Top Navigation Bar

ASAE Journal Article

Soil Nitrogen Content Influence on Canopy Reflectance Spectra

Y. Shao, Y. Bao, Y. He


Published in Transactions of the ASABE Vol. 54(6): 2343-2349 ( Copyright 2011 American Society of Agricultural and Biological Engineers ).

Submitted for review in July 2010 as manuscript number IET 8688; approved for publication by the Information & Electrical Technologies Division of ASABE in September 2011.

The authors are Yongni Shao, Doctoral Student, Yidan Bao, Associate Professor, and Yong He, ASABE Member, Professor, College of Biosystems Engineering and Food Science, Zhejiang University. Hangzhou, China. Corresponding author: Yong He, College of Biosystems Engineering and Food Science, Zhejiang University. 866 Yuhangtang Road, Hangzhou 310058, China; phone: 086-571-88982143; fax: 086-571-88982143; e-mail: yhe@zju.edu.cn.


Abstract. Making nitrogen (N) recommendations without knowing the N supply capability of a soil can lead to inefficient use of N and potential pollution of the groundwater. Conventional soil test techniques are destructive and time-consuming. Remote sensing of canopy reflectance has the potential capability of non-destructive and rapid estimation of crop total N. In addition, this technique could be used for evaluation on soil N availability. This study was conducted on an experimental field at Zhejiang University with rice in the tillering and booting stages because these stages require maximum N for proper growth. The soil total N content of variable N application treatments was measured at the two stages. The rice canopy reflectance spectra were measured by visible and near-infrared spectroscopy (Vis/NIRS, 350 to 1075 nm). The partial least squares (PLS) method was used to build a calibration model between rice canopy reflectance and soil total N. The model was optimized with four latent variables (LVs), with coefficient of prediction (r p ), root mean square error of prediction (RMSEP), and bias of 0.81, 8.44, and 2.03 for the tillering stage and 0.91, 7.01, and -1.50 for the booting stage, respectively. Moreover, independent component analysis (ICA) was used to select several sensitive wavelengths (SWs) based on loading weights. The optimal least squares support vector machines (LS-SVM) model was achieved with SWs (560 nm, 720-730 nm, and 655-680 nm) selected by ICA. This model had better performance for soil N estimation in both the tillering and booting stages, with correlation coefficient (r), RMSEP, and bias of 0.83, 7.80, and 2.15, respectively. The results show that ICA was effective with respect to the selection of SWs. In addition, the use of Vis/NIRS canopy reflectance spectra can effectively estimate the soil total N content.

Keyword. Independent component analysis (ICA), Least squares support vector machines (LS-SVM), Rice, Soil nitrogen.

Near-infrared reflectance spectroscopy (NIRS) has historically been used in the detection of several soil properties directly (Mouazen et al., 2007). Fidencio et al. (2002) adopted NIRS to determine the organic matter in soils, and the model built by radial basis function networks was better than multi-layer perceptor and partial least squares regression. Russell et al. (2002) used NIR spectroscopy to predict nitrogen mineralization in rice soils, and the results showed that similar wavelengths were correlated with both plant N uptake and mineralization. Lee et al. (2003) estimated chemical properties of Florida soils using spectral reflectance, and the prediction model built by partial least squares for measured soil chemical properties for the three soil orders yielded R 2 values of 0.24 to 0.88. Brown et al. (2005) used visible and near-infrared diffuse reflectance spectroscopy for soil C prediction in Montana, and the results demonstrated that the spatial structure of calibration and validation samples is vital.

