Performance Of Partial Least Squares Research Articles

Determination of watermelon soluble solids content based on visible/near infrared spectroscopy with convolutional neural network

The soluble solids content（SSC）of fruits is an essential indicator of fruit taste and flavor. The near-infrared spectroscopy (NIRS) is widely used in fruit SSC detection. A common method to deal with such predictive modelling is the partial least-squares (PLSR). Wavelength selection algorithms are often employed to eliminate redundant wavelengths before constructing the model. However, the wavelength selection algorithms can not only increase the model complexity but also may lose valuable information. In this work, 1D-CNN was used to improve the prediction model of the watermelon SSC. First, the best parameter combination for 1D-CNN was found. The root mean square error of prediction (RMSEP) was 0.21 and the determination coefficient of prediction (Rp2) was 0.97. Then, the performance of PLSR and backpropagation neural network (BPNN) are compared. The results show that the Rp2 of 1D-CNN is improved by 14.1% and 6.6%, respectively. Finally, a gradient-based method for visualizing features of CNN regression models, namely regression activation mapping (RAM), was proposed. The results show that the features of 1D-CNN are consistent with the feature peaks of the spectrum, while the features of the BPNN and PLSR are not obvious.

A novel spectral index for estimation of relative chlorophyll content of sugar beet

Fast and non-destructive estimation of canopy chlorophyll content from hyperspectral sensing is essential to monitor the physiological status of vegetation or to estimate crop above ground nitrogen status. The objective of this study is to propose an optimal method for the relative chlorophyll content (SPAD) estimation of sugar beet canopy using ground-based hyperspectral imagery. Field experiments were conducted over three years at three different growth stages, across three different sites, using different cultivars and nitrogen (N) application rates. Quantitative correlations between SPAD value and canopy hyperspectral reflectance of sugar beet canopy after different pretreatment algorithms were established. Ten classical spectral indexes selected from the literature for estimating SPAD value in sugar beet canopy were evaluated and compared to a novel modified chlorophyll index (MCI) produced in this study by introducing a parameter to chlorophyll index (CI) to improve the estimation accuracy. Normalized difference vegetation index (NDVI) and chlorophyll index (CI) were optimized by using all possible combinations of spectral bands from the range of 390 nm to 990 nm. The prediction performance of partial least squares (PLS) regression models for optimized indexes (e.g., NDVI, CI and MCI), compared to the corresponding classical spectral indexes (e.g., ND550, ND705, CIgreen and CIred edge) was examined. Results showed that standard normal variate transformation (SNV) was the best pretreatment method for the hyperspectral data of this study. Models resulted after bands combinations optimization were found to be more accurate than models developed using the classical spectral indexes. The performance of proposed spectral index, MCI (R747, R839), MCI (R861, R884) and MCI (R931, R770), were best for the prediction accuracy of SPAD value in sugar beet for the validation set with the coefficient of determination (R2) of 0.83, 0.70 and 0.75, the root mean square error (RMSE) of 2.37, 3.11 and 2.78, and the relative root mean square error (RRMSE) of 4.95%, 6.05% and 5.75%, for the rapid growth stage of leaf cluster, sugar growth stage and sugar accumulation stage, respectively. It can be concluded that the index proposed can be implemented for the prediction of SPAD value of sugar beet using proximal hyperspectral sensors under a wide range of environmental conditions.

Comparison of the Performance of Partial Least Squares and Support Vector Regressions for Predicting Fatty Acids and Fatty Acid Classes in Marine Oil Dietary Supplements by Using Vibrational Spectroscopic Data

Simple, fast, and accurate analytical techniques for verifying the accuracy of label declarations for marine oil dietary supplements containing eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA) are required because of the increased consumption of these products. We recently developed broad-based partial least squares regression (PLS-R) models to quantify six fatty acids (FAs) and FA classes by using the spectroscopic data from a portable Fourier transform infrared (FTIR) device and a benchtop Fourier transform near infrared (FT-NIR) spectrometer. We developed an improved quantification method for these FAs and FA classes by incorporating a nonlinear calibration approach based on the machine learning technique support vector machines. For the two spectroscopic methods, high accuracy in prediction was indicated by low root mean square error of prediction and by correlation coefficients (R2) close to 1, indicating excellent model performance. The percent accuracy of the support vector regression (SV-R) model predicted values for EPA and DHA in the reference material was 90 to 110%. In comparison to PLS-R, SV-R accuracy for prediction of FA and FA class concentrations was up to 2.4 times higher for both ATR-FTIR and FT-NIR spectroscopic data. The SV-R models also provided closer agreement with the certified and reference values for the prediction of EPA and DHA in the reference standard. Based on our findings, the SV-R methods had superior accuracy and predictive quality for predicting the FA concentrations in marine oil dietary supplements. The combination of SV-R with ATR-FTIR and/or FT-NIR spectroscopic data can potentially be applied for the rapid screening of marine oil products to verify the accuracy of label declarations.

