John E. Fels, Kris C. Matson

A Cognitively-based Approach for Hydrogeomorphic Land Classification using Digital Terrain Models

The continuous surface of the landscape is often described in discrete terms -- ridge, sideslope, flat, etc. -- that represent broadly recognizable cognitive types. The classification of land in this manner is useful for a variety of purposes, including the mapping of hydrogeomorphic land types associated with specific ranges of water table depth. These land types can be mapped from digital terrain models by applying decision-tree classification procedures based on pertinent geomorphometric parameters. The classification process is both interactive and interpretive, requiring repeated visualization of the resulting classification maps and adjustment of classification parameters. Moreover, it is specific to physiographic provinces having internally similar topographic and geologic characteristics affecting geomorphic and hydrologic processes.

In this study, hydrogeomorphic land types were mapped for all of North Carolina using digital terrain models with a resolution of 300 feet. Separate land type classifications were developed for each of the six major physiographic provinces of the state -- Blue Ridge, Piedmont, Sandhills, Upper Coastal Plain, Lower Coastal Plain, and Coastal Islands. The resulting classifications were then expressed in terms of expected ranges of water table depth for each land type in each province to provide an essential GIS layer for the statewide DRASTIC mapping program. These land type maps, while awaiting accuracy assessment, may also serve as a useful component of North Carolina's Corporate Geographic Database. Since these land type classifications are cognitively based, they can be broadly interpreted and might be useful in other studies requiring a regional land classification framework.

Introduction

Ground water vulnerability to contamination from the land surface has become an important area of hydrogeological research. To date, most of this research has focused on small sites of special interest using intensive sampling techniques and complex modeling procedures. There is also a need, however, for more extensive evaluations of ground water vulnerability that might help guide land planning and management at regional scales. To that end, the North Carolina Department of Environment, Health, and Natural Resources, Division of Environmental Management has undertaken the mapping of ground water vulnerability throughout the state, based on the DRASTIC system developed by the US Environmental Protection Agency (Aller et al. 1987).

DRASTIC is an acronym that represents seven constituent variables used in creating an overall index of ground water vulnerability to surface contamination. These variables are: average annual depth to the water table (D); net recharge (R); aquifer media (A); soil media (S); topography, expressed as land surface slope (T); impact of the vadose zone (I); and hydraulic conductivity (C). For local applications, these variables are typically determined from on-site measurements, but for regional applications they must be interpreted from existing thematic map layers. For this project, most of these variables could be reliably interpreted from geologic and soil maps but, for depth to the water table, no statewide data were available. It became necessary, then, to develop a new approach to mapping water table depth which would provide information appropriate to the DRASTIC assessment scheme, express this information at a useful level of resolution, and be applicable to very large subject areas.

Several possible mapping approaches were considered, including deterministic modeling, statistical modeling, and landscape classification (Matson and Fels 1996). Although modeling approaches would likely provide the most accurate estimations of ground water depth, they would also require large amounts of field data -- far more than is currently available for the state. In addition, the time and costs required for modeling approaches were beyond the scope of the project. A classification approach, however, could be implemented within the scope of the project using existing data and knowledge. Hydrogeomorphic land classification based on digital terrain models could be applied to large areas at a suitable level of resolution and, since the DRASTIC system only requires ranged estimates of water table depth, would provide the information necessary for vulnerability assessment.

Hydrogeomorphic land classification divides the continuous surface of the landscape into discrete landscape elements (land types) such as ridges, sideslopes, coves, or flats, which are associated a priori with specific ranges of average annual water table depth. Water table depth not only varies among different land types, but among similar land types in different physiographic provinces, so it was necessary to develop a separate classification scheme for each of the major physiographic provinces of North Carolina, reflecting the particular geomorphic settings and hydrologic conditions found in each. Within each province, classification was accomplished by applying a cognitive schema differentiating land types on the basis of pertinent geomorphometric attributes. Since these land types are cognitive elements, classifications were refined through a process of repeated visualization and landscape interpretation. The present classification scheme was developed through consultation with ground water and mapping scientists from the Groundwater Section of the NC Division of Environmental Management, the Water Resource Division of the US Geological Survey, and the Design Research Laboratory at North Carolina State University using professional experience of the relationship of water table depth to different geomorphic settings among North Carolina's physiographic provinces. The classification hierarchy was limited to those land type classifications thought to be achievable using one-degree digital elevation models (the only available statewide) and classification procedures realizable within the scope of the project. It is anticipated that the classification scheme used in this first iteration of the methodology will be revised pending the results of accuracy assessment.

