For sample diversity, three types of samples were collected, which included fresh tea leaves, manufactured green tea and partially processed green tea. The total number of samples was 738. The general information of samples was summarized in Table 1. Hereinto, the fresh leaves of type I were picked from five varieties of tea plants, and these samples were comprised of different tenderness leaves including young shoot, mature leaves and senescent leaves. The detailed information of samples in type I is shown in Table 2. Type II contained Xi-hu-long-jing tea of seven grades, and their detailed description is given in Table 3. Type III included eight kinds of partially processed green tea from eight processing procedures, as shown in Table 4.

In modeling, all 738 samples were divided into the calibration set and the prediction set with a ratio of 2:1. To avoid bias in subset partition, all samples were first arranged in an ascending order according to their respective MC values, then one sample was picked out from every three samples consecutively, resulting in 246 samples of prediction set, and the remaining 492 samples formed calibration set. The statistical information of Y-value of each set was shown in Table 5.

In this study, a Vis/NIR spectroradiometer (FieldSpec®3, Analytical Spectral Devices, Inc., Boulder, CO, USA) was adopted for Vis/NIR spectroscopy acquisition. This spectroradiometer has high sensitivity in the range of 325–1,075 nm with a 512 photodiode array detector, while the field-of-view is 10°, the spectral resolution is 3.5 nm, and the interval of sampling is 1.5 nm. A 150 watt halogen lamp was used to provide uniform light in the visible and short-wave near infrared range. When scanning spectrum, the spectroradiometer was fixed on a tripod with 45° between the spectroradiometer axis and horizontal line, and fixed at approximately 100 mm above samples. After each sample was scanned, it was taken away to empty the position for the next sample, this movement might lead to a change in the measurement system. In order to reduce this influence, the spectroradiometer was calibrated every half hour by a 100-mm² white standard panel with approximately 100% reflectance across the entire spectrum. So, relative reflectance was calculated with measurements from both the samples and the standard panel as shown in Figure 1. With respect to each sample, a mean spectrum was averaged by 30 scans. Besides, there were obvious noises at the beginning and the end of the spectrum, so only spectral bands of 400–1,050 nm were taken for further analysis.

The reference MC was measured by the gravimetric method according to the Chinese National Standard GB8304-87. In detail, every sample was heated in a constant temperature oven at 103 °C for 4 h, and weighed before and after the heating by an electronic balance with an accuracy of 0.0001 g. All the measurements were carried out in a room at approximate constant temperature of 25 °C and relative humidity of 40–55%.

The quality of the regression model was quantified by root mean squared error of calibration (RMSEC), root mean squared error of prediction (RMSEP), and the correlation coefficient (r) between the predicted and measured parameters [27]. A good model should have a low RMSEC, a low RMSEP, a high r, and a small difference between RMSEC and RMSEP [14].

Vis/NIR diffuse reflectance spectra of the three types of samples are shown in Figure 1. Similar contours were seen for all three types of samples. An obvious absorption peak was detected at 680 nm which was caused by the intense absorptivity of chlorophyll in the red light range. After 680 nm, the absorbance sharply declined as the wavelength increased from 680 nm to 750 nm. Then the absorbance was flat and low throughout the whole near infrared region. It could be found that the tea samples mainly absorbed the visible light in the range of 400–680 nm, especially at 680 nm. This phenomenon may be caused by the strong absorption of pigments in tea samples, while the absorptions of near infrared light (750–1,050 nm) were relative lower.

Except of the above similarities, many differences also existed in the spectra among the three types of samples. Comparing Figure 1(a) with Figure 1(b), there were many different absorptive responses within the range from 540 nm to 640 nm. In detail, two small absorption peaks were detected at 540 nm and 610 nm in Figure 1(b), but these absorptive responses did not exist in Figure 1(a). This phenomenon might be caused by the color change along with the variation of MC between type I and type II. The MCs of samples in type I were all bigger than 50%, while those in type II didn't exceed 7%, as shown in Tables 2 and 3. The big variation of MC caused by heating and drying led to huge concentration changes of chromogenic compositions in tea leaves. Former researchers have found that the chlorophyll a and chlorophyll b gradually degrade, and the contents of pheophytin a and pheophytin b increase in manufacturing process [28]. In type III, samples came from eight kinds of processing procedures, and the MCs were distributed in a broad range from 3.7% to 67% as shown in Table 4, so those curves were dispersing in Figure 1(c).

Multi-signal wavelet decomposition was realized to expose the internal structure of all the spectral data of the 738 samples. After WT, the spectrum of each sample was decomposed to four sets of wavelet coefficients, including approximation coefficients cA₃ and detail coefficients cD₁, cD₂, cD₃ as shown in Figure 3. It could be found that cA₃ had the same trend with the original spectra, and it was very similar to the original spectra. While cD₁, cD₂, cD₃ contained much high-frequency information, especially in the beginning. In order to evaluate the information contained in the four sets of wavelet coefficients in this decomposition, the percentages of energy of the four sets of wavelet coefficients were calculated. And the energy percentages of wavelet coefficients for all the 738 samples were plotted in Figure 4.

Figure 4 shows the energy distribution of the wavelet coefficients including cA₃, cD₁, cD₂, and cD₃, where it can be seen that the energy percentages of the cD₁, cD₂ and cD₃ are very close to zero, while the wavelet coefficients of cA₃ correspond to most of spectral energy. Furthermore, Figure 4(b) shows the energy distribution of cD₁, cD₂, and cD₃ in detail. It can be seen that their percentages of energy are very small, and there are relatively high-energy coefficients at the beginning of the three sets of detail coefficients. In other words, at the beginning of these detail coefficients contain a wealth of high-frequency information, which indicates that there is some high-frequency information at the beginning of the spectra [29]. Actually, due to potential system imperfection and limitation of spectroradiometer measurement, the scattering ray usually results in noise and disturbance at the beginning and the end of the spectral data [14], so this information at the beginning of these detail coefficients is likely caused by imperfections of the system and the spectroradiometer used in this research, so only approximate coefficients cA₃ are taken as characteristic features for further analysis.

