SPECTRAL MEASUREMENT OF COMMON SOIL PHOSPHATES I. Bogrekci, W. S. Lee
ABSTRACT. Compounds of Al, Fe, Ca, and Mg phosphates were particularly found in the Lake Okeechobee drainage basin in Florida. Identification of the spectral characteristics of these compounds would improve sensing of phosphorus (P) concentration in soil samples. This study investigated the effects of common soil P compounds on reflectance spectra of sandy soils using ultraviolet (UV), visible (VIS), and near-infrared (NIR) reflectance spectroscopy. Pure sandy soil was leached to remove all nutrients and organic matter. Effects of four different P compounds (CaPO4 , AlPO4 , FePO4 2H2 O, and Mg3 (PO4 )2 2H2 O) at 0.0, 12.5, 62.5, 175.0, 375.0, 750.0, and 1000.0 mg kg −1 were investigated. Actual P concentrations of the soil samples were analyzed. Reflectance of the samples was measured between 175 and 2550 nm with 1 nm intervals. The highest absorbance peaks were found at 286, 2548, 2516, and 228 nm for FePO4 2H2 O, Mg3 (PO4 )2 2H2 O, CaPO4 , and AlPO4 , respectively. These wavelengths may be used for a calibration to detect or determine phosphates of soils. Correlation coefficients, standard deviations, and first derivatives of absorbance spectra were computed. Wavelengths were selected using a stepwise discriminant analysis to build calibration models for P prediction. Classification of P compounds was tested using discriminant analysis. The results indicated that the Fe, Ca, Mg, and Al associated phosphates could be detected with a classification error of 1.9%. Partial least squares (PLS) analysis results for the dry soil samples with four compounds yielded R 2 of 0.48 to 0.75 and root mean square error (RMSE) of 27 to 43 mg/kg. Keywords. Lake Okeechobee, NIR, P compounds, Phosphates, Phosphorus, Reflectance, Sensor, Spectroscopy, UV, VIS.
P
hosphorus (P) is commonly found in nature as phosphates. Corbridge (1990) stated that P chemistry is dominated by compounds containing phosphorusoxygen linkages and these are termed as phosphates. Nair et al. (1995) investigated the P forms in soil profiles from dairies in south Florida. They stated that most P derived from the Ca-Mg associated phosphates for the A horizon. Brookes et al. (1982) measured P from microbial biomass in soils. They found that most of the P released was in inorganic form, and the proportion increased with the duration of fumigation. Tyler (2002) studied the P fractions in grassland soils. Total mineral (inorganic) P was calculated as the sum of P-Ca, P-Al, and P-Fe. In addition, total P was calculated as the sum of mineral P and organic P. Tyler (2002) studied the phosphorus fractions in grassland soils. He mentioned that results of soil P fractionations are often related to general soil chemical properties, such as soil acidity or weathering state. Calcium phosphate fraction was high in soils of low acidity, and the concentration of iron (III) phosphate dihydrate was high in very acid soils. Advances in spectroscopy have provided new methods to determine concentration of elements in chemistry. Therefore, it may be possible to use UV-VIS-NIR spectroscopy for
Article was submitted for review in October 2004; approved for publication by the Information & Electrical Technologies Division of ASABE in September 2005. Presented at the 2004 ASAE Annual Meeting as Paper No. 043114. The authors are Ismail Bogrekci, ASABE Member Engineer, Postdoctoral Research Associate, and Won Suk Lee, ASABE Member Engineer, Assistant Professor, Department of Agricultural and Biological Engineering, University of Florida, Gainesville, Florida. Corresponding author: I. Bogrekci, Frazier Rogers Hall, Museum Rd., Gainesville, FL 32611-0570; phone: 352-392-1864; fax: 352-392-4092; e-mail:
[email protected].
