Micro-FTIR spectroscopy and partial least-squares regression for rapid determination of moisture content of nanogram-scaled heat-treated wood

Moisture sorption has a significant impact on the performance of heat-treated wood. In order to better characterize moisture sorption of heat-treated wood, a method for rapid determination of moisture content (MC) of nanogram-scaled heat-treated wood is proposed in this work. During moisture adsorption process, micro-Fourier transform infrared (FTIR) spectra of heat-treated wood were recorded. Spectral analysis was applied to these measured spectra, and then moisture adsorption sites and spectral ranges affected by moisture sorption were identified. Meanwhile, moisture contents (MCs) of heat-treated wood at various relative humidity (RH) levels were measured by using dynamic vapor sorption (DVS) setup. Based on these spectral ranges and MCs, a quantitative forecasting model was established using partial least-square regression (PLSR). Furthermore, the developed forecasting model was applied to acquire moisture sorption isotherm of heat-treated wood, in which a very positive correlation between predicted and measured MCs was observed. It was confirmed that this method was effective for rapid detection of MC of nanogram-scaled heat-treated wood which had unique advantages of rapid analysis (second level) and less sample consumption (nanogram level).


Introduction
Wood is a green and renewable building material, and it has been widely used in construction industry, furniture production and pulping and papermaking industry [1][2][3]. Heat treatment is considered to be an effective technique for wood physical modification, while moisture sorption has a significant impact on the performance of heat-treated wood [4][5][6][7]. Hence, a deeper research on moisture sorption of heat-treated wood is extremely important.
Many spectroscopic methods such as near-infrared spectroscopy [19,20], Fourier transform infrared spectroscopy (FTIR) [21][22][23] and Raman spectroscopy [24] have been employed to study moisture sorption of heattreated wood at molecular level. For example, Esteves et al. [25] demonstrated that the near-infrared spectroscopy has enough ability to predict MC of heat-treated pine (Pinus pinaster) and eucalypt (Eucalyptus globulus). Boonstra et al. [26] took advantage of FTIR spectroscopy to study the moisture sorption, and showed that the decrease in MC of heat-treated wood was attributed to the cross-linking of lignin and the reducing of OH group. Guo et al. [24] examined moisture sorption using Raman spectroscopy. Among these spectroscopic methods, FTIR spectroscopy has been widely applied, for it has many merits, for example: high spectral quality [27,28], fast data collection speed [29][30][31][32], higher signal-to-noise ratio [33,34], high sensitivity for the detection of moisture [35]. Moreover, micro-FTIR spectroscopy is a superior analytical technique for investigating micron-sized sample [36]. Through the use of a light microscope, an infrared spectrophotometer, a mercury cadmium telluride detector, and an extensive on-line software library of organic chemical spectra, this new technique is capable of identifying micron-scaled sample.
Considering that micro-FTIR spectroscopy has the ability to study micron-sized sample [37,38], in this study, we developed a method for rapid detection of MC of nanogram-scaled heat-treated wood. First, we collected micro-FTIR spectra of nanogram-scaled heattreated wood during moisture adsorption process. Second, these collected spectra were used to determine moisture sorption sites and spectral ranges correlated with moisture sorption. Third, on account of these determined spectral ranges and measured MCs, the micro-FTIR forecasting model was generated. Finally, the developed forecasting model was applied to acquire moisture sorption isotherm of nanogram-scaled heattreated wood.

Sample preparation
Wood specimens (dimensions 100 × 30 × 10 mm in length, width, and thickness) were collected from straight stem of Ginkgo biloba L. (Ginkgoaceae). Then heat treatment was used for these wood specimens in electric vacuum drying oven under controlled condition of 180 ± 1 °C. This heat treatment lasted 4 h. From these heat-treated wood specimens, transverse sections were prepared without embedding and any chemical treatment. These sections, 5 mm × 5 mm × 10 μm, were cut using a manual rotary microtome (Leica RM2135). Prior to the spectral measurement, the transverse section of heat-treated wood specimen was dried at 102 ± 3 °C in the oven for 2 h. Figure 1a shows an experimental setup for measurement of micro-FTIR spectroscopy. The main section of the experimental setup was a spectrometer (Nicolet IN 10). This spectrometer included one microscope which provided a new function for selecting observation area. During the spectral measurement, one observation area (30 μm by 30 μm) was randomly selected in the transverse section of heat-treated wood specimen, in which small quantities (~ 1 ng) of heat-treated wood was present. The micro-FTIR spectroscopy in the wavenumber range from 720 to 4000 cm −1 was recorded with 4 cm −1 resolution; 32 scans were collected. Meanwhile, the spectral data were recorded at a constant temperature of 25 °C. Figure 1b shows the sample cell. One prepared transverse section was placed on the base of sample cell which was made of ZnSe plate. Then this sample cell was mounted on the stage of the spectrometer. Meanwhile, through this sample cell, nitrogen gas with specific RH was circulated.

