Uma estratégia de validação híbrida para calibração multivariada baseada na seleção de amostras fixadas pelo algoritmo SPXY

Abstract

In multivariate analysis, it is usual to come across the recurring problem of partitioning the dataset in order to obtain, for the calibration subset, the most representative samples that cover the boundary of the multidimensional sample space.The present work aimed to develop a hybrid validation strategy for multivariate calibration (such as MLR-APS and PLS group-out), which prevents extrapolation problems and provides models with greater predictive capacity and robustness.The proposed strategy uses the SPXY algorithm to select the experimental space boundary samples for calibration and which are more representative for exploring the statistics of X (instrumental responses) and Y (parameter of interest). To assess its performance, two sets of NIR reflectance data were used.The first involves the analysis of wheat samples in which the protein content has been determined; the second refers to the determination of the moisture content in corn. In the hybrid validation applied to the PLS grou-out modeling, it was not possible to assess the variability of the RMSEP as a function of the sample index. This is because it did not present a significant variability of results for 20 executions, making it impossible to better evaluate the proposed strategy. In the MLR-APS modeling, a variability was observed in terms of RMSECV and RMSEP, making it possible to assess the influence of fixed samples on the predictive capacity of the models.The models resulting from the fixation of frontier samples in the calibration showed the highest correlation coefficients, which were equal to, respectively, 0.9996 and 0.9934 for the corn and wheat dataset. The values of RMSEP and RMSECV for the two sets showed a significant decrease.In fact, the values of 0.194 and 0.163 % (m/m) were obtained, respectively, for the protein content in wheat and 0.0121 and 0.0061 % (m/m) for moisture in the corn samples. A smaller number of variables was also obtained. The hybrid validation strategy is a viable alternative for multivariate calibration, providing more parsimonious models with greater robustness and predictive capacity.