A Hybrid Multivariate Calibration Optimization Method for Visible Near Infrared Spectral Analysis

A modified simple-to-use interactive self-modeling mixture analysis (SIMPLISMA) combined with successive projections algorithm (SPA) is proposed for robust optimization of calibration model. It is not as SIMPLISMA do originally for pure variables selection, the modified SIMPLISMA is for feature samples selection, and getting the calibration sample set with minimum colinear. At meanwhile, SPA is introduced for informative wavelength variables selection. Under the determination algorithm of SIMPLISMA-SPA, the calibration model is optimized with selected samples and variables simultaneously. Two visible near infrared (VIS-NIR) spectra analysis experiments (distilled water pH detection experiment by spectrophotometer and distilled water pH detection experiment by grating spectrograph) are employed to evaluate the performance of SIMPLISMA-SPA. The results indicate that, for the experiment with spectrophotometer, under the SIMPLISMA-SPA, 21 samples and 18 variables are selected from the original 34 samples and 351 variables for constructing multiple linear regression (MLR) model, the root mean square errors (RMSE) of cross validation is 0.047, it is higher than the model of sample set partitioning based on joint x-y distance combined with successive projections algorithm (SPXY-SPA); for the experiment with grating spectrograph, under the SIMPLISMA-SPA, 34 samples and 33 variables are selected from the original 60 samples and 2860 variables for constructing MLR model, the RMSE of cross validation is 0.008, the performance is similar to the MLR model with SPXY-SPA. The conclusions show that, SIMPLISMA-SPA can be as an alternative multivariate calibration model optimization method for VIS-NIR spectral analysis.