Evaluation of predictive performance of PLS regression models after being transferred from benchtop to handheld NIR spectrometers

With increasing availability of handheld near infrared (NIR) spectrometers, it would be beneficial to transfer an existing calibration model for a benchtop NIR to another spectrometer with minimal loss in prediction performance. In this study, a benchtop FOSS XDS Rapid Content Analyzer (B1) was calibrated to estimate nitrogen content (N) of warmseason grasses using partial least squares (PLS) regression. This model had a coefficient of determination of validation (r2) of 0.928 and ratio of standard deviation to standard error of prediction (RPD) of 3.78. The calibration model was transferred to another benchtop, FOSS 6500 (B2) and two handheld spectrometers, ASD QualitySpec (R) Trek and Tellspec Enterprise (H1 and H2, respectively). B1, B2, and H1 had similar specifications, while H2 had a shorter spectral range with a wider spectral interval. When the calibration model for B1 was transferred to B2 and H1, both r2 and RPD decreased slightly. However, transferring the model to H2 required truncating and standardizing the B1 spectra using piecewise reverse standardization and more preprocessing than was done for B2 and H1. The calibration model developed for the standardised B1 spectra had an r2 = 0.889 and RPD = 1.79, which were lower than those for the original calibration model developed for B1. When this new model was transferred and applied to H2, r2 decreased further to 0.726 while RPD remained 1.79. Based on these values, the model to H2 spectrometer can be used for rough screening purposes.

Automated summary

This study explores the feasibility of transferring a calibration model for nitrogen content estimation in warmseason grasses from one benchtop near infrared (NIR) spectrometer to another. The results show that the calibration model can be successfully applied to different spectrometers, although there may be a slight decrease in prediction performance during the transfer process.