• E-mail
• Print
• Comment
• Font Size
• Digg
• del.icio.us
• Discuss article

Patterns of Watershed Urbanization and Impacts on Water Quality1

Posted on: Saturday, 16 July 2005, 03:00 CDT

ABSTRACT:

Urban runoff contributes to nonpoint source pollution, but there is little understanding of the way that pattern and extent of urbanization contributes to this problem. Indicators of type and density of urbanization and access to municipal services were examined in six urban watersheds in Durham, North Carolina. Principal components analysis (PCA) was used to identify patterns in the distribution of these variables across the urban landscape. While spatial variation in urban environments is not perfectly captured by any one variable, the results suggest that most of the variation can be explained using several variables related to the extent and distribution of urban development. Multiple linear regression models were fit to relate these urbanization indicators to total phosphorus, total kjeldahl nitrogen, total suspended solids, and fecal coliforms. Development density was correlated to decreased water quality in each of the models. Indicators of urbanization type such as the house age, amount of contiguous impervious surface, and stormwater connectivity explained additional variation. In the nutrient models, access to city services was also an important factor. The results indicate that while urbanization density is important in predicting water quality, indicators of urbanization type and access to city services help explain additional variation in the models.

(KEY TERMS: urban water management; watershed management; water quality; land use planning; nonpoint source pollution; nutrients; runoff.)

INTRODUCTION

Since the U.S. Congress passed the Clean Water Act in 1972, environmental managers and policy makers have made substantial headway in identifying and reducing point source discharge into the nation's waterways (USEPA, 2000). Yet many of these water bodies remain impaired, often because of nonpoint source pollution (Carpenter et al., 1998). According to the U.S. Environmental Protection Agency (USEPA), nonpoint source pollution is the reason that as many as 40 percent of the nation's surveyed water bodies are not meeting their designated uses (USEPA, 1996). Nonpoint source pollution from urban areas poses a particular problem for water quality management because responsibility is spread among entire populations, complicating source identification and reduction.

The impacts of urbanization on water quality have been well documented. Regional case studies demonstrate that urban streams exhibit increased levels of phosphorus (Novotny, 1991; Soranno et al., 1996; May et al., 1997), nitrogen (Novotny, 1991; Lenat and Crawford, 1994; McMahon and Harned, 1998; Basnyat et al, 2000), total suspended solids (Novotny, 1991; May et al, 1997; McMahon and Harned, 1998), biochemical oxygen demand (BOD; Fitzpatrick, 1995), metals (Lenat and Crawford, 1994; Mumley, 1995; Fitzpatrick, 1995; May et al, 1997; Bhaduri et al, 2000), oil and grease (Fitzpatrick, 1995), and fecal coliform bacteria (Schueler, 1994; Duda et al., 1998) relative to reference streams. While specific sources are difficult to identify, potential pollutants in urban environments include street litter, end products from fossil fuel combustion, rubber and metal eroded from vehicles, corrosion of galvanized roofing materials, pet wastes, and fertilizers and pesticides applied to lawns (Loehr, 1974; Cole et al., 1984; Bannerman et al, 1993).

In urban watersheds, water chemistry depends not only on the extent of urbanization but also on the type of urbanization (Paul and Meyer, 2001; Strayer et al., 2003), its arrangement relative to other land uses, and differences in management practices (Jones et al., 2001). Percent impervious surface cover can be used to represent a continuum of urbanization not taken into account by traditional land use classifications. For example, a study conducted in the city of Olympia, Washington, found that urban areas ranged from 10 to 90 percent impervious depending on the specific land use. Many studies suggest that impervious surfaces exhibit a threshold effect, with major stream degradation occurring where as little as 10 percent of the surface area is impervious (Brun and Band, 2000; Wang et al., 2001; Jennings and Jarnagin, 2002). Impervious surfaces can be further classified based on the type of surface, its function, and its connectivity to the stream. Runoff from streets and parking lots will have different properties, from a water quality standpoint, than runoff from residential rooftops (Arnold and Gibbons, 1996). Additionally, impervious surface that is directly connected to the stream via a stormwater drain will have a greater impact than impervious surfaces that drain to an adjacent lawn (COPWD, 1995; Booth and Jackson, 1997; Wang et al, 2001).

Landscape metrics quantifying the distribution of land use classes have been used to examine the relationship between landscape pattern and accepted measures of stream health (Hunsaker and Levine, 1995; Richards et al, 1996; Jones et al, 2000, 2001; He et al., 2000; Gergel et al, 2002). Preliminary research suggests that landscape metrics quantifying the arrangement of human altered land in a catchment can improve water chemistry predictions offered by proportion based land use models. However, little is known about the underlying mechanisms of such relationships (Gergel et al., 2002), and further study will be necessary before refined spatial models can be developed to effectively predict water chemistry based on measures of land use arrangement (Jones et al., 2000). Most studies have focused on the arrangement of broad land use classes (Hunsaker and Levine, 1995; Richards et al, 1996; Jones et al, 2000, 2001; He et al, 2000; Gergel et al, 2002). The potential to apply such metrics to the analysis of impervious surfaces and other indicators of urbanization has not been sufficiently explored.

The objective of this study was to quantify the relationship between stream water quality and the type and pattern of watershed urbanization. Urbanization was defined using metrics for the type and distribution of impervious surfaces, household census data, city infrastructure, and natural features. Principal components analysis was used to determine the spatial distribution of urbanization metrics. In addition, multiple regression analysis was used to relate the urbanization metrics to different measures of water quality.

METHODS

Study Site

This study examined six urban and urbanizing watersheds in Durham, North Carolina. Durham is a medium-sized city of 223,314 inhabitants (TJCOG, 2000), which, along with the cities of Raleigh and Chapel Hill, makes up the rapidly growing Research Triangle area in the central North Carolina Piedmont. The Piedmont ecoregion of central North Carolina is characterized by gently sloping, well rounded hills with long valleys and ridges (Daniel and Dahlen, 2002). Due to the high clay and low permeability of soils in this region, smaller streams are susceptible to both high runoff during storm events and low flow during dry periods (NCDENR, 2001).

