CovSel : Variable selection for highly multivariate and multi-response calibration.

J.M. Roger (1) , B. Palagos (1) , D. Bertrand (2)

1 UMR ITAP - Information – Technologies – Analyse Environnementale – Procédés Agricoles
2 StatSC - Statistique, Sensométrie et Chimiométrie

Variable selection is of major interest for NIR calibration, either as a feature selection or for the design of multi-wavelength devices. Some dedicated methods have been developed in chemometrics, but none of them addresses the case of multi response calibration. Variable selection for NIR spectroscopy must face two problems: (1) the huge number of variables yields a very large solution space; (2) the variables are highly correlated, and if no special attention is paid the model built on the selection may be inconsistent. In this article we present a new method, CovSel, which tackles these two problems by following this procedure: (1) Variable selection step by step on the basis of their global covariance with all the responses; (2) Projection of the data orthogonally to the selected variable. CovSel was applied on three problems: the first one concerns a one-response MIR calibration (Brix degree content in apricot), the second one concerns a multi-response NIR calibration (4 main constituents in corn) and the last application concerns the NIR discrimination of 3 wine grape varieties.

Mots clés: SPECTROSCOPIE; INFRAROUGE

Domaines: Sciences de l'environnement

J.M. Roger, B. Palagos, D. Bertrand. CovSel : Variable selection for highly multivariate and multi-response calibration.. The 14th International Conference on Near Infrared Spectroscopy (NIR-2009), Nov 2009, Bangkok, Thailand. p. - p. ⟨hal-00473261⟩