Iterative Interval Backward Selection Algorithm and Its Application in Calibration Transfer of Near Infrared Spectra
Zheng Kai-yi,Feng Yu-hang,Zhang Wen,Huang Xiao-wei,Li Zhi-hua,Zhang Di,Shi Ji-yong,Zou Xiao-bo
DOI: https://doi.org/10.3964/j.issn.1000-0593(2021)06-1789-06
2021-01-01
Spectroscopy and spectral analysis
Abstract:The near-infrared spectra (NIR) with advantages of fastness, non-destructiveness and easy operation have been widely used in food analysis. As an indirect analysis method, NIR should calibrate the model between spectra and concentrations for analysis. Thus, the maintenance of the model can ensure high accuracy. The changes of external conditions, including the changes of samples characters, the variations of functions between physical and chemical indicators and the changes of the environment such as humidity and temperature, can diverge the spectra of the same samples and then decrease the prediction accuracy of the original model. To solve this problem, recalibration can eliminate the chances of spectra butcost huge laborious and economic expense. Thus, calibration transfer can correct the spectral divergence and improve model prediction accuracy without the expense of recalibration. In previous work, the calibration transfer algorithms usually use full spectra variables to transfer, which increase computation burden and not find spectra intervals with chemical information. Thus, this paper proposed a variable selection method called iterative interval backward selection (JIBS) for calibration transfer. IIBS firstly calculates the importance vectors of variable intervals in spectra, including regression coefficients (beta), residual errors (Res) and VIP (VIP) vectors. Then set the geometric mean of the important values of variables in each interval as the corresponding interval' s importance. Moreover, based on the importance values of intervals, remove the smallest one. After that, repeat the above procedure iteratively for both primary and secondary spectra, including computing the importance and values of variables and intervals and remove the intervals with minimal importance value. Finally, compute the root mean squared error of validation (RMSEV) for each interval subsets combination of both primary and secondary spectra and choose the intervals combination with minimal RMSEV as the best one. Two datasets, including corn and wheat datasets, were executed to test this algorithm. The results show that compared with the spectra with full intervals, the beta, Res and VIP can select fewer but more important variable intervals from whole spectra to improve the calibration transfer accuracy. In contrast with different variable importance vectors, the beta can select variables intervals with low prediction errors.