Professional Documents
Culture Documents
Research article
a r t i c l e i n f o a b s t r a c t
Article history: Understanding the relationship between land use and water quality is essential to improve water quality
Received 27 October 2017 through carefully managing landscape change. This study applies a linear mixed model at both water-
Received in revised form shed and hydrologically sensitive areas (HSAs) scales to assess such a relationship in 28 northcentral New
16 February 2018
Jersey watersheds located in a rapidly urbanizing region in the United States. Two models differ in terms
Accepted 21 February 2018
of the geographic scope used to derive land use matrices that quantify land use conditions. The land use
matrices at the watershed and HSAs scales represent the land use conditions in these watersheds and
their HSAs, respectively. HSAs are the hydrological “hotspots” in a watershed that are prone to runoff
Keywords:
Hydrologically sensitivity areas
generation during storm events. HSAs are derived using a soil topographic index (STI) that predicts
Soil topographic index hydrological sensitivity of a landscape based on a variable source area hydrology concept. The water
Linear mixed model quality indicators in these models are total nitrogen (TN), total phosphorus (TP) and total suspended
Land use matrix solids (TSS) concentrations in streams observed at the watershed outlets. The modeling results suggest
Water quality that presence of low density urban land, agricultural land and wetlands elevate while forest decreases
Northcentral New Jersey TN, TP and/or TSS concentrations in streams. The watershed scale model tends to emphasize the role of
agricultural lands in water quality degradation while the HSA scale model highlights the role of forest in
water quality improvement. This study supports the hypothesis that even though HSAs are relatively
smaller area compared to watershed, still the land uses within HSAs have similar impacts on down-
stream water quality as the land uses in entire watersheds, since both models have negligible differences
in model evaluation parameters. Inclusion of HSAs brings an interesting perspective to understand the
dynamic relationships between land use and water quality.
© 2018 Elsevier Ltd. All rights reserved.
1. Introduction events (Shuster et al., 2005; Walsh et al., 2005), which causes
streambank erosion, modifies channel morphology, transports
Rapid urbanization has significantly transformed the landscape nutrients, metals, pharmaceuticals, and toxic substances to streams
in the U.S. during last four decades (USDA, 2015). New Jersey being (Fitzpatrick et al., 2005; Hatt et al., 2004; and Kolpin et al., 2002),
the most densely populated state in the U.S. represents such a and directly or indirectly changes hydrological, biological, and
typical landscape transformation. From 1986 to 2012, approxi- chemical processes of an aquatic ecosystem (Li and Zhang, 2011; Li
mately 29 percent increase in urban lands was observed in New et al., 2009; Yu et al., 2013). Water quality degradation has
Jersey, accompanied by 26.7 percent decrease in agricultural lands, prompted an increasing interest in better understanding how land
6.7 percent decrease in forest, and 5.4 percent loss in wetlands uses in a landscape affect downstream water quality (Huang et al.,
(Lathrop et al., 2016). 2013; Li et al., 2009; Pratt and Chang, 2012; Wan et al., 2014; Wilson
Urbanization increases impervious surface area and alters and Weng, 2010). There are numerous researches that have
magnitude, volume, frequency, and timing of high streamflow attempted to better understand the effects of land use on water
quality (Giri et al., 2012, 2016; Huang et al., 2013; Li et al., 2009;
Nejadhashemi et al., 2011; Wilson and Weng, 2010). Two most
commonly used approaches are biophysical watershed modeling
* Corresponding author.
and statistical modeling. The biophysical watershed modeling
E-mail address: subhasis.giri@rutgers.edu (S. Giri).
https://doi.org/10.1016/j.jenvman.2018.02.075
0301-4797/© 2018 Elsevier Ltd. All rights reserved.
