Journal of Hydrology 583 (2020) 124584

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Spatial-temporal rain field generation for the Guangdong-Hong Kong-Macau T

Greater Bay Area considering climate change
Yejia Qianga,b, Limin Zhanga,b, , Te Xiaoa,b
⁎

a
Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Hong Kong, China
b
HKUST Shenzhen Research Institute, Shenzhen, China

ARTICLE INFO ABSTRACT

This manuscript was handled by G. Syme, A stochastic rainfall generator is required to provide rainfall inputs for the analysis and mitigation of such
Editor-in-Chief hydrological or geologic hazards as floods and rain-induced landslides. This paper presents a new spatial-tem-
Keywords: poral rainstorm generator for generating simultaneous rainfall processes at numerous locations considering the
Spatial-temporal rain field generator spatial correlation among these locations and interpolating the point processes into an areal rain field. The
Rainstorm generator is able to include the effect of climate change by adjusting the parameters of the marginal distributions
Climate change of variables constituting rainfall events. A case study on the Guangdong-Hong Kong-Macau Greater Bay Area
Natural hazards (GBA), one of the regions that are most prone to storm-related disasters in the world, is presented. The per-
Guangdong-Hong Kong-Macau Greater Bay Area
formance of the proposed generator is excellent in reproducing the historical statistical characteristics of re-
Flooding
gional rainfall. The model is adapted to climate change through extrapolation of the variation trend of the model
parameters in the observation period to explore possible future scenarios of regional rainfall in GBA. The si-
mulation results indicate a significant increase in rainfall extremes, especially for short-duration rainfall, at the
end of 21st century in GBA.

1. Introduction from analysis of historical daily precipitation data (Zhang et al., 2009).
Projections of heavy rainfall events from general circulation models of
The Guangdong-Hong Kong-Macau Greater Bay Area (GBA) (Fig. 1) the Fourth Assessment Report released by Intergovernmental Panel on
is a very densely populated metropolitan area consisting of 11 cities, Climate Change indicate a substantial increase in heavy rainfall fre-
including Hong Kong, Macau, Guangzhou and Shenzhen. It is one of the quency (35.9–50.2% increase over that in the period of 1980–1999)
regions with the strongest economic vitality in China. However, GBA is and intensity (2.8–6.4% increase over that in the period of 1980–1999)
one of the most disaster-prone regions in the world, and frequently in South China at the end of 21st century (Chen et al., 2012).
suffers storm-related hazards (Swiss Re, 2013; Hallegatte et al., 2013). Apart from the total rainfall amount, the spatial-temporal variation
According to Swiss Re (2013), GBA is the world’s most flood-exposed of rainfall will also affect the response of hydrogeological systems by
urban area, ranking first among major metropolitan areas around the affecting the soil moisture, groundwater level and discharge of catch-
world, with around 12 million people potentially affected. Intense ments (Schuurmans and Bierkens, 2007; Zhu et al., 2018). To in-
rainstorms are one of the major triggers of hydrological or geologic vestigate the hazards related to rainfall, it is necessary to project future
hazards, such as floods (Wright et al., 2014; Gao et al., 2019), landslides regional rainstorm scenarios for engineering design and hazard pre-
(Chen and Zhang, 2014; Ran et al., 2018; Gao et al., 2018b), debris vention purposes, considering the trend of climate change. Stochastic
flows (Chen et al., 2016; Iadanza et al., 2016; Shen et al., 2018; Zhou rainfall simulation models can be used to generate continuous rainfall
et al., 2019), etc. The risks of storm-related hazards are likely to become processes to provide input for hazard investigation. Stochastic simula-
higher in GBA in the future as a result of potentially more frequent and tion of spatial-temporal process of rainfall is challenging due to its in-
intense rainstorms. Due to global warming, the capacity of the atmo- termittency and the spatial-temporal cross-correlation. The inter-
sphere to hold moisture will enhance, which implies that extreme mittency characteristics of rainfall can be modelled by describing
rainstorms will increase in terms of both frequency and intensity. In rainfall occurrence and rainfall intensity separately. For rainfall oc-
South China, a remarkable increasing trend in rainfall intensity and currence, based on point rainfall observations, simple models for gen-
rainfall amount from extreme events is found in the Pearl River Basin erating point synthetic rainfall sequences have been developed, such as

⁎
Corresponding author at: Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Hong Kong, China.
E-mail addresses: yqiang@connect.ust.hk (Y. Qiang), cezhangl@ust.hk (L. Zhang), xiaote@ust.hk (T. Xiao).

https://doi.org/10.1016/j.jhydrol.2020.124584
Received 9 October 2019; Received in revised form 11 December 2019; Accepted 13 January 2020
Available online 20 January 2020
0022-1694/ © 2020 Elsevier B.V. All rights reserved.
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

Zhaoqing
Guangzhou
Huizhou

Foshan Dongguan

Zhongshan Shenzhen

Jiangmen Hong Kong

Zhuhai
Macau
Pearl river estuary

Fig. 1. The Guangdong-Hong Kong-Macau Greater Bay Area.