Experimental setup for measurement of micro-FTIR spectroscopy
Before the spectral measurement, a kinetic spectroscopy test was conducted to calculate the balance time. Figure 2 shows typical changes of the set RH, the real RH, temperature and peak height of real-time spectrum with time. When the set RH was changed to a new value (such as 5 and 10%), the latency of 3-4 min appeared. During the latency period, the real RH was more and more close to the set RH, and then arrived at stabilization. Meanwhile, 15 min later the real-time spectrum nearly did not change (peak height of the main peak at 3352 cm −1 associated with moisture sorption was advocated to detect spectral change). Based on these results, 60 min were set as the balance time.

Determination of MC using DVS apparatus
Moisture content was measured using DVS apparatus (DVS AdvantagePlus). First, a heat-treated wood sample was put on the sample tray which was hung on the microbalance which was situated in a thermostatically controlled cabinet. The apparatus recorded sample mass at the set RH (adsorption): 0, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90 and 95%, at a temperature of 25 °C and in the reverse sequence for the desorption isotherm. Meanwhile, every set RH was maintained for enough time until the sample mass of the heat-treated wood specimen altered less than 0.002% per minute during 10 min.
Moisture content was calculated using the following formula: where m d was dry sample mass and m is real-time sample mass. Figure 3 shows typical changes of MC and RH with time. When the set RH was changed to the next value (for example, 5 and 10%), a time delay of 4-10 min appeared. During the latency period, the real RH got close to the set RH, and then arrived at stabilization afterward. Therefore, MC would increase to a steady value at the set RH which was recorded as reference value. It should be noticed that three replicates were exposed and each MC collected as reference value was mean of three replicates.

Spectral data processing Acquiring difference spectrum
To further analyze the moisture sorption of heat-treated wood qualitatively and quantitatively, a difference spectrum technique was introduced.

Establishment of the forecasting model based on micro-FTIR spectra
A micro-FTIR forecasting model was generated using TQ Analyst ™ qualitative and quantitative analysis software, one of the OMINIC software suites. In description table, quantitative analysis was set to "partial least squares". In pathlength table, pathlength type was set to "constant". In components table, component was set to "moisture content" whose maximum and minimum values were both entered, and these two values were acquired using DVS apparatus. In standards table, sample spectra collected at 35 different relative humidities were introduced. Fourfifths of sample spectra measured at 28 different relative Fig. 2 Typical changes of the set RH, the real RH, temperature and peak height of real-time spectrum with time Fig. 3 Typical changes of MC and RH vs. time humidities with six replicates (i.e., 168 spectra) were assigned to the calibration set, and the remaining sample spectra acquired at seven different relative humidities with six replicates (i.e., 42 spectra) were allocated to the validation set. In spectra table, no smoothing and multipoint baseline correction were programmed. In regions table, spectral range was adjusted according to the qualitative description of heat-treated wood moisture sorption. When these parameters were enabled, the TQ Analyst ™ qualitative and quantitative analysis software program was run, and one forecasting model was established. Model performance was estimated using the correlation of determination (R 2 ), root-mean-square error of cross-validation (RMSECV) and root-mean-square error of prediction (RMSEP).

