Moisture Content Measurement of Broadleaf Litters Using Near-Infrared Spectroscopy Technique

Abstract

Near-infrared spectroscopy (NIRS) was implemented to monitor the moisture content of broadleaf litters. Partial least-squares regression (PLSR) models, incorporating optimal wavelength selection techniques, have been proposed to better predict the litter moisture of forest floor. Three broadleaf litters were used to sample the reflection spectra corresponding the different degrees of litter moisture. The maximum normalization preprocessing technique was successfully applied to remove unwanted noise from the reflectance spectra of litters. Four variable selection methods were also employed to extract the optimal subset of measured spectra for establishing the best prediction model. The results showed that the PLSR model with the peak of beta coefficients method was the best predictor among all of the candidate models. The proposed NIRS procedure is thought to be a suitable technique for on-the-spot evaluation of litter moisture.

1. Introduction

Floor litter refers to the relatively fresh organic residue on the uppermost layer of soil; it plays an important role in the water dynamics of the forest floor [1]. It can retain water within the layer during storm periods and deplete stored moisture through evaporation [2]. The moisture variation of litter can influence the hydrologic, carbon, and nutrient cycles of forests by altering the wetting and drying phases [3,4]. Moreover, litter moisture content is one of the critical determinants for fire ignition and spread in forests [5].

Several attempts have been made to quantitatively measure the moisture variation of floor litter over the past couple of decades [6]. Among them, the gravimetric method is the most common technique that determines the moisture amount from the difference in the litter weight of wet and dry conditions [7]. Yet, it is a highly cumbersome and labor-intensive measurement. Some researchers have attested the continuous non-destructive techniques for litter moisture measurement. Gillespie and Kidd proposed the electrical impedance grids with mock leaf sensors to monitor the leaf wetness of crops [8]. This technique was further improved by Hanson et al. [9]. On the other hand, time-domain reflectometry (TDR) probes have been employed in litter moisture measurement [10,11]. Ataka et al. developed a rather simple technique to continuously detect moisture content in litters using the modified capacitance sensors [1]. Robichaud and Bilskie developed a commercial device for measuring dead fuel moisture with a frequency domain (FD) sensor [12]. However, these methods exhibit inherently electrical bias and limitations for the in situ measurement of litter moisture, because floor litter is highly heterogeneous and porous material with lower density and compactness.

Recently, spectroscopy techniques have been widely used to nondestructively analyze the quality of crops, chemicals, and biomaterials [13,14,15,16,17,18]. Among them, a near-infrared spectroscopy (NIRS) technique is being considered as a promising spectroscopic method that treats the near-infrared (NIR) region of an electromagnetic spectrum, which corresponds to the wavelength range of 780 to 2500 nm [19]. NIRS can obtain a signal of relatively high energy when compared to far-infrared and microwave radiation, and also detect the special spectrum from the intrinsic component of object because each functional group in a molecule has characteristic absorption frequencies in the NIR spectrum by their vibrational characteristics of several chemical bonds (O-H, C-H, and N-H). The exact frequency at which a given vibration occurs is determined by the strengths of the bonds involved and the mass of the component atoms. Moreover, a NIR device structurally consists of a light splitter, a fiber optic cable, a monochromator, and a detector. Since this simple assembly is relatively easy to implement, the NIRS technique is being widely used in nondestructive quality analysis for food, feed, pharmaceutical, and agricultural industries [20,21,22]. In ecological studies, NIRS has been successfully used for the determination of organic compounds in soils [23,24,25], or to characterize the composition of forest foliage and forest litters [26,27,28,29]. Gillon et al. predicted litter decomposability more accurately by NIRS than by the initial litter composition measured by standard chemical methods [30]. Petra et al. investigated the ability of near infrared reflectance spectroscopy for the rapid determination of several chemical properties of plant litters, and found good relationships between the chemical properties and NIR spectra using partial least squares regression method by showing that the r² that varied from 0.80 to 0.99 [31].

