Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Journal of Hydro-environment Research

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

Research papers

A new method for estimation of spatially distributed rainfall through

merging satellite observations, raingauge records, and terrain digital
elevation model data
⁎ ⁎
Haiyun Shia,b,c,d, , Ji Chena, , Tiejian Lib,c,d, Guangqian Wangb,c,d
a
Department of Civil Engineering, The University of Hong Kong, Hong Kong, China
b
State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing, China
c
State Key Laboratory of Plateau Ecology and Agriculture, Qinghai University, Xining, Qinghai, China
d
Sanjiangyuan Collaborative Innovation Center, Qinghai University, Xining, Qinghai, China

A R T I C L E I N F O A B S T R A C T

Keywords: This study develops a new method to estimate spatially distributed rainfall through merging the satellite ob-
Rainfall servation, the raingauge record, and the terrain digital elevation model (DEM) data, including the following four
Satellite observation steps: (1) to select a suitable satellite observation dataset, (2) to downscale the selected satellite observation
Raingauge record dataset with the DEM data, (3) to determine the weighted differences between the raingauge record and the
Elevation
downscaled satellite observation dataset, and (4) to calculate the spatially distributed rainfall through merging
Spatial distribution
Data merging
the downscaled satellite observation dataset and the weighted differences. The rainstorm occurred on 21 July
2012 in Beijing, China, was considered as a case study to validate the method. Three satellite observation
datasets (i.e., TMPA 3B41RT, 3B42RT and CMORPH) were compared with the related raingauge record. Using
the new method, this study generated the spatially distributed rainfall data, which were further compared with
the three rainfall datasets, i.e., two original rainfall datasets (the selected satellite observation dataset and the
raingauge record) and one merged rainfall dataset without consideration of topographic influence. The result
revealed that the merged spatially distributed rainfall data is a more rational representation of the actual rainfall
than the three other datasets. Furthermore, using this data merging method and a hydrological model, the
Digital Yellow River Integrated Model (DYRIM), this study simulated the streamflow process at the Dashi River
basin in the southwest of Beijing and the Qingjian River basin in the middle Yellow River. The simulation results
showed that the spatially distributed rainfall data could have better performance than those three other datasets,
especially for the peak flow simulation. Overall, it is concluded that this data merging method can enhance our
capability in estimating the spatial distribution of rainfall.

1. Introduction regarded to be relatively accurate and reliable at the point where a

raingauge station locates, the density of the raingauge network is
As a fundamental meteorological variable, rainfall is a primary generally not high enough to describe the spatial distribution of rainfall
input for hydrological models; specifically, for distributed hydrological accurately. Hence, the satellite observation, which can provide areal
models, the estimation of spatially distributed rainfall is necessary. rainfall estimates, is regarded as a viable substitute or supplement for
Normally, the spatial distribution of rainfall can be estimated from a the raingauge record (e.g., Nezlin and Stein, 2005; Mishra et al., 2011;
variety of sources, such as the raingauge record and the satellite ob- Li et al., 2012). However, as a grid dataset, the satellite observation
servation. Several studies (e.g., Tabios and Salas, 1985; Dirks et al., may not be as accurate as the raingauge record for a designated point
1998; Nezlin and Stein, 2005; Mishra et al., 2011; Li et al., 2012) fo- because the rainfall estimate recorded at each grid is the result of the
cused on estimating the spatial distribution of rainfall. With different retrieval over the radiance that is scattered by the scene as a whole and
spatial interpolation methods, spatially distributed rainfall over a de- includes considerable uncertainties. Therefore, to enhance our cap-
signated region can be derived from the raingauge record (Tabios and ability in evaluating the spatial distribution of rainfall, we need merge
Salas, 1985; Dirks et al., 1998). Although the raingauge record is these two sources of rainfall data.

⁎
Corresponding authors at: Department of Civil Engineering, The University of Hong Kong, Hong Kong, China.
E-mail addresses: [email protected] (H. Shi), [email protected] (J. Chen).

http://dx.doi.org/10.1016/j.jher.2017.10.006

1570-6443/ © 2017 International Association for Hydro-environment Engineering and Research, Asia Pacific Division. Published by Elsevier B.V. All rights reserved.

