The Importance of the Reference Method Accuracy and Why it is Important Not to Cut Corners While Developing a Model

Abstract

The author has spent the last three years as part of a team developing a handheld near infrared (NIR) fruit spectrometer for use across the agricultural supply chain. In this article, Ryan provides examples of calibrations and discusses some of the finer issues of good analytical technique to answer some of the most common questions asked by users of NIR spectroscopy in developing robust and reliable models.A s near infrared (NIR) spectroscopy moves from the laboratory to field-ready, handheld spectrometer units, the NIR community needs to develop simple examples and language suitable for farmers, QA technicians and other layperson operators. As I interact with people new to the use of NIR for agriculture, one question that always arises is: "how good a model can I make?". As practitioners of NIR, we know the answer to this question depends on many factors; the largest being the cost and time for the level of effort they are willing to put into the calibration process. Yet, we as a community have stock answers such as "the model can only be as good as the reference method", or "the model can be better than the reference method, if you keep only the well-predicted values". As accurate as these statements might be they are less than useful for someone without a background in NIR.To help customers visualise the importance of the reference method accuracy and why it is important not to cut corners while developing a model, we discuss the transflection sugar water data in Table 1. By taking a set of 18 spectra, shown in Figure 1, and inputting either the estimated sugar concentration or the actual sugar concentration into a partial least squares regression (PLSR), we can directly see the impact of error in the reference values. From the results shown in Figure 2, it becomes obvious that the error of estimation in the reference is carried over to the model. Specifically, that the root mean square error of cross-validation (RMSECV ), for the model built using the estimated sugar concentration is nearly double that of the model developed on the basis of the actual sugar concentration. Both models were built using the