ESTIMATING CHEMICAL PROPERTIES OF FLORIDA SOILS USING SPECTRAL REFLECTANCE W. S. Lee, J. F. Sanchez, R. S. Mylavarapu, J. S. Choe ABSTRACT. This study was conducted to develop fundamental relationships between soil properties from three representative Florida soil orders and their spectral characteristics. The ultimate goal of this work is to develop a real–time soil property sensor for use in effective farm management. A total of 270 samples collected from the three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida were used for analysis. Soil samples were obtained from 0 to 15 cm depth at 15 sampling points within three specific fields of 2.0 ha each of the three soil orders at six different times of the year, assuring a wide range of sample variability in sampling times and locations. Reflectance of the soil samples was measured in the range of 400 to 2498 nm with a 2 nm increment, and the corresponding nutrient content (P, K, Ca, and Mg) along with pH and soil organic matter content was measured for each of the samples. Partial least squares analysis was used to build prediction models with a calibration data set of 180 randomly chosen samples. The remaining 90 samples were used to validate the models. The prediction models for measured soil chemical properties for the three soil orders yielded R2 values of 0.24 to 0.88. This result could be useful in the development of a soil nutrient sensor for site–specific crop management. Keywords. NIR spectroscopy, PLS, Precision agriculture, Reflectance, Sensor, Soil nutrients, Soil property.

P

recision agriculture has been widely used to improve farm management practices and thus increase profits by increasing yield and reducing manage– ment costs. Improvements in sensor technologies and GPS/GIS have enabled growers to use precision technologies, such as yield monitoring and mapping, variable–rate application (VRA), and remote sensing. Among sensor development efforts, many research activities have focused on developing plant and soil nutrient sensing systems, especially for nitrogen (N) (Stone et al., 1996; Sui et al., 1998; Lee at al., 1999; Lee and Searcy, 2000). However, few investigators have reported on sensing phosphorus (P), potassium (K), and micro–nutrient elements. Phosphorus has been identified as a pollutant carried into waterbodies in Florida, causing eutrophication of lakes, rivers, and streams (Lemunyon and Daniel, 1998). Excess algal bloom causes a decline in the aesthetic quality of waters along with hypoxia (depletion in dissolved oxygen levels), seriously impacting aquatic life. Agricultural operations have been determined to be major contributors to nutrient loadings into water bodies (U.S. EPA, 1996). Effective soil testing to determine actual crop needs of P, and soil

Article was submitted for review in December 2002; approved for publication by the Information & Electrical Technologies Division of ASAE in July 2003. Presented at the 2001 ASAE Annual Meeting as Paper No. 011179. The authors are Won Suk Lee, ASAE Member Engineer, Assistant Professor, Department of Agricultural and Biological Engineering, Jaime F. Sanchez, Graduate Research Assistant, Department of Soil and Water Science, and Rao S. Mylavarapu, Assistant Professor, Department of Soil and Water Science, University of Florida, Gainesville, Florida; and Jung Seob Choe, Assistant Professor, Department of Mechanical Engineering, Sangju National University, Sangju, South Korea. Corresponding author: Won Suk Lee, Department of Agricultural and Biological Engineering, Frazier Rogers Hall, Museum Road, Gainesville, FL 32611–0570; phone: 352–392–1864, ext. 227; fax: 352–392–4092; e–mail: [email protected].

test–based nutrient management (either through fertilizer or manure), will potentially minimize unnecessary P applications and thereby prevent pollution caused by buildup of P in the soils. One practical technique to assess plant and soil nutrient status is near–infrared (NIR) spectroscopy, due to its conceivable use for rapid and non–destructive determination of the concentration of certain constituents in a sample and its significant labor and cost savings. NIR spectroscopy has been extensively used for many agricultural applications such as water and nutrient stress sensing for agricultural crops (Thomas and Oerther, 1972; Stafford et al., 1989a, 1989b; Blackmer et al., 1994; Masoni et al., 1996; Sudduth and Hummel, 1996; Bausch et al., 1998), and weed detection (Wang et al., 2000). There were several attempts to assess soil properties among these studies. Sudduth and Hummel (1993a) designed and tested a portable NIR spectrophotometer in the range of 1650 to 2650 nm. They reported that soil reflectance curves obtained with this instrument agreed well with data obtained using a research–grade spectrophotometer. Sudduth and Hummel (1993b) further tested the portable NIR spectrophotometer for quantifying soil organic matter content (SOM), CEC, and moisture content. They reported that laboratory calibrations yielded an R2 of 0.89 and a standard error of prediction (an error term for prediction performance of the calibration models) of 0.40% organic matter. However, limited in–furrow field tests produced much higher errors due to the movement of soil past the sensor. Ehsani et al. (1999) investigated the possibility of rapidly sensing soil mineral–N content using NIR reflectance and reported that reflectance in the 1800–2300 nm range could be used to determine the nitrate content of soil successfully, although calibration equations were needed for the same field where validation samples were acquired.