Please cite this article as: Shi, H., Journal of Hydro-environment Research (2017), http://dx.doi.org/10.1016/j.jher.2017.10.006
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Several methods for merging rainfall data, e.g., the statistical ob- geostationary satellites are used for making up the rainfall data (e.g.,
jective analysis technique (Pereira et al., 1998; Kalinga and Gan, 2012), Miller et al., 2001; Huffman et al., 2007; Kidd et al., 2009; Michaelides
the Bayesian merging technique (Todini, 2001), the conditional mer- et al., 2009; Kidd and Levizzani, 2011).
ging technique (Ehret, 2002; Sinclair and Pegram, 2005), and other The CMORPH has been available since November 2002. One of its
techniques (e.g., Vila et al., 2009; Woldemeskel et al., 2013), have been products has the finest spatial (8 km) and temporal (30 min) resolutions
developed. The statistical objective analysis technique developed by among all of the satellite observation datasets, and is available by only
Pereira et al. (1998) is considered to be a purely statistical method for 3 h after a rainfall event; besides, it covers most part of the earth sur-
merging multiple rainfall datasets, and it does not account for the face, including 60°S-60°N and 180°E-180°W. By contrast, the TMPA has
factors which may influence rainfall spatial distribution (e.g., topo- been available several years earlier. Its standard product, 3B42, which
graphy). Todini (2001) developed a technique to combine areal rainfall is considered to be the most reliable one, has been produced since 1998,
and point rainfall in a Bayesian sense, based on the use of block-Kriging but it is unavailable until several months after a rainfall event, which
and Kalman filtering. Ehret (2002) proposed the conditional merging makes it impossible for the rapid analysis of the rainfall events. While
technique, which could extract the optimal information from the rain- its real-time products, including 3B40RT, 3B41RT, and 3B42RT, have
gauge record and the radar data, preserving the mean value based on been produced since 2000, and are available by 8 h after a rainfall
the raingauge record and the spatial structure of the radar data. event. It is worth noting that all of these TMPA products have lower
Moreover, Vila et al. (2009) proposed a method for merging the daily spatial and temporal resolutions than the above-mentioned CMORPH
raingauge record and satellite observation based on a combination of product. All of the TMPA products have the same spatial resolution of
additive and multiplicative bias correction to gain the lowest bias; 0.25° (about 25 km), while their temporal resolutions are different (e.g.,
Woldemeskel et al. (2013) developed a method for merging gauge and 3 h for 3B40RT and 3B42RT, 1 h for 3B41RT). Moreover, they cover the
satellite rainfall with specification of associated uncertainty. The above- range of 50°S-50°N and 180°E-180°W, which is smaller than that of the
mentioned methods indicate that rainfall data derived from different CMORPH. In this study, 3B40RT will not be used because it has been
sources at different spatial scales can be merged effectively. However, included in 3B42RT, which is a combination of 3B40RT and 3B41RT.
the current available merging methods normally lack the representation That is, for each grid box of 3B42RT, the value of 3B40RT is used if
of the orographic influence on rainfall. Numerous investigations have available, and otherwise the value of 3B41RT is used as a replacement
confirmed that topography may have a great impact on rainfall (e.g., (Qi and Wang, 2015). Therefore, this study will use the two TMPA real-
Daly et al., 1994; Wotling et al., 2000; Naoum and Tsanis, 2004). time products (i.e., 3B41RT and 3B42RT) and the above-mentioned
Generally, rainfall may increase along with the increase of elevation CMORPH product to develop the data merging method.
below a certain altitude, and then decrease above that altitude
(Schermerhorn, 1967). 3. Methodology
As mentioned above, the raingauge record can give relatively ac-
curate and reliable point rainfall estimates, the satellite observation can To merge the rainfall data from different sources and to include the
roughly describe the spatial distribution of rainfall, and orographic orographic influence, this study employs the following steps. First, pair-
influence on rainfall is considerable. This motivates the present study. wise comparisons of the raingauge record against various satellite ob-
With the availability of the terrain digital elevation model (DEM) data servation datasets are implemented to determine the most suitable sa-
which are comprised of exhaustive topographic information (Chen and tellite observation for further data merging. Second, the procedures
Kumar, 2001; Jia et al., 2011; Chen and Wu, 2012), this paper con- shown in Fig. 2 are used to include topographic information (using the
tributes to develop a new data merging method to estimate spatially DEM data), to separate the selected satellite observation grids into sub-
distributed rainfall through merging a variety of rainfall data (i.e., the grids, and then to merge the raingauge record and the selected satellite
raingauge record and multiple satellite observation datasets) and the observation sub-grid data. Third, validation of the data merging method
topographic information (i.e., the DEM data). is conducted through comparing the merged rainfall data against the
two original rainfall datasets (i.e., the raingauge record and the satellite
2. Study area and data observation) and another merged rainfall dataset without consideration
of elevation influence. Finally, to further validate the method, the
This study focuses on the rapid analysis of the rainfall events and merged rainfall data and the original datasets are used to drive a hy-
takes the rainstorm in Beijing occurred on 21 July 2012 as a case study. drological model and to simulate the streamflow.
To validate the newly-developed data merging method, the streamflow
process in the Dashi River basin in the southwest of Beijing is simulated.
3.1. Analysis of rainfall data from different sources
The study period is from July 21 (00:00 UTC) to July 22 (00:00 UTC),
2012. In addition, the Advanced Spaceborne Thermal Emission and
Each raingauge station corresponds to a certain satellite observation
Reflection Radiometer (ASTER) Global DEM dataset (ASTER GDEM
grid, and then the rainfall values recorded at that grid can be extracted
Validation Team, 2009, 2011) with the grid size of 30 m is used.
by using the spatial analyst tools in ArcGIS. For comparison, the rainfall
This study employs two sources of rainfall data, including the
depths recorded at the raingauge stations and the corresponding sa-
raingauge record and multiple satellite observation datasets. The hourly
tellite observation grids are compared; further, the differences of rain-
raingauge record data are obtained from 115 raingauge stations inside
fall depths among those different satellite observation datasets are
or around Beijing provided by the Beijing Water Authority (Fig. 1a).
computed. Then, the most suitable satellite observation for estimating
The satellite observation datasets used are the Climate Prediction
spatially distributed rainfall, which best matches with the raingauge
Center Morphing method (CMORPH) dataset produced by the National
record, can be determined.
Oceanic and Atmospheric Administration (NOAA) of the USA (Joyce
et al., 2004) and the TRMM Multi-satellite Precipitation Analysis
3.2. The data merging method
(TMPA) dataset produced by the National Aeronautics and Space Ad-
ministration (NASA) of the USA (Huffman et al., 2007; Huffman and
Generally, the raingauge record data, Pgau , and the satellite ob-
Bolvin, 2012). Generally, rainfall estimates from the CMORPH and
servation data, Psat , can be expressed as follows:
TMPA datasets are derived primarily by merging various passive mi-
crowave observations from the low-earth orbit satellites (e.g., TMI, Pgau = Ptrue + εgau (1)
SSM/I, AMSU-B and AMSR-E). For the locations where the passive
microwave observations are unavailable, infrared data from the Psat = Ptrue + εsat (2)

2
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Fig. 1. The raingauge stations and one hydrological station

in (a) Beijing and (b) the Dashi River basin.

where Ptrue denotes the true rainfall depth; εgau and εsat represent the where L is the mean distance between stations; S is the area of the
errors of the raingauge record and the satellite observation deviating smallest circle which can cover all the raingauge stations; N is the
from the true value, respectively. It is well known that the satellite number of raingauge stations. For any interest point, only raingauge
observation have systematic errors and random errors that are caused stations within the EIR of this point are used for data merging.
by various factors (Nair et al., 2009; Li et al., 2012). In this study, εsat in The second assumption is that spatial variation of rainfall can be
Eq. (2) includes both of these two types of errors. For a designated represented by a function of elevation. The functional forms may be
location, the difference between the two rainfall data, Δ , can be ex- various, such as linear, logarithmic and exponential (Osborn, 1984;
pressed as follows: Daly et al., 1994; Naoum and Tsanis, 2004). As a result, for a designated
region, a suitable functional form can be determined by using the ob-
Δ = Pgau−Psat = εgau−εsat (3) served rainfall data and the DEM data.
To obtain the merged spatially distributed rainfall data, two as- As the grid size of the original satellite observation dataset is usually
sumptions are adopted. The first one is that any region has an effective large (e.g., 8 km, or 25 km), it will not be suitable to provide a higher
influence radius (noted as EIR hereafter) for any rainfall event, which spatial resolution of rainfall data through data merging, which can be
reflects the influence of a rain bandwidth, and such an EIR is not larger used by hydrological models. Therefore, spatial downscaling, which can
than a certain distance. Qian (2004) indicated that the rain bandwidths separate an original satellite observation grid into a number of smaller
in most regions are within 10–50 km. Then, the following rule is used to sub-grids, is necessary (e.g., Mehrotra et al., 2014). The size of the se-
determine the EIR value. If the mean distance between stations com- parated satellite observation sub-grid can be determined from the scale
puted by Eq. (4) (Lin et al., 1997) is less than 50 km, the mean distance of the study area and the spatial resolution of the digital drainage
is the EIR value; otherwise, the EIR value is assigned as 50 km. network. If the rainfall data recorded at an original satellite observation
grid, Psat,ori , the average elevation over an original satellite observation
S grid, Hsat,ori , and a separated satellite observation sub-grid, Hsat,sub , are
L=
N (4) known, the downscaled rainfall data at each separated satellite