The ridgeline between the Neuse and Cape Fear River basins divides the City of Durham, with 40 percent of the city's area contained in the Neuse River basin and the remaining 60 percent contained in the Cape Fear River basin. Six streams were considered in this analysis: Ellerbe Creek, Little Lick Creek, and Stirrup Iron Creek from the Neuse River basin and New Hope Creek, Third Fork Creek, and Northeast Creek from the Cape Fear River basin (Figure 1). Of these six watersheds, Ellerbe Creek and Third Fork Creek contain the highest density of urban development (Table 1), as they each drain substantial portions of the downtown area. While Third Fork Creek watershed contains slightly more impervious surface per unit area than Ellerbe Creek watershed, the latter contains the highest population density and on average the oldest development. Little Lick Creek and Stirrup Iron Creek flow mostly outside of the city, and each watershed contains less than 10 percent impervious surface cover. New Hope Creek and Northeast Creek watersheds fall between these extremes, with 10 to 20 percent impervious surface area.

Ellerbe Creek, Third Fork Creek, New Hope Creek, and Little Lick Creek are all currently listed on North Carolina's 303(d) list of impaired water bodies not meeting water quality standards as defined by the Federal Clean Water Act. These streams do not fully support aquatic life, as indicated by poor benthic macroinvertebrate samples and fish community assessments. The North Carolina Department of Natural Resources has identified urban nonpoint source pollution as a probable source of water quality impairment (NCDENR, 2001).

Figure 1. Location of the Six Watersheds Draining Durham, North Carolina.

TABLE 1. Summary Characteristics for the Six Watersheds Draining the City of Durham, North Carolina.

Data Sources

The City of Durham Public Works Department provided planimetries files containing impervious surface data digitized from 1994 black and white digital orthophotography at 0.46 m ground resolution (City of Durham, 1994). Additionally, the city provided a 0.61 m vertical resolution point elevation coverage (Marcus Bryant, City of Durham, \GIS Department, July 30, 2001, unpublished data), stormwater inlet and outlet locations (John Cox, City of Durham, Stormwater Services, June 18, 2002, unpublished data), and sewer lines (City of Durham, 1994). Other geographic information system (GIS) datasets included 1:24,000 streams (USGS, 2001), 2000 digital census data (U.S. Census Bureau, 2001, 2002), and county level soils (USDA-NRCS, 1998).

City of Durham Stormwater Services supplied total phosphorus (TP), total kjeldahl nitrogen (TKN), total suspended solids (TSS), and fecal coliform data collected at 22 monitoring points throughout the city from February 2000 through June 2002 (John Cox, City of Durham, Stormwater Services, June 18, 2002, unpublished data). The data are from seasonal grab samples performed as part of Durham's stormwater National Pollutant Discharge Elimination System (NPDES) compliance program. During 2000, the sites were sampled monthly from January through November. In 2001, sampling was restricted to three times per monitoring station from February to June. In 2002, the city settled on the current seasonal sampling regime of once every three months. Due to changes in monitoring frequency, as well as the addition and removal of sampling stations during this period, the quantity of data is not consistent across sites. The number of samples at each site ranges from 4 to 12.

Grab samples were collected during low flow conditions, which is classified as receiving less than 0.1 inch (2.54 mm) of rainfall during the 24-hour period prior to sampling (NCDENR, 2003). Total P was measured according to the colorimetric, two-reagent method (USEPA, 1983). Total kjeldahl nitrogen was measured with the potentiometric method by ion selective electrode (USEPA, 1983). Total suspended solids were measured using the non-filterable residue, gravimetric method by drying oven at 103 to 105C (USEPA, 1983). Fecal coliforms were counted using the membrane filtration procedure (Clesceri et al., 1998). These four parameters were chosen from a full suite of water chemistry parameters measured for each sample and were selected because nutrients, sediments, and fecal coliforms are of particular concern in the study region.

Spatial Analysis

The point elevation data were used to construct a 1.524 m ground resolution digital elevation model (DEM) using inverse distance weighted interpolation. A flow accumulation grid derived from the digital elevation model was used to determine if the actual sampling locations could be used to delineate watersheds based on the DEM. Because the actual stream channel did not perfectly match the stream channel as predicted by the DEM, the monitoring point locations were shifted so that they fell inside of the theoretical stream channel (average shift of 5.52 m). The corrected monitoring points were used to delineate a series of 22 nested watersheds. While this nested design may impose a slight spatial dependence among the observations within a watershed, the nested structure was necessary to fully capture the data available for each watershed.

A series of 14 GIS derived variables was calculated for each of the nested watersheds (Figure 2). Indicators of urbanization included two variables relating to urbanization density (percent impervious surface area and household density), four variables related to type of urbanization (percent connected impervious surface area, mean impervious surface patch size, median impervious surface patch size, and median house age), and four variables related to access to city services (density of sewer system connections, septic tank density, stormwater outfall density, and percent of the watershed inside city limits). To account for natural differences between watersheds, an additional five variables were calculated related to natural watershed features (hydric soil density, mean saturated hydraulic conductivity, mean soil erosivity, and wetland density).

Impervious surface percentages were derived from the planimetries dataset, which was clipped to each individual watershed. Connected impervious surface area was determined by intersecting the planimetries data with stormwater inlet locations provided by Durham Stormwater Services (John Cox, City of Durham, Stormwater Services, June 18, 2002, unpublished data). Mean and median impervious surface patch sizes were calculated by grouping neighboring impervious surface cells using an eight-direction neighborhood analysis to form patches of contiguous impervious surface. To prevent the formation of one contiguous patch for each watershed, linear features such as roads and sidewalks were excluded from the analysis.