Model performance of partial least squares in utilizing the visible spectroscopy data for estimation of algal biomass in a photobioreactor

Abstract Spectroscopy technology and statistical methods (Partial Least Squares) have been integrated to develop a model that allows estimating the microalgal biomass in a photobioreactor. The model employing PLS combines the absorption spectrum measurements in the visible range (400–750 nm) with a microalgae cell density in a water sample. First, a calibration model was constructed using a calibration data set, and then, the predictive capacity of the model was determined by cross validation. Finally, an external validation of the predictive performance of the model was carried out with an independent data set. To test the accuracy of the model it was applied to different culture conditions yielding a predictive capacity of 96.7%. The results achieved are highly satisfactory due to the good lineal adjustment between observed cell densities vs. predicted ones obtained. According to the results obtained, the application of the model is a useful tool for the management and the decision-making process when operating a photobioreactor. Moreover, this model may boost the real-time measurements and may represent a previous step for further technical development in the “internet of things” applied to the management of the photobioreactor.

An improved partial least-squares regression method for Raman spectroscopy

It is known that the performance of partial least-squares (PLS) regression analysis can be improved using the backward variable selection method (BVSPLS). In this paper, we further improve the BVSPLS based on a novel selection mechanism. The proposed method is based on sorting the weighted regression coefficients, and then the importance of each variable of the sorted list is evaluated using root mean square errors of prediction (RMSEP) criterion in each iteration step. Our Improved BVSPLS (IBVSPLS) method has been applied to leukemia and heparin data sets and led to an improvement in limit of detection of Raman biosensing ranged from 10% to 43% compared to PLS. Our IBVSPLS was also compared to the jack-knifing (simpler) and Genetic Algorithm (more complex) methods. Our method was consistently better than the jack-knifing method and showed either a similar or a better performance compared to the genetic algorithm.

Variables selection for quantitative determination of cotton content in textile blends by near infrared spectroscopy

Investigations were initiated to develop near infrared (NIR) techniques coupled with variables selection method to rapidly measure cotton content in blend fabrics of cotton and polyester. Multiplicative scatter correction (MSC), smooth, first derivative (1Der), second derivative (2Der) and their combination were employed to preprocess the spectra. Monte Carlo uninformative variables elimination (MCUVE), successive projections algorithm (SPA), and genetic algorithm (GA) were performed comparatively to choose characteristic variables associated with cotton content distributions. One hundred and thirty-five and fifty-nine samples were used to calibrate models and assess the performance of the models, respectively. Through comparing the performance of partial least squares (PLS) regression models with new samples, the optimal model of cotton content was obtained with spectral pretreatment method of 2 Der-Smooth-MSC and variables selection method of MCUVE-SPA-PLS. The correlation coefficient of prediction (rp) and root mean square errors of prediction (RMSEP) were 0.988% and 2.100%, respectively. The results suggest that NIR technique combining with variables selection method of MCUVE-SPA has significant potential to quantitatively analyze cotton content in blend fabrics of cotton and polyester; moreover, it could indicate the related spectral contributions.

Using Vis/NIR Diffuse Transmittance Spectroscopy and Multivariate Analysis to Predicate Soluble Solids Content of Apple