Results and discussion
Qualitatively analyzing moisture sorption in heat-treated wood Figure 4 shows heat-treated wood spectra collected at the moisture adsorption process. The development of micro-FTIR spectra could be seen over a range of MC from 0 to 15.0% in this figure. At the MC of 15.0%, the main band at 3358 cm −1 assigned to O-H stretching vibration increased, showing the OH group was moisture adsorption sites of heat-treated wood. The band at 1736 cm −1 was assigned to C=O stretching vibration of carboxylic acid, the 1600 cm −1 band belonged to the aromatic skeletal vibration plus the C=O stretching vibration, and the band at 1158 cm −1 was attributed to the glucosidic C-O-C vibration. For comparison, heattreated wood spectrum measured at 0% MC was also displayed, in which these three bands appeared at 1739, 1604, and 1160 cm −1 . When RH increased, the band positions of these three bands exhibited continuous red shifts. These peak shifts indicated the carbonyl and C-O groups were moisture adsorption sites of heattreated wood. What's more, it was confirmed that two micro-FTIR spectral ranges correlated with moisture sorption were 3700-3100 and 1780-1700 cm −1 . Figure 5 presents difference spectra at various MC levels during the moisture adsorption process. The broad envelope range of 3700-2800 cm −1 containing many component bands was observed to rise which resulted from moisture sorption. Further, the first spectral range correlated with moisture sorption was precisely identified as from 3700 to 2800 cm −1 . Meanwhile, the 1755 cm −1 band belonged to free carbonyl group decreased, while the 1725 cm −1 band was assigned to hydrogen bonded carbonyl group had reverse trend. Meanwhile, the band around 1642 cm −1 was assigned to H-O-H bending vibration. Therefore, second spectral range was precisely identified as from 1770 to 1580 cm −1 . Moreover, with an increase of RH, the 1171 cm −1 band was shown to decrease, while the 1142 cm −1 band was found to rise. The same reflections were happened in two bands located at 1171 and 1142 cm −1 . This further suggested the third spectral range correlated with moisture sorption could be precisely confirmed as from 1180 to 1140 cm −1 .
Further, the variation in the peak height for three peaks impacted by water sorption vs. MC is shown in Fig. 6. Clearly the variation for three peaks was diverse, suggesting that water molecule was absorbed by all these sorption sites. However, none of these three peaks could predict moisture sorption isotherm. Therefore, a method for determination of MC of heat-treated wood is needed urgently. Fig. 4 Heat-treated wood spectra collected during the moisture adsorption process Fig. 5 Difference spectra collected in the MC range from 1.5 to 15.0%

Quantitative detection of MC of heat-treated wood
As shown previously, DVS has offered vast amounts of moisture sorption isotherms [39][40][41]. Hence, this technique was introduced to collect MCs as measured values. Figure 7 shows the experimental sorption isotherm. This isotherm curve presented a typical sigmoidal shape commonly observed for other lignocellulosic materials [42].
In order to establish a method for rapid determination of MC, the forecasting model based on micro-FTIR spectroscopy ought to be determined firstly. As mentioned earlier, spectral range was an important parameter for the forecasting model. Three spectral ranges of 3700-2800, 1770-1580, and 1180-1140 cm −1 correlated with moisture sorption were proposed as Case A. Further, the widened and narrowed spectral ranges were introduced separately as Case B and C (Case B: 3700-2800, 2800-2700, 1770-1580, and 1180-1140 cm −1 ; Case C: 3700-3000, 1770-1580, and 1180-1140 cm −1 ). In all three cases, the micro-FTIR forecasting model was generated and corresponding parameters such as RESECV, RESEP, and R 2 could be acquired in TQ Analyst ™ software (as shown in Table 1). The established forecasting model in Case A had highest forecast accuracy, for it possessed highest values of R 2 as well as the lowest values of RMSEP and RMSECV. What's more, this model made use of the whole spectral ranges correlated with moisture sorption, and the change of spectral range (increase and decrease) could decrease the accuracy of the established forecasting model in Case B and C.
Based on the established forecasting model, MCs of heat-treated wood were predicted. Moreover, measured values using DVS setup are displayed in Fig. 8. During the moisture adsorption process, the predicted MCs were much closed to the measured values (relative error was lower than 3%). Results indicated that this method for rapid detection of MC in nanogram-scaled heat-treated wood using micro-FTIR spectroscopy and partial leastsquares regression was effective and efficient. Compared to the traditional DVS, it has unique advantages of rapid analysis (second level) and less sample consumption (nanogram level).

Conclusion
One method for rapid determination of moisture content of nanogram-scaled heat-treated wood was proposed here. Micro-FTIR spectra were measured during the moisture adsorption process. An analysis of these spectra confirmed that hydroxyl and carbonyl groups were moisture sorption sites of heat-treated wood. Moreover, three spectral ranges, such as 3700-2800, 1770-1580, and 1180-1140 cm −1 were identified that related to moisture sorption. Based on these three spectral ranges and referential values, a quantitative forecasting model was built using PLSR. Further, the developed forecasting model was applied to acquire moisture sorption isotherm of heat-treated wood, in which a very positive correlation between the forecasts and recorded values. It was confirmed that this method  Equilibrium moisture content of heat-treated wood over the full set RH range in the adsorption process for rapid detection of moisture content in nanogramscaled heat-treated wood was effective which had unique advantages of rapid analysis (second level) and less sample consumption (nanogram level).

Table 1 PLSR quality parameters for cross-and test set-validation for the proposed three cases
RMSECV: root-mean-square error of cross-validation; RMSEP: root-mean-square error of prediction; R 2 : coefficient of determination