Weighted averages were estimated for all census data variables (population density, household density, median house age, septic tank density, sewer system density) by weighting the value of each census block partially or completely contained within the watershed by the percentage of the watershed area that it represents. When appropriate, this value was divided by watershed area to yield a density estimate. The same approach was used to calculate variables associated with soil properties for each watershed. All other spatial variables were calculated by clipping the dataset to the individual watersheds and dividing the quantity of each variable (km^sup 2^ of wetland, km of sewer lines) by watershed area to yield a density measurement.

Principal Components Analysis

Principal components analysis (PCA) was used to explore variation of the explanatory variables across space in the urban environment. Principal components analysis is a statistical data reduction technique that forms a new set of orthogonal variables that are linear composites of the original variables (Sabins, 1987; Sharma, 1996). The new variables, called principal components, are derived by rotating the axes in the multidimensional space created by the original set of variables. The first principal component is chosen such that it accounts for the maximum variance in the original dataset. The second principal component accounts for the maximum variation not accounted for by the first principal component and is orthogonal to the first principal component. Additional principal components can be derived following this same method until most of the variation in the original dataset has been captured (Sharma, 1996; Johnson and Wichern, 1998). The result is a smaller number of uncorrelated variables that captures the distribution of the initial dataset (Sabins, 1987). Loadings values are generated that represent correlations between the principal components and the original variables (Sharma, 1996).

Principal components analysis can be performed on raster data by treating each individual grid cell as a member of the population sample. This method is commonly used in remote sensing to determine the optimal combinations of light spectrum channels to separate land cover classes based on reflectance values (Sabins, 1987). The resulting principal components represent land features such as water, natural vegetation, and urban areas that cannot be mapped using reflectance values from only one portion of the light spectrum. Applying this same method to a range of variables representing urbanization density, extent, and natural features may reveal patterns in the urban landscape that are not visible using any one indicator variable alone.

Prior to performing the PCA, each explanatory variable was converted to a raster grid clipped to the entire study area. Neighborhood analysis was used to calculate local densities for density dependent variables (e.g., impervious surface cover, sewer line density). A neighborhood sum function was run using a 4.57 m^sup 2^ neighborhood window to yield a grid representing neighborhood sums across the study area. This grid was divided by the neighborhood window size to generate a grid representing neighborhood densities across the study area. Principal components analysis was performed using the PCA function in the ERDAS Imagine 8.6 software package with grids representing each of the 14 explanatory variables as input data. Eigenvalues and loadings were calculated for each of the principal components, and output grids were generated to map the distribution of the first three principal components across the study area.

Figure 2. Examples of Urbanization Indicators Used in the Subwatershed Analysis: (a) Household Density; (b) House Age; (c) Impervious Surface; (d) Connected Impervious Surface; (e) Impervious Surface Patches; and (f) Stormwater Outfall Density.

Statistical Analysis

Multiple linear regression models were fit to relate 16 candidate predictor variables to TP, total nitrogen (TN), TSS, and fecal coliforms from the 22 ambient monitoring locations. The candidate predictor variables included the 14 spatially derived variables plus variables for recent rainfall and watershed area (Table 2). Recent rainfall data were calculated as the amount of rainfall, in inches, during the 72 hours prior to each sampling event. Mean recent rainfall was used as an explanatory variable to account for slight variations in dry-weather sampling condition among sites. Watershed area was included to account for the influence of drainage area on stream water quality.

TABLE 2. Legend for Explanatory Variables Considered in the Multiple Regression Analysis, Arranged by Thematic Group.

To remain consistent with the analytical techniques used by the City of Durham Stormwater Services (CDSS, 1999), the three-year geometric mean TP, TN, TSS, and fecal coliform count was calculated at each site for all samples collected from 2000 to 2002. While averaging across the study time period ignored seasonal effects at each individual sampling location, all watersheds were treated equally with regard to data averaging. This provides a sufficient basis for comparing water quality among watersheds during the study period.

Initially, full models were considered \that related all 16 candidate explanatory variables to the water quality response variables. Mallows' Cp (Mallows, 1973, 1995), Bayesian Information Criterion (BIC; Schwartz, 1978), and adjusted r^sup 2^ (Ezekiel, 1930) were calculated for all subset models. Subset models selected for further investigation were those with the best Mallows' Cp (closest to the number of model parameters), lowest BIC, highest adjusted r^sup 2^, and in which all of the explanatory variables were statistically significant. The final models were selected by examining residual plots to determine which models best met the assumptions of the standard linear regression model.

In theory, Mallows' Cp balances bias due to excluding important explanatory variables and the extra variance created by including too many (Ramsey and Schafer, 1997). Models with redundant variables should not receive high Cp ratings because the additional variables create variance in the model without explaining additional variation in the response variable. Despite receiving high Cp values, the final models contained several variables in the same thematic group, indicating a potential for collinearity. Collinearity between two or more variables can cause instability in the estimates of the coefficients for the explanatory variables (Montgomery and Peck, 1982). This complicates the interpretation of the relationship between the collinear variables and the response variable.

TABLE 3. Correlation Matrix for Explanatory Variables [high correlations are shown in bold (> 0.70)].

The final models were screened for collinearity by examining a correlation matrix of the explanatory variables (Table 3). Problems associated with collinearity increase as the correlation coefficient nears one, but there is no widely accepted cutoff for detecting collinearity based on pairwise correlation (Belsley, 1991). Conventionally, correlation coefficients ranging from 0.7 to 0.9 have been used to screen variables for potential collinearity (Gunst and Mason, 1980). For the purposes of this study, variables with pairwise correlations greater than 0.7 were chosen for further investigation. These variables were sequentially dropped from each of the water quality models to test their effect on the coefficient estimates for the other variables. Generally, if the removal of one variable causes a sign change or large shifts in the coefficients of one or more of the other variables, it is an indication that the model is affected by collinearity (Gunst and Mason, 1980).