It is known that NIR spectra possess complex and overlapping absorption bands, so that mathematical procedures are needed for turning the spectra into meaningful information. Therefore, a multivariate analysis method is commonly used for the analysis of data sets with more than one variable because it is capable of describing how the measured spectral features are related to the property of interest [32,33]. Until now, some multivariate techniques, such as partial least-squares regression (PLSR), principal component regression (PCR), and multiple linear regression (MLR), are commonly used to extract quantitative and qualitative information from NIR spectra [34,35]. However, NIRS technique is occasionally facing an undesired situations in use, i.e. noise can be induced into reflectance spectra by environment and instruments. This undesired influences can reduce the detection accuracy of a prediction model because NIR spectrum comprises various combinations and overtones of fundamental vibrational transitions that are overlapped. Therefore, suitable preprocessing techniques were developed and applied to eliminate the inappropriate noise components [36]. Rinnan et al. provided an overview of preprocessing techniques that are widely used in NIRS applications [37].

In addition to the application of suitable preprocessing techniques, an optimal variable selection method is being significantly considered as a critical step to establish a reliable and robust model because it can extract a key subset of optimal spectral frequencies to produce better prediction results [38,39]. Moreover, the optimally squeezed variable can be usefully applied to on-the-spot evaluation of dead leaf litter moisture. For this purpose, many mathematical approaches for optimal wavelength selection have been studied from the scientific knowledge on the spectroscopic properties of materials [40]. It was revealed that the selection of informative variables can produce better and simpler prediction models if the wavelengths carrying the most valuable information are carefully selected. Indeed, the removal of redundant, irrelevant, and noisy wavelengths can typically enhance the models in terms of accuracy and robustness, and also reduce their complexity [41,42]. Therefore, we employed four different variable selection methods that refer to the optimal separation of classes and the selection of spectral bands to demonstrate the accuracy enhancement of prediction models.

The main objective of this study is to develop a nondestructive and non-contact moisture measuring method of broadleaf litters using the NIRS technique through the selection of optimal wavelength. Floor litter samples were taken from deciduous forests, and the gravimetric method was used to obtain the reference moisture content of each sample. PLSR multivariate analysis with several preprocessing techniques has been employed to quantitatively predict the moisture content of floor litter. In addition, four different optimal variable selection methods were used to extract the optimal wavelength range through model development, and the model performance was also evaluated in conformity with variable selection methods.

2. Materials and Methods

2.1. Litter Moisture Measurement

The NIR spectra were acquired from the current year’s (2017 years) litters, collected in the Seoul National University Arboretum in Seoul, Korea. The broadleaf litters were manually gathered under deciduous trees, such as Chinese cork oak (Quercus variabilis), Sawtooth oak (Quercus acutissima), and Mongolian oak (Quercus mongolica). Sampled litters were then sorted to three groups according to tree species. All of the samples were immersed in water for 12 h in order to fully saturate, and were then placed in a screened plate for a certain times to naturally drain water from samples. During the spectra measurements of 10 h, all of the samples were naturally dried in the laboratory environments that were controlled at a temperature of 24 °C and relative humidity of 30%, and three leaves were extracted from each species group and the reflectance spectrum of single leaf litter was measured at every one hour. The weight of each litter was immediately measured to obtain the initial wet weight of litter when NIR measurement was done. After the experiment was completed, all of the litter samples were placed in an oven at a temperature of 70 °C for 48 h, and then measured the weight of oven-dried litter to obtain the dry weight. The moisture content of litter was calculated based on the gravimetric method as follow:

$$\mathrm{Moisture}\mathrm{content}(\%)=\frac{m_w-m_d}{m_d}\times 100,$$

(1)

where m_w and m_d are the litter weight of wet and oven-dried conditions, respectively. The moisture content histogram of broadleaf litters used in this study is shown in Figure 1, and the data are summarized in Table S1.

