Statistical tests for comparison of quantitative and qualitative models developed with near infrared spectral data

Near-infrared spectroscopy (NIRS) has been applied for both qualitative and quantitative evaluation of sugar beet. However, chemometrics methods are numerous and a choice criterion is sometime difficult to determine. In order to select the most accurate chemometrics method, statistical tests are developed. In the first part, quantitative models, which predict sucrose content of sugar beet, are compared. To realize a systematic study, 54 models are developed with different spectral pre-treatments (Standard Normal Variate (SNV), Detrending (D), first and second Derivative), different spectral ranges and different regression methods (Principal Component Regression (PCR), Partial Least Squares (PLS), Modified PLS (MPLS)). Analyze of variance and Fisher's tests are computed to compare respectively bias and Standard Error of Prediction Corrected for bias (SEP(C)). The model developed with full spectra pre-treated by SNV, second derivative and MPLS methods gives accurate results: bias is 0.008 and SEP(C) is 0.097 g of sucrose per 100 g of sample on a concentration range between 14 and 21 g/100 g. In the second part, McNemar's test is applied to compare the classification methods. The classifications are used with two data sets: the first data set concerns the disease resistance of sugar beet and the second deals with spectral differences between four spectrometers. The performances of four well-known classification methods are compared on the NIRS data: Linear Discriminant Analysis (LDA), K Nearest Neighbors method (KNN), Simple Modeling of Class Analogy (SIMCA) and Learning Vector Quantization neural network (LVQ) are computed. In this study, the most accurate method (SIMCA) has a prediction rate of 81.9% of good classification on the disease resistance determination and has 99.4% of good classification on the instrument data set. (C) 2003 Elsevier Science B.V. All rights reserved.