The correlation analysis revealed strong bivariate correlations between household density, median house age, sewer line density, percent impervious surface area, and percent connected impervious surface area. Strong correlations also existed between hydric soil density and both watershed area and wetland density. Two or more correlated variables were contained in the fitted models for TKN (household density and median house age), TSS (household density and percent connected impervious surface area; hydric soil density, wetland density, and watershed area), and fecal coliforms (household density, house age, and percent impervious surface area; hydric soil density and watershed area). In many cases, sequentially dropping these potentially collinear variables from the models resulted in changes in the sign or value of the coefficients estimates for the other variables. This suggests that these three models are influenced by collinearity. Reliable estimates cannot be generated for the coefficients of the correlated variables. Collinearity does not, however, influence the predictive power of linear models (Montgomery and Peck, 1982). Nor does it indicate that the correlated variables contribute no useful information to the predictive model (Bowerman and O'Connell, 1990). Since the correlated variables represent aspects of watershed urbanization that may influence water quality, removing them from the models would exclude potentially important predictive information. To avoid this, the correlated variables were kept in the model with the understanding that caution should be taken in the interpretation of coefficient estimates for those variables.

RESULTS

Principal Components Analysis

The results of the principal components analysis show that most of the spatial variation in the explanatory variables can be reduced to the first three principal components (Figure 3). Eighty-five percent of the variation in the explanatory variables is resolved by the first principal component, which has nearly equal positive loadings from indicators of urbanization density (household density), indicators of urbanization type (mean and median impervious surface patch size), access to city services (sewer density, septic system density, and stormwater outfall density), and soil properties (overall soil erosivity, top layer erosivity, and mean saturated hydraulic conductivity). An additional 7 percent of the variation is resolved by the second principal component, which is dominated by equal negative loadings from two indicators of urbanization type - mean and median impervious surface patch size. The third principal component contributes an additional 7 percent of the variation in the explanatory variables and consists of negative loadings from indicators of urbanization density (household density), access to city services (sewer density, septic system density), and soil properties (overall soil erosivity, top layer soil erosivity, and saturated hydraulic conductivity) and positive loadings associated with indicators of urbanization type (mean and median impervious surface patch size). Together, the first three principal components account for 99 percent of the overall spatial variation in the explanatory variables (Table 4).

Total Kjeldahl Nitrogen Model

The TKN multiple regression model included one indicator of urbanization density (household density), one indicator of urbanization type (median house age), one indicator of access to city services (percent of the watershed within city limits), and a variable for recent rainfall (Figure 4a). Of these variables, household density and recent rainfall were associated with increased TKN concentrations in the presence of the other variables, while house age and the percent watershed within city limits were associated with decreased TKN (Table 5). The model predicts that with all other variables held constant, densely developed watersheds that have received more recent rainfall will have higher TKN concentrations, while watersheds with older development and that are largely contained within city limits will have lower TKN. However, the correlation between household density and median house age makes the coefficients for these two variables unstable (Table 3). It is not possible to determine the exact relationship between each of these variables and TKN concentrations when the other variables are held constant.

Total Phosphorus Model

The total phosphorus model revealed significant correlations to one indicator of urbanization density (household density), one indicator of urbanization type (median impervious surface patch size), one indicator of access to city services (percent of the watershed inside city limits), and a variable for watershed area (Figure 4b). In the presence of the other variables, household density and median impervious surface patch size were associated with an increase in TP, while watershed area and the amount of the watershed inside city limits were associated with a decrease in TP (Table 5). With all other variables held constant, densely developed watersheds with more large impervious surfaces tended to have the highest TP levels, while larger watersheds mostly inside city limits tended to have lower TP levels. None of the explanatory variables in this model had bivariate correlations greater than 0.7.

Figure 3. Mapped Principal Components: (a) First Principal Component; (b) second Principal Component; and (c) Third Principal Component.

Total Suspended Solids Model

The TSS model (Table 5, Figure 4c) showed significant correlations with one indicator of urbanization density (household density), one indicator of urbanization type (percent connected impervious surface area), one indicator of access to city services (stormwater outfall density), and three variables related to natural watershed features (watershed area, hydric soil density, and wetland density). In the presence of the other variables, an increase in household density and hydric soil density was associated with an increase in suspended solids. Concurrently, increases in watershed area, wetland density, stormwater outfall density, and percent connected impervious surface area were associated with decreases in suspended solids. With all other variables held constant, densely developed watersheds and watersheds with more hydric soils had higher suspended solids, while large watersheds, watersheds with more wetlands, greater stormwater outfall densities, and more connected impervious surface area had lower suspended solids. Strong correlations were found among watershed area, hydric soil density, and wetland density, and between household density and percent connected impervious surface area (Table 3). The coefficients for these variables are unstable, making their individual relationships with TSS difficult to interpret.

TABLE 4. Original Variable Loadings and the Cumulative Percent of Variation Explained by the First Three Principal Components.

Fecal Coliform Model

The fecal coliform model (Table 5, Figure 4d) included two indicators of urbanization density (household density and percent impervious surface area), two indicators of urbanization type (median house age and median impervious surface patch size), and three variables related to natural watershed features (watershed area, hydric soil density, and recent rainfall). In the presence of the other variables, increases in watershed are\a, housing density, impervious surface area, and rainfall were associated with an increase in fecal coliform count. Increases in median house age, hydric soil density, and median impervious surface patch size were associated with a decrease in fecal coliform count. With all other variables held constant, larger watersheds with denser housing, more impervious surface area, and that had received more recent rainfall tended to have higher fecal coliform counts. Watersheds with older development, more hydric soils, and larger impervious surfaces tended to have lower fecal coliform counts. The correlation matrix for the explanatory variables revealed strong bivariate correlations between watershed area and hydric soil density and among household density, median house age, and percent impervious surface area (Table 3). Due to collinearity, the coefficient estimates for these variables are unstable and may not accurately reflect their relationship to fecal coliform levels.

