PLS pruning: a new approach to variable selection for multivariate calibration based on Hessian matrix of errors

In this article, a new approach called partial least squares (PLS) pruning is described for variable selection in PLS modeling. The aim of the method is the deletion of unimportant PLS coefficients of regression by using information from all second derivatives of the error function. The proposed approach was applied to Brix determination in sugar cane juice by near infrared spectroscopy. The results obtained were promising, leading to a meaningful variable reduction of 96% without loss of model prediction capability.