Journal of Hazardous Materials 209–210 (2012) 48–58

Contents lists available at SciVerse ScienceDirect

Journal of Hazardous Materials

journal homepage: www.elsevier.com/locate/jhazmat

Effects of land cover, topography, and built structure on seasonal water

quality at multiple spatial scales
Bethany Pratt a , Heejun Chang a,b,∗
a
Department of Geography, Portland State University, Portland, OR 97201, USA
b
Institute for Sustainable Solutions, Portland State University, Portland, OR 97201, USA

a r t i c l e i n f o a b s t r a c t

Article history: The relationship among land cover, topography, built structure and stream water quality in the Portland
Received 14 July 2011 Metro region of Oregon and Clark County, Washington areas, USA, is analyzed using ordinary least squares
Received in revised form (OLS) and geographically weighted (GWR) multiple regression models. Two scales of analysis, a sectional
22 December 2011
watershed and a buffer, offered a local and a global investigation of the sources of stream pollutants.
Accepted 25 December 2011
Model accuracy, measured by R2 values, fluctuated according to the scale, season, and regression method
Available online 10 January 2012
used. While most wet season water quality parameters are associated with urban land covers, most
dry season water quality parameters are related topographic features such as elevation and slope. GWR
Keywords:
Geographically weighted regression
models, which take into consideration local relations of spatial autocorrelation, had stronger results than
Water quality OLS regression models. In the multiple regression models, sectioned watershed results were consistently
Development better than the sectioned buffer results, except for dry season pH and stream temperature parameters.
This suggests that while riparian land cover does have an effect on water quality, a wider contributing
area needs to be included in order to account for distant sources of pollutants.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction In humid temperate climates, where there are substantial sea-

sonal variations in precipitation and temperature, constituent
The effect of land development on natural systems is frequently concentrations vary due to flow regimes. These seasonal regimes
quantified by examining the relationship between land cover and need to be considered when studying particulate and other pol-
streams [1–4]. Land development, in the form of urban and agricul- lutants’ concentrations in order to account for dilution and runoff
tural land usage, increases impervious surfaces, has been found to [7,15,18]. Water quality measures, such as phosphorus and stream
have a negative correlation with stream health, typically increas- water temperature, peak during the low flow, warm season,
ing flash runoff and nutrient and heavy metal loads [5–9]. This unlike many other parameters that fluctuate based on runoff.
association implies that without abatement efforts, increases in Studies have found that the influencing land cover for specific
development can lead to decreases in water quality, which affects water quality measurements changes according to the season
safe drinking water availability, recreational opportunities, flood- [15,19].
plains, and habitat [10–13]. Selecting sample sites near anthropogenic or natural sources
With this connection between land development and poor of elements such as nitrogen or phosphorus can yield valuable
water quality, land managers are being encouraged to use geo- results [19]. Wang et al. [19] found that industrial sites along the
graphic information systems (GISs) to identify problem areas and Grand Canal in China consistently displayed higher gasoline and
develop projects to improve stream health [10,11,14–16]. Spatial metal levels than samples taken elsewhere. A similar finding was
analysis techniques allow users to view and analyze geographic reported in the study of the lower Han River basin in South Korea,
data, which includes, water quality, climate, topographic, and land- where sites downstream of industrial and urban areas were found
scape variables, quickly and efficiently. Restoration projects or to have worse water quality than those upstream [20]. Located at
storm water retention ponds can be planned by using GIS to identify the confluence of the Willamette and Columbia rivers, Portland
stream reaches and areas that would most benefit [17]. – Vancouver’s urban and industrial development has negatively
impacted the first and second-order urban streams. Lower levels
of DO and higher levels of nutrients and water temperatures are
common in these low elevation streams [21,22].
The scale of analysis is important because it determines what
∗ Corresponding author. Tel.: +1 503 725 3162; fax: +1 503 725 3166. area researchers use to link land cover with a stream site’s chem-
E-mail address: [email protected] (H. Chang). ical and physical properties. By using the watershed scale, an area

0304-3894/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhazmat.2011.12.068
 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58 49

not located along or even near the stream might be attributed to

being the source of a pollutant. Including an entire stream reach
might also be unreasonable, as the stream dilutes pollutants before
they are sampled and may be caught up and absorbed by plants or
soils along the stream [e.g. 23, 24]. Cunningham et al. [2] suggests
that watershed managers focus on simple projects in riparian areas
to improve water quality in the general area being targeted and
downstream. Removing a parking lot next to a stream, for exam-
ple, would reduce flashy flooding and allow runoff to seep into the
ground before joining the stream.
Multivariate regression model allows researchers to assume a
diverse array of landscape parameters in order to derive the causes
of pollutants. Broad categories, such as urban or residential land
covers, assume a heterogenic landscape [25,26]. These categories
ignore, or do not fully incorporate changes in density or simply
differences in anthropogenic development and physical geogra-
phy across a space [17]. By using a spatially explicit multivariate
approach, non-point sources such as agriculture as well as variables
such as street density may be incorporated into a finer resolution
analysis [27].
The objectives of this paper are to examine the relationship
between landscape variables and water quality in the Portland,
Oregon and Clark County, Washington area by answering three
questions. First, does the season matter in determining what land
cover has the most influence on water quality? Second, does the
scale of analysis have an effect on the results, and if so, what are
they? And third, does GWR, which incorporates spatial autocorre-
lation, offer betters predictive power than a global OLS regression
model?
Fig. 1. Map of the study area.