DISCUSSION

Many different aspects of urban environments interact to influence the water quality of urban streams in ways that are not easily captured by a single land use indicator. Various pollution sources exist in the urban environment. Not all of these sources are directly related to urbanization density, the most common indicator of urbanization used in watershed modeling. The inclusion of variables describing urbanization type and access to city services such as sewer and stormwater systems helps to account for a wide range of sources. Taken together, these variables provide a more comprehensive picture of a complex urban environment with respect to its impact on water quality than any single variable taken alone.

The complexity of urban areas is seen in the results of the PCA, in which various indicators of urbanization density and type are needed to separate spatial trends related to urban sprawl from trends related to urbanization in general. Based on equal, positive loadings from most urbanization indicators, the first principal component roughly corresponds to degree of urbanization within the study area. The second principal component represents rural areas characterized by only a few small, isolated impervious surfaces. The third principal component has negative loadings for variables related to urbanization density and positive loadings for variables related to impervious surface patch size. This is characteristic of the sprawling, "box-style" development that has grown up around the core urban area in Durham. While residential development is more spread out in these areas, commercial development is dominated by large shopping centers and their associated parking lots.

By identifying spatial trends in the urban landscape, the first three principal components provide valuable information that is useful in interpreting the results of the four multiple regression models. The second and third principal components, while representing only a small portion of the variation in the original dataset, provide information that is not readily captured by traditional urbanization indicators. Overall urbanization is relatively easy to quantify, but distinguishing between types of urbanization and their relative impacts on natural systems is not. The Triangle region of North Carolina was recently rated the third most "sprawling" community in the United States (Ewing et al., 2002). The relative impact of this style of urban development on water resources has not been reliably measured. The urbanization indicators examined in this study, when used in combination, can effectively distinguish suburban sprawl from general landscape urbanization. Identifying and quantifying rural development patterns is also helpful, because it provides a baseline for monitoring landscape change over time.

Figure 4. Response Versus Fit Plots for the Multiple Regression Models: (a) TKN Model; (b) TP Model; (c) TSS Model; and (d) Fecal Coliform Model. Solid line represents the correlation between predicted and actual values for the response variables. Dotted line represents perfect correlation between the predicted and actual values.

Within the four statistical models, indicators of urbanization density (household density, impervious surfaces) had significant, positive correlations with nonpoint source pollution. While the individual coefficients cannot be trusted due to collinearity in three of the models, the selection of models with these variables supports the results of previous studies that have linked urbanization with water quality impairment (Lenat and Crawford, 1994; Schueler, 1994; Fitzpatrick, 1995; Mumley, 1995; May et al., 1997; Duda et al, 1998; Bhaduri et al, 2000). In most models, household density was the preferred indicator of urbanization density. This highlights the importance that residential development, in particular, has on the generation of many urban pollutants (Loehr, 1974; Bannerman et al., 1993). It also demonstrates that indicators of urbanization density derived from census data can be effective replacements for impervious surface data, particularly in models where impervious surfaces are also being used to calculate indicators of urbanization type.

TABLE 5. Multiple Regression Models for Total Kjeldahl Nitrogen (TKN), Total Phosphorus (TP), Total Suspended Solids (TSS), and Fecal Coliforms.

The four regression models suggest that while the quantity of watershed urbanization impacts water quality, the type of urban development within a watershed is also important. Three indicators of urbanization type (median house age, percent connected impervious surface area, and median impervious surface patch size) were included in one or more of the water quality models. Median house age was included in the TKN model and the fecal coliform model. In both cases, younger residential development was associated with increased pollutant levels, but the extent of this association cannot be quantified due to collinearity with indicators of household density in these models. Only the TSS model contained a variable for connected impervious surface area. In this model, connected impervious surface was associated with decreased suspended solid loads. In Durham, most connected impervious surface area is located in older, built-out sections of the city where ongoing development is less common. Strict stormwater regulations for new development limits connected impervious area in new neighborhoods, but the high concentration of new construction in these areas can result in overall higher TSS loads.

Median impervious surface patch size was included in the models for total phosphorus and fecal coliforms. In the TP model, areas with more large impervious surfaces tended to have higher TP loads. Since the PCA results show that impervious surface patch size is correlated with areas of urban sprawl in Durham, this indicates that sprawling development is associated with elevated TP loads. The opposite correlation was observed in the fecal coliform model. Areas dominated by smaller patches - rural areas and older neighborhoods, as indicated by the results of the PCA - contribute to high fecal coliform counts. This is consistent with the results of Durham's fecal coliform source identification and elimination program, which has located leaking sewer lines and illicit sewage connections to the stormwater system in older neighborhoods and ineffective septic tanks in rural communities (CDSS, 1999).

Of the variables related to access to city services, only the variables for city limits and stormwater outfall density were included in the regression models. The negative correlation between city limits and both nutrient variables is indicative of the role city services and regulations play in determining water quality in urban streams. The models predict that areas in the study watersheds lying beyond city limits tend to contribute higher nutrient and fecal bacteria loads than their urban counterparts, when developed in the same manner and density as areas within city limits. This points to the relative effectiveness of municipal wastewater treatment as compared to on-site wastewater treatment. Stormwater outfall density was important only in the TSS model, where higher stormwater outfall densities were associated with lower TSS loads.

Watershed area, hydric soil density, wetland density, and recent rainfall were the most important natural features in the four water quality models. Watershed area was included in all but one model. The relationship between watershed size and nonpoint source pollutant loads has been well documented (Novotny and Olem, 1994). Interestingly, this association is reversed in the two nutrient models, with larger watersheds corresponding to lower nutrient loads. Due to the nested watershed design, the largest watersheds were those with sampling points farthest downstream. These watersheds provide more opportunity for instream processing than smaller watersheds draining the center of the city. Recent rainfall was included in the TKN and fecal coliform models. Samples taken after even light rainfall had higher TKN and fecal coliform levels than samples collected during the driest periods. This reflects the rapid instream processing for these pollutants. Finally, hydric soil distribution and wetland density were both important variables in the TSS model. Despite a strong correlation between hydric soils and wetland density, these two variables have opposite correlations with measured TSS loads. Given the instability of the coefficients for these two variables, it is not feasible to clearly determine the individual effect each variable has on TSS loads when the other variable is held constant.