The random selection of non-independent test samples can overestimate predictive accuracy relative to independent validation. Stenberg et al. (2005) used near-infrared reflectance spectroscopy to predict nitrogen uptake by winter wheat. They found that NIR spectroscopy integrates information on organic C with other relevant soil components, and it has good potential ability to predict complex functions of soils, such as N mineralization. Brown et al. (2006) applied visible and near-infrared diffuse reflectance spectroscopy to predict organic and inorganic C content in soils, and the validation root mean squared deviations were 54 g kg -1 for clay, 7.9 g kg -1 for soil organic C, and 5.6 g kg -1 for inorganic C. Khanna et al. (2007) developed angle indexes for soil moisture estimation, dry matter detection, and land cover discrimination based on NIRS, and the results showed that the shortwave angle slope index is a good indicator of soil moisture. The angle at NIR is promising in discrimination of dry plant matter from soil and in quantifying the amount of dry matter. Waiser et al. (2007) investigated soil clay content with visible and near-infrared diffuse reflectance spectroscopy, and the validated clay predictions had a root mean squared deviation of 61 and 41 g kg -1 dry soil for the field-moist and air-dried in situ cores, respectively. Wetterlind et al. (2008) used near-infrared reflectance spectroscopy to predict nitrogen uptake in cereals, and the cross-validated NIR calibrations for plant N uptake within fields for separate years resulted in r 2 values of 0.75 to 0.85 and average cross-validation errors of 11 to 16 kg N ha -1 for two fields.

NIRS has also been used to detect rice plant status in recent years. Batten et al. (1991) used near-infrared reflectance spectroscopy to determine shoot nitrogen status in rice and achieved a standard error of prediction of 0.15% N. Our group at Zhejiang University also investigated nitrogen status in rice using NIRS. We found that the wavelengths of 560 nm, 575-580 nm, 700 nm, 730 nm, and 740 nm were better for nitrogen detection (Shao et al., 2010). Liu et al. (2010) discriminated rice panicles using reflectance data based on principal component analysis and support vector classification. The overall accuracies of support vector classification with principal component spectra derived from the raw, first, and second reflectance spectra for the testing dataset were 96.55%, 99.14%, and 96.55%, respectively. Zhang et al. (2010) estimated leaf nitrogen content in rice with near-infrared reflectance spectroscopy. They found that the performance of the models with leaf powder was better than with fresh leaf.

There has been significant investigation on soil properties using NIRS to evaluate soil directly in the last decade. However, there is little literature available concerning the application of NIRS to the measurement of plant canopy reflectance for the purpose of determining soil properties. Xue et al. (2006) measured soil-available N content using rice canopy reflectance spectra. In their study, possible normalized difference vegetation indices (NDVIs), ratio vegetation indices, and hybrid vegetation indices were calculated. Correlation was made between NDVIs and soil-available N, and the regression equation was analyzed. The physical explanation is that nitrogen accumulation for a crop is generally 50% to 80% of the soil nitrogen. In addition, nitrogen fertilizer uptake by a crop needs the soil media. The soil nutrient supply directly affects the quality of the crop and determines the crop yield, and the growing status of the crop has an influence on the soil nutrient supply. Therefore, there may be a relationship between canopy reflectivity and soil nutrient content, and it is meaningful to use canopy reflectance spectra to estimate soil nitrogen content.

Least squares support vector machines (LS-SVM) are capable of calculating linear and nonlinear relationships between spectra and soil chemical constituents (Suykens and Vanderwalle, 1999; Suykens et al., 2002). Based on this, a new combination of independent component analysis (ICA) with LS-SVM is proposed as a nonlinear calibration model for quantitative analysis using spectroscopic techniques. The performance of ICA-LS-SVM is then used to estimate soil N content in the tillering and booting stages of rice.

This study aims to use canopy reflectance for non-destructive and rapid estimation of plant total N during the rice tillering and booting stages. In addition, the ICA-LS-SVM model is used to find sensitive wavelengths for estimation of soil N content.

Materials and methods

Experimental Design

The experimental samples in this study were 18 basins of rice, which were planted in conditioned soil with three nitrogen levels: 0, 120, and 240 N kg h m -2 . To avoid accidental damage to the basins or samples, a duplicate set of basins was prepared, so there were 36 basins in total. For each nitrogen level, there were 12 basins including the additional basins. Each basin's inner diameter and height were 30 and 45 cm, respectively. Each basin contained 10 kg soil and four rice plants. The basins were placed in a slotted field using the surrounding soil for backfill. The soil used in this experiment was from the 20 to 40 cm depth of the experimental field.

