Abstract
On-line near infrared (NIR) spectroscopic analysis systems play an important role in assessing the quality of sugarcane in Australia. As quality measures are used to calculate the payment made to growers, it is imperative that NIR models are both accurate and robust. Machine learning and non-linear modelling approaches have been explored as methods for developing improved NIR models in a variety of industrial settings, yet there has been little research into their application to cane quality measures. The objective of this paper was to compare chemometric models of commercial cane sugar (CCS) based on four calibration techniques. CCS was estimated using partial least squares regression (PLS), support vector regression (SVR), artificial neural networks (ANNs) and gradient boosted trees (GBTs). Model performance was assessed on an independent validation data set using root mean square error of prediction (RMSEP) and r2 values. SVR (RMSEP = 0.37%; r2 = 0.92) and ANN (RMSEP = 0.36%; r2 = 0.93) performed similarly to PLS (RMSEP = 0.37%; r2 = 0.92) on the validation data set, while GBT exhibited a much lower skill (RMSEP = 0.51%; r2 = 0.85). Analysis of important wavelengths in each model showed that PLS regression, SVR and ANN techniques emphasized the importance of similar spectral regions. Future research should consider testing model robustness over seasons and/or regions. Comparisons of chemometric models should consider reporting variable importance as a way of understanding how models use spectral information.
© 2018 The Author(s)
PDF Article
More Like This
Cited By
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Contact your librarian or system administrator
or
Login to access Optica Member Subscription