2.2.2. Other Input Data

In addition to soil data, application of the SWAT to the Urban Torrens catchment requires a number of input data types and maps:


## *2.3. Soil and Water Assessment Tool (SWAT) Model Set-Up*

SWAT (ArcSWAT version 2012 revision 637, USDA, Washington, DC, USA) is a continuous-time, semi-distributed simulator developed to assist water resource managers in predicting impacts of land management practices on water quality, including various species of nitrogen and phosphorous [13,15]. Spatially, the model subdivides a catchment into sub-basins, which are further delineated into hydrological response units (HRUs) based on physical characteristics of topography, soil, and land uses. In this study, application of the SWAT model resulted in a subdivision of the Urban Torrens catchment into 23 sub-basins and further into 125 HRUs using the multiple HRU thresholds method of soil, land use, and slope at 10, 20, and 10%, respectively. A modified Soil Conservation Service (SCS) curve number technique was used to estimate the streamflow, while the instream processes of TN and TP loads were estimated using the Enhanced Stream Water Quality Model (QUAL2E) [37]. Local information on management practices was imported into the model on the basis of expert knowledge. All land operations were scheduled by specific application date [15]. The growing season was defined from 1 June to 30 May for all urban land categories. In order to simulate management activities along land uses by agriculture, pasture, and orchards, the approach designed by Nguyen et al. [28] has been applied.

The parameter optimization of the SWAT model was based on sensitivity analysis, model calibration, model validation, and uncertainty analysis. These steps are in accordance with Neitsch et al. [15] and Arnold et al. [38], and will be discussed in the following section.

## 2.3.1. Parameter Sensitivity Analysis

The sequential uncertainty fitting (SUFI2) algorithm [38] of the SWAT Calibration and Uncertainty Program (SWAT-CUP, EAWAG, Dübendorf, Switzerland) allows analysis of global and one-at-a-time sensitivity. Here we applied the global sensitivity analysis to identify parameters for the calibration and validation steps.

#### 2.3.2. Model Calibration, Validation and Uncertainty

The parameter optimization was performed on a monthly time step using the generalized likelihood uncertainty (GLUE) algorithm that showed better calibration results for this case study when compared to the results of the SUFI2 program. GLUE performs a combined calibration and uncertainty analysis and accounts for all sources of uncertainties [39–41]. The calibration was conducted consecutively beginning with the streamflow followed by loads of sediment (TSS), TN and TP by means of the observed data from 2008 to 2015 and using the first three years as a warm-up period for model stabilization. Data from 2011 to 2013 were used for calibration, and validation was performed for the years 2014 and 2015. 5000 iterations were applied and the Nash-Sutcliffe (NS) [42] behavioral threshold of 0.5 was used for both streamflow and nutrient simulations. The coefficient of determination (R2), percent bias (PBIAS) [43], and NS efficiency coefficient were used as statistical criteria for evaluation of simulated results.

$$\text{NS} = 1 - \frac{\sum\_{i} \left( Q\_{m,i} - Q\_{s,i} \right)^{2}}{\sum\_{i} \left( Q\_{m,i} - \overline{Q\_{m}} \right)^{2}} \tag{6}$$

where: *Q* is the streamflow variable, m and s are measured and simulated values respectively, *i* is the *i*th datum, and the bar stands for average values.

The threshold for R<sup>2</sup> and NS greater than 0.5 for streamflow, TN and TP loads, and PBIAS ranging between ±25% for streamflow and ±70% for TN and TP loads, respectively, were considered as satisfactory modelling results [44]. The model uncertainty was expressed using the 95% prediction uncertainty index (95PPU) and statistically was evaluated based on the percentage of observation points bracketed by the prediction uncertainty band (p-factor) and the degree of uncertainty (r-factor). The values close to 1 were selected as satisfactory criteria for p- and r-factors [45].

#### *2.4. Scenario Analysis*

The calibrated model was used to simulate present and future scenarios of urbanization, and determine best-management practices (BMPs). The past (P) and present (BS) urbanization scenarios were represented through land use maps generated in ArcGIS, which indicated a substantial shift in the period from 2001 to 2015 from grassland to urban lands of low residential, institutional, and commercial lands (Figure 2). For the future urbanization scenario (FS0), the urban land budget will not change significantly according to the 'The 30 year Plan for Greater Adelaide' report, even though the urban population density is expected to triple [46]. Therefore, we maintained the relative percentage of land uses from 2015 (Figure 2), and reclassified the land use from low residential to high residential. The change in residential land use was reflected by an increase in the fraction of total impervious areas (FIMP) from 0.12 to 0.6, the amount of solids allowed to build up on impervious area (DIRTMX) from 125 to 225 kg/curb km, TN concentration in suspended solid loads from impervious area (TNCONC) from 360 to 550 mg N/kg sediment, and TP concentration in suspended solid loads from impervious area (TPCONC) from 96 to 223 mg P/kg sediment [15]. Values of parameters for the high-residential land use were extracted from the default database, while data for the low-residential land use were manually calibrated prior the auto-calibration step [47]. Meteorological input data were kept unchanged for all urbanization scenarios.

In order to determine potential BMPs for mediating water deterioration issues by urbanization, the following scenarios were designed:


Results for the past and future urbanization scenarios (P and FS) were compared with results of the present urbanization scenario BS. Results of the scenarios S1, S2 and S3 were compared with the scenario FS0. The statistical significance of scenarios of urbanization and BMPs were evaluated by means of a paired Wilcoxon test using an R tool according to the criteria ρ < 0.05.

**Figure 2.** Characteristics of past (P), present (BS), and future (FS0) land use changes in the Urban Torrens catchment from 2001 to 2045.

#### **3. Results**

#### *3.1. Model Sensitivity*

The global sensitivity analysis identified the runoff curve number (CN2), the baseflow alpha factor for bank storage (ALPHA\_BNK) and the moist bulk density (SOL\_BD) as most sensitive parameters for streamflow simulation whereas soil parameters SOL\_BD, SOL\_K, and SOL\_AWC were amongst the 10 most sensitive parameters (Table 2). In contrast, the organic N in the baseflow (LAT\_ORGN), the denitrification exponential rate coefficient (CDN) and denitrification threshold water content (SDNCO) proved most sensitive parameters for TN-load.


Note: <sup>a</sup> Values in parentheses indicate the soil layer; <sup>b</sup> Indicated value refers to a relative change in the parameter.

#### *3.2. Model Calibration, Validation and Uncertainty*

Calibrations for streamflow, TN, and TP resulted in coefficients of determination R<sup>2</sup> of 0.77, 0.62, and 0.56, NS of 0.77, 0.62, and 0.51, and PBIAS of −4.18, −2.91, and 24.87 respectively (Figure 3) that according to Moriasi et al. [44] indicate to be satisfactory. Validation for streamflow achieved R2 = 0.97, NS = 0.96, and PBIAS = −9.21, for TP R2 = 0.88, NS = 0.84, and PBIAS = −28.4 and for TN with <sup>R</sup><sup>2</sup> = 0.67, NS = 0.66, and PBIAS = −2.60. The p-factor for the uncertainty for flow ranged between 0.39 and 0.42, for TN between 0.83 and 0.71, for TP between 0.56 and 0.54, and the r-factor ranged between 0.75 and 0.79 for flow, 1.32 and 0.96 for TN, and 1.00 and 0.83 for TP during calibration and validation, respectively. The simulated peaks of streamflow, TN and TP loads corresponded well with monthly average precipitation in this urbanized catchment.

**Figure 3.** Hydrographs of observed and simulated streamflow and TN and TP loads of the Urban Torrens catchment during the calibration (2011–2013) and validation (2014–2015) periods.

## *3.3. Urbanization Scenarios*

Results of the scenario BS indicated an overall increase of 0.6% in monthly streamflow due to an increase of surface streamflow by 1% and a decrease of baseflow by approximately 2% (Figure 4). Whilst scenario BS also predicted an increase of the TP load by the 2.9% forecasted, TN loads changed insignificantly compared to the past urbanization scenario. The trends for streamflow, TN and TP are relatively similar for all months of the year. The scenario FS0 (future urbanization) suggested a significant increase in total runoff by 13.3% when compared to present urbanization. The partitioning of streamflow under the scenario FS0 (Figure 4) indicates a similar trend of an increasing surface streamflow from 77 to 82%, while baseflow is further decreasing from 23 to 18%. There is also a significant increase by 36.4% of the TP-loads at the catchment outlet suggested. Meanwhile, model results suggest a noticeable decrease in TN loads of 6.9%. From the results of the future urbanization scenario it is also evident that higher rates of nutrient load variations are observed for the rainy period in winter (June to August). Overall, the trend is clear and similar when the effects of past, present and future urbanization scenarios are compared with more pronounced effects of future urbanization versu s present urbanization.

*Water* **2018**, *10*, 1846

**Figure 4.** Streamflow, TN and TP responses to scenarios past urbanisation (P), present urbanisation (BS) and future urbanization (FS0). Pie charts show the relative proportion of different hydrological components. Bar graphs show the average streamflow, TN, and TP loads. Error bars show one standard deviation.

#### *3.4. Scenarios of Management Practices*

The Table 3 suggests that the scenario '30-m buffer strips' may achieve the highest reduction of the TN loads by 19.88% and of the TP loads by 4.13% compared to 1.22% and 2.73%, respectively, by the scenario 'river bank stabilization'. However, both scenarios predicted statistically insignificant changes in the catchment outflow. The scenario 'wetland development' showed a slight decrease in TN and TP loads, and buffering effects for the increased run-off into the main stream. The scenario that combined the three feasible management practices predicted a decreased runoff and the highest reduction in nutrient loads compared to results of the scenarios of the three single measures.


**Table 3.** Results of best management scenarios for flow, TN and TP loads at the Urban Torrens catchment. The relative change of best-management practices (BMP) scenarios are compared with the results of the FS0 scenario.

<sup>a</sup> indicates a significant different value (*p*-value < 0.05) for a BMP scenario as compared with the FS0 scenario based on the paired Wilcoxon test.

#### **4. Discussion**

This study applied SWAT for modelling impacts of urbanization on the Torrens catchment that is of high relevance Australia-wide.

With regards to model optimization, it proved to be advantageous to include field-based soil database of the Torrens catchment as model input that resulted in satisfactory streamflow simulation of both peak and base flows (Figures 3 and 5) and improved simulation results for nutrient loads when compared with results for the urban catchment Aldgate of a previous study by Shrestha et al. [27] that was based on a coarser representation of soils.

**Figure 5.** Flow duration curve of observed and simulated streamflow of the Urban Torrens catchment for the period from 2011 to 2015.

All urbanization-related scenarios predicted increased streamflow as a result of increased surface flow and decreased baseflow that corresponds well with findings by Richards et al. [16] and Sunde et al. [19]. The trends of predicted TP loads as appeared to be strongly positively correlated with streamflow since phosphorus is primarily transported by sediments in surface streamflow. The model predicted annual increases of TP loads by 4 g/ha/year in scenario BS and 65 g/ha/year in scenario FS0. In contrast, the scenario results showed that urbanization may decrease TN load most likely because of reduced soil leaching by up to 26 g/ha/year and up to 2 g/ha/year less nitrogen in the baseflow as revealed by the comparison between the scenarios FS0 and BS. The highest changes in

nutrient loads were recorded during autumn and winter months when pollutants are often released and transported in river catchments during short periods of intensive rainfall [8,54].

The comparison between the scenarios P and FS0 revealed significant increases in streamflow by 13.3% and in TP loads by 36.4% whilst TN-loads decreased by 6.9%. A possible explanation lies in the fact that pervious urban lands are modelled in SWAT as Bermuda grass, which in this study is configured similarly to pasture and grassland. Thus, the conversion of the low-residential land use accounting for 38.6% of the total land budget of scenario P to high-residential land by scenario FS0 corresponded to an increase of overall impervious surface in the study by approximately 20%. According to the study of Brun and Band [55], 20% is the threshold at which a dramatic change in runoff can be observed. It is also important to mention that in the case of the Urban Torrens catchment, the sewage system is completely separated from the stormwater drainage network. Therefore, an increasing urban population is projected to cause more fragmented housing sites and smaller-sized yards but not necessarily an increase in surface flow by waste water, and simulated streamflows and nutrient loads are only driven by stormwater.

In an attempt to determine measures for counteracting the impacts of urbanization we have examined three management options. The scenario that simulated the extension of the grassed buffer zone proved to be efficient in reducing TP loads whilst developing wetlands may buffer the flow into the main rivers. However, the implementation of these two measures in combination with river bank stabilization promises to be the best management practice in response to future urbanization of the Urban Torrens catchment.

#### **5. Conclusions**

As outcomes of this study, the following conclusions were drawn:


The SCS curve number approach performed well in this case study, but the results of streamflow calibration can be improved for densely urbanized sub-catchments by the Green and Ampt method in the SWAT model as suggested by Tasdighi et al. [56]. The results of scenario analyses in this study are restricted by simplified assumptions related to the default configuration of urban land uses in SWAT, and affected by some uncertainty. However, the results are showing most likely trends and magnitudes of expected effects of different land use developments and mitigation solutions on the catchment. Future research will build on outcomes of this study by extending the research to the downstream estuary region in order to address the effects of urbanization, and other potential sources that could combine with urbanization to cause significant threat to the riparian and coastal ecosystems.

**Author Contributions:** All authors contributed to the study design. Data acquisition and preparation were performed by H.H.N. Development of soil data for modelling was assisted by W.M. Modelling and writing of the original draft was done by H.H.N., F.R. and W.M. reviewed thoroughly the manuscript for submission. The final version of the manuscript was prepared by H.H.N.

**Funding:** This research was funded by the SA Water Corporation (grant number SW100309) and the Department of Environment, Water and Natural Resources (grant number F0000108271).

**Acknowledgments:** The authors acknowledge Malcolm Sheard for his professional advice and data support during the development of soil database for this study. We would like to thank Leon van der Linden, Brooke A. Swaffer and Robert Daly for their constructive comments on the study outputs. The authors also wish to thank Margaret Cargill for editing the manuscript for linguistic quality.

**Conflicts of Interest:** The authors declare no conflict of interest.