2. Methods

2.1. Study area

The Portland Metropolitan area in Oregon and Clark County,

Washington, lie on opposite sides of the Columbia River in the
Pacific Northwest (Fig. 1). This region experiences a Marine west
coast climate with wet, mild winters and cool, dry summers. Tem-
peratures average 15–27 ◦ C in the summer months, to 1–10 ◦ C in
the winter. Peak urban stream flow occurs during the wet win-
ter months, while the dry summer months experience lower flows
(Fig. 2). Individual stream’s average flows vary from 0.5 to 30 m3 /s
annually [28,29].
Streams in both areas are listed on the federal 303d list for water
quality violations. In Portland, nearly the entire stream reaches Fig. 2. Average monthly rainfall and precipitation for Portland, November
of Tryon Creek and Johnson Creek are listed for water tempera- 2006–April 2009.
ture [22]. Johnson Creek is also listed for high E. coli levels, and
on the west side Fanno Creek is listed for toxins. In Clark County, Twenty-one sites from Portland and 30 sites from Clark County
the upper half of Burnt Bridge Creek, an urban stream running were selected based on the available sample dates, parameters,
through Vancouver, is listed for multiple reasons, including water and watershed characteristics. Each government agency collected
temperature, pH, and E. coli [30]. Portions of Salmon Creek are listed data based their sampling methods and quality control on USEPA
for temperature, pH, and dissolved oxygen violations. Other listed standards (Table 1).
streams in the county appear for DO and temperature violations.
Two watersheds, Jones Creek and Chelatchie Creek, are not listed;
their primary land covers are forest and agriculture. Table 1
Sampling methods.

Portland Clark County Burnt Bridge

2.2. Stream data
EC SM 2510 B Electrode – on site SM2510B
DO SM 4500-O G Membrane electrode – on site On site
We collected water quality data from several government agen- NN EPA 300.0 EPA 353.2 SM4500NO3I
cies, and downloaded spatial stream data from the US Geological pH SM 4500-H B Glass electrode – in situ On site
Survey (USGS) National Hydrography Dataset [31]. Washington TP EPA 365.4 EPA 365.1 SM4500PF
water quality data was collected from Washington Department of TS SM 2540 B EPA 160.3 –
Temp SM 2550 B Thermistor in situ
Ecology [30], as well as Clark County Environmental Services [32].
The Portland Bureau of Environmental Services has been collecting Portland data from the Bureau of Environmental Services; Clark County data, except
monthly stream data at select sites consistently since 1998–2010. Burnt Bridge, via CCES; Burnt Bridge data via WADE.
50 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58

Seven water quality parameters were chosen based on its impor-

tance to human and aquatic life in both Portland and Vancouver
study sites. Nitrogen nitrate NO3 + –N (NN) and total phosphorus
(TP) are generally considered to be direct measures of human
activity in an area, as fertilizers, vehicle emissions, and impervi-
ous surfaces increase the amount of NN and TP in their respective
natural cycles [12]. Total solids (TS) can be used as a quantita-
tive measure of aesthetics as suspended sediments in streams
make the water appear cloudy. This study also used conduc-
tivity (EC), dissolved oxygen (DO), pH, and water temperature
(Temp). These measurements are associated with predicting algae
bloom likelihood and habitat quality for fish and other aquatic
animals.
In order to account for the seasonal variation in stream flows, the
data were split into wet (November–April) and dry (May–October) Fig. 3. Contribution watershed areas to sampling points.
seasons. The seasonal data were aggregated to a geometric mean
for the entire period. The geometric mean was used because it is
a slightly more conservative estimate of aggregated water quality
issue is partly resolved. A consistent distance, however, has not
parameters than an arithmetic mean. It is also more appropriate
been agreed on in literature and riparian buffers range from 8 m to
to use a geometric mean when data are not normally distributed,
200 m [24,34–36].
which was the case for two parameters.
In order to determine the association between landscape vari-
ables and water quality at each monitoring site, this study uses
2.3. Independent variables sectioned watersheds and riparian buffers. These sectioned zones
limit the area associated with the sample site to the next site
Topographic and landscape variables are shown in Table 2. immediately upstream (Fig. 3). The riparian buffer was used to
The standard deviation of slope, derived here from a 10 m digi- determine if the immediate environment surrounding the stream
tal elevation model, has been used in past studies as a measure has a stronger relationship than the entire area. Watersheds were
of topography complexity where the study area is relatively flat, delineated from the 51 sample sites using the 10 m DEM in ArcGIS
which is the case in many of the urban watersheds [25]. The v10.0, while the riparian areas were created by buffering the
2006 US National Land Cover Dataset was used to categorize per- streams 100 m. The downstream watersheds and riparian areas
cent urban, forest, agriculture, and wetlands in each area, with were clipped to the upstream watershed, where applicable, to
areas of less than 0.1% not included for analysis. Structural vari- create sectioned watersheds and buffers. The area of SFR was nor-
ables include single family residential (SFR) taxlots and street malized to a percent coverage of the area, and the streets layer
density. These spatial data allowed researchers a finer scale with was normalized to length (m)/(1000) area (m2 ). SFR house age was
which to examine land development within the study area. The averaged from SFR taxlots present in the sectioned area. Land cover
percent area of SFR provided a measure of residential housing and topographic variables were calculated using the Spatial Analyst
impact. Average building age of SFR homes built before 2010 tools in ArcGIS.
was used as a measure of historical development. Street density
provides a measure of habitat fragmentation as well as imper-
vious surfaces. We used the 2010 taxlot and streets datasets 2.5. Normalization and variable statistics
produced by Clark County and the Portland Metropolitan Author-
ity. All variables were tested for a normal distribution using the one-
sample Kolmogorov–Smirnov test, which tests for normality by
2.4. Spatial analysis examining the observed and theoretical distributions and deter-
mining if the difference between them is significant [37]. Of the
The land area associated with a sample point is often the sub- seven seasonal water quality parameters, wet and dry for each
watershed upstream of the sampling site [8,10,33]. The size of this (total 14), three were found to be skewed. The dry season NN and
area varies based on the size of the watershed and the position of DO were transformed exponentially and logarithmically, respec-
the sampling point. Distant sources of pollutants, whether agricul- tively, while a single record was removed from the dry season EC
ture or urban in nature, may be diluted in the stream or stored in to resolve skewness. Of the independent variables, normalization
the soils before reaching the sample station, thereby giving an inac- was achieved by removing records less than 10% and performing
curate measure of association between the land cover and water log transformations (Table 3). Soil type A was removed entirely
quality parameter [8]. By examining just the riparian area around because it was present in only two watersheds and one buffer
a stream, delineated as a constant distance from the stream, this area.