The objectives of this research were to compare the effect of different fruit orientations on the quality of acquired spectra and to provide a suitable calibration model for further online determination of soluble solids content (SSC) of “Fuji” apples using visible and near-infrared (Vis/NIR) diffuse transmittance. The diffuse transmittance spectra between 650 and 910 nm were collected with the designed spectrum measurement system in two fruit orientations: stem-calyx axis horizontal (T1) and stem-calyx axis vertical (T2). Area change rate (ACR) was used to evaluate the stability of spectra collected in two fruit orientations. Results showed that the fruit orientation T1 was better for our designed spectrum measurement system. Then, the performance of partial least squares (PLS) models based on spectral data after the pretreatment of several preprocessing methods was analyzed and compared. Finally, the modified competitive adaptive reweighted sampling (MCARS), successive projection algorithm (SPA), and their combination were investigated to select the effective variables for the determination of SSC. It concluded that the MCARS-SPA-PLS model based on the spectra after preprocessing of Savitzky-Golay (SG) smoothing achieved better results for SSC prediction. The correlation coefficients between measured and predicted SSC were 0.962 and 0.946, and the root mean square errors were 0.510 and 0.527°Brix for calibration and prediction set, respectively. Moreover, the physicochemical properties of 27 variables selected by MCARS-SPA were discussed to obtain a better interpretation of the calibration model. The overall results indicated that the designed diffuse transmittance spectrum measurement system together with the PLS calibration model with 27 effective variables selected by MCARS-SPA method had a potential application for online SSC detection of apple.

Comparison of Chemometric Algorithms for Multicomponent Analyses and Signal Processing: An Example from 4-(2- Pyridylazo) Resorcinol-Metal Colored Complexes

Chemometrics, a relatively young area of analytical chemistry involves extracting meaningful information from experimental data by utilizing mathematical, statistical, and computational methods with the ultimate goal of evaluating and interpreting analytical data or signals. In this article, we provide a comparison of the performance of various chemometrics techniques using data obtained from the colored complexes formed between the ligand 4-(2-pyridylazo)resorcinol (PAR) and zinc, copper, nickel, manganese, and cobalt metals. Using a reduced Box-Behnken experimental design, we developed calibration models (n=15) and compared the performance of partial least squares (PLS), principal component regression (PCR), multiple linear regression (MLR) (K-matrix), and ridge regression (RR) chemometric techniques in an independent test set (n=7). Our results show that PLS mostly outperformed the other techniques in almost all metal components in the test set. This article provides an overview of various multivariate regression techniques used in chemometrics and provides a comparison of their performance for multi-analyte detection. A discussion of some recently developed multivariate regression techniques in chemometrics as well as challenges and future directions in this field are also discussed. Keywords: Box-Behnken design, chemometrics, multi-component detection, partial least squares, principal component regression, ridge regression.

The Elephant in the Room: Evaluating the Predictive Performance of Partial Least Squares (PLS) Path Models

Attempts to introduce predictive performance metrics into Partial Least Squares (PLS) path modeling have been slow and fall short of demonstrating impact on both practice and scientific development in PLS. This study contributes to PLS development by offering a comprehensive framework that identifies different dimensions of prediction and their effect on predictive performance evaluation with PLS. We use this framework to contextualize prior efforts in PLS and prediction, and to highlight potential opportunities and challenges. Our second contribution is to formally propose procedures to generate and evaluate different types of predictions from PLS models. The procedures account for the best practices identified in our framework. Our third contribution is outlining the many powerful ways in which predictive PLS methodologies can strengthen theory­ building research. Our framework, procedures, and research guidelines will hopefully form the basis for a more informed and unified development of the rigorous theoretical and practical applications of PLS.

Conflating Antecedents and Formative Indicators: A Comment on Aguirre-Urreta and Marakas

Aguirre-Urreta and Marakas [Aguirre-Urreta MI, Marakas GM (2014) Research note—Partial least squares and models with formatively specified endogenous constructs: A cautionary note . Inform. Systems Res. 25(4):761–778] aim to evaluate the performance of partial least squares (PLS) path modeling when estimating models with formative endogenous constructs, but their ability to reach valid conclusions is compromised by three major flaws in their research design. First, their population data generation model does not represent “formative measurement” as researchers generally understand that term. Second, their design involves a PLS path model that is misspecified with respect to their population model. Third, although their aim is to estimate a composite-based PLS path model, their design uses simulation data generated via a factor analytic procedure. In consequence of these flaws, Aguirre-Urreta and Marakas' (2014) study does not support valid inference about the behavior of PLS path modeling with respect to endogenous formatively measured constructs.