Field Data Acquisition

At each growing stage, two canopy tissue samples from each of 16 basins (the rice samples in the other two basins were incompletely growing, which may have been caused by bad seedlings) were selected for spectral measurement. Samples were also selected from another 16 replicated basins, and a total of 64 samples were obtained. Measurements were conducted at the tillering (June 15) and booting (July 2) stages of the 2009 growing season. A total of 128 samples were prepared.

All rice canopy reflectance measurements were made using a portable spectroradiometer (FieldSpec Vis/NIR, Analytical Spectral Devices, Boulder, Colo.). The instrument uses a 512-element photo-diode array with a resolution of 3.5 nm and a range of 325 to 1075 nm. The scan number for each spectrum at the same position was set to ten, and three reflectance spectra were taken for each sample. Thus, a total of 30 individual measurements were stored for later analysis. Considering its 25° field of view, the spectroradiometer was placed above the rice canopy at a distance of 70 cm from the top of the canopy. To achieve accurate relative reflectance measurements, a white reference panel was scanned before scanning the samples until a clean 100% reference line was obtained. The 64 samples were divided into a calibration set of 48 samples and a prediction set of 16 samples. In order to compare the performance of different calibration models, the samples in the calibration and prediction sets were used for all calibration models.

Soil Total N Content Measurement

The rapid N cube (Elementar Analysensysteme, Hanau, Germany) was used to measure total soil N content. Its basic concept is combustion of the sample and catalytic post-combustion, followed by reduction of the combustion gases. The formed N 2 flows through a thermoconductivity detector (TCD) and produces an electrical signal that correlates to the N content in the sample.

The Dumas combustion method was used to determine soil N content 20 cm beneath the soil surface in the rice basins, and the soil samples were collected and investigated in the laboratory. All samples were first oven-dried and then ground. Subsamples (1 g) were prepared and pressed into tablets for measurement. The soil N contents of all the samples were in the range of 97.1 to 216.5 mg kg -1 , and the standard deviation was 25.5 mg kg -1 .

Data Pretreatment

The first and last 75 wavelengths from the spectral measurements were eliminated to improve the measurement accuracy. All visible and NIR spectroscopy analysis was based on 400 to 1000 nm. Spectral data were first processed with ViewSpec Pro V4.02 (Analytical Spectral Device, Inc.) and then smoothed through a Savitzky-Golay filter with a window width of 7 (3-1-3) points. Data preprocessing was managed with Unscrambler V 9.6 (CAMO Software AS, Oslo, Norway).

Partial Least Squares

Partial least squares (PLS) is a bilinear modeling method in which the original independent information ( X -data) is projected onto a small number of latent variables (LVs) to simplify the relationship between X and Y for predicting with the smallest number of LVs.

In the development of the PLS model, calibration models were built between the canopy spectral data and the soil total N content. Full cross-validation was used to evaluate the quality and to prevent overfitting of calibration models. The optimal number of LVs was determined by the lowest value of predicted residual error sum of squares (PRESS). The prediction performance was evaluated by the correlation coefficient (r), the root mean square error of calibration (RMSEC) or prediction (RMSEP), and bias. The ideal model should have a higher r value, but smaller differences between RMSEC and RMSEP, and lower bias as well. The RMSEC, RMSEP, and bias are calculated as:

IET8688_files\eqn\eqn1.gif (1)

IET8688_files\eqn\eqn2.gif (2)

where IET8688_files\eqn\eqn3.gif is the predicted value of each observation in calibration set (RMSEC) or prediction set (RMSEP), y i is the measured value of the each observation, and I p is the sample number in the prediction set. The models were carried out using Unscrambler V9.6 (CAMO Software AS, Oslo, Norway).