sensing different phosphate molecules. Janik et al. (1998) studied the diffuse reflectance analysis in comparison with soil extractions. They stated that recent developments in infrared spectroscopy resulted in an increase in the potential for soil analysis. They also mentioned that infrared spectroscopy in both the near- and mid-infrared ranges allowed rapid acquisition of soil information at quantitative, qualitative, or indicator levels for use in agricultural and environmental monitoring. Nyquist and Kagel (1971) investigated the infrared spectra of inorganic compounds. In particular, they listed characteristic frequencies for inorganic ions such as phosphide hypophosphite, orthophosphite, metaphosphite, orthophosphates, pyrophosphate, phosphorothioate, phosphorofluoridate, and phosphorodifluoridate. Dunn et al. (2002) investigated the potential of near-infrared reflectance spectroscopy for soil analysis in southeastern Australia. They suggested that near-infrared spectros− copy could successfully determine some soil properties in both the topsoil and subsoil. Russell et al. (2002) studied the potential of near-infrared (NIR) spectroscopy to predict nitrogen mineralization in rice soils. Lee et al. (2003) conducted research to estimate chemical properties of Florida soils using VIS and NIR spectroscopy. They produced prediction models to estimate pH, organic matter (OM), phosphorus (P), potassium (K), calcium (Ca), and magnesium (Mg) concentrations of three representative soil orders such as Alfisol, Entisol, and Ultisol. Bogrekci et al. (2003) studied the wet and dry reflectance spectra of soil samples collected from Okeechobee County in Florida and correlated them with P concentrations of soil samples in order to produce a prediction model for P in soils. Currently, there are many chemical extraction methods in determining the concentrations of soil P for various soil types (Bray and Kurtz, 1945; Nelson et al., 1953; Olsen et al., 1954; Carter, 1993; Mehlich, 1984; Pierzynski, 2000). The imple-
Transactions of the ASAE Vol. 48(6): 2371−2378
E 2005 American Society of Agricultural Engineers ISSN 0001−2351
2371
mentation of these extraction methods is laborious, costly, and time consuming. Therefore, the success of this investigation into using spectroscopy to determine different P compounds would be useful in designing a P sensor for agricultural and environmental use.
measurement, soil samples were oven-dried at 104°C for 24 h. The soil samples were sent to a laboratory for chemical analysis of P concentrations. All soil samples were analyzed for total P using the acid digestion method and inductively coupled plasma (ICP).
OBJECTIVE This study proposed first to detect the existence of soil phosphorus compounds (CaPO4, AlPO4, FePO42H2O, and Mg3(PO4)22H2O) commonly found in the Lake Okeechobee basin with the ultimate aim of designing a P sensor for that region, and second to determine the concentrations of each P compound. Specific objectives were to: S Identify significant absorption bands and spectral signature for each compound. S Classify each P compound in soils. S Quantify P concentration of each compound.
REFLECTANCE MEASUREMENT A spectrophotometer (Cary 500, Varian, Inc., Palo Alto, Cal.) equipped with an integrating sphere (DRA-CA-5500, Labsphere, Inc., North Sutton, N.H.) was used to collect spectral reflectance for each soil sample. The instrument had the wavelength accuracy of 0.020 nm for data interval and 0.20 nm for spectral bandwidth with an averaging time of 0.033 s. Photometric errors of the instrument were less than 0.0003% and 0.003% transmission in ultraviolet-visible and near-infrared, respectively. A 50 mm diameter polytetrafluoroethylene (PTFE) disk (Spectralon, Labsphere, Inc., North Sutton, N.H.) was used to obtain the optical standard of the system before spectral measurements. A 27.5 g soil sample was placed into a sample holder, and the spectral signature for each soil sample was collected using baseline correction mode. Reflectance for each soil sample with different P compounds was measured in 175-2550 nm with an increment of 1 nm. In addition, reflectance for each P compound (CaPO4, AlPO4, FePO42H2O, and Mg3(PO4)22H2O) was measured in 175-2550 nm with an increment of 1 nm. Since spectrum of each soil sample had noise in 175-225 nm and 2525-2550 nm, only the spectra of soil samples between 225 and 2525 nm were used. Reflectance of each compound had noise in 175-208 nm; therefore, reflectance data in 2082550 nm were used for P compounds’ signatures. Reflectance of the soil samples was measured before and after drying.