3
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Fig. 2. Flow chart of the data merging method.

Fig. 3. A conceptual diagram of Eq. (6).

observation sub-grid, Psat,sub , can be computed as follows according to no rainfall observed by the satellite over the grid). Further, if there is no
the second assumption: raingauge in an original satellite observation grid and Psat,ori is 0, the
rainfall depth in that satellite observation grid is 0; otherwise, the
Psat,sub = f (Hsat,sub,Hsat,ori,Psat,ori) (5) spatially distributed rainfall in that satellite observation grid can be
where f is a function derived from the relationship between rainfall and obtained from the raingauge record data only.
elevation. Usually, an original satellite observation grid or a separated In addition, according to the first assumption, the number of rain-
gauge stations used for data merging can be determined by the EIR
satellite observation sub-grid can include a number of the DEM grids.
Therefore, the value of Hsat,ori or Hsat,sub can be obtained by averaging value. Then, the data merging method need consider three situations
for including the raingauge record data, and the merged rainfall data at
the elevations of those DEM grids within the original satellite ob-
servation grid or the separated satellite observation sub-grid. It is worth any sub-grid, i, can be obtained using the following equation:
noting that the above downscaling process will not be conducted if the
value of Psat,ori at an original satellite observation grid is 0 (i.e., there is

4
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

⎧ Pi,sat ,sub K=0 stations within the EIR of the sub-grid; β is the power of D, which is a
⎪ Pi,sat ,sub + Δk K=1 positive real number. Following common practices (e.g., Nalder and
Pi,mer K Wein, 1998; Goovaerts, 2000; Mito et al., 2011), this study adopted the
⎨
⎪ Pi,sat ,sub + ∑ αk Δk K ⩾ 2 value β to be 2, and the IDW method turns into the so-called inverse
⎩ k=0 (6) distance squared method.
where Pi,mer and Pi,sat ,sub represent the merged rainfall data and the se-
parated satellite observation at the sub-grid i, respectively. K is the 3.4. The DYRIM hydrological model
number of raingauge stations within the EIR of the sub-grid. αk is the
weight of the k-th raingauge station, which can be computed by using To evaluate the accuracies of the merged spatially distributed
an interpolation method. Δk (= Pk,gau−Pk,sat ,sub , see Eq. (3)) is the dif- rainfall data, raingauge record data and satellite observation data, the
ference between the two rainfall datasets at the k-th raingauge station, Digital Yellow River Integrated Model (noted as DYRIM hereafter) was
which is equal to the raingauge record, Pk,gau , minus the separated sa- used to simulate runoff. The DYRIM is a distributed and physically-
tellite observation, Pk,sat ,sub . As the errors of the two rainfall datasets based hydrological model developed by Tsinghua University in Beijing,
(see Eqs. (1) and (2)) are assumed independent (Li and Shao, 2010), China, for simulating large-scale hydrological processes, soil erosion
data merging may compensate part of those errors and improve the and sediment transport (Wang et al., 2007, 2015; Li et al., 2009; Shi
accuracy of rainfall estimates to some extent; however, it is possible et al., 2015, 2016a). The DYRIM uses the high-resolution digital drai-
that data merging may enlarge those errors under certain conditions nage network, which is extracted from the DEM data and coded by the
such as the extreme rainfall events because both the raingauge record binary tree method (Li et al., 2010), to simulate the hydrological pro-
and satellite observation data may underestimate the rainfall amount in cesses. Moreover, the dynamic parallel computing technology was de-
that situation (Jiang et al., 2016a, 2016b). veloped to speed up the simulation (Li et al., 2011; Wang et al., 2011,
Fig. 3 is a conceptual diagram of Eq. (6). From Eq. (6), it can be seen 2012).
that if there is no raingauge station within the EIR of the interest sub- The DYRIM hydrological model takes hillslope-channel as a basic
grid (i.e., K = 0), the merged rainfall data will be equal to the separated hydrological unit, in consideration of different hydrological response
satellite observation at that interest sub-grid, taking no account of the mechanisms of hillslope and channel. The runoff-yield model is estab-
influence of the raingauge record. If there is only one raingauge station lished on the hillslope unit, where the soil mass is divided into topsoil
within the EIR of the interest sub-grid (i.e., K = 1), the merged rainfall layer and subsoil layer. A variety of hydrological processes can be si-
data will be equal to the sum of the separated satellite observation at mulated, including vegetation interception, evapotranspiration, in-
that interest sub-grid and the difference between the raingauge record filtration-excess runoff yield over the surface, subsurface flow in those
and the separated satellite observation at that raingauge station. If there two layers, and water exchange between those two layers.
are more than one raingauge station within the EIR of the interest sub- The parameters in the DYRIM hydrological model can be divided
grid (i.e., K ≥ 2), the merged rainfall data can be computed by using a into two types. One is the invariant parameters used for describing the
linear combination of the separated satellite observation at that interest properties of land use and soil type, including field capacity of topsoil
sub-grid and the weighted differences between the raingauge record layer, free water content of topsoil layer, field capacity of subsoil layer,
and the separated satellite observation at these raingauge stations. free water content of subsoil layer, depth of topsoil layer and water
Using Eq. (6), for any interest sub-grid over a region, we can merge capacity of unit leaf area index. They can be determined from literature
the raingauge record, the satellite observation and the DEM data, and and field survey. The other type is the adjustable parameters, including
then the spatial distribution of the merged rainfall data over the whole the vertical saturated conductivity of topsoil layer (KV1) which can
region can be obtained. In this study, in order to identify the im- control the surface infiltration rate and primarily influence the in-
portance of including the orographic influence, the merged rainfall data filtration-excess runoff, the vertical saturated conductivity from topsoil
without the orographic influence (excluding the DEM data) (noted as layer to subsoil layer (KV2) which can control the rate of water redis-
MWD hereafter) are also generated. Using Eq. (6), this study produces tribution between the two soil layers, and the horizontal saturated
the MWD data through taking the values of the separated satellite ob- conductivities of topsoil layer (KH1) and subsoil layer (KH2) which can
servation at the interest sub-grid (Pi,sat ,sub) and the separated satellite control the lateral subsurface flow rates of the two soil layers. Those
observation at the k-th raingauge station (Pk,sat ,sub) to equal the values of adjustable parameters can be calibrated with the observed rainfall and
the original satellite observation grids which cover that interest sub- streamflow data.
grid or raingauge station.
3.5. Automatic optimization technique of model calibration
3.3. Interpolation method
The adjustable parameters mentioned in the preceding subsection
are calibrated using an automatic optimization technique available in
The weight αk in Eq. (6) can be computed by using a spatial inter-
HPC (High Performance Computing) system, which is a double-layer
polation method, such as the inverse distance weighting (noted as IDW
parallelization of model calibration (Zhang et al., 2016). The lower-
hereafter) method. As a widely-used geometric method, the IDW
layer parallelism is brought by using the Message Passing Interface for
method cannot include the topographic influence, but its interpolation
the DYRIM hydrological model (Li et al., 2011), and the upper-layer
result is smooth (e.g., Goovaerts, 2000; Shi, 2013; Shi et al., 2014,
parallelism is achieved by using the job scheduling functions of an HPC
2016b, 2017a, 2017b). For the accuracy, several studies (e.g., Dirks
system for the genetic algorithm. Those adjustable parameters of the
et al., 1998; Lin et al., 2002) showed that the IDW method can provide
hydrological model at the lower-layer parallelism are calibrated by
better results than other methods. As a result, this paper selects the IDW
using the genetic algorithm at the upper-layer parallelism.
method to compute the weight assigned to each sampled sub-grid (αk )
in Eq. (6) as follows:
3.6. Assessment criteria
1
Dkβ
αk = K
In order to evaluate the performance of the data merging method
1
∑ j=1 and the hydrological simulation, four objective functions are used. They
Djβ (7)
are r (Correlation Coefficient), RE (Relative Error), RMSE (Root Mean
where Dk and Dj are the distances from the k-th and j-th sampled sub- Square Error) and NSCE (Nash-Sutcliffe Coefficient of Efficiency) (Nash
grids to the interest sub-grid, respectively. K is the number of raingauge and Sutcliffe, 1970). The equations for computing these objective