Table 2
Independent variables used in analysis.

Agency Source Data Derived variable Original data

USGS National elevation dataset (10 m) Mean slope Elevation

Slope standard deviation
Mean elevation
USGS National land cover dataset (30 m) Agriculture Pasture, cultivated crops.
Forest Deciduous forest, evergreen forest, mixed forest.
Urban Low, medium, high intensity developed, open space.
Wetland Woody wetlands, emergent herbaceous wetlands.
NRCS Soil types ABCD hydrologic soil groups Soil survey geographic (SSURGO) database
 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58 51

Table 3
Independent variable statistics.

Buffer Watershed

Min Max Average Stdev n Min Max Average Stdev n

Urban 12.71b 100 73.53 27.83 48 0.89 100 58.68 31.05 51

Forest 0.1a 85.41 20.26 22.29 43 0.7 93.71 26.47 22.95 41
Agriculture 0.16a 43.14 14.31 14.12 26 0.23a 56.65 21.29 17.53 23
Wetland 0.01a 10.83 2.9 2.91 34 0.68a 19.39 7.22 5.19 33
Mean slope 0.31 33.5 13.07 7.85 51 1.11 33.53 10.49 7.8 51
Slope StDev 1.2 25.62 11.95 5.73 51 2.07 25.38 10.55 5.22 51
Mean elev. 10.38 546.47 114.34 93.52 51 39.69 593.85 130.18 95.68 51
Street density 0.63 15.37 6.92 4.11 51 1.04 21.84 10.45 5.04 51
%SFR 2.8 72.3 33.13 19.07 50 14.98 71.14 42.95 14.03 50
SFR age 2.8a 72.3 33.13 19.07 50 14.98 71.14 42.95 14.03 50
per b 0.84a 100 58.84 33.76 34 – – – – –
per c – – – – – 1.69a,c 100 56.66 39.89 36
per d 0.22a 100 28.14 27.8 39 0.28a 93.84 14.37 18.39 42

Variables before log or exponential transformation.

a
Log 10 transformed for analysis.
b
The 3 lowest values were removed to resolve non-normality and are not included.
c
The 4 lowest values were removed to resolve non-normality and are not included.

2.6. Statistical analysis density of sample sites varied across the study area. The GWR out-
puts include local residuals and R2 results, as well as a global R2
Multivariate OLS regression and GWR models were developed to [38]. The global R2 generated by ArcGIS is defined as the propor-
examine the relationship between the independent variables and tion of variable variance the regression model accounts for [39].
the water quality parameters. Multivariate analysis filters out the
significant variables across the landscape [17,19]. To find the inde-
pendent variables with the strongest correlation with the water Table 4
parameters, stepwise multiple linear regression (SMLR) was run Multivariate OLS regression models.
in PASW Statistics 17. With seven water quality parameters, two Regression Model
seasons, and two scales of analysis, 28 OLS models were generated
Buffer
(Table 4). SMLR models run using only those independent variables
Wet Season
identified as significant at the 95% confidence level. These variables EC 0.974 (Urban) + 58.896
were then used to run GWR and OLS regressions in ArcMap. An DO 1.417 (Forest) + 2.615 (SFR age) + 5.901
advantage of running an OLS regression in ArcMap is the output NN −0.038 (%SFR) – 0.053 (Mean slope) + 3.751
pH 0.031 (StDev slope) + 0.006 (Urban) + 6.024
from the process includes the residual for each site, allowing the
TP 0.042 (Urban) + 4.486
researcher to more easily test the residuals for spatial autocorrela- TS 0.533 (Urban) + 83.439
tion. Temp 0.03 (Urban) – 0.089 (StDev slope) + 0.005 (Mean
A law of geography is that things that are closer together are elevation) + 5.615
more likely to be related than things that are far apart. GWR cap- Dry Season
EC 0.486 (Urban) – 0.124 (Mean elevation) + 139.099
tures the local variations by weighting closer observations greater
DOa 627.083 (Mean slope) + 10,871.260 (SFR age) –
than those further away. OLS models are like the following. 11,571.635
NNb −0.003 (Mean elevation) – 0.008 (%SFR) + 0.59

p
pH 0.295 (SFR age) – 0.002 (Mean elevation) + 0.029
 = ˇ0 + ˇi xi + ε (1) (StDev slope) + 6.963
i=1 TP −0.017 (Mean elevation) + 0.346 (Streets) + 9.713
TS −0.409 (Mean elevation) + 5.63 (Mean slope) – 6.176
 represents the dependent variable, ˇ0 is the intercept, and ˇ0 xi (StDev Slope) + 186.291
are the coefficient and the independent variable. ε represents the Temp −0.01 (Mean elevation) + 0.191 (StDev slope) – 0.017
(%SFR) – 0.096 (Mean slope) + 15.06
error term, and p is the number of independent variables. Watershed
The GWR equation differs in that it incorporates the coordinates Wet Season
of each location. EC 5.315 (Streets) – 2.229 (StDev slope) + 93.006
DO 0.13 (Mean slope) + 0.021 (SFR age) + 8.838
p
NN 0.013 (SFR age) – 0.006 (Mean elevation) – 0.022
j = ˇ0 (uj , vj ) + ˇi (uj , vj )xij + εj (2) (%SFR) + 2.816
i=1
pH 0.006 (Urban) + 6.934
TP 0.053 (Urban) + 4.024
where j represents the location, the coordinates (uj , vj ) for each TS −1.728 (StDev slope) + 3.604 (Streets) + 99.903
Temp 0.152 (Streets) – 0.019 (%SFR) + 6.481
location are taken and multiplied by the local independent variable
Dry Season
xij . The model is calibrated using an exponential distance decay EC 2.596 (Streets) – 0.193 (Mean elevation) + 155.069
function. DOa 1248.699 (StDev slope) + 186.205
NNb −0.003 (Mean elevation) – 0.009 (%SFR) + 0.912
−dij2 pH −0.003 (Mean elevation) + 0.035 (StDev slope) + 7.458
Wij = exp (3) TP 0.076 (Urban) + 5.72
b2
TS 2.194 (Streets) – 0.219 (Mean elevation) + 150.29
The weight of site j as it effects site i, Wij , is calculated using the Temp −0.014 (Mean elevation) + 0.12 (StDev slope) + 14.961
distance (d) between sites i and j with b acting as the kernel band- All TP values were multiplied by 100 to accommodate software limitations.
width. The weight decreases rapidly when the kernel is smaller than a
Exponentially transformed.
b
the distance. For this study, an adaptive band was used because the Log 10 transformed.
52 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58