MATERIALS AND METHODS SOIL SAMPLE PREPARATION In order to study the effects of different P compounds on reflectance spectra of soils, pure sandy soil was donated by a company located in Edgar, Putnam County, Florida. Sandy soil was graded into three particle sizes using a sieve shaker (Ro-Tap, W. S. Tyler, Inc., Mentor, Ohio). Sandy soil particle sizes were 125, 250, and 600 m for fine, medium, and coarse, respectively. A total of 112 medium-size sandy soil samples were used in this study (table 1). Four replications of seven P application rates and four different P compounds generated 112 samples for this study. Soil samples were leached using 0.1 molar HCl acid solutions and de-ionized water in order to remove existing P. After leaching, pH and P concentration of the sandy soils were analyzed. Soil pH was measured using a pH/temperature meter (HI 991000, Hanna Instruments, Woonsocket, R.I.). The pH meter accuracy was ±0.01. Soil P was determined using a soil test kit (Luster Leaf Products, Inc., Atlanta, Fla.). Soils were leached with de-ionized water to obtain pH of 6, if pH was detected lower than 6. Further leaching was applied if P was detected in the soil samples. Different amounts of compounds were mixed into the soil thoroughly, and the soil samples were wetted to field capacity level (8% w.b. moisture content for Okeechobee county soils in Florida). These compounds were CaPO4 (calcium phosphate), AlPO4 (aluminum phosphate), FePO42H2O (iron (III) phosphate dihydrate), Mg3(PO4)22H2O (magnesium phosphate dihydrate). All compounds were purchased in powder form (Sigma-Aldrich, Inc., Milwaukee, Wisc.). The compound rates are given in table 1. Soil samples were incubated for seven days with all P compounds. After reflectance
P Compounds
CaPO4 AlPO4 FePO42H2O Mg3(PO4)22H2O
2372
Particle Size (Μm)
250
DATA ANALYSIS Data Processing Reflectances of all samples were converted into absorbance before further analysis in order to find a relationship between P concentrations and absorbance of the samples at different wavelengths using Beer-Lambert’s law (Williams and Norris, 2001). Absorbance was calculated using following formula: Abs = Log(1/Ref)
where Abs is absorbance and Ref is reflectance. The data were filtered using the Savitzky-Golay polynomial convolution filter to remove the noise in the signal (Savitzky and Golay, 1964). Smoothing was controlled by the degree of polynomial (2) and the number of points (50). In order to find the significant bands for each compound, correlation coefficient and standard deviation spectra were
Table 1. Experimental treatments. Moisture Content Spectral Range (% w.b.) (nm) pH
8
(1)
6
175-2550
P Compounds Application Rates (mg kg−1) (four replications) No compounds Very low Low Medium High Very high Extremely high
0.0 12.5 62.5 175.0 375.0 750.0 1000.0
TRANSACTIONS OF THE ASAE
Table 2. Compound application rates and average actual total P concentration of soil samples. Compound Average Actual Application Rate Total P Concentration (mg kg−1) (mg kg−1) Compound
computed. Correlation coefficients were computed between absorbance and actual total P concentration of the soil samples. Higher correlation coefficients indicate that P concentrations of the samples could be estimated indirectly by measuring absorbance without any chemical analysis. In addition, in order to observe the relationships between absorbance and P compound concentration, first derivatives of absorbance were computed. Standard deviations of absorbance were computed for soil samples at four P compounds with different concentrations. Data were divided into two sets as calibration and validation. The calibration and validation data sets were obtained using simple random sampling. The calibration data set had 56 soil samples, and the validation data set had 53 soil samples. Three samples were discarded due to experimental error and outliers. The SAS STEPDISC procedure was used to select important wavelengths and to reduce the number of independent variables. The SAS REG and SAS PRINCOMP procedures were used to select and reduce the number of independent variables (SAS, 1999). The main aim in computing the correlation, standard deviation, and first derivative, and conducting stepwise discrimination, stepwise multiple linear regression, and principal component analysis was to reduce the number of wavelengths before model building. Statistical Analysis and Model Building In order to build robust prediction models, both calibration and validation data sets should be sampled using statistical sampling methods. Each data set should represent the whole population. During model building, the SAS DISCRIM procedure was used to classify each compound (SAS, 1999). Wavelengths selected after correlation, stepwise discrimination, stepwise multiple linear regression, and principal component analysis were used as independent variables in the calibration and validation data sets for discriminant analysis. The percentage error rate was computed as the number of misclassified samples over the total number of samples, multiplied by 100. The SAS PLS procedure was used to calibrate and predict P concentrations of soils using spectral data (SAS, 1999). The same number of calibration and validation samples was used as in the data processing step. The number of extracted factors was determined by cross-validation, that is, fitting the model to part of the data and minimizing the prediction error for the unfitted part. The predicted residual sum of squares (PRESS) was used to determine the number of factors. The NIPALS algorithm was used. For cross-validation, the split-sample validation method was used.