Independent Component Analysis

Independent component analysis (ICA) is a well-established statistical signal processing technique that aims to decompose a set of multivariate signals into a base of statistically independent components with minimal loss of information content (Fiori, 2003; Kano et al., 2004; Virtanen et al., 2009). The independent components are LVs, meaning they cannot be directly observed, and the independent component must have a non-Gaussian distribution. A chief explanation of a noise-free ICA model could be written as the following expression:

x = As (3)

where x denotes the recorded data matrix, s represents the independent components, and A represents the coefficient matrix. The independent components (ICs) are obtained by a high-order statistic, which is a much stronger condition than orthogonality. This goal is equivalent to finding a separating matrix W that satisfies:

IET8688_files\eqn\eqn4.gif = Wx (4)

where IET8688_files\eqn\eqn5.gif is the estimation of s . The separating matrix W can be trained as the weight matrix of a two-layer feedforward neural network in which x is the input and IET8688_files\eqn\eqn6.gif is the output.

There are many algorithms for performing ICA (Hyvarinen et al., 2001; Lee, 1998). Among these algorithms, the fast fixed-point algorithm (FastICA), developed by Hyvärinen and Oja (2000), is highly efficient for performing the estimation of ICA. Fast ICA was chosen for ICA and carried out in Matlab 7.0 (The Math Works, Natick, Mass.).

Least Squares Support Vector Machines

LS-SVM can work with linear or non-linear regression or multivariate function estimation relatively quickly (Borin et al., 2006). The LS-SVM method uses a linear set of equations instead of a quadratic programming (QP) problem to obtain the support vectors (SVs) (Guo et al., 2006; Chen et al., 2007). The LS-SVM model can be expressed as:

IET8688_files\eqn\eqn7.gif (5)

where K ( x , x i ) is the kernel function, x i is the input vector, a i is Lagrange multipliers called support value, and b is the bias term.

In the model development using LS-SVM and radial basis function (RBF) kernel, the optimal combination of the gamma ( ? ) and sigma 2 ( s 2 ) parameters was selected that resulted in the smallest root mean square error of cross-validation (RMSECV). Parameter ? determines the trade-off between structural risk minimization and empirical risk minimization. Parameter s 2 controls the value of the function regression error and directly influences the number of initial eigenvalues or eigenvectors (Wu et al., 2007). In this study, gamma ( ? ) was optimized in the range of 2 -1 to 2 10 and sigma 2 ( s 2 ) was optimized in the range of 2 to 2 15 with adequate increments (Li et al., 2009). The grid search had two steps: the first step was a crude search with a large step size, and the second step was the specified search with a small step size (for details, see Xie and Ying, 2009). The free LS-SVM toolbox (LS-SVM v 1.5, Suykens, Leuven, Belgium) was applied with Matlab 7.0 to develop the calibration models.

Results and Discussion

Reflectance Spectra Investigation

The reflectance spectra shown in figure 1 represent the average spectral characteristics of 64 samples of rice canopy in the tillering and booting stages. The variable nitrogen contents were similar in trend but different in their corresponding reflectivity. The booting stage was greater in corresponding reflectivity than the tillering stage. In the blue (400-500 nm) and red (600-700 nm) regions, the low reflectance (<10%, fig. 1) was due to the strong absorbtion of blue and red light from crop photosynthesis. In the green band (560-570 nm), a small peak appeared because of this absorption. Reflectance increased rapidly at about 690-760 nm (red edge) from 10% to 50% to 70% (booting stage). The reflectance in the booting stage was obviously higher than in the tillering stage. A possible reason was that the character of the rice canopy was more affected by environmental factors (soil, etc.) in the tillering stage than in the booting stage. The higher reflectance in the near-infrared region in the booting stage was related to the increases of plant biomass and leaf area index (LAI).

IET8688_files/image8.gif IET8688_files/image9.gif

Figure 1. Reflectance spectra of rice canopy with different rates of nitrogen in the (a) tillering stage and (b) booting stage.

The various nitrogen treatments caused significant variation in the reflectance of the rice canopy. The canopy reflectance increased in both the visible and near-infrared regions as nitrogen availability increased. An increased nitrogen supply always causes lower reflectance on average in the visible range. However, in our study, the reflectance in the visible range was similar among the different nitrogen levels, and the increased nitrogen supply had higher reflectance. A possible reason is that the result was affected by soil factors or by the measurement date, and it was an average of multiple samples of rice canopy reflectance, which affects the final result. The increase in reflectance with increased nitrogen supply in the near-infrared region is related to the increases in plant biomass, leaf area index (LAI), and water content at the high nitrogen rate (Hansen and Schjoerring, 2003).

