1. Standardization of NIR Instruments: How Useful Are the Existing Techniques? Benoit Igne (igneb@iastate.edu), Glen R. Rippke (rippke@iastate.edu), Charles R. Hurburgh, Jr. (tatry@iastate.edu) Department of Agricultural and Biosystems Engineering, Iowa State University, Ames, Iowa. Introduction Objectives Materials and Methods Conclusions Results The development of a near infrared (NIR) prediction model is a long and costly process. Often, a calibration is developed on a master unit and transferred to secondary or slave units using standardization techniques. Those techniques have shown their usefulness when employed within the same instrument brand. Increasingly, network managers would like to be able to add to their network instruments from different manufacturers. This requires chemometrics experts to develop techniques that are either able to combine historical databases from both brands or to transfer a model from one brand to another. Evaluate the usefulness of existing standardization techniques within and among instrument brands for soybeans protein and oil prediction models. Data collection Samples: Calibration set: 638 samples from 2002 to 2006 crop years. Validation sets: Set 1: 20 samples representative of the variability of the calibration set. Set 2: 40 very diverse samples from the 2006 crop year. Spectral data 4 transmittance units, spectral range: 850 – 1048 nm with 2 nm increment. 2 Foss Infratec™ (Foss North America, Eden Prairie, MN): Foss Infratec™ 1229 and Foss Infratec 1241™. 2 Dickey-john OmegAnalyzer G (Dickey-john ® corporation, Auburn, IL). Reference analysis: Protein content by Combustion (AOAC 990.03), Eurofins, Des Moines. Oil content by ether extract (AOCS Ac 3-44), Eurofins, Des Moines. Calibration method Partial Least Squares Regression. Data Preprocessing Second derivative (5-point window) + normalize (to unit area). Standardization techniques Optical techniques: Direct Standardization (DS) and Piecewise Direct Standardization (PDS). DS: Reconstruct slave unit’s spectra to match the master unit based on the whole slave unit’s spectrum. PDS:Reconstruct slave unit’s spectra to match the master unit based on a small size window on the slave unit’s spectrum. Slope and/or offset/bias post regression correction: Correction to slave unit predictions based on comparison between a standardization set prediction on slave units and the reference. Robust calibration: All available spectral data from various units (same of various brands) are combined to form the calibration set. Model comparison Relative Predictive Determinant (RPD) was used to determine model precision. RPD is a standardized parameter allowing an easy comparison between models validated on the same set. Initial calibration performances Each instrument was calibrated on its own database: Protein (RPD) Infratec 1229Infratec 1241OmegAnalyzer G 6110OmegAnalyzer G 6118 Set 19.7310.418.9010.78 Set 23.623.933.734.01 Oil (RPD) Set 14.285.744.935.36 Set 23.894.605.124.80 Use of existing standardization techniques: within brands -DS and PDS gave significantly lower RPDs. -Standardization techniques gave either as good as or better RPDs than when slaves units were calibrated on their original calibration sets. -Slope and Bias techniques appeared to give higher RPDs for both master and slave units. Use of existing standardization techniques: across brands Foss Infratec 1241 master of the network: OmegAnalyzer G 6118 master of the network: Infratec 1241 and OmegAnalyzer G 6118 were selected as masters. -PDS and DS gave significantly lower RPD for among brand standardization. -Other techniques were not significantly different from when each unit was calibrated on its own calibration set. -Master units gave significantly higher results than slave units of the same brand, no matter the network’s master. -Intermediate parameters for optical standardization (DS, PDS) increased error significantly. -Differences among standardization methods were proven consistent across standardization sets, across instrument brands and across standardization strategies. -For within brand standardization, slope and bias correction appeared to give better results while for across brand, robust techniques were more suitable.