2.2. NIR Spectra Measurement

NIR reflectance spectra of litter samples were sampled in spectral range from 904 to 1707 nm with the NIR spectrometer (CDI-NIR128, Control Development Inc., South Bend, IN, USA). The tungsten-halogen lamp (LS-1, Ocean Optics, Largo, FL, USA) was installed as a light source, as shown in Figure 2. The reflectance spectrum was taken at three different positions on the litter surface, and then the averaged values were recorded according to the moisture content of litter samples. The integration time was set to 0.25 s, and the distance between the sample and a probe was maintained at 1 cm. The reflectance of measured spectra was calibrated by using both white-referenced and dark-referenced spectra, as shown in Equation (2). The white-referenced spectra was measured from the reflectance from a spectralon panel, and the dark-referenced spectra was measured with a completely blocked fiber optic cable in a dark room.

$$\mathrm{Reflectance} (\%)=\frac{O_i-D}{R-D}\times 100,$$

(2)

where O_i is the raw reflectance of the ith sample, and D and R represent the raw reflectance of dark- and white-referenced spectra, respectively.

2.3. Multivariate Model Development

The pretreatment of raw NIR spectra is the first step for model development and optimization to achieve the better NIR veracity. The instrument and environment causes may lead to sample-to-sample variations, such as noise, light scattering, and optical path changes. Generally, spectral preprocessing is strongly demanded to exclude noise components in reflectance spectra. In this study, two scatter correction techniques (multiplicative scatter correction (MSC), standard normal variate (SNV)), three normalization techniques (maximum normalization, mean normalization, range normalization), and three Savitzky-Golay (SG) filters (smoothing, the first- and second-derivative with a cubic polynomial and a window size of five points) were used for spectra pretreatment.

The PLSR model was also employed to establish the relationship between litter moisture content and reflectance responses. PLSR is a classical and widely used statistical method that bears a relation between independent and dependent variables in large data sets. The general structure of PLSR model is written as follow:

$$\begin{array}{l}X=T{P}^T+E\\ Y=T{Q}^T+F,\end{array}$$

(3)

where X is the n by m matrix of predictors, Y is the n by p matrix of responses, T is the n by 1 matrix of the score matrix, E and F are Gaussian error terms, and P and Q are the m by 1 and p by 1 loading matrices.

The performance of PLSR model is quantitatively assessed with the root mean square error (RMSE) and the coefficient of determination (R²). For calibration purposes, the regression model with a small RMSE value and a large value of R² could be selected as an appropriate model. In this study, latent variables (LVs) set with the smallest RMSE value was determined through the validation procedure, and then the prediction model performance was evaluated through the further test process.

A calibration model has to be evaluated with a validation set of samples to get an impression of its predictive ability. Model test has sometimes conducted in the NIRS research to further ensure model performance when the calibration and validation have completed. Among 99 spectra dataset, model calibration, validation, and test have conducted from 60, 15, and 24 dataset, respectively. Both preprocessing and regression analysis were done by using MATLAB commercial software (ver. R2016, MathWorks, Natick, MA, USA).

2.4. Estimation of Prediction Uncertainty

Uncertainty estimation is an important criterion in the prediction model, and in the case of ordinary regression, there is a standard expression that quantifies uncertainty. However, it is known that there is no generally accepted uncertainty estimation method in chemometrics, such as PLSR model. For uncertainty estimation of PLSR model, various methods such as linearization-based methods, re-sampling method, and U-deviation have been proposed [43]. In this study, the uncertainty of the PLSR model was estimated by the re-sampling method using bootstrap residual [44]. As a first step, residues of the constructed PLSR model were calculated, and in this study, pseudo-degrees of freedom was used to obtain degree of freedom in multivariate regression [45]. We constructed bootstrap residues F* through random substitutions with replacement for the calculated residues, and then added the constructed bootstrap residues to the Y_PLS matrix to compute the Y* matrix as follow:

$$Y*=Y_{PLS}+F*,$$

(4)

