Jean-Michel Roger

EPO-PLS external parameter orthogonalisation of PLS application to temperature-independent measurement of sugar content of intact fruits

Abstract Near Infrared (NIR) spectrometry would present a high potential for on-line measurement if the robustness of multivariate calibration was improved. The lack of robustness notably appears when an external parameter varies—e.g. the product temperature. This paper presents a preprocessing method which aims at removing from the X space the part mostly influenced by the external parameter variations. This method estimates this parasitic subspace by computing a PCA on a small set of spectra measured on the same objects, while the external parameter is varying. An application to the influence of the fruit temperature on the sugar content measurement of intact apples is presented. Without any preprocessing, the bias in the sugar content prediction was about 8° Brix for a temperature variation of 20 °C. After External Parameter Orthogonalisation (EPO) preprocessing, the bias is not more than 0.3° Brix, for the same temperature range. The parasitic subspace is studied by analysing the b -coefficient of a Partial Least Square Regression (PLS) between the temperature and the influence spectra. Further work will be achieved to apply this method to the case of multiple external parameters and to the calibration transfer issue.

Authenticating white grape must variety with classification models based on aroma sensors, FT-IR and UV spectrometry

Abstract This paper aims at assessing the capability of high-speed analytical devices, such as aroma sensors (“electronic noses”), Fourier Transform InfraRed (FT-IR) and ultraviolet spectrometers to classify white grape musts (grape juices before fermentation) in variety categories. Due to the complexity of the signal generated, specific data processing techniques have been developed and are described here. First, a pre-processing technique, based on Genetic Algorithms , is applied to spectra to improve spectrometer efficiency without expert knowledge in spectrometry; by selecting the most discriminant subsets of wavelengths, this stochastic method tends to reduce over-fitting and improves classification results. Secondly, the Partial Least Squares Regression technique is adapted to a pattern recognition problem, using Partial Least Squares-Discriminant Analysis , a multivariate classification technique. These devices and data processing techniques are applied to more than 100 must samples. FT-IR spectrometry is the most satisfactory technique with a 9.6% classification error level. Finally, outputs of the three individual sensors are combined in a “low-level” fusion method, by concatenating the individual sensor signals. This straightforward fusion method does not significantly improve results.

Robustness of models developed by multivariate calibration. Part I: The assessment of robustness

Abstract Monitoring products for quality assurance in real-time during industrial processes has become of great importance in recent years. Infrared spectroscopic (IRS) techniques combined with multivariate calibration methods are primarily used for on-line analysis, in situ sensors or automatic sampling. In order to ensure the correct use of these methods for routine industrial use, all the mechanical and the environmental conditions need to be taken into account, as well as the introduction of time delays and signal bias during sampling. This requires a robustness study of the IRS measurement and the calibration model used. In this review, we focus on both identifying the “robustness” used for multivariate calibration and the different methods applied to evaluate this robustness, especially with regard to the IRS technique used in industry. We also present and discuss various criteria intended for robustness assessment.

Fusion of aroma, FT-IR and UV sensor data based on the Bayesian inference. Application to the discrimination of white grape varieties

Abstract The objective of this study is to present a fusion method based on the Bayesian inference to combine the outputs of various sensors. The sensors studied here are aroma sensors, FT-IR and UV spectrometers. The application deals with classifying musts of white grapes according to their variety. The fusion procedure is not based on the combination of the signals, but of the class assignments provided individually by each sensor. Two methods have been developed based on the Bayesian inference: the Bayesian minimum error fusion rule and the minimum risk rule. The latter involves both experimental knowledge, in computing error probability values, and expert knowledge, through the level of error costs. The paper presents the mathematical theory concerning the Bayesian approach and the results obtained on white grape classification. This effective fusion method leads to a significant improvement in the grape variety discrimination: the final misclassification error is 4.7%, whereas the best individual sensor (FT-IR) gave a misclassification error twice as high, i.e. 9.6%. Bayesian fusion proved to be very well suited to the combination of all kinds of analytical measurements or sensors (curves or single value outputs), as long as they provide individual classification outputs. Furthermore, Bayesian fusion is able to cope with sensors providing large, noisy and redundant data as well as sensors showing very dissimilar efficiency levels.

Evaluation of oil-palm fungal disease infestation with canopy hyperspectral reflectance data.

Fungal disease detection in perennial crops is a major issue in estate management and production. However, nowadays such diagnostics are long and difficult when only made from visual symptom observation, and very expensive and damaging when based on root or stem tissue chemical analysis. As an alternative, we propose in this study to evaluate the potential of hyperspectral reflectance data to help detecting the disease efficiently without destruction of tissues. This study focuses on the calibration of a statistical model of discrimination between several stages of Ganoderma attack on oil palm trees, based on field hyperspectral measurements at tree scale. Field protocol and measurements are first described. Then, combinations of pre-processing, partial least square regression and linear discriminant analysis are tested on about hundred samples to prove the efficiency of canopy reflectance in providing information about the plant sanitary status. A robust algorithm is thus derived, allowing classifying oil-palm in a 4-level typology, based on disease severity from healthy to critically sick stages, with a global performance close to 94%. Moreover, this model discriminates sick from healthy trees with a confidence level of almost 98%. Applications and further improvements of this experiment are finally discussed.

