Published online 21 June 2006
Published in Soil Sci Soc Am J 70:1295-1302 (2006)
DOI: 10.2136/sssaj2005.0297
© 2006 Soil Science Society of America
677 S. Segoe Rd., Madison, WI 53711 USA
Soil Physics
Characterization of Soil Water Content Using Measured Visible and Near Infrared Spectra
A. M. Mouazen*,
R. Karoui,
J. De Baerdemaeker and
H. Ramon
Division of Mechatronics, Biostatistics and Sensors (MeBioS), Faculty of Bioscience Engineering, Kasteelpark Arenberg 30, B-3001 Heverlee, Belgium
* Corresponding author (abdul.mouazen{at}biw.kuleuven.be)
 |
ABSTRACT
|
|---|
Soil water content (WC) affects the accuracy of the visible (VIS) and near infrared (NIR) spectroscopic measurement of other soil properties, for example, C, N, and other nutrients. This study was conducted to subtract the WC contribution to VIS-NIR spectra by classifying soil spectra into different WC groups. This classification might improve the accuracy of prediction of other soil properties with calibration models established separately for each group of WC. A mobile, fiber-type, VIS-NIR spectrophotometer (Zeiss Corona 1.7 visnir fiber), with a measurement range of 306.5 to 1710.9 nm was used to measure the light reflectance of two sample sets: one (275 samples) collected from a single field and the other (360 samples) collected from multiple fields in Belgium and northern France. The partial least squares (PLS) regression analysis and factorial discriminant analysis (FDA) were applied to the VIS-NIR spectra to quantify WC and classify spectra into different WC groups, respectively. Samples were divided into calibration and validation sets with ratios of 10:1 and 3:1 for the PLS and FDA, respectively. The PLS for the single-field sample set provided better estimation of WC (R2 = 0.98) than for the multiple-field sample set (R2 = 0.88). For the single-field sample set, spectra were successfully classified into six WC groups with correct classification (CC) of 94.1 and 95.6% for the calibration and validation datasets, respectively. Due to the large variability in the multiple-field sample set, soils were successfully classified into three WC groups only. The CC obtained were 88.1 and 79.7% for the calibration and validation sets, respectively. These results suggested that the FDA can be successfully used to classify soil VIS-NIR spectra into different WC levels, particularly when soil variability is minimal.
Abbreviations: CC, correct classification • FDA, factorial discriminant analysis • NIR, near infrared • PCA, principal component analysis • PC, principal component • PLS, partial least squares • VIS, visible • WC, water content
|
INTRODUCTION
|
|---|
SOIL WC is one of the most critical soil components for successful plant growth and land management, particularly in arid lands. Measurement of soil WC can be very beneficial for site-specific irrigation, seeding, and land management. The conventional method to determine WC by oven drying of samples collected from fields is a difficult, costly, and time-consuming procedure. The VIS and NIR spectroscopy is a proven technique for the measurement of soil WC, as it is fast, nondestructive, and cost effective, although a VIS-NIR instrument is expensive. This VIS-NIR measurement of soil WC includes offline laboratory (Bowers and Hanks, 1965; Skidmore et al., 1975; Kano et al., 1985; Dalal and Henry, 1986; Slaughter et al., 2001; Lobell and Asner, 2002; Whiting et al., 2004) and on-the-go field conditions (Mouazen et al., 2005; Shibusawa et al., 2005). Although WC can be successfully measured with VIS-NIR spectroscopy, it is considered as one of the most critical factors affecting the accuracy of VIS-NIR models developed for the determination of other soil properties.