310 S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319
involves extensive input data, model calibration, and in-depth 2. Materials and methods
modeling knowledge and has become a specialized expertise in
water resource management. Statistical modeling offers a simpler 2.1. Study area
alternative to biophysical watershed modeling by directly linking
land use conditions and water quality that can be applied efficiently This study was conducted in 28 watersheds located in the
to a large region. Most statistical modeling studies use non-spatial northcentral New Jersey including Essex, Hunterdon, Mercer,
land use matrices such as percentage of urban land and impervious Middlesex, Monmouth, Morris, Somerset, Sussex, Union, Passaic,
surface cover in a watershed to determine the effects of land use on Burlington, and Ocean Counties (Fig. 1). All 28 watersheds were
downstream water quality (e.g. Comelo et al., 1996; Johnson et al., located in three physiographic regions including Valley and Ridge,
1997). Such non-spatial landscape indicators assume that each Highlands, and Piedmonts where VSA hydrology is considered to be
part of landscape has equal potential to affect stream water quality, a dominant hydrological process for runoff generation (Qiu, 2009;
which is overly simplistic and might misunderstand the relation- Walter et al., 2002). There were 68 watersheds with the long-
ship between land use and water quality (Williams et al., 2005). term water quality monitoring stations in the region, but we only
Two types of progress have been made to overcome the limi- selected these 28 watersheds having no water transfer and located
tation of such simplistic non-spatial landscape representation. First, entirely in New Jersey.
an inverse distance weighted statistical approach was developed to The watersheds were delineated using the Streamstats, a web-
assess the relationship between land use and water quality. This based watershed delineation tool developed and maintained by
approach assigned more weights to land uses closer to a stream the U.S. Geological Survey. Latitude and longitude of each water
than land uses further away from the stream when evaluating the quality monitoring station were submitted through the Streamstats
impacts of land use on stream water quality (Kennen et al., 2008; interface to obtain the watershed boundary. An example of delin-
Peterson et al., 2010). Second, critical areas such as riparian zones eated watershed using the Streamstats is presented in Fig. 2.
of streams were considered to have the most significant impacts on These watersheds vary in size from 24 square kilometers (km2)
stream water quality (NRC, 2002). The land uses within the riparian to 2062 km2. These watersheds had experienced varying degrees of
zones instead of whole watershed were consequently used to urbanization. The primary urbanization form is the expansion of
assess the impacts of land use on stream water quality (Baker et al., low density residential areas, accompanied by losses in agricultural
2006; Paringit and Nadaoka, 2003). Although the riparian zone land, forest, and wetlands (Lathrop et al., 2016). Such changes in
approach is interesting, it has some drawbacks (Qiu, 2009). First, land uses have altered watershed hydrology as well as the physical,
there is no uniform way of defining the width of riparian zones. chemical, and biological condition of streams in New Jersey
Second, riparian zone is not a surrogate measure of hydrological (Kennen et al., 2008, 2010).
sensitivity considering spatially varying hydrological connectivity
in landscape. 2.2. Soil topographic index
We use a concept called hydrologically sensitive areas (HSAs) to
improve the understanding of hydrological connectivity between Soil topographic index (STI) is an indicator of hydrological
terrestrial landscapes and aquatic stream. HSAs are the areas in a sensitivity of a landscape and is calculated using following equation
watershed having higher propensity to generate runoff. This is (Giri et al., 2017; Buchanan et al., 2014; Qiu, 2009; Walter et al.,
consistent with the variable source area (VSA) hydrology concept 2002):
where the primary source of runoff is saturated areas in uplands,
whose scale varies depending on storm intensity (Qiu et al., 2014; a
STI ¼ ln (1)
Walter et al., 2000). These smaller saturated areas in a watershed T tanðbÞ
primarily generate and transport pollutants to streams and influ-
ence stream hydrographs (Hewlett, 1982). where a is the upslope contributing area per unit contour length
Although some studies (e.g. Easton et al., 2008; Heathwaite (m), b is the local surface slope (mm1), T is a soil transmissivity
et al., 2005; Qiu, 2009; Walter et al., 2000, 2009) have used VSA (m2/day) computed as the product of the saturated hydraulic
hydrology to identify HSAs and target best management practices conductivity (m/day) and the depth to a restrictive layer (m). STI
(BMPs) to control nonpoint source pollution, however, they were indicates the likelihood of a point in a watershed to generate runoff
conducted on agricultural field or small watershed scale with im- and is used to identify spatial distribution of runoff contributing
plicit assumption that BMPs within HSAs would be more effective a
areas in watersheds (Qiu, 2009). Wetness index, ln tanð bÞ , was the
in improving water quality. No study has attempted to empirically
test and validate whether the high-intensity land uses within HSAs most common form of topographic index (Beven and Kirkby, 1979).