Some of the collinearity in the water quality models results from common zoning ordinances across most of Durham's watersheds. This would be alleviated by developing a larger dataset involving watersheds from a variety of cities withdifferent zoning ordinances. A significant challenge with developing a dataset involving more than one municipality is the availability of consistent GIS layers and sufficient water quality data. This problem may be alleviated as GIS technologies continue to become more accessible to local governments and as more cities are required to develop stormwater NPDES monitoring programs. However, some degree of correlation will always be present among indicators of the type and density of urban development. Principal components regression should also be explored as a method for dealing with inherent collinearity in urban watershed analysis. This method produces a new set of orthogonal variables that can be used in regression analysis and may result in more robust water quality models (Montgomery and Peck, 1982). Since the effects of the original variables are difficult to interpret using principal components regression, further studies will also be needed to clarify the roles of individual variables.

Both older, densely developed neighborhoods and younger, low density suburban sprawl contribute to the water quality impairment observed in the streams in this study - albeit in different ways. This study shows that while overall urbanization is important, management of urban nonpoint source pollution should also focus on low-density newer urban development and watersheds with a dominance of large impervious surfaces. While collinearity complicated the understanding of individual coefficients for many of the variables used in this study, it is clear that measures such as impervious surface patch size, connected impervious surfaces, extent and distribution of sewer and stormwater systems, and census data are useful tools for predicting water quality in urban and urbanizing watersheds. Urban watersheds are complex systems, and there is inherent difficulty in modeling how streams respond to different patterns of urbanization. Gaining a better understanding of how different aspects of urbanization interact will allow for the development of better predictive tools for modeling urban nonpoint source pollution. That knowledge can then be incorporated into the decision making processes that are critical to addressing urban runoff issues.

CONCLUSIONS

With principal components analysis, it is possible to separate urban sprawl from spatial trends related to general urbanization and rural development using variables related to density and type of urbanization as well as the extent of city services. In multiple linear regression models, development density was shown to be correlated to increased nitrogen, phosphorus, suspended solids, and fecal coliform bacteria. Indicators of urbanization type such as the house age, amount of contiguous impervious surface, and stormwater connectivity explained additional variation for some of these pollutants. In nutrient models, access to city services was also an important factor. The results indicate that while urbanization density remains important in predicting water quality, urban watershed models can be improved by including indicators of urbanization type and access to city services.

ACKNOWLEDGMENTS

This research was supported by a conservation fellowship from the Doris Duke Charitable Foundation to Melissa Vernon Carle and by a National Science Foundation Urban Research Initiative Program grant to Patrick N. Halpin. The authors would like to thank the City of Durham Stormwater Services, Planning, and GIS Departments for providing the data used in this project. Special thanks go to John Cox, Chris Outlaw, Robert Louque, Jane Korest, and George Norris for their support during various stages of this research. The authors are grateful to the Duke University Landscape Ecology lab group for their thoughtful critique of an early draft and to the editor and three reviewers whose comments substantially improved the final draft of this paper.

Carle, Melissa Vernon, Patrick N. Halpin, and Craig A. Stow, 2005. Patterns of Watershed Urbanization and Impacts on Water Quality. Journal of the American Water Resources Association (JAWRA) 41 (3):693-708.

1 Paper No. 04044 of the Journal of the American Water Resources Association (JAWRA) (Copyright 2005). Discussions are open until December 1, 2005.

LITERATURE CITED

Arnold, C.L. and C.J. Gibbons, 1996. Impervious Surface Coverage: The Emergence of a Key Environmental Indicator. Journal of the American Planning Association 62(2):243-258.

Bannerman, R.T, D.W. Owens, R.B. Dodds, and N.J. Hornewer, 1993. Sources of Pollutants in Wisconsin Stormwater. Water Science and Technology 28(3-5):241-259.

Basnyat, P., L.D. Teeter, E.G. Lockaby, and K.M Flynn, 2000. The Use of Remote Sensing and GIS in Watershed Level Analysis of Nonpoint Source Pollution Problems. Forest Ecology and Management 128:65-73.

Belsley, D.A, 1991. Conditioning Diagnostics: Collinearity and Weak Data in Regression. Wiley, New York, New York, 396 pp.

Bhaduri, B., J. Harbor, B. Engel, and M. Grove, 2000. Assessing Watershed-Scale, Long-Term Hydrologie Impacts of Land-Use Change Using a GIS-NPS Model. Environmental Management 28(6):643-658.

Booth, D.B. and C.R. Jackson, 1997. Urbanization of Aquatic Systems: Degradation Thresholds, Stormwater Detection, and Limits of Mitigation. Journal of the American Water Resources Association (JAWRA) 33(5): 1077-1090.

Bowerman, B.L. and R.T. O'Connell, 1990. Linear Statistical Models: An Applied Approach (second Edition). PWS-Kent, Boston, Massachusetts, 1,024 pp.

Brun, S.E. and L.E. Band, 2000. Simulating Runoff Behavior in an Urbanizing Watershed. Computers, Environment, and Urban Systems 24(l):5-22.

Carpenter, S.R., N.F. Caraco, D.L. Correll, R.W. Howarth, A.N. Sharpley, and V.H. Smith, 1998. Nonpoint Pollution of Surface Waters With Phosphorus and Nitrogen. Ecological Applications 8(3):559-568.

City of Durham, 1994. Utility Structures: Water, Sewer, Storm Water (digital dataset). Public Works Department, Durham, North Carolina.