Methods and Results

Land type classifications were derived from 27 one-degree digital elevation models distributed by the US Geological Survey. The one-degree digital elevation models cover areas of one degree of latitude by one degree of longitude with land elevation sampled at intervals of three seconds of latitude and longitude (USGS 1990). These tiles were reprojected, using Arc/Info GRID (ESRI 1994), to the North Carolina State Plane coordinate system with a sampling interval of 300 feet. Altogether, these 27 tiles consisted of roughly 34 million model points. In addition to the gridded elevation data, corresponding tiles were also produced for land surface slope (expressed in slope percent), for landscape position, and for the major physiographic provinces of the state (Figure 1). Within each of the major physiographic provinces -- Blue Ridge, Piedmont, Sandhills, Upper Coastal Plain, Lower Coastal Plain, and Coastal Islands -- procedures for land type classification were based on various combinations of two differentiating criteria: land slope and landscape position.

Figure 1. The major physiographic provinces of North Carolina

Calculation of Landscape Position

Various quantitative methods have been developed for characterizing the morphology of land surfaces (Evans 1972, Mark 1975, Dole and Jordan 1978, Papo and Gelman 1984, Elghazali and Hassan 1986, Zevenbergen and Thorne 1987, McNab 1989, 1993, Fels 1994) and for extracting hydrologic characteristics from digital topographic models (Jenson and Domingue 1988, Skidmore 1990). However, since hydrogeomorphic classifications are not based on morphology alone but also on the position of the land surface in relation to its surroundings, a method recently developed by Fels for ecological land classification was adopted. This method yields a quantitative index of landscape position by evaluating elevation differences between a given point and other model points within a specified search radius (USFS 1995, Fels 1995). Specifically, landscape position is calculated as:

where

E0 = elevation of the model point under evaluation

En = elevation of a surrounding model point

d = horizontal distance between the two model points

n = the total number of surrounding points employed in the evaluation

The value calculated is the mean of the distance-weighted elevation differences between a given point and all other model points within a specified search radius. Greater positive values indicate lower topographic positions (proximal to streams) and greater negative values indicate higher landscape positions (ridges, summits) while values approaching zero indicate mid-slope positions. Where relief is minimal within the search radius, values will also tend to approach zero.

The extent of the search area is an important consideration, since the evaluation of landscape position will be most meaningful when confined to a single landform. In principle, the radius of search should be one-half of the fractal dimension of the landscape, that is, one half of the ridge-to-stream distance in that landscape. Under such circumstances, a point located at mid-slope position will be evaluated with respect to points extending from the stream at the bottom of the slope to the ridge at the top of the slope. Average ridge-to-stream distance varies considerably among different landscapes but is fairly consistent within a particular physiographic province. Estimates of ridge-to-stream distance were obtained for the various physiographic provinces by visualizing digital terrain models within each region, measuring ridge-to-stream distance for a number of typical landforms, and taking the mean of these measurements to obtain a representative value. Average ridge-to-stream distances and search radii for the six physiographic provinces are shown in Table 1.

Table 1. Average ridge-to-stream distances and search radii for the major physiographic provinces.

Algorithms were developed to calculate landscape position index for every point in the digital terrain models, using appropriate search radii within the various physiographic provinces. Applying these algorithms to the one-degree elevation tiles produced 27 binary files of landscape position values . These were then converted to ASCII text files compatible with ARC/Info GRID.

Land Type Classification and Mapping

For each physiographic province, a classification scheme was developed to identify discrete land types and to relate these to estimated water table depth ranges for that province (Table 2). For each land type in each province, depth ranges were determined based on field experience, recorded observations, and values provided in the EPA DRASTIC Guidelines (Aller et al. 1987).