The residuals may be used to test the model’s accuracy at predict- the case for wetland or forested land covers, where wetland and
ing local conditions by running a test for spatial autocorrelation. forested area percentages decreased at the buffer scale in nearly
We used Global Moran’s I for the residuals of both OLS and GWR every case. Agriculture land increased at the buffer scale, albeit to
models to test spatial dependence. Moran’s I is calculated from the less than 1%.
following formula.
n n 3.3. Influence of scale and season
n i=1
W (Xi − X)(Xj − X)
j=1 ij
I= n 2
 n n (4)
(Xi − X) i=1 j=1
Wij The landscape variables identified as being significant in pre-
i=1
dicting water quality varied depending on the scale of analysis and
where, Xi and Xj refer to water quality at station i and station j, season. At the watershed scale, street density, a measure of imper-
respectively. X is the overall mean water quality, and Wij is the vious surfaces and landscape fragmentation, was significant in five
weight matrix. Like correlation coefficient, I is positive if both Xi cases (Tables 4 and 5). Likewise, forest land cover was not signif-
and Xj lie either above or below the mean, while it is negative if icant at the watershed scale, but came up as a significant variable
one station is above the mean and the adjacent stations are below for explaining variations in wet season DO at the buffer scale.
the mean [40]. Out of all the predicted correlations, 11 relationships between
Global models assume that relationships between water quality water quality parameters and explanatory variables occurred, and
and explanatory variables are the same across space. This is partic- the sign of the coefficients remained the same at both scales. Only
ularly problematic given the variation in land cover and multiple one model, dry season pH, switched from negative at the water-
sources of pollutants. In a study of Northwest England, urban and shed scale to positive at the buffer scale for standard deviation
agriculture land covers were determined to be highly correlated of slope. While the models were not consistent in identifying the
with NN levels [17]. Li et al. [24] found that NN was correlated with same independent variables for each parameter in both scales, the
forest cover during the wet season and bare land during the dry variables’ relationships with the water quality parameters remain
season. With multiple sources of pollutants and patchwork land- unchanged.
scapes, it is necessary to consider local conditions when modeling Dry season results at both scales are dominated with topo-
pollutant loads because pollutants can have multiple sources, both graphic variables. Mean elevation is negatively related to dry season
from direct point discharge and non-point sources. EC, NN, pH, TS, and TD. As shown in Fig. 6, the areas with a lower
mean elevation have higher NN values. In several cases on Johnson
3. Results Creek and Salmon Creek, this may be due to the higher proportion
of agriculture present (Fig. 5). Both the mean and standard devia-
3.1. Spatial and seasonal variations of water quality tion of slope appear multiple times in dry season models. Higher
standard deviations of slope is credited with increasing Temp in
As shown in Fig. 4, there are substantial variations in water the dry season at both scales, and decreasing TS at the dry sea-
quality over space and season. Generally, higher concentrations of son buffer scale and the wet season watershed scale. Gentle slope
conductivity, NN, pH, TP, and TS are associated with low elevation may be acting as a literal sink for TS, slowing runoff or stopping it
urban land covers, and EC, TS, and TP values are lower during the entirely and allowing the water to percolate through the surface.
wet periods than during the dry season (Fig. 4). Such spatiotempo- Less slope variability may be increasing stream water temperature
ral variations of water quality are related to the fact that catchment in a similar fashion, as slow-moving water has more residence time
characteristics are heterogeneous in space and time. Precipitation to absorb sunlight and heat from pavement.
in forested areas can dilute TS and TP concentrations in winter Structural variables – urban, percent SFR, and SFR age –
months [1,2,19], while in agricultural-dominant catchment, rain appeared in multiple models at both scales. The urban variable is
runoff can contribute to elevated level of nutrients and TS. Dur- present in EC, pH, TP, TS, and Temp wet season buffer models, as
ing the dry season, microorganisms in the water may be absorbing well as the pH and TP watershed models, positively correlated. Per-
more DO and NN, causing lower levels, while high TP values are cent SFR is negatively correlated to NN in all four models, as well
due to runoff from fertilized lawns and agricultural fields [7]. In as wet season buffer Temp and dry season watershed Temp.
most cases, NN is higher during the wet season, probably because The parameter’s season with the higher concentration did not
of runoff from impervious surfaces and the effect of dormant plants necessarily have higher R2 values (Table 6). DO and pH had stronger
not absorbing nitrogen. models during the wet season than during the dry season at both
scales, while NN and TS had higher values during the dry season
3.2. Land cover at the two spatial scales than during the wet season at both scales. At the watershed scale,
the wet season acted as a better predictor of water quality than the
The land cover percentages for each buffer and watershed varied dry season except for TS and NN. At the buffer scale, the dry season
across the region (Fig. 5). Portland generally had less agriculture and data with lower flows was a better predictor for five parameters,
forest present in the watersheds. Tryon Creek (TC) and Fanno Creek but not for DO or pH.
(FC) have some forest present, but are mostly urban, unlike Balch
Creek (BC) and the upper reaches of Johnson Creek (JC). The Salmon 3.4. Comparison of OLS and GWR models
Creek (SC) watershed in Clark County is similar to Johnson Creek,
as the upper reaches are more heavily forested and farmed than In every case, the global R2 value for GWR models was higher
the lower reaches that pass through urban areas. The Burnt Bridge than that for the OLS models (Table 6). Improvements of over 20%
(BB) watershed also follows this gradient, with a few sites in the occurred in 18 models, 11 at the watershed scale. Of these, half are
upper part of the watershed having some agriculture present, but for dry season and half for wet season models. The GWR analy-
mostly dominated with urban development. Other watersheds in sis includes local R2 values for each variable. The smallest ranges,
Clark County had a mixture of land cover types, with the exception under 0.20, occur in the buffer GWR analyses, and with the excep-
of Jones Creek, which is almost entirely forested. tion of three cases, the buffer models have the smallest ranges of
Every land cover present in the watersheds was also present in local R2 values. The smaller range of values implies that the local
the buffer areas. In almost every case, there is higher % urban land GWR model might be applied globally with the derived coefficients
cover at the buffer scale then at the watershed scale. This is not and constants with general consistency. Wider ranges imply that
 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58 53