Magnesium phosphate dihydrate
0.0 12.5 62.5 175.0 375.0 750.0 1000.0
11 18 23 35 73 108 174
Iron (III) phosphate dihydrate
0.0 12.5 62.5 175.0 375.0 750.0 1000.0
21 15 23 36 48 103 139
Calcium phosphate
0.0 12.5 62.5 175.0 375.0 750.0 1000.0
14 16 22 32 58 112 153
Aluminum phosphate
0.0 12.5 62.5 175.0 375.0 750.0 1000.0
11 9 20 42 73 125 168
are shown in figure 1. Iron (III) phosphate dihydrate had the highest absorbance values in the UV and VIS ranges. The highest absorbance was at 286 nm for iron (III) phosphate dihydrate. Absorbance values for iron (III) phosphate dihydrate sharply decreased from 286 nm to 667 nm. Iron (III) phosphate dihydrate had three recognizable peaks at 867, 1464, and 1944 nm from visible to NIR. Absorbance values for iron (III) phosphate dihydrate increased steadily from 2152 to 2550 nm. Magnesium phosphate hydrate had lower absorbance values in the UV range. It showed almost the same 1.0 Iron (III) phosphate dihydrate Magnesium phosphate hydrate Calcium phosphate Aluminum phosphate
RESULTS AND DISCUSSION Compound application rates and average chemical analysis results are given in table 2. The highest P concentration for soil samples was 174 mg kg−1 and the lowest was 9 mg kg−1. Although the samples were leached to remove P concentration, reference method showed that soils still had P in the concentration range from 10 to 21 mg kg−1. Since the soil test kit showed no P while the reference method indicated low concentration of P in the samples, these testing methods seem to have different accuracies and sensitivities. Average absorbance spectra for calcium phosphate, aluminum phosphate, iron (III) phosphate dihydrate, and magnesium phosphate hydrate in the UV-VIS-NIR regions
Vol. 48(6): 2371−2378
Absorbance
0.8
0.6
0.4
0.2
0.0 500
1000
1500
2000
2500
Wavelength (nm)
Figure 1. Average absorbance of four replications of the four P compounds in 208-2550 nm.
2373
extremely high P compound concentrations for UV-VIS-NIR (225-2525 nm). There was a positive linear relationship between absorbance and P concentration of the wet sandy soil samples with iron (III) phosphate dihydrate from 225 to 365 nm. The lower the P concentration, the lower the absorbance was from 225 to 365 nm. The results from the reflectance measurement of the wet sandy soil with iron (III) phosphate dihydrate suggested that P concentrations of iron (III) phosphate dihydrate could be determined using absorbance in the UV region. Absorbance spectra for the wet sandy soils with calcium phosphate from zero to extremely high P compound concentrations in 225-2525 nm are given in figure 2c. There were three distinguishable peaks at 271, 1454, and 1936 nm. The peaks at 1454 and 1936 nm indicate the water content of the wet sandy soil with calcium phosphate. The average absorbance spectra for the wet sandy soil with aluminum phosphate from zero to extremely high P compound concentrations in 225-2525 nm are given in figure 2d. The peaks at 1454 and 1936 nm indicate the water content of the wet sandy soil with aluminum phosphate. For all wavelengths, extremely high concentration of aluminum phosphate in soil samples produced the highest absorbance values. On the other hand, zero concentration yielded lower absorbance values between 225 and 2525 nm. Figure 3a shows the average spectra for the dry sandy soil with magnesium phosphate hydrate from zero to extremely
absorbance from 387 to 1122 nm. Since magnesium phosphate hydrate contains water molecules, two peak absorbance values were detected in 1419 and 1938 nm. The highest absorbance values were observed between 1938 and 2550 nm. The calcium phosphate spectra showed three distinct peaks at 283, 1439, and 1948 nm. The highest absorbance was observed at 2516 nm for calcium phosphate. Aluminum phosphate showed lower absorbance values between 1374 and 2550 nm. The highest absorbance values in 228 nm were observed in the UV range for aluminum phosphate. The spectral measurement results for the wet sandy soil with magnesium phosphate hydrate at zero to extremely high P compound concentrations in the range of 225-2525 nm are shown in figure 2a. The wet sandy soil with magnesium phosphate hydrate resulted in four recognizable peaks for all P concentrations at 271, 1452, 1931, and 2206 nm. P concentration of the wet sandy soil with magnesium phosphate hydrate had a positive linear relationship with absorbance in the UV range. In particular, high P concentration of the wet sandy soil with magnesium phosphate hydrate caused the dilution of absorbance spectra to almost medium P concentrations in the NIR region. Water absorption bands were also very distinct at 1452 and 1931 nm. Figure 2b shows the average absorbance for the wet sandy soil samples with iron (III) phosphate dihydrate from zero to
1.0
1.0 No Compound Very low Low Medium High Very high Extremely High
0.6
0.8
Absorbance
Absorbance
0.8
0.4
0.2
0.6
0.4
0.2
0.0
0.0 500
1000
1500
2000
2500
500
Wavelength (nm)
1000
1500
(a)
2500
(b) 1.0
0.8
0.8
Absorbance
1.0
0.6
0.4
Absorbance
Absorbance
2000
Wavelength (nm)
0.2
0.0
0.6
0.4
0.2
0.0 500
1000
1500
Wavelength (nm)
(c)
2000
2500
500
1000
1500
2000
2500
Wavelength (nm)
(d)
Figure 2. Average absorbance of wet soil samples with different concentrations of: (a) magnesium phosphate hydrate, (b) iron (III) phosphate dihydrate, (c) calcium phosphate, and (d) aluminum phosphate in 225-2525 nm. Each spectrum is an average of four samples.
