In this paper, we propose a novel approach for the estimation of the concentration of chemical components through spectrophotometric measurements. It is based on the exploitation of the whole spectral information available in the original spectral data space by means of a Multiple Regression System (MRS) whose design is performed in three successive steps. The first one aims at a simple partitioning of the original spectral data space into subspaces of reduced dimensionality. The second step consists in training a (linear or nonlinear) regression method in each of the subspaces obtained in the previous step. In the third and final step, the estimates provided by the ensemble of regressors are combined in order to produce a global estimate of the concentration of the chemical component of interest. For this purpose, two linear and one nonlinear combination strategies are explored.

The experimental assessment of the MRS was carried out on two different datasets: 1) a wine dataset for the determination of alcohol concentration by mid-infrared spectroscopy; and 2) an orange juice dataset where near-infrared reflectance spectroscopy is used to estimate the saccharose concentration. The obtained results suggest that the proposed MRS approach represents a promising alternative to the traditional regression methods.

Spectrophotometric data often comprise a great number of numerical components or variables that can be used in calibration models. When a large number of such variables are incorporated into a particular model, many difficulties arise, and it is often necessary to reduce the number of spectral variables. This paper proposes an incremental (Forward–Backward) procedure, initiated using an entropy-based criterion (mutual information), to choose the first variable. The advantages of the method are discussed; results in quantitative chemical analysis by spectrophotometry show the improvements obtained with respect to traditional and nonlinear calibration models.

Data from spectrophotometers form spectra that are sets of a great number of exploitable variables in quantitative chemical analysis; calibration models using chemometric methods must be established to exploit these variables. In order to design these calibration models which are specific to each analyzed parameter, it is advisable to select a reduced number of spectral variables. This paper presents a new incremental method (step by step) for the selection of spectral variables, using linear regression or neural networks, and based on an objective validation (external) of the calibration model; this validation is carried out on data that are independent from those used during calibration. The advantages of the method are discussed and highlighted, in comparison to the current calibration methods used in quantitative chemical analysis by spectrophotometry.

RBFN (Radial-Basis Function Networks) represent an attractive alternative to other neural network models. Their learning is usually split into an unsupervised part, where center and widths of the basis functions are set, and a linear supervised part for weight computation. Although available literature on RBFN learning widely covers how basis function centers and weights must be set, little effort has been devoted to the learning of basis function widths. This paper addresses this topic: it shows the importance of a proper choice of basis function widths, and how inadequate values can dramatically influence the approximation performances of the RBFN. It also suggests a one-dimensional searching procedure as a compromise between an exhaustive search on all basis function widths, and a non-optimal a priori choice.