Transfer of a calibration model from a benchtop to a handheld NIR spectrometer to predict nitrogen content of warm-season forage

Authors: RUKUNDO, ISAAC - University Of Nebraska; DANAO, MARY-GRACE - University Of Nebraska; Mitchell, Robert - Rob; WELLER, CURTIS - University Of Nebraska

Submitted to: Biosystems Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/12/2022
Publication Date: 5/6/2022
Citation: Rukundo, I., Danao, M., Mitchell, R., Weller, C. 2022. Transfer of a calibration model from a benchtop to a handheld NIR spectrometer to predict nitrogen content of warm-season forage. Biosystems Engineering. https://doi.org/10.1016/j.biosystemseng.2022.04.014.
DOI: https://doi.org/10.1016/j.biosystemseng.2022.04.014

Interpretive Summary: Portable or handheld near infrared (NIR) spectrometers are becoming increasingly available. It would be beneficial to transfer an existing calibration model for a benchtop instrument to a handheld instrument with minimal loss in prediction performance. In this study, a calibration model to predict N content of warm-season forage developed on a benchtop spectrometer was successfully transferred to another benchtop spectrometer and a transportable spectrometer with minimal loss in performance. Transferring the same calibration model to a smartphone spectrometer was successful, but had some loss in performance. The final transfer also required a different approach due to wider instrument variation. With the increasing availability of low-cost handheld NIR instruments on the market, users will benefit from using existing calibration models to predict spectra collected with the handheld device. This ability facilitates adopting a low-cost handheld NIR spectrometer, enabling the user to bring their lab analysis to the field instead of bringing their samples to the lab, at least for screening purposes.

Technical Abstract: With increasing availability of portable or handheld near infrared (NIR) spectrometers, it would be beneficial to transfer an existing calibration model for a benchtop instrument to a handheld instrument with minimal loss in prediction performance. In this study, a partial least squares (PLS) regression model developed to calibrate a benchtop NIR (B1) to estimate nitrogen content (N) of forage was transferred to another benchtop (B2) and two handheld NIR spectrometers (H1 and H2). B1, B2, and H1 had similar specifications, while H2 had a shorter spectral range with a wider interval. After each transfer, changes in the model’s performance were evaluated in terms of model parameters, including coefficient of determination of validation (r^2) and ratio of standard deviation to standard error of prediction (RPD). The PLS regression model developed for B1 had r^2 = 0.928 and RPD = 3.78. When transferred to B2, there was only 0.22% and 1.32% decrease in r^2 and RPD, respectively. Likewise, when Model B1 was transferred to H1, r^2 and RPD decreased by 1.62% and 10.3% decrease, respectively. Transferring the model to H2 required application of a second-derivative transformation to the spectra and yielded a model with lower prediction performance (r^2 = 0.571 and RPD = 1.6). The transferred model, however, was still useful for screening of samples based on r^2. The predicted N values were compared before and after calibration transfer using the Bland-Altman plot analysis. This analysis revealed a large bias (11.05 g/kg) between B1 and H2, but this could be addressed by a correction factor.