the Markov renewal model (Haan et al., 1976; Foufoula-Georgiou and with historical data in Section 3, and the possible future variation of
Lettenmaier, 1987) and the alternating renewal model (Acreman, 1990; rainfall characteristics in GBA will be systematically investigated in
Haberlandt, 1998; Haberlandt et al., 2008). The variation of rainfall Section 4. The spatial-temporally varying rain fields produced by the
intensity (rainfall pattern) can be modelled with such methods as the proposed model can be further applied to large-scale hydrological and
random cascade model (Tang et al., 2018) and the cluster analysis (Wu geologic hazard analyses using distributed models, such as rainfall-
et al., 2006). To capture the spatial and temporal characteristics of runoff-inundation analysis and regional rainfall-induced landslide
rainfall, multisite models are widely used when point observations are analysis, which is beyond the scope of this paper.
available. For example, Wilks (1998) established a chain dependent
process model for simultaneous simulation of daily precipitation oc- 2. Spatial-temporal rain field generation considering climate
currences and amounts at multiple locations. Serinaldi (2009) devel- change
oped a model based on bivariate copula-based mixed distributions to
generate daily rainfall time series at multiple sites with no need for The proposed spatial-temporal rain field generator generates rain-
separating rainfall occurrence and rainfall amount. In addition to the fall sequences at multiple stations considering the inter-station spatial
multisite models, some models included space-time correlation of rain correlation. The framework of the generator is shown in Fig. 2.
events with random field theory. Bell (1987) established a model to
describe the spatial pattern of precipitation by rescaling part of a cor- 2.1. Temporal rainfall modelling
related Gaussian random field into a lognormal type. Allcroft and
Glasbey (2003) transformed rainfall to a thresholded Gaussian variable The intermittency feature of rainfall can be approximated as an
and modelled the space-time dependence structure as a Gaussian alternating renewal process (Acreman, 1990; Haberlandt, 1998) which
Markov random field. models a system alternating between two states over time, i.e. raining
The models mentioned above attempt to reproduce the character- or not raining. The occurrence, interval time and rainfall amount can be
istics of point or regional precipitation. Many of them involve a large characterised by three variables, i.e. wet spell duration (W), dry spell
number of parameters and complex estimation processes. Moreover, duration (D) and average intensity (I) (Fig. 3). Herein, the wet and dry
how to take climate change into stochastic rainfall simulation models is spell durations are referred to as the duration of continuous raining (i.e.
still a challenge. To investigate rainfall-triggered hazards in GBA, which the duration of a rainstorm event) and the inter-event duration, re-
may become more severe in the future, a rain field generator with a spectively. A basic assumption of the alternating renewal process is that
logical structure and the flexibility to include climate change is parti- pairs of the two states are independent. When applying the alternating
cularly necessary. renewal process in modelling rainfall, all the samples of wet spell and
The primary objective of this paper is to develop a model for gen- dry spell durations are often considered to be independent, i.e. the
erating spatial-temporal rainfall processes as inputs for hazard analysis, neighbouring W and D are independent as well (Acreman, 1990;
with its parameters estimated from historical point observations from Haberlandt, 1998). However, the average intensity of rainfall during
rainfall stations and adjustable for the trend of climate change. Section each wet spell is not independent of the wet spell duration. The joint
2 will present the detailed development of the model, followed by distribution of W and I with the consideration of their correlation can
model parameter estimation and adaptation to climate change. The be described with a two dimensional copula (Nelsen, 2007; Bezak et al.,
performance of the model will be validated through a case study in GBA 2016; Gao et al., 2018a):

2
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

Fig. 2. Schematic diagram of the spatial-temporal rain field generator.

50 exponentially (Haberlandt et al., 2008):

W1 D1 W2 D2 W3 i (t ) = im × e | t tm |
(2)
40
where im = a × ib is the peak instantaneous rainfall intensity during the
wet spell and assumed to be only related to the average rainfall in-
Hourly rainfall (mm)

30 tensity within the wet spell in a power form, in which parameters a and
b can be estimated by regression analysis; tm is the time of peak in-
tensity that is assumed to be uniformly distributed within the wet spell
20 and generated randomly during the simulation; and γ is an event-spe-
cific parameter and can be obtained according to the balance of rainfall
amount:
10 I3 tm w
I1 I2 im × e (t tm) dt + im × e (tm t ) dt = wi
0 tm (3)
0
0 10 20 30 40 50 60 70 80 Once w, i, im and tm have been determined, Eq. (3) can be solved
through the Newton-Raphson method. Even for the same set of W and I,
Time (h)
different simulated values of tm still lead to temporally varying rainfall
Fig. 3. Illustration of a precipitation process. patterns. Combining the event sequences and the simulated rainfall
patterns, a time series of rainfall intensity reflecting the temporal var-
F (w, i) = C [FW (w ), FI (i); ] (1) iation of rainfall can be obtained.

where F(w,i) is the joint cumulative distribution function of W and I; 2.2. Rain field generation including multisite spatial correlation
FW(w) and FI(i) are the marginal cumulative distribution functions of W
and I, respectively; C is the copula function that relates F(w,i) to FW(w) The rain field for a specific area of interest can be described by the
and FI(i); and θ is the copula parameter representing the dependency. temporal rainfall processes at numerous stations. The time series of
The marginal distributions of W, D and I can be estimated from his- rainfall for each station should be generated at the same time due to the
torical point observations within the area of interest, and rainfall event connection between temporal and spatial variations of rainfall. Under
sequences can be produced by assembling sets of the three variables the assumption that the spatial dependences of W, D and I among sta-
sampled from the corresponding distributions. tions are the same (Xiao et al., 2017a; Zhu et al., 2017), the inter-station
The temporal variation of rainfall intensity within a wet spell can be spatial correlation can be described with a prescribed spatial correla-
simplified as a predefined form of rainfall pattern. Although the rainfall tion function, e.g., a Gaussian correlation function:
pattern does have influence on the occurrence and process of rainfall- = exp[ ( )2]
pq pq (4)
triggered hazards, a prescribed pattern form is easier to describe when
the focus is on a relatively long-term rainfall process. For example, the where ρpq is the correlation coefficient of the three variables between
instantaneous rainfall intensity i(t) within each wet spell duration can stations p and q; τpq is the separation distance between stations p and q;
be assumed to increase exponentially first and then decrease and δ is the spatial correlation parameter.

