| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
a Dep. of Agronomy, The Pennsylvania State Univ., 116 A.S.I. Bldg., Univ. Park, PA 16802
b Pasture Systems & Watershed Management Research Unit, USDA–ARS, Curtin Rd., Univ. Park, PA 16802
* Corresponding author (ban127{at}psu.edu)
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Abbreviations: Log M3P, logarithm of M3P • M3P, Mehlich-3 extractable soil P • MAE, mean absolute error
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
There is evidence that the great majority of agricultural P export originates from a small portion of the landscape in humid, upland agricultural watersheds (Gburek and Sharpley, 1998). These areas have been termed critical source areas and are characterized by having high potential to release P into surface or subsurface runoff in conjunction with hydrologic connectivity with streams or ditches. Targeting critical source areas would increase the efficiency and reduce the economic costs of control. In response, a site vulnerability assessment tool, the P index, has been developed to target P management (Lemunyon and Gilbert, 1993). The P index accounts for source (soil P and rate, method, and timing of applied P) and transport (surface runoff, erosion, leaching, and landscape position) factors controlling P loss in surface runoff and ranks sites for their potential risk of P loss.
In areas with large fields, the mean or median soil test value is generally used as the best estimate of P concentration in a field, except in cases where precision sampling and fertilizer application are used. In areas with small fields, such as Pennsylvania, a single bulk composite or the mean or median soil test value is traditionally used as the best estimate of P concentration. Under these models, information on farm- and field-scale variability is not used for the estimation of P distribution. More complex interpolation methods, such as those from the disciplines of geostatistics and precision agriculture, incorporate spatial variability into estimates of P distribution. Field-scale variability, which is confined to field boundaries, may be caused by uneven fertilizer distribution or movement within fields. Farm-scale variability, which is not confined to field boundaries, is likely caused by larger scale management factors such as distances to roads or manure storage facilities. Natural factors, such as variations in weathering, soil parent material, erosion, and water movement patterns, may also influence soil P distribution (Larson et al., 1997). The influence of management is probably stronger than natural factors in fields with very high soil P and a history of large P applications. The choice of an interpolation method should be based on an assessment of the scale and strength of autocorrelation present and the costs associated with the sampling design.
The variogram is an important tool to detect the presence of spatial autocorrelation and to estimate the variability structure of soil properties (McBratney and Webster, 1986). A global variogram can be used to assess the variability structure of a soil property across a watershed, but it does not account for smaller-scale factors such as field boundaries. Variograms can also be developed for each field individually (Goovaerts, 1997). Several researchers have used within-field, more generally termed within-stratum, variography to estimate the spatial variability structure of soil properties (Stein et al., 1988; Boucneau et al., 1998). However, data sparsity may prevent the reliable estimation of spatial semivariance functions within each stratum (Webster and Oliver, 1992). At least 50 to 100 data points may be necessary to achieve a stable variogram, depending on lag spacing and the smoothness of the spatial variation (Voltz and Webster, 1990; Burrough and McDonnell, 1998).
Most research concerning soil nutrient distribution has focused on individual fields (Pierce et al., 1995; Gupta et al., 1997). For a 16-ha field, a common field size in the Midwest, about three samples per hectare are needed for 50 data points per field (
50-m grid). For a 2-ha field, a field size common in the northeastern USA and other parts of the country, a sampling intensity of about 25 samples per hectare is required to produce the 50 data point minimum (18-m grid). This sampling intensity is not economically feasible for many agronomic and environmental applications. In such cases, a single pooled within-stratum variogram can be computed, based on the assumption that the spatial variability structure is the same within each stratum. The pooled within-stratum approach has been used to interpolate soil textural fractions across soil mapping units (Voltz and Webster, 1990; Van Meirvenne et al., 1994).
Soil P distribution maps calculated using a field mean are generally used for the P index. However, some researchers have used P distribution maps with subfield scale variability (Eghball and Gilley, 1999; Gburek et al., 2000a). Gburek et al. (2000b) applied the P index at field and 25-m2 cell scales across the same watershed that we are investigating in this study. Results were generally similar, yet there were some differences resulting from the different soil P map and finer resolution of runoff and erosion characteristics based on locally steeper slopes within fields. The authors raised the question whether a subfield resolution will be necessary for P index application or whether other proposed P index modifications will be sufficient to account for fine-resolution factors.