Modifications for removing the influence of WC on the accuracy of VIS-NIR measurement of soil properties are needed because WC is an important cause of changes in the shape of soil spectra. Such a change is attributed to modification of soil color and water absorbance bands with changing WC level. The increase in reflectance (or decrease in absorbance) is an indication of decreasing soil WC and vice versa. Smith et al. (1987) used the measured total water potential to increase the accuracy of measurement of organic matter with a spectrophotometer. Shonk and Gaultney (1988) reported that WC and surface preparation significantly affected their real-time soil organic matter sensor output, recommending that calibration should be done for similar conditions encountered during real-time measurement. Lobell and Asner (2002) noticed a great reduction in the strength of the important absorption feature for N and C at a wavelength of 2200 nm in mineral soils with increasing WC. Using hyperspectral images of soils, Kooistra et al. (2003) reported that WC had a negative effect on prediction capabilities for organic matter and clay content. This implies that the effect of WC on the shape of spectra should be removed in order to improve the accuracy of quantifying soil properties (Whiting et al., 2004). One option to remove the effect of WC is to construct dry soil spectra from wet spectra, as adopted by Bogrekci and Lee (2005a) to improve the accuracy of P prediction. Others developed quantitative models of soil properties based on scanning dry soil samples (Ben-Dor and Banin, 1995; Chang et al., 2001; Walvoort and McBratney, 2001; Cozzolino and Morón, 2005; Bogrekci and Lee, 2005b). Calibrations based on dry scanning are only useful for laboratory measurement. For in situ and on-the-go measurements, calibration models should be developed based on scanning wet soils. Alternative methods are needed to minimize the effect of WC without the need for scanning dry soils or for spectra transformations that can destroy the original shape of the spectra, leading to the loss of important information on other components to be quantified. As a first step, spectra can be classified into a few classes corresponding to a specific number of soil WC groups, using a qualitative analysis technique. Principal component analysis has been used for different applications to classify soil spectra into different groups (Stenbergh et al., 1995; Leone and Sommer, 2000; Chang et al., 2001). The PCA alone is insufficient, however, since it is a descriptive and not a predictive technique that can classify new individuals into prior established groups. Therefore, it is essential to use a more advanced technique such as the combination of PCA and FDA to establish classification models of WC. A new measured spectrum could be classified into a WC group using the classification models developed. After classification of the new spectrum is done, quantitative determination of other soil properties could be performed using quantitative calibration models developed separately for each WC group. These group-wise quantitative calibration models are expected to provide more accurate determination of other soil properties. This approach to VIS-NIR calibration is particularly important for in situ and on-the-go measurement of soil properties.
The scope of this study was the use of the combination of PCA and FDA for classifying VIS-NIR soil spectra into different WC groups, using samples collected from a single field and multiple fields covering large areas in Belgium and northern France. The secondary objective of the study was to evaluate the accuracy of VIS-NIR PLS models for quantifying WC for the single-field and multiple-field sample sets.
|
MATERIALS AND METHODS
|
|---|
Soil Samples
Two sets of soil samples were used. The first sample (single-field) set was collected in 2003 from a relatively uniform field (70000 m2) in the Zoutleeuw region of Belgium. The second sample (multiple-field) set was collected by the Soil Service of Belgium (Haverlee, Belgium) during the spring and summer of 2004 from multiple fields in Belgium and northern France. A total of 46 soil samples was collected from the single field from the upper layer of 15 to 20 cm during the wheat (Triticum aestivum L.) growing season, based on a 50 by 50 m grid. Soil samples were dried for 24 h at 60°C, after which they were ground and sieved with a 2-mm sieve (Slaughter et al., 2001). The multiple-field samples were also collected from the upper soil layer (15–25 cm) during a relatively long period of time (5 mo), which resulted in 360 samples with a wide range in WC. These samples were stored wet in plastic bags at 4°C until measurement with a VIS-NIR spectrophotometer.
The soil of the single field is an Arenic Cambisol, according to the FAO (Food And Agriculture Organization) classification, with a sandy loam texture dominating in the field. The texture of the multiple-field sample set determined in a sensory way (White, 1997) indicated that the set included soils with a wide range of texture (Table 1) compared with the limited variations for the single-field sample set. Using the Munsell soil color chart, variation in color was also found among the multiple-field samples. Under dry measurement conditions, all 360 samples belong to the 2.5Y or 10YR hue.
Spectrophotometer and Scanning
A mobile, fiber-type VIS-NIR spectrophotometer developed by Zeiss Company (Zeiss Corona 45 visnir fiber, Germany) was used. It is fast, precise, and robust, without moving parts, which makes it suitable for alignment on mobile machines to measure WC on the go (Mouazen et al., 2005). In addition to the InGaAs diode array for measurement in the NIR region (944.5–1710.9 nm), a Si array is available for measurement in the VIS and short infrared wavelength region (306.5–1135.5 nm). The light source is a 20 W tungsten halogen lamp illuminating the targeted soil surface, which is in direct contact with the optical measurement unit. Within the optical unit, the light illumination and reflectance fibers are collected together at a 45° angle.