3
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

With the spatial correlation function, any number of simultaneous Furthermore, the spatial correlation parameter δ for modelling the
rainfall events at numerous stations can be generated. To generate inter-station spatial correlation of W, D and I is required. However, δ is
rainfall sequences at ns stations, a ns × ns auto-correlation matrix difficult to be directly estimated from observations. Instead, it can be
R = [ρpq] is formed first according to the spatial correlation function. determined by minimizing the difference of a simultaneous rainfall
Then, a ns × 1 correlated standard normal random vector N can be feature between the generated rainfall and the observed rainfall. In this
obtained using the covariance matrix decomposition method (Li et al., paper, the simultaneous rainfall feature is described by the correlation
2016; Zhu et al., 2017): coefficients of the instantaneous rainfall intensity (ρI) (approximated as
(5) the cumulative rainfall depth in a short period of time) between dif-
N = LU
ferent stations when rainfall occurs simultaneously. Given the marginal
where L is a lower triangular matrix satisfying Cholesky decomposition distributions of W, D and I as well as the dependency between W and I,
R = LLT; and U is a ns × 1 independent standard normal random the decay trend of the correlation coefficient against the separation
vector. N can be transformed into other types of random vectors ac- distance between rainfall generation stations (τ) varies as the spatial
cording to the variable distribution types. correlation parameter δ changes. For simplicity, the pairs of stations are
Note that D is independent of W and I, while W and I are cross- grouped into several lag bins according to seperation distances, and the
correlated. For one rainfall event, the ns × 1 vectors of D, W and I can difference of the decay trends between the generated and observed
be obtained as follows: rainfall is quantified by an objective index Δ:

1. Randomly generate ns × 3 potential samples of a standard uni- ( )= wj | I ( j ) I* ( j | )|

j (7)
formly distributed variable at ns stations, with no dependence
among all the samples. where τj is the mean separation distance of pairs of stations in the jth lag
2. Transform the first column of the potential samples into an in- bin; ρI(τj) is the mean value of ρI in the jth lag bin obtained from ob-
dependent standard normally distributed vector UD firstly; then servation data; ρI*(τj|δ) is the mean value of ρI in the jth lag bin ob-
transform it into a standard normally distributed vector ND with tained from generated data given a value of δ; and wj is the weight of
inter-station spatial correlation according to Eq. (5). The ns × 1 the jth lag bin to emphasize the importance of realistic reproduction of
vector of D for ns stations can be obtained by transforming ND to the the simultaneous rainfall properties at close stations, and can be taken
selected marginal distribution of D. as, for example, wj = (τmax − τj)/τmax, where τmax is the maximum
3. Transform the last two columns of the potential samples into a pair value of τj. The most plausible value of δ is the one corresponding to the
of standard uniformly distributed variables obeying a given copula minimum objective index, which will be further illustrated in Section
function with independency among stations. 3.2.
4. Similar to Step 2, transform the two vectors firstly into two standard
normally distributed vectors UW and UI with spatial independency 2.4. Adaptation to climate change
among stations, and then into two standard normally distributed
vectors NW and NI with inter-station spatial correlation using Eq. There is sufficient flexibility for the proposed model to consider the
(5). The wet spell duration series W and average rainfall intensity variations of rainfall characteristics due to climate change. For ex-
series I can then be obtained by transforming NW and NI to the se- ample, adjustments can be made to the marginal distributions of W, D
lected marginal distributions of W and I. and I for future scenarios to capture the variations of rainfall duration,
interval time and intensity, while the correlation between W and I, as
Repeatedly performing the above steps, the matrices D, W and I for well as the inter-station spatial correlation can be considered as sta-
ns stations with many realizations of rainfall events can be obtained. tionary for simplicity (Peres & Cancelliere, 2018; Lee & Singh, 2019).
After generating rainfall processes at multiple stations, the areal dis- Trend of regional climate change is usually identified based on histor-
tribution of rainfall at each snapshot can then be obtained by spatial ical data analysis or climate projections from a regional climate model.
interpolation using kriging (Oliver and Webster, 1990; Xiao et al., When long-term historical observations are available, the variation
2017b), for which the parameters of range, sill, and nugget are esti- trends of W, D and I can be explored and the parameters of the marginal
mated by fitting the empirical semivariogram with commonly used distributions can be adjusted through extrapolation. Besides, the sta-
theoretical semivariogram models. tistic moments are usually easy to obtain from rainfall projection, and
the change of parameters can be transformed from projected statistic
2.3. Estimation of model parameters moments. As the types of marginal distributions can be different for
different samples, the changing rates of the three variables can be
To generate a rain field, the first step is to select appropriate mar- evaluated with the change rates of statistical moments instead of the
ginal and joint distributions for the three variables W, D and I, and to parameters of the marginal distributions:
estimate the marginal distribution parameters and the copula para- Fn = Mn, f Mn, r (8)
meter. Several candidate parametric distributions and copula models
commonly used in hydroclimate study can be tried to fit the data where Fn is the change rate of the nth moment of a variable; Mn,f and
samples. The optimal model can be identified according to the Akaike Mn,r are the nth moment of the variable in a future period and a re-
information criterion (AIC) defined as (Tang et al., 2013; Chen and ference period, respectively. For simplicity, they are obtained through
Singh, 2018): linear regression, as demonstrated later in this paper.

AIC = 2k 2 ln L (6)
3. Rain field generation for the Greater Bay Area with historical
where k is the number of model parameters; and L is the likelihood data
function reflecting the goodness-of-fit and is maximized through max-
imum likelihood estimation. Given a set of candidate models for fitting 3.1. Study area and historical data
the data, the preferred model is the one with the minimum AIC value.
Regarding the rainfall pattern, the parameters in the power form of The proposed rain field generator is applied to GBA in South China.
peak intensity can be obtained using regression analysis based on the GBA has a humid subtropical monsoon climate characterised by clear
pairwise data of the peak and average rainfall intensities for every wet rainy and dry seasons, and the average annual rainfall is over 1300 mm.
spell. Generally, the monthly rainfall amount in GBA has two peaks annually,

4
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

50

40
Guangzhou station
30

20

10

50

40 Dongguan station
30
Rainfall intensity (mm/h)

20

10

50

40 Shenzhen station
30

20

10

50

40 Zhuhai station
30

20

10

0
1998 2003 2008 2013 2017
Year
Fig. 4. Rainfall observations in rainy seasons at 4 representative stations in GBA.