2374
TRANSACTIONS OF THE ASAE
high P compound concentrations in 225-2525 nm. The highest absorbance values occurred between the UV and VIS regions. Low peaks at 1414 and 1934 nm indicate traces of water in the dried soil sample. The other distinguishable peaks were at 269 and 2211 nm. The results point out that overall absorbance values of the dry sandy soil with magnesium phosphate hydrate are lower than those of the wet sandy soil with magnesium phosphate hydrate. Comparing the average absorbance spectra between the wet samples (fig. 2) and the dry samples (fig. 3), the spectral difference mainly comes from the water in the wet samples. Water is a strong absorber and increases the overall absorbance. The most dominant water absorption bands are at 1450 and 1940 nm (Hruschka, 2001). The wet spectra in figure 2 show higher absorbance than those of the dry samples in fig. 3. The wet spectra in figure 2 also show two main water absorption bands at 1450 and 1940 nm, which do not exist in the dry spectra in figure 3. However, the spectral characteristics in the UV region remain the same between the wet and dry samples. The average spectra for the dry sandy soil with iron (III) phosphate dihydrate from zero to extremely high P compound concentrations in 225-2525 nm are given in figure 3b. A positive linear relationship exists between absorbance and P compound concentrations of the dry sandy soil with iron (III) phosphate dihydrate, particularly in 225-490 nm. The results from the reflectance measurement of dry sandy soil with iron (III) phosphate dihydrate suggest that the con-
centration of the Fe-associated P could be determined by measuring absorbance in the UV region. Apart from the most significant peak at 281 nm, four small peaks occurred at 871, 1413, 1934, and 2213 nm. Average dry sandy soil spectra with calcium phosphate from zero to extremely high P compound concentrations in 225-2525 nm are presented in figure 3c. Eight small peaks were observed for the dry soil with calcium phosphate at 268, 875, 1415, 1932, 2212, 2314, 2382, and 2500 nm. The highest absorbance values were seen in the UV range. Absorbance values decreased sharply from 268 to 375 nm. Decreasing of absorbance values continued steadily from VIS to NIR. Figure 3d shows the average dry soil spectra with aluminum phosphate from zero to extremely high P compound concentrations in 225-2525 nm. Dry soil sample with aluminum phosphate absorbed more light energy in the UV and VIS ranges than in the NIR. The highest absorbance (Abs = 0.44) was observed in the UV range. Seven low-amplitude peaks in the NIR range and one high-amplitude peak in the UV range occurred. Correlation coefficient spectra for the wet sandy soils with four compounds with regard to P concentrations are presented in figure 4. The highest correlation coefficient (r = 0.84 at 317 nm) appeared in the UV range for iron (III) phosphate dihydrate, but correlation coefficients decreased in the VIS and NIR regions. Correlation coefficients for aluminum phosphate, calcium phosphate, and magnesium phosphate hydrate did not vary much about the average correlation 0.7
No compounds Very low Low Medium High Very high Extremely high
0.6 0.5
Absorbance
Absorbance
0.6
0.4
0.4 0.3 0.2
0.2
0.1 0.0
0.0 500
1000
1500
2000
2500
500
Wavelength (nm)
1000
(a)
2000
2500
2000
2500
(b)
0.7
0.7
0.6
0.6
0.5
0.5
Absorbance
Absorbance
1500
Wavelength (nm)
0.4 0.3
0.4 0.3
0.2
0.2
0.1
0.1
0.0
0.0 500
1000
1500
Wavelength (nm)
(c)
2000
2500
500
1000
1500
Wavelength (nm)
(d)
Figure 3. Average absorbance of dry soil samples with different concentrations of: (a) magnesium phosphate hydrate, (b) iron (III) phosphate dihydrate, (c) calcium phosphate, and (d) aluminum phosphate in 225-2525 nm. Each spectrum is an average of four samples.