Within each province, land types were classified using a divisive procedure applied in successive stages (decision tree), with different criteria distinguishing land types at each stage. In the Sandhills province, for example, if landscape position at a particular point was less than a critical value, then that point was classified as a hill or ridge, otherwise if land slope exceeded another critical value that point was classified as a sideslope, otherwise it was classified as a flat (Figure 2). Classification procedures for the Coastal Plain and Coastal Islands provinces were simpler than this, while those for the Piedmont and Blue Ridge provinces were considerably more complex. Classification procedures for all six physiographic provinces are described by Fels (1995).

Table 2. Physiographic provinces, land types, and water table depth ranges.

Figure 2. Classification matrix and decision tree for the Sandhills physiographic province

Critical values for differentiating criteria were determined by visualizing the resulting classifications, and comparing these to interpretations of land type based on familiarity with the geomorphology of the landscape and knowledge of ground water behavior in that landscape. Several representative tiles were used to determine critical values for each physiographic province. This process was repeated many times, gradually optimizing critical values and refining the classification procedures.

A special problem arose in the final stages of classification for the Blue Ridge and Piedmont provinces. This required differentiating upland flats from lowland flats based on landscape position. In these provinces, especially in the Piedmont, extensive areas of upland flat are separated from extensive areas of lowland flat by shallow to moderate slopes, forming what appear as very large terraces in the landscape. Since these 'terraces' are considerably more extensive than the typical ridge and valley landforms used to determine the search radii for landscape position in these provinces, the resulting tiles of landscape position were inappropriate for differentiating the larger upland and lowland flats. Thus, lowland classifications sometimes appeared in the center of upland flats, and vice-versa, when the dimensions of these areas were considerably larger than the specified search radius. This problem was redressed by producing a second set of landscape position tiles (termed "relative position") employing a much larger search area and providing a much more general appraisal of landscape position. That variable was employed only in the Blue Ridge and Piedmont classifications, and only to distinguish upland from lowland flats.

Land type mapping was accomplished by applying the classification procedures for each province to the 27 digital terrain models, with repeated visualization of the resulting classifications and adjustment of differentiating values. The final maps of hydrogeomorphic land types (see Figure 3) contained two-digit codes representing both physiographic provinces and land types. Again, the binary output files were converted to ASCII text files compatible with ARC/Info GRID. Since each land type was related a priori to a specific depth range, these tiles could then be reclassified to map water table depth throughout the state. To complete the statewide map of ground water depth, values for hydrologic settings (lakes, wetlands, etc.) were determined using digital hydrographic data from the North Carolina Corporate Geographic Database (NC CGIA 1995).

Figure 3. Classified land type map for a portion of the Blue Ridge physiographic province

Conclusions

These methods represent one solution among a spectrum of possible solutions to the mapping of water table depth for large areas. Other possible approaches include deterministic modeling, trend surface interpolation of ground water levels, regression modeling based on ground water data and larger sets of geomorphometric parameters, and landscape classification based on discriminant analysis. This approach was chosen primarily for its efficacy, given a limited timeframe and the logistic difficulty of assembling sufficient field data for direct statistical modeling. Adequate field data do exist, however, to support an accuracy assessment of the methods and assumptions embodied in this approach and, in the future, the realization and comparison of different approaches using higher resolution digital models and more complex conceptual models.

The methods employed here are clearly heuristic in nature, relying on digital media in the interactive development and testing of classification procedures. The classifications themselves are based on cognitive schemata widely employed by field scientists but, until now, not expressed in a spatially explicit context. While modeling procedures were implemented using custom software designed for that purpose and commercial GIS was used mainly to integrate modeling results, future GIS applications might better serve environmental modeling efforts by facilitating interactive modeling and visualization scenarios. Until GIS becomes a comprehensive modeling platform, environmental scientists will continue to rely on task-specific modeling software coupled with general- purpose geographic information systems.

This project has produced, in addition to its primary goal of water table mapping, a map of classified land types for the entire state of North Carolina. While this classification was designed to reflect the influence of topography on ground water depth, and was not meant to be all inclusive with respect to other fields of interest, it does provide a level of geomorphic description previously unavailable at the state level. In addition to the application described here, land type mapping might provide a useful framework for organizing ecological inventory and monitoring programs, interpreting and analyzing land use patterns, or developing land planning and management strategies. Beyond applications in ground water analyses, this information also might serve as an important component in the study and management of North Carolina's many other natural resources.

