Science of the Total Environment 753 (2021) 142007

Contents lists available at ScienceDirect

Science of the Total Environment

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

Robustness analysis of storm water quality modelling with LID

infrastructures from natural event-based field monitoring
Sijie Tang a,b, Jiping Jiang a,b,⁎, Yi Zheng a, Yi Hong c, Eun-Sung Chung d, Asaad Y. Shamseldin e,
Yan Wei f, Xiuheng Wang b
a
Shenzhen Municipal Engineering Lab of Environmental IoT Technologies, School of Environmental Science and Engineering, Southern University of Science and Technology, Shenzhen
518055, China
b
State Key Laboratory of Urban Water Resource and Environment, School of Environment, Harbin Institute of Technology, Harbin 150090, China
c
School for Environment and Sustainability, University of Michigan, Ann Arbor, MI 49109-1041, USA
d
Department of Civil Engineering, Seoul National University of Science and Technology, Seoul 01811, Republic of Korea
e
Department of Civil and Environmental Engineering, The University of Auckland, Auckland 1010, New Zealand
f
Shenzhen Howay Technology CO., LTD, Shenzhen 518000, China

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• ‘Robustness’ of LID design revealed from

the view point of SWMM model calibra-
tion
• Automatic calibration technique with
intensive field data makes objective
evaluation.
• Water quality process modelling of
greenbelts and rain gardens is of most
robust.
• Runoff of SS, TN and COD is better
captured by the SWMM models than
NH3-N and TP.
• Rain features and pollutant concentra-
tion level affect the reliability of LID
design.

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

Article history: Sponge city construction (SCC) in China, as a new concept and a practical application of low-impact development
Received 12 June 2020 (LID), is gaining wide popularity. Modelling tools are widely used to evaluate the ecological benefits of SCC in
Received in revised form 21 August 2020 stormwater pollution mitigation. However, the understanding of the robustness of water quality modelling
Accepted 25 August 2020
with different LID design options is still limited due to the paucity of water quality data as well as the high cost
Available online 26 August 2020
of water quality data collection and model calibration. This study develops a new concept of ‘robustness’ mea-
Editor: Ashantha Goonetilleke sured by model calibration performances. It combines an automatic calibration technique with intensive field
monitoring data to perform the robustness analysis of storm water quality modelling using the SWMM (Storm
Keywords: Water Management Model). One of the national pilot areas of SCC, Fenghuang Cheng, in Shenzhen, China, is se-
Robustness analysis lected as the study area. Five water quality variables (COD, NH3-N, TN, TP, and SS) and 13 types of LID/non-LID
Stormwater quality infrastructures are simulated using 37 rainfall events. The results show that the model performance is satisfactory
SWMM for different water quality variables and LID types. Water quality modelling of greenbelts and rain gardens has the
Sponge city best performance, while the models of barrels and green roofs are not as robust as those of the other LID types. In
Nash-Sutcliffe efficiency coefficient (NSE)
urban runoff, three water quality parameters, namely, SS, TN and COD, are better captured by the SWMM models
than NH3-N and TP. The modelling performance tends to be better under heavy rain and significant pollutant

⁎ Corresponding author at: Shenzhen Municipal Engineering Lab of Environmental IoT Technologies, School of Environmental Science and Engineering, Southern University of Science
and Technology, Shenzhen 518055, China.
E-mail address: jiangjp@sustech.edu.cn (J. Jiang).

https://doi.org/10.1016/j.scitotenv.2020.142007
0048-9697/© 2020 Elsevier B.V. All rights reserved.
2 S. Tang et al. / Science of the Total Environment 753 (2021) 142007

concentrations, denoting a potentially more stable and reliable design of infrastructures. This study helps to
improve the current understanding of the feasibility and robustness of using the SWMM model in sponge city
design.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction kilometres in area (Mei et al., 2018; Randall et al., 2019), smaller catch-
ments (Luan et al., 2019) and site-scale individual LID facilities (Zhang
Stormwater management is important for the safety, sustainability et al., 2018). In terms of the temporal scales, the time of duration
and resilience of urban water (Barbosa et al., 2012). There have been range from minutes up to years. SWMM supports both continuous
many concepts and practices proposed to manage urban stormwater, and event-based simulations. In the case of event-based simulations,
such as Best Management Practices (BMPs) and Low-Impact Develop- it can use one or several rainfall events (Alfredo et al., 2010; Fioretti
ment (LID) in the USA, Sustainable Urban Drainage System (SUDS) in et al., 2010; Wang and Altunkaynak, 2012). In the case of continuous
the UK, Water Sensitive Urban Design (SWUD) in Australia, Low Impact simulation, the data record length ranges from one day up to several
Urban Design and Development programme (LIUDD) in New Zealand, years (Ashbolt et al., 2013; Zhang and Shuster, 2014).
and Active, Beautiful, Clean Waters programme (ABC) in Singapore In terms of water quality, SWMM uses a simplified approach to sim-
(Fletcher et al., 2015). Among them, the LID concept is most widely ulate the pollutant build-up and wash-off process, which results in a
used. However, the latest development in urban stormwater manage- large internal uncertainty in the characterization of the time-variant
ment is the concept of sponge city proposed by the Chinese government stormwater quality process. Most recent studies have noted and sum-
in 2013 as a localized and updated version of LID (Wang et al., 2018). marized the uncertainty associated with pollutant build-up and wash-
Since the introduction of the sponge city concept, approximately 30 off modelling (Wijesiri et al., 2016a, 2016b). Thus, some studies have
pilot sponge city projects have been constructed in major cities in tried to revise the pollutant build-up and wash-off modules in SWMM
China, such as Beijing, Shanghai and Shenzhen. In early 2020, the Chi- (Baek et al., 2020) and proposed new physical-based modelling tools
nese Ministry of Housing and Urban-Rural Development led a perfor- for the pollutant build-up and wash-off process (Hong et al., 2016).
mance appraisal of these pilot sponge city projects. Moreover, external factors, such as the high time resolution of
One of the major ecological benefits of sponge city construction and stormwater events, expensive water quality data collection and large
the associated stormwater management practice is the ability to reduce uncertainty regarding pollution sources, make things worse (Baek
stormwater pollution, thereby promoting the urban water environment et al., 2020; Zhang et al., 2017). This results in large challenges for the
and assisting with the regulation of biogeochemical cycles (Barbosa robust design of LID infrastructures and the practical effects of
et al., 2012; Eckart et al., 2017; Yang and Toor, 2018; Zhang et al., stormwater pollutant mitigation. Previous studies reported uncertainty
2020; Zheng et al., 2014). To some degree, sponge city construction is assessment by rainfall simulators with small road surface plots (Wijesiri
very well linked to urban water pollution mitigation and pollution con- et al., 2016a) as experiment-oriented investigations and uncertainty
trol (Deletic and Wang, 2019; Ren et al., 2017). However, characterizing quantification by the GLUE method (Freni et al., 2009) as theory-
the water quality of LID infrastructures, such as ecological tree pools, oriented investigations. The Markov Chain Monte Carlo (MCMC) is
green roofs, rain gardens and permeable pavements, with field observa- one of most popular tools, deriving a series of methods, such as Shuffled
tions is more difficult and expensive than measuring the water quantity Complex Evolution Metropolis algorithm (SCEM-UA), which is well
characteristics, namely, hydrological processes (Yang and Toor, 2018; suited to infer the posterior distribution of hydrologic model parame-
Zeng et al., 2019). Moreover, there is no universal consensus on the ters (Vrugt et al., 2003). More research benefits from the combination
evaluation of the pollution mitigation performance of LID infrastruc- of MCMC and Bayesian subsequently (Beven, 2010). Besides, the multi-
tures (Zeng et al., 2019). Very few studies have attempted to quantify objective optimization was also introduced (Vrugt and Robinson,
the effects of sponge city construction because of the lack of large- 2007). Total suspended solids are the most concerning water quality var-
scale synchronized water quality monitoring of diverse LID infrastruc- iables. Few studies have been conducted based on full-scale field moni-
tures. In the development of the pilot sponge city project in China, toring and covering C, N, and P nutrients.
there is a compulsory water quantity and water quality monitoring SWMM calibration can be traditionally performed manually by
programme linking a smart decision support system for urban water changing one parameter at a time and comparing the simulation with
management. This compulsory monitoring programme provides an ex- observations. This “trial-and-error” method is a subjective procedure
cellent opportunity for enriching sponge city data by capturing a vast that relies heavily on the experience and judgement of practitioners
amount of spatial field observations made at different locations within (James and Kuch, 1998; Shamsi and Koran, 2017; Zhang et al., 2017).
a sponge city. The investment of such a compulsory monitoring pro- Moreover, there is no guarantee that the parameter values are the
gramme in each pilot city is approximately 30 million Chinese Yuan. “best-fit” and reproducible. Recently, automatic calibration tools have
To design, assess and manage LID practices, many modelling tools been reported, which benefit essentially from optimization and search
have been developed around the world, such as MOUSE, MUSIC, and algorithms (Shahed Behrouz, 2018; Shahed Behrouz et al., 2020;
SWMM (Elliott and Trowsdale, 2007). The Storm Water Management Shamsi and Koran, 2017). Combining automatic calibration tools with
Model (SWMM) can be regarded as one of the most widely recognized field observations on large-scale sponge city practices would help to re-
models for the simulation of water quantity and quality (Gironás et al., veal the structured uncertainty of the water quality modelling of LID
2010; Obropta and Kardos, 2007). SWMM has been used for urban infrastructures and guide drainage system design (Li et al., 2018).
drainage flooding analysis and water quality simulation. It has also The primary objective of the investigation described in this paper is
been used for quantifying the impacts of different LID options, sewer to quantitatively assess the robustness of water quality characterization
systems, land use and climate change using hypothetical scenarios with LID infrastructure modelling from natural storm event-based field
(Niazi et al., 2017). In China, SWMM has been widely used in sponge monitoring. This paper tries to assess this robustness using the calibra-
city planning and design (Guo et al., 2019; Jia et al., 2017). tion performance of SWMM models. The robustness characteristics of
For urban water management, the SWMM integrated stormwater stormwater quality modelling for different LID and non-LID infrastruc-
modelling tools provide reliable results in modelling hydrologic pro- tures and different water quality variables are expected to be revealed.
cesses at different spatial and temporal scales. The SWMM spatial The effects of rainfall features and pollutant concentrations on model-
scale encapsulates large catchments that can extend hundreds of square ling robustness are analysed. This research could support the post-
 S. Tang et al. / Science of the Total Environment 753 (2021) 142007 3