From each sample of the single-field sample set, 35 g of sieved soil was packed in a 1.5-cm-deep, 6-cm-diameter plastic cup. Six WC levels were established by adding water (by weight) to the soil in the plastic cups. The six WC levels were labeled Group 1 (0.0–0.025 kg kg–1), 5 (0.026–0.075 kg kg–1), 10 (0.076–0.125 kg kg–1), 15 (0.126–0.175 kg kg–1), 20 (0.176–0.225 kg kg–1), and 25 (0.226–0.275 kg kg–1). No further WC beyond 0.26 kg kg–1 was considered because the VIS-NIR spectroscopy is more sensitive to WC differences at the dry end. The selected WC range of 0.005 to 0.26 kg kg–1 (Table 2) covers those most required for plant activities during the cropping season. The soil samples were scanned with the VIS-NIR spectrophotometer when the equilibrium WC was reached 24 h after adding water. After completing the scanning of a WC level, a new amount of water was added to the soil samples to be scanned after another 24 h. The sample surface was carefully leveled before scanning to increase light reflectance. A total of 275 scans was performed, resulting from six WC for each of the 46 samples (one sample from Group 25 was lost).
View this table:
[in this window]
[in a new window]
|
Table 2. Comparison of sample statistics of water content between the single-field and multiple-field sample sets.
|
|
For the multiple-field sample set, different amounts of fresh soil according to different textures were packed in 1.0-cm-deep, 3.6-cm-diameter plastic cups. No pretreatment was done before scanning except removing plant roots. Soil in the cup was first shaken and a gentle pressure was applied on the surface before the surface was carefully leveled to enhance light reflectance. After scanning, the WC was determined with the oven-drying method (105°C for 24 h). The 360 samples of the multiple-field sample set were classified into five different WC groups, labeled Group 5 (0.0–0.060 kg kg–1), 10 (0.061–0.120 kg kg–1), 15 (0.121–0.180 kg kg–1), 20 (0.181–0.240 kg kg–1), and 25 (0.241–0.400 kg kg–1).
For both sample sets, three reflectance readings were taken from each soil specimen at three different spots by rotating the plastic cups 120°. Each spectrum was an average of five successive spectra measured for 2.5 s. An overall average spectrum, representing 15 spectra, was then obtained from the three measured spectra to be considered for spectra pretreatment and model establishment.
Principal Component Analysis
The PCA was applied on the VIS-NIR spectra recorded on soil samples to extract information on the different soil WC groups. The PCA transforms the original independent variables (wavelengths) into new axes, or principal components (PCs). These PCs are orthogonal, so that the datasets presented on these axes are uncorrelated with each other (Jolliffe, 1986; Martens and Naes, 1989). Therefore, the PCA expresses the total variation in the dataset in only a few PCs and each successively derived PC expresses decreasing amounts of the variance. The first PC covers as much of the variation in the data as possible. The second PC is orthogonal to the first PC and covers as much of the remaining variation as possible, and so on. By plotting the PCs, one can view interrelationships between different variables, and detect and interpret sample patterns, groupings, similarities, or differences. The spectral patterns corresponding to the PCs provide information about the characteristic peaks, which are the most significant ones for discriminating soil samples according to a property. While similarity maps allow comparison of the spectra in such a way that two neighboring points represent two similar spectra, the spectral patterns exhibit the absorption bands that explain the similarities observed on the maps.
Before the PCA was performed, VIS-NIR spectra were first normalized to remove the scattering effect. For each spectrum, the reflectance at each wavelength was divided by the sum of the reflectance values of the entire wavelengths, which resulted in reducing the area under each spectrum to a value of 1 (Bertrand and Scotter, 1992), as shown in Fig. 1. This normalization produced mainly a shift in the spectrum peak maximum and peak width. After normalization, the PCA was performed on the normalized VIS-NIR spectra to investigate differences between soil WC groups. A total of 148 wavelength points (variables) was included in the PCA. The PCA was performed using StatBoxPro software, Version 5 (Karoui and Dufour, 2003).
View larger version (20K):
[in this window]
[in a new window]
|
Fig. 1. Average visible and near infrared spectra recorded on soil samples of the single-field dataset for Groups 1 (0.0–0.025 kg kg–1), 5 (0.026–0.075 kg kg–1), 10 (0.076–0.125 kg kg–1), 15 (0.126–0.175 kg kg–1), 20 (0.176–0.225 kg kg–1), and 25 (0.226–0.275 kg kg–1) (a) before normalization, and (b) after normalization (normalization adopted before principal component analysis).
|
|
Factorial Discriminant Analysis
The aim of the FDA analytical technique was to predict the membership of an individual soil sample following the definition of different WC classes created according to the WC of the soils. The FDA is a multivariate method that allows the testing of hypotheses (Le Bart et al., 1977). This method can show the presence of certain relationships between a qualitative explanatory criterion and a group of quantitative explanatory characters, and it allows one to describe these relationships. The extraction of a qualitative variable within a population allows the division of this population into different groups, with each individual assigned to one group. Discrimination of the groups consists in maximizing the variance between their centers of gravity; one can then clarify the properties that distinguish the different groups. If the individual is close to the center of gravity of its group, it is "correctly classified." In the case where the distance to the center of gravity of its group is larger than the distance to the center of gravity of another group, the individual is "poorly classified" and it will be reassigned to this other group.