Fig. 4. Spatial variations of wet and dry season water quality.

the outlier sites have other sources of pollutants and ought to be did not have spatial autocorrelation. At these exceptions, the
examined more closely. The watershed scale had higher R2 values results were found to be positively autocorrelated, clustered at the
than the buffer scale in 10 out of 14 cases, three in the dry season 5% significance level. Fig. 7 maps the GWR and OLS residuals of wet
when runoff is less frequent or voluminous. season DO and pH at the watershed scale. In OLS models, similar
As described in Table 7, except the three watershed scale OLS residuals values are clustered in a few watersheds. DO residuals
models, wet season DO and pH and dry season Temp, the models are clustered in Johnson Creek and Fanno Creek in Portland, while
54 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58

Fig. 5. Percent land cover at watershed and buffer scales.

Table 5
Regression model coefficients for explanatory variables.

Land cover Structure Topographic

Forest Urban Street density % SFR SFR age Mean elevation Mean Slope StDev slope

GWR OLS GWR OLS GWR OLS GWR OLS GWR OLS GWR OLS GWR OLS GWR OLS

Buffer Wet Season EC + +

DO + + + +
NN − − − −
pH + + + + + +
TP + +
TS + +
Temp + + + + − −

Dry Season EC + + # −
DO + + + +
NN − − − −
pH + + − − + +
TP + + − −
TS − − + + − −
Temp − − − − − − + +

Watershed Wet Season EC + + − −

DO # + + +
NN − − + + − −
pH # +
TP + +
TS + + # −
Temp + + − −

Dry Season EC # + # −
DO + +
NN − − − −
pH − − + −
TP + +
TS + + − −
Temp − − + +

− indicates a negative coefficient; + indicates a positive coefficient; # indicates there is no majority positive or negative coefficient present.
 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58 55

Table 6
Coefficient of determination (R2 ) in OLS and GWR models at two different spatial scales.

GWR R2 range Buffer Watershed

Buffer Watershed GWR OLS n GWR OLS n

Wet season
EC 0.12–0.9 0.35–0.83 0.69 0.48 48 0.72 0.60 51
DO 0.24–0.33 0.26–0.53 0.72 0.55 50 0.73 0.41 51
NN 0.55–0.63 0.57–0.72 0.58 0.42 32 0.59 0.51 31
pH 0.48–0.69 0.01–0.66 0.66 0.59 50 0.77 0.31 51
TP 0.16–0.51 0.07–0.49 0.38 0.24 30 0.52 0.37 30
TS 0.11–0.9 0.01–0.79 0.50 0.26 30 0.73 0.40 30
Temp 0.62–0.76 0.06–0.87 0.70 0.65 51 0.58 0.48 51
Dry season
EC 0.04–0.9 0.05–0.74 0.72 0.28 47 0.69 0.31 48
DO 0.25–0.63 0.04–0.56 0.27 0.26 42 0.34 0.30 50
NN 0.34–0.6 0.39–0.51 0.62 0.58 32 0.67 0.63 31
pH 0.5–0.67 0–0.65 0.55 0.49 48 0.61 0.33 51
TP 0.09–0.33 0.13–0.52 0.48 0.38 27 0.44 0.29 30
TS 0.05–0.47 0.01–0.75 0.86 0.76 27 0.80 0.56 30
Temp 0.58–0.78 0.3–0.55 0.70 0.66 46 0.78 0.49 48

0 values indicate a value less than .01

Table 7
Spatial autocorrelation of residuals in GWR and OLS models.