such as agricultural and urban lands contribute more to water STI extends this topographically based wetness index by consid-
quality degradation across watersheds. The objectives of this study ering soil water storage capacity above a restrictive layer and is
is to assess the impact of land use at both HSAs and watershed scale more applicable to the hydrological process in the Northeast in the
on water quality using a linear mixed model and to test the hy- U.S. (Buchanan et al., 2014; Qiu, 2009; Walter et al., 2002).
pothesis that even though HSAs represent a small fraction of a
watershed, the land uses within HSAs have the similar impacts on 2.2.1. Soil transmissivity
downstream water quality as the land uses in the watershed. Soil transmissivity was based on soil saturated hydraulic con-
Validation of such a hypothesis is a critical step to the development ductivity and soil depth (Buchanan et al., 2014) of topsoil layers in
of efficient watershed management strategies for water quality the soil survey geographic (SSURGO) database downloaded from
improvement. For example, watershed manager can strategically the U.S. Department of Agriculture (USDA) Geospatial Gateway. The
target HSAs for implementing BMPs to enhance their effectiveness information on saturated hydraulic conductivity and the soil depth
and cost-effectiveness in improving water quality if the hypothesis in the SSURGO database was extracted using a soil data viewer
is deemed to be true. developed by USDA Natural Resources Conservation Service. This
saturated hydraulic conductivity for calculating T in Equation (1) is
the geometric mean of the saturated hydraulic conductivity of all
soil layers above a restrictive layer (Qiu, 2009). A correction factor
S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319 311
of 0.000864 was multiplied to the soil transmissivity layer to merged into a single soil transmissivity layer for the region, which
convert the unit into m2/day. The soil transmissivity layer was is finally clipped into 28 soil transmissivity layers for those selected
created for each County. All soil transmissivity layers were then watersheds based on the watershed boundaries.
312 S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319
Fig. 2. The U.S. Geological Survey Streamstats interface for watershed delineation.
2.2.2. Wetness index than or equal to ten were the HSAs. The areas of the delineated HSA
The wetness index was based on the light detection and ranging made up approximately 27 percent of the watersheds for 28 wa-
(LiDAR) digital elevation model (DEM) at a 3-m resolution. The tersheds, which is slightly larger than 20 percent, the potential
LiDAR DEM for New Jersey was obtained from New Jersey saturation areas under typical topographic and climate conditions
Department of Environmental Protection (NJDEP). This fine scale suggested by Herron and Hairsine (1998). Different STI threshold
elevation data better represents the complexity of the terrain values were assessed in our other watershed studies with signifi-
condition in the study area (Valley and Ridge, Highlands, and cant local stakeholder engagement in the region and the threshold
Piedmonts). Similar to soil transmissivity layers, separate LiDAR value of ten seemed to delineate HSAs that are empirically verifi-
DEMs were created for each of those 28 watersheds by clipping the able and administratively manageable (Qiu, 2015; Qiu et al., 2014).
statewide LiDAR DEM using the watershed boundaries. Fill, calcu-
lation of slope, and calculation of catchment area were performed 2.4. Land use matrix
one-at-a-time to generate the wetness index for each watershed. To
ensure a continuous flow in each grid of LiDAR DEM, the depression Land use matrices for these watersheds and their HSAs was
was filled using fill sinks (Planchon and Darboux, 2001). Least extracted from the 2007 land use/cover data maintained by NJDEP
Square Fitted Plane method (Horn, 1981; Costa-Cabral and Burges, to understand the relationship between land use and water quality.
