2.5.2. UVE Combined with Monte Carlo Sampling (MCUVE) and PLSR

The informative wavelengths were selected by UVE based on the regression coefficients of PLSR models. The Monte Carlo sampling method was used to randomly select *N* kinds of sample groups. The PLSR regression coefficient vector β(j,:) was obtained from the spectra and corresponding SCC vector of the jth group. The stability value C(k) at the kth wavelength was calculated by Equation (5). Wavelengths were sorted according to the values of vector C from the largest to the smallest. Evaluating models were established by adding new spectra of one wavelength, which had a smaller stability value. The wavelengths were selected as characteristic wavelengths with the minimum value of RMSEV.

$$\mathbf{C}(\mathbf{k}) = \frac{\text{mean}(\boldsymbol{\beta} \,\mathrm{(k,:)})}{\text{std}(\boldsymbol{\beta} \,\mathrm{(k,:)})} \tag{5}$$

where mean(β(k,:)) and std(β(k,:)) refer to the mean coefficient and standard deviation at the kth wavelength, respectively.

#### *2.6. Modeling Algorithm*

The radial basis function (RBF) is a good generalization of the kernel function of SVR. The particle swarm optimization algorithm (PSO) was used to determine the optimal combination of C and g [41] in order to obtain a model with good performance. The model was evaluated by RMSEV and R.
