An effective method for determining the hyperparameter of spectral space transformation models with few standardization samples in near infrared spectroscopic analysis

Abstract

Spectral space transformation (SST) is one of the most popular methods for eliminating spectral differences induced by the changes in instruments or measurement conditions in field of in-situ analysis of complex systems using near infrared spectroscopy. The implementation of SST involves the tuning of a hyperparameter value, which generally requires the collection of a sufficient number of standard samples. However, in practice, the preparation of large number of standardization samples and their long-term storage poses many problems for the users of SST, and greatly limits its wider application. For situations where only few standardization samples are available for SST model building, there is no effective method for determining the hyperparameter value of SST. To meet this practical need, a simple but effective method was developed in this contribution. In the proposed method, the optimal hyperparameter value of SST was determined by scrutinizing four correlation coefficients among some loading vector pairs as well as some score vector pairs. The performance of the method has been tested on three NIR data sets. The experimental results show that the SST models with the hyperparameter values recommended by the proposed method could effectively correct the spectral variations induced by changes in instruments. It can be expected that with the help of the proposed method, SST will have a wider application prospect.