Acknowledgments

The authors wish to thank the Groundwater Section of the North Carolina Department of Environment, Health, and Natural Resources, and the United States Environmental Protection Agency, for the opportunity and funding to conduct this work. We thank Carl Bailey, Ted Mew, Arthur Mouberry, and Perry Nelson for conceptualizing and supporting efforts to characterize the hydrogeology of the shallow aquifer system in North Carolina. We also thank Ralph Heath, consultant, and Bruce Lloyd, Charles Daniel, Tim Spruill, and Gerry McMahon of the United States Geological Survey for their insights into regional water table mapping and the approaches discussed here. We are also indebted to Scott Huffman and Kai-Ping Wang of the Groundwater Section for their technical support in this mapping project. Computing hardware was provided in part by Apple Corporation, Community Affairs Division, through their EarthGrants Program.

References

Aller, L., Bennet, T., Lehr, J.H., Petty, R.J., and Hackett, G. (1987) DRASTIC: A Standardized System for Evaluating Ground Water Pollution Potential using Hydrogeologic Settings. Washington, DC: US Environmental Protection Agency. Document EPA/600/2-85-018.

Dole, W.E. and Jordan, N.F. (1978) Slope mapping. The American Association of Petroleum Geologists Bulletin 62: 2427-2440.

Elghazali, M.S. and Hassan, M.M. (1986) A simplified terrain relief classification from DEM data using finite differences. Geo-Processing 3: 167-178.

Fels, J.E. (1994) Modeling and Mapping Potential Vegetation using Digital Terrain Data: Applications in the Ellicott Rock Wilderness of North Carolina, South Carolina, and Georgia. Raleigh, NC: North Carolina State University. Ph.D. dissertation.

Fels, J.E. (1995) Landscape Position and Classified Landtype Mapping for the Statewide DRASTIC Mapping Project. Raleigh, NC: North Carolina State University. Technical report to the North Carolina Department of Environment, Health, and Natural Resources, Division of Environmental Management.

ESRI. (1994) Arc-Info Version 7. Redlands, CA: Environmental Systems Research Institute.

Evans, I.S. (1972) General geomorphometry, derivations of altitude, and descriptive statistics. Spatial Analysis in Geomorphology, ed. Chorley, R.J. New York: Harper and Row.

Jenson, S.K. and Domingue, J.O. (1988) Extracting topographic structure from digital elevation data for geographic information system analysis. Photogrammetric Engineering and Remote Sensing 54: 1593-1600.

Mark, D.M. (1975) Geomorphometric parameters: a review and evaluation. Geografiska Annaler 3-4, Series A: 165-177.

Matson, K.C. and Fels, J.E. (1996) Approaches to automated water table mapping. Proceedings, Third International Conference on Integrating GIS and Environmental Modeling. Santa Barbara, CA: National Center for Geographic Information and Analysis. WWW, CD.

McNab, W.H. (1989) Terrain Shape Index: Quantifying effect of minor landforms on tree height. Forest Science 35: 253-258.

McNab, W.H. (1993) A topographic index to quantify the effect of mesoscale landform on site productivity. Canadian Journal of Forest Resources 23: 1100-1107.

NC CGIA (1995) North Carolina Geographic Data Catalog. Raleigh, NC: North Carolina Center for Geographic Information and Analysis.

Papo, H.B. and Gelman, E. (1984) Digital terrain models for slopes and curvatures. Photogrammetric Engineering and Remote Sensing 50: 695- 701.

USFS. (1995) Ecological Classification, Mapping, and Inventory for the Chattooga River Watershed. Atlanta, GA: US Forest Service, Southern Region, Chattooga Ecological Classification Team. Draft report.

USGS. (1990) Digital Elevation Models, Data Users Guide 5. Reston, VA: US Department of the Interior, US Geological Survey.

Zevenbergen, L.W. and Thorne, C.R. (1987) Quantitative analysis of land surface topography. Earth Surface Processes and Landforms 12: 47- 56.