Then, we constructed a new PLSR model by using the obtained Y* and calculated the predicted value of the model for each sample i using the new model Beta coefficient β* to calculate the residues of the predicted value of the existing model and the predicted value of the new model. This bootstrap residue was calculated by repeating N times. The 95% bootstrap confidence intervals were computed during the calculation and the confidence interval was calculated using the percentile method in the distribution of y_i values of each sample as follows:

$$f_i^{*}{}_{N(\frac{\alpha }{2})}\le y_i\le f_i^{*}{}_{N(1-\frac{\alpha }{2})},$$

(5)

where α is the degree of confidence, and N is the repeating number of the bootstrap calculation.

2.5. Optimal Wavelength Selection

The reflectance spectra, ranging 904 to 1707 nm, were used in this study for establishing a regression model to predict the moisture content of litter. PLSR technique generally uses the entire range of wavelength’s spectra, but recent studies have employed the most appropriate wavelength. It proved that data analysis involving only a representative part of the spectra can lead to a better prediction performance. Selection of optimal wavelength is to carefully pick the subset of spectral data, which closely relate to the property of the interest. If the number of wavelength bands are greater than that of spectral samples, then the predictive model is likely to enhance its capability with the optimal wavelength. In this study, various techniques such as the peak of beta coefficients, variable importance in projection (VIP), bootstrap of beta coefficients, and interval PLS (iPLS) were applied to determine the optimal wavelength from preprocessed spectra information.

2.5.1. Peak of Beta Coefficient

The peak of beta coefficients (beta-peak method) extracts the optimal wavelength corresponding to the peak of beta coefficients. The beta coefficient, the regression coefficient of PLSR model, is a linear vector between the measured spectra and the predicted values. It measures the contribution of each wavelength set on the model’s predictive ability so as to select the optimum wavelength.

2.5.2. Variable Importance in Projection

VIP is a method that evaluates the contribution of a variable on the weight matrix of a multivariate analysis, such as a PLSR model. The idea of this method is to accumulate the importance of each variable j being reflected by w from each component. It has been widely applied to determine major wavelength bands when the variable matrix X defined in Equation (3) was considered [46]. The VIP measure v_j is defined as follows [47]:

$$v_j=\sqrt{\frac{p{\displaystyle \sum _{a=1}^A[S{S}_a(w_{ak}/{\Vert w_a\Vert }^2)}]}{{\displaystyle \sum _{a=1}^A(S{S}_a)}}},$$

(6)

where v_j represents the VIP value, p is the number of variables, SS_a is the sum of squares explained by the ath component, w_a is the weight of ath component, and w_ak is the weight of the ath component at the kth variable.

Hence, the v_j is a measure of the contribution of each variable according to the variance explained by each PLS component where (w_ak/ǁw_aǁ²) represents the importance of the kth variable. The variable selection process by a VIP technique is terminated when the calculated VIP value approaches to the threshold. In general, a threshold value of 1 is set in many studies. In this study, we compared the accuracies of developed models for selected wavelength sets by adjusting a threshold value.

2.5.3. Bootstrap of Beta Coefficients

The bootstrap of beta coefficients (bootstrap method) is a useful technique to set the confidence interval, and to estimate the significance level. It is mainly used when it is difficult to estimate the distribution of samples. In the bootstrap method, a new dataset is first obtained by randomly re-sampling the sample population. Thereafter, statistical values are obtained from the selected data sets, and then the confidence interval is set [48]. The PLSR-bootstrap method can re-sample the sample data n times and perform a PLSR for each re-sampled data set to obtain the beta coefficient. From the distribution of beta coefficients, the confidence interval is obtained according to the significance level of the beta coefficient of the specific variable. Finally, optimal wavelength is determined by removing the variables with zero value in its confidence interval. The confidence interval is computed by

$$I_k=\overline{b_k}\pm c{s}_k,$$

(7)

where I_k represents the confidence interval, $\overline{b_k}$ is the mean of beta coefficients at the kth variable, c is a constant that determines the confidence interval, and S_k is the standard deviation of beta coefficients at the kth variable.