5
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Table 1
Statistical results for the comparison of the three satellite observation datasets against the
raingauge record data during the 21 July 2012 rainstorm in Beijing, China (see Fig. 4 for
the scatter plot of the comparison).

Data r Range of RE RMSE (mm)

CMORPH 0.70 −0.66 to 1.41 77.20

3B41RT 0.52 −0.77 to 1.87 83.65
3B42RT 0.65 −0.74 to 1.56 77.09

Fig. 4. A scatter plot of the comparison of the three satellite observation datasets against
the raingauge record data based on the total rainfall depth for the 21 July 2012 rainstorm
in Beijing, China.

functions are given in Appendix A.

4. Results and discussion

4.1. Selection of a suitable satellite observation dataset

In recent years, several studies (e.g., Shen et al., 2010; Qin et al.,
2014; Jiang et al., 2016a, 2016b) on the evaluation of TMPA and
CMORPH products have been conducted, which indicate that regarding
the near real-time datasets, CMORPH performs better than TMPA on
the whole. In this study, we evaluated two TMPA real-time products
(i.e., 3B41RT and 3B42RT) and one CMORPH product with the rain-
gauge record, and Fig. 4 shows the comparison of the total rainfall
depth recorded at each raingauge station against those recorded at the
corresponding grids of the three satellite observation datasets (i.e.,
TMPA 3B41RT, 3B42RT, and CMORPH) during the study period. From
Fig. 4, it can be observed that all of the three satellite observation da-
tasets had positive correlation with the raingauge record, but most of
the raingauge record values were larger than the corresponding satellite
observation values; however, some of the raingauge record values were
smaller than the corresponding satellite observation values. The reason
for this is that the raingauge record values represent point rainfall es- Fig. 5. Comparison of the CMORPH data against the raingauge record data during the 21
July 2012 rainstorm: (a) the total rainfall depth for each raingauge station; (b) the re-
timates but the satellite observation values represent areal rainfall es-
lationship between the raingauge record data and the RE values of the CMORPH data
timates. If a raingauge station is located near the rainfall center, the against the raingauge record data.
rainfall data recorded at the station may be much larger than the cor-
responding satellite observation data; in contrast, the raingauge rainfall
data may be smaller, or even equal to 0, if the station happens to locate data merging with the raingauge record.
far away from the rainfall center. Moreover, from Fig. 4, it can be found To understand more about the CMORPH data, we took the rainfall
that the total rainfall depth could be effectively estimated by both the totals during the study period recorded at all of the raingauge stations
raingauge record and the satellite observation datasets when the rain- as the X-axis and the rainfall totals recorded at the corresponding grids
fall is less than 200 mm; however, when the total rainfall depth re- of the CMORPH as the Y-axis. Then, the best fitted linear regression
corded at a raingauge station exceeded 200 mm, the corresponding between the two rainfall datasets was computed, and the line slope was
satellite observation values were not greater than 200 mm. 0.64 (Fig. 5a). Moreover, the relation between the raingauge record
Table 1 lists the statistical results from the comparison of the three data and the RE values of the CMORPH data against the raingauge
satellite observation datasets against the raingauge record; the values record data was explored (Fig. 5b), and it can be observed that most of
were obtained by using the equations in Appendix A. Among those RE values are negative. Positive RE values only appear when the rain-
three satellite observation datasets, the CMORPH dataset is the best to gauge record values are smaller than 200 mm. Furthermore, it is clear
match the raingauge record, with the highest r (0.70), the smallest that the RE values of the CMORPH data against the raingauge record
range of RE (−0.66 to 1.41) and the second smallest RMSE (77.20, near data presented a trend of linear decrease along with the increase of the
to the smallest one, 77.09). Moreover, in consideration of its highest raingauge record data. That is because the local large values which
spatial and temporal resolution, the CMORPH was regarded as the most were recorded by the raingauge record data could not be recorded by
suitable satellite observation for this case study and was selected for the CMORPH data; then, the larger raingauge record data would cause

6
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Table 2
Relation between rainfall (P) and elevation (H) described by different functional forms derived from the total CMORPH rainfall data and the related DEM data during the study
period.