The FDA was performed on the first five PCs resulting from the PCA of the VIS-NIR spectral data recorded from the soil samples. Considering the first five PCs only was justified by the fact that they cover most of the variation (>99% of the total variance) contained in the raw data. The VIS-NIR spectral collections were divided into two datasets for calibration and validation. Two-thirds of the samples were used for the calibration set and one-third for the validation set. Like the PCA, the FDA was performed using StatBoxPro software, Version 5 (Karoui and Dufour, 2003). The precision of results obtained by the FDA was evaluated using the correct classification (CC) index. The percentage of CC of any WC group (CCg) was calculated as follows:
 | [1] |
where CCng is the number of correctly classified samples out of the total number of samples (ng) of that group. The overall CC was calculated using the following formula:
 | [2] |
where CCn is the number of correctly classified samples out of the total number of samples (n) in the dataset.
Partial Least Squares Regression Analysis
The PLS regression analysis was used in this study to establish calibration models for quantifying WC. It is a bilinear modeling method where information in the original x data is projected onto a small number of underlying ("latent") variables called PLS components. The y data are actively used in estimating the "latent" variables to ensure that the first components are those that are most relevant for predicting the y variables. Interpretation of the relationship between x data and y data is then simplified as this relationship is concentrated on the smallest possible number of components. More detailed information about the PLS can be found in Martens and Naes (1989).
Figure 2 shows the pretreatment steps performed on five arbitrarily selected soil raw spectra (Fig. 2a). These pretreatment steps considered before PLS regression analysis consisted of the following successive steps: (i) spectra reduction to cut the noise at the two ends of the spectra (Fig. 2b); (ii) maximum normalization to arrange spectra in almost the same scale (Fig. 2c); (iii) application of the Savitzky–Golay first derivative through use of the second-order polynomial. A smoothing factor was included with the first derivation (Fig. 2d).
View larger version (13K):
[in this window]
[in a new window]
|
Fig. 2. Pretreatment steps of five arbitrarily selected soil visible and near infrared spectra considered before the partial least squares regression analysis: (a) raw spectra; (b) spectra reduction; (c) maximum normalization; (d) first derivative with the Savitzy–Golay method.
|
|
The spectra were first reduced from 306.5 to 1710.9 nm to 401.4 to 1699 nm to eliminate the noise at edges of each spectrum, which might reduce the accuracy of the WC model. More data points were removed from the low end of the spectra because of a lower signal-to-noise ratio at that end (Fig. 2b). These spectra were then normalized by maximum normalization. Normalization is typically used to get all data to approximately the same scale (Fig. 2c), or to get a more even distribution of the variances and the average values. Normalization considered for the PLS was different from the one considered for the PCA, as the two analyses are absolutely separated. The maximum normalization process is performed on a sample spectrum by dividing each reflection value by the sample maximum absolute reflection, as follows:
 | [3] |
where X'(i,k) is the maximum normalized spectrum, X(i,k) is the sample reflection (%) and X(i,*) is the sample maximum absolute reflection (%).
The maximum normalization is a normalization that "polarizes" the spectra. The peaks of all spectra with positive values touch 1, while spectra with values of both signs touch –1. Since all the soil spectra in this study have positive values, the peaks of these spectra touched 1, as shown in Fig. 2c.
After maximum normalization, soil spectra were subjected to the first derivative using the Savitzky–Golay method (Hruschka, 2001). The Savitzky–Golay method enables computation of first or higher order derivatives, including a smoothing factor that determines how many adjacent variables (wavelengths) will be used to estimate the polynomial approximation used for derivation. In this study, a second-order polynomial approximation was selected (Fig. 2d).
After all pretreatment steps were performed, the PLS regression analysis was used to develop quantitative models relating the variations in WC (y variable) to the variations in wavelengths (290 wavelengths, the x variables). It was performed separately on the single- and multiple-field sample sets to built two separate calibration models. In each sample set, spectra were divided into two groups for calibration and validation with a ratio of 10:1. The first group was used for the establishment of the PLS model, while the second group was used for its validation. Since successful establishment of the PLS model consists of modeling and validation procedures, the leave-one-out cross-validation method was used together with the PLS analysis. With leave-one-out cross validation, the same calibration samples were used for both model estimation and testing. One sample was left out from the calibration dataset and the model was calibrated on the remaining samples. Then the value for the left-out sample was predicted and the prediction residual was computed. The process was repeated with another sample of the calibration set, and so on until every object had been left out once; then all prediction residuals were combined to compute the validation residual variance and root mean square error of prediction. As the final step, the validation set was used to validate the PLS-cross validation developed WC models.