Transactions of the ASAE Vol. 46(5): 1443–1453

E 2003 American Society of Agricultural Engineers ISSN 0001–2351

1443

Table 1. Soil sampling schedule in 2001. Set 1 Set 2 Set 3 Set 4 Set 5

Soil Alfisol Entisol Ultisol

Feb. 28 Feb. 6 Feb. 5

March 7 March 7 March 7

April 2 April 2 April 2

May 31 May 31 May 31

July 6 July 6 July 6

Set 6 Aug. 17 Aug. 17 Aug. 17

Ingleby and Crowe (2000) developed linear regression models to predict soil organic carbon content in five Saskatchewan fields and reported that the best model for each field varied in number of regressors and reflectance wavelengths, suggesting that site–specific models may be needed. Shibusawa et al. (1999) developed a real–time portable spectrophotometer for measuring underground soil reflectance in the 400–1700 nm range. They reported that R2 values between reflectance spectra and soil properties (moisture, pH, EC, SOM, NO3–N) ranged from 0.19 to 0.87. Thomasson et al. (2001) investigated the relationships between soil properties and reflectance spectra using samples obtained from two fields in northeastern Mississippi. They reported that only Ca and Mg in one field and clay and pH in the other could be predicted by multiple regression models with R2 values greater than 0.50. They found that the 400– 800 nm and 950–1500 nm regions had high discriminatory power. Hummel et al. (2001) used an NIR soil sensor to predict soil moisture and organic matter content of surface and subsurface soils. Spectral reflectance data in the 1623– 2467 nm range were used in stepwise multiple linear regression analysis. They reported that standard errors of prediction for organic matter and soil moisture were 0.62% and 5.31%, respectively. OBJECTIVE The objective of this research was to develop fundamental relationships between soil properties from representative soil orders in Florida and the spectral characteristics of the soil samples, as a preliminary step toward development of a real–time soil property sensor for use in site–specific crop management. Once a real–time soil sensor system is developed, it will greatly decrease the time and drudgery required for collecting and analyzing soil samples. The sensor system could eventually contribute to efficient water and nutrient management and potentially increase yield and profit, as well as reduce the risks of environmental contamination by excessive nutrients.

MATERIALS AND METHODS

SOIL SAMPLING Soil samples were collected using a soil sampling auger from fields representative of three predominant soil orders in

Florida: Alfisol, Entisol, and Ultisol. Order is the highest level of soil classification categories, and brief technical characteristics of each selected soil order are given below. Alfisols are soils that have an ochric epipedon (surface layer), an argillic horizon (subsoil with a higher clay content than the layers above), and a moderate to high level of bases (Ca, Mg, K, etc.). Alfisols range from well–drained soils on ridges and knolls to very poorly drained soils on flats, depressions, and floodplains. Limestone sometimes occurs below the argillic horizons. Alfisols cover an area of approximately 1.9 million ha in the state of Florida. Entisols are generally young, sandy soils that show little or no development of pedogenic horizons except for ochric epipedons and albic horizons. Entisols are excessively to poorly drained. Entisols cover approximately 3.0 million ha in Florida. Ultisols are soils that have an argillic or kandic horizon (layer in the subsoil with a higher clay content) and a low level of bases such as Ca, Mg, K, etc. Ultisols are deep soils on upland ridges, rises, and knolls to very poorly drained soils and cover approximately 2.8 million ha in Florida. Soils covered by these three orders are some of the major agricultural production areas in the state, and most of these are well–drained to suit commercial crop production. Field sites were located at the research farm of the University of Florida Plant Science Research and Education Unit (PSREU) at Citra, Florida. Three specific fields of 2.0 ha each of the three soil orders were selected for our study based on soil mapping as per USDA Classification (Collins, 2000). The sampling sites were fallow lands, with no agricultural activity during the sampling period. This helped ensure that the study sites were not influenced by any management practices. At each sampling event, 15 sampling points were randomly established for each soil order within each site. Soil samples were collected at a 0–15 cm depth at each sampling point. Sites were sampled at six different times of the year at approximately one–month intervals between February and August 2001, which assured a wide range of sample variability in sampling times and locations (table 1). These data have been categorized as sets 1, 2, 3, 4, 5, and 6, respectively. A total of 270 samples were collected during this period. ANALYSIS OF SOIL PROPERTIES All soil samples were analyzed for pH (soil:water 1:2), soil organic matter, and Mehlich–1 extractable Ca, Mg, P, and K at the UF/IFAS Analytical Research Laboratory in Gainesville, Florida. Soil organic matter was determined by the Walkley–Black method and was expressed as a percentage.