Rapid Determination of Process Variables of Chinese Rice Wine Using FT-NIR Spectroscopy and Efficient Wavelengths Selection Methods

There is a growing need for the effective fermentation monitoring during the manufacture of wine due to the rapid pace of change in the wine industry. In this study, Fourier transform near-infrared (FT-NIR) spectroscopy combined with chemometrics was applied to monitor time-related changes during Chinese rice wine (CRW) fermentation. Various wavelength selection methods and support vector machine (SVM) algorithm were used to improve the performances of partial least squares (PLS) models. In total, ten different calibration models were established. It was observed that the performances of models based on wavelength variables selected by variable selection methods were much better than those based on the full spectrum. In addition, nonlinear models outperformed linear models in prediction of fermentation parameters. After systemically comparing and discussing, it was found that for both ethanol and total acid, genetic algorithm-support vector machine (GA-SVM) models obtained the best result with excellent prediction accuracy. The correlation coefficients (R2 (pre)), root mean square error of prediction (RMSEP), and the residual predictive deviation (RPD) for the prediction set were 0.94, 3.02 g/L, and 8.7 for ethanol and 0.97, 0.10 g/L, and 6.1 for total acid, respectively. The results of this study demonstrated that FT-NIR could monitor and control CRW fermentation process rapidly and efficiently with efficient variable selection algorithms and nonlinear regression tool.

Research on outlier detection methods for determination of oil yield in oil shales using near-infrared spectroscopy

In the present paper, the outlier detection methods for determination of oil yield in oil shale using near-infrared (NIR) diffuse reflection spectroscopy was studied. During the quantitative analysis with near-infrared spectroscopy, environmental change and operator error will both produce outliers. The presence of outliers will affect the overall distribution trend of samples and lead to the decrease in predictive capability. Thus, the detection of outliers are important for the construction of high-quality calibration models. The methods including principal component analysis-Mahalanobis distance (PCA-MD) and resampling by half-means (RHM) were applied to the discrimination and elimination of outliers in this work. The thresholds and confidences for MD and RHM were optimized using the performance of partial least squares (PLS) models constructed after the elimination of outliers, respectively. Compared with the model constructed with the data of full spectrum, the values of RMSEP of the models constructed with the application of PCA-MD with a threshold of a value equal to the sum of average and standard deviation of MD, RHM with the confidence level of 85%, and the combination of PCA-MD and RHM, were reduced by 48.3%, 27.5% and 44.8%, respectively. The predictive ability of the calibration model has been improved effectively.

Assessment of rind quality of ‘Nules Clementine’ mandarin fruit during postharvest storage: 2. Robust Vis/NIRS PLS models for prediction of physico-chemical attributes

The robustness of visible to near infrared spectroscopy (Vis/NIRS) models is a crucial requirement for assessment of fruit quality parameters. This study was conducted to investigate the performance of partial least squares (PLS) models developed with data from individual orchard locations with those developed from combined orchard locations and two seasons in predicting postharvest rind physico-chemical properties related to susceptibility of ‘Nules Clementine’ mandarins to progressive rind breakdown disorder (RBD). Vis/NIRS signals were acquired on freshly harvested fruit and reference physico-chemical properties were measured after 8 weeks of storage at 8±0.5°C, including incidence of RBD, rind hue angle (h°), rind dry matter, and non-structural carbohydrates (sucrose, glucose, fructose, total carbohydrates) concentration. PLS regression with leave-one-out full cross validation was used to develop calibration models of studied parameters. The models were externally validated using data from a different location or season not included in the calibration. Prediction performance of PLS models of a single orchard location validated with data of an independent location was low but encouraging, with residual predictive deviation (RPD) values ranging from 0.95 to 1.58 for fructose models. The fructose calibration models developed using two combined orchard locations had higher prediction accuracy (RPD ranging between 1.32 and 1.97) than models of one orchard location. The performance of models developed from three orchard locations in 2012 to predict parameters of 2011 was better (RPD for fructose model=2.50) than models developed from individual orchards and two combined orchards. Results from this study demonstrated that Vis/NIRS models offer considerable robustness for non-invasive prediction of rind quality attributes which might predispose ‘Nules Clementine’ mandarin fruit to RBD.

