An efficient wavelength selection method based on the maximal information coefficient for multivariate spectral calibration

Spectral data on the modern spectroscopic instrument are commonly of high co-linearity and contain a large number of spectral variables, which may cause the poor predictive performance of the developed model. To address this problem, a novel method for wavelength selection, named maximal information coefficient screening combined with PLS regression (MICPLS), is proposed. MIC can capture a wide range of relationships between feature variables and target variable, including both functional and non-functional relationships. By employing the simple but effective MIC screening, MICPLS can rapidly and efficiently select strongly correlated spectral variables related to the response. The performance of MICPLS is investigated with two real spectroscopic datasets. Compared with commonly used methods for variable selection based on informative criterions including variable importance in projection (VIP) and selectivity ratio (SR), the results show the proposed method can improve the prediction performance, which is a good and promising strategy for variable selection in spectral analysis.