Buffer Watershed

GWR OLS GWR OLS

Wet season
EC −0.14 0.01 0.02 0.11
DO −0.01 0.03 0.10 0.39
NN −0.04 0.11 −0.24 −0.21
pH 0.01 0.12 0.01 0.24
TP 0.04 0.15 0.08 0.19
TS 0.19 0.21 0.13 0.19
Temp −0.07 0.04 0.05 0.16
Dry season
EC −0.23 0.06 −0.19 −0.06
DO 0.00 0.02 −0.01 0.01
NN 0.03 0.09 0.02 0.09
pH 0.16 0.19 0.08 0.18
TP 0.05 0.10 0.06 0.07
TS −0.17 −0.15 −0.11 0.00
Temp 0.03 0.05 0.11 0.27

Underlined values indicates the model is autocorrelated, with a p < .05

Fig. 6. Nitrogen nitrate, dry season, sorted low to high watershed mean elevation.
56 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58

collected multiple times a day during high and low flow events
found that TP loads increased with the rise in streamflow during a
high-flow event, and speculated that the TP was trapped within the
river bed sediments [41]. Although TP concentration is generally
lower during the wet season in our study watersheds, in Fanno
Creek, high levels of phosphorus in soils can contribute to elevated
levels of TP concentrations during high flow events [42].
Forest land cover is positively associated with wet season DO
at the buffer scale. This result was found in another study, where
high DO levels were associated with unmanaged forest land [13].
Given that the present vegetation ought to be absorbing the oxygen
in the water, this leads to speculation that the DO values might be
much higher or vegetation may not be acting efficiently [7]. In Mis-
erendino’s study [13], EC was found to be positively correlated with
forest land cover, however other studies found a negative relation-
ship [11,24]. The type of plant may be affecting the water chemistry,
and so a finer scale of analysis may be useful with consideration as
to the type of vegetation present.
Urban land cover at the buffer scale was found to be positively
correlated with EC [5,8,13]. A study in Boston found that residential
land had a positive correlation with conductance however this was
not seen here using the %SFR variable [38]. Urban land cover has
been found to be positively correlated with a number of parame-
ters, including EC, DO, NN, and TS [3,5,7–10,43]. This result is not
unexpected, as surfaces collect particulates and chemicals that are
flushed out during rain events into nearby streams. Residential land
cover, however, can act as a sink for NN. This is similar to Li et al.
[44] who found that NN had a negative correlation with vegetation
coverage. However, the actual makeup of the residential land in
Fig. 7. Residuals of wet season dissolved oxygen and pH at the watershed scale. this case is unknown. These areas may be completely paved over
or have large, fertilized lawns. Its appearance in models for NN and
Temp suggest that lawns are retaining water and allowing runoff
pH residuals are clustered in Salmon Creek in Vancouver. With
to slow down and percolate through the soil.
generally lower residual values, GWR model residuals, however,
The importance of topography has been noted in several other
do not show any spatial clustering.
studies [15,25,34,36]. Slope can act as sinks or sources for particu-
Fig. 8 illustrates the shift in coefficient values for EC, NN, DO, pH,
lates, as areas with high variability may act as sinks and steep slopes
TS, and Temp at both scales. The EC map is the most striking, with a
can hurry runoff to the stream, picking up particulates along the
clear divide between Clark County and Portland coefficient values
way. Besides the topographic variables, street density was found
for mean elevation. Elevation is credited as increasing the level of EC
to be significant at the watershed scale in 10 models. This is likely
in the waters in the urban watersheds of Tryon and lower Johnson
tied to the surface of trafficked streets funneling particulate-laden
Creek, as well as the more forested Balch Creek watersheds. North of
runoff into streams [45].
the river in Clark County, mean elevation is found to have a negative
relationship with the levels of EC. Elevation is positively associated
4.2. Scale
with dry season pH and DO in middle Johnson Creek, but it is neg-
atively related to pH and Do in the rest of other study streams. e
The watershed scale of analysis clearly generated a stronger
Elevation is consistently negatively associated with Temp and NN
model than the buffer scale. Carter et al. [46] noted that storm water
in all streams.
mitigation practices were not being fully exploited at the water-
In the pH, TS, and Temp maps, clusters of higher R2 values are
shed scale. Instead of identifying non-point sources, direct inputs to
evident in watersheds across the study area. Despite the diverse
streams were identified and mitigation projects built around them.
landscape present in Johnson Creek, relatively high R2 values were
In this study, street density, urban and residential land covers, as
still reported. In Burnt Bridge Creek, a gradient of R2 values is evi-
well as topographic variables played important roles in predicting
dent, as R2 values decrease from upstream forests to downstream
stream water quality. The finer resolution of this data, both in terms
urban areas. This may be due to the sectioned method used; per-
of spatial as well as categorically, indicates that general land cover
haps the GWR model did not capture the upstream variables as
categories, i.e., urban, do not capture key variances in land uses
thoroughly due to the distance decay built into the analysis. Addi-
affecting water quality. Lee et al. [8] had a similar finding, where
tionally, downstream urban areas may contain other sources of
analysis of land use patterns suffered due to poor spatial resolution
pollutants that were not included in our models.
and the generalization of urban land cover.

4. Discussion 4.3. Spatial autocorrelation

4.1. Water quality parameters and predicting variables The GWR models performed as expected by accounting for local
variance in generating local coefficients and constants. If the resid-
Sources of TP vary. Some studies claim underlying geology in uals are autocorrelated, then it is likely that the model missed a
areas for high TP loads, while others believe that the decrease in variable that explains the variation. Several OLS models in this
TP is due to the TP attaching itself to sediment and settling into study had residuals that were spatially autocorrelated. The GWR
the streambed [i.e., 36, 14, 18]. Bowes, Smith, and Neal used data models functioned correctly accounting for local variability in land
 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58 57

Fig. 8. Mean elevation coefficient signs and local R2 of GWR.