Functional form P-H function for downscaling r2

Linear P = −0.08H + 143 Psat,sub = Psat,ori + 0.08(Hsat,ori−Hsat,sub) 0.44

Logarithmic P = −21ln(H ) + 223 Psat,sub = Psat,ori + 21ln(Hsat,ori/ Hsat,sub) 0.50
Exponential P = 141exp(−0.0008H ) Psat,sub = Psat,oriexp(0.0008(Hsat,ori−Hsat,sub)) 0.46

the larger difference between the CMORPH data and the raingauge for the spatial downscaling of the CMORPH data. Then, the relationship
record data. between the rainfall data recorded at an original grid of the CMORPH
dataset Psat,ori and the average elevations within that original grid Hsat,ori
4.2. The merged rainfall data can be expressed as follows:
Psat,ori = −21ln(Hsat,ori) + 223 (8)
Based on the 115 raingauge stations available over the whole city of
Beijing, the mean distance between stations (see Eq. (4)) is 20 km. As it Furthermore, this study assumes that the rainfall data recorded at a
was smaller than 50 km, the EIR in this study was set to be 20 km. sub-grid of the CMORPH dataset Psat,sub and the average elevations
Moreover, the studies of Sangati and Borga (2009) and Lobligeois et al. within that sub-grid Hsat,sub have the same relationship as that of an
(2014) revealed that the spatial resolution of 1 km is suitable for hy- original grid, the following equation can be obtained:
drological simulation, and, therefore, this study selected the sub-grid Psat,sub = −21ln(Hsat,sub) + 223 (9)
resolution of 1 km to estimate the spatial distribution of rainfall. Sub-
sequently, each original grid of the CMORPH would be separated into As Psat,ori,Hsat,ori and Hsat,sub were available, Psat,sub can be obtained by
64 (=8 × 8) smaller sub-grids, and totally there are 20,817 separated subtracting the above two equations:
sub-grids over the whole city of Beijing. Then, the merged rainfall data
Psat,sub = Psat,ori + 21ln(Hsat,ori/ Hsat,sub) (10)
in each sub-grid can be computed by using Eqs. (5)–(7). To merge the
raingauge record and satellite observation data, the form of Eq. (5) Theoretically, we can use Eq. (9) to compute Psat,sub directly since
should be determined first. Table 2 lists the relation between rainfall Hsat,sub can be computed using the DEM data. However, there are two
and elevation described by different functional forms derived from the constraints to use Eq. (9) directly. First, the r2 value in developing Eq.
CMORPH observations and the DEM data during the study period, and (8) is only 0.50 so that the errors involved into the description of the
the logarithmic function gives the best performance for this case study relationship between rainfall and elevation were rather large. Thus,
(r2 = 0.50). It is worth noting that rainfall had an inverse relationship including Psat,ori into Eq. (10) can reduce those errors in downscaling the
with elevation for this case study, mainly due to that the location of the CMORPH data. Second, it is worth noting that Eq. (8) was developed
main rainfall center in the lower elevation areas (i.e., the southern part based on total rainfall depth, and the downscaled scheme (i.e., Eq. (10))
of Beijing (see Fig. 6 for details)). Subsequently, this function was used is used to downscale the CMORPH data at hourly time step. Therefore,

Fig. 6. Comparison of the spatial distribution of (a) the merged rainfall data with the consideration of elevation influence, (b) the original CMORPH data, (c) the interpolated raingauge
record data with the IDW method directly, and (d) the MWD data, at the same spatial resolution, over the whole city of Beijing during the 21 July 2012 rainstorm.