PLS Regression Model

A PLS model between the rice canopy reflectance spectra and the soil total N content was developed after spectral data preprocessing by Savitzky-Golay smoothing and multiplicative scatter correction (MSC). Different LVs were applied to build the calibration models, and no outliers were detected in the calibration set during development of the PLS models. Of the 64 samples, 48 were used as the calibration set and the remaining 16 samples were the prediction set. Among the calibration models, the models with four LVs were the best for predicting total N in both the tillering and booting stages using the evaluation standards discussed earlier in the Partial Least Squares section. In the prediction models, the correlation coefficient (r p ), RMSEP, and bias for the optimal PLS models were 0.81, 8.44, and 2.03 for the tillering stage and 0.91, 7.01, and -1.50 for the booting stage, respectively. Figure 2 shows the measured versus predicted values for soil total N content by the PLS models. The diagonal dotted line shows the ideal result, i.e., the predicted values are equal to the measured values. The closer the sample plots are to this line, the better the model is. Hence, an acceptable prediction performance was achieved by the PLS models.

IET8688_files/image10.gif IET8688_files/image11.gif

Figure 2. Measured versus predicted values plots for soil total N content by PLS model in the (a) tillering stage and (b) booting stage.

Figure 2 shows separate models for each stage, and we also built a PLS model for the total N in both stages. All 128 samples (64 in the tillering stage and 64 in the booting stage) were divided into a calibration set of 88 samples and a prediction set of 40 samples, and the performance of the Vis/NIR model was evaluated using the 40 samples in the prediction set. The results showed that the r p , RMSEP, and bias were 0.85, 7.67, and 1.88 for total N content in the tillering and booting stages.

Sensitive Wavelengths Analysis Based on Regression Coefficients

The sensitive wavelengths reflecting the characteristics of the reflectance spectra for total N content in the tillering and booting stages were obtained based on regression coefficients (fig. 3) using the PLS method. From figure 3, we can find that wavelengths of 560-580 nm and 730-750 nm might be of particular importance for the N content calibration in both the tillering and booting stages (Johnson et al., 2008; Yao et al., 2010).

The reflectance spectra curves of plants have significant characteristics. At 350 to 490 nm, the reflectance spectra curves have a gentle shape (no evident peaks and troughs) and a low reflectance value. At 490 to 600 nm, the curves have a peak shape and a medium reflectance value. At 600 to 700 nm, due to the strong absorption of light, the curves have a trough shape, and the majority of plants have a minimum reflectance value at 680 or 670 nm. At wavebands from 700 to 750 nm, the curves have a sharp curve and are closer to a straight line. At wavebands from 750 to 1000 nm, the reflectance spectra curves of plants have a wavy undulating shape and a high reflectivity, indicating a low absorption rate.

IET8688_files/image12.gif IET8688_files/image13.gif

Figure 3. Sensitive wavelengths for total N content in the (a) tillering stage and (b) booting stage based on regression coefficients.

ICA-LS-SVM Models

ICA was applied for the selection of sensitive wavelengths (SWs), which could reflect the main features of the raw absorbance spectra, for total N content in the tillering and booting stages. FastICA was used on the preprocessed spectra data, and the main absorbance peaks and valleys were indicated by the spectra of the ICs. The SWs were selected by the weights of the first three ICs, and the wavelengths with the highest weights of each IC were selected as the SWs. Three SWs were selected, corresponding to the first three ICs, and they were 560 nm, 720-730 nm, and 655-680 nm, as shown in figure 4. In order to evaluate the performance of the SWs, they were applied as the input data matrix to develop the ICA-LS-SVM models.

IET8688_files/image14.gif

Figure 4. Three ICs for soil total N content estimation by the ICA-LS-SVM model.