Table 2. Means separation by Tukey, range, and CV for pH, organic matter (OM) content, and nutrient contents for three soil orders. Alfisol Entisol Ultisol

[a]

Variable

Mean

Range

CV

Mean

Range

CV

Mean

Range

CV

pH OM (%) P (mg/kg) K (mg/kg) Ca (mg/kg) Mg (mg/kg)

5.2 b[a] 2.1 a 118.7 a 50.6 a 273 a 45.7 a

4.6 – 6.0 0.9 – 3.2 65.2 – 236.0 10.34 – 152.0 105.9 – 612.0 9.8 – 112.0

6.0 18.0 27.1 50.8 42.8 52.1

6.0 a 1.7 b 52.2 b 39.9 a 460 a 74.3 a

5.1 – 6.7 0.3 – 2.8 32.2 – 142.0 11.0 – 89.7 98.8 – 1140.0 6.5 – 186.0

6.6 25.3 27.6 36.1 42.6 45.3

5.4 b 1.8 b 56.1 b 37.1 a 265 a 38.3 a

4.9 – 6.8 0.5 – 3.1 27.4 – 130.8 16.8 – 83.3 83.0 – 1184.0 3.4 – 143.3

7.6 24.5 32.3 37.0 66.5 74.0

Means within a row followed by the same letter are not significantly different (P > 0.05).

1444

TRANSACTIONS OF THE ASAE

Table 3. Number of extracted factors and PRESS value for the prediction models for calibration data. PLS Models

1.2

....

log(1/Reflectance)

1.0

Alfisol Entisol Ultisol

0.8

Water

0.6

Water

0.4

0.2

500

1000

1500

2000

Constituent

Number of Extracted Factors

PRESS for Validation

pH OM P K Ca Mg

12 9 9 5 20 6

0.436 0.882 0.536 0.886 0.304 0.526

Table 4. R2, root mean square error (RMSE), and coefficient of variation (CV) for calibration data. Prediction Models

2500

Wavelength (nm) Figure 1. Average absorbance of the samples acquired from three different soil orders in February 2001.

For pH, organic matter, and nutrient content, a univariate analysis of variance was performed to distinguish differences among the soil orders. For the analysis of variance, a nested structure was used, where the soil sampling dates were nested within the soil orders. Additionally, coefficient of variation (CV) and range of soil characteristics were calculated. REFLECTANCE MEASUREMENT All soil samples were oven dried at 104°C for 24 hours, ground, and sieved with a 0.6 mm sieve in order to remove any effect of particle size before measuring reflectance. Spectral reflectance (R) was measured for all soil samples using a spectrophotometer (model 6500, FOSS NIRSystems, Inc., Silver Spring, Md.) from 400 to 2498 nm with a 2 nm increment. Absorbance for the scanned range was recorded (log[1/R]). A reference scan was measured before each sample was scanned. One reflectance spectrum contained 1050 measurements at 1050 different wavelengths. SPECTRAL DATA ANALYSIS The objective of the spectral data analysis was to find an optimal number of factors that would best explain soil properties for different soil orders, and to develop prediction models for the different soil chemical properties analyzed. The data set of 270 samples, 90 for each soil order, was split into a calibration set of 180 randomly selected samples and a validation set with the remaining 90 samples. The selection was randomly stratified by soil order, with each order contributing the same number of samples for both sets, i.e., 60 samples for calibration and 30 samples for validation. PARTIAL LEAST SQUARES (PLS) REGRESSION ANALYSIS Partial least squares analysis (PLS) was used to find an optimal number of factors to explain the variability observed (SAS, 1999). The optimum number of factors for the PLS model was obtained by a cross–validation approach to select the best–fitted model for each soil property (pH, organic matter, P, K, Ca, and Mg). A verification process was performed using regression analysis (PROC GLM) between the actual values and the predicted values from the PLS model. Spectral patterns were established plotting loadings of the first three factors with their respective loadings (Martens and Næs, 1989).