The studied watershed, FD-36, is the site of ongoing USDA–ARS research on chemical and hydrologic factors controlling P transport. A primary objective of the project is to delineate critical source areas of P, areas both high in soil P and within runoff producing zones (Gburek and Sharpley, 1998). The objectives of the study reported in this paper were to detect and analyze the spatial autocorrelation of soil P in the watershed, and to compare and validate three interpolation models (one classical and two geostatistical) for the estimation of soil P distribution in the watershed.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
FD-36 has mixed land use (about 50% soybean [Glycine max (L.) Merr.], wheat [Triticum aestivum L.], or corn [Zea mays L.], 30% woodland, and 20% pasture, meadow, or turfgrass). The watershed has 22 cropped fields with an average field size of 1.0 ha. In many cases, these fields are laid out in strips. However, we decided to analyze each of these strips as individual fields because they are not managed in a coherent rotation. Rather, each field (or strip) is essentially managed as a separate unit. In the 5 yr prior to sampling, selected fields north of the stream received about 60 m3 ha-1 yr-1 swine (Sus scrofa) slurry in spring and no fertilizer P. This amounts to about 100 kg P ha-1 yr-1 assuming a slurry P concentration of 1.6 g L-1 (Eck and Stewart, 1995; Sharpley et al., 1998). South of the stream, 5 Mg ha-1 yr-1 of poultry manure was applied to cropland in the spring. This amounts to 85 kg P ha-1 yr-1 assuming a manure P concentration of 16.9 g kg-1 (Eck and Stewart, 1995; Sharpley et al., 1998).
In July 1996, a total of 301 soil samples (0–5-cm depth) were collected in the cropped areas of the watershed on a roughly 30-m grid (Fig. 1) . Sampling locations were altered on several parts of the watershed to provide better coverage within variable field boundaries. Six cores (0–5-cm depth) were taken using a 2-cm auger within a 1-m radius of the sampling location and composited. This depth of soil sampling is environmentally based and represents the depth of soil interacting with rainfall and surface runoff that controls P release and transport in runoff (Sharpley et al., 1996). The samples were air dried and sieved (2 mm). Mehlich-3 soil P concentration was determined by extraction of 1 g soil with 10 mL of 0.2 M CH3COOH, 0.25 M NH4NO3, 0.015 M NH4F, 0.013 M HNO3, and 0.001 M EDTA for 5 min (Mehlich, 1984). Phosphorus in filtered and neutralized extracts was determined by the method of Murphy and Riley (1962).
|
In the global model, the spatial autocorrelation between points is a function of the distance between points and is not affected by field boundaries. A global omnidirectional variogram was generated based on all the points in the watershed, with semivariance, 
(h), estimated as:
 | (1) |
In the within-field model, the autocorrelation between points is modeled only for point pairs within the same field. Individual variograms for each field were lumped to create a single pooled within-stratum variogram (Goovaerts, 1997, p. 187). This can also be viewed as a global variogram restricted to point pairs within the same field. The pooled within-stratum variogram, ws(h), was estimated using the following equation:
 | (2) |
where h; sk) is the variogram value for the kth stratum, and N(h; sk) is the number of pairs of data locations a distance h apart that jointly belong to the kth stratum. Ordinary kriging was applied within each field individually using the parameters for the pooled within-stratum variogram.
Validation was performed by randomly removing one-third of the data points from each field for use as a validation data set. The three interpolation models were used to estimate M3P concentration at the validation data points and residuals were recorded. Predictions were calculated solely on the nonvalidation data set. Note that this also included a reestimation of the variogram. This estimation was performed with weighted least squares regression and was checked visually. This process was repeated 25 times so that at least two residual estimates were obtained for each data point. Residuals were averaged for a generalized residual estimate. Interpolation methods were compared based on residuals and MAE. Statistical analyses were conducted with S-Plus 2000 and S+Spatial-Stats v.1.1 (Mathsoft, Inc., 1996, 1997) and the SAS System (SAS Institute, 1990).
|
RESULTS AND DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Values were based on exploratory data analysis and the Shapiro–Wilk statistic, it was determined that the M3P was not normally distributed either globally or within fields. The logarithm of M3P (LogM3P) was determined to be normally distributed within fields, although it was negatively skewed globally (skewness = 2.4). The field means of LogM3P (rather than individual values) were also determined to be normally distributed. The mean and variance of LogM3P were 5.18 and 0.55, respectively, and when averaged by field were 5.27 and 0.17, respectively. Analyses were run using both the M3P and LogM3P variables. Results and conclusions were not substantially different between the two variables, therefore only M3P results are presented.