Before building the ICA-LS-SVM calibration models, three steps are crucial for the optimal input feature subset, proper kernel function, and optimal kernel parameters. First, the three SWs of 560 nm, 720-730 nm, and 655-680 nm from the ICA analysis should be used as the input data set. Second, RBF can handle the nonlinear relationships between the spectra and target attributes. Finally, two important parameters, gamma ( ? ) and sigma 2 ( s 2 ), should be optimized for the RBF kernel function, as mentioned earlier.

All 128 samples (64 in the tillering stage and 64 in the booting stage) were divided into a calibration set of 88 samples and a prediction set of 40 samples, and the performance of the Vis/NIR models was evaluated using the 40 samples in the prediction set. The results showed that the r p , RMSEP, and bias were 0.83, 7.80, and 2.15 for total N content in the tillering and booting stages. Figure 5 shows the predicted versus measured values. The ICA-LS-SVM models achieved better performance for total N content in both the tillering and booting stages.

IET8688_files/image15.gif

Figure 5. Measured versus predicted values plots for soil total N content by ICA-LS-SVM model in both tillering and booting stages.

Xue et al. (2006) showed that the transformed soil adjusted vegetation index calculated with 870 and 710 nm wavelengths obtained a coefficient of determination (R 2 ) that increased from 0.46 to 0.6. The ratios of vegetation index and new soil adjusted vegetation index calculated with 1220 and 760 nm were well related to soil-available N content at the heading and filling stages. In our study, the optimal LS-SVM model was achieved with the SWs (560 nm, 720-730 nm, and 655-680 nm) selected by ICA, and it had better performance for soil content estimation in both the tillering and booting stages, with r p , RMSEP, and bias of 0.83, 7.80, and 2.15, respectively. In addition, the R 2 was 0.69, which was better than the results of Xue et al. (2006).

Wavelengths near 560 nm were close to the green peaks, and wavelengths at 660-680 nm or 720-730 nm were close to the red edge position. Therefore, the selection of SWs was suitable for such a situation in the present study, and the effectiveness of the SWs was also validated. The SWs represented most of the features of the original spectra and could replace the whole wavelength region to predict the total N content in the tillering and booting stages of rice.

The ICA-LS-SVM models predicted the soil total N content in both the tillering and booting stages of rice. The LS-SVM models utilized the nonlinear information of the spectral data, improving the prediction precision. The ICs from ICA were obtained by a high-order statistic, providing a much stronger condition than orthogonality. Therefore, the SWs selected from the ICs were more effective and may prove helpful for the development of portable instruments to predict soil total N content during rice cultivation.

The practical advantage of the proposed method is its simplicity. Once the model is build, it only needs the reflectance value of the sample, and then the prediction results, such as soil N content, can be obtained. No special preparation of the samples was required. Because of its simplicity and fast response, the proposed method can realize the goal of real-time analysis in agricultural. After extracting the sensitive wavelengths by ICA, the method was capable of estimating soil N content using Vis/NIRS canopy reflectance spectra. The limitation of the method is that the model may be affected by the growing period of the rice. Therefore, a large number of samples in different growing periods were required for calibration.

Conclusions

Vis/NIR spectroscopy was successfully utilized for the determination of soil total N content in the tillering and booting stages of rice. PLS was applied to build calibration models between the rice canopy reflectance and soil total N content. Optimal results were obtained with r p , RMSEP, and bias of 0.81, 8.44, and 2.03 for the tillering stage and 0.91, 7.01, and -1.50 for the booting stage, respectively. The SWs reflecting the characteristics of the reflectance spectra for total N content in the tillering and booting stages were analyzed as well, and the analysis showed that wavelengths of 560-580 nm and 730-750 nm might be of particular importance for N content calibration in both the tillering and booting stages.