Through feature extraction, WT, PCA and KPCA produced 89-dimensional new synthetic variables from original 651-dimensional spectral data respectively. Thus, samples can be represented with these new variables. Figure 5 shows the descriptions of tea samples in these new synthetic variable spaces. It can be found that the samples are described in the similar way by PCA and KPCA. Obviously, there are sharp peaks and valleys at the beginning of these curves in Figure 5(A,B), and then the curves gradually tend to 0, it can be concluded that most of the variance is centralized in the first tens of PCs and KPCs respectively. While in Figure 5(C) the 89-wavelet coefficients description of samples is very similar to the original spectral, indicating that the WT effectively captures the trend and characteristic information of the original spectra in low dimension.

To evaluate the performances of WT, PCA and KPCA, three regression models (Models 1, 2 and 3) were respectively developed with the three sets of newly synthesized variables as predictors. Moreover, the original 651-dimensional spectra were also taken as predictor to develop regression model (Model 4). PLSR was adopted to establish regression models based on the full cross-validation method. The results of the above four models are shown in Table 6.

In Table 6, all four models afford excellent results. In detail, Model 4 outperforms Model 1 and Model 2 with much higher accuracy and smaller error. It can be concluded that there is much more useful information in the original spectral data than those in PCs and KPCs. In other words, PCA and KPCA result in loss of useful information through compressing the 651-dimensional spectral data into the 89-dimensional PCs and KPCs. Moreover, Model 2 is slightly better than model 1, which indicates that the nonlinear algorithm of KPCA catches more useful information than the linear algorithm of PCA. Model 3 based on the 89-dimensional cA₃ obtains the optimal result in the four models, which suggests that WT algorithm is more superior than KPCA and PCA algorithms for extraction of useful information. Especially, Model 3 is much better than Model 4, which indicates that the approximate coefficients of cA₃ not only cover the characteristic information of spectra, but also avoid the interference of noise in the spectra, and WT is a powerful tool for extraction of characteristic information from spectral data.

As shown above, the 89-dimensional coefficients cA₃ were proved to be the optimal characteristics of spectroscopy, thus these coefficients were set as independent variables for further analysis. To obtain the optimal measurement, three regression algorithms including PLSR, MLR and LSSVM were adopted to develop regression models. Furthermore, LSSVM model was also based on RBF kernel function, and the kernel parameters of gam (γ) and sig2 (σ²) were optimized as 111,570 and 972.655 by grid-search which was a two-dimensional optimization procedure based on exhaustive search in a limited range [30]. The determination results of these three models are listed in Table 7. In detail, the MLR model obtains outstanding result with high correlation (r), and low root mean squared error (RMSE). Moreover, LSSVM model acquires excellent results in calibration stage, but the prediction results of the LSSVM model is slightly worse than that of the MLR model. And the PLSR model performs relative worse in both calibration and prediction stages comparing to MLR and LSSVM models. It may be concluded that the MLR model is the most proper description for the relationship between spectroscopy and MC. The results of the MLR model are plotted in Figure 6.

In the MLR model, the relationship between wavelet coefficients cA₃ and response variable (MC) could be represented by a set of regression coefficients seen in Figure 7. It can be seen that the B-coefficients of many wavelet approximation coefficients are close to zero, and intense jagged peaks and valleys can be seen at the beginning and in the middle of the regression line. The B-coefficients represent the independent contributions of each independent variable to the prediction of the dependent variable. However, the amplitude of B-coefficients is related to the amplitude of the corresponding independent variables. So it is improper to detect fingerprint wavelength solely based on B-coefficients. In this manuscript, characteristic wavelength is obtained through combination of B-coefficients and experience as well as repeated attempting. Afterwards six determination models were established based on six sets of independent variables respectively, and the results are listed in Table 8.

In Table 8, Model 13 based on the 65th–83th coefficients of cA₃ obtains excellent determination results in both calibration and prediction stages, and the prediction accuracy (r = 0.991, RMSE = 0.034) is very close to that of Model 6 based on all the 89 coefficients of cA₃. This phenomenon indicates that the cA₃ in the range of 65th–83th play an important role for determination of MC. What is the hidden meaning of this finding? As the wavelet approximation coefficients cA₃ is dimensionless, which is mathematic derived from original spectral data. Even though there is a clear linear formula relationship between wavelet coefficient and the MC of samples, the characteristic spectral absorbance of chemical bond O–H of water in the samples is obscure. However, there is a clear decomposition structure in WT, and WT has an outstanding reconstruction capability, so the relationship between spectral absorbance and MC might be detected by wavelet reconstruction. Figure 8 shows the reconstructed spectra, and the spectra in the range of 888–1,007 nm are generated from the cA₃ of 65th–83th based on wavelet reconstruction, so the absorption spectra of 888–1,007 nm might be the fingerprint wavelengths for characterization of MC. To test this hypothesis, a determination model based on these wavelengths (888–1,007 nm) was developed, and the correlation coefficient (r), RMSE of prediction and bias were 0.986, 0.046 and −1.450e⁻⁰² respectively. This result indicates that the spectra in the range of 888–1,007 nm are significantly correlated to MC of tea. This finding is corresponding to the strong and characteristic second overtone absorption position of O–H (960 nm).