Absorption characteristics and quantitative contribution of overtones and combination of NIR: Method development and validation

The present study focuses on the interpretation of the relationship between absorption characteristics and quantitative contribution in each near-infrared (NIR) frequency range. NIR spectra were measured for citric acid in each mixture solvent of water and ethanol. Accuracy profile theory was successfully utilized to validate the performance of partial least squares (PLS) models in each frequency range. It was found that saturated absorption characteristic (i.e. intensity) occurred in combination region (CR) and proportion of first combination-overtone (FCOT) region. The result of chemometric indicators, i.e. correlation coefficient (r), the root mean square errors of calibration and prediction (RMSEC and RMSEP), revealed the importance of second combination-overtone (SCOT) region in the calibration model performance. Simultaneously, it demonstrated that PLS model in FCOT or CR region could provide the lowest RMSEC and RMSEP, only when FCOT or CR region has rich transmission information. According to validation methodology parameters, i.e. relative bias, repeatability, intermediate precision, risk, lower limits of quantification (LLOQ), these results also indicated that FCOT or CR region was preferred as variables selection region, but if saturated absorption FCOT or CR region occurred, SCOT region was a better choice. In addition, the validated results testified that absorption intensities of solvent seriously interfered with model performance, especially for FCOT or CR region. Therefore, this revelation work provided a guideline for variables selection in each NIR frequency range.

Improved variable reduction in partial least squares modelling based on Predictive-Property-Ranked Variables and adaptation of partial least squares complexity

The calibration performance of partial least squares for one response variable (PLS1) can be improved by elimination of uninformative variables. Many methods are based on so-called predictive variable properties, which are functions of various PLS-model parameters, and which may change during the variable reduction process. In these methods variable reduction is made on the variables ranked in descending order for a given variable property. The methods start with full spectrum modelling. Iteratively, until a specified number of remaining variables is reached, the variable with the smallest property value is eliminated; a new PLS model is calculated, followed by a renewed ranking of the variables. The Stepwise Variable Reduction methods using Predictive-Property-Ranked Variables are denoted as SVR-PPRV. In the existing SVR-PPRV methods the PLS model complexity is kept constant during the variable reduction process. In this study, three new SVR-PPRV methods are proposed, in which a possibility for decreasing the PLS model complexity during the variable reduction process is build in. Therefore we denote our methods as PPRVR-CAM methods (Predictive-Property-Ranked Variable Reduction with Complexity Adapted Models). The selective and predictive abilities of the new methods are investigated and tested, using the absolute PLS regression coefficients as predictive property. They were compared with two modifications of existing SVR-PPRV methods (with constant PLS model complexity) and with two reference methods: uninformative variable elimination followed by either a genetic algorithm for PLS (UVE-GA-PLS) or an interval PLS (UVE-iPLS). The performance of the methods is investigated in conjunction with two data sets from near-infrared sources (NIR) and one simulated set. The selective and predictive performances of the variable reduction methods are compared statistically using the Wilcoxon signed rank test. The three newly developed PPRVR-CAM methods were able to retain significantly smaller numbers of informative variables than the existing SVR-PPRV, UVE-GA-PLS and UVE-iPLS methods without loss of prediction ability. Contrary to UVE-GA-PLS and UVE-iPLS, there is no variability in the number of retained variables in each PPRV(R) method. Renewed variable ranking, after deletion of a variable, followed by remodelling, combined with the possibility to decrease the PLS model complexity, is beneficial. A preferred PPRVR-CAM method is proposed.

Nonlinear Robust PLS Modeling of Wastewater Effluent Quality Indices

To solve the strong nonlinearity and data deterioration due to missing, outliers contained in the training data, this paper combines robust EMPCA (Expectation Maximization Principle Component Analysis) and the error-based input weights updating NNPLS (Neural Network Partial Least Square) to build a nonlinear and robust model as a software sensor of effluent quality indices for the anoxic-aeration activated sludge with nitrogen removal process in wastewater treatment pants. As the first step, data preprocessing based on the modified robust EMPCA is used to eliminate gross error, estimate missing data. Then an error-based input weights updating NNPLS (EB-NNPLS) is further used to predict effluent quality indices. This study compares the performance of partial least squares (PLS) regression analysis, polynomial PLS, NNPLS and EB-NNPLS with robust nonlinearity for the prediction of effluent quality. Simulations results for industrial process data show that the proposed method outperforms basic PLS, the polynomial PLS and NNPLS for the prediction of effluent quality indices.

Selection of Variables Based on Most Stable Normalised Partial Least Squares Regression Coefficients in an Ensemble Monte Carlo Procedure

A modification of ensemble Monte Carlo uninformative variable elimination (EMCUVE) is proposed, which does not involve the use of random variables, with the aim of improving the performance of partial least squares (PLS) regression models, increasing the consistency of results and reducing processing time by selecting the most informative variables in a spectral dataset. The proposed method (ensemble Monte Carlo variable selection—EMCVS) and the robust version (REMCVS) were compared to PLS models and with the existing EMCUVE method using three near infrared (NIR) datasets, i.e. prediction of n-butanol in a five-solvent mixture, moisture in corn and glucosinolates in rapeseed. The proposed methods were more consistent, produced models with better predictive accuracy (lower root mean squared error of prediction) and required less computational time than the conventional EMCUVE method on these datasets. In this application, the proposed method was applied to PLS regression coefficients but it may, in principle, be used on any regression vector.