Application of independent component analysis on Raman images of a pharmaceutical drug product: Pure spectra determination and spatial distribution of constituents

Independent component analysis (ICA) was used as a blind source separation method on a Raman image of a pharmaceutical tablet. Calculations were performed without a priori knowledge concerning the formulation. The aim was to extract the pure signals from the initial data set in order to examine the distribution of actives and major excipients within the tablet. As a method based on the decomposition of a matrix of mixtures of several components, the number of independent component to choose is a critical step of the analysis. The ICA_by_blocks method, based on the calculation of several models using an increasing number of independent components on initial matrix blocks, was used. The calculated ICA signals were compared with the pure spectra of the formulation compounds. High correlations between the two active principal ingredient spectra and their corresponding calculated signals were observed giving a good overview of the distributions of these compounds within the tablet. Information from the major excipients (lactose and avicel) was found in several independent components but the ICA approach provides high level of information concerning their distribution within the tablet. However, the results could vary considerably by changing the number of independent components or the preprocessing method. Indeed, it was shown that under-decomposition of the matrix could lead to better signal quality (compared to the pure spectra) but in that case the contributions due to minor components or effects were not correctly identified and extracted. On the contrary, over-decomposition of the original dataset could provide information about low concentration compounds at the expense of some loss of signal interpretability for the other compounds.

Correction of the temperature effect on near infrared calibration: application to soluble solid content prediction

Non-destructive fruit quality assessment in packing houses can be carried out using near infrared spectroscopy. However, the prediction performance can be affected by measurement conditions (“external parameters”), such as temperature or variation in stray light. In this work we propose a methodology to reduce the effect of the fruit temperature on sugar content prediction. Two approaches were used to correct for the temperature effect depending on whether the temperature is measurable or not. In the case of measurable temperature, three methods were tested. The first uses a spectrum correction while the second and third are based on regression coefficients which vary with temperature. In case of non-measurable temperature, the studies have led to robust calibration models and to a self-correcting model (where fruit temperature is estimated using spectral data). All techniques were tested with “Golden Delicious” apples. Our study has shown that the most efficient models remove the temperature from the spectral space. These methodologies can be used to minimise other external parameters in NIR calibration.

Major Issues of Diffuse Reflectance NIR Spectroscopy in the Specific Context of Soil Carbon Content Estimation: A Review

Abstract Soil carbon sequestration is one possible way of reducing greenhouse gas emissions in the atmosphere. However, to evaluate the real benefits offered by these methods (new agricultural practices, reforestation, etc.), there is a need in rapid, precise, and low-cost analytical tools. Near-infrared spectroscopy (NIRS) is now commonly used to measure different physical and chemical parameters of soils, including carbon content. However, prediction model accuracy is insufficient for NIRS to replace routine laboratory analysis and/or to make in situ measurements, whatever the type of soil. One of the biggest issues that need to be addressed concerns the calibration process: how does the mathematical method or the sample selection influence the model quality? In most cases, there are not a lot of thoughts put into the choice of the mathematical method, which is often made empirically (test and try). It is therefore essential to return to fundamental laws governing spectrum formation in order to optimize calibration. Indeed, the light/matter interactions are at the basis of the resulting linear modeling. This chapter reviews and discusses the basic theoretical concepts underpinning NIRS and linear chemometric modeling in the specific context of soil: (i) light scattering due to soil particles causes departure in the assumed linear relationship between the spectrum and the carbon content, and (ii) the other classical linear regression assumptions (constant residual variance, normal error distribution, etc.) are also put into question. Regarding these specific issues, the different chemometric methods presented as possible solutions to perform better calibration model are discussed, from linear methods associated with various preprocessing, local methods, or nonlinear methods.

Applicability of Vis-NIR hyperspectral imaging for monitoring wood moisture content (MC)

Abstract Visible-near-infrared hyperspectral imaging was tested for its suitability for monitoring the moisture content (MC) of wood samples during natural drying. Partial least-squares regression (PLSR) prediction of MC was performed on the basis of average reflectance spectra obtained from hyperspectral images. The validation showed high prediction accuracy. The results were compared concerning the PLSR prediction of MC mapping from raw spectra and standard normal variate (SNV) treatment. SNV pretreatment leads to the best results for visualizing the MC distribution in wood. Hyperspectral imaging has a high potential for monitoring the water distribution of wood.

Raman spectroscopy and multivariate analysis for the rapid discrimination between native-like and non-native states in freeze-dried protein formulations

This study investigates whether Raman spectroscopy combined with multivariate analysis (MVA) enables a rapid and direct differentiation between two classes of conformational states, i.e., native-like and non-native proteins, in freeze-dried formulations. A data set comprising of 99 spectra, both from native-like and various types of non-native freeze-dried protein formulations, was obtained by freeze-drying lactate dehydrogenase (LDH) as model protein under various conditions. Changes in the secondary structure in the solid freeze-dried proteins were determined through visual interpretation of the blank corrected second derivative amide I band in the ATR-FTIR spectra (further called FTIR spectra) and served as an independent reference to assign class labels. Exploratory analysis and supervised classification, using Principal Components Analysis (PCA) and Partial Least Squares - Linear Discriminant Analysis (PLS-LDA), respectively, revealed that Raman spectroscopy is with 95% accuracy able to correctly discriminate between native-like and non-native states in the tested freeze-dried LDH formulations. Backbone (i.e., amide III) and side chain sensitive spectral regions proved important for making the discrimination between both classes. As discrimination was not influenced by the spectral signals from the tested excipients, there was no need for blank corrections. The Raman model may allow direct and automated analysis of the investigated quality attribute, opening possibilities for a real time and in-line quality indication as a future step. However, the sensitivity of the method should be further investigated and where possible improved.