Robustness of Partial Least-Squares Models to Change in Sample Temperature: I. A Comparison of Methods for Sucrose in Aqueous Solution

Abstract

Sample temperature is well known to impact model performance for prediction of chemical attributes in high-moisture-content samples when using short-wave near infrared spectroscopy. A number of methods proposed to reduce this effect were compared in this study. A short-wave near infrared spectroscopy system operating in transflectance geometry was used to record spectra of sucrose solutions (mean = 4.16% and SD = 5.8% w/v) at different temperatures. Partial least-squares regression models were developed using spectra of sucrose solutions collected at 15°C and tested on spectra of an independent set of sucrose solutions at a sample temperature of 35°C. As sample temperature impacts the water peak, the performance of a model of sucrose content is perturbed, mainly through an increase in bias. Addition of a relatively small number of spectra of the same set of samples at different temperatures facilitated a model robust to temperature, but continued addition of samples at 15°C beyond the ratio of 1:125 overwhelmed the compensation effect, resulting in a model that was not robust to temperature. The use of orthogonal scatter correction (OSC), generalised least square weighting (GLSW), external parameter orthogonalisation (EPO) and repeatability file were considered. Of these methods, OSC corrected bias but impacted bias corrected root mean square error of prediction ( SEP) and r2, EPO performed better but still with some bias, GLSW gave the best r2 and SEP result but still with bias, while use of the repeatability file method gave the best overall result.