Robustness of NIR calibrations for assessing fruit quality

Abstract

Predictive models based on near infra-red spectroscopy for the assessment of fruit internal quality attributes must exhibit a degree of robustness across the parameters of variety, district and time to be of practical use in fruit grading. At the time this thesis was initiated, while there were a number of published reports on the development of near infra-red based calibration models for the assessment of internal quality attributes of intact fruit, there were no reports of the reliability ("robustness") of such models across time, cultivars or growing regions. As existing published reports varied in instrumentation employed, a re-analysis of existing data was not possible.

An instrument platform, based on partial transmittance optics, a halogen light source and (Zeiss MMS 1) detector operating in the short wavelength near infra-red region was developed for use in the assessment of intact fruit. This platform was used to assess populations of macadamia kernels, melons and mandarin fruit for total soluble solids, dry matter and oil concentration. Calibration procedures were optimised and robustness assessed across growing areas, time of harvest, season and variety. In general, global modified partial least squares regression (MPLS) calibration models based on derivatised absorbance data were better than either multiple linear regression or `local' MPLS models in the prediction of independent validation populations . Robustness was most affected by growing season, relative to the growing district or variety . Various calibration updating procedures were evaluated in terms of calibration robustness. Random selection of samples from the validation population for addition to the calibration population was equivalent to or better than other methods of sample addition (methods based on the Mahalanobis distance of samples from either the centroid of the population or neighbourhood samples). In these exercises the global Mahalanobis distance (GH) was calculated using the scores and loadings from the calibration population on the independent validation population. In practice, it is recommended that model predictive performance be monitored in terms of predicted sample GH, with model updating using as few as 10 samples from the new population undertaken when the average GH value exceeds 1 .0 .