7
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

it is necessary that Eq. (10) needs Psat,ori to determine Psat,sub at each 4.3. Hydrological simulation
hourly time step. Using Eq. (10), the grid size of the CMORPH data can
be separated from 8 km to 1 km, and the rainfall data within each sub- To further evaluate the performance of the data merging method,
grid of the CMORPH can be computed. Then, following the data mer- the merged rainfall data were used for streamflow simulation in the
ging equations (see Eqs. (6) and (7)), a combination of the raingauge Dashi River basin in Beijing. There are 4 raingauge stations (i.e.,
record and the separated CMORPH at any sub-grid over Beijing can be Daanshan, Manshuihe, Nanjiao and Xiayunling stations) and 7 rain-
obtained. gauge stations (i.e., Datai, Hongshuiyu, Huangta, Liangxiang, Puwa,
Fig. 6 shows the spatial distribution of the merged rainfall data with Nanxinfang and Zhaitang stations) inside or near to the river basin,
the consideration of elevation influence (Fig. 6a), the original CMORPH respectively (Fig. 1b). Hourly rainfall data are available at those 11
data (Fig. 6b), the interpolated raingauge record data with the IDW stations, which can be used to compute the merged rainfall data for
method directly (Fig. 6c), and the MWD data (Fig. 6d) over the whole each hour during the rainstorm. Moreover, there is one hydrological
city of Beijing during the rainstorm. It is worth noting that all of the station, the Manshuihe station, in this river basin (see Fig. 1b), and the
rainfall data from Fig. 6a–d have the same spatial resolution of 1 km. It observed streamflow data at this station are used to evaluate the si-
is clear that the merged rainfall data were characterized by more sig- mulation results. Fig. 1b also shows the high-resolution drainage net-
nificant spatial variation than the original CMORPH data. Two rainfall work of the Dashi River basin; there are 3304 river reaches and nearly
centers were found in Fig. 6a, locating in the southwest and the 8260 hillslopes in total, with an average hillslope area of 6.2 ha.
northeast of Beijing, respectively, while they could not be observed in It is worth noting that the merged rainfall data obtained in Section
Fig. 6b. That is mainly because the merged rainfall data can include the 4.2 were derived from all the 115 raingauge stations over the whole city
features of the raingauge record data which showed two rainfall centers of Beijing. As mentioned before, the density of raingauge stations in
in Fig.6c. In addition, for the merged rainfall data in both Fig. 6a and d, different part of Beijing was quite different (Fig. 1a), which might affect
rainfall values over some sub-grids were distinct from those over their the computation of the mean distance between raingauge stations, thus
surrounding sub-grids, and it is mainly due to the distinct differences of affecting the EIR. Therefore, we would reproduce the merging rainfall
the CMORPH data over those related grids in Fig. 6b. data for hydrological simulation. Based on the 11 raingauge stations
Table 3 lists the comparisons of the average rainfall depth computed and Eq. (4), the mean distance between raingauge stations for the river
by using those four rainfall datasets for the districts in Beijing as well as basin is 15 km. As it was smaller than 50 km, the EIR was set to be
the whole city (see Fig. 1). Overall, the values of the merged rainfall 15 km. Moreover, there are 836 sub-grids in total when the separated
data (see column (a) of Table 3) were mostly in between of the original sub-grid size was set to be 1 km. For each hour during the study period,
CMORPH data (see column (b) of Table 3) and the interpolated rain- the data merging procedures ((5)–(7)) were used to compute the
gauge record data (see column (c) of Table 3). The values of the MWD merged rainfall data. Fig. 7a shows the hourly merged rainfall data in
data (see column (d) of Table 3) are much closer to the interpolated the Dashi River basin from 02:00 to 17:00 UTC, July 21, 2012; Fig. 7b
raingauge record data, and generally larger than the merged rainfall gives the corresponding rainfall hyetograph during this period. From
data with topographic influence. This is mainly because of the form of Fig. 7, it is observed that there were two rainfall peaks. The first rainfall
Eq. (6), which indicates that the values of the merged rainfall data were peak began at 05:00 UTC in the downstream, lasting only one hour; the
dominant by the larger one of the two original rainfall datasets. As the second one began at 08:00 UTC in the upstream, and gradually moved
original CMORPH data were generally smaller than the original rain- to the downstream, lasting three hours. Moreover, from Fig. 7a, it can
gauge record data, the MWD data were dominant by the latter. In be observed that the maximum hourly rainfall (116 mm) appeared in
contrast, the downscaled CMORPH data with using the DEM data in- the upstream at 08:00 UTC, while the second largest hourly rainfall
creased the rainfall values in certain sub-grids. Further, it is observed (100 mm) occurred in the downstream at 10:00 UTC. After 12:00 UTC,
that the values of the original CMORPH data in most districts were no more heavy rainfall occurred in this river basin. In addition, the
much smaller than the other three rainfall datasets, except for Daxing rainfall hyetographs derived from the other three datasets (i.e., the
District. It is observed from Fig. 6b that large rainfall occurred inside original CMORPH data, the raingauge record data, and the MWD data)
Daxing district (described by the original CMORPH data); however, the are shown in Fig. 7b for comparison.
rainfall depths given by the raingauge record data (see Fig. 6c) were not With the high-resolution digital drainage network and the rainfall
large. The reason is that the density of raingauge stations inside this data, the DYRIM was used to simulate streamflow process over the
district (i.e., six in total) was lower and could not properly capture the Dashi River basin. Since only one rainfall case (i.e., 21 July 2012) was
occurred large rainfall (see Fig. 1a). available, this study adopted the following method to evaluate the ef-
fectiveness of the data merging method. First, the four rainfall datasets
(i.e., the merged rainfall data, the original CMORPH data, the raingauge
record data, and the MWD data) might have significantly different
Table 3 features of spatial-temporal patterns, it would be inappropriate to use
Comparisons of the average rainfall depth (in the unit of mm) computed by using (a) the
the same parameters for streamflow simulation. Therefore, using the
merged rainfall data against (b) the original CMORPH data, (c) the interpolated raingauge
record data with the IDW method directly, and (d) the MWD data, for the districts in automatic optimization technique of model calibration on HPC system,
Beijing as well as the whole city. the adjustable parameters were calibrated with each of the four rainfall
datasets, respectively. Second, with their own calibrated adjustable
District (a) (b) (c) (d) parameters, the streamflows computed by using each of the four rainfall
Fangshan 283 157 291 290 datasets were obtained; then, this study had four groups of streamflow
Pinggu 236 155 233 233 simulations. Finally, all the computed streamflows were compared with
Shunyi 218 126 218 217 the observed streamflow data. Table 4 lists the calibrated adjustable
Main urban area 219 135 217 220 parameters in those four groups, and it can be found that the major
Daxing 191 191 184 186
Tongzhou 171 172 180 178
difference was the parameter KV1, which can influence the infiltration-
Mentougou 144 102 151 148 excess runoff yield significantly.
Miyun 149 116 154 152 Table 5 lists the statistical results of comparison between stream-
Changping 138 101 145 144 flow simulations and observations using the four rainfall datasets. Fig. 8
Huairou 102 83 108 106
shows the comparisons of the streamflows simulated by using those
Yanqing 69 59 72 70
The whole city 167 119 171 170 four rainfall datasets with their own calibrated adjustable parameters
against the observed streamflow at the Manshuihe hydrological station.

8
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Fig. 7. (a) The hourly merged rainfall data and (b) the rainfall hyetograph of the four datasets in the Dashi River basin from 02:00 to 17:00 UTC, July 21, 2012.