Table 1
Optimal model identification for marginal distributions and copulas.
AIC values of candidate models (×103)

Lognormal Generalized extreme value Exponential Inverse Gaussian Geometric

Dry spell duration, D 415.9 423.0 426.8 415.6 430.9

Wet spell duration, W 165.7 267.1 162.9 166.1 161.8
Average intensity, I 244.0 241.8 240.0 280.6 −

Gaussian Frank Clayton Gumbel

Copula, C −5.2 −7.2 −1.6 −2.4

Note: the model with value in bold italic is the optimal one with minimum AIC value.

Table 2
Parameter estimates for rain field generation in GBA.
Variable Description Equation Parameter Value

Dry spell duration, D Inverse Gaussian distribution μ 10.598

FD (d ) =
d ( d
µ
1) + exp ( ) 2
µ d ( d
µ )
+1
λ 3.987
Wet spell duration, W Geometric distribution FW (w ) = 1 (1 p) w p 0.548
Average intensity, I Exponential distribution FI (i) = 1 e i β 2.374
Copula, C Frank copula 1 (e u1 1)(e u2 1) θ 2.687
C (u1, u2; ) = ln 1 +
e 1
Peak intensity, im Power form im = a × i b a 1.605
b 0.934

Note: d* = d/3; w* = w/3.

5
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

(a) 0.6 (b) 0.20

0.5
0.15

Probability density

Probability density
0.4

0.3 0.10
Geometric distribution Inverse Gaussian distrbution
0.2
0.05
0.1

0.0 0.00
0 10 20 30 40 50 0 30 60 90 120 150
Wet spell duration, W (h) Dry spell duration, D (h)
(c) 0.5 (d) 25

Average intensity, I (mm/h)

0.4 20
Probability density

Frank copula
0.3 15 10-5

10-4
0.2 Exponential distribution 10
-3
0.01 10
0.1 5
0.05
0.0 0
0 4 8 12 16 20 3 10 20 30 40

Average intensity, I (mm/h) Wet spell duration, W (h)

Fig. 5. Marginal distributions for (a) wet spell duration; (b) dry spell duration; (c) average intensity; and (d) joint distribution for wet spell duration and average
intensity.

and the largest rainfall amounts occur in June, May and August. Most of
GBA is delta plain with smooth terrain, surrounded by intermittent
mountains and hills. The topography of GBA and locations of 41 rainfall
stations are shown in Fig. 1. Hourly rainfall data is recorded by these
rainfall stations and daily rainfall data is available to the public. Ad-
ditionally, high resolution rainfall data is also available from the da-
taset of NASA Tropical Rainfall Measuring Mission (Huffman et al.,
2007). The temporal and spatial resolutions of the rainfall data are 3 h
and 0.25° × 0.25°, respectively. The observation periods are the rainy
seasons (from May to August) of each year from 1998 to 2017. The
rainfall data in this paper is obtained from the NASA dataset. Ob-
servations at four representative stations (Guangzhou, Dongguan,
Shenzhen and Zhuhai) in GBA are shown in Fig. 4.

3.2. Determination of model parameters

For GBA, the marginal distributions of variables W, D and I are

chosen based on the historical data at the locations of rainfall stations.
Data from all the stations is pooled together to estimate the parameters
for the marginal distributions and the joint distribution because GBA is
a meteorological homogeneous region (Chong & Lee, 2015) and a large
number of rain stations are located on the delta plain with low eleva-
tions, which makes the effect of topography to be very low. Five can-
didate marginal distributions (i.e., lognormal, generalized extreme
value, exponential, inverse Gaussian, and geometric distributions) and
four candidate copulas (i.e., Gaussian, Frank, Clayton and Gumbel co-
pulas) are compared to fit the data in GBA. Note that W and D are
normalized by 3.0 in this study because 3-hour interval rainfall data is
used. As shown in Table 1, the geometric distribution, inverse Gaussian
distribution and exponential distribution are chosen to model the wet
spell duration, dry spell duration and average intensity, respectively;
Fig. 6. Determination of the spatial correlation parameter: (a) identification and the Frank copula is selected to model the joint distribution between
according to the minimum objective index; (b) decay trends of the correlation W and I due to their respective minimum AIC values. The corresponding
coefficient with the separation distance. model parameters, as well as the parameters for peak intensity, are
provided in Table 2. Fig. 5 illustrates that all the selected probability
models fit the data very well. With respect to the spatial correlation

6
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

(a) 50 (b)
Guangzhou station Day 1
40

30

20

10

50
Dongguan station
Rainfall intensity (mm/h)

40

30

20 Day 2
10

50

40 Shenzhen station

30

20

10

50
Day 3
40 Zhuhai station

30

20

10

0
1998 2003 2008 2013 2017
Year
Fig. 7. Example of generated rainfall: (a) one realization of rainfall generation in rainy season at four representative stations for a period of 20 years; (b) spatial
distributions of 24-hour cumulative rainfall during an intense rainstorm.

