Combined Multi-Instrument NIRS Calibration: A Case Study

Between instrument difference is a major hurdle to transferability of NIR equations. Building robust calibrations by including spectra from instruments of the same type are often used as a solution to this problem.
In this work, we have tried to demonstrate feasibility of extending this situation to a case where NIR spectrometers of different types and/or vendors are involved and also tried to compare the performance of global vs. Local modelling approaches in this context. A multi-instrument calibration based on large datasets especially when combined with powerful non-linear methods such as Local regression can go a long way towards achieving truly universal NIR equations.
Datasets from four different NIR spectrometers (Buchi NIRFlex, Bruker MPA, Foss NIRS6500 and Thermo Antaris)- spanning three FT types and one dispersive type - were merged to produce a large dataset comprising over 30,000 cereal samples from INGOT database (AuNIR). All spectra were measured in reflection mode covering wavelength range 1100-2500 nm. About 6,000 samples covering all four instrument types uniformly were randomly selected from the above pool for the purpose of independent validation. Moisture, protein and starch were chosen as calibration parameters for this study. Global PLS and Local regressions (WinISI software, FOSS Analytical) were employed for the chemometrics modelling work.
Initially individual PLS models were made for individual instruments and their average predictive performance (average RMSEPs) for moisture, protein and starch were found to be 0.46, 0.39 and 1.58 respectively when run against individual instrument validation sets. The combined multi-instrument calibration provided acceptable RMSEP figures of 0.55, 0.50 and 1.79 respectively when run against the combined validation set.
The best combined performance was achieved after performing a Local calibration on the combined datasets (RMEPs= 0.41, 0.37 and 1.65). These figures were found to be comparable to dedicated Local model performance built for individual instruments.