GIS and Risk Assessment:
A Fruitful Combination
William W. Hargrove, Daniel A. Levine,
Michael R. Miller, Phil R. Coleman,
Daniel L. Pack, and Richard C. Durfee
Results from evaluations of human health risks associated with environmental contamination are traditionally presented non-spatially. Because of the expense associated with sampling and analysis, samples of environmental contaminants are often taken from relatively few spatial locations, such as test wells. Moreover, in many cases, the environmental gradient of contamination is known to be anisotropic due to directional forces like groundwater flow or wind. In an effort to avoid presenting misleading information, non-spatial tabular reporting of single values, such as increased incidence of human cancer, has become the widely-accepted convention for communication of human health risk results.
We are exploring ways that GIS can be used to recover the spatial component of risk without extrapolating beyond the known data. Ultimately, interpolation of contaminant data which are rare in space and time is necessary for full evaluation of human risks. Spatial interpolation techniques make assumptions, and may therefore be misleading, incomplete, or incorrect. But to encapsulate human health risk into a single value in a table may be at least as incomplete or misleading, since the spatial relationships among contaminant values are not retained. The presentation of an interpolated contaminant layer together with bar and pie chart symbols placed at actual sample locations distinguishes between measured and derived concentration values, and provides a means of qualitatively evaluating uncertainty. Such a presentation also communicates the spatial weighting of the sampling design.
We present an alternative spatial format for the communication of contamination and risk; an array of maps which we call a 'Map Spreadsheet.' The Map Spreadsheet is analogous to a data spreadsheet, except that each cell is a spatially co-registered interpolated map of contaminant concentration. Columns and rows in the Map Spreadsheet could represent alternative ingestion pathways, chemical classes and/or species of contaminants, or years through time, for example. Just as a typical spreadsheet might have a column to the right that represents a row sum or average, or a row at the bottom that represents a column sum or average, so the Map Spreadsheet has column and row totals represented by map algebraic sums or averages of the maps in that row or column.
Thus, it would be possible, for example, to view a site contaminated with cesium as a series of yearly maps, as well as a map summarizing average cesium concentration over all years. Alternatively, if columns represent risk due to alternative ingestion pathways, the spatially-explicit risk from cesium inhalation, ingestion, and direct exposure, as well as a summary map of cesium risk from all pathways, could be simultaneously viewed. The 2-dimensional map spreadsheet analogy can be extended into the third dimension as a 'stack' of such spreadsheets, creating a map cube. Three full 'faces' of the Map Cube potentially represent summation map 'cells.' Of course, only a single 2-D 'slice' of a Map Cube could be viewed at a time.
The viewer of a Map Spreadsheet obtains a perspective of the combined risk from exposure to multiple contaminants, as well as an idea of the heterogeneity of contamination across space. Such insights are usually not possible with a tabular presentation of single values of human health risk. Several examples, both hypothetical and real-world, are shown to illustrate some of the potentially useful products resulting from the marriage of GIS and risk assessment.
Environmental risk professionals typically report risk as single values or tables of values. Often, separate numbers are calculated and reported for different pathways, each of which are considered independently. No specific spatial information is included, beyond a general idea of the spatial extent to which the evaluation applies.
Yet it is clear that risk assessments have an important spatial component. Risk evaluations are generated by reference to specific environmental data collected from specific locations. Environmental samples may be scarce in space, since they may only be accessible at test wells, for example. Analysis for contaminants may be expensive, further limiting sample size.
It may be a desire to avoid presenting misleading information that causes risk professionals to hesitate to map risk. Contamination scenarios within fluid media are likely to show directional tendencies as that medium flows, causing anisotropic effects in contaminant concentrations. These directional effects complicate interpolation among sample points. In the case of groundwater contamination, the subsurface geology, which is often unknown, has profound effects on subsurface flow and movement of contaminants. In the face of such complex and unaccounted for compounding factors, most risk professionals have elected to present results in a non-spatial way, perhaps thinking that this represents a conservative approach.
We argue that the presentation of risk benefits greatly from a spatial approach, and that the marriage of GIS and risk analysis is a sensible one. Although difficulties and problems certainly exist in the spatial analysis of risk, we suggest that it is a better - and more conservative - strategy to openly present these details and assumptions spatially rather than allowing them to remain implicit. Risk professionals will not mislead by presenting maps - they mislead by not presenting maps.
