Selection of calibration sub-sets to predict ryegrass quality using principle component analysis for near infrared spectroscopy

Abstract

Near infrared reflectance spectroscopy (NIRS) has become the routine method of assessing forage quality on grass evaluation and breeding programmes. NIRS requires predictive calibration models that relate spectral data to reference values developed using a calibration set (Burns et al. 2013). The samples that form the calibration set influence the accuracy and reliability of these models and need to be representative of samples that will likely be analysed (Shenk and Westerhaus, 1991; Burns et al. 2014). Analysing samples from the calibration set using reference techniques has a significant cost and time associated and needs to be considered in the context of the desired accuracy and robustness of calibration models. Calibration selection techniques can therefore maximise the accuracy and robustness of calibration models whilst reducing the number of samples requiring reference analysis. One such method is principal component analysis (PCA; Shenk and Westerhaus, 1991) whereby Shetty et al. (2012) reported that the number of samples could be reduced by up to 80% with a minimal loss in accuracy of calibration model. PCA selects representative calibration sub-sets through plotting all the samples in hyper-dimensional space, based on spectral data, and a sample is selected to represent a local neighbourhood cluster of samples for reference analysis. The aim of this research was to assess the accuracy of NIRS calibration models for buffering capacity, in vitro dry matter digestibility (DMD) and water soluble carbohydrate (WSC) content developed using calibration sub-sets selected by PCA.