cover [38]. This suggests that GWR models, by taking into account variables than the smaller buffer scale in 10 cases. Riparian restora-
spatial autocorrelation, give higher predictive power than tradi- tion projects should not be discounted, nor should their impact on
tional OLS regression models. The GWR models also suggest how stream water quality be exaggerated. The cumulative effect of non-
the relation between water quality and each explanatory variable point sources across a watershed cannot be negated within a zone
might vary at a local scale. a set distance from a stream. Management practices must incorpo-
rate abatement projects watershed-wide instead of focusing on the
5. Conclusions stream’s immediate area [11].
The GWR models did generate higher R2 values than the OLS
This study found that depending on the type of analysis being models, thus implying that GWR models account for more varia-
performed, the parameter itself being examined, the season does tions in local areas. However, without examining the distribution
affect the results. While selecting the time period where the water of local coefficients and independent variables, the results may
quality parameter had a higher concentration or value generally be misinterpreted. Additional analysis would benefit from a wider
improved model strength. Topographic variables clearly appear to sample pool in the watersheds and the ability to remove site out-
be important in determining water quality parameters during the liers due to extreme land cover differences, e.g. Jones Creek. While
dry season at both scales. urban land cover and other structural variables did appear as sig-
Across season and parameters, the scale of analysis did matter. nificant in several models, their presence varies greatly across the
Higher R2 values were generated using sectioned watershed scale study area and also their effect. The GWR models developed in the
58 B. Pratt, H. Chang / Journal of Hazardous Materials 209–210 (2012) 48–58