FD-36 Data Analyses
Both the global and the pooled, within-field omnidirectional variograms are presented in Fig. 2
. The nonwithin-field semivariance values, based on point pairs that are not in the same field, are also presented. Numbers near points indicate the number of data pairs within the lag increment (only labeled when <100 data pairs). A spherical semivariance function provided a better fit than did exponential, gaussian, or linear functions for both variograms based on the Akaike's information criteria (Akaike, 1973).
|
Interpolation Results
Prediction surfaces based on the three interpolation models are presented in Fig. 3
. Mehlich-3 P divisions are based on agronomic limits and proposed environmental limits. The 30 mg kg-1 level is the minimum optimal M3P test recommended by the Pennsylvania State College of Agricultural Sciences (Serotkin and Tibbetts, 1998). Between 30 and 100 mg kg-1 M3P, there will generally be little crop response to P fertilizer, but little enrichment of P in surface runoff is expected (Sharpley et al., 1996; Weld et al., 2000). Between 100 and 200 mg kg-1 M3P, no crop response is expected and some enrichment of P in surface runoff is expected to occur, while between 200 and 400 mg kg-1 M3P, considerable enrichment is expected. A > 400 mg kg-1 M3P level was also included to represent the areas of very high P concentration observed in the watershed.
|
The distribution of generally high P concentrations in the watershed are likely influenced only by management practices. There does not seem to be an impact of landscape position, surface soil texture, and other natural factors (data not presented).
Validation
The above analyses demonstrated that there is spatial variability in the watershed at farm- and field-scales. Validation was used to assess the importance of this spatial variability. In Table 1, validation results are presented by field and as an overall average. In Fig. 4
, the average residuals are presented for each model. Classification breaks at 71 and 142 correspond to 1 and 2 times the average MAE across all models. The average MAE of the within-field model is 6.7 and 11.1 mg kg-1 M3P better than the global and field classification models, respectively. These differences are small and indicate that, overall, the three models performed similarly. From the complete data set, the average deviance from the mean within fields was 71 mg kg-1. Although this is a biased estimator, this value roughly compares with the average MAE values observed in the validation data sets, an indication that the removal of one-third of the data did not severely degrade prediction precision.
|
|
For discussion purposes, two models we considered to have performed differently within a field when the difference between the average MAE between the models was >15 mg kg-1. Field identification numbers are georeferenced in Fig. 1, and match those used by Gburek et al. (2000a). The field classification model performed worse in five fields relative to the kriging models (Fields 16, 21, 26, 29, and 30). All three models poorly characterized one of these fields, Field 16, which contains only three sample points and has the greatest within-field variability in the watershed. The other four fields that were poorly described by the field classification exhibit substantial within field spatial autocorrelation (compare Table 1 and Fig. 1). All five of these fields are within 50 m of the stream, and therefore may be among the most hydrologically-active in the watershed.
The global model performed worse than did the within-field model in Fields 28, 31, and 32. The spatial variability in these fields is confined to field boundaries (Fig. 1). The inclusion of farm-scale effects in the modeling of these fields caused a poorer characterization by the global model. The global model also performed worse than the field classification model in fields 31 and 32; in these fields the incorporation of farm-scale variability into the modeling is worse than assuming no spatial autocorrelation. The global model provided the best fit for two fields with very small sample sizes, Fields 16 and 18. In this region of the watershed, the field classification and within-field models failed to incorporate the farm-scale effect of high M3P concentration. The within-field model performed worse than did the other models in Field 18, but was better than both models in Fields 20, 28, and 32. Field 18 is one of the fields with a low sample size (6 samples). Fields 28 and 32 have high within-field spatial autocorrelation that does not extend beyond field boundaries.
Of the 11 fields located within 150 m of the stream, eight were fields that were mapped substantially different by the three models based on the 15 mg kg-1 MAE criterion. These fields are particularly important to the characterization of the watershed because the near-stream areas are the most hydrologically active in this region (Gburek and Sharpley, 1998). Points in the landscape must be hydrologically active for the transport of P. Generally, soils in the near stream area had lower M3P concentrations (Fig. 1 and 3). We suggest that management in these fields may have been affected by poor soil drainage, limiting productivity and accessibility to farm equipment. As a result, these near-stream fields received less P and have lower M3P concentrations than adjecent fields.
An additional analysis was conducted to analyze the influence of sampling intensity on validation results. The validation procedure was repeated except the removal rate for the validation data set was increased from 33 to 67%. In many cases, there was too much scatter in the pooled-within stratum variogram generated from the remaining sample points (100 points) to adequately fit the spherical semivariance function. Based on results from those cases where an adequate fit was possible, the average MAEs increased from 76 to 154 mg kg-1 for the field classification model, 71 to 153 mg kg-1 for the global model and 66 to 150 mg kg-1 for the within-field model. The average deviance from the field mean for the full data set was 71 mg kg-1. The precision decrease was substantially greater as sampling intensity decreased from 200 to 100 sampling points in comparison to a decrease from 301 to 200 sampling points.
The P index, as described by Sharpley (2000), was applied to the FD-36 watershed on a 5-m grid using the three prediction surfaces (Fig. 3) generated from the interpolation models. Though slight differences in P index values were observed, no differences in P index classifications were observed in any field in the watershed between the three prediction surfaces. Therefore, for applications that are not sensitive to small errors in soil P concentration estimates, such as the P index, the field classification method should provide adequate results.
Researchers have found that soil P and runoff P are closely related. This relationship has been shown to vary with soil type and management and P application as manure or fertilizer (Sharpley et al., 1996; Sharpley and Tunney, 2000). This variability should be quantified and integrated with error arising from interpolation to characterize the uncertainty inherent in environmental soil P limits.
|
CONCLUSIONS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
The importance of the spatial autocorrelation observed in the watershed is small. Overall, the kriging models performed only slightly better than the field classification model in terms of estimating M3P concentrations at unknown locations. There were several fields, particularly those in the near-stream zone, that were modeled substantially better by the kriging models because of strong spatial autocorrelation in these fields. If strong autocorrelation is found to be common in hydrologically-active watershed areas, perhaps because of soil wetness or stewardship practices, then the increased cost and complexity of the kriging models may be warranted for some environmental P management applications. Nonetheless, the strength of the spatial autocorrelation observed in this watershed was not sufficient to affect P index classifications.
|
ACKNOWLEDGMENTS |
|---|
Received for publication September 13, 2000.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  L. M. Dudley, A. Ben-Gal, and U. Shani Influence of Plant, Soil, and Water on the Leaching Fraction Vadose Zone J., April 14, 2008; 7(2): 420 - 425. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. L. Little, S. C. Nolan, J. P. Casson, and B. M. Olson Relationships between Soil and Runoff Phosphorus in Small Alberta Watersheds J. Environ. Qual., July 17, 2007; 36(5): 1289 - 1300. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. E. Vaughan, B. A. Needelman, P. J. A. Kleinman, and A. L. Allen Spatial Variation of Soil Phosphorus within a Drainage Ditch Network J. Environ. Qual., May 25, 2007; 36(4): 1096 - 1104. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
T. Page, P. M. Haygarth, K. J. Beven, A. Joynes, T. Butler, C. Keeler, J. Freer, P. N. Owens, and G. A. Wood Spatial Variability of Soil Phosphorus in Relation to the Topographic Index and Critical Source Areas: Sampling for Assessing Risk to Water Quality J. Environ. Qual., November 7, 2005; 34(6): 2263 - 2277. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| The SCI Journals | Agronomy Journal | Crop Science | |||
| Journal of Natural Resources and Life Sciences Education |
Vadose Zone Journal | ||||
| Journal of Plant Registrations | Journal of Environmental Quality |
The Plant Genome | |||