|
RESULTS AND DISCUSSION
|
|---|
Classification of Soils into Different Water Content Levels
The PCA applied on the VIS-NIR spectra of the soil samples provided descriptive information of the different soil groups. For the single-field sample set, the first two PCs accounted for 96.4% of the total variance, with PC1 accounting for 70.6% of the total variance. Figure 3a shows a clear classification of soil samples into six different WC groups. The groups located on the right side of the PC1, including those labeled 1, 5, and 10, were well separated, while a small overlap between groups on the left side can be observed, particularly between Groups 20 and 25. This means that when WC is >0.15 kg kg–1, the sensitivity of VIS-NIR spectra to variable WC decreases. The spectral patterns associated with the PCs provide the characteristic wavelengths that may be used to discriminate between spectra. For the single-field sample set, the Spectral Patterns 1 and 2 associated with the PC1 and PC2, respectively, are shown in Fig. 3b. Both spectral patterns indicate a significant wavelength of soil WC, which is associated with the water absorption band (1450 nm) in the second overtone region. The two spectral patterns show indirect, weak correlation between soil WC and spectra in the VIS wavelength range (400–750 nm). This indirect correlation is associated with color, since soil darkness changes with variable WC.
View larger version (19K):
[in this window]
[in a new window]
|
Fig. 3. (a) Principal component similarity maps determined for visible and near infrared spectra of the single-field dataset, showing Groups 1 (0.0–0.025 kg kg–1), 5 (0.026–0.075 kg kg–1), 10 (0.076–0.125 kg kg–1), 15 (0.126–0.175 kg kg–1), 20 (0.176–0.225 kg kg–1), and 25 (0.226–0.275 kg kg–1), and (b) spectral patterns SP1 and SP2 corresponding to the principal components PC1 and PC2, respectively.
|
|
For the multiple-field sample set, soil samples were classified into five groups only. Overlap between different groups appears clearly on the similarity map defined by the first two PCs, which accounted for 92.6% of the total variance (Fig. 4a). This can be attributed to the large variation in texture, color, and origin of the investigated soils, since they were collected from many fields spread across a large geographical region. Similar to the spectral patterns obtained from the single-field dataset, the spectral patterns of the multiple-field dataset exhibit a strong feature associated with the water absorption band (1450 nm) in the second overtone region. This wavelength is the most significant one in the wavelength range of 306.5 to 1710.9 nm for the discrimination of soil samples into different WC levels (Fig. 4b). A very limited effect of the VIS wavelength range on discrimination can be seen on Spectral Pattern 1.
View larger version (18K):
[in this window]
[in a new window]
|
Fig. 4. (a) Principal component similarity maps determined for visible and near infrared (VIS-NIR) spectra of the multiple-field dataset, showing Groups 5 (0.0–0.060 kg kg–1), 10 (0.061–0.120 kg kg–1), 15 (0.121–0.180 kg kg–1), 20 (0.181–0.240 kg kg–1), and 25 (0.241–0.400 kg kg–1), and (b) spectral patterns SP1 and SP2 corresponding to the principal components PC1 and PC2, respectively.
|
|
The single- and multiple-field samples sets were classified into six and five WC groups using the FDA. For the single-field sample set, the map defined by the discriminant factors F1 and F2 accounted for 99.7% of the total variance, with F1 accounting for 96.8% (Fig. 5). Considering the F1, the soil groups labeled 1, 5, and 10 are located on the right side, while the remaining three groups are observed on the left side. The F2 discriminates soil groups labeled 1 and 25 with positive scores from the groups labeled 10 and 15 with mostly negative scores. For the multiple-field sample set, the map defined by the F1 and F2 takes into account 97.9% of the total variance, with F1 accounting for 81.5%, as shown in Fig. 6. A clear overlap can be observed on the map, particularly between adjacent soil groups 5 and 10 and between adjacent groups 15 and 20.