Vol. 48(6): 2371−2378
2375
coefficients of 0.58, 0.21, and 0.52, respectively. The highest correlation coefficients of the wet sandy soil samples were 0.62 for aluminum phosphate at 1213 nm, 0.26 for calcium phosphate at 1384 nm, and 0.58 for magnesium phosphate hydrate at 831 nm. Results of the correlation between the absorbance and P concentrations for the wet sandy soils with four different P compounds suggested that the concentrations of Fe, Mg, and Al associated P in pure sand soil can be determined by developing calibration models for individual compounds. Figure 5 shows the correlation coefficient spectra for the dry sandy soils with four compounds with regard to P concentrations. Dry sandy soil with iron (III) phosphate dihydrate produced the highest correlation coefficient (r = 0.93 at 343 nm) in the UV range. The highest correlation coefficient between the wet and dry soils for iron (III) phosphate dihydrate noticeably shifted from 317 to 343 nm. This shift might be due to molecular structure change during water loss. Mean correlation coefficient of calcium phosphates for the dry soil increased from 0.21 to 0.41. The highest correlation coefficients of the dry sandy soil samples were 0.68 for aluminum phosphate at 2278 nm, 0.54 for calcium phosphate at 2500 nm, and 0.73 for magnesium phosphate hydrate at 2520 nm. 1.0
0.84@317
Aluminum phosphate Calcium phosphate Iron (III) phosphate dihydrate Magnesium phosphate hydrate
0.62@1213 0.6
0.30
0.4
Aluminum phosphate Calcium phosphate Iron (III) phosphate dihydrate Magnesium phosphate hydrate
0.25
0.58@831
Standard deviation
Correlation coefficient (r)
0.8
Close to the 2500 nm region, absorbance spectra (fig. 2) show positive linear relationship with P concentrations, with one or two exceptions. These exceptions caused lower correlation coefficients for the correlation coefficient spectra of the wet soils (fig. 4). When absorbance variation for iron (III) phosphate dihydrate in the UV region (figs. 2b and 3b) is compared with the correlation coefficient spectra (figs. 4 and 5), the absorbance spectra have a positive linear relationship with P concentrations without any exceptions. Therefore, values for the correlation coefficient are observed to be higher. The standard deviation spectra for the wet and dry soils with Fe, Mg, Al, and Ca associated P compounds are given in figure 6a and 6b. Standard deviations of absorbance for each sample with different concentrations and types of phosphates were computed in 225-2525 nm. Regions that have large positive peaks are regions where the spectra vary the most. Higher standard deviations indicate wavelength regions, which can be utilized effectively for developing calibration models since they contain more variability than other regions. The highest standard deviation was observed at 1922 nm for aluminum phosphate and at 274 nm for iron (III) phosphate dihydrate for the wet and dry sandy soils, respectively. There was a general trend of higher standard deviation values with increasing wavelength for regions more than 1000 nm in figure 6a. When absorbance variation for iron (III) phosphate dihydrate in the UV region (figs. 2b and 3b) is compared with the standard deviation spectra (figs. 6a and 6b), figures 2b
0.2
0.26@1384 0.0
−0.2 500
1000
1500
2000
0.20
0.15
0.10
2500 0.05
Wavelength (nm)
Figure 4. Correlation coefficient spectra between chemical results (P) and their absorbance for wet sandy soil samples with four phosphorus compounds.
0.00 500
1000
1500
2500
(a)
1.0
0.93@343
0.10
0.73@2520
0.8
Aluminum phosphate Calcium phosphate Iron (III) phosphate dihydrate Magnesium phosphate hydrate
0.08
0.6 0.4 0.2 0.0
0.68@2278
−0.2
0.54@2500
Aluminum phosphate Calcium phosphate Iron (III) phosphate dihydrate Magnesium phosphate hydrate
−0.4
Standard deviation
Correlation coefficient (r)
2000
Wavelength (nm)
0.06
0.04
0.02
0.00
−0.6
500 500
1000
1500
2000
2500
1000
1500
2000
2500
Wavelength (nm)
Wavelength (nm)
(b) Figure 5. Correlation coefficient spectra between chemical results (P) and their absorbance for dry sandy soil samples with four phosphorus compounds.
2376
Figure 6. Standard deviation spectra for (a) wet and (b) dry sand soil samples with four phosphates.