Moreover, independent component analysis (ICA) was executed to select several SWs based on loading weights. The combined ICA-LS-SVM model was achieved with SWs (560 nm, 720-730 nm, and 655-680 nm) selected by ICA, and a two-step grid search technique was used for the optimal RBF kernel parameters ( ? and s 2 ). ICA-LS-SVM models were developed with r p , RMSEP, and bias of 0.83, 7.80, and 2.15 for soil N content estimation in the tillering and booting stages. The overall results indicate that ICA was an effective tool with respect to the selection of SWs. The proposed method has potential ability to estimate soil total N content by using Vis/NIRS canopy reflectance spectra in the tillering and booting stages of rice.

Acknowledgements

This study was supported by the Natural Science Foundation of China (Project No. 61134011), Zhejiang Provincial Natural Science Foundation of China (Project No. Z3090295), 863 National High-Tech Research and Development Plan (Project No. 2011AA100705), and Zhejiang Provincial Science and Technology Innovation Team (Project No. 2009R50001).

REFERENCES

Batten, G. D., A. B. Blakeney, M. Glennie-Holmes, R. J. Henry, A. C. McCaffery, P. E. Bacon, and D. P. Heenan. 1991. Rapid determination of shoot nitrogen status in rice using near-infrared reflectance spectroscopy. J. Sci. Food and Agric. 54(2): 191-197.

Borin, A., M. F. Ferrao, C. Mello, D. A. Maretto, and R. J. Poppi. 2006. Least squares support vector machines and near-infrared spectroscopy for quantification of common adulterants in powdered milk. Anal. Chim. Acta 579(1): 25-32.

Brown, D. J., R. S. Bricklemyer, and P. R. Miller. 2005. Validation requirements for diffuse reflectance soil characterization models with a case study of VNIR soil C prediction in Montana. Geoderma 129(3-4): 251-267.

Brown, D. J., K. D. Shepherd, M. G. Walsh, M. D. Mays, and T. G. Reinsch. 2006. Global soil characterization with VNIR diffuse reflectance spectroscopy. Geoderma 132(3-4): 273-290.

Chen, Q. S., J. W. Zhao, C. H. Fang, and D. M. Wang. 2007. Feasibility study on identification of green, black, and Oolong teas using near-infrared reflectance spectroscopy based on support vector machine (SVM). Spectroc. Acta A 66(3): 568-574.

Fidencio, P. H., R. J. Poppiand, and J. C. De Andrade. 2002. Determination of organic matter in soils using radial basis function networks and near-infrared spectroscopy. Anal. Chim. Acta 453(1): 125-134.

Fiori, S. 2003. Overview of independent component analysis technique with an application to synthetic aperture radar (SAR) imagery processing. Neural Networks 16(3-4): 453-467.

Guo, H., H. P. Liu, and L. Wang. 2006. Method for selecting parameters of least squares support vector machines and application. J. System Simulation 18(7): 2033-2036, 2051.

Hansen, P. M., and J. K. Schjoerring. 2003. Reflectance measurement of canopy biomass and nitrogen status in wheat crops using normalized difference vegetation indices and partial least squares regression. Remote Sens. Environ. 86(4): 542-553.

Hyvarinen, A., and E. Oja. 2000. Independent component analysis: Algorithms and applications. Neural Networks 13(4-5): 411-430.

Hyvarinen, A., J. Karhunen, and E. Oja. 2001. Independent Component Analysis. New York, N.Y.: John Wiley and Sons.

Johnson, R. M., R. P. Viator, J. C. Veremis, E. P. Richard Jr., and P. V. Zimba. 2008. Discrimination of sugarcane varieties with pigment profiles and high-resolution, hyperspectral leaf reflectance data. J. American Soc. Sugar Cane Tech. 28: 63-75.

Kano, M., S. Hasebe, I. Hashimoto, and H. Ohno. 2004. Evolution of multivariate statistical process control: Application of independent component analysis and external analysis. Comput. Chem. Eng. 28(6-7): 1157-1166.

Khanna, S., A. Palacios-Orueta, M. L. Whiting, S. L. Ustin, D. Riano, and J. Litago. 2007. Development of angle indexes for soil moisture estimation, dry matter detection, and land cover discrimination. Remote Sens. Environ. 109(2): 154-165.

