Local Regression approaches for large NIRS databases based on PLS scores

ALLEGRINI FRANCO; OLIVIERI ALEJANDRO C.; FERNANDEZ PIERNA, JUAN ANTONIO; FRAGOSO WALLACE; BAETEN VINCENT; DARDENNE PIERRE

Barcelona

Congreso; XVI Chemometrics in Analytical Chemistry; 2016

Universitat de Barcelona

Extensive spectral databases compiling thousands of NIR reflectance spectra have been created during the last 30 years. Between these libraries, the ones related to typical parameters used in grains monitoring (like moisture, proteins, fat, starch and ash) are the most extended. In order to deal with this kind of databases, many chemometric tools have been developed, with Partial Least Squares (PLS) regression being the most popular because of its robustness, simplicity and efficiency.The approaches we consider here is based on local regression. This is a group of methods based on selecting from a large database, a set of samples spectrally similar to an unknown sample whose properties are to be predicted. Following this strategy, a specific local calibration is then developed for that sample using the previously selected neighbourhood samples as calibration set. This means that each sample is predicted with a different calibration equation. Up to now, the main types of local regression methods described in the literature are Locally Weighted Regression (LWR), the LOCAL algorithm, Comparison Analysis using Reconstructed Near Infrared and Constituent Data (CARNAC), and the formulation of complex indexes to measure distances between samples after the application of data reduction (Principal Component Analysis or Fast Fourier Transform).In this work, we present two new approaches to perform sample selection in LOCAL regression methods based on PLS scores. These approaches differ in the way they automatically fix and select a certain number of samples to optimally predict the unknown sample. Preliminary results obtained after applying these strategies over a large corn dataset, shows to improve predictions if compared with a standard PLS model using the complete set of samples. The proposed methodologies could be extended to the prediction of more than one product from a unique and large data sets, with the consequent savings in time and effort required development of individual calibration models.