City of Durham, 2000. City Limits (digital dataset). Public Works Department, Durham, North Carolina.

CDSS (City of Durham Stormwater Services), 1999. NPDES MS4 Permit Renewal Application. Durham, North Carolina.

COPWD (City of Olympia Public Works Department), 1995. Impervious Surface Reduction Study. Olympia, Washington.

Clesceri, L.S., A.E. Greenberg, and A.D. Eaton (Editors), 1998. Standard Methods (20th Edition). American Public Health Association, American Water Works Association, and Water Environment Federation, Washington, D.C, 1,220 pp.

Cole, R.H., R.E. Frederick, R.P. Healy, and R.G. Rolan, 1984. Preliminary Findings of the Priority Pollutant Monitoring Project of the Nationwide Urban Runoff Program. Journal of the Water Pollution Control Federation 56(7):898-908.

Daniel III, C.C. and P.R. Dahlen, 2002. Preliminary Hydrogeological Assessment and Study Plan for a Regional Ground- Water Resource Investigation of the Blue Ridge and Piedmont Provinces of North Carolina. U.S. Geological Survey Water Resources Investigations Report 02-4105, Raleigh, North Carolina.

Duda, A.M., D.R. Lenat, and D.L. Penrose, 1998. Water Quality in Urban Streams - What We Can Expect. Journal of the Water Pollution Control Federation 54(7):1139-1147.

Ewing, R., R. Pendall, and D. Chen, 2002. Measuring Sprawl and Its Impact. Smart Growth America, Washington, D.C.

Ezekiel, M., 1930. Methods of Correlation Analysis. Wiley, New York, New York, 406 pp.

Fitzpatrick, W.P., 1995. Wisconsin Smart Program: Starkweather Creek. In: United States Center for Environmental Research Information, National Conference on Urban Runoff Management: Enhancing Urban Watershed Management at the Local, County, and State Levels. U.S. Environmental Protection Agencey, EPA/625/R-95/003, Cincinnati, Ohio, pp. 173-179.

Gergel, S.E., M.G. Turner, J.R. Miller, J.M. Melack, and E.H. Stanley, 2002. Landscape Indicators of Human Impacts to Riverine Systems. Aquatic Sciences 64:118-128.

Gunst, R.F. and R.L. Mason, 1980. Regression Analysis and Its Application. Marcel Dekker, New York, New York, 402 pp.

He, C., S.B. Malcolm, K.A. Dahlberg, and B.Fu, 2000. A Conceptual Framework for Integrating Hydrological and Biological Indicators Into Watershed Management. Landscape and Urban Planning 49:25-34.

Hunsaker, C.T. and D.A. Levine, 1995. Hierarchical Approaches to the Study of Water Quality in Rivers. BioScience 45(3):193-211.

Jennings, D.B. and S.T. Jarnagin, 2002. Changes in Anthropogenic Impervious Surfaces, Precipitation and Daily Streamflow Discharge: A Historical Perspective in a Mid-Atlantic Subwatershed. Landscape Ecology 17:471-489.

Johnson, R.A. and D.W. Wichern, 1998. Applied Multivariate Statistical Analysis. Prentice-Hall, Upper Saddle River, New Jersey, 816 pp.

Jones, K.B., D.T. Heggem, T.G. Wade, A.C. Neale, D.W. Ebert, M.S. Nash, M.H. Mehaffey, K.A. Hermann, A.R. Selle, S. Augustine, I.A. Goodman, J. Pedersen, D. Bolgrien, J.M. Viger, D. Chiang, C.J. Lin, Y. Zhong, J. Baker, and R.D. Van Remortel, 2000. Assessing Landscape Condition Relative to Water Resources in the Western United States: A Strategic Approach. Environmental Monitoring and Assessment 64:227- 245.

Jones, K.B., A.C. Neale, M.S. Nash, R.D. Van Remortel, J.D. Wickham, KH. Riiters, and R.V. O'Neill, 2001. Predicting Nutrient and Sediment Loadings to Streams From Landscape Metrics: A Multiple Watershed Study From the United States Mid-Atlantic Region. Landscape Ecology 16:301-312.

Lenat, D.R. and J.K. Crawford, 1994. Effects of Land Use on Water Quality and Aquatic Biota of Three North Carolina Piedmont Streams. Hydrobiologia 294:185-199.

Loehr, R.C., 1974. Characteristics and Comparative Magnitude of Nonpoint Sources. Journal of the Water Pollution Control Federation 46(8):1849-1872.

Mallows, C.L., 1973. Some Comments on Cp. Technometrics 15:661676.

Mallows, C.L., 1995. More Comments on Cp. Technometrics 37:362372.

May, C.W., R.R. Horner, J.R. Karr, B.W. Mar, and E.B. Wel\ch, 1997. Effects of Urbanization on Small Streams in the Puget Sound Lowland Ecoregion. Watershed Protection Techniques 2(4):483494.

McMahon, G. and D.A. Harned, 1998. Effect of Environmental Setting on Sediment, Nitrogen, and Phosphorus Concentrations in Albemarle-Pamlico Drainage Basin, NC and VA, USA. Environmental Management 22(6):887-903.

Montgomery, D.C. and E.A. Peck, 1982. Introduction to Linear Regression Analysis. Wiley, New York, New York, 504 pp.

Mumley, T.E., 1995. Urban Runoff Pollution Prevention and Control Planning: San Fransisco Bay Experiences. In: United States Center for Environmental Research Information, National Conference on Urban Runoff Management: Enhancing Urban Watershed Management at the Local, County, and State Levels. U.S. Environmental Protection Agency, EPA/625/R-95/003, Cincinnati, Ohio, pp. 106-108.

NCDENR (North Carolina Department of Environment and Natural Resources), 2000. Cape Fear River Basinwide Water Quality Plan. Division of Water Quality, Raleigh, North Carolina.