Lee, T. W. 1998. Independent Component Analysis: Theory and Application. Boston, Mass.: Kluwer.

Lee, W. S., J. F. Sanchez, R. S. Mylavarapuand, and J. S. Choe. 2003. Estimating chemical properties of Florida soils using spectral reflectance. Trans. ASAE 46(5): 1443-1453.

Li, Z. F., J. J. Yu, and Y. He. 2009. Use of NIR spectroscopy and LS-SVM model for the discrimination of varieties of soil. IFIP Advances in Info. and Comm. Tech. 293: 97-105.

Liu, Z. Y., J. J. Shi, L. W. Zhang, and J. F. Huang. 2010. Discrimination of rice panicles by hyperspectral reflectance data based on principal component analysis and support vector classification. J. Zhejiang Univ. Sci. B 11(1): 71-78.

Mouazen, A. M., M. R. Maleki, J. De Baerdemaeker, and H. Ramon. 2007. On-line measurement of some selected soil properties using a VIS-NIR sensor. Soil Tillage Res. 93(1): 13-27.

Russell, C. A., J. F. Angus, G. D. Batten, B. W. Dunn, and R. L. Williams. 2002. The potential of NIR spectroscopy to predict nitrogen mineralization in rice soils. Plant Soil 247(2): 243-252.

Shao, Y. N., C. J. Zhao, Y. D. Bao, and Y. He. 2010. Quantification of nitrogen status in rice by least squares support vector machines and reflectance spectroscopy. Food Bioproc. Tech. doi: 10.1007/s11947-009-0267-y.

Stenberg, B., A. Jonsson, and T. Borjesson. 2005. Use of near-infrared reflectance spectroscopy to predict nitrogen uptake by winter wheat within fields with high variability in organic matter. Plant Soil 269(1-2): 251-258.

Suykens, J. A. K., and J. Vanderwalle. 1999. Least squares support vector machine classifiers. Neural Proc. Lett. 9(3): 293-300.

Suykens, J. A. K., T. Van Gestel, J. De Brabanter, B. De Moor, and J. Vandewalle. 2002 . Least Squares Support Vector Machines. Singapore: World Scientific.

Waiser, T. H., C. L. S. Morgan, D. J. Brown, and C. T. Hallmark. 2007. In situ characterization of soil clay content with visible near-infrared diffuse reflectance spectroscopy. SSSA J. 71(2): 389-396.

Wetterlind, J., B. Stenberg, and A. Jonsson. 2008. Near-infrared reflectance spectroscopy compared with soil clay and organic matter content for estimating within-field variation in N uptake in cereals. Plant Soil 302(1-2): 317-327.

Wu, D., S. Feng, and Y. He. 2007. Infrared spectroscopy technique for the nondestructive measurement of fat content in milk powder. J. Dairy Sci. 90(8): 3613-3619.

Xie, L. J., and Y. B. Ying. 2009. Use of near-infrared spectroscopy and least squares support vector machine to determine quality change of tomato juice. J. Zhejiang Univ. Sci. 10(6): 465-471.

Xue, L. H., P. Lu, L. Z. Yang, Y. H. Shan, X. H. Fan, and Y. Han. 2006. Estimation of soil nitrogen status with canopy reflectance spectra in rice. J. Plant Ecol. 30: 675-681 (in Chinese).

Virtanen, H., T. Noponen, and P. Merilainen. 2009. Comparison of principal and independent component analysis in removing extracerebral interference from near-infrared spectroscopy signals. J. Biomed. Opt. 14(5): 054032.

Yao, X., Y. Zhu, Y. C. Tian, W. Feng, and W. X. Cao. 2010. Exploring hyperspectral bands and estimation indices for leaf nitrogen accumulation in wheat. Intl. J. Applied Earth Obs. and Geoinfo. 12(2): 89-100.

Zhang, Y. S., X. Yao, Y. C. Tian, W. X. Cao, and Y. Zhu. 2010. Estimating leaf nitrogen content with near-infrared reflectance spectroscopy in rice. Chinese J. Plant Ecol. 34(6): 704-712.