Analytica Chimica Acta 555 (2006) 286–291 Robustness of calibration models based on near infrared spectroscopy for the in-line grading of stonefruit for total soluble solids content M. Golic ∗ , K.B. Walsh Plant Sciences Group, Central Queensland University, Bruce Highway, N. Rockhampton, Qld 4702, Australia Received 18 March 2005; received in revised form 2 August 2005; accepted 9 September 2005 Available online 19 October 2005 Abstract The utility of near infrared spectroscopy as a non-invasive technique for the assessment of internal eating quality parameters of stonefruit (peaches, nectarines and plums) was assessed. Calibration model performance for the attributes of total soluble solids (TSS) was encouraging (typical R 2 > 0.88, RMSECV 0.53–0.88%TSS, SDR CV 2.9–3.7). Model performance was acceptable using a combined multi-variety peach–nectarine data set, but it was advantageous to maintain a separate multi-variety plum model. Model robustness to temperature was achieved by including into the calibration set samples scanned at a range of temperatures, with less than 5% of total population required to be treated in this way. Similarly, where models incorporated the range of TSS seen in the validation population, prediction performance was good. Model performance was stable over several seasons in terms of R 2 (typical R 2 > 0.8), with bias corrected SEP varying in proportion to population S.D. Prediction bias for new populations could be corrected by model updating or direct bias adjustment. © 2005 Elsevier B.V. All rights reserved. Keywords: NIR; Stonefruit; Peach; Nectarine; Plum; Total soluble solids content; Robustness 1. Introduction Fruit can be non-invasively assessed for internal quality attributes, such as total soluble solids content (TSS) through correlation with their near infrared (NIR) spectra (e.g. Kawano [1,2]). Partial least-squares (PLS), multiple linear regression (MLR) or other multivariate techniques are typically employed in this process. Many variables, for example temperature, geo- graphic region, picking time, cultivar, data pre-treatment and model algorithm, can affect the performance of a predictive model [3–6]. It is therefore expected that the model prediction statistics for a truly independent population will be poorer than the calibration statistics (i.e. RMSEP > RMSECV, |bias| > 0; where RMSEP is the root mean standard error of prediction, RMSECV is the root mean standard error of cross-validation and bias is the average difference between predicted and actual values). ∗ Corresponding author at: P.O. Box 8124, Allenstown, 4700, Australia. Tel.: +61 7 4939 7315; fax: +61 7 4930 6536. E-mail address: golicm@iinet.net.au (M. Golic). The RMSEP is an estimate of total prediction error for an independent data set. RMSECV describes total error for samples within the calibration data set. Leave-one-out cross-validation method (used in this work) involves prediction of each sample using a model created on all other calibration data set samples. Standard error of calibration (SEC) is a standard deviation of the fit residuals. Bias (a systematic part of prediction error) between the actual and predicted value is a common component of the increased prediction error. This source of error contributes to the RMSEP value. Sources of error other than bias are estimated by the bias-corrected standard error of prediction (SEP). The relationship between RMSEP, SEP and bias parameters can be described as: RMSEP 2 = SEP 2 + bias 2 (1) SEP and R 2 (squared correlation coefficient) are related through population standard deviation (S.D.) as: R 2 = 1 - (SEP/S.D.) 2 (2) The SDR statistic (Eq. (3a and 3b)(3)) enables comparison of model performances across the populations with different 0003-2670/$ – see front matter © 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.aca.2005.09.014