In Equation (7), the constant c is explicitly determined regarding the level of significance. A higher significance level could select fewer wavelength bands. In this study, the number of re-sampling is set to 1000, and the model accuracy with the selected wavelength is evaluated by changing the c values [49].

2.5.4. Interval PLS

iPLS executes a PLSR model with the wavelength of a specific interval rather than the entire wavelength band. It removes unnecessary wavelength bands and uses only the certain bands of a given interval. The interval is repeatedly sampled to achieve the lowest RMSE value. In this study, the entire wavelength band divided into 20 intervals and PLSR analysis was conducted to acquire the optimal wavelength by shifting the wavelength range to the right or the left until the smallest RMSE value was pursued [50].

3. Results

3.1. Reflectance Spectra of Litters

Figure 3 depicts the raw and average NIR reflectance spectra of litter samples in the wavelength range of 900–1700 nm prior to the pretreatment. There seems to be a slight variation in all of the reflectance spectra through the litter moisture (Figure 3a). Figure 3b illustrates the averaged spectra with different degrees of litter moisture (0–50%, 50–100%, 100–150%, and 150–200%). The spectral difference appeared more clearly in Figure 3b. The absorption rate of NIR spectrum increased in the wavelength range of 1400–1500 nm regardless of litter moisture. It is commonly known that NIR spectrum is mostly absorbed at 980, 1222, and 1450 nm wavelength by the overtone and combination of the vibrational transitions of water molecules [51]. On the NIR reflectance of litters, the absorption coefficients at 980 and 1222 nm wavelength were estimated to be less than 1.3 cm⁻¹, while the value was estimated to be 29.8 cm⁻¹ at a wavelength of 1450 nm [51]. Figure 3b exhibits that the reflectivity of NIR spectrum decreased sharply beyond the wavelength of 1400 nm and had the lowest value at the wavelength of 1450 nm for all of the litter samples.

3.2. PLSR Model for Different Preprocessing Methods

A number of pretreatment techniques for NIR spectra have been tested to achieve better performance in litter moisture spectral model.

Table 1. Partial least-squares regression (PLSR) results with different preprocessing techniques.

Preprocessing Method	LVs	Calibration		Validation		Test
		Rc2	RMSEc	Rv2	RMSEv	Rt2	RMSEt
Raw data	9	0.930	12.757	0.920	19.041	0.884	19.170
SG smoothing	8	0.918	13.848	0.915	19.571	0.897	18.103
SG-1st derivative	6	0.918	13.882	0.925	18.376	0.883	19.259
SG-2nd derivative	6	0.937	12.114	0.913	19.792	0.840	22.506
MSC	8	0.926	13.132	0.933	17.432	0.914	16.482
SNV	8	0.927	13.069	0.943	16.106	0.915	16.426
Max. normalization	7	0.920	13.699	0.938	16.797	0.922	15.711
Mean normalization	9	0.926	13.149	0.926	18.308	0.892	18.479
Range normalization	7	0.922	13.535	0.931	17.722	0.920	15.970

compares the results of PLSR model with different spectral preprocessing techniques. The entire range of NIR dataset was treated and used to develop PLSR models with the different preprocessing techniques. The performance of the model was evaluated with two statistical criteria, coefficient of determination (R²) and root mean square error (RMSE), between the reference and prediction of litter moisture content. The PLSR models treated by the preprocessing methods of MSC, SNV, maximum normalization, and range normalization have a good prediction performance for validation. The test results of the models showed that maximum normalization achieved the highest performance (R_t²: 0.922, RMSE_t: 15.711), and followed by the model with range normalization preprocessing (R_t²: 0.920, RMSE_t: 15.970). Figure 4 shows the prediction result of two most highly correlated PLSR models for litter moisture estimation with the uncertainty estimation for samples shown by the error bar. The PLSR model preprocessed by the maximum normalization show better confidence in the calibration set, and the lower and upper confidence limits that are calculated for this set ranged from 11.3 to 10.9, and all of the samples were fairly predicted.

3.3. PLSR Models for Optimal Wavelength Selection Methods

Four variable selection methods were used to set the optimal wavelength bands from around 800 data set of wavelength (900–1700 nm). NIR dataset that was treated by maximum normalization was used in the works of optimal wavelength selection because the PLSR model, which developed by the preprocessing method of maximum normalization, showed the highest performance in the moisture estimation of broadleaf litter in previous result. The selected wavelength set was then input the PLSR model for predicting the moisture content of floor litter. Figure 5 illustrates the beta coefficients and VIP values of PLSR model that were derived from the entire wavelength’s spectra. As shown in Figure 5a, the peak points of beta coefficients were evenly distributed over the entire wavelength band, and the largest peak point was discovered around 1450 nm. But, large VIP values are observed around only in the range of 1400–1500 nm (exactly in the range of 1370–1580 nm) because a threshold value of 1 is set in this study.

Figure 6 shows the first four weight vectors with respect to equivalent latent variables of the PLSR model that were obtained from the entire wavelength’s spectra. The weight vector can imply the correlation between the measured spectrum and calculated latent variables through PLSR analysis. It revealed how much the measured spectrum contributed to the formation of corresponding latent variables. As shown in Figure 6a–d, the highest peak is observed around the wavelength of 1400–1500 nm, which is highly correlated with the largest absorption coefficient of a water molecule.

Figure 7 shows the optimal wavelength bands that were determined by four variable selection methods. By applying the beta-peak method, 63 optimal wavelength bands corresponding to all of the peak points of the beta coefficient were selected, and they were uniformly distributed over the entire wavelength, as shown in Figure 7a. VIP method was implemented with different threshold values (v = 0.7, 1.0, 1.5, 2.0), and the wavelength ranged from 1375 to 1577 nm was finally selected for v = 1.0 (Figure 7b). In the case of the bootstrap method, the significant level constant c was set to four cases (c = 1.0, 1.3, 1.6, 1.9). For c = 1.3, the extracted wavelength bands were relatively evenly distributed over the entire range, as shown in Figure 7c. After applying the iPLS technique with several intervals, 150 variables were selected in the wavelength range of 1394–1543 nm (Figure 7d).

Table 2. PLSR performance comparison with variable selection methods.

Method		LVs	No. of Variables	Calibration		Validation		Test
				Rc2	RMSEc	Rv2	RMSEv	Rt2	RMSEt
Full PLSR		7	804	0.920	13.699	0.938	16.797	0.922	15.711
iPLS		7	150	0.900	17.825	0.930	17.825	0.918	16.115
VIP	v = 0.7	7	316	0.924	13.391	0.934	17.276	0.923	15.597
	v = 1.0	7	203	0.905	20.131	0.910	20.131	0.918	16.130
	v = 1.5	7	150	0.909	14.620	0.926	18.314	0.920	15.916
	v = 2.0	7	40	0.884	16.522	0.853	25.818	0.859	21.167
Boots-trap	c = 1.0	7	305	0.919	13.795	0.941	16.392	0.927	15.164
	c = 1.3	7	192	0.916	14.012	0.940	16.472	0.924	15.526
	c = 1.6	7	106	0.909	14.586	0.929	17.954	0.918	16.128
	c = 1.9	7	35	0.875	17.1670	0.884	22.913	0.878	19.655
Beta-peak		6	63	0.918	13.834	0.932	17.494	0.930	14.905

presents the calibration, validation, and test performance of the PLSR models developed from the optimally selected wavelength’s spectra. Predictive abilities of different models have been compared against the PLSR model with the entire wavelength’s spectra (hereafter, the full PLSR model). As 63 wavelength bands, selected by the beta-peak method, were input in the PLSR model, the model could better predict the litter moisture than the full PLSR model. The VIP method also confirmed that the performance of the PLSR model taken from the 316 spectra showed better prediction performance (R_t²: 0.923, RMSE_t: 15.597) than that of the full PLSR_model. VIP value of 1.5 marked 150 wavelength set among the entire wavelength range, and had a similar ability for litter moisture estimation (R_t²: 0.920, RMSE_t: 15.916) with the entire wavelength PLSR_model.

The bootstrap method was applied to select the optimal wavelength. The PLSR model with 305 spectra data (c = 1.0) showed better capability in data analysis (R_t²: 0.927, RMSE_t: 15.164), while the PLSR model with 106 spectra (c = 1.6) also showed a similar prediction performance (R_t²: 0.918, RMSE_t: 16.128) when compared to the full PLSR model. A total of 150 wavelength bands was optimally selected in the iPLS technique application, and the prediction performance of PLSR model achieved the R_t² value of 0.918, RMSE_t value of 16.115.

As presented in Table 2, the predictive abilities of the PLSR models with reduced wavelength bands by employing the variable selection methods showed similar or better performance with the full PLSR_model, excepting only two cases of VIP (v = 2.0) and bootstrap method (c = 1.9). But, optimal wavelength selection can improve the prediction accuracy by effectively identifying the best subset of candidate spectra. It also eliminates unnecessary information so as to enhance the efficiency and effectiveness of model run. Figure 8 shows the prediction results of the best PLSR models, which were derived from the optimally selected wavelength’s spectra with the uncertainty estimation for samples as shown by the error bar. In common with the results of Table 2, the PLSR model developed by the variable selection of beta-peak method showed the best confidence in the test set, and the lower and upper confidence limits that were calculated for this test set ranged from 11.3 to 10.8, while worst confidence of the model developed by the VIP ranged from 13.7 to 13.2. However, whole samples of Figure 8 were fairly predicted when compared to the result of Table 2, and the results show the reliability of the model to predict the moisture content of floor litter.

4. Conclusions

Litter refers to the organic residue on the forest floor, and plays an important role in forest water cycle. Litter moisture is one of the critical determinants for evaporation, infiltration, and fire ignition in a forest environment. In this study, the commonly used technique, NIRS, has been implemented to fast, non-destructively monitor the moisture content of broadleaf litters. Several techniques for data pretreatment and optimal wavelength selection have been tested to establish a reliable and robust multivariate model.

Prior to establishing a PLSR model, the pretreatment was conducted to eliminate the unwanted noise in raw NIR spectra. Maximum normalization preprocessing, one of well-known normalization techniques, achieved the best prediction ability in litter moisture estimation among eight pretreatment methods. The entire wavelength may uphold the whole information of target samples, but the redundant data causes the decline in model accuracy. The peak of beta coefficients seems to be the best variable selector to extract the optimal wavelength bands as to yield the better predictive ability in litter moisture estimation.

The NIRS has been commonly used in both qualitative and quantitative analysis of target samples with respect to the vibrational energy of molecules. The low optical absorbance characteristics of NIR rays facilitates deeper penetration than mid-infrared rays. This optical method has some merits in reducing time-consuming work and supporting reliable results by minimizing the errors that may originate from the quick changes of litter moisture on the spot.

The NIRS also has limited ability to resolve noise that is caused by the instrument and environment causes. NIR spectra tends to be highly complex and over-parameterized, which sometime yields a poor prediction. In the face of its drawback, multivariate analysis through variable reduction and proper pretreatment have been successfully introduced in practical uses. It is believed that this research provides us with some great opportunities for multicomponent analysis of agricultural products and foods, as well as many other scientific sectors, such as life sciences and biomaterial research.