Table 4 the peak value and the appearance time of the first peak flow was not
The adjustable parameters in the DYRIM calibrated by using (a) the merged rainfall data, consistent with the observation. In contrast, only the second peak flow
(b) the original CMORPH data, (c) the raingauge record data, and (d) the MWD data,
could be simulated properly by using the raingauge record data and the
respectively.
MWD data. Moreover, although both of the two peak flows could be
Scenario KV1 (mm/hr) KV2 (mm/hr) KH1 (mm/hr) KH2 (mm/hr) simulated by using the CMORPH data, the appearance times were not
consistent with the observations.
(a) 10.0 100.0 18.0 72.0 Second, known from the statistical results in Table 5, the peak va-
(b) 1.8 50.0 18.0 72.0
(c) 35.0 220.0 18.0 72.0
lues of the simulation results by using these four rainfall datasets were
(d) 22.0 170.0 18.0 72.0 nearly the same, and the RE values were all within ± 10% (+8.4% for
the raingauge record data, 0% for the CMORPH data, +9.5% for the
MWD data, and −1% for the merged rainfall data). However, the NSCE
Overall, among those four datasets, the simulation result computed by values and the streamflow volume of the simulation results (i.e., from
using the merged rainfall data had the best performance, while that by 4:00 UTC, July 21, 2012 to 16:00 UTC, July 22, 2012) were different.
using the CMORPH data had the worst performance. The following The NSCE value of the simulation result by using the merged rainfall
gives two aspects of the comparison to indicate the performance of the data was 0.66, which was larger than that by using the raingauge record
merged rainfall data. data (0.51), the CMORPH data (0.38), or the MWD data (0.52). The RE
First, from Fig. 8, it can be found that by using the merged rainfall value of streamflow volume of the simulation result by using the
data both of the two observed peak flows could be simulated, and, merged rainfall data was −17.9%, which was much smaller than that
specifically, the second peak flow could be captured properly; however,

9
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Table 5
The statistical results of streamflow simulations.

Scenario Data Peak flow RE Difference to peak time Streamflow volume RE NSCE
(m3/s) (%) (hr) (106 m3) (%)

(a) Merged rainfall 1099 −1 0 23 −17.9 0.66

(b) CMORPH 1110 0 +1 34 21.4 0.38
(c) Raingauge record 1203 8.4 0 20 −28.6 0.51
(d) MWD 1216 9.5 0 20 −28.6 0.52
Observed streamflow 1110 / 15:00 UTC 28 / /

sub-grid more accurate and the addition of the DEM data basically
presents the impact of local topography, resulting in the more sig-
nificant spatial distribution of rainfall which is much closer to the ac-
tual situation.

4.4. Applications

For the 21 July 2012 rainstorm in Beijing, this data merging method
has been proved to have satisfactory performance. In order to further
evaluate the applicability of this method for different rainfall events in
other regions, the Qingjian River basin (109°12′-110°24′ E, 36°39′-
37°19′ N), a river basin in the middle Yellow River, was selected
(Fig. 9). There are two hydrological stations inside this river basin,
Fig. 8. Comparisons of the streamflows simulated by using the raingauge record data, the
named the Yanchuan station and the Zichang station, respectively. As
original CMORPH data, the MWD data, and the merged data with their own calibrated the density of the raingauge station network in the upstream of the
parameters against the observed streamflow data at the Manshuihe hydrological station at Zichang station is higher than that in the downstream, only the up-
the Dashi River basin in Beijing. stream region was used for streamflow simulation. Moreover, two
rainfall events occurred in 2006.8 (noted as Event I hereafter) and
by using the raingauge record data (−28.6%), the CMORPH data 2006.9 (noted as Event II hereafter) were selected. The rainfall data
(+21.4%), or the MWD data (−28.6%). recorded at relevant 11 raingauge stations (their names are given in
It is worth noting that the merged rainfall data can include the Fig. 9) were used to compute the merged rainfall data, and Fig. 10
features of the raingauge record and the satellite observation, as well as shows the hourly merged rainfall data in the upstream of the Zichang
the impact of elevation on rainfall. Based on the rough spatial dis- station during these two selected rainfall events. Streamflow simulation
tribution of rainfall described by the areal satellite observation, the was conducted for each rainfall event using the same way as above, and
addition of the raingauge record makes the estimates for a designated the streamflow data recorded at the Zichang station were used to
evaluate the simulation results. Only the raingauge record data, the

Fig. 9. The raingauge stations and hydrological stations in the Qingjian River basin.

10
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Fig. 10. The hourly merged rainfall data in the upstream of the Zichang station for (a) Event I (2006.8) and (b) Event II (2006.9).

MWD data and the merged data were used for streamflow simulation Event II by using the raingauge record data and −26% for Event I and
because we focused on the improvement brought by this data merging 16% for Event II by using the MWD data; specifically, for Event I, the
method to the raingauge record data, which is still the most widely- second observed peak flow could only be simulated by using the
used rainfall data. merged rainfall data. Through analyzing the raingauge record data, the
Fig. 11 shows the comparisons of the streamflows computed by reason for this was that only three stations (i.e., Zichang, Xinzhuangke
using these three rainfall data with their own optimal parameters and Hecaogou) had rainfall depths larger than 5 mm before the second
against the observed streamflow data, and Table 6 lists several statis- peak flow appeared; while the rainfall depths of other three stations
tical results. Overall, the merged rainfall data showed the best perfor- (i.e., Jinzeyan, Sanshilipu and Yangkelangwan) were all 0. It indicated
mance and could improve the simulation accuracy to some extent, that the raingauge stations might have happened to locate far away
especially in capturing the peak flows. First, the NSCE values of the from the rainfall center and failed to record the local heavy rainfall. As
simulation results by using the merged rainfall data were both larger the merged rainfall data combined the raingauge record with the sa-
than those by using the raingauge record data and the MWD data; tellite observation, the rainfall condition over the whole study area
specifically, for Event I, the NSCE value reached 0.94. Second, the peak might be better reflected so that better simulation result could be ob-
flows of these two events could be well simulated by using the merged tained. However, the RE values of streamflow volumes by using the
rainfall data, with the RE values of 1% for Event I and 2% for Event II, merged rainfall data were larger than those by using the raingauge
respectively, comparing to those of −35% for Event I and 21% for record data and the MWD data.

11
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

Fig. 11. Comparisons of the streamflows computed by

using the raingauge record, the MWD data and the
merged rainfall data with their own optimal parameters
against the observed streamflow data for (a) Event I
(2006.8) and (b) Event II (2006.9).