Vol. 46(5): 1443–1453

Constituent

R2

RMSE

CV (%)

pH OM P K Ca Mg

0.89 0.47 0.81 0.40 0.97 0.79

0.17 0.22 (%) 15.59 (mg/kg) 9.22 (mg/kg) 31.74 (mg/kg) 13.77 (mg/kg)

3.03 12.08 20.41 21.51 9.43 25.52

Table 5. R2, root mean square error (RMSE), and coefficient of variation (CV) for validation data. Prediction Models Constituent

R2

RMSE

CV (%)

pH OM P K Ca Mg

0.78 0.49 0.72 0.24 0.88 0.77

0.22 0.23 (%) 18.60 (mg/kg) 8.76 (mg/kg) 63.61 (mg/kg) 12.87 (mg/kg)

3.97 12.63 24.80 21.65 19.84 26.44

The PLS procedure was developed first in chemometrics to evaluate multivariate data (Martens, 2001; Wold et al., 1983), and it is an extension of the econometric path modeling established during the 1970s by H. Wold (Wold, 1975). In situations where there are many independent variables explaining a dependent factor, PLS can be an appropriate multivariate technique to use (Sundberg, 1999). PLS is a two–step process. The first step is a reduction of matrix dimensions or finding the number of relevant components (Helland, 2001), and the second step involves identification of latent structure models in the data matrix (Lingærde and Christophersen, 2000). An ordinary least squares procedure is not applicable when one or more columns of the matrix data can be expressed as a linear combination of other data columns of the same matrix, a condition called “collinearity.” In such cases of high collinearity, unbiased predictors have high variance, which generally leads to unrealistic regression coefficients (Björkström and Sundberg, 1999). The PLS technique uses steps to find the best model, using the residual error of each step as an input of the next (Björkström and Sundberg, 1999; Stoica and Söderström, 1998). An optimal number of factors is generally obtained by a cross–validation procedure, where the calibration set is submitted to “leave one out” comparisons until the predicted residual sum of squares (PRESS) is at minimum (Sundberg, 1999). At this point, the model has an optimal number of factors.

1445

Figure 2. Prediction of pH with the samples from three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida by PLS regression. The solid line represents the regression line, and the dotted line shows the 95% confidence interval of the predicted values.

Figure 3. Prediction of organic matter with the samples from three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida by PLS regression. The solid line represents the regression line, and the dotted line shows the 95% confidence interval of the predicted values.

The PLS procedure is widely used in multivariate calibration (Martens and Næs, 1989) because another goal of the procedure is to take into account the variation in the predictors and improve the prediction of new observations during the validation process (SAS, 1999). After the unknown structure of the data has been detected with a PLS model, the model can be used to predict new values for a new data set from the same population. The PLS procedure is also a good alternative to the traditional multiple regression analysis and principal component regression because the model parameters obtained by PLS are more robust, which means that the parameters do not change extensively when new calibration samples are submitted to the model (Geladi and Kowalski, 1986). Finally, the goodness of fit of the model, observed in the test set, is evaluated with the model:

1446

So = Sp + Ů

(1)

where So = observed nutrient content in the test set Sp = predicted value with the PLS model selected Ů = error term.

RESULTS AND DISCUSSION

SOIL NUTRIENT ANALYSIS Table 2 shows the analysis of variance results and the Tukey test results for separation of means for the data sets obtained at different times, using the nested analysis procedure. Each value in the table is an average of all 90 samples from the same soil order. The samples of Entisol showed a significantly higher pH value than those of the other

TRANSACTIONS OF THE ASAE

Figure 4. Prediction of P with the samples from three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida by PLS regression. The solid line represents the regression line, and the dotted line shows the 95% confidence interval of the predicted values.