1996) was used to calculate slope of each gird while multiple The 2007 land use/cover data were produced by visually inter-
triangular flow direction method (Seibert and McGlynn, 2007) was preting color infrared aerial photography. Through this process,
performed to obtain the catchment area of each grid. The wetness photo interpreters examine each image and classify the image into
index was generated using the SAGA geographic information sys- various land use/cover categories based on a modified Anderson
tem in R. The soil transmissivity was then combined to the wetness classification system (NJDEP, 2010). The areas of various land use
index to create the soil topographic index for each of the 28 wa- categories were extracted from the 2007 land use/cover data
tersheds following Equation (1). Fig. 3 presents an example of how (agricultural land, forest, high density urban land, medium density
wetness index and soil transmissivity are combined to develop STI urban land, low density urban land, rural residential, wetlands, and
for one of the watersheds in the study area. water) for these watershed and their HSAs. High density urban land
is located at densely populated urban areas having either high
2.3. Hydrologically sensitive areas (HSAs) density single units or multiple units on one eighth to one fifth acre
lots (NJDEP, 2007). Medium density urban land is found in urban/
HSAs are the areas in a watershed having high potential to suburban areas and consists of residential units on greater than one
generate runoff (Qiu, 2009). Identification of HSAs is essential to eighth acre up to half acre lots while low density urban land con-
optimize the use of resources for both stakeholders and policy tains single residential unit on greater than half acre to one acre lots
makers. Qiu (2009) discussed several ways of delineating HSAs in a (NJDEP, 2007). Rural residential consists of single residential unit
watershed and recommended a simple threshold value approach. between one acre to two acre lots (NJDEP, 2007). Land use matrices
Following the approach, the areas in a watershed having the STI include the percentage of each type of land use in watersheds and
values greater than a threshold value are considered to be HSAs in their HSAs. High density urban land and medium density urban
that watershed. This study delineated the HSAs using the STI land were then combined into high medium density urban land to
threshold value of ten, i.e. the areas having the STI values greater capture the effect of urbanization with higher land use intensity on
S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319 313
Fig. 3. Spatial distribution of (a)wetness index, (b)soil transmissivity and (c) soil topographic index for a selected watershed.
Table 1
Descriptive statistics of percentages of land uses in watersheds and HSAs.
forest, 28 percent urban, 13 percent agriculture, 22 percent wet- characteristics of watershed i, and εij is the residual. We
lands, and 7 percent water, where urban included high, medium, assumed that εij ~ N (0, s2) and ui ~ N (0, t2) for this analysis with
and low density urban lands and rural residential. t2 capturing the variation due to different watershed charac-
teristics, and s2 measuring the unexplained variation.
2.5. Water quality monitoring data
A Log transformation of water quality data was performed to
The locations of the water quality monitoring stations were avoid the impacts of the skewness in the response variables in
downloaded from NJDEP and the water quality monitoring data statistical analysis. The predictors were expressed in percentage
including total nitrogen (TN), total phosphorus (TP) and total sus- and rural residential is considered as baseline. The parameters in
pended solid (TSS) for each station were obtained from National the model were estimated using lme function in nlme package by
Water Quality Monitoring Council data portal. Available water Maximum likelihood method in R. A backward stepwise elimina-
quality monitoring data for each station was downloaded and was tion of predictors was performed using Akaike information crite-
expressed as concentration (mg/l). The downloaded water quality rion (AIC) based on an initial regression analysis to assess the
data were further processed to keep the data only from 2006 to individual predictors' influence for each water quality indicator. In
2008 for each station for statistical analysis as described below to the initial step, all predictors were added in the regression equation
assess the effects of the 2007 land use conditions on water quality. and AIC was calculated. In the next step, one of the predictor was
The water quality monitoring data in three years (2006e2008; one dropped from the regression equation and a new AIC was calcu-
prior and one after) instead of only one year (2007) were used to lated. If the new AIC is smaller than the previous AIC, then the
capture the broad impacts of land use on water quality due to un- dropped predictor was removed permanently from the regression
known and uncertain land use and climate conditions. The water equation. However, if the new AIC is greater than the previous AIC,
quality at these monitoring stations were sampled 4e5 times a year the dropped predictor was added back to the regression equation.