One challenge for the effective spatial presentation of environmental risk is the juxtaposition of multiple variables within a recognizable spatial context. Risk analyses must deal with widely disparate types of data, including multiple contaminants sampled from multiple locations at different intervals. This multivariate nature of contamination data sets is a double-edged sword; the same multidimensionality that is difficult to present can, when couched in the right spatial context, effectively communicate nuances about the risk evaluation process and lead to insights for making decisions.
One approach that we have used to convey multivariate data spatially is to create a series of hybrid maps which are combinations of charts and maps. This approach embraces the multiple dimensions of the data, attempting to present spatially as many related facets of the data as possible.
Another approach with which we have experimented is to collapse the multiple dimensions of contaminant data down to the single common currency of human health risk. Different types of data are directly comparable when converted directly to risk. Contaminant concentrations usually end up in terms of risk; we simply suggest that this conversion be made early in the analysis, so that the concomitant unification of data can be enjoyed. The ultimate extrapolation of this collapsing approach results in what we call a ``Map Spreadsheet'', in which arrays of maps of risk are spatially summed across rows and columns.
We suggest here a number of techniques for presenting complex multivariate data sets regarding environmental contamination. Although we will make reference to particular environmental risk problems, these references are examples only; our purpose is to illustrate generally-appilcable techniques.
New techniques in GIS are not required to support risk analysis. The spatial presentation of simple univariate data as standard maps can be revealing for risk evaluation. For example, this map of population density relative to Oak Ridge National Laboratory is useful for risk estimations simply because it locates population centers within 50 and 100 miles of Department of Energy facilities. The map clearly shows a non-uniform distribution of population relative to the DOE facilities. Before any additional considerations, this map indicates that risk (or at least cost) will not be uniform in all directions with distance. Terrain forms and hydrographic features shown on the map might represent significant conduits for human population exposure. One could easily imagine more sophisticated analyses which considered, for example, population movement with time of day.
The placement of icons on maps is not a new idea; it is as old as cartography itself. In its simplest form, this approach includes, within a standard map, small icons at particular sampled locations. The size and color of the icons is used to reflect location and concentration of various contaminants.
This series of maps shows the spatial relationships bewteen chemical concentrations in a series of test wells and a manufacturing facility, a series of buried pipelines, and a group of holding ponds. Concentration data are presented spatially for 1,1-dichloroethene, trichloroethene, and vinyl chloride in this K-1420 industrial area. Although we chose to plot each contaminant on separate maps, the icons could have been combined on a single map, and differentiated by the shape of the icons used.
Only slightly more sophisticated is the plotting not of simple contaminant concentrations, but of risk due to the contamination at that concentration. In this map of total human health risk due to vinyl chloride contamination, human risk values calculated for the water ingestion pathway are mapped directly. Mapped units are in terms of likelihood of ``excess'' human cancers.
All of these maps present contamination information against a referential backdrop of recognizable features, including roads, buildings, pipelines, and bodies of water. It is this spatial context which defines the value of the data; one cannot help mentally comparing contaminant concentrations with their spatial proximity to potentially related features like pipelines and holding ponds, for example.
The combination of charts with maps allows for the inclusion of additional information regarding contaminants. The simple maps above presented a static snapshot of contamination; no information was included about the way that contaminant concentrations changed through time.
We have combined small bar chart icons with maps to spatially present contamination profiles through time in this map of water samples from the Clinch River/Watts Bar Reservoir system. Several contaminants are presented simultaneously, and samples from 3 different times are included with this technique. Map insets allow emphasis on particular widely-separated areas within the large spatial extent of the map.
A longer time series is shown for multiple contaminants in this map of contaminant concentrations in test wells from the K-1420 area. Histogram bars are shown for a series of ten measurements made between 1986 and 1995. Color as well as labels serve to distinguish among various contaminants.
All maps to this point have shown total contamination or risk. By plotting size-dependent pie diagrams at sampled locations, we can communicate not only the total but the proportion of total contamination or risk contributed by each of several contaminants. This map of the X-10 area shows proportional human health risk by all pathways posed by test well concentrations. The size of each pie reflects total relative human health risk in terms of increased incidence of human cancer, while the color and proportion of each pie slice indicate each contaminants' contributions. The locations of each data point are maintained, so that the spatial distribution of the test wells and their proximity to prominent features is obvious.