Table 6 designated sub-grid more accurate. Third, the relationship between

The statistical results of streamflow simulations for the two rainfall events in the Qingjian rainfall and elevation established in this study was simply based on
River basin.
regression method, which was not physically based. The negative effect
Event Data Peak flow RE Streamflow RE NSCE of elevation on rainfall was found in the first event, which contradicted
(m3/s) (%) volume (%) with the common understanding to the orographic effect on rainfall.
(106 m3) This should be further investigated using more events in this region and
elevation may not be the only factor to be considered. However, due to
Event I Raingauge 110 −35 2.35 −3 0.79
(2006.8) record
the large grid size of the satellite observation data, the effects of other
MWD 125 −26 2.63 9 0.85 factors (e.g., mountain aspect) on rainfall could not be fully reflected.
Merged rainfall 171 1 3.03 25 0.94 Fourth, this method is event- and area-dependent. For different regions
Observed 170 / 2.41 / / and time scales, the relationships between rainfall and elevation should
streamflow
be reestablished based on the observed rainfall data and the DEM data.
Event II Raingauge 266 21 6.55 1 0.14 The simulation results were better to some extent if the DEM data was
(2006.9) record
considered, especially in capturing the peak flows, which indicated that
MWD 255 16 6.66 2 0.17
Merged rainfall 224 2 6.69 3 0.18 the DEM data might play a certain role in improving the simulation
Observed 220 / 6.51 / / accuracy although such contribution was limited. However, in some
streamflow cases (e.g., the simulations of streamflow volume in the second and
third events), the results were better if the DEM data was not used.
Therefore, the question about adding the DEM data into the merged
4.5. Discussion rainfall data should be further discussed.
Therefore, with the awareness of the above limitations, the new
Applying the proposed method for obtaining more reliable spatially method can provide more reliable spatially distributed rainfall. This
distributed rainfall, we also need to be fully aware of the limitations of would be useful for providing the more accurate input data for dis-
this method, which are mainly related to the following four aspects. tributed hydrological models (e.g., Wang et al., 2009), evaluating cli-
First, this study used the total rainfall observations to downscale the mate change (e.g., Zope et al., 2012), forecasting floods (e.g., Lobligeois
satellite observation data and to merge raingauge observations, and it et al., 2014), and exploring water resources sustainability (e.g., Tang
still needs more systematic researches to downscale and merge satellite et al., 2007). Overall, this study paved the way for our further devel-
observation and raingauge record rainfall data at each time step (e.g., opment in generating more rational spatially distributed rainfall.
one hour). Second, the density of raingauge stations was relatively high
in the selected regions, and it still needs further exploration on the
5. Conclusions
applicability of this newly-developed method in regions with lower
density of raingauge stations. Nevertheless, it is worth noting that based
This paper developed a new method for generating spatially dis-
on the coarse spatial distribution of rainfall described by the areal sa-
tributed rainfall with the integration of the raingauge record and the
tellite observation and topographic information of the DEM data, the
satellite observation and the consideration of the topographic effect on
addition of the raingauge record makes the rainfall estimates for a
rainfall. The findings of this paper can be concluded as follows: first, of

12
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

all the three satellite observation datasets (i.e., TMPA 3B41RT, TMPA valuable spatially distributed rainfall which indicated the significance
3B42RT, and CMORPH), the CMORPH best matched the raingauge of topographic influence and resulted in more rational flood flow si-
record for the study period; moreover, due to the highest spatial and mulation.
temporal resolutions, the CMORPH was selected for data merging with
the raingauge record. Second, different functional forms for describing Acknowledgements
the relation between elevation and CMORPH rainfall were obtained,
and the most suitable one was the logarithmic form, which was then This study was supported by the Non-profit Fund Program of the
used to downscale the CMORPH data by using the DEM data. Third, the Ministry of Water Resources of China (Grant No. 201501028), the
merged rainfall data, with and without the consideration of topographic Natural Science Foundation of Qinghai Province project (Grant No.
influence, were compared with the CMORPH and the raingauge record 2017-ZJ-911), the Natural Science Foundation of China project (Grant
datasets, and the merged rainfall data with topographic influence could No. 51579131) and the State Grid Corporation of China funded project
effectively comprise the satellite observations and raingauge records. (Grant No. SGQHJY00GHJS1500057). We are also grateful to the two
Fourth, the hydrological model simulation results by using those four anonymous reviewers who offered the insightful comments leading to
rainfall datasets revealed that the newly-developed method can provide improvement of this paper.

Appendix A

The equations used to compute r (correlation coefficient), RE (Relative Error), RMSE (Root Mean Square Error) and NSCE (Nash-Sutcliffe
Coefficient of Efficiency) are given as follows:
N
∑i = 1 (Pi,sat−Psat )(Pi,gau−Pgau )
r=
N N
∑i = 1 (Pi,sat−Psat )2 ∑i = 1 (Pi,gau−Pgau )2 (A.1)

RE = Pi,sat / Pi,gau−1 or Qi,sim/ Qi,obs−1 or Vi,sim/ Vi,obs−1 (A.2)

N
1
RMSE =
N
∑ (Pi,sat−Pi,gau )2
i=1 (A.3)
N
∑ (Qi,obs−Qi,sim )2
NSCE = 1− i =N1
∑i = 1 (Qi,obs−Qobs )2 (A.4)
where Pi,sat and Pi,gau are the i-th satellite observation and raingauge record (mm), respectively; Qi,obs and Qi,sim are the i-th observed and simulated
streamflows (m3/s), respectively. Vi,obs and Vi,sim are the i-th observed and simulated flood volumes (106m3), respectively. N is the sample size; Psat ,
Pgau and Qobs are the mean values of the satellite observation, the raingauge record and the observed streamflow, respectively.

References global precipitation estimates from passive microwave and infrared data at high
spatial and temporal resolution. J. Hydrometeorol. 5, 487–503.
Kalinga, O.A., Gan, T.Y., 2012. Merging WSR-88D stage III radar rainfall data with rain
ASTER GDEM Validation Team, 2009. ASTER Global DEM Validation Summary Report. gauge measurements using wavelet analysis. Int. J. Remote Sens. 33 (4), 1078–1105.
METI & NASA. Kidd, C., Levizzani, V., Turk, F.J., et al., 2009. Satellite precipitation measurements for
ASTER GDEM Validation Team, 2011. ASTER Global DEM Version 2 – Summary of water resource monitoring. J. Am. Water Resour. Assoc. 45, 567–579.
Validation Results. METI & NASA. Kidd, C., Levizzani, V., 2011. Status of satellite precipitation retrievals. Hydrol. Earth
Chen, J., Kumar, P., 2001. Topographic influence on the seasonal and interannual var- Syst. Sci. 15, 1109–1116.
iation of water and energy balance of basins in North America. J. Clim. 14 (9), Li, M., Shao, Q.X., 2010. An improved statistical approach to merge satellite rainfall es-
1989–2014. timates and raingauge data. J. Hydrol. 385, 51–64.
Chