SSSAJ Grow Your Career with SSSA
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online 25 August 2005
Published in Soil Sci Soc Am J 69:1590-1599 (2005)
DOI: 10.2136/sssaj2003.0264
© 2005 Soil Science Society of America
677 S. Segoe Rd., Madison, WI 53711 USA
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Ziadat, F. M.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Ziadat, F. M.
GeoRef
Right arrow GeoRef Citation
Agricola
Right arrow Articles by Ziadat, F. M.
Related Collections
Right arrow Frozen Soils
Right arrow Predictive soil mapping
Right arrow Soil Models

Soil & Water Management & Conservation

Analyzing Digital Terrain Attributes to Predict Soil Attributes for a Relatively Large Area

Feras M. Ziadat*

Dep. of Land, Water and Environment, Faculty of Agriculture, Univ. of Jordan, P.O. Box 13693, Amman 11942, Jordan

* Corresponding author (fziadat{at}ju.edu.jo)


   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Accurate information about soil attributes, presented in a spatially continuous form, is prerequisite for many land resources management applications. The availability of detailed soil maps and its ability to supply such information for modern tools and applications are questionable. Some alternatives, based on using terrain attributes derived from digital elevation model (DEM) to predict soil attributes are investigated in this study. A study area of 148 km2 in the northern part of Jordan was used. The area is covered by detailed soil map and 2193 field observations, from which the soil attributes were extracted. Terrain attributes derived from 20-m resolution DEM were utilized to predict soil attributes by implementing different statistical and clustering techniques. The use of multiple linear regression models within small watershed subdivisions enabled the prediction of soil depth for 89.3% of the field observations within ±50 cm, the water-holding capacity (WHC) for 75.8% within ±50 mm cm–1 and the surface cover percentage for 78.7% within ±10%. The models also predicted surface cover type for 94.5% of the field observations, erosion type for 48.4%, erosion class for 98.0%, and soil texture for 90.3%, within one class difference between predicted and field estimated classes. Comparing these results with estimates of soil attributes using the soil map indicated that the modeling of the soil-landscape relationships within small watershed subdivisions is a promising approach to predict soil attributes for large areas. An important feature is the spatial distribution of the predicted soil attributes, which is provided in more detailed form than what the soil map provides.

Abbreviations: CTI, compound topographic index • DEM, digital elevation model • RMSE, root mean square errors • WHC, water-holding capacity


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
DETAILED SOIL information is indispensable for land resources management and environmental modeling (Gobin et al., 2000). In most cases, the soil information is derived from soil maps produced by conventional soil surveys (Ahn et al., 1999). However, detailed soil maps cover only very limited areas and its production is costly (McKenzie et al., 2000). Modeling tools using geographic information systems (GIS) require more information than available even in detailed soil maps (Ahn et al., 1999; Cook et al., 1996; Gessler et al., 1995). Soil maps show polygons in which changes in soil properties are considered to be abrupt, when in most cases changes in properties are gradual across the boundary and generally across the polygon as well (Zhu et al., 2001). Since the spatial distribution of soil properties is important for many applications (Young and Hammer, 2000a), alternative approaches are required (McKenzie et al., 2000; Cook et al., 1996; Moore et al., 1993).

Topographic attributes control the differential distribution of water, sediments, and dissolved material, which in turn result in soil differentiation (Pachepsky et al., 2001; Young and Hammer, 2000a; Girgin and Frazier, 1996). Soil surveyors often map soils by interpreting the implicit relationship with landscape characteristics, but this is usually a qualitative approach (Zhu et al., 2001; McKenzie et al., 2000; Hudson, 1992). Some of these relationships have been quantified by measurements over small areas, but to understand soil-landscape relationships, more rigorous models are still required (Gessler and Chadwick, 1997; Girgin and Frazier, 1996). Digital elevation model is one of the important GIS applications from which landscape attributes are derived and utilized in landforms characterization (Dobos et al., 2000; Caccetta, 1997). Different statistical models have been used to establish quantitative relationships between landscape attributes and the distribution of soil properties (McKenzie et al., 2000; Gessler et al., 1995). The basic assumption behind these models is the link between landform characteristics and soil properties, which enable the inference of the latter by knowing the former (Zhu et al., 2001). This approach is helpful on soil survey for explaining relationships between landscapes and soil properties, estimating the uncertainly area associated with soil mapping activity, providing more easy and rapid updates of soil maps, providing better representation of the gradual and continuous changes in soil properties, providing digital information product, and reduced cost (Zhu et al., 2001; Cheng et al., 1997).

Many researchers have found satisfactory statistical relationships between easily DEM-derived terrain attributes and different soil attributes, but in most cases for specific, localized landscapes. Soil depth was significantly correlated (R2 = 0.30) with slope angle, absolute and relative height (Goodman, 1999). Soil depth, A-horizon depth, and presence/absence of E horizon were correlated with plan curvature, compound topographic index (CTI), and upslope mean plan curvature (Gessler et al., 1995). Models that utilize only CTI could account for 84 and 71% of variations in soil depth and A-horizon depth, respectively (Gessler et al., 2000). The spatial distribution of E horizon can be estimated using the terrain variables surface curvature, slope, and aspect (Girgin and Frazier, 1996). Moore et al. (1993) found that slope and wetness index accounted for one-half of the variability in A-horizon depth, sand content, and other soil properties. Pachepsky et al. (2001) indicated that slope, tangential, and profile curvatures are good predictors of soil texture, and explained more than 60% of variations in soil water content at 10 and 33 kPa.

Researchers have utilized different statistical methods to study the above relationships, such as multiple regression, stepwise regression, stepwise principle component regression, and correlation analysis (Dai-Fuchu and Lee, 2002; Chaplot et al., 2001; Dobos et al., 2000; Gessler et al., 2000; Gobin et al., 2000; Moore et al., 1993). Generally, linear regression terms are preferred because of more generality that enables prediction at regional scales, more understandable relationships, and the possibility of accounting for estimations of errors (Gessler et al., 1996). Researchers also have been investigating the use of terrain attributes to produce classes, through cluster analysis, and to link these classes with the distribution of individual soil attributes (Lark, 1999; De Bruin and Stein, 1998; Cheng et al., 1997; Gessler et al., 1996; Bell et al., 1994), or with the distribution of taxonomic soil classes (Zhu et al., 2001; Giles and Franklin, 1996; Irvin et al., 1995; Edmonds et al., 1985). Gessler et al. (1996) suggested that the development of quantitative and continuous predictions of individual soil attributes rather than the use of taxonomic soil classes enable broader spatial application of the established relationships.

Classification based on landform variables significantly enhances the ability of these variables to predict soil water content and sand and clay contents (Lark, 1999). An overall agreement of 70% in recording drainage class was achieved using a soil-landscape model, compared with 50% agreement using published soil survey report (Cheng et al., 1997). A study by De Bruin and Stein (1998) indicated that fuzzy c-means clustering of terrain attributes enhances conventional soil-landscape modeling. A high degree of association between three clusters and measured topsoil clay (R2 = 0.68) was reported.

There are certain aspects that need more investigations to improve these models and enhance their applicability. For example, the magnitudes and patterns of variations in soil properties within specific landforms, the extrapolation of models for wider areas than those used to establish a particular model and the relationship between soil attributes derived from clusters (produced from classification of terrain attributes) and those derived from conventional soil mapping (Gessler et al., 2000; Young and Hammer, 2000a, 2000b). The objectives of this research were to investigate the relationships between digital terrain attributes and soil attributes over a relatively large area, and, to establish models that predict soil attributes using terrain attributes with the minimum input from the field. The ultimate use of this approach is to provide a tool for making land-use decisions.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
The study area is located in the northern part of Jordan between the latitudes 32°22' and 32°45' N and the longitudes 36°22' and 36°45' E (Fig. 1) . It is centered to the northeast of Mafraq city, covering an area of 148 km2. The bulk of the area is dominated by Calciorthids with inclusions of Camborthids (USDA, Soil Conservation Service, Soil Survey Staff, 1990). Generally, it is very gently undulating lava plain with slopes ranging 1 to 4% and altitudes within the range 650 to 750 m above sea level. Although the area is covered by gently sloping land, high variability in soil depth and stoniness was recorded among different slopes. Annual precipitation is about 175 mm. Most of the soils have transition aridic-xeric moisture regime and thermic temperature regime. The study area is covered by detailed soil survey (scale 1:10000), in addition to 2193 soil observations (borehole or pit) (MoA, 1995). Boreholes involved the digging of small "chisel pits" to the 40- to 50-cm depth, followed by auguring to the maximum depth of the auger or to an impeding layer (rock or large stones). Pits were excavated to 200 cm, or an impeding layer. Soil descriptions were based on the FAO terminology (FAO, 1977). For each soil observation, the following soil attributes and site characteristics were collected: limiting soil depth (depth to 200 cm or to rock or large stones); slope percentage (estimated using the Abbney level); slope shape (using three classes, convex, concave, and linear); texture (determined by feeling and cross-checked with laboratory analysis for some samples); WHC (calculated by considering soil texture and depth of each horizon and corrected to take account of stone and gravel contents); erosion class (erosion severity judged in the field: nil, slight, moderate, and severe); erosion type (dominant type judged in the field: nil, sheet, rill, gully, wind, and undifferentiated); surface cover type (dominant type of non-soil material covering the surface: nil, rock, boulder, stone, gravel, and grit); surface cover percentage (percentage of the dominant surface cover type); aspect (determined using compass).



View larger version (84K):
[in this window]
[in a new window]
  		Fig. 1. The spatial distribution of field observations and the subdivision of the study area into watersheds using the digital elevation model (DEM) for Al-Mafraq study area.

 
Generation of Digital Elevation Model
Contour lines were automatically digitized (using TRACE command in Arc/Info software, version 8.0.2, Environmental Systems Research Institute, Inc. [ESRI 2000]) from topography maps (scale 1:25000), with 5-m vertical intervals. An updated TOPOGRID algorithm in Arc/Info GRID was used to create a 20-m resolution DEM. The algorithm is based on the ANUDEM program (Hutchinson, 1989), updated to account for the magnitude of random errors in the input data and to control the degree of data smoothing. Filling the sinks using the FILL command generated a depressionless DEM. The accuracy of the generated DEM was verified by three methods. First, following the USGS procedure (United States Geological Survey [USGS], 1997), the RMSE was calculated using 33 spot heights digitized from the original topography maps. Second, by calculating the elevation difference between interpolated contours and the original contours. Third, by visually comparing digitized streamlines with those derived from the DEM.

Derivation of Terrain Attributes From Digital Elevation Model
The following terrain attributes were derived from the DEM using standard commands in Arc/Info GRID module: slope (percentage and degrees); profile, plan, and mean curvatures; aspect; and the contributing area (accumulation area). To assess the accuracy and reliability of the DEM-derived terrain attributes, the derived values of slope, curvature, and aspect were compared with field-recorded values for each soil observation. The above attributes were calculated based on a window of nine pixels surrounding each pixel. This does not considers the characteristics of the upslope contributing area of each pixel, nor does it consider the relative position of each pixel within the toposequence. Therefore, the average value of these terrain attributes for the upslope contributing area of each pixel was calculated. This was done in Arc/Info by repeating the calculation of the upslope contributing area, each time one of the above attributes' grids was considered as a "weighting grid." The average value assigned for any pixel is calculated from the values of all pixels of the weighting grid located within the contributing area of that pixel. In addition to the above attributes, the average relative altitude of the upslope contributing area (relief potential) was calculated (Caccetta, 1997). Compound topographic index (CTI) for each pixel was also calculated using the average upslope contributing area (As) and the slope degree (B), according to the formula (CTI = ln [As/tan B]) (Moore et al., 1993).

A stream network was derived by connecting all pixels that accumulate flow from 100 pixels or more. Flow accumulation grid and digitized outlets from the stream network were used to automatically subdivide the whole area into small watersheds. Each watershed was subdivided into two facets, separated by the streamline passing through the watershed. Field observations with flow accumulation more than 100 pixels were excluded from the analysis. Observations close to a stream are usually selected to represent the adjacent area. However, because of the coarse resolution of the DEM (20 m), these observations were considered as part of the stream, and consequently were assigned larger contributing areas than the actual.

Statistical Analysis
Different statistical analyses were utilized to investigate the relationships between soil attributes and terrain attributes derived from the DEM. First, the correlation between DEM-derived terrain attributes and soil attributes was examined using the whole set of the field observations. Stepwise and multiple linear regression were used to verify the dependence of the soil attributes (dependent variables) on the terrain attributes (independent variables). Second, an unsupervised classification algorithm was used to group the derived terrain attributes into clusters. In Arc/Info this is done using the GRID ISOCLUSTER command, which is based on the ISODATA algorithm (Iterative Self-Organizing Data Analysis Techniques). This process produces a signature file that is used as input to a maximum likelihood classifier (MLCLASSIFY), which classifies each pixel into single class. There is no rule for the selection of the number of classes to cluster the data. Therefore, the procedure was repeated many times with various numbers of classes. Regression analysis between soil depth and the membership of soil observations in different classes was used to evaluate the clusters produced each time. It was assumed that the most appropriate classification produces the largest regression coefficient because the classes would have low within-classes variability and high among-classes variability (De Bruin and Stein, 1998; Giles and Franklin, 1996). Before running the classification procedure, two tasks were undertaken. The first was the selection of the terrain attributes that were used as an input for the clustering process. The terrain attributes that showed high correlation with soil attributes and low correlation with other selected terrain attributes were selected, these were: slope percentage, slope aspect, CTI and the average curvature of the upslope contributing area. Second, all attributes, except the CTI, were transformed using log transformation to create a Gaussian distribution, because this is an important assumption for this type of classification (Irvin et al., 1995).

The third analysis was to verify the relationship between terrain and soil attributes within each watershed. A watershed is comprised of continuum of hillslopes, which are useful to understand flow pattern and connectivity, and hence the relationships between landscape and soil pattern (Moore et al., 1993). To investigate the role of watershed subdivisions in the soil-landscape relationships, two approaches were used. First, by repeating the regression analysis between terrain attributes and soil attributes within each watershed and within each watershed subdivision (facet), which contained a substantial number of soil observations (at least 15 observations). The other method was to classify watersheds based on their characteristics (area and slope), using an unsupervised classification procedure, and to run multiple regression analysis within each class. The regression coefficient between soil depth and terrain attributes within each class was used as the criteria to choose the optimum number of classes. Regression models were established using 32 observations within each class selected randomly. The models were applied in ArcInfo GRID to predict the soil attributes within each class, and then by compiling the predictions for all classes, layers of predicted soil attributes for the whole area were generated. The prediction accuracy was verified by comparing the predicted and measured values using the whole set of observations, which provides an unbiased testing procedure.


   RESULTS AND DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
CONCLUSIONS
 REFERENCES
 
Accuracy of the Generated Digital Elevation Model
The root mean square errors (RMSE) of the difference between the 33 spot heights and elevation values estimated by the DEM was 1.2 m. This indicates high accuracy considering the resolution of the DEM (20 m). In several studies the errors in vertical elevation estimated by DEMs were typically between 0.5 and 1.5 times the pixel size (Giles and Franklin, 1996; Bolstad and Stowe, 1994). Furthermore, the recommended error for Level 1 USGS 7.5 min DEMs is <7 m (USGS, 1997). The RMSE for the difference between elevation values of the original contour lines and elevation from DEM was 0.92 m. The visual comparison between digitized streamlines from topography map and those generated from DEM indicated good coincidence and match between both layers. These results indicate that the generated DEM was accurate in representing terrain elevation and suitable to undertake the subsequent analysis.