This technique only works when disparate types of contaminant data are unified by conversion to risk. Because of the inequalities among individual contaminants, this technique cannot be used with absolute concentrations. Concentration levels causing concern and requiring action differ widely among contaminants; including raw concentrations in the same pie would be comparing apples and oranges. Conversion to total risk allows direct comparison in the terms that are most immediately important.
Interpolation, implicit or explicit, is inherent in risk analysis. Risk analysts are already extrapolating and interpolating from limited environmental data, whether mentally or with some software tool. It may be better to be objective than subjective.
The selection of the interpolation technique is important. Contaminants may not be equally likely to move in any direction; assymetric probabilities of movement, or anisotropy, may be caused by the flow of the surrounding medium of air, water, or groundwater. Such flow will draw the contaminant along preferentially in a single direction, creating a plume.
Simple kriging is not likely to be adequate for interpolating under such circumstances. Radial basis function interpolation methods, such as regularized spline with adjustable tension, do not assume stationarity, and allow the underlying statistical distributions to vary as called for by local data as needed across the area, rather than relying on a single semivariogram to represent the spatial autocorrelation in all directions across the entire map. These methods have been developed by Lubos Mitas at NCSA , (see also Mitasova and Mitas 1993, Mitasova et al. 1995) and Mitas and Mitasova 1988, Computers and Mathematics with Applications, v16, p. 983.
In this particular example of risk distribution at the K-25 facility based on sample well data, a simple triangular irregular network was constructed with the sample well locations, and a quintic polynomial surface was created to extrapolate a risk surface from the known data. The surface was converted to integer values, and an isoopleth map was drawn.
At least two gross simplifying assumptions were made in producing this map; the spatial distribution of risk was assumed to be symmetrical in all directions, and the risk medium was assumed to be uniform and homogeneous. Clearly, both of these assumptions are violated, since a river and a creek pass through the interpolated area. In a more sophisticated approach, special considerations could be made for spreading risk directionally, through rivers and streams versus groundwater, etc. But, at some level, we still gain insight from the simple interpolation model over what we would have from looking at a map of spot data.
In the K-25 risk map, one can easily see the spatial arrangement of sample locations (the black dots are sampled test wells) within the interpolated area for which the risk assessment is being made. By combining this interpolation approach with some of the techniques already described for the presentation of multivariate data, we could have shown the actual values at each well, so that users could have judged for themselves the accuracy with which the synthetic interpolation represented the actual measured data values.
Extending this idea, we could generate maps quantifying such error or uncertainty. Although not specifically for risk, we have presented a similar map on the Clinch River Web page, specifically at URL http://www.esd.ornl.gov/programs/CRERP/SEDIMENT/PINHEAD.HTM. Far from hiding uncertainty in interpolated maps, these techniques use maps to clearly highlight and quantify interpolation error.
Once the value of explicit interpolation in risk assessment is accepted, we can use a familiar mental model, in concert with the powerful tool of map algebra to produce a technique which we have called the Map Spreadsheet. A Map Spreadsheet is comprised of a series of interpolated maps which are co-registered in space. As with a standard GIS stack of overlays, each map is in the same physical location and of the same extent. However, instead of stacking vertically, the maps are spread out in a flat array using another familiar context - that of a digital spreadsheet. Each of the cells in the Map Spreadsheet consist of co-registered maps.
Just as rows and columns can have sums and averages in a standard spreadsheet, so a Map Spreadsheet can use map algebra to sum maps across rows or down columns to produce another row or column of maps representing a spatial additive sum of the other maps in that series. For some map variables, spatial averages may be more appropriate than spatial sums.
We have constructed a Map Spreadsheet for human health risk from organic contaminants in the groundwater contamination in the K-1420 wells. In this example, rows are particular contaminants and columns are years sampled. Some map ``cells'' have no data; not all wells were sampled in all years, and the same wells were not sampled for each contaminant each year.
Despite the confusing nature of the well contamination data set, a number of risk characteristics are evident from the Map Spreadsheet. The vinyl chloride contaminant (top row) contributes most of the risk to human health in this area, with some increased contribution from 1,1-dichloroethene (second row) in later years. Earlier dates tend to have higher concentrations, but this may be an artifact of the historical data set; contaminant concentrations below the detection limit are set to the detection limit, and improvements in analytical capabilities have resulted in a decrease in detectable levels.