NCDENR (North Carolina Department of Environment and Natural Resources), 2001. Basinwide Assessment Report - Neuse River Basin. Division of Water Quality, Raleigh, North Carolina.

NCDENR (North Carolina Department of Environment and Natural Resources), 2002. Neuse River Basinwide Water Quality Plan. Division of Water Quality, Raleigh, North Carolina.

NCDENR (North Carolina Department of Environment and Natural Resources), 2003. Ellerbe Creek Local Watershed Plan: Draft. Wetlands Restoration Program, Raleigh, North Carolina.

Novotny, V., 1991. Urban Diffuse Pollution: Sources and Abatement. Water Environment and Technology 3(12):60-65.

Novotny, V. and H. Olem, 1994. Water Quality: Prevention, Identification, and Management of Diffuse Pollution. Van Nostrand Reinhold, New York, New York, 1,054 pp.

Paul, M.J. and J.L Meyer, 2001. Streams in the Urban Landscape. Annual Review of Ecological Systems 32:333-365.

Ramsey, FL. and D.W. Schafer, 1997. The Statistical Sleuth: A Course in Methods of Data Analysis. Duxbury Press, Belmont, California, 742 pp.

Richards, C., L.B. Johnson, and G.E. Host, 1996. Landscape-Scale Influences on Stream Habitats and Biota. Canadian Journal of Fisheries and Aquatic Science 53(Suppl. 1):295-311.

Sabins, F.F., 1987. Remote Sensing: Principles and Interpretation (second Edition). WH. Freeman, New York, New York, 449 pp.

Schueler, TR, 1994. The Importance of Imperviousness. Watershed Protection Techniques 1(3):100-111.

Schwartz, G., 1978. Estimating the Dimension of a Model. Ann. Statist. 6:461-464.

Sharma, S., 1996. Applied Multivariate Techniques. Wiley, New York, New York, 493 pp.

Soranno, P.A., S.L. Hubler, S.R. Carpenter, and R.C. Lathrop, 1996. Phosphorus Loads to Surface Waters: A Simple Model to Account for Spatial Pattern of Land Use. Ecological Applications 6(3):865- 878.

Strayer, D.L., R.E. Beighley, L.C. Thompson, S. Brooks, C. Nilsson, G. Pinay, and R.J. Naiman, 2003. Effects of Land Cover on Stream Ecosystems: Roles of Empirical Models and Scaling Issues. Ecosystems 6(5):407-423.

TJCOG (Triangle J Council of Governments), 2000. Riparian Buffer Preservation and Streambank Restoration in the Upper Neuse Basin: A Blueprint for the Future. Research Triangle Park, North Carolina.

U.S. Census Bureau, 2001. TIGER/Line Files - Census 2000 (digital dataset). Available at http://www2.census.gov/geo/tiger/ tiger2k/nc/ tgr37063.zip. Accessed in June 2002.

U.S. Census Bureau, 2002. Census 2000 Summary File 3 - North Carolina. Available at http://www2.census.gov/census_2000/ datasets/ Summary_File_3/North Carolina/all_North_Carolina. zip. Accessed in June 2002.

USDA-NRCS (U.S. Department of Agriculture, Natural Resources Conservation Service), 1998. Detailed County Soils - Durham County, North Carolina (digital dataset). USDA-NRCS, Raleigh, North Carolina.

USEPA (U.S. Environmental Protection Agency), 1983. Methods for Chemical Analysis of Water and Wastes. Office of Research and Development, EPA/841/S-83/101, Washington, D.C.

USEPA (U.S. Environmental Protection Agency), 1996. Nonpoint Source Pollution: The Nation's Largest Pollution Problem. Nonpoint Source Control Branch, EPA/841/F-96/004a, Washington, D.C.

USEPA (U.S. Environmental Protection Agency), 2000. National Water Quality Inventory: 1998 Report to Congress, Office of Water, EPA/841/R-02/001, Washington, D.C.

USFWS (U.S. Fish and Wildlife Service), 1992. National Wetlands Inventory: Northwest Durham. Available at http://wetlands.fws. gov/ arcdata/greensboro/nwdurh.eOO. Accessed in June 2002.

USFWS (U.S. Fish and Wildlife Service), 1994. National Wetlands Inventory: Northeast Durham. Available at http://wetlands.fws. gov/ arcdata/greensboro/nedurh.eOO. Accessed in June 2002.

USFWS (U.S. Fish and Wildlife Service), 1995a. National Wetlands Inventory: Southwest Durham. Available at http://wetlands.fws. gov/ arcdata/raleigh/swdurh.eO. Accessed in June 2002.

USFWS (U.S. Fish and Wildlife Service), 1995b. National Wetlands Inventory: Southeast Durham. Available at http://wetlands.fws. gov/ arcdata/raleigh/sedurh.eOO. Accessed in June 2002.

USGS (U.S. Geologic Survey), 2001. National Hydrography Dataset (digital dataset). Available at http://nhd.usgs.gov/data.html. Accessed in June 2002.

Wang, L., J. Lyons, and P. Kanehl, 2001. Impacts of Urbanization on Stream Habitat and Fish Across Multiple Spatial Scales. Environmental Management 28(2):255-266.

Melissa Vernon Carle, Patrick N. Halpin, and Craig A. Stow2

2 Respectively, Wetlands Specialist, North Carolina Division of Coastal Management, Mail Service Center 1638, Raleigh, North Carolina 27669 (formerly Research Associate, Duke University); Assistant Professor of the Practice of Landscape Ecology, Nicholas School of the Environment and Earth Sciences, Duke University, Box 90328, Durham, North Carolina 27708; and Associate Professor, Department of Environmental Health Sciences, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina 29208 (Email/Carle: melissa.carle@ncmail.net).

Copyright American Water Resources Association Jun 2005

Source: Journal of the American Water Resources Association

More News in this Category