Accuracy of Deriving Terrain Attributes from Digital Elevation Model
The results indicated that the slope percentage values that were derived from the DEM are in agreement with field estimates for 26.7% of the observations if no difference is allowed, while this agreement increased to 66.4% if a slope difference of 1% is allowed between DEM and field estimates. The standard deviation between DEM-derived slope values and slope recorded in the field was 1.4%. These results compare favorably with those reported by Barringer and Lilburne (1997), Bolstad and Stowe (1994), and Giles and Franklin (1996). Researchers indicated that slope values derived from DEM might be better than those estimated in the field, because the former utilizes fixed length to estimate slope, while the latter is based on an integrated slope length judged by the observer (Ziadat et al., 2003; Barringer and Lilburne, 1997). The results indicated that the DEM-derived curvature and aspect were in agreement with field estimates for 50.3 and 34.6% of the observations, respectively, if no difference was allowed, while 96 and 59.2% of the observations' curvature and aspect, respectively, agreed with field estimates if a difference of one class was allowed. These results confirm the ability of the DEM to estimate terrain attributes and the possibility of using these in the subsequent analysis.

Statistical Analyses
The result shows a significant but rather low correlation between soil attributes and terrain attributes (Table 1). The ability of all terrain attributes to predict any of the soil attributes was also very low, with a maximum R2 of 0.19 for the surface cover percentage. These results contradict with findings reported by other researchers (Pachepsky et al., 2001; Gessler et al., 2000, 1995; Goodman, 1999). The results reported by these researchers were drawn from small areas and, in most cases, from a single landscape facet. In this study, the relationships are investigated over a larger area that is extended over many watersheds, which encompasses high heterogeneity of landscape-related processes. Therefore, these results are not surprising because a relationship that is valid for an individual landscape setting is not necessarily valid for others. It has been indicated by other researches that the soil-landscape relationships were not examined for large scale areas, and hence, the extrapolation of models established for an individual hillslope with given characteristics to other hillslopes needs some investigations (Gessler et al., 2000; Young and Hammer, 2000b). Johnson et al. (2000) examined the relationship over large area (214 ha) and reported that five terrain attributes were ineffective in predicting the soil chemical properties; together they explained only 4 to 25% of the variance in these properties.


View this table:
[in this window]
[in a new window]
  		Table 1. Correlation and regression coefficients between soil attributes and terrain attributes using the whole set of field observations.

 
To investigate the above explanations, the relationships between soil and terrain attributes were investigated within selected watersheds rather than for the whole area. Watershed boundaries and the watershed subdivisions (facets) were delineated from the DEM (Fig. 1). The regression coefficients between terrain attributes and soil attributes within each of the selected watershed (Table 2) indicate much stronger relationships compared with the above results. For example, the minimum regression coefficient for surface cover percentage is 0.28 (watershed No. 102) (Table 2), while for the whole study area the regression coefficient for the surface cover percentage was the highest among other variables and was only 0.19 (Table 1). Furthermore, the relationships were further strengthened when investigated within each of the two facets rather than within the whole watershed (Table 2). For example, regression coefficient increased from 0.46 for predicting soil depth within Watershed No. 300 to 0.54 and 0.68 within Facet 1 and Facet 2 of the same watershed. These findings indicate that terrain attributes could be used to predict soil attributes, given consistent and uniform landscape-related processes that influence soil variation are represented. This was achieved by minimizing the range of variability in terrain attributes, which also coincided with the range of variability in soil attributes, within each watershed subdivision. Many researchers emphasized the role of terrain attributes in controlling the movement of water and material over the landscape, which in turn, control the variability of soil attributes. These findings were mostly based on two-dimensional models, or three-dimensional models for a single hillslope facet (Pachepsky et al., 2001; Dobos et al., 2000; Gessler et al., 2000; Young and Hammer, 2000a; Gessler and Chadwick, 1997; Girgin and Frazier, 1996; Brubaker et al., 1993). Watershed subdivisions used in this study could be viewed as a three-dimensional representation of hillslope models, which maintain a systematic flow of water and sediments in one direction over the landscape. Therefore, it provides consistent relationships between the position and characteristics of the terrain units on one hand, and the soil characteristics of the terrain unit on the other hand.


View this table:
[in this window]
[in a new window]
  		Table 2. Regression coefficients between soil attributes and terrain attributes within selected watersheds and watershed subdivisions using the number of points indicated for each line.