TRANSACTIONS OF THE ASAE
and 3b show the absorbance variation in the UV region, and figures 6a and 6b agree with the absorbance variation by yielding higher standard deviations. Derivatives are often used to enhance the appearance of spectra (Williams and Norris, 2001). The first derivatives of average absorbance of the dry soil samples with iron (III) phosphate dihydrate in 225-2525 nm are shown in figure 7a. Figure 7b shows the first derivative of average absorbance of the dry soil samples with iron (III) phosphate dihydrate in 225-450 nm. There were two higher-amplitude peaks at 257 and 354 nm. At 257 nm, the amplitudes of absorbance presented a direct positive linear relationship with iron (III) phosphate dihydrate rates, whereas at 354 nm, the amplitudes of absorbance produced a direct negative linear relationship with iron (III) phosphate dihydrate. Table 3 shows the classification results for the dry soil samples with four compounds using absorbance in selected wavelengths and considering the given criteria. Increasing the threshold level of the correlation coefficient (r = 0.4) and using the stepwise discriminant procedure reduced the number of selected wavelengths. However, this reduction in the number of wavelengths caused a high classification error (54.7%). Reducing the threshold level of the correlation coefficient (r = 0.3) and using the stepwise discriminant
CONCLUSIONS
0.002
Absorbance first derivative
procedure increased the number of selected wavelengths. Although the classification error was reduced to 15.1%, a lower classification error (1.9%) was observed when the correlation coefficient was equal to 0.35 and stepwise discriminant analysis was used. The results in table 3 indicated that when the number of independent variables (number of selected wavelengths) was high, the classification error increased, and when the number of independent variables was small, the classification error again increased. Fifteen was the optimum number of wavelengths and produced lower classification errors. Partial least squares analysis (PLS) results from the calibration and validation data sets for soils with four P compounds are presented in table 4. The number of factors was six. The results showed that PLS predicted P concentration of soils with root mean square error (RMSE) ranging from 27 to 43 mg kg−1. Prediction performance of iron phosphate seems to be reasonable, as the model validation produced R2 of 0.75 and RMSE of 27 mg kg−1. For the other three phosphates, there seems to be more room for improvement of the validation models, which yielded lower values of R2 and higher values of RMSE than those for iron phosphate. Ultimately, the developed PLS models could be useful in designing a P sensor for different applications and could be used to improve sensing of P concentrations in soil samples.
0.000
−0.002
No Compounds Very low Low Medium High Very high Extremely high
−0.004
−0.006 500
1000
1500
2000
2500
Absorbance spectra of the leached sandy soil samples changed with added P compounds. The highest change in absorbance values was observed in the UV region of the electromagnetic spectrum for the sandy soil samples with iron (III) phosphate dihydrate. Iron (Fe) associated phosphates showed absorption bands in UV regions of the electromagnetic spectrum, which other investigated phosphates in this study did not indicate. Maximum correlation coefficients of the dry sandy soil samples were 0.93 for iron (III) phosphate dihydrate at 343 nm, 0.68 for aluminum phosphate at 2278 nm, 0.54 for
Wavelength (nm)
(a) 0.002
Absorbance first derivative
Criteria
Table 3. Classification results for the dry soil samples with four compounds. Number of Selected Wavelengths or Principal Components
Correlation coefficient >0.3 and stepwise discriminant analysis Correlation coefficient >0.35 and stepwise discriminant analysis Correlation coefficient >0.4 and stepwise discriminant analysis Stepwise multiple linear regression (SMLR) Principal component analysis
0.000
−0.002
−0.004
−0.006 250
300
350
400
450
Wavelength (nm)
(b) Figure 7. First derivative of average absorbance of dry soil samples with different concentrations of iron (III) phosphate dihydrate in (a) 2252525 nm and (b) 225-450 nm. Each spectrum is an average of four samples.