Laser-Induced Breakdown Spectroscopy of Steel: A Comparison of Univariate and Multivariate Calibration Methods

Laser-induced breakdown spectroscopy (LIBS) was carried out on twenty-three low to high alloy steel samples to quantify their concentrations of chromium, nickel, and manganese. LIBS spectral data were correlated to known concentrations of the samples and three calibration methods were compared. A standard LIBS calibration technique using peak area integration normalized by an internal standard was compared to peak area integration normalized by total light and the multivariate statistical technique of partial least squares. For the partial least squares analysis, the PLS-1 algorithm was used, where a predictive model is generated for each element separately. Partial least squares regression coefficients show that the algorithm correctly identifies the atomic emission peaks of interest for each of the elements. Predictive capabilities of each calibration approach were quantified by calculating the standard and relative errors of prediction. The performance of partial least squares is on par with using iron as an internal standard but has the key advantage that it can be applied to samples where the concentrations of all elements are unknown.

Near-infrared multivariate calibration updating using placebo: A content uniformity determination of pharmaceutical tablets

This paper presents an application of near-infrared (NIR) multivariate calibration for the determination of content uniformity of pharmaceutical tablets. Data used in this study were acquired from 3 NIR instruments obtained from 2 different vendors. One of these spectrometers is a dispersive NIR system configured for reflectance measurements while the other two, are Fourier transform (FT) based configured for both transmittance and reflectance. The performance of partial least squares (PLS) calibration models generated from the primary instrument (A) was investigated using samples measured on secondary instruments (B and C). The cross-instrument predictions without standardization or updating was not successful in the reflectance mode for both instrument B and C producing root mean square error of prediction (RMSEP) values of 2.78 and 5.78%, respectively. Similar deterioration of the calibration model was evidenced for instrument B in the transmittance mode with an RMSEP value of 2.65%. To minimize the effect of instrument response variations, piecewise direct standardization (PDS) was used. This lowered the RMSEP in the reflectance mode for the secondary instruments B and C to 1.19 and 1.58%, respectively. Similar improvement was obtained in the transmittance mode for instrument B producing RMSEP of 1.12%. For cases in which the stability of drug substance is poor and the on-demand manufacture of transfer samples is difficult, a model updating strategy that uses placebo samples is proposed. This technique also outperformed the direct cross-instrument predictions with RMSEP values of 1.25 and 1.70% for instruments B and C, respectively, in the reflectance mode and 1.42% for instrument B in the transmittance mode.

The prediction of total anthocyanin concentration in red-grape homogenates using visible-near-infrared spectroscopy and artificial neural networks

This study compares the performance of partial least squares (PLS) regression analysis and artificial neural networks (ANN) for the prediction of total anthocyanin concentration in red-grape homogenates from their visible-near-infrared (Vis-NIR) spectra. The PLS prediction of anthocyanin concentrations for new-season samples from Vis-NIR spectra was characterised by regression non-linearity and prediction bias. In practice, this usually requires the inclusion of some samples from the new vintage to improve the prediction. The use of WinISI LOCAL partly alleviated these problems but still resulted in increased error at high and low extremes of the anthocyanin concentration range. Artificial neural networks regression was investigated as an alternative method to PLS, due to the inherent advantages of ANN for modelling non-linear systems. The method proposed here combines the advantages of the data reduction capabilities of PLS regression with the non-linear modelling capabilities of ANN. With the use of PLS scores as inputs for ANN regression, the model was shown to be quicker and easier to train than using raw full-spectrum data. The ANN calibration for prediction of new vintage grape data, using PLS scores as inputs, was more linear and accurate than global and LOCAL PLS models and appears to reduce the need for refreshing the calibration with new-season samples. ANN with PLS scores required fewer inputs and was less prone to overfitting than using PCA scores. A variation of the ANN method, using carefully selected spectral frequencies as inputs, resulted in prediction accuracy comparable to those using PLS scores but, as for PCA inputs, was also prone to overfitting with redundant wavelengths.