View larger version (14K):
[in this window]
[in a new window]
|
Fig. 5. Discriminant analysis similarity map determined by discriminant factors F1 and F2 for calibration set of the single-field soil samples, showing Groups 1 (0.0–0.025 kg kg–1), 5 (0.026–0.075 kg kg–1), 10 (0.076–0.125 kg kg–1), 15 (0.126–0.175 kg kg–1), 20 (0.176–0.225 kg kg–1), and 25 (0.226–0.275 kg kg–1).
|
|
View larger version (15K):
[in this window]
[in a new window]
|
Fig. 6. Discriminant analysis similarity map determined by discriminant factors F1 and F2 for calibration set of the multiple-field soil samples, showing Groups 5 (0.0–0.060 kg kg–1), 10 (0.061–0.120 kg kg–1), 15 (0.121–0.180 kg kg–1), 20 (0.181–0.240 kg kg–1), and 25 (0.241–0.400 kg kg–1).
|
|
Tables 3 and 4 provide the CC of the calibration and validation datasets for the single- and multiple-field sample sets, respectively. Considering six WC levels for the single-field set, CC of 94.1 and 95.6% is observed for the calibration and validation samples, respectively (Table 3). For the multiple-field sample set with five WC levels, CC of 69.8 and 62.7% is observed for the calibration and validation samples, respectively (Table 4). For the calibration samples, the classification of the Group 25 is successfully established compared with the less successful classification of the other four groups (Table 4). The worst classification is observed for Groups 10 and 20, with CC of 62.5% and 69%, respectively. This suggests reducing the number of soil WC groups to three instead of five groups. Samples of Groups 5 and 10 and Groups 15 and 20 were combined to obtain three groups labeled 8 (0.0–0.120 kg kg–1), 18 (0.121–0.240 kg kg–1), and 25 (0.241–0.400 kg kg–1), respectively. Indeed, the classification into three groups of dry (Group 8), moderately wet (Group 18), and wet (Group 25) led to improvement in the results obtained from the FDA analysis (Table 5). Comparing the overall CC of five WC groups (Table 4) with three WC groups (Table 5) indicates considerable improvement in the classification accuracy for both the calibration and validation datasets; however, the CC of Group 25 in the validation set is very poor due to the smaller number of samples considered compared with Groups 8 and 18. This suggests that the FDA did not allow for a valuable amount of CC for all groups of the multiple-field sample set, since a large diversity in color, texture, and origin of soils exists among samples collected from a large geographical area; however, the results obtained are still promising for the investigated soil WC levels.
View this table:
[in this window]
[in a new window]
|
Table 3. Classification of the single-field dataset into six water-content groups: 1 (0.0–0.025 kg kg–1), 5 (0.026–0.075 kg kg–1), 10 (0.076–0.125 kg kg–1), 15 (0.126–0.175 kg kg–1), 20 (0.176–0.225 kg kg–1), and 25 (0.226–0.275 kg kg–1) using factorial discriminate analysis.
|
|
View this table:
[in this window]
[in a new window]
|
Table 4. Classification of the multiple-field dataset into five water-content groups: 5 (0.0–0.060 kg kg–1), 10 (0.061–0.120 kg kg–1), 15 (0.121–0.180 kg kg–1), 20 (0.181–0.240 kg kg–1), and 25 (0.241–0.400 kg kg–1) using factorial discriminate analysis.
|
|
View this table:
[in this window]
[in a new window]
|
Table 5. Classification of the multiple-field dataset into three water content groups: 8 (0.0–0.120 kg kg–1), 18 (0.121–0.240 kg kg–1), and 25 (0.241–0.40 kg kg–1) using factorial discriminant analysis.
|
|
Quantitative Estimation of Water Content
Quantifying soil WC with the NIR spectroscopy relies on existing overtones of water absorption bands at 1450, 1950, and 2950 nm (Plamer and Williams, 1974), which result from resonance in the molecular vibration of the illuminated water molecule (Whalley and Stafford, 1992). Since the spectrophotometer used in this study covers the wavelength range of 306.5 to 1710.9 nm, only one water absorption band of 1450 nm in the second overtone region has significant influence on the prediction accuracy of the PLS-developed WC models. The average spectra of the single-field dataset indicate decreases in reflectance (Fig. 1a) and increases in the absorption depth at 1450 nm (Fig. 1b) with increasing WC. The PLS cross-validation model of WC provides good estimation for both calibration and validation datasets (Table 6); however, the limited number of soil textures and minimal variation in color of the single-field sample set makes the prediction of WC much better than that of the multiple-field sample set (Table 6). This is mainly due to soil texture, which has an important effect on the accuracy of VIS-NIR measurement of soil properties similar to the effect of WC. Indeed, Krishnan et al. (1980) found that large particle sizes (<1.6 mm) may at least partly account for the failure to predict organic matter by reflectance technique in the NIR. Dalal and Henry (1986) experienced a much higher standard error of prediction of WC, organic C, and total N in coarsely ground (<2 mm) than in finely ground (<0.25 mm) soil samples; however, other properties that affect the color of soil (organic matter, Fe oxides, etc.) could also influence the measurement accuracy of WC with VIS-NIR spectroscopy. The higher accuracy of quantitative prediction of WC for the single-field sample set than the multiple-field set is in line with the classification results, showing more successful quantification of WC for the single-field than the multiple-field set.