 
In spite of the good regression coefficients obtained here, there are two limitations that prevent the use of these models to predict soil attributes. The first is the variability of regression coefficients among watershed and among watershed subdivisions. This indicates variability in the soil-landscape relationships among these subdivisions. This is not surprising because each watershed subdivision (representing landscape facet) is unique in terms of terrain components and the sequence in which these components are arranged, and therefore in terms of landscape processes that differentiate soils. The second, the watershed number (a serial number) assigned here does not indicate any meaning in terms of the watershed characteristics, and therefore, the extrapolation of these models for other watersheds is impossible. An attempt of using one model established for any watershed to predict soil attributes for the whole study area revealed very low prediction accuracy. Two approaches were investigated to overcome these limitations. First, by applying cluster analysis to group terrain attributes into natural clusters within which variability in soil attributes can be reduced. The result indicates a significant but weak correlation between soil attributes and the membership of soil observations in classes produced by cluster analysis (Table 3). Generally, cluster analysis that produces 70 classes has the higher correlation coefficients, ranging from 0.22 for the surface cover percentage to zero for soil texture. The rather low correlations here contradict the idea of high correlations between the membership of an observation in clusters and some soil properties, reported by other researchers (Lark, 1999; De Bruin and Stein, 1998; Cheng et al., 1997; Gessler et al., 1996). Again, the results reported by these researchers were drawn from small study areas and from a single landscape facet. It can be concluded that cluster analysis implemented in this study was not efficient in capturing soil variability for such a large area, even with a wide range of classes used (from 10 to 100). It can be argued that a continuous (fuzzy) classification, implemented by other researchers (De Bruin and Stein, 1998; Irvin et al., 1995), might be better than the unsupervised classification used in this study. However, the advantage of the unsupervised classification is the capability of providing summary information about classes that is easily viewed and understood in polygonal form (Irvin et al., 1995).


View this table:
[in this window]
[in a new window]
  		Table 3. Correlation coefficients between soil attributes and the membership of soil observations in classes derived by using an unsupervised classification based on the homogeneity of terrain attributes and using different number of classes.

 
The second approach was the clustering of watershed subdivisions based on some important characteristics (area and average slope percentage). The analysis was repeated to generate different numbers of classes (from 15 to 40), which was verified to select the best classification. Twenty-six classes were found optimum for the study area because this classification indicated the highest regression coefficient between soil depth and terrain attributes within each of the 26 classes shown in Fig. 2 . Linear models that utilize terrain attributes to predict each soil attribute within each class were generated. An example of these models to predict soil depth indicates that, for all classes, the models are significant at the 0.05 probability level (Table 4). Some classes (Class 3, 9, 11, 19, and 20), indicated low regression coefficient, but these occupy a relatively small area. The predicted soil attributes using these models were compared with field values to estimate the accuracy of the models.



View larger version (109K):
[in this window]
[in a new window]
  		Fig. 2. The watershed subdivisions classified into 26 classes using an unsupervised classification, and the average slope and area of each class.

 

View this table:
[in this window]
[in a new window]
  		Table 4. Regression models to predict soil depth for each of the 26 classes derived from an unsupervised classification based on the average slope and area of watershed subdivisions.

