting pasting properties in terms of quality index of rice flour. Two methods were adopted as the texture measurement to acquire predictors for the models. In the calibration set, all the multiple regression models by a single‐grain method exhibited a higher R² than those by a three‐grain method. Each of the former models also showed a lower SEP and a higher RPD in the validation set. The prediction performance was best for consistency (RPD = 2.4). The single‐grain method was more advantageous for the pasting prediction. These results suggest that the models based on grain texture could predict rice flour quality.     

不同品种间的猪肉含水率高光谱模型传递方法研究

针对目前的模型传递方法研究大多为不同仪器间的近红外光谱模型传递,该文采用高光谱技术建立猪肉含水率定量检测模型,并针对不同品种间的模型传递提出了一种分段直接校正结合线性插值(piecewise direct standardization combine with linear interpolation,PDS-LI)的传递算法。以杜长大、茂佳山黑猪和零号土猪3个品种为研究对象,以杜长大作为主品种,茂佳山黑猪和零号土猪作为从品种,采用偏最小二乘回归(partial least squares regression,PLSR)法建立猪肉含水率主模型,经PDS-LI算法对主模型进行传递后,主模型对茂佳山黑猪和零号土猪样品的预测决定系数R2p分别由传递前的0.263和0.507提高到0.832和0.848,预测均方根误差分别由传递前的1.151%和0.857%降低到0.470%和0.440%,剩余预测偏差(residual prediction deviation,RPD)分别由传递前的1.000和1.214提高到2.447和2.364。结果表明,PDS-LI传递算法能够实现杜长大对茂佳山黑猪和零号土猪样品的模型传递。研究结果为提高猪肉含水率模型适配性问题提供参考。     

Performance comparison of UV and FT-Raman spectroscopy in the determination of conjugated linoleic acids in cow milk fat

The determination of conjugated linoleic acids (CLA) in cow milk fat was studied by using UV (210-250 nm) and Fourier transform (FT)-Raman (900-3400 cm (-1)) spectroscopy in order to determine the best spectrophotometric technique for routine analysis of milk fat. A collection of 57 milk fat samples was randomly divided into two sets, a calibration set and a validation set, representing two-thirds and one-third of the samples, respectively. All calculations were performed on the calibration set and then applied to the validation set. The CLA content ranged from 0.56 to 4.70%. A comparison of various spectral pretreatments and different multivariate calibration techniques, such as partial least-squares (PLS) and multiple linear regression (MLR), was done. This paper shows that UV spectroscopy is as reliable as FT-Raman spectroscopy to monitor CLA in cow milk fat. The best calibration for FT-Raman was given by a PLS model of seven factors with a standard error of prediction (SEP) of 0.246. For UV spectroscopy, PLS models were also better than MLR models. The most robust PLS model was constructed with only one factor and with SEP=0.288.     

Joint Modeling of Spatial Variability and Within-Row Interplot Competition to Increase the Efficiency of Plant Improvement

Trials in the early stages of selection are often subject to variation arising from spatial variability and interplot competition, which can seriously bias the assessment of varietal performance and reduce genetic progress. An approach to jointly model both sources of bias is presented. It models genotypic and residual competition and also global and extraneous spatial variation. Variety effects were considered random and residual maximum likelihood was used for parameter estimation. Competition at the residual level was examined using two special simultaneous autoregressive models. An equal-roots second-order autoregressive (EAR(2)) model is proposed for trials where competition is dominant. An equal-roots third-order autoregressive (EAR(3)) model allows for competition and spatial variability. These models are applied to two yield data sets from an Australian sugarcane selection program. One data set is in the paper and the other is in supplementary material available online. To determine the effect of simultaneously adjusting for spatial variability and interplot competition on selection, the percentages of superior varieties in common in the top 15% for the joint model and classical approaches were compared. Agreement between the two approaches was 45 and 84%. Hence, for some trials there are large differences in varieties advanced to the next stage of selection.     

Rapid estimation of taro (Colocasia esculenta) quality by near-infrared reflectance spectroscopy

The aim of the present study is to develop a methodology for the rapid estimation of taro (Colocasia esculenta) quality. Chemical analyses were conducted on 315 accessions for major constituents (starch, total sugars, cellulose, proteins, and minerals). NIRS calibration equations, developed on a calibration set composed of 243 accessions, showed high explained variances in cross-validation (r(2)(cv)) for starch (0.89), sugars (0.90), proteins (0.89), and minerals (0.90) but poor response for amylose (0.44) and cellulose (0.61). The predictions were tested on an independent set of 58 randomly selected accessions. The r(2)(pred) values for starch, sugars, proteins, and minerals were, respectively, of 0.76, 0.74, 0.85, and 0.85 with ratios of performance to deviation (RPD) of 3.41, 4.01, 3.78, and 3.64. New calibration equations developed on 303 accessions confirmed good RPD values for starch (3.30), sugars (4.13), proteins (3.61), and minerals (3.74). NIRS could be used to predict starch, sugars, proteins, and minerals contents in taro corms with reasonably high confidence.     

基于近红外光谱技术的三种硬实种子无损鉴定

摘要：为探讨近红外光谱技术在鉴定种子硬实特性上的普遍性,该文采用近红外光谱法结合偏最小二乘法建立了大豆、苦豆子和决明子单粒种子硬实特性的定性分析模型,每种种子均选择120粒种子进行近红外定性分析,种子分为建模集、检验集2组,建模集80粒,检验集40粒,各组中硬实与非硬实种子的比例为1:1,比较了光谱重复次数、光谱范围以及不同建模样品的建模效果。结果表明：采用二次平均光谱所建模型的鉴别率优于单次光谱;大豆采用4 000～5 000 cm-1光谱范围,矢量校正预处理,主成分为8时,建模集与检验集鉴别率均在85%以上;决明子采用4 000～8 000 cm-1光谱范围,一阶导数预处理,主成分为4时,模型建模集与检验集鉴别率均在90%左右;苦豆子采用4 000～8 000 cm-1光谱范围,二阶导数预处理,主成分为8时,模型的建模集与检验集鉴别率均在95%以上。以上结果表明近红外光谱技术可以很好地应用于单粒种子硬实特性的判断,有助于硬实机理的深入研究。