View this table:
[in this window]
[in a new window]
|
Table 6. Validation results of the partial least squares cross-validation model for quantifying soil water content.
|
|
|
CONCLUSIONS
|
|---|
The FDA and PLS regression analysis were performed on VIS-NIR spectra (306.5– 1710.9 nm) to provide qualitative and quantitative information, respectively, about soil WC. Results showed that both the quantitative and qualitative evaluation of WC can be successfully performed for soil samples collected from a single field with limited variation in texture and color. Due to variable color, texture, and origin of soil samples for the multiple-field sample set, less successful quantitative and qualitative results were obtained; however, the quality of the quantitative calibration model of WC is good enough (R2 = 0.88) to be recommended for future prediction of WC for new soil samples collected from any field in Belgium and northern France. The classification of soil spectra from the multiple-field samples could be successfully done for a smaller number of WC groups (three groups) compared with the larger number of groups (six groups) obtained for the single-field sample set. With only 360 samples in the multiple-field set, the current model is not yet very robust for successful classification of soils into different WC groups. More research is needed to include other samples, aiming at obtaining a uniform sample distribution of all WC groups that can increase the classification accuracy. When a classification system of soil spectra into separate WC groups is achieved, groupwise separate quantitative models are expected to increase the accuracy of VIS-NIR measurement for other soil properties (C, N, etc.). This calibration system of VIS-NIR spectroscopy (classification of spectra to WC groups followed by quantitative prediction of soil properties) should be automated to facilitate the in situ or on-the-go measurement of soil properties. Furthermore, it is worthwhile to explore the potential of the classification methodology (PCA and FDA) adopted in this study in the domain of soil type classification. The inclusion of other properties such as texture, color, organic matter, pH, and Fe oxides are recommended for a successful classification system.
|
ACKNOWLEDGMENTS
|
|---|
We gratefully acknowledge Mr. Jan Bries and Mr. Danny Van de Put at the Soil Service of Belgium for analyzing and sharing the soil samples. We also acknowledge the financial support of the IWT-Flanders (Projects no. IWT/20711 and IWT/30836) as well as the financial support of the research council of the K.U. Leuven.
Received for publication September 8, 2005.
|
REFERENCES
|
|---|
- Ben-Dor, E., and A. Banin. 1995. Near-infrared analysis as a rapid method to simultaneously evaluate several soil properties. Soil Sci. Soc. Am. J. 59:364–372.[Abstract/Free Full Text]
- Bertrand, D., and C.N.G. Scotter. 1992. Application of multivariate analyses to NIR spectra of gelatinized starch. Appl. Spectrosc. 46:1420–1425.[CrossRef]
- Bogrekci, I., and W.S. Lee. 2005a. The effects of water content on reflectance spectra of soils using UV-VIS-NIR spectroscopy. p. 1307–1317. In Proc. Int. Conf. on Precision Agric., 7th, Minneapolis, MN. 25–28 July 2004. Univ. of Minnesota, Minneapolis.
- Bogrekci, I., and W.S. Lee. 2005b. Spectral phosphorus mapping using diffuse reflectance of soils and grass. Biosyst. Eng. 91:305–312.[CrossRef]
- Bowers, S.A., and R.J. Hanks. 1965. Reflection of radiant energy from soils. Soil Sci. 100:130–138.
- Chang, C.-W., D.A. Laird, M.J. Mausbach, and C.R. Hurburgh. 2001. Near-infrared reflectance spectroscopy–principal components regression analyses of soil properties. Soil Sci. Soc. Am. J. 65:480–490.[Abstract/Free Full Text]
- Cozzolino, D., and A. Morón. 2005. Potential of near-infrared reflectance spectroscopy and chemometrics to predict soil organic carbon fractions. Soil Tillage Res. 85:78–85.[CrossRef]
- Dalal, R.C., and R.J. Henry. 1986. Simultaneous determination of moisture, organic carbon, and total nitrogen by near infrared reflectance spectrophotometer. Soil Sci. Soc. Am. J. 50:120–123.[Abstract/Free Full Text]
- Deckers, J. 1996. Bodemgeografie. (Soil geography.) J. Decker, Inst. of Land and Water Manage., Katholieke Univ., Leuven, Belgium.
- Hruschka, W.R. 2001. Data analysis: Wavelength selection methods. p. 39–58. In P. Williams, and K. Norris (ed.) Near-infrared technology in agriculture and food industries. Am. Assoc. Cereal Chemists, St. Paul, MN.
- Jolliffe, I.T. 1986. Principal component analysis. Springer, New York.
- Kano, Y., W.F. McClure, and R.W. Skaggs. 1985. A near infrared reflectance soil moisture meter. Trans. ASAE 28:1852–1855.
- Karoui, R., and E. Dufour. 2003. Dynamic testing rheology and fluorescence spectroscopy investigations of surface to center differences in ripened soft cheeses. Int. Dairy J. 13:973–985.[CrossRef]
- Kooistra, L., J. Wanders, G.F. Epema, R.S.E.W. Leuven, R. Wehrens, and L.M.C. Buydens. 2003. The potential of field spectroscopy for the assessment of sediment properties in river floodplains. Anal. Chim. Acta 484:189–200.[CrossRef]
- Kirshnan, P., J.D. Alexander, B.J. Butler, and J.W. Hummel. 1980. Reflectance technique for predicting soil organic matter. Soil Sci. Soc. Am. J. 44:1282–1285.[Abstract/Free Full Text]
- Le Bart, L., A. Morineau, and N. Tabard. 1977. Technique de la description statistiques, methodes et logiciel pour l'analyse des grand tableaux. (In French.) Dunod, Paris.
- Leone, A.P., and S. Sommer. 2000. Multivariate analysis of laboratory spectra for the assessment of soil development and soil degradation in the southern Apennines (Italy). Remote Sens. Environ. 72:346–359.[CrossRef]
- Lobell, D.B., and G. Asner. 2002. Moisture effect on soil reflectance. Soil Sci. Soc. Am. J. 66:722–725.[Abstract/Free Full Text]
- Martens, H., and T. Naes. 1989. Multivariate calibration. 2nd ed. John Wiley & Sons, Chichester, UK.
- Mouazen, A.M., J. De Baerdemaeker, and H. Ramon. 2005. Towards development of on-line soil moisture sensor using a fiber-type NIR spectrophotometer. Soil Tillage Res. 80:171–183.[CrossRef]
- Plamer, K.F., and D. Williams. 1974. Optical properties of water in the near infrared. J. Opt. Soc. Am. 64:1107–1110.
- Shibusawa, S., K. Ehara, T. Okayama, H. Umeda, and S. Hirako. 2005. A real-time multi-spectral soil sensor: Predictability of soil moisture and organic matter content in a small field. p. 495–502. In J.V. Stafford (ed.) Precision agriculture '05. Wageningen Acad. Publ., Wageningen, the Netherlands.
- Shonk, J.L., and L.D. Gaultney. 1988. Spectroscopic sensing for the determination of organic matter content. ASAE Pap. no. 88–2142. ASAE, St. Joseph, MI.
- Skidmore, E.L., J.D. Dickerson, and H. Shimmelpfennig. 1975. Evaluation surface-soil water content by measuring reflectance. Soil Sci. Soc. Am. J. 39:238–242.[Abstract/Free Full Text]
- Slaughter, D.C., M.G. Pelletier, and S.K. Upadhyaya. 2001. Sensing soil moisture using NIR spectroscopy. Appl. Eng. Agric. 17:241–247.
- Smith, D.L., C.R. Womer, and J.W. Hummel. 1987. Soil spectra reflectance relationship to organic matter content. ASAE Pap. no. 87–1608. ASAE, St. Joseph, MI.
- Stenbergh, B., E. Nordkvist, and L. Salomonsson. 1995. Use of near infrared reflectance spectra of soils for objective selection of samples. Soil Sci. 159:109–114.
- Walvoort, D.J.J., and A.B. McBratney. 2001. Diffuse reflectance spectrometry as a proximal sensing tool for precision agriculture. p. 503–507. In G. Grenier and S. Blackmore (ed.) ECPA 2001, Proc. European Conf. Precision Agric., 3rd, Montpellier, France. Agro Montpellier, Montpellier, France.
- Whalley, W.R., and J.V. Stafford. 1992. Real-time sensing of soil water content from mobile machinery: Options for sensor design. Comput. Electron. Agric. 7:269–284.
- White, R.E. 1997. Principles and practice of soil science: The soil as a nature resource. 3rd ed. Blackwell Publ., Oxford, UK.
- Whiting, M.L., L. Li, and S.L. Ustin. 2004. Predicting water content using Gausian model on soil spectra. Remote Sens. Environ. 89:535–552.[CrossRef]
TOP
ABSTRACT
INTRODUCTION