current study can identify such local variations, which can inform [19] X.L. Wang, J.Y. Han, L.G. Xu, Q. Zhang, Spatial and seasonal variations of the
managers of the role of different landscape variables affecting water contamination within water body of the Grand Canal, China, Environ. Pollut.
158 (2010) 1513–1520.
quality in urban streams. [20] H. Chang, Spatial and temporal variations of water quality in the Han river
and its tributaries, Seoul, Korea, 1993-2002, Water Air Soil Pollut. 161 (2005)
Acknowledgments 267–284.
[21] CCES, Clark County NPDES Long-term Index Site Project: Quality Assur-
ance Project Plan, http://www.co.clark.wa.us/water-resources/documents/
The US National Science Foundation funded this study through Monitoring/LISP%20QAPP%20version%202.pdf (2004) (last accessed 23.05.11).
the Portland-Vancouver ULTRA-Ex grant (grant # 0948983). Addi- [22] Oregon Department of Environmental Quality (ODEQ). Water Quality
Assessment Database. http://www.deq.state.or.us/wq/assessment/rpt0406/
tional support was provided by Portland State University, Oregon search.asp (last accessed 25.05.11).
State University, Washington State University, and Reed College. [23] K.K. Gardner, B.L. McGlynn, Seasonality in spatial variability and influence of
Local agencies, the Portland Bureau of Environmental Services and land use/land cover and watershed characteristics on stream water nitrate con-
centrations in a developing watershed in the Rocky Mountain West, Water
Clark County Environmental Services, provided stream water qual-
Resour. Res. 45 (2009) W08411.
ity data. We also appreciate the Johnson Creek Watershed Council [24] S. Li, S. Gu, X. Tan, Q. Zhang, Water quality in the upper Han River basin, China:
and Portland Metro for sharing GIS data. Thanks also go to two the impacts of land use/land cover in riparian buffer zone, J. Hazard. Mater. 165
(2009) 317–324.
anonymous reviewers whose comments helped clarify some points
[25] L. Sliva, D.D. Williams, Buffer zone versus whole catchment approaches
of the paper. Views expressed are our own and are not those of the to studying land use impact on river water quality, Water Res. 35 (2001)
sponsoring agencies. 3462–3472.
[26] A.J. Wade, C. Neal, D. Butterfield, M.N. Futter, Assessing nitrogen dynamics in
European ecosystems, integrating measurement and modeling: conclusions,
References Hydrol. Earth Syst. Sci. 8 (2004) 846–857.
[27] S. Su, D. Li, Q. Zhang, R. Xiao, F. Huang, J. Wu, Temporal trend and source appor-
[1] C. Reimann, T.E. Finne, O. Nordgulen, O.M. Saether, A. Arnoldussen, D. Banks, tionment of water pollution in different functional zones of Qiantang River,
The influence of geology and land-use on inorganic stream water quality in the China, Water Res. 45 (2011) 1781–1795.
Oslo region, Norway, Appl. Geochem. 24 (2009) 1862–1874. [28] United State Geologic Survey Washington Water Science Center, Water,
[2] M.A. Cunningham, K.M. Menking, D.P. Gillikin, K.C. Smith, C.P. Freimuth, S.L. http://wa.water.usgs.gov/data/ (last accessed 26.05.11).
Belli, A.M. Pregnall, M.A. Schlessman, P. Batur, Influence of open space on water [29] Clark County Environmental Services (CCES), Flow Monitoring, 2011 (last
quality in an urban stream, Phys. Geogr. 31 (2010) 336–356. accessed 26.05.11) http://www.clark.wa.gov/water-resources/monitoring/
[3] C.P. Tran, R.W. Bode, A.J. Smith, G.S. Kleppel, Land-use proximity as a basis for flow.html.
assessing stream water quality in New York State (USA), Ecol. Indicators 10 [30] Washington Department of Ecology (WADE), Current Washington
(2010) 727–733. State Water Quality Assessment, 2008 (Last accessed 25 May 2011)
[4] R. Utz, K.N. Eshleman, R.H. Hilderbrand, Variation in physiochemical responses http://www.ecy.wa.gov/programs/wq/303d/2008/index.html.
to urbanization in streams between two Mid-Atlantic physiographic regions, [31] United States Geologic Survey (USGS), National Hydrography Dataset, 2011,
Ecol. Appl. 21 (2011) 402–415. http://nhd.usgs.gov/.
[5] X. Wang, Z. Yin, Using GIS to assess the relationship between land use and water [32] WADE, Environmental Information Management System, 2011 (last accessed
quality at a watershed level, Environ. Int. 23 (1997) 103–114. 21.05.11) http://www.ecy.wa.gov/eim/.
[6] M.J. Paul, J.L. Meyer, Streams in the urban landscape, Annu. Rev. Ecol. Syst. 32 [33] X. Yang, W. Jin, GIS-based spatial regression and prediction of water qual-
(2001) 333–365. ity in river networks: a case study in Iowa, J. Environ. Manage. 91 (2010)
[7] T. Tsegaye, D. Sheppard, K.R. Islam, A. Johnson, W. Tadesse, A. Atalay, L. Marzen, 1943–1951.
Development of chemical index as a measure of in-stream water quality in [34] H. Chang, Spatial analysis of water quality trends in the Han River basin, South
response to land-use and land cover changes, Water Air Soil Pollut. 174 (2006) Korea, Water Res. 42 (2008) 3285–3304.
161–179. [35] M.S. Nash, D.T. Heggem, D. Ebert, T.G. Wade, R.K. Hall, Multi-scale landscape
[8] S. Lee, S. Hwang, S. Lee, H. Hwang, H. Sung, Landscape ecological approach to factors influencing stream water quality in the state of Oregon, Environ. Monit.
the relationships of land use patterns in watersheds to water quality charac- Assess. 156 (2009) 343–360.
teristics, Landsc. Urban Plan. 92 (2009) 80–89. [36] F.B. Daniel, M.B. Griffith, M.E. Troyer, Influences of spatial scale and soil per-
[9] P. Zeilhofer, E.B.N.R. Lima, G.A.R. Lima, Land use effects on water quality in the meability on relationships between land cover and baseflow stream nutrient
urban agglomeration of Cuiaba and Varzea Grande, Mato Grosso State, central concentrations, Environ. Manage. 45 (2010) 336–350.
Brazil, Urban Water J. 7 (2010) 173–186. [37] R.A. Olea, V. Pawlowsky-Glahn, Kolmogorov–Smirnov test for spatially corre-
[10] M.H. Mehaffey, M.S. Nash, T.G. Wade, D.W. Ebert, K.B. Jones, A. Rager, Link- lated data, Stoch. Environ. Res. Risk Assess. 23 (2009) 749–757.
ing land cover and water quality in New York City’s water supply watersheds, [38] J. Tu, Z. Xia, Examining spatially varying relationships between land use and
Environ. Monit. Assess. 107 (2005) 29–44. water quality using geographically weighted regression I: model design and
[11] V.L. Versace, D. Ierodiaconou, F. Stagnitti, A.J. Hamilton, M.T. Walter, B. Mitchell, evaluation, Sci. Total Environ. 407 (2008) 358–378.
A.-M. Boland, Regional-scale models for relating land cover to basin surface- [39] Environmental Services Research Institute, Interpreting GWR Results, http://
water quality using remotely sensed data in GIS, Environ. Monit. Assess. 142 help.arcgis.com/en/arcgisdesktop/10.0/help/005p/005p00000032000000.htm
(2008) 171–184. (2011) (last accessed 4.06.11).
[12] A.L. Heathwaite, Multiple stressors on water availability at global to catchment [40] D. O’Sullivan, D.J. Unwin, Geographic Information Analysis, Wiley, Hoboken,
scales: understanding human impact on nutrient cycles to protect water quality 2010.
and water availability in the long term, Freshw. Biol. 55 (2010) 241–257. [41] M.J. Bowes, J.T. Smith, C. Neal, The value of high-resolution nutrient monitor-
[13] M.L. Miserendino, R. Casaux, M. Archangelsky, C.Y. Di Prinzio, C. Brand, A.M. ing: a case study of the River Frome, Dorest, UK, J. Hydrol. 378 (2009) 82–
Kutschker, Assessing land-use effects on water quality, in-stream habitat, ripar- 96.
ian ecosystems and biodiversity in Patagonian northwest streams, Sci. Total [42] C.W. Anderson and S.A. Rounds, Phosphorus and E. coli and their relation to
Environ. 409 (2011) 612–624. selected constituents during storm runoff conditions in Fanno Creek, Oregon,
[14] J.J. Rothwell, N.B. Dise, K.G. Taylor, T.E.H. Allott, P. Scholefield, H. Davies, C. Neal, 1998–99. US Geological Survey Water Resources Investigation Report, 2002,
A spatial and seasonal assessment of river water chemistry across North West http://pubs.usgs.gov/wri/wri024232/.
England, Sci. Total Environ. 408 (2010) 841–855. [43] M. Boeder, H. Chang, Multi-scale analysis of oxygen demand trends in an urban-
[15] J. Kang, S.W. Lee, K.H. Cho, S.J. Ki, S.M. Cha, J.H. Kim, Linking land-use type and izing Oregon watershed, USA, J. Environ. Manage. 87 (2008) 567–581.
stream water quality using spatial data of fecal indicator bacteria and heavy [44] S. Li, S. Gu, W. Liu, H. Han, Q. Zhang, Water quality in relation to land
metals in the Yeongsan river basin, Water Res. 44 (2010) 4143–4157. use and land cover in the upper Han River Basin, China, Catena 75 (2008)
[16] J. Tu, Spatially varying relationships between land use and water quality across 216–222.
an urbanization gradient explored by geographically weighted regression, [45] C. Wilson, Q. Weng, Assessing surface water quality and its relation with urban
Appl. Geogr. 31 (2011) 376–392. land cover changes in the Lake Calumet Area, Greater Chicago, Environ. Manage.
[17] J.J. Rothwell, N.B. Dise, K.G. Taylor, T.E.H. Allott, P. Scholefield, H. Davies, C. 45 (2010) 1096–1111.
Neal, Predicting river water quality across North West England using catchment [46] T.C. Carter, R. Jackson, A. Rosemond, C. Pringle, D. Radcliffe, W. Tollner, J. Maerz,
characteristics, J. Hydrol. 395 (2010) 153–162. D. Leigh, A. Trice, Beyond the urban gradient: barriers and opportunities for
[18] T.J. Sullivan, K.U. Snyder, E. Gilbert, J.M. Bischoff, M. Wustenberg, J. Moore, D. timely studies of urbanization effects on aquatic ecosystems, North Am. Ben-
Moore, Assessment of water quality in association with land use in the Tillam- thol. Soc. 28 (2009) 1038–1050.
ook Bay Watershed, Oregon, USA, Water Air Soil Pollut. 161 (2005) 3–23.