Figure 5. Prediction of K with the samples from three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida by PLS regression. The solid line represents the regression line, and the dotted line shows the 95% confidence interval of the predicted values.

two soils but with similar coefficient of variation (CV) and range to the other two soil orders. Alfisol soils had higher soil organic matter (OM), with a similar range of values but lower variability than the other two soil orders. Soil P concentrations were significantly higher and varied over a wider range for Alfisols; however, the CV range among the soil orders was closer. Concentrations of K were not significantly different among the three soil orders but were slightly higher for the Alfisol samples. The higher variability observed in Ca and Mg resulted in no significant differences between the three soils evaluated, but the Entisol samples tended to have higher Ca and Mg contents. The maximum coefficient of variation was observed for Mg in Ultisol, and the lowest was for pH in Alfisol. The higher range was detected for Ca in Ultisol, with very close values in Entisol. The CV values and ranges in

Vol. 46(5): 1443–1453

table 2 indicate high variation both within and among soil orders, resulting in a data set that covered a wide range of observations. CHARACTERISTICS OF SOIL ABSORBANCE SPECTRA Figure 1 shows average absorbance curves of different soil orders acquired in February 2001. Soil samples from the three soil orders had similar absorbance characteristics. The figure shows water absorption bands at 1414 and 1928 nm. Since two different detectors were used in the spectrophotometer, there was a discontinuity at 1100 nm in the absorbance measurement. PLS MODELS FOR SOIL NUTRIENT CONTENT PREDICTION The objective of the PLS analysis was to find an optimal number of factors that would best explain the variability among soil properties for different soil orders, and to develop 1447

Figure 6. Prediction of Ca with the samples from three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida by PLS regression. The solid line represents the regression line, and the dotted line shows the 95% confidence interval of the predicted values.

Figure 7. Prediction of Mg with the samples from three representative soil orders (Alfisol, Entisol, and Ultisol) in Florida by PLS regression. The solid line represents the regression line, and the dotted line shows the 95% confidence interval of the predicted values.

Constituent pH OM P K Ca Mg

Factor 1

Table 6. Selected wavelengths (nm) with peak loadings for different soil properties. Factor 2 Factor 3

770, 1066, 1450, 2064, 2256 512, 1412, 2036, 2256 522, 1412, 2036 522, 1412, 1722, 2040, 2256 530, 2046, 2256 530, 2046, 2256

426, 940, 1412, 1912, 2204 428, 612, 748, 1140, 1504, 1914, 2206 428, 612, 742, 1100, 1498, 1912, 2206, 2256 428, 690, 740, 1110, 1502, 1912, 2204 418, 612, 742, 1118, 1412, 1914, 2206 418, 612, 742, 1124, 1412, 1914, 2206

prediction models for the soil properties. Table 3 shows the results for the selection of the optimal number of factors from the PLS process. The PRESS statistic (predicted residual sum of squares) is a result of the cross–validation process carried out to obtain an optimal number of factors in the PLS process (SAS, 1999) and judges the goodness of prediction of the process (Sundberg, 1999). Lower PRESS values indicate better fit of

1448

664, 1362, 1614, 1746, 2052, 2232 624, 1376, 1626, 1816, 1860, 2072, 2278, 2376 430, 602, 880, 1722, 1948, 2066, 2272 444, 828, 1356, 1604, 1716, 1722, 1916, 2064, 2206 626, 1366, 1622, 1810, 2064, 2234 626, 1368, 1626, 1812, 2066, 2232

the model obtained in the PLS process. For this reason, a rule of thumb is to select an optimal number of factors when PRESS is at minimum value. For organic matter, P, K, and Mg, less than 10 factors were necessary to obtain the best PLS model. For pH and Ca, 12 and 20 factors were selected, respectively, indicating probably a more complex interaction with soil color.

TRANSACTIONS OF THE ASAE

Figure 8. Loadings from the first three PLS factors for soil pH in the validation data set.

Figure 9. Loadings from the first three PLS factors for organic matter in the validation data set.

Table 4 shows the performance of the calibration data set with the PLS procedure. The models for pH, P, Ca, and Mg accounted for more than 79% of the variation observed in the calibration data set, and those for soil organic matter (OM) and K accounted for less than 50% of the variation. Chemical stability of organic matter is very low in the southeastern U.S. soils due to high temperature and rainfall conditions, resulting in rapid alterations of the decomposition products. As a result, variability in soil organic matter is high, which can explain the lower R2 obtained in the model. Soil K is highly soluble and mobile, and therefore is subject to high leaching in sandy soils, which resulted in low K concentrations in all the soil orders. The magnitude of leaching is directly influenced by soil texture, which therefore resulted in lower concentrations and higher variability (lower R2) in all the soil orders.

Vol. 46(5): 1443–1453

The prediction results with the validation data set (table 5) were very promising for pH, Ca, Mg, and P as they yielded reasonably high R2. The good result for P would be very useful in Florida with its unique problem regarding the high P concentration around the Lake Okeechobee drainage basin. The model could be used to predict P concentration in soil by measuring surface reflectance instead of collecting samples from the field and analyzing them in a laboratory. The variability observed in table 5 for the model validation reflects complexity resulting from both individual and cumulative effects of factors influencing soil color such as sampling times, mineralogy, soil management, etc. The spectroscopic techniques also have their own limitations. Figures 2 through 7 show the relationships between observed and predicted values for each soil characteristic obtained from the validation data. The solid lines in the figures are regression lines. The dotted lines show 95%

1449

Figure 10. Loadings from the first three PLS factors for P in the validation data set.

Figure 11. Loadings from the first three PLS factors for K in the validation data set.

confidence limits for predicted values. The R2 values followed by “**” indicate that the simple linear models are significant at the 0.01 level. The Alfisol samples showed higher P levels than the other soils (table 2 and fig. 4). For this reason, the soil type apparently had a strong influence on the regression obtained for P. The same tendency was observed for organic matter (fig. 3). The Entisol and Ultisol samples were well–distributed over the entire range for soil pH (fig. 2), whereas the Alfisol samples tended to be lower. For all the other soil characteristics, the sample values were distributed more or less randomly, occupying all the range without any tendency of grouping (figs. 3, 5, 6, and 7). Figures 8 through 13 show loading spectra for each soil characteristic for the first three PLS factors. For all soil characteristics, the loadings at all wavelengths for factor 1 seemed to be similar (i.e., a mostly flat line). However, the

1450

loadings for factors 2 and 3 varied greatly from wavelength to wavelength for the different soil properties. Loading spectra from PLS processes have been used to identify relationships between wavelengths and specific characteristics of the samples under study (Martens and Næs, 1989). It is difficult to link these peaks with chemical or physical soil characteristics that can affect the values of each variable from the current data analysis. However, these peaks contributed to establishing the relationship between spectral patterns and factors that influence soil nutrient concentrations. For example, we have identified peaks for P at 522, 1412, and 2036 nm from factor 1; 428, 612, 742, 1100, 1498, 1912, 2206, and 2256 nm from factor 2; and 430, 602, 880, 1722, 1948, 2066, and 2272 nm from factor 3. Table 6 summarizes the peaks identified for all the other characteristics under study. More work is necessary to identify the importance of these peaks in the characterization of the

TRANSACTIONS OF THE ASAE

Figure 12. Loadings from the first three PLS factors for Ca in the validation data set.

Figure 13. Loadings from the first three PLS factors for Mg in the validation data set.

useful spectra to find the actual values from soil samples using NIR techniques. It is also important to note that some of those peaks identified can be the result of unidentified noise or interference with other aspects of the sample, including the sampling and preparation of the specimens submitted to analysis.

SUMMARY AND CONCLUSIONS

This study was conducted to find fundamental relationships between chemical properties of soil samples from major soil orders in Florida and their spectral characteristics, as a preliminary step toward a real–time soil property sensor designed for efficient nutrient management on Florida’s agricultural farms.

Vol. 46(5): 1443–1453

A total of 270 samples were obtained from fields representative of three predominant soil orders in Florida: Alfisol, Entisol, and Ultisol. Field sites were located at the research farm of the University of Florida Plant Science Research and Education Unit (PSREU) at Citra, Florida. Three specific fields of 2.0 ha each of the three soil orders were selected for our study based on soil mapping as per USDA Classification. The sampling sites were fallow lands, with no agricultural activity during the sampling period. At each sampling event, 15 sampling points were randomly established for each soil order within each site. Soil samples were collected at a 0–15 cm depth at each sampling point. Sites were sampled at six different times of the year at approximately one–month intervals between February and August 2001, which assured a wide range of sample variability in sampling times and locations.

1451

All samples were analyzed for their soil characteristics (pH, organic matter content, P, K, Ca, and Mg), and reflectance was measured from 400 to 2498 nm with a 2 nm increment. PLS regression analyses were conducted with 180 randomly selected samples to build prediction models. The remaining 90 samples were used to validate the models. The major findings from this research were: S The mean pH, soil organic matter, and P values among the soil orders were significantly different at the 0.05 level. The mean K, Ca, and Mg values, however, were not significantly different (a = 0.05) among the soil orders. These observations reflected spatial and temporal variability among soil types. S The prediction models developed by the PLS procedure were able to explain more than 79% of the variation observed in the calibration data set for soil pH, P, Ca, and Mg. The models for soil organic matter and K accounted for less than 50% of the variation observed in the calibration data set. S The PLS prediction models accounted for more than 72% of the variation observed in the validation set for soil pH, P, Ca, and Mg. For K and soil organic matter, the models accounted for less than 50% of the variation observed in the validation data set. S Wavelengths were identified from the factors extracted with PLS analysis for different soil properties. Further studies are needed to increase the spectral resolution of each nutrient in different soils. These may be used to develop a soil nutrient sensor for efficient site–specific farming. ACKNOWLEDGEMENTS The authors would like to thank Joseph Nguyen (Senior Chemist, Department of Soil and Water Science, University of Florida) and Bob Tonkinson (Engineer, Department of Agricultural and Biological Engineering, University of Florida) for soil sampling, Juan Herrera (undergraduate student, Department of Agricultural and Biological Engineering, University of Florida) for preparing soil samples for spectral measurements, and Dr. Sam Coleman of USDA–ARS at Brooksville, Florida, for access to the NIRS spectrophotometer. This research was supported by the Florida Agricultural Experiment Station and approved for publication as Journal Series No. R–09212.

REFERENCES

Bausch, W. C., K. Diker, A. F. H. Goetz, and B. Curtis. 1998. Hyperspectral characteristics of nitrogen–deficient corn. ASAE Paper No. 983061. St. Joseph, Mich.: ASAE. Björkström, A., and R. Sundberg. 1999. A generalized view on continuum regression. Scandinavian J. Statistics 26(1): 17–30. Blackmer, T. M., J. S. Schepers, and G. E. Varvel. 1994. Light reflectance compared with other nitrogen stress measurements in corn leaves. Agronomy J. 86(6): 934–938. Collins, M. 2000. Personal communication: Soil mapping at the UF/IFAS PSREU. Gainesville, Fla.: University of Florida, Department of Soil and Water Science, Soil Pedology Program. Ehsani, M. R., S. K. Upadhyaya, D. Slaughter, S. Shafii, and M. Pelletier. 1999. An NIR technique for rapid determination of soil mineral nitrogen. Precision Agric. 1(2): 219–236.

1452

Geladi, P., and B. R. Kowalski. 1986. Partial least squares regression: A tutorial. Analytica Chimica Acta 185: 1–17. Helland, I. S. 2001. Some theoretical aspects of partial least squares regression. Chemometrics and Intelligent Laboratory Systems 58(2): 97–107. Hummel, J. W., K. A. Sudduth, and S. E. Hollinger. 2001. Soil moisture and organic matter prediction of surface and subsurface soils using an NIR soil sensor. Computers and Electronics in Agric. 32(2): 149–165. Ingleby, H. R., and T. G. Crowe. 2000. Reflectance models for predicting organic carbon in Saskatchewan soils. Canadian Agric. Eng. 42(2): 57–63. Lee, W. S., and S. W. Searcy. 2000. Multispectral sensor for detecting nitrogen in corn plants. ASAE Paper No. 001010. St. Joseph, Mich.: ASAE. Lee, W. S., S. W. Searcy, and T. Kataoka. 1999. Assessing nitrogen stress in corn varieties of varying color. ASAE Paper No. 993034. St. Joseph, Mich.: ASAE. Lemunyon, J., and T. C. Daniel. 1998. Phosphorus management for water quality protection: A national effort. In Soil Testing for Phosphorus: Environmental Uses and Implications, 1–4. J. T. Sims ed. Southern Coop. Series Bulletin No. 389, SERA–IEG–17. Lindaerde, O. L., and N. Christophersen. 2000. Shrinkage structure of partial least squares. Scandinavian J. Statistics 27(3): 459–473. Martens, H. 2001. Reliable and relevant modeling of real world data: A personal account of the development of PLS regression. Chemometrics and Intelligent Laboratory Systems 58(2): 85–95. Martens, H., and T. Næs. 1989. Multivariate Calibration. Chichester, U.K.: John Wiley and Sons. Masoni, A., L. Ercoli, and M. Mariotti. 1996. Spectral properties of leaves deficient in iron, sulfur, magnesium, and manganese. Agronomy J. 88(6): 937–943. SAS. 1999. The PLS Procedure. In SAS/STAT User’s Guide. Version 8. Cary, N.C.: SAS Institute, Inc. Shibusawa, S., M. Z. Li, K. Sakai, A. Saao, H. Sato, S. Hirako, and A. Otomo. 1999. Spectrophotometer for real–time underground soil sensing. ASAE Paper No. 993030. St. Joseph, Mich.: ASAE. Stafford, J. V., G. S. Weaving, and J. C. Lowe. 1989a. A portable infrared moisture meter for agricultural and food materials: Part 1. Instrument development. J. Agric. Eng. Research 43(1): 45–56. Stafford, J. V., C. R. Bull, and G. S. Weaving. 1989b. A portable infrared moisture meter for agricultural and food materials: Part 2. Field evaluation on grass. J. Agric. Eng. Research 43(1): 57–66. Stoica, P., and T. Söderström. 1998. Partial least squares: A first–order analysis. Scandinavian J. Statistics 25(1): 17–24. Stone, M. L., J. B. Solie, R. W. Whitney, W. R. Raun, and H. L. Lees. 1996. Sensors for detection of nitrogen in winter wheat. SAE Technical Paper Series. SAE Paper No. 961757. Warrendale, Pa.: SAE. Sudduth, K. A., and J. W. Hummel. 1993a. Portable, near–infrared spectrophotometer for rapid soil analysis. Trans. ASAE 36(1): 185–193. Sudduth, K. A., and J. W. Hummel. 1993b. Soil organic matter, CEC, and moisture sensing with a portable NIR spectrophotometer. Trans. ASAE 36(6): 1571–1582. Sudduth, K. A., and J. W. Hummel. 1996. Geographic operating range evaluation of an NIR soil sensor. Trans. ASAE 39(5): 1599–1604. Sui, R., J. B. Wilkerson, W. E. Hart, and D. D. Howard. 1998. Integration of neural networks with a spectral reflectance sensor to detect nitrogen deficiency in cotton. ASAE Paper No. 983104. St. Joseph, Mich.: ASAE.

TRANSACTIONS OF THE ASAE

Sundberg, R. 1999. Multivariate calibration – Direct and indirect regression methodology. Scandinavian J. Statistics 26(2): 161–207. Thomas, J. R., and G. F. Oerther. 1972. Estimating nitrogen content of sweet pepper leaves by reflectance measurements. Agronomy J. 64(1): 11–13. Thomasson, J. A., R. Sui, M. S. Cox, and A. Al–Rajehy. 2001. Soil reflectance sensing for determining soil properties in precision agriculture. Trans. ASAE 44(6): 1445–1453. U.S. EPA. 1996. Environmental indicators of water quality in the United States. EPA 841–R–96–002. USEPA Office of Water (4503F). Washington, D.C.: U.S. Government Printing Office.

Vol. 46(5): 1443–1453

Wang, N., N. Zhang, D. E. Peterson, and F. E. Dowell. 2000. Testing of a spectral–based weed sensor. ASAE Paper No. 003127. St. Joseph, Mich.: ASAE. Wold, H. 1975. Soft modeling by latent variables: The nonlinear iterative partial least squares approach. In Perspectives in Probability and Statistics. Papers in honor of M. S. Barlett. J. Gani, ed. New York, N.Y.: Academic Press. Wold, S., H. Martens, and H. Wold. 1983. The multivariate calibration method in chemistry solved by the PLS method. In Proc. Conference in Matrix Pencils, 286–293. A. Ruhe and B. Kagstrom, eds. Heidelberg, Germany: Springer–Verlag.

1453

1454

TRANSACTIONS OF THE ASAE