depending the available funding in New Jersey. There were 12e15 This process would repeat until minimum AIC was obtained. Two
water quality measurements per site during the three-year period. additional model evaluation measures, Bayesian information cri-
terion (BIC) and log-likelihood (LogLik), were also used to deter-
mine which land use matrix (within HSAs or whole watershed)
2.6. Statistical analysis
would project a stronger relationship between land use and water
quality. To visualize the data and assess their relationship for
A statistical analysis was conducted to understand the rela-
selecting predictors, Spearman's rank correlation coefficient, scat-
tionship between the land use matrices in the watersheds and the
terplot, and histogram were also plotted prior to the linear mixed
HSAs to the water quality measurements (TSS, TN, and TP) observed
modeling.
at the watershed outlets. The land use matrix was a set of predictors
and the water quality measurements were response variables. A set
of predictors (i.e. the land use matrix in 2007) would correspond to 3. Results and discussion
multiple responses (i.e. several water quality measurements
observed from 2006 to 2008) and the number of responses would 3.1. Visualization of data and correlations
also vary across watersheds because different numbers of moni-
toring could be conducted at different monitoring states. Therefore, Fig. 4 and Fig. 5 present the Spearman's correlation matrices
unequal repeated measures (the water quality data) were modeled among water quality indicators (TN, TP, and TSS) and all land use
in R using the following linear mixed model for each of the three categories for the watersheds and their HSAs, respectively. Spear-
water quality indicators at both watershed and HSA scales: man's rank correlation coefficients are presented in the lower
panel, scatter plots are located in the upper panel, and histograms
Yij ¼ bX i þ ui þ εij (2) are presented in the diagonal. The line in the scatter plot represents
the simple linear regression between a pair of two variables. The
where i was the index for watershed and j was the index for the correlation between two variables is presented by Spearman's rank
number of measurements on water quality varying by watershed. correlation coefficients while the distribution of data is visualized
Yij was the observed water quality measurement for watershed i, by histogram. According to Mukaka (2012), the absolute value of
the coefficient below 0.30 indicates negligible correlation, between
i ¼ 1, …, 28, on its jth measurement, j ¼ 1, 2, …, ni as repeated 0.30 and 0.50 low correlation, between 0.50 and 0.70 moderate
measures; Xi was a land use matrix, i.e., the percentage of correlation and above 0.70 high correlation.
agricultural land, forest, high medium density urban land, low Fig. 4 indicates there are positive moderate correlation between
density urban land, rural residential, wetlands, and water in TN and both high medium density and low density urban lands,
watershed i or its HSAs, b represented the fixed effects of these negative low correlation between TN and forest, and negligible
predictors, ui represents the random effect due to the unique correlation between TN and other land uses including agricultural
S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319 315
Fig. 4. Spearman's Correlation Matrix for water quality indicators and land uses in watersheds: A for agricultural land, B high medium density urban land, C low density urban land,
D water, E wetlands, F forest, and G rural residential; Bold numbers indicate that coefficients are statistically significant at 10 percent level of confidence.
lands, water, wetlands and rural residential at the watershed scale. and adds forest in its contributing land use matrix. Unlike the high
The similar relationships between TN and land uses are observed at medium density urban land at the watershed scale that has posi-
the HSA scale except that the correlation between TN and low tive, but insignificant impact on TN concentration, forest has
density urban land, which is downgraded from moderate to low negative and insignificant impact on TN concentration in streams at
(Fig. 5). the HSA scale. Similar to the watershed-scale model, the HSA scale
The correlation analyses also identify positive moderate corre- model shows that agricultural land and low density urban land
lation between TP and both high medium density and low density have positive and significant impacts, and wetland has positive, but
urban lands, negative moderate correlation between TP and forest, insignificant impacts on TN concentration (Table 1).
and negligible correlation between TP and other land uses at the The selection of the land use categories as predictors are eval-
watershed scale (Fig. 4). Similarly, the correlation between TP and uated by using three evaluation parameters (AIC, BIC, and Loglik) as
low density urban land is downgraded from moderate to low at the discussed above. Smaller AIC and BIC and larger Loglik value
HSA scale (Fig. 5). There are negligible correlations between TSS represent better model prediction. As shown in Table 2, the
and all land uses at both watershed and HSA scales. watershed scale model seems to have better predicting power than
the HSA scale model; however, the differences are generally very
small and mostly negligible.
3.2. Relationship between TN and land use matrix
The significant impacts of agricultural land and low density
urban land on TN concentration are understandable. Nitrogen fer-
Table 2 presents the fixed effect coefficient of different land uses
tilizer application in agricultural land is often one of the primary
on TN concentration in streams along with their p-values and three
sources to the elevated TN concentration in streams. The low
model evaluation parameters at both watershed and HSA scales. At
density urban land is a predominant urban development in New
the watershed scale, both agricultural land and low density urban
Jersey (Lathrop et al., 2016). Higher TN concentration from the low
land have significant and positive impacts on TN concentration. The
density urban land may be due to fertilizer application in lawn care
results indicate that the presence of low density urban land in
and presence of animal waste such as horses in the study area.
watersheds tends to have greater impacts on TN concentration than
Another possible explanation of higher TN concentration from low
agricultural land. For an example, one percent increase in agricul-
density urban land may be due to legacy nitrogen in the form of
tural land in a watershed would increase TN concentration by 0.263
biogeochemical nitrogen such as retention of nitrogen in the root
percent while one percent increase in low density urban land in a
zone from past agricultural activities. As most low density urban
watershed would elevate TN concentration by 0.424 percent. High
lands in New Jersey were converted from agricultural land. Such
medium density urban land and wetlands seem to have positive
biogeochemical legacy nitrogen might act as a long-term source of
impacts on TN concentration, however, such impacts are deemed to
mineralization and leaching to waterways resulting in increased TN
be statistically insignificant.
concentration in streams (Van Meter et al., 2016). Tsegaye et al.
The HSA scale model drops high medium density urban land
316 S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319
Fig. 5. Spearman's Correlation Matrix for water quality indicators and land uses in HSAs: A.1 for agricultural land, B.1 high medium density urban land, C.1 low density urban land,
D.1 water, E.1 wetlands, F.1 forest, and G.1 rural residential; Bold numbers indicate that coefficients are statistically significant at 10 percent level of confidence.
Table 2 Table 3
Land uses and their associated fixed effects and the model evaluation statistics on TN Land uses and their associated fixed effects and the model evaluation statistics on TP
concentration in streams. concentration in streams.
Predictors Watershed Scale HSA Scale Model Predictors Watershed Scale HSA Scale Model
Model Model
reactive phosphorus and TP loads to streams than agricultural land quality issue in the study area (NJDEP, 2014). Second, TSS concen-
in a Timber Lake restoration project in North Carolina. In the study tration in streams may primarily be due to the instream processes
area, the nutrient storing capacity of wetland might be surpassed, rather than overland processes in the study area. Qiu and Wang
resulting in the release of phosphorus into streams. (2014) found that more than 60 percent of the sediment load to
The HSA scale model shows the similar result as in the water- streams came from stream banks erosion and streambed sediment
shed scale model with a few exceptions. First, agricultural land resurfacing during large storm events in Neshanic River Watershed,
shows positive, but insignificant impacts on TP concentration. one of these 28 selected watersheds in this study. Furthermore,
Second, forest is added into the contributing land use matrix and Nagel et al. (2007) also found that more than 50 percent of sedi-
has a negative, but insignificant impact on TP concentration. ment loads in Cayuga and Catskills Watersheds in New York was
This TP modeling results are similar to the results found in other due to stream bank erosion where VSA hydrology was the domi-
studies. Pratt and Chang (2012) and Wan et al. (2014) observed that nant process for runoff formation. The model evaluation statistics
agricultural land and urban land were the primary contributor to shows the mixed results regarding the predicting power of the two
phosphorus loading to streams in the Portland metropolitan area in models. The watershed scale model shows better predicting power
Oregon and Xitiaoxi River Watershed in China, respectively. Tu in terms of AIC and Loglik and worse in terms of BIC, but their
(2011) found that the presence of forest significantly decreased differences are small and negligible.
phosphorus concentration in streams around the Boston metro-
politan area in eastern Massachusetts. 4. Conclusion
Similarly, the watershed scale model shows a better predicting
power than the HSA scale model because of the better model Understanding the effect of land use on water quality is vital to
evaluation statistics, i.e. lesser AIC and BIC, and greater Loglik sustainable water resources management. This study assessed the
values, but the differences in those values are small and negligible relationship between land use and water quality in watersheds
(Table 3). located in northcentral New Jersey using a regional linear mixed
model. Specifically, this study tried to identify which types of land
3.4. Relationship between TSS and land use matrix uses contribute to stream water quality degradation. The response
variables were TN, TP, and TSS concentration whereas the pre-
The selected land use predictors and their fixed effect on TSS dictors were agricultural land, forest, high medium density urban
concentration in streams, and associated p-values are presented in land, low density urban land, rural residential, wetlands, and water.
Table 4. The watershed scale model shows agricultural land, forest, The analysis was performed at both watershed and HSA scales
and high medium density urban land has negative impacts on TSS where HSAs of each watershed were delineated using STI. The areas
concentration in streams. The positive impacts of forest and high in a watershed having STI greater than ten were considered to be
medium density urban land on TSS concentration are significant. HSAs in that watershed.
Trees and shrubs in forest generally reduce soil erosion. High per- Although urbanization is considered to be a major cause of
centage of impervious surface area in high medium density urban stream water quality degradation, but no study further delineates
land density also helps prevent soil erosion. The impact of agri- the impacts of different urbanization forms on water quality. This
cultural lands on TSS concentration in streams is negative, but study first attempts to consider how different types of urban lands
insignificant in the study area. affect water quality. In this study, the urban land was further
The HSA scale model drops out agricultural land and only in- classified into high medium density urban land and low density
cludes forest, and high medium density urban land in its contrib- urban land to assess their impacts on water quality. The results at
uting land use matrix. Like the watershed scale model, the presence both watershed and HSA scales suggested that low density urban
of both forest and high medium density urban land density in HSAs land significantly contributed to elevated TN and TP concentration
reduce TSS concentration in streams. The impacts of forest are in streams. The watershed scale models also implied that high
significant while the impacts of high medium density urban land medium density urban land increased TN concentration, but
density are insignificant at the HSA scale (Table 4). significantly reduced sediment loads to streams. The HSA scale
Both watershed and HSA modeling results agree that the pres- models did not identify the significant impacts of other forms of
ence of forest significantly reduces TSS concentration in streams, urban land uses except the positive and significant impact of low
but disagree in every other aspect, which could be due to the density urban land on TN and TP. The modeling results also sug-
insignificant source contribution from upland in the study area. gested that agricultural land and wetlands increased while forest
First, TSS concentration in streams is generally not significant water reduced TN, TP or TSS concentration in streams with varying levels
of statistical significance.
This is the first study that tries to compare the impacts of land
Table 4 uses on water quality between the watershed and HSA scales. The
Land uses and their associated fixed effects and the model evaluation statistics for
modeling results suggested that the predicting power of linear
TSS concentration in stream.
mixed models both at HSA and watershed scale were similar
Predictors Watershed Scale HSA Scale Model because of negligible difference in three model evaluation statistics.
Model
This supports our hypothesis that even though HSAs represent
b-value p-value b-value p-value much smaller areas compared to watershed, the land uses in HSAs
Intercept 1.302 ***
0.000 1.303 ***
0.000 have similar impacts on downstream water quality as the land uses
Agricultural land 0.248 0.108 in entire watershed do. Furthermore, the HSA scale modeling re-
Forest 0.505*** 0.009 0.254 **
0.020 sults show some interesting insights into the dynamic relationship
High medium density urban land 0.446** 0.036 0.161 0.132
between land use and water quality. For example, the HSA scale
Model Evaluation Statistic
AIC 657.55 658.42 models downplayed the impacts of agriculture on water quality.
BIC 678.61 675.97 Both models identified agricultural land as a significant contributor
Loglik 322.77 324.21 to an increase in TN concentration in streams, but the HSA scale
***
significant with at least 1 percent level of confidence; **
5 percent level of con- model showed that agricultural land insignificantly contributed to
fidence; and * 10 percent level of confidence. increase in TP concentration and dropped out agricultural land as a
318 S. Giri et al. / Journal of Environmental Management 213 (2018) 309e319