Vol. 48(6): 2371−2378
Error (%)
51
15.1
15
1.9
1
54.7
52 5
32.1 18.5
Table 4. PLS results for the dry soil samples with four compounds (RMSE is in mg kg−1, and the number of factors was six). Calibration Validation P Compound CaPO4 FePO42H2O AlPO4 Mg3(PO4)22H2O
R2
RMSE
R2
RMSE
0.72 0.76 0.68 0.93
29 23 38 17
0.63 0.75 0.48 0.56
31 27 40 43
2377
calcium phosphate at 2500 nm, and 0.73 for magnesium phosphate hydrate at 2520 nm. Discriminant analysis results showed that the existence of Fe, Mg, Al, and Ca associated P in a sandy soil sample can be estimated with a classification error of 1.9% using reflectance spectroscopy. In addition, PLS predicted P concentration of soils with root mean square error (RMSE) ranging from 27 to 43 mg kg−1. ACKNOWLEDGEMENTS The authors would like to thank to Drs. Donald A. Graetz, Vimala D. Nair, and Willie G. Harris of the Department of Soil and Water Science at the University of Florida for their invaluable suggestions about experimental design and soil properties. This research was supported by the Florida Agricultural Experiment Station and a grant from the Florida Department of Agriculture and Consumer Services, and approved for publication as Journal Series No. R-10447.
REFERENCES Bogrekci, I., W. S. Lee, and J. Herrera. 2003. Assessment of P concentrations in the Lake Okeechobee drainage basins with spectroscopic reflectance of VIS and NIR. ASAE Paper No. 031139. St. Joseph, Mich.: ASAE. Bray, R. H., and L. T. Kurtz. 1945. Determination of total, organic, and available forms of phosphorus in soils. Soil Sci. 59(1): 39-45. Brookes, P. C., D. S. Powlson, and D. S. Jenkinson. 1982. Measurement of microbial biomass phosphorus in soil. Soil Biol. Biochem. 14(4): 319-329. Carter, M. R. 1993. Soil Sampling and Methods of Analysis. Boca Raton, Fla: Lewis Publishing. Corbridge, D. E. C. 1990. Studies in Inorganic Chemistry: Phosphorus, An Outline of Its Chemistry, Biochemistry, and Technology. 4th ed. New York, N.Y.: Elsevier. Dunn, B. W., H. G. Beecher, G. D. Batten, and S. Ciavarella. 2002. The potential of near-infrared reflectance spectroscopy for soil analysis: A case study from the Riverine Plain of southeastern Australia. Australian J. Exp. Agric. 42(5): 607-614.
2378
Hruschka, W. R. 2001. Chapter 3: Data analysis: Wavelength selection methods. In Near-Infrared Technology in the Agricultural and Food Industries, 39-58. 2nd ed. P. Williams and K. Norris, eds. St. Paul, Minn.: American Association of Cereal Chemists. Janik, L. J., R. H. Merry, and J. O. Skjemstad. 1998. Can mid-infrared diffuse reflectance analysis replace soil extractions? Australian J. Exp. Agric. 38(7): 681-696. Lee, W. S., J. F. Sanchez, R. S. Mylavarapu, and J. S. Choe. 2003. Estimating chemical properties of Florida soils using spectral reflectance. Trans. ASAE 46(5): 1443-1453. Mehlich, A. 1984. Mehlich 3 soil test extractant: A modification of Mehlich 2 extractant. Comm. Soil Sci. Plant Analysis 15(12): 1409-1416. Nair, V. D., D. A. Graetz, and K. M. Portier. 1995. Forms of phosphorus in soil profiles from dairies of South Florida. SSSA J. 59: 1244-1249. Nelson, W. L., A. Mehlich, and E. Winters. 1953. The development, evaluation, and use of soil tests for phosphorus availability. Agronomy 4: 153-188. Nyquist, R. A., and R. O. Kagel. 1971. Infrared Spectra of Inorganic Compounds. New York, N.Y.: Academic Press. Olsen, S. R., C. V. Cole, F. S. Watanabe, and L. A. Dean. 1954. Estimation of available phosphorus in soils by extraction with sodium bicarbonate. Circular No. 939. Washington, D.C.: USDA. Pierzynski, G. M. 2000. Methods for phosphorus analysis for soils, sediments, residuals, and waters. Southern Cooperative Series Bulletin No. 396. Southern Association of Agricultural Experiment Station Directors (SAAESD). 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 and Soil 247(2): 243-252. SAS. 1999. SAS/STAT User’s Guide. Ver. 8. Cary, N.C.: SAS Institute, Inc. Savitzky, A., and M. J. E. Golay. 1964. Smoothing and differentiation of data by simplified least squares procedure. Anal. Chem. 36(8): 1627-1638. Tyler, G. 2002. Phosphorus fractions in grassland soils. Chemosphere 48(3): 343-349. Williams, P., and K. Norris. 2001. Near-Infrared Technology in the Agricultural and Food Industries. 2nd ed. St. Paul, Minn.: American Association of Cereal Chemists.
TRANSACTIONS OF THE ASAE