Table 3 cumulative rainfall amount during a generated intensive rainstorm are

Comparison of rainfall characteristics. shown in Fig. 7.
Rainfall characteristics Observation Simulation
Table 3 compares four characteristics of the generated and observed
rainfalls, including average number of rainfall events, average single
Range Average event rainfall, standard deviation of single event rainfall, and total
rainfall amount within 4 months. All the four characteristics are well
Average number of events 78.2 75.1–87.3 81.3
Average single event rainfall (mm) 15.0 14.3–17.3 15.6
reproduced by the proposed rainfall generator. In addition, frequency
Standard deviation of single event 23.5 21.9–28.6 24.6 analysis is conducted with the annual maximum time series of observed
rainfall (mm) and generated rainfall amounts in four common durations, i.e., 6, 12,
Total rainfall amount within 4 months 1170.7 1136.3–1405.8 1271.2 24 and 48 h. It should be noted that the generated rainfall data from all
(mm)
the stations is pooled together as a single sample in the frequency
analysis because of the short length of observation period. In fact,
simply pooling data from different stations will include some samples
parameter δ, a trial-and-error method is adopted for a series of discrete with strong dependency, i.e. the spatial correlation among stations can
values with a certain interval to minimize the objective index Δ, be- reduce the number of effective samples. This kind of effect is ignored
cause the variation trend of Δ with δ is of concern and minor differences here because the analysis is just used for validation purpose. For each
in δ have little impact on the simulation results. The objective index Δ is realization, the generalized extreme value distribution is used to fit the
calculated from the mean value of a number of simulations with nine data and the rainfall amounts corresponding to several exceedance
possible values of δ ranging from 200 to 600 km with an interval of probability values are calculated. The results are shown in Fig. 8. For
50 km, and a value of δ = 400 km corresponding to the minimum Δ is different rainfall durations, the observed values are mostly inside the
chosen (Fig. 6a). Fig. 6b demonstrates a consistent decay trend of si- 90% confidence interval of simulated values. For a short duration (i.e.,
multaneous rainfall intensity with the separation distance based on si- 6-hour), the observed values are a bit larger than the simulated results
mulated and observed data. or near the upper bound. These are understandable because the pat-
terns of rainfall events are assumed to have a similar predefined profile
3.3. Validation of rainfall characteristics (i.e., an exponentially increase to the peak value first, followed by an
exponentially decrease). The difference between the real temporal
With all the parameters determined, the rainfall processes of any variation of rainfall and the assumed profile may cause bias especially
duration can be generated within the study area. One hundred reali- for the cumulative rainfall amount in a short duration. As a whole, the
zations of 20-year rainfall (4-month for each year) are generated, and generated rainfall is able to reproduce the statistical characteristics of
an example of generated rainfall at the four representative stations and observations, including extreme events.
the variation of the interpolated spatial distribution of 24-hour

7
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

(a) 1000 (b) 1000

6-hour Simulation 12-hour Simulation
800 Observation 800 Observation

Rainfall amount (mm)

Rainfall amount (mm)
600 600

400 400

200 200

0 0
1E-3 0.01 0.1 0.5 1E-3 0.01 0.1 0.5
Frequency Frequency
(c) 1000 (d) 1000
24-hour Simulation 48-hour Simulation
800 Observation 800 Observation

Rainfall amount (mm)

Rainfall amount (mm)

600 600

400 400

200 200

0 0
1E-3 0.01 0.1 0.5 1E-3 0.01 0.1 0.5
Frequency Frequency

Fig. 8. Frequency distributions of observed and simulated cumulative rainfall amounts in different durations (the shadows are 90% confidence intervals): (a) 6 h; (b)
12 h; (c) 24 h; (d) 48 h.

4. Generation of future rainfall scenarios for the Greater Bay Area

4.1. Adaptation to regional climate change

In addition to learning the past characteristics of rainfall, possible

rainfall scenarios in the future are of great interest because it will sig-
nificantly influence the magnitudes of rainfall-triggered hazards and
the strategies of risk mitigation in GBA. Evidences of climate change in
terms of precipitation in the study area have been documented in the
literature (Zhang et al., 2009; Lenderink et al., 2011; Chen et al., 2012;
Du et al., 2016). A clear decreasing trend in rainy days and increasing
trend in the intensity and amount of extreme rainstorms have been
found in the past decades. The parameters of the proposed model can be
adjusted to consider the future trend of rainfall under climate change.
Although the observation period of data in this paper is relatively
short, some trends can still be found. The trends of inter-annual var-
iations of the parameters of the marginal distributions for the three
variables characterising the process of rainfall are shown in Fig. 9. The
inverse Gaussian distribution is used to fit the samples of dry spell
duration, and the probability density function of the dry spell duration
exhibits a clear trend to become less positively skewed, which indicates
that the probability of longer dry spell has increased over the past
20 years. Similarly, the probability density function of the average in-
tensity shows a trend that the rainfall has become more intense during
the past two decades.
To consider the future rainfall scenarios of the study area, the
parameters of the marginal distributions of the three variables are ex-
trapolated assuming the trend to be linear, while the other parameters
(i.e., the dependency between W and I, the relation between peak in-
tensity and average intensity, and the spatial correlation) are assumed
unchanged. Rainfall is generated with the adjusted parameters of the
marginal distributions of W, D and I during the period of 2080–2100.
The moments of the three variables in the future period and the asso-
ciated changing rates can be calculated. Frequency analysis is then
Fig. 9. Variation of parameters of marginal distributions: (a) wet spell duration carried out with the results of 100 realizations. Fig. 10 compares the
(geometric distribution); (b) dry spell duration (inverse Gaussian distribution); frequency distributions of simulated rainfall in the past and the future.
(c) average intensity (exponential distribution). The upper bounds of the 90% intervals, as well as the medians of

8
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

(a) 1000 (b) 1000

6-hour 1998-2017 12-hour 1998-2017
800 2081-2100 800 2081-2100

Rainfall amount (mm)

Rainfall amount (mm)

600 600

400 400

200 200

0 0
1E-3 0.01 0.1 0.5 1E-3 0.01 0.1 0.5
Frequency Frequency
(c) 1000 (d) 1000
24-hour 1998-2017 48-hour 1998-2017
800 2081-2100 800 2081-2100
Rainfall amount (mm)

Rainfall amount (mm)

600 600

400 400

200 200

0 0
1E-3 0.01 0.1 0.5 1E-3 0.01 0.1 0.5
Frequency Frequency
Fig. 10. Comparison of the frequency distributions of cumulative rainfall amount of different durations in the past (1998–2017) and future (2081–2100) (the solid
lines are median lines and the shadows are 90% confidence intervals): (a) 6 h; (b) 12 h; (c) 24 h; (d) 48 h.

Fig. 11. Comparison of the frequency distributions of rainfall amounts of different durations generated with different spatial correlation parameters (the solid and
dashed lines are median lines and 90% confidence intervals, respectively): (a) 6 h; (b) 12 h; (c) 24 h; (d) 48 h.

rainfall amount for all the four durations show a significant upside shift, of future rainfall scenarios. More rigorous adjustments can be made
especially for low frequency events. For example, the medians of future based on projections from climate models.
rainfall at the frequency of 0.01 increase by 11.3%, 8.1%, 8.9% and
7.6% for durations of 6, 12, 24 and 48 h, respectively. Based on the 4.2. Effect of spatial correlation
trend extrapolation of the past 20 years, the simulation results imply a
major increase in rainfall extremes at the end of 21st century. The in- In the proposed model, the spatial correlation parameter δ is used to
creasing trend of short-duration rainfall is particularly significant, characterise the inter-station correlation. In Section 4.1, the spatial
which indicates that flash floods and shallow slope failures related to correlation is assumed to be stationary in adaptation to climate change
short-term and high intensity rainfall will become more frequent in the given that the variation of spatial correlation is difficult to evaluate. In
future in GBA. The extrapolation is suitable for preliminary evaluation this section, sensitivity analysis is carried out to investigate the effect of

9
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

spatial correlation on the model performance in terms of the frequency- GBA, especially for short-duration rainfall.
based regional rainfall characteristics. Parameter δ is scaled to be one 3. The sensitivity of spatial correlation parameter quantifying the
half and twice of the value (i.e., 200 and 800 km) estimated from the inter-station correlation is investigated. The uncertainty of gener-
historical data. Different from the frequency analysis in Section 3.3, the ated rainfall will increase significantly for low frequency events as
generated annual maximum time series from different stations are not the correlation becomes stronger, which underlines the importance
simply pooled together in this case, because the difference of inter- of spatial correlation in regional rainfall modelling. However, the
station correlation will lead to different sample numbers. Instead, a differences in the magnitudes of regional rainfall induced by spatial
method using regional L-moments modified from the method of prob- correlation are limited especially for short-duration rainfall.
ability weighted moments estimation (Hosking & Wallis, 2005) is 4. The proposed rain field generator can be combined with physically-
adopted here for regional frequency analysis. The method of regional based distributed models for hydrological and geologic hazard
frequency analysis is used to estimate the frequency distribution of analyses in GBA to investigate the future hazard scenarios under
rainfall from the generalized extreme value distribution with para- climate change in the next step.
meters determined using regional averages of the mean, L-variance and
L-skewness. This method is based on the meteorological homogeneity of CRediT authorship contribution statement
the study area, which has been verified by Chong and Lee (2015).
The results of the regional frequency distribution with different Yejia Qiang: Data curation, Formal analysis, Writing - original
rainfall durations are shown in Fig. 11. A stronger spatial correlation (a draft. Limin Zhang: Conceptualization, Methodology, Funding acqui-
larger δ value) leads to higher uncertainties, especially for low fre- sition, Supervision, Writing - review & editing. Te Xiao: Software,
quencies. The differences of uncertainty intervals induced by δ decrease Validation, Investigation.
and tend to be similar as the frequency becomes larger. For long-
duration rainfall (i.e., 48-hour), the uncertainty growth due to a larger Declaration of Competing Interest
δ for low frequency events is quite remarkable; the upper bound of
which increases by 44.7% for events with a frequency of 0.001 when δ The authors declare that they have no known competing financial
is doubled. Except for the uncertainty, there are no obvious differences interests or personal relationships that could have appeared to influ-
between the magnitudes (medians) of 6-hour and 12-hour rainfall ence the work reported in this paper.
generated with different δ values, which indicates that the variation of δ
controlling the spatial distribution of rainfall within each specific time Acknowledgements
slot will not have a significant influence on the statistical characteristics
of short-duration regional rainfall. For rainfall of long durations (i.e., This work was supported by the Science and Technology Plan of
24-hour and 48-hour), the magnitudes (medians) of low frequency Shenzhen, China (Project No. JCYJ20180507183854827) and the
events will be smaller as the spatial correlation becomes stronger. The Research Grants Council of the Hong Kong SAR Government (Project
inter-station spatial correlation is very important in terms of un- Nos. C6012-15G and 16206217). The rainfall data used in this paper
certainty of statistical characteristics of generated rainfall, and it is can be downloaded from the dataset of NASA Tropical Rainfall
desirable to take the variation of spatial correlation parameter into Measuring Mission (https://trmm.gsfc.nasa.gov/).
account when adapting the model to climate change. However, since
the differences of rainfall magnitudes caused by variation of δ are not Appendix A. Supplementary data
very large, for problems that concern only about the overall regional
characteristics of rainfall, the spatial correlation parameter still can be Supplementary data to this article can be found online at https://
assumed to be stationary and even selected based on experience instead doi.org/10.1016/j.jhydrol.2020.124584.
of choosing from a wide range, and a predefined spatial correlation
function suffices to capture the regional spatial variation feature. On References
the other hand, more attention should be paid to the description of
spatial correlation to make the features of generated areal rainfall of Acreman, M.C., 1990. A simple stochastic model of hourly rainfall for Farnborough,
each specific time slot more realistic, and the effect of terrain should be England. Hydrol. Sci. J. 35 (2), 119–148. https://doi.org/10.1080/
02626669009492414.
considered in the spatial correlation analysis in areas with large var- Allcroft, D.J., Glasbey, C.A., 2003. A latent Gaussian Markov random-field model for
iations in topography. spatiotemporal rainfall disaggregation. J. R. Statistical Soc.: Series C (Appl. Statistics)
52 (4), 487–498. https://doi.org/10.1111/1467-9876.00419.
Bell, T.L., 1987. A space-time stochastic model of rainfall for satellite remote-sensing
5. Summary and conclusions studies. J. Geophys. Res.: Atmos. 92 (D8), 9631–9643. https://doi.org/10.1029/
JD092iD08p09631.
A spatial-temporal rain field generator based on point observations Bezak, N., Šraj, M., Mikoš, M., 2016. Copula-based IDF curves and empirical rainfall
thresholds for flash floods and rainfall-induced landslides. J. Hydrol. 541, 272–284.
is developed and applied to the Guangdong-Hong Kong-Macau Greater
https://doi.org/10.1016/j.jhydrol.2016.02.058.
Bay Area (GBA) which frequently suffers storm-related hazards. The Chen, L., Singh, V.P., 2018. Entropy-based derivation of generalized distributions for
major conclusions are as follows: hydrometeorological frequency analysis. J. Hydrol. 557, 699–712. https://doi.org/
10.1016/j.jhydrol.2017.12.066.
Chen, H., Sun, J., Chen, X., Zhou, W., 2012. CGCM projections of heavy rainfall events in
1. The rain field generator simulates the spatial-temporal rain field China. Int. J. Climatol. 32 (3), 441–450. https://doi.org/10.1002/joc.2278.
through an alternating renewal process in the time domain with the Chen, H.X., Zhang, L.M., 2014. A physically-based distributed cell model for predicting
consideration of inter-station spatial correlation in the space do- regional rainfall-induced shallow slope failures. Eng. Geol. 176, 79–92. https://doi.
org/10.1016/j.enggeo.2014.04.011.
main. The statistical characteristics of the rainfall in GBA is well Chen, H.X., Zhang, S., Peng, M., Zhang, L.M., 2016. A physically-based multi-hazard risk
reproduced by the rain field generator, which demonstrates the assessment platform for regional rainfall-induced slope failures and debris flows. Eng.
model’s ability as a tool for hazard analysis in GBA. Geol. 203, 15–29. https://doi.org/10.1016/j.enggeo.2015.12.009.
Chong, S.N., Lee, T.C., 2015. Regional Frequency Analysis of Extreme Rainfall in Hong
2. The rain field generator can be adapted for climate change with the Kong, Macao and Guangzhou. Hong Kong Special Administration Region: Hong Kong
parameters of the marginal distributions of three variables de- Observatory.
scribing duration, interval and intensity of rainfall events adjusted. Du, Y.D., Yang, H.L., Cao, C.X., Zhang, L.J., 2016. Future change of precipitation extremes
over the Pearl River Basin from regional climate models. J. Tropical Meteorol. 22 (1),
Extrapolating the variation trends of the marginal distributions ob- 57–65. https://doi.org/10.16555/j.1006-8775.2016.01.007.
served in the past 20 years in GBA to the end of 21st century, the Foufoula-Georgiou, E., Lettenmaier, D.P., 1987. A Markov renewal model for rainfall
simulation results show a significant increase in rainfall extremes in occurrences. Water Resour. Res. 23 (5), 875–884. https://doi.org/10.1029/

10
Y. Qiang, et al. Journal of Hydrology 583 (2020) 124584

WR023i005p00875. 2018.10.036.
Gao, C., Xu, Y.P., Zhu, Q., Bai, Z., Liu, L., 2018a. Stochastic generation of daily rainfall Ran, Q., Hong, Y., Li, W., Gao, J., 2018. A modelling study of rainfall-induced shallow
events: a single-site rainfall model with Copula-based joint simulation of rainfall landslide mechanisms under different rainfall characteristics. J. Hydrol. 563,
characteristics and classification and simulation of rainfall patterns. J. Hydrol. 564, 790–801. https://doi.org/10.1016/j.jhydrol.2018.06.040.
41–58. https://doi.org/10.1016/j.jhydrol.2018.06.073. Schuurmans, J.M., Bierkens, M.F.P., 2007. Effect of spatial distribution of daily rainfall on
Gao, L., Zhang, L.M., Cheung, R.W.M., 2018b. Relationships between natural terrain interior catchment response of a distributed hydrological model. Hydrol. Earth Syst.
landslide magnitudes and triggering rainfall based on a large landslide inventory in Sci. Discuss. 3 (4), 2175–2208. https://doi.org/10.5194/hess-11-677-2007.
Hong Kong. Landslides 15 (4), 727–740. https://doi.org/10.1007/s10346-017- Serinaldi, F., 2009. A multisite daily rainfall generator driven by bivariate copula-based
0904-x. mixed distributions. J. Geophys. Res.: Atmospheres 114, D10103. https://doi.org/10.
Gao, L., Zhang, L., Li, X., Zhou, S.Y., 2019. Evaluating metropolitan flood coping cap- 1029/2008JD011258.
abilities under heavy storms. J. Hydrol. Eng. 24 (6), 05019011. https://doi.org/10. Shen, P., Zhang, L.M., Chen, H.X., Fan, R.L., 2018. EDDA 2.0: integrated simulation of
1061/(ASCE)HE.1943-5584.0001793. debris flow initiation and dynamics considering two initiation mechanisms. Geosci.
Haan, C.T., Allen, D.M., Street, J.O., 1976. A Markov chain model of daily rainfall. Water Model Dev. 11 (7), 2841–2856. https://doi.org/10.5194/gmd-11-2841-2018.
Resour. Res. 12 (3), 443–449. https://doi.org/10.1029/WR012i003p00443. Swiss Re, 2013. Mind the risk: a global ranking of cities under threat from natural dis-
Haberlandt, U., 1998. Stochastic rainfall synthesis using regionalized model parameters. asters. Switzerland: Swiss Re.
J. Hydrol. Eng. 3 (3), 160–168. https://doi.org/10.1061/(ASCE)1084-0699(1998) Tang, G., Huang, J., Sheng, D., Sloan, S.W., 2018. Stability analysis of unsaturated soil
3:3(160). slopes under random rainfall patterns. Eng. Geol. 245, 322–332. https://doi.org/10.
Haberlandt, U., Ebner von Eschenbach, A.D., Buchwald, I., 2008. A space-time hybrid 1016/j.enggeo.2018.09.013.
hourly rainfall model for derived flood frequency analysis. Hydrol. Earth Syst. Sci. 12 Tang, X.S., Li, D.Q., Rong, G., Phoon, K.K., Zhou, C.B., 2013. Impact of copula selection
(6), 1353–1367. https://doi.org/10.15488/549. on geotechnical reliability under incomplete probability information. Comput.
Hallegatte, S., Green, C., Nicholls, R.J., Corfee-Morlot, J., 2013. Future flood losses in Geotech. 49, 264–278. https://doi.org/10.1016/j.compgeo.2012.12.002.
major coastal cities. Nat. Clim. Change 3 (9), 802–806. https://doi.org/10.1038/ Wilks, D.S., 1998. Multisite generalization of a daily stochastic precipitation generation
NCLIMATE1979. model. J. Hydrol. 210 (1–4), 178–191. https://doi.org/10.1016/S0022-1694(98)
Hosking, J.R.M., Wallis, J.R., 2005. Regional Frequency Analysis: An Approach Based on 00186-3.
L-moments. Cambridge University Press, UK. Wright, D.B., Smith, J.A., Baeck, M.L., 2014. Flood frequency analysis using radar rainfall
Huffman, G.J., Bolvin, D.T., Nelkin, E.J., Wolff, D.B., Adler, R.F., Gu, G., et al., 2007. The fields and stochastic storm transposition. Water Resour. Res. 50 (2), 1592–1615.
TRMM multisatellite precipitation analysis (TMPA): quasi-global, multiyear, com- https://doi.org/10.1002/2013WR014224.
bined-sensor precipitation estimates at fine scales. J. Hydrometeorol. 8 (1), 38–55. Wu, S.J., Tung, Y.K., Yang, J.C., 2006. Stochastic generation of hourly rainstorm events.
https://doi.org/10.1175/JHM560.1. Stoch. Env. Res. Risk Assess. 21 (2), 195–212. https://doi.org/10.1007/s00477-006-
Iadanza, C., Trigila, A., Napolitano, F., 2016. Identification and characterization of 0056-3.
rainfall events responsible for triggering of debris flows and shallow landslides. J. Xiao, T., Zhang, L.M., Li, X.Y., Li, D.Q., 2017b. Probabilistic stratification modeling in
Hydrol. 541, 230–245. https://doi.org/10.1016/j.jhydrol.2016.01.018. geotechnical site characterization. ASCE-ASME J. Risk Uncertainty Eng. Syst., Part A:
Lee, T., Singh, V.P., 2019. Discrete k-nearest neighbor resampling for simulating multisite Civ. Eng. 3 (4), 04017019. https://doi.org/10.1061/AJRUA6.0000924.
precipitation occurrence and model adaption to climate change. Geosci. Model Dev. Xiao, T., Li, D.Q., Cao, Z.J., Tang, X.S., 2017a. Full probabilistic design of slopes in
12 (3), 1189–1207. https://doi.org/10.5194/gmd-12-1189-2019. spatially variable soils using simplified reliability analysis method. Georisk: Assess.
Lenderink, G., Mok, H.Y., Lee, T.C., Van Oldenborgh, G.J., 2011. Scaling and trends of Manage. Risk Eng. Syst. Geohazards 11 (1), 146–159. https://doi.org/10.1080/
hourly precipitation extremes in two different climate zones–Hong Kong and the 17499518.2016.1250279.
Netherlands. Hydrol. Earth Syst. Sci. 15 (9), 3033–3041. https://doi.org/10.5194/ Zhang, Q., Xu, C.Y., Becker, S., Zhang, Z.X., Chen, Y.D., Coulibaly, M., 2009. Trends and
hess-15-3033-2011. abrupt changes of precipitation maxima in the Pearl River basin, China. Atmos. Sci.
Li, D.Q., Xiao, T., Cao, Z.J., Zhou, C.B., Zhang, L.M., 2016. Enhancement of random finite Lett. 10 (2), 132–144. https://doi.org/10.1002/asl.221.
element method in reliability analysis and risk assessment of soil slopes using Subset Zhou, S.Y., Gao, L., Zhang, L.M., 2019. Predicting debris-flow clusters under extreme
Simulation. Landslides 13 (2), 293–303. https://doi.org/10.1007/s10346-015- rainstorms: a case study on Hong Kong Island. Bull. Eng. Geol. Environ. 78,
0569-2. 5775–5794. https://doi.org/10.1007/s10064-019-01504-3.
Nelsen, R.B., 2007. An Introduction to Copulas. Springer Science & Business Media, Zhu, Z., Wright, D.B., Yu, G., 2018. The Impact of rainfall space-time structure in flood
Berlin. frequency analysis. Water Resour. Res. 54 (11), 8983–8998. https://doi.org/10.
Oliver, M.A., Webster, R., 1990. Kriging: a method of interpolation for geographical in- 1029/2018WR023550.
formation systems. Int. J. Geogr. Inf. Syst. 4 (3), 313–332. https://doi.org/10.1080/ Zhu, H., Zhang, L.M., Xiao, T., Li, X.Y., 2017. Generation of multivariate cross-correlated
02693799008941549. geotechnical random fields. Comput. Geotech. 86, 95–107. https://doi.org/10.1016/
Peres, D.J., Cancelliere, A., 2018. Modeling impacts of climate change on return period of j.compgeo.2017.01.006.
landslide triggering. J. Hydrol. 567, 420–434. https://doi.org/10.1016/j.jhydrol.

11