SO UTH AFRICAN AVO CADO G RO WERS’ ASSO CIATIO N YEARBO O K 36, 2013 33 ABSTRACT This project continued from 2011 and it was found that by including fruit from the 2012 season, the calibration of the handheld near-infrared spectrometer (NIR) was made more robust, but this did increase the standard error of prediction to 3.3% MC (R 2= 72%). This project will conclude in 2013, with the addition of the third full season’s data. The effect of the skin will also be examined, to see if the accuracy of the handheld NIR can be improved by removing the skin – although this means the NIR is no longer non-destructive, and is just a tool for the rapid analysis of moisture content. Calibration of handheld NIR to determine avocado maturity – Progress report R Blakey Westfalia Technological Services PO Box 1103, Tzaneen 0850, South Africa E-mail: robert.blakey@westfalia.co.za INTRODUCTION This progress report follows on from the previous re- ports (Blakey & Van Rooyen, 2011; Blakey, 2012), where it was found that with a range of 50 to 87% moisture content (MC), the Root Mean Squares Error of Prediction (RMSEP) = 2.8% and R 2 = 78%. With a narrower range of 70 to 87% MC, which is more criti- cal for avocado fruit quality and maturity, the RMSEP = 2.3% MC and R 2 = 74%. The aim of this project is to develop robust cali- bration models for the handheld NIR for ‘Hass’ and ‘Fuerte’ that have a commercially useful Standard Er- ror of Prediction (SEP). MATERIALS AND METHODS Instrument: Fruit were again scanned at four to six locations around the equator using the Phazir 1018 handheld near-infrared spectrometer (NIR; Thermo Scientific, Wilmington, MA, USA). Further details about the instrument are available in the previous reports. Fruit: Twelve orchards (‘Hass’ and ‘Fuerte’) on vari- ous Westfalia farms were sampled monthly (12 to 15 fruit) from January 2012 until harvest – usually in May. Further samples were taken from the Westfalia pack house until mid-September. Sampling: The area to be sampled was scanned with the skin (exocarp) intact and then again with the skin removed. Flesh samples of approximately 1.0 g were taken from the same areas that were scanned with the handheld NIR. The samples were then oven-dried at 70°C for at least 24 hours. Analysis: The chemometric analysis was done us- ing Polychromix Method Generator™ version 3.101. The 2012 data set was randomly and equally divided between the calibration and external validation set, while the 2011 data set was used in the calibration. Various pre-treatments were tested and it was found that a Savitzky-Golay (SG) derivative and a Standard Normal Variate (SNV) transformation were the most suitable pre-processing treatments. The SG deriva- tive was a five point smoothing with a second order derivative with a third order polynomial smoothing. Spectra with a Mahalanobis distance greater than 3.0 were deemed as outliers and were mostly because of poor spectral quality. RESULTS AND DISCUSSION Calibration: By including fruit from 2012 in the models, the robustness of the models for estimat- ing MC increased, as (external validation) fruit from the 2012 season were predicted with equal accuracy (Standard Error of Prediction; SEP) as fruit from 2011 (Fig. 1). However, the SEP increased to 3.3% MC. Optimally this would be lower than 1% MC, but this is unlikely with this handheld NIR. It is more likely that an SEP of approximately 2.5% MC is achievable. This is higher because the spectrometer in the hand- held NIR has a resolution of 11 nm, while high-end spectrometers – such as the Matrix-F FT-NIR (Bruk- er Optics, Ettlingen, Germany) instrument used by