A Map Spreadsheet could be constructed using contaminant concentrations, although the technique is most powerful if risk is directly mapped. Plotting concentrations of a particular contaminant precludes direct comparison across various contaminants and exposure pathways, one of the best characteristics of the Map Spreadsheet.
Risk assessments typically consider health risks of single contaminants and particular exposure pathways serially. Yet it is clear that environmantal exposure to suites of contaminants via multiple pathways. The Map Spreadsheet, with its rows and columns of summed risk maps, is a way to present risks from multiple contaminants and pathways in an understandable way.
The Map Spreadsheet permits many other possible array configurations. Columns or rows could include various exposure pathways (e.g. ingestion, inhalation, dermal contact), chemical classes (i.e. radionuclides, organics, metals) or particular contaminants. In fact, the Map Spreadsheet concept can be extended into a third dimension for simultaneous consideration of additional variables. Such a Map Cube would allow, in addition to the solid cube of map ``cells'', 3 complete planes of single sum maps, 6 vectors of double sum maps, and one triple sum map. The triple sum map would represent total spatial risk from all contaminant variables considered in the cube.
The Map Spreadsheet is an ultimate way to present multivariate data. The map array contains more useful information than could be included in any single map. Map Spreadsheets are an ideal format for the spatial presentation of risk assessments.
Although we have concentrated on human health risks here, the benefits of a spatial treatment of risk may be particularly important for ecological risk analysis. Rather than having the luxury of concentrating on human effects, ecological risk analysts face the daunting task of considering effects on endpoints throughout the biotic spectrum. Endpoints for some biota are available at URL www.hsrd.ornl.gov/eco risk/. The Map Spreadsheet allows consideration of risk from multiple contaminants, via multiple pathways, across multiple years, and for multiple biotic endpoints.
It is clearly not sufficient to report risk non-spatially. Risk is an inherently spatial phenomenon. A risk map should be considered the ultimate product of any risk investigation, and should be the first resource sought for any risk decision or evaluation. GIS techniques can be central to these important and critical processes of risk identification, quantification, and evaluation.
Thanks to Wilson McGinn for contaminant and risk data and initial funding to explore these ideas.
Mitas, L., and Mitasova H. 1988. General variational approach to the interpolation problem. Computers and Mathematics with Applications 16:983.
Mitasova, H. 1992. New capabilities for interpolation and topographic analysis in GRASS. GRASSclippings 6(2):13.
Mitasova, H. 1992. Surfaces and modeling. GRASSclippings 6(3):16-18.
Mitasova, H., and Hofierka, J. 1993. Interpolation by Regularized Spline with Tension. II. Application to Terrain Modeling and Surface Geometry Analysis. Mathematical Geology 25:657-667.
Mitasova, H., Mitas, L., Brown, W.M., Gerdes, D.P., Kosinovsky, I., Baker, T. 1995. Modeling spatially and temporally distributed phenomena: new methods and tools for GRASS GIS. International Journal of Geographical Information Systems 9(4):433-446.
Mitasova, H., and Mitas, L. 1993. Interpolation by Regularized Spline with Tension. I. Theory and Implementation. Mathematical Geology 25:641-655.
Rood, S.G, Smith, V.E, Filkins, J.C, Wildhaber, M.L., and Schmitt, C.J. 1994. Assessment and Remediation of Contaminated Sediments (ARCS) Program: Assessment Guidance Document. EPA 905-B94-002. U.S. Environmental Protection Agency, Great Lakes National Program Office, Chicago, IL.
Talmi, A., and Gilat, G. 1977 Method for Smooth Approximation of Data. Journal of Computational Physics23:93-123.
Wahba, G. 1990. Spline Models for Observational Data. CNMS-NSF Regional Conference series in applied mathematics, SIAM, Philadelphia, Pennsylvania.
Research sponsored by Office of Environmental Restoration and Waste Management, U.S. Department of Energy under contract DE-AC05-84OR21400 with Martin Marietta Enargy Systems, Inc.
"The submitted manuscript has been authored by a contractor of the U.S. Government under contract No. DE-AC05-84OR21400. Accordingly, the U.S. Government retains a nonexclusive, royalty-free license to publish of reproduce the published form of this contribution, or allow others to do so, for U.S. Government purposes."
For additional information contact:
