Averaging a local PLSR pipeline to predict chemical compositions and nutritive values of forages and feed from spectral near infrared data

Abstract

Partial least squares regression (PLSR) is a reference method in chemometrics. In agronomy, it is used for instance to predict components of chemical composition (response variables y) of vegetal materials from spectral near infrared (NIR) data X collected from spectrometers. The principle of PLSR is to reduce the dimension of the spectral data X by computing vectors that are then used as latent variables (LVs) in a multiple linear model. A difficulty is to determine the relevant dimensionality (number of LVs) of the model for the given available data. This step can also become time consuming when many different datasets have to be processed and/or the datasets are frequently updated. An alternative to determinate the relevant PLSR dimensionality is the ensemble learning method "PLSR averaging". In the past, this method has been demonstrated to be efficient for complex biological materials such as mixed forages, and facilitates to automatize predictions (e.g. in user-friendly web interface platforms). This article presents the extension of the PLSR averaging to a k-nearest neighbors locally weighted PLSR pipeline (kNN-LWPLSR). The kNN-LWPLSR pipeline has the advantage to account for non-linearity between X and y existing for instance in heterogeneous data (e.g. mixing of vegetal species, collection from different geographical areas, etc.). In the article, kNN-LWPLSR averaging is applied to an extensive NIR database built to predict the chemical composition of European and tropical forages and feed. The main finding of the study was the overall superiority of the averaging compared to the usual kNN-LWPLSR. Averaging may therefore be recommended in local PLSR pipelines to predict NIR forage and feed data.