 
The results for the continuous variables indicate a low number of observations where soil attributes were correctly predicted, being 2.3, 1.0, and 6.6%, for soil depth, WHC, and surface cover percentage, respectively (Table 5). However, the percentage of agreement increases sharply as the difference between predicted and measured attributes increased. For example, soil depth indicated 34.5% agreement within ±10 cm and 89.3% agreement within ±50 cm. Comparing this result with the best estimate of soil depth that could be achieved using the published soil map (scale 1:10000), clearly highlights the benefits of this approach. The soil series in the published soil map give an estimate of soil depth classified into three groups; shallow (<50 cm), moderately deep (50–90 cm), and deep (>90 cm). To illustrate the accuracy of the predicted depth values, the soil depth of the 40 observations where predicted and measured soil depth is equal (zero difference) were compared with values of soil depth from the soil map. For only 23 observations out of the 40, the soil depth value agreed with the soil map. For example, the predicted and measured soil depth for one observation is 70 cm, while the same observation is classified as deep soil (>90 cm) by the soil map. Furthermore, the spatial variability of soil attributes within the mapping units is not indicated. The approach followed in this study predicted soil depth with good accuracy, and most importantly, shows the spatial distribution of soil attributes. This is an important requirement for generating accurate site-specific information for land resources management and modeling applications. The spatial distribution of the predicted soil depth clarifies some relationship between soil depth and the position on the terrain and its characteristics (Fig. 3) . Not all of the underlying terrain characteristics (slope, curvature, aspect ... etc.) are clearly shown in this view. However, these were used to generate this continuous distribution of soil depth, and would certainly be useful in clarifying the pattern. The agreement for WHC follows a similar pattern of the soil depth, with slightly lower percentages. This is because WHC is calculated from other soil attributes (soil depth, texture, and stone content), and therefore, its variation depends also on the variations of these soil attributes. The agreement for the surface cover percentage was the highest because it is a surface property and therefore, its relationship with terrain attributes was clearer. Park and Burt (2002) indicated that topsoil properties were better correlated with terrain attributes than subsoil properties.


View this table:
[in this window]
[in a new window]
  		Table 5. Agreement between measured and predicted (using regression models similar to that in Table 4) values of soil attributes.

 


View larger version (162K):
[in this window]
[in a new window]
  		Fig. 3. The relationship between the position and characteristics of the terrain and the spatial distribution of the predicted soil depth using regression models.

 
The ordinal soil attributes also indicated good agreement between predicted and measured soil attributes (Table 6). The models correctly predicted 59.4% of the observations for the surface cover type, 18.6% for erosion type, 61.9% for erosion class, and 53.0% for soil texture. Furthermore, the models predicted, within ±1 class, 94.5% of the observations for the surface cover type, 48.4% for erosion type, 98.0% for erosion class and 90.3% for soil texture. Erosion type has the minimum agreements because it is a result of interaction between many soil and terrain attributes, and therefore is not directly linked to terrain attributes only.


View this table:
[in this window]
[in a new window]
  		Table 6. Agreement between measured and predicted (using regression models similar to that in Table 4) classes of ordinal soil attributes. (Note: non-parametric [Spearman] correlation between predicted and measured classes indicated highly significant correlation at the 0.01 level.)

 

   CONCLUSIONS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
REFERENCES
 
The results of this study indicated that some DEM-derived terrain attributes could be used as predictors of soil attributes. Subdividing large areas into smaller watershed subdivisions (landscape facets) and implementing multiple linear regression models that utilize terrain attributes, within these subdivisions, is a promising approach to undertake this prediction. Each subdivision encompasses a unique set of characteristics and sequential distribution of terrain units, which results in better representation of the soil-landscape relationships. Comparing the predicted values of soil attributes with estimates from a published soil map indicated an acceptable result. In addition, the spatial distribution of the predicted soil attributes is provided in a better resolution than what the soil map provides. This is important for modern environmental modeling tools and site-specific management activities. However, it should be emphasized that these predictions are not a substitute of soil maps. Nevertheless, the generated predictions would be very useful to supply valid information for specific-purpose modeling activities. Particularly in areas where soil maps and field observations are both unavailable, such models to predict soil attributes might be indispensable to make decisions about the utilization and management of land resources. Future work is recommended to apply the approach over larger areas that encompass more diversity of landscape features to validate their results and widen their applications. More importantly, the benefits or shortcomings of utilizing the predicted values as inputs to land resources management activities, such as land evaluation, must be verified, as this represents the end use of these predictions.

Received for publication October 6, 2003.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 





This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Ziadat, F. M.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Ziadat, F. M.
GeoRef
Right arrow GeoRef Citation
Agricola
Right arrow Articles by Ziadat, F. M.
Related Collections
Right arrow Frozen Soils
Right arrow Predictive soil mapping
Right arrow Soil Models

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
The SCI Journals Agronomy Journal Crop Science
Vadose Zone Journal Journal of Plant Registrations
Journal of Natural Resources
and Life Sciences Education		 Journal of
Environmental Quality