Spectral Reconstruction and Quantitative Analysis by B-Spline Transformations and Penalized Partial Least Squares Approach / 分析化学

ABSTRACT

Taking into account the near infrared spectra(NIR) on numerous predictor variables with serious collinearity and having nonlinear quantitative relationship with the chemical compositions, a novel nonlinear partial least squares(PLS) approach, termed as Spline-PPLS, was constructed by combining the penalized partial least squares(PPLS) regression with B-splines transformation. Firstly, the observed spectral predictors were considered as discrete observations of curves of the wavelength and were nonlinearly transformed using B-spline basis functions. The choice of the degree of the polynomial pieces and of the number of knots was performed using the cross-validation strategy. Then, the PPLS algorithm was performed on the high dimensional transformed data matrix to build the calibration model by imposing a penalty term to the optimization criterion of PLS. The roughness penalty term indeed controlled the curvature of the functions and its smoothing parameter could also be obtained by the cross-validation. Finally, the proposed Spline-PPLS approach was applied to the wheat NIR diffuse reflectance spectra reconstruction and quantitative analysis. The result indicates that the Spline-PPLS approach not only can yield high accuracy reconstructing spectrum, but also improves the model prediction accuracy in the case of nonlinear relationships.