Transfer of Multivariate Classification Models between Laboratory and Process Near-Infrared Spectrometers for the Discrimination of Green Arabica and Robusta Coffee Beans

Abstract

Analogous to the situation found in calibration, a classification model constructed from spectra measured on one instrument may not be valid for prediction of class from spectra measured on a second instrument. In this paper, the transfer of multivariate classification models between laboratory and process near-infrared spectrometers is investigated for the discrimination of whole, green Coffea arabica (Arabica) and Coffea canefora (Robusta) coffee beans. A modified version of slope/bias correction, orthogonal signal correction trained on a vector of discrete class identities, and model updating were found to perform well in the preprocessing of data to permit the transfer of a classification model developed on data from one instrument to be used on another instrument. These techniques permitted development of robust models for the discrimination of green coffee beans on both spectrometers and resulted in misclassification errors for the transfer process in the range of 5-10%.

The vector of calibration ratios: A simple transfer method for mass spectra

“…Alternatives to local models may be the use of flexible multivariate regression models such as partial least squared regression (PLSR) [32], direct and piecewise direct standardization [28,35,36], neural networks [30], orthogonal least squares [27] and slope/bias corrections [26,27]. Multivariate calibration strategies offer more flexibility than the use of a single vector, because individual calibration ratios are calculated for each spectrum.…”

“…In the field of calibration this is commonly referred to as standardization of instruments [25]. Standardization and calibration transfers have been applied to correct for instrument differences in near infrared spectroscopy [25][26][27], nuclear magnetic resonance spectroscopy [28,29] pyrolysis mass spectra [30,31] and gas sensor arrays [32]. Because of different noise patterns, standardization and calibration methods are transferable only to a limited degree between these techniques.…”

Transfer of Multivariate Calibration Models

“…84 Both modified SBC and orthogonal signal correction (OSC) methods can be used to correct NIR spectra used in a classification model to allow transfer of the classification between a laboratory and a process instrument. Given the close similarity of regression and some classification models, it is not too surprising that some of the same considerations discussed above in the context of calibration models also apply to a classification model.…”

Practical comparison of sparse methods for classification of Arabica and Robusta coffee species using near infrared hyperspectral imaging

“…Classical point-wise NIR spectroscopy has been widely used to discriminate Arabica and Robusta species, both on green coffee [35,36] and roasted coffee [37,38]. Moreover some research works have been also reported where hyperspectral imaging is used to characterize these coffee species [39,40].…”