assessment of sponge city construction and identify the challenges for mainly monitors the precipitation and flow rate of storm runoff
follow-up design and planning. flowing through typical infrastructures. The water quantity is re-
corded by the level gauges located at the outlets of infrastructures,
2. Materials and methods while water quality measurements rely on manual sampling and lab
analysis. The sampling usually begins with surface runoff, and 12
2.1. Study area and field monitoring campaign samples are collected at variable intervals. At the same time, a
rain gauge keeps a record of the precipitation per minute. The mon-
The Fenghuang Cheng region (means Phoenix City in Chinese), itoring campaign has been continued for more than 1 year since July
located in Guangming New District of Shenzhen city, Southeast 2018. In total, more than one thousand water quality samples were
China (Fig. 1), was selected as the study area in this study. Shenzhen collected, analysed and used in this study. Fig. S1 shows some
is one of the 30 national demonstration cities for sponge city. More photos of field sampling. All the sampling process and laboratory
than 20 community-scale LID construction projects have been analysis, implemented by Shenzhen Howay Technology Co., Ltd.,
implemented in this area since 2016. It has an area of approximately are under quality control by the local government according to na-
24.65 km2, with a mean annual precipitation of 1654.2 mm. Accord- tional standards.
ing to the Köppen-Geiger climate classification system, it has a We mainly focus on five water quality variables, COD (chemical
temperate oceanic climate. In fact, Guangming New District has con- oxygen demand), NH3-N (ammonia nitrogen), TN (total nitrogen),
ducted a LID pilot project since 2008, eight years before Shenzhen TP (total phosphorus) and SS (suspended solids), due to their large
city was enrolled in the pilot cities. composition in stormwater runoff, major impacts on biogeochemical
In this study, we selected 13 types of LID and non-LID infrastruc- cycles and national regulation of construction impact assessment
tures, including permeable pavements, sunken greenbelts, ecologi- (Cederkvist et al., 2017; Ekanayake et al., 2019; Gavrić et al., 2019).
cal tree pools, green roofs, rain gardens, ordinary roads, ordinary Samples were not filtered for the accurate measurements of SS, and
roofs, ordinary pavements, lawns, barren areas, grassy swales and nutrients that are partly attached to the particles (Vaze and Chiew,
squares, as well as a part of Xincheng Park (consisting of 6 sub- 2004). Analyses follow the methods required by the national stan-
catchments), as shown in Fig. 1. Up to 155 automatic monitoring dards of the P.R.C. (Environmental quality standards for surface
devices are distributed to cover the demonstration area, which water, GB3838-2002).

a) d)

b) e)

f)
c)

Fig. 1. The location of the study area in Shenzhen, China, and the location of infrastructures: a) ordinary pavement and monitoring device; b) grassy swale and monitoring device;
c) ordinary road and monitoring device; d) ecological tree pool and monitoring device; e) rain garden; and f) green roof.
4 S. Tang et al. / Science of the Total Environment 753 (2021) 142007

2.2. Setup of the SWMM models and calibrated parameters Besides, DE is powerful in situations in which the objective function is
stochastic, noisy, or difficult to differentiate (such as the SWMM output
SWMM models of individual infrastructures were set up based on in this study) (Mullen et al., 2011).
the facilities and associated upstream catchments. The modified Here, a package named “swmmr” based on R developed by Leutnant
Horton method was used to model the infiltration. In the water qual- et al. (2019) was used to automatically calibrate the SWMM models. A
ity module, the exponential model was used for the pollutant wash- set of functions in this package is provided to read/rewrite the input
off processes since they are representative and commonly used. Gen- files, run the SWMM models, and read/present the output files in the
erally, the build-up process is vital for water quality modelling, and R environment. The DE algorithm was realized by utilizing another R
the complex relationship between build-up and wash-off should be package called “DEoptim” (Mullen et al., 2011).
carefully considered beyond the simplified numerical model. How- For simplicity, the autocalibration optimizes the hydrologic and
ever, we did not enable the accumulation module in view of the water quality model parameters simultaneously. In addition, since this
short simulation duration, which make the build-up process contrib- work mainly focuses on the calibration performance, we did not con-
ute few to the pollutant amount. The simulation was carried out on duct validation after calibration. In early attempts, we managed to cali-
an event scale, namely, a single rainfall process. In addition, we ig- brate the parameters manually. For water quantity data, the simulation
nored the street-sweeping process (the typical sweeping frequency can fit monitoring data very well. For water quality data, the gap be-
in Shenzhen is 1/day) and the decay of pollutants in view of the tween simulation and observation are hard to fill which amplified un-
short time span. Here, the SWMM LID module was also used to certainties. Overall, the manual calibration is time-consuming and
setup the hydrological process. Based on field investigation to LID difficult to obtain acceptable fitness (see Fig. S3). In this work, the fol-
practices and relevant research (Mangangka et al., 2016), for the eco- lowing sections mainly emphasize the concept of ‘robustness’ from
logical tree pool, green roofs and permeable pavement, underdrains, auto-calibration, and the physical insights of the water quality part.
following infiltration and percolation, dominate the hydrological
process, while surface overflow plays a vital role in grassy swales, 2.4. Robustness analysis of SWMM-based LID infrastructure modelling
rain gardens, and sunken greenbelts. A typical example of the
SWMM structure of LID infrastructure can refer to Fig. S2, Table S1, 2.4.1. Concept of ‘robustness’ based on ‘goodness of fit’ index
and relevant statement. We use the term ‘robustness’ here to describe how the stormwater
Based on previous studies, a list of SWMM parameters was se- model performs under a large variability of practical conditions, such
lected for calibration (Li et al., 2014; Sun et al., 2014; Vanrolleghem as the rainfall features and concentration levels of pollutants. In this
et al., 2015). Their ranges are given in Table 1. The parameters, % work, we tried to evaluate such robustness by the goodness of fit be-
Zero-imperv, Width, %Slope, N-Perv, N-Imperv, Dstore-Perv, Dstore- tween the model and observation data, especially the field data. Here
Imperv, Max.Infil. Rate and Decay Constant, are considered to be the we emphasize that the goodness of fit should come from the automatic
most sensitive parameters in the simulation of hydrological pro- calibration rather than manual calibration to avoid the man-made sub-
cesses, although their sensitivities vary with specific rainfall scenar- jectivism. The proposed robustness concept is a new measurement of
ios. For water quality parameters, the sensitivities are much more uncertainty, which differs from traditional Monte Carlo-based struc-
complex than those of the hydrological parameters. Sensitivity is re- tured analysis (Aronica et al., 2005; Freni et al., 2009; Sun et al., 2014;
lated to rain intensity, underlying surface type and model inputs. The Wijesiri et al., 2016a).
parameters Coefficient, Exponent, Cleaning Effic., and BMP Effic. are Rainfall features, including hourly intensity, rainfall patterns, and
thought to have a greater impact on the simulation of the water qual- number of antecedent dry days (ADD) prior to storm events, impact
ity process (Niazi et al., 2017). Because the time span in this study is the flushing process and the accumulation of nonpoint source loading
the duration of a single rainfall event (approximately several hours), (Zeng et al., 2019). Moreover, the pollutant mean concentration in
the Initial Buildup and Rain Concen parameters were selected and urban runoff is significant for nonpoint pollution management, and
calibrated. The parameters of build-up process related to the accu- partly reflects the accumulation and decomposition of pollutants in
mulation of pollutants were not used in this study for the same some degree. As stated in Section 2.2, some model parameters related
reason. The value ranges of the parameters referred to the literature to the accumulation and decomposition of pollutants were ignored, as
and the official user's manual (Bonhomme and Petrucci, 2017; Niazi they do not play an important role in the simulation of a single rainfall
et al., 2017; Rossman, 2010). How these parameters work in SWMM event. Hence, we can consider that the water quality processes of the
refer to the governing equations in SI (Eqs. S1–S4) and official refer- five given variables are consistent and can be described by the same
ence manuals. model formulation. Multiple linear regression and k-means clustering
were used to capture the potential impact on model robustness. Here,
2.3. Optimization objectives, algorithm and auto-calibration tool the NSE values from calibration indicate the robustness of the water
quality models.
The Nash-Sutcliffe efficiency coefficient (NSE) of water quality vari-
ables, as the objective function, was used to evaluate the model perfor- 2.4.2. Rainfall features
mance. It can be calculated as follows (Nash and Sutcliffe, 1970): Rainfall intensity can be easily calculated with rainfall data. By
analysing many rainfall events, Bogomazova from the USSR divided
T
2
them into seven types of hyetographs as shown in Fig. S4 (Molokov
∑t¼1 Cto −Ctm
NSE ¼ 1−  2 : ð1Þ and Shtigorin, 1956). The rainfall patterns here were determined by
T
∑t¼1 Cto −C o matching actual rainfall events and the seven established rainfall
modes (Molokov and Shtigorin, 1956). In addition, we collected
where Co is the observation data, Cm is the simulation data, Ct represents historical meteorological records from the Shenzhen Meteorological
a certain value of time i, and C o represents the average of the observa- Data System belonging to the Meteorological Bureau of Shenzhen Mu-
tion data. The range of NSE is negative infinity to 1, and the closer to 1 nicipality. Then, the number of antecedent dry days (ADD) for each
the value is, the better the model and the higher the reliability. In this event was extracted after matching the record with the simulation.
study, an evolutionary algorithm named differential evolution (DE)
(Storn and Price, 1997) was selected to carry out the optimization. DE 2.4.3. Pollutant concentration
is similar to the popular genetic algorithm (GA), which generally con- Event mean concentration (EMC) is frequently used to charac-
sists of 4 steps, i.e., initialization, mutation, crossover, and selection. terize stormwater loadings and can be multiplied by the runoff
 S. Tang et al. / Science of the Total Environment 753 (2021) 142007 5

Table 1
Parameters to be calibrated and default ranges.

Para.Type Para. Abbr. Range Unit Description

Hydrology %Zero-imperv (0,100) % Percent of impervious area with no depression storage

Width (0,100) m Width of overland flow path
% Slope (0,50) % Average surface slope
N-Imperv (0,0.5) m/s1/3 Manning's n for impervious area
N-Perv (0,0.5) m/s1/3 Manning's n for pervious area
Dstore-Imperv (0,100) mm Depth of depression storage in impervious area
Dstore-Perv (0,100) mm Depth of depression storage in pervious area
Max.Infil.Rate (50,100) mm/h Maximum rate on the Horton infiltration curve
Min.Infil.Rate (0,50) mm/h Minimum rate on the Horton infiltration curve
Decay Constant (0,10) 1/h Decay constant for the Horton infiltration curve
Water quality Coefficient (0,0.05) Wash-off coefficient
Exponent (0,2) Runoff exponent in wash-off function
Cleaning Effic. (0,100) % Street cleaning removal efficiency for the pollutant
BMP Effic. (0,100) % Removal efficiency associated with any best management practice utilized
Initial Buildup (0,100) kg/ha Initial pollutant build-up on catchment
Rain Concen. (0,10) mg/L Concentration of pollutant in rainwater

volume to estimate the mass discharge (Sun et al., 2015). In this measure norm indicating the distance between data points and their re-
study, the classical EMC was used to characterize pollutant loading spective cluster centroids. This is calculated according to:
as follows:
1X
Ni

RT Pn mi ¼ Q ; i ¼ 1; 2; 3; …; k ð5Þ
M C ðt ÞQ ðt Þdt CiV i Ni j¼1 ij
EMC ¼ C ¼ ¼ 0R T ≈ Pi¼1
n ð2Þ
V Q ðt Þdt i¼1 V i
0

where M is the total pollutant loading, V is the total runoff volume, C(t) Table 2
NSE values of SWMM calibration results.
is the distribution of the pollutant concentration with time, Q(t) is the
distribution of the runoff volume with time, T is the total duration of LID type IDa COD NH3-N TP TN SS Mean Median
runoff, n is the number of time segments, Ci is the pollutant concentra- Ecological tree STSC-1-0412 0.90 0.18 0.38 0.81 0.62 0.67 0.79
tion at time i, and Vi is the runoff volume in time segment i. As a flow- pool STSC-1-0416 0.89 0.45 0.81
weighted average concentration in the whole rainfall-runoff process, STSC-1-1125 0.58 0.79 0.73 0.78 0.87
EMC is defined as the total pollution load mass divided by the total run- STSC-1-1127 0.94 0.30 0.12 0.91 0.95
Grassy swale C1-1-0218 0.99 0.56 0.61 0.85 0.99 0.74 0.77
off volume, which can be accurately described by the integral in Eq. (2).
C1-1-0612 0.89 0.92 0.85 0.69
However, limited by the practical conditions, we can only approximate C1-2-0412 0.46 0.61 0.81 0.72 0.42
the integral with a weighted sum of finite samples. Green roof LWD-1-0307 0.02 0.10 0.19 0.01 0.66 0.27 0.19
LWD-1-0412 0.21 0.35 0.19 0.45 0.63
LWD-3-0412 0.17 0.03 0.08 0.26 0.68
2.4.4. Multiple linear regression
Permeable TSPZ-2-0412 0.85 0.57 0.34 0.65 0.21 0.60 0.62
Multiple linear regression (MLR) was used to examine the relation- pavement TSPZ-3-0611 0.57 0.49 0.26 0.71
ship between the model robustness and rainfall features. MLR models TSPZ-3-1125 0.70 0.61 0.62 0.82 0.93
are as follows. Rain garden YSHY-1-0218 0.86 0.77 0.50 0.98 0.95 0.81 0.86
Sunken L1-2-1125 0.92 0.90 0.90 0.73 0.86 0.90
n
y ¼ β0 þ ∑i¼1 βi xi þ ε ð3Þ greenbelt
Barren area LD-2-0611 0.14 0.60 0.58 0.24 0.39 0.41
Lawn CD-1-0611 0.76 0.82 0.19 0.59 0.76
where y represents the dependent variable, β0 is the intercept, xi is the
Impervious PZ-1-0218 0.90 0.61 0.41 0.66 0.64 0.72 0.82
independent variable, βi is the corresponding coefficient, ε represents pavement PZ-1-0416 0.43 0.90 0.84 0.88
the error term, and n is the number of independent variables. PZ-1-0427 0.44 0.45 0.99 0.91 0.97
PZ-2-0611 0.80 0.96 0.14 1.00
2.4.5. K-means clustering Road LM-1-0223 0.72 0.44 0.75 0.13 0.84 0.60 0.64
LM-1-0218 0.93 0.58 0.42 0.74 0.70
K-means was used to identify potential patterns of robustness. It was LM-1-0307 0.62 0.53 0.41 0.08 0.49
originally created in 1967 to process the problem of clustering data LM-3-0303 0.80 0.66 0.29 0.91 0.87
(MacQueen, 1967). Currently, it is still a well-known clustering analysis Roof WM-1-0218 0.59 0.66 0.91 0.72 0.69 0.62 0.67
method that is widely adopted in scientific fields due to its ease of im- WM-2-1225 0.46 0.98 0.64 0.84 0.88
WM-2-0307 0.67 0.20 0.25 0.43 0.43
plementation, simplicity and high efficiency in application (Han et al.,
WM-3-0412 0.76 0.73 0.19 0.80
2011). This algorithm is a process that splits the data into k clusters Square GC-1-0611 0.68 0.44 0.48 0.78 0.48 0.48
(in this study, we set k as 3) according to a distance function after having GC-1-0612 0.60 0.39 0.52 0.29
determined the dataset and random assignment of clusters to each data GC-2-0611 0.49 0.50 0.22
point. Second, data are rearranged within the clusters by assigning them GC-2-0612 0.45 0.23 0.81 0.28
Xincheng XC-1-0612 0.36 0.12 0.88 0.41 0.36
to the nearest cluster centroid. This is achieved by minimizing the objec- Parkb XC-3-0612 0.34 0.56 0.25 0.49 0.42
tive function as follows: XC-4-0612 0.25 0.73 0.22 0.29 0.41
XC-6-0612 0.11 0.54 0.32 0.61
X
k X
Ni  
J¼ Q ij −mij 2 ð4Þ
a
Note: the letters in the ID column denote the type of LID infrastructure (combined by
i¼1 j¼1 the first letter of Chinese Pinyin of infrastructure name), the digit in the middle denotes
the location relative to given infrastructure, and the last four digits denote the month
and date of the rainfall event. Twenty-four sites in total are included.
where Ni means that there are Ni data in the cluster centroid mi, Qij b
In Xincheng Park, the four catchments (XC-1, 2, 3, 6) were mainly covered by green
means the jth data in cluster centroid mi and ‖Qij − mij‖2 is the distance land; therefore, in the SWMM model, the averaged underlying surface was used.
6 S. Tang et al. / Science of the Total Environment 753 (2021) 142007

where clustering terminates when the predefined maximum number of in Table 2. Twenty-eight percent of NSE values are greater than 0.8, 61%
iterations is reached. In addition, the pollutant concentration data were are greater than 0.5, and only 11% are less than 0.2. The maximum NSE
log-normalized before k-means. value can be up to 1.00, and even the minimum is not less than 0. Figs. 2
and 3 show typical examples of calibration results with good and poor
3. Results model performance, respectively.
Table S2 and Fig. 4a further show the basic statistics that suggest the
Based on relevant references and tests, we determined the parame- model performance from different types of infrastructures on different
ters of DE algorithms: the number of population members (NP) is water quality variables (see Section 2.4). The relatively high median
recommended to be 160, 10 times the length of the vector (i.e. the hy- and mean NSE values show the good fitness of the models. The averaged
drological and water quality parameters); the crossover probability NSE of LID infrastructures is 0.60, which is slightly better than 0.56 from
(CR) is set to be 0.5; the differential weighting factor (F) is 0.8; the max- non-LID infrastructures. The model performance of sunken greenbelt
imum iteration (population generation) is 500. The automatic calibra- and rain garden areas was better than that of the other models, while
tion framework performs well under this parameter set. A typical the models of green roofs did not perform well, with an average NSE
convergence curve is as shown as Fig. S5, in which most of the genera- of 0.27. In addition, the mean value of NSE in Xincheng Park is 0.41,
tions stay on a plateau and sharp increase of NSE occurs in only several which is not outstanding.
iterations. Though the nature of SWMM make the curve unsmooth, 500 From the perspective of water quality variables, the simulation of SS
population generations is great enough to get the potential global opti- with a mean of 0.64 and a median of 0.68 is better than that of others.
mum. Total optimization duration is 2 h 4 min 49 s, namely about 15 s The TP shows the worst modelling performance, with a mean of 0.48
per generation. We used RStudio to run .r script in the environment of and a median of 0.41.
Window 10 with an Intel Core i5-7400, 3.00GHz CPU and an 8GB mem-
ory. Multithreading computing enables us to call all the 4 cores of pro- 3.2. Impacts of rainfall features on modelling robustness
cessor for optimization.
Fig. 4c shows the NSE statistics under several rainfall types (see
3.1. Modelling robustness of different LID infrastructure and water quality Fig. S4). The model robustness of types 1, 3, and 5 is slightly better
variables than that of types 2 and 7. Only one rainfall event was recognized as
type 4 and one event as type 6, which resulted in the abnormal boxplots.
A total of 166 NSE values (from 5 water quality variables in 37 rain- The boxplots for single-peak mode (i.e., types 1, 2, and 3) and double-
fall events) are generated during the calibration investigation, as shown peak mode (i.e., types 4, 5, and 6) do not show a significant difference,

Fig. 2. Some of the positive results from model auto-calibration show good model performance (NSE > 0.8). Green dots represent the monitoring data. Orange lines represent the
simulation data. Blue straight lines represent the rainfall every minute.
 S. Tang et al. / Science of the Total Environment 753 (2021) 142007 7

Fig. 3. Some of the negative results from model auto-calibration show poor model performance (NSE < 0.2). Green dots represent the monitoring data. Orange lines represent the
simulation data. Blue straight lines represent the rainfall every minute.

consistent with the results from the analysis of variance. Using F-test, dynamic due to its unique concern of particle size distribution
the computed f value is 1.80, less than the corresponding critical 2.74 (Wijesiri et al., 2016a), which may result in the abnormal robustness
under a 90% confidence level. indicator of SS. The difference in transport between the dissolved and
To analyse the potential robustness pattern among different rainfall particulate phases of pollutants has been demonstrated (Sansalone
intensities, 37 rainfall events were divided into 3 groups according to av- et al., 1998) and suggests that the total mass of pollutants is dominated
erage rain intensity, including light (<2 mm/h), moderate (2– 4 mm/h), by the dissolved part within this study.
and heavy (>4 mm/h) rain. Here, we refer to standards recommended As the ADD data are not sufficiently differentiated (Fig. S6), we com-
by the United States Geological Survey and China Meteorological Ad- bined the ADD with rainfall intensity to carry out the MLR analysis. The
ministration, in view of the lack of a general classification method for results shown in Fig. 6 and Table S3 indicate that the introduction of
rainfall intensity. Fig. 4d shows that the averaged NSEs of the 3 groups ADD enhances the interpretability of the rainfall features to NSE, espe-
vary from 0.49 to 0.61 to 0.68, and the SD falls from 0.28 to 0.26 and cially for SS. The MLR models of COD, TN, and SS show identical patterns
then to 0.20, which means that the model is more stable and reliable in which the NSE increases with the ADD and rainfall intensity. Fig. 6b, c
under moderate and heavy rain. suggests that the model robustness of NH3-N and TP is inversely related
Fig. 5 depicts how model performance changes under different rain- to ADD. However, the significance test between only ADD and NSE did
fall intensities. Generally, the linear relationship between NSE and rain- not support the trend. The greater ADD values mean a longer period
fall intensity is not as good as expected. However, the averaged model for the build-up process of pollutants and a considerable initial stock
performance significantly improves with the increase in rainfall inten- for wash-off, which may relatively reduce the uncertainty from other
sity, where the correlation coefficient is up to 0.44, p-value<0.01 physical properties in the wash-off process. All the MLR models are
(Fig. 5f). The model performance of TP and NH3-N also significantly in- acceptable to some degree according to the significance test, especially
creased with rainfall intensity, with a p-value<0.05 (Fig. 5b&c). Al- for COD and TN with a p value less than 0.05. In addition, the MLR anal-
though the correlation coefficient R values relative to TN and COD are ysis also re-examined and confirmed the relationship between rainfall
not very convincing, the significance tests have p-values less than 0.1, intensity and NSE.
which would confirm the existence of the correlations to some degree.
In addition, the significance test for SS does not suggest a correlation 3.3. Impacts of pollutant concentration on modelling robustness
due to the high p-value. The higher rainfall intensity is thought to be
more powerful in driving the wash-off process of pollutants by elimi- As shown in Fig. 4b, the EMCs of COD and SS increased up to
nating uncertainty from the surface physical properties. However, the 175.73 mg/L. The concentrations of NH3-N and TP are mostly on the
wash-off process of the most important SS was confirmed to be more order of 0.1 mg/L. The small catchments and short duration of sampling
8 S. Tang et al. / Science of the Total Environment 753 (2021) 142007

Fig. 4. a) Boxplots of NSE for different water quality variables and facility types; b) boxplots for pollutant concentrations; c) boxplots of NSE with different rainfall types; and d) scatterplots
of rainfall intensity and NSE and boxplots of NSE under different rainfall intensity groups. Boxplots show the median, first and third quartiles, and whiskers extend to the furthest data point
that is within 1.5 times the interquartile range; triangles give the 95% confidence intervals for the medians; the crosses are outliers.

result in a large number of outliers. Fig. 7a shows the distribution of Although, the EMC is universally used and considered capable of
EMC. In logarithmic coordinates, the cumulative distribution function characterizing the entire hydrological process. The SMC (site mean con-
(CDF) is roughly a straight line, indicating that the EMC values follow centration) was also used for the same analysis in this study, and the
a power law distribution instead of a normal distribution. Half of the results (see Figs. S7 & S8) show similar patterns.
values are not greater than 1.84 mg/L, and only approximately one
quarter are greater than 10 mg/L. 4. Discussion
The k-means clustering results shown in Fig. 7b indicate the poten-
tial relationship of model robustness and EMC within a certain range. 4.1. Impacts on LID design and management
The 3 clusters are approximately divided by the EMC, namely, each clus-
ter represents a range of EMC. Cluster 1 almost consists of dots of TP and This study reveals the feasibility of using the popular SWMM model
NH3-N with relatively low EMC. The majority in Cluster 2 is TN, which is on runoff pollutant control for sponge city design and construction. The
roughly located in the range of 1–10 mg/L. SS and COD make up Cluster robustness of SS, COD, and TN control from the design of drainage infra-
3 with EMC greater than 10 mg/L. Linear regression analysis was used structures is relatively more stable than that of NH3-N and TP. Sunken
for each cluster, but only in Cluster 1 was the relationship between greenbelts, rain gardens, grassy swales, and ecological tree pools have
NSE and EMC identified as significant (R = 0.32, p < 0.05). less uncertainty in LID design, probably due to better water storage ca-
The dots of Clusters 2 and 3 are distributed in the area with a concen- pacity fitting the SWMM module (Table 2). The green roof has the worst
tration greater than 1, and Cluster 2 with a relatively low average NSE is calibration performance, but the traditional roof has satisfactory calibra-
the smallest group. Therefore, we believe that SWMM is generally reli- tion performance. This result may be attributed to the slight pollutant
able enough to simulate water quality, but once the concentration load from atmospheric deposition in Shenzhen (Berndtsson, 2010;
range is reduced below 1 mg/L, the model performance will also rapidly Rowe, 2011). The pollutant abatement of green roofs in Shenzhen is
decline to an unusable level. The concept of uncertainty shall be a pow- not significant. In fact, there are two stormwater reuse projects under-
erful tool to explain this pattern. When the pollutant concentration is as way in the study area based on the green roof system. Their design
low as the resolution limit of the wash-off model, it is difficult to capture can neglect nutrient pollution.
the water quality process that is likely hidden by other uncertainties. As global warming intensifies in the future, the frequency of extreme
Therefore, the increase in concentration, in a certain range, can effec- rainfall events will continue to increase (Westra et al., 2014). This trend
tively make the potential process appear from the noise caused by has been shown around the world in many countries and regions, e.g., in
other influence factors. the USA (Armal et al., 2018), Asia (Roxy et al., 2017), and Europe (Larsen
 S. Tang et al. / Science of the Total Environment 753 (2021) 142007 9

Fig. 5. Scatterplots for correlation analysis between NSE and rainfall intensity. Orange lines present the linear regression. Regression equations, correlation coefficients, and p-values are
given in the lower right corner of each small chart. Here, NSE values of 5 water quality variables are averaged for a single rainfall event to complete chart a.

et al., 2009). Especially in China, the South China region where our study (2020) recently reported a successful effort to revise SWMM for more ac-
area is located is thought to play an important role in the increase in curate prediction of LID water quality. The other solution is to develop an
heavy rain and large rainfall amounts in China (Chen, 2013). Mean- integrated urban water model (Bach et al., 2014) by connecting the dif-
while, the distribution of ADD will also become more extreme. Hence, ferent hydrodynamic processes of stormwater runoff.
given the widespread use of SWMM in urban stormwater drainage de-
sign, we need to re-examine the design plans when considering the 4.2. Study limitations and future works
resilience of green infrastructures facing future climate challenges. On
the other hand, extreme stormwater events are more threatening to The stormwater sampling and laboratory analysis in the monitoring
us because of flood risks than because of runoff pollutants. The potential campaign were all implemented by the same staff from the same orga-
poor NSE performance at low levels of pollutant concentrations needs to nization with strict quality control. The data quality is hence relatively
be considered in sophisticated management. The design effects will be reliable. Moreover, it can be assumed that uncertainties related to
more stable in areas with higher pollutant concentrations where the input data could be ignored for output uncertainty (Lee et al., 2005).
LID practices are more meaningful. Though the low R values for correla- For the variability in build-up and wash-off processes, we should
tion analysis may undermine the investigation from statistical aspect, make it clear that the washed-off amount of pollutants is only a partial
more researches proved the low level of correlation are still meaningful indication of the build-up. This is valid for build-up events that precede
and acceptable in some degree, especially in environmental and ecolog- and follow the current wash-off events. Generally, this relationship is far
ical field where uncertainty and variability are hard to describe (Fan too complex to ignore. In fact, this ignorance is, however, quite common
et al., 2020). Certainly, more decisive conclusion to be reported is relied in most studies that only focus on designing or proposing new pollution
on the analysis of more factors and powerful statistics support. control measures. The key issue here in the designing and performance
In addition, the modification of the water quality module in SWMM is assessment of stormwater treatment measures is that the tools
necessary to improve water quality predictions. It is also the common (models) do not adequately account for the natural variability in pro-
shared disadvantage of other widely used LID modelling tools. One solu- cesses such as build-up and wash-off. This issue can get worsened
tion is to update the current open-source models; for example, Baek et al. when relied on malformed models and then further ignore certain
10 S. Tang et al. / Science of the Total Environment 753 (2021) 142007

Fig. 6. MLR of model robustness related to rainfall intensity and ADD. The planes were drawn by the MLR models, which are from the data points represented by the small magenta circles.
The X axes represent the ADD, the Y axes represent rainfall intensity, and the Z axes represent the NSE. The ADD and rainfall intensity data were all min-max normalized to the 0–1 range.

conditions. In view of the model's features and objective fact of short 2005); 3) model calibration and validation with more comprehensive
simulation duration, we did not work on the build-up process in this approaches; and 4) case studies on urban renewal areas.
study. And further research should focus on how it will affect the overall Several issues are worthy of further discussion. For example, could
robustness assessment. the different structures and materials of the same facility affect the
Except for features of rainfall and pollutant concentration, more robustness? Do the geographical characteristics of different regions af-
factors can be considered in future studies, including 1) the impacts fect the robustness of the design of stormwater pollution control?
from land use in each catchment and the seasonal variations of land Conducting a big data analysis based on 30 pilot sponge city projects
cover (Afed Ullah et al., 2018; Berndtsson, 2010); 2) the uncertainty is very promising. In addition, since this study only considered indepen-
associated with the temporal resolution of rainfall (Aronica et al., dent LID infrastructures with averaged parameters of the runoff

Fig. 7. Cumulative distribution of the EMC and patterns of joint distribution of the EMC and NSE. a) Cumulative distribution of EMC referring to 5 water quality variables. Dotted lines
indicate the 3 quartiles. b) Scatterplot of the EMC and NSE. A total of 166 dots are grouped into 3 clusters by k-means. The dotted line indicates the linear regression of Cluster 1. It
should be noted that the pollutant concentration data were transformed by log normalization before k-means clustering.
 S. Tang et al. / Science of the Total Environment 753 (2021) 142007 11

processes, what will be the robustness of a catchment-scale design Appendix A. Supplementary data
coupled with different LID infrastructure? In situ monitoring based on
natural rainfall events is irreplaceable, but the experimental design Supplementary materials include 2 table, 6 figures, and 4 equations.
can hardly be controlled. Smart integration of natural rainfall-based The details of each are as follows: photographs of sampling sites
analysis and rainfall simulator tests probably provides a good solution (Fig. S1); model structure of sunken greenbelt (Fig. S2); design param-
(Wijesiri et al., 2016a). These abovementioned considerations also sug- eters of sunken greenbelt (Table S1); governing equations of SWMM
gest directions for future research. (Eq. S1–S4); statistics for NSE values (Table S2); manual calibration re-
sults (Fig. S3); hyetograph types of rainfall (Fig. S4); convergence curve
5. Conclusion of NSE (Fig. S5); scatterplot of ADD vs. NSE (Fig. S6); statistics for MLR
models (Table S3); statistics for site mean concentration (SMC) of pol-
In this study, a new robustness analysis approach for evaluating the lutants (Fig. S7); cumulative distribution for the concentrations of all
pollutant mitigation performance of LID infrastructure was proposed the pollutants (Fig. S8a), and scatterplot of concentration and NSE
based on subjective automatic calibration and natural event-based (Fig. S8b). Supplementary data to this article can be found online at
field monitoring. In a national sponge city demonstration area in https://doi.org/10.1016/j.scitotenv.2020.142007.
Shenzhen, China, 166 calibrations of 24 SWMM models under 37 rain-
fall events were analysed. Some key findings in the following will References
guide the design and re-evaluation practices.
Afed Ullah, K., Jiang, J., Wang, P., 2018. Land use impacts on surface water quality by sta-
(1) The robustness indicator, i.e., the NSE during the water quality tistical approaches. Global Journal of Environmental Science and Management 4 (2),
231–250.
process mainly ranged from 0.4 to 0.8. Water quality modelling Alfredo, K., Montalto, F., Goldstein, A., 2010. Observed and modeled performances of pro-
of greenbelts and rain gardens has the best performance, while totype green roof test plots subjected to simulated low-and high-intensity precipita-
the models of barren land and green roofs are not as robust as tions in a laboratory experiment. J. Hydrol. Eng. 15 (6), 444–457.
Armal, S., Devineni, N., Khanbilvardi, R., 2018. Trends in extreme rainfall frequency in the
the other LID types. SWMM models are better at capturing SS,
contiguous United States: attribution to climate change and climate variability
TN and COD than NH3-N and TP in urban runoff. modes. J. Clim. 31 (1), 369–385.
(2) Generally, rainfall features show correlations with model ro- Aronica, G., Freni, G., Oliveri, E., 2005. Uncertainty analysis of the influence of rainfall time
bustness. Rainfall intensity and ADD are both considered ben- resolution in the modelling of urban drainage systems. Hydrol. Process. 19 (5),
1055–1071.
eficial for model robustness, which can be partly explained Ashbolt, S., Aryal, S., Petrone, K., McIntosh, B., Maheepala, S., Chowdhury, R., Gardner, T.,
physically. 2013. Can stormwater harvesting restore pre-development flows in urban catch-
(3) K-means clustering demonstrates that model robustness in- ments in South East Queensland? Water Sci. Technol. 67 (2), 446–451.
Bach, P.M., Rauch, W., Mikkelsen, P.S., McCarthy, D.T., Deletic, A., 2014. A critical review of
creases with the pollutant concentration in a certain range integrated urban water modelling – urban drainage and beyond. Environ. Model
(<1 mg/L), which is likely caused by external uncertainty as well. Softw. 54, 88–107.
(4) The robustness analysis for the water quality model leads to the Baek, S.-S., Ligaray, M., Pyo, J., Park, J.-P., Kang, J.-H., Pachepsky, Y., Chun, J.A., Cho, K.H.,
2020. A novel water quality module of the SWMM model for assessing low impact
concern of re-examining existing and prospective LID designs
development (LID) in urban watersheds. J. Hydrol. 586, 124886.
considering the different scenarios of application conditions, es- Barbosa, A.E., Fernandes, J.N., David, L.M., 2012. Key issues for sustainable urban
pecially in the context of ecological city construction and climate stormwater management. Water Res. 46 (20), 6787–6798.
change. Berndtsson, J.C., 2010. Green roof performance towards management of runoff water
quantity and quality: a review. Ecol. Eng. 36 (4), 351–360.
Beven, K., 2010. Environmental Modelling: An Uncertain Future? CRC press
Bonhomme, C., Petrucci, G., 2017. Should we trust build-up/wash-off water quality
models at the scale of urban catchments? Water Res. 108, 422–431.
CRediT authorship contribution statement
Cederkvist, K., Jensen, M.B., Holm, P.E., 2017. Method for assessment of stormwater treat-
ment facilities—synthetic road runoff addition including micro-pollutants and tracer.
Sijie Tang: Methodology, Formal analysis, Software, Writing - origi- J. Environ. Manag. 198, 107–117.
nal draft, Writing - review & editing, Visualization. Jiping Jiang: Concep- Chen, H., 2013. Projected change in extreme rainfall events in China by the end of the 21st
century using CMIP5 models. Chin. Sci. Bull. 58 (12), 1462–1472.
tualization, Methodology, Data curation, Writing - original draft, Deletic, A., Wang, H., 2019. Water pollution control for sustainable development. Engi-
Writing - review & editing, Supervision, Project administration, Funding neering 5 (5), 839–840.
acquisition. Yi Zheng: Methodology, Supervision, Writing - review Eckart, K., McPhee, Z., Bolisetti, T., 2017. Performance and implementation of low impact
development - a review. Sci. Total Environ. 607, 413–432.
& editing, Funding acquisition. Yi Hong: Methodology, Writing - review Ekanayake, D., Aryal, R., Johir, M.A.H., Loganathan, P., Bush, C., Kandasamy, J.,
& editing. Eun-Sung Chung: Conceptualization, Writing - review & Vigneswaran, S., 2019. Interrelationship among the pollutants in stormwater in an
editing. Asaad Y. Shamseldin: Conceptualization, Writing - review & urban catchment and first flush identification using UV spectroscopy. Chemosphere
233, 245–251.
editing. Yan Wei: Data curation, Validation, Investigation, Resources.
Elliott, A.H., Trowsdale, S.A., 2007. A review of models for low impact urban stormwater
Xiuheng Wang: Conceptualization, Project administration, Funding drainage. Environ. Model Softw. 22 (3), 394–405.
acquisition. Fan, D., Kong, W., Wang, F., Yue, L., Li, X., 2020. Fencing decreases microbial diversity but
increases abundance in grassland soils on the Tibetan Plateau. Land Degrad. Dev.
1–14.
Declaration of competing interest Fioretti, R., Palla, A., Lanza, L.G., Principi, P., 2010. Green roof energy and water related
performance in the Mediterranean climate. Build. Environ. 45 (8), 1890–1904.
Fletcher, T.D., Shuster, W., Hunt, W.F., Ashley, R., Butler, D., Arthur, S., Trowsdale, S.,
The authors declare that they have no known competing financial Barraud, S., Semadeni-Davies, A., Bertrand-Krajewski, J.-L., Mikkelsen, P.S., Rivard,
interests or personal relationships that could have appeared to influ- G., Uhl, M., Dagenais, D., Viklander, M., 2015. SUDS, LID, BMPs, WSUD and more –
ence the work reported in this paper. the evolution and application of terminology surrounding urban drainage. Urban
Water J. 12 (7), 525–542.
Freni, G., Mannina, G., Viviani, G., 2009. Uncertainty in urban stormwater quality model-
Acknowledgements ling: the influence of likelihood measure formulation in the GLUE methodology. Sci.
Total Environ. 408 (1), 138–145.
Gavrić, S., Leonhardt, G., Marsalek, J., Viklander, M., 2019. Processes improving urban
This work was supported by the National Natural Science Founda-
stormwater quality in grass swales and filter strips: a review of research findings.
tion of China (Grant No. 51979136), Innovation Project of Universities Sci. Total Environ. 669, 431–447.
in Guangdong Province‐Natural Science(Grant No. 2018KTSCX201), Gironás, J., Roesner, L.A., Rossman, L.A., Davis, J., 2010. A new applications manual for the
and Science, Technology and Innovation Commission of Shenzhen Mu- Storm Water Management Model (SWMM). Environ. Model Softw. 25 (6), 813–814.
Guo, X., Du, P., Zhao, D., Li, M., 2019. Modelling low impact development in watersheds
nicipality (Grant No. KQJSCX20180322152024270). We thank the help using the storm water management model. Urban Water J. 16 (2), 146–155.
of Yang Gu on raw data organization. Han, J., Pei, J., Kamber, M., 2011. Data Mining: Concepts and Techniques. Elsevier.
12 S. Tang et al. / Science of the Total Environment 753 (2021) 142007

Hong, Y., Bonhomme, C., Le, M.-H., Chebbo, G., 2016. A new approach of monitoring and Shahed Behrouz, M., 2018. Automatic Calibration of Storm Water Management Model
physically-based modelling to investigate urban wash-off process on a road catch- (SWMM) With Multi-objective Optimization. State University of New York at Buffalo.
ment near Paris. Water Res. 102, 96–108. Shahed Behrouz, M., Zhu, Z., Matott, L.S., Rabideau, A.J., 2020. A new tool for automatic
James, W., Kuch, A., 1998. Sensitivity-calibration decision-support tools for continuous calibration of the Storm Water Management Model (SWMM). J. Hydrol. 581, 124436.
SWMM modeling: a fuzzy-logic approach. Journal of Water Management Modeling Shamsi, U.M.S., Koran, J., 2017. Continuous calibration. Journal of Water Management
6, R200–R209. Modeling 25, C414.
Jia, H., Wang, Z., Zhen, X., Clar, M., Yu, S.L., 2017. China’s sponge city construction: a dis- Storn, R., Price, K., 1997. Differential evolution—a simple and efficient heuristic for global
cussion on technical approaches. Frontiers of Environmental Science & Engineering optimization over continuous spaces. J. Glob. Optim. 11 (4), 341–359.
11 (4), 18. Sun, S., Barraud, S., Castebrunet, H., Aubin, J.-B., Marmonier, P., 2015. Long-term
Larsen, A.N., Gregersen, I.B., Christensen, O., Linde, J.J., Mikkelsen, P.S., 2009. Potential fu- stormwater quantity and quality analysis using continuous measurements in a
ture increase in extreme one-hour precipitation events over Europe due to climate French urban catchment. Water Res. 85, 432–442.
change. Water Sci. Technol. 60 (9), 2205–2216. Sun, N., Hong, B., Hall, M., 2014. Assessment of the SWMM model uncertainties within the
Lee, J.-T., Hur, S.-C., Kim, T.-H., 2005. Sensitivity analysis of the SWMM model parameters generalized likelihood uncertainty estimation (GLUE) framework for a high-
based on design rainfall condition. Journal of Korea Water Resources Association 38 resolution urban sewershed. Hydrol. Process. 28 (6), 3018–3034 (n/a-n/a).
(3), 213–222. Vanrolleghem, P.A., Mannina, G., Cosenza, A., Neumann, M.B., 2015. Global sensitivity
Leutnant, D., Döring, A., Uhl, M., 2019. SWMMR - an R package to interface SWMM. Urban analysis for urban water quality modelling: terminology, convergence and compari-
Water J. 16 (1), 68–76. son of different methods. J. Hydrol. 522, 339–352.
Li, C., Wang, W., Xiong, J., Chen, P., 2014. Sensitivity analysis for urban drainage modeling Vaze, J., Chiew, F.H., 2004. Nutrient loads associated with different sediment sizes in
using mutual information. Entropy 16 (11), 5738–5752. urban stormwater and surface pollutants. J. Environ. Eng. 130 (4), 391–396.
Li, L., Chung, E.-S., Jun, K.S., 2018. Robust parameter set selection for a hydrodynamic Vrugt, J.A., Robinson, B.A., 2007. Improved evolutionary optimization from genetically
model based on multi-site calibration using multi-objective optimization and mini- adaptive multimethod search. Proc. Natl. Acad. Sci. 104 (3), 708–711.
max regret approach. Water Resour. Manag. 32 (12), 3979–3995. Vrugt, J.A., Gupta, H.V., Bouten, W., Sorooshian, S., 2003. A Shuffled Complex Evolution
Luan, B., Yin, R., Xu, P., Wang, X., Yang, X., Zhang, L., Tang, X., 2019. Evaluating Green Metropolis algorithm for optimization and uncertainty assessment of hydrologic
Stormwater Infrastructure strategies efficiencies in a rapidly urbanizing catchment model parameters. Water Resour. Res. 39 (8).
using SWMM-based TOPSIS. J. Clean. Prod. 223, 680–691. Wang, K.-H., Altunkaynak, A., 2012. Comparative case study of rainfall-runoff modeling
MacQueen, J., 1967. Some Methods for Classification and Analysis of Multivariate Obser- between SWMM and fuzzy logic approach. J. Hydrol. Eng. 17 (2), 283–291.
vations. pp. 281–297 (Oakland, CA, USA). Wang, H., Mei, C., Liu, J., Shao, W., 2018. A new strategy for integrated urban water man-
Mangangka, I.R., Liu, A., Goonetilleke, A., Egodawatta, P., 2016. Enhancing the Storm agement in China: sponge city. SCIENCE CHINA Technol. Sci. 61 (3), 317–329.
Water Treatment Performance of Constructed Wetlands and Bioretention Basins. Westra, S., Fowler, H., Evans, J., Alexander, L., Berg, P., Johnson, F., Kendon, E., Lenderink,
Springer. G., Roberts, N., 2014. Future changes to the intensity and frequency of short-
Mei, C., Liu, J., Wang, H., Yang, Z., Ding, X., Shao, W., 2018. Integrated assessments of green duration extreme rainfall. Rev. Geophys. 52 (3), 522–555.
infrastructure for flood mitigation to support robust decision-making for sponge city Wijesiri, B., Egodawatta, P., McGree, J., Goonetilleke, A., 2016a. Assessing uncertainty in
construction in an urbanized watershed. Sci. Total Environ. 639, 1394–1407. stormwater quality modelling. Water Res. 103, 10–20.
Molokov, M., Shtigorin, Γ., 1956. The Rain Water and Confluent Channel. Architectural En- Wijesiri, B., Egodawatta, P., McGree, J., Goonetilleke, A., 2016b. Understanding the uncer-
gineering Press, Beijing. tainty associated with particle-bound pollutant build-up and wash-off: a critical re-
Mullen, K., Ardia, D., Gil, D.L., Windover, D., Cline, J., 2011. DEoptim: an R package for view. Water Res. 101, 582–596.
global optimization by differential evolution. J. Stat. Softw. 40 (6), 1–26. Yang, Y.-Y., Toor, G.S., 2018. Stormwater runoff driven phosphorus transport in an urban
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models part I—a residential catchment: implications for protecting water quality in urban watersheds.
discussion of principles. J. Hydrol. 10 (3), 282–290. Sci. Rep. 8 (1), 11681.
Niazi, M., Nietch, C., Maghrebi, M., Jackson, N., Bennett, B.R., Tryby, M., Massoudieh, A., Zeng, J., Huang, G., Luo, H., Mai, Y., Wu, H., 2019. First flush of non-point source pollution
2017. Storm water management model: performance review and gap analysis. Jour- and hydrological effects of LID in a Guangzhou community. Sci. Rep. 9 (1), 13865.
nal of Sustainable Water in the Built Environment 3 (2), 04017002. Zhang, Y., Shuster, W., 2014. The comparative accuracy of two hydrologic models in sim-
Obropta, C.C., Kardos, J.S., 2007. Review of urban stormwater quality models: determinis- ulating warm-season runoff for two small, hillslope catchments. J. Am. Water Res.
tic, stochastic, and hybrid approaches 1. J. Am. Water Res. Assoc. 43 (6), 1508–1523. Assoc. 50 (2), 434–447.
Randall, M., Sun, F., Zhang, Y., Jensen, M.B., 2019. Evaluating Sponge City volume capture Zhang, J., Zhou, Y., Liu, C., Lu, X., Wang, Z., Zhao, S., 2017. Determination of water quality
ratio at the catchment scale using SWMM. J. Environ. Manag. 246, 745–757. parameters of swmm model for water quality simulation of rainfall-runoff. Environ.
Ren, N., Wang, Q., Wang, Q., Huang, H., Wang, X., 2017. Upgrading to urban water system Sci. Technol. 40 (05), 165–170.
3.0 through sponge city construction. Frontiers of Environmental Science. Engineer- Zhang, L., Lu, Q., Ding, Y., Peng, P., Yao, Y., 2018. Design and performance simulation of
ing 11 (4), 9. road bioretention media for sponge cities. J. Perform. Constr. Facil. 32 (5), 04018061.
Rossman, L.A., 2010. Storm Water Management Model User’s Manual, Version 5.0. Na- Zhang, K., Bach, P.M., Mathios, J., Dotto, C.B.S., Deletic, A., 2020. Quantifying the benefits of
tional Risk Management Research Laboratory, Office of Research and …. stormwater harvesting for pollution mitigation. Water Res. 171, 115395.
Rowe, D.B., 2011. Green roofs as a means of pollution abatement. Environ. Pollut. 159 Zheng, Y., Lin, Z., Li, H., Ge, Y., Zhang, W., Ye, Y., Wang, X., 2014. Assessing the polycyclic
(8–9), 2100–2110. aromatic hydrocarbon (PAH) pollution of urban stormwater runoff: a dynamic
Roxy, M.K., Ghosh, S., Pathak, A., Athulya, R., Mujumdar, M., Murtugudde, R., Terray, P., modeling approach. Sci. Total Environ. 481, 554–563.
Rajeevan, M., 2017. A threefold rise in widespread extreme rain events over central
India. Nat. Commun. 8 (1), 1–11.
Sansalone, J.J., Koran, J.M., Smithson, J.A., Buchberger, S.G., 1998. Physical characteristics of
urban roadway solids transported during rain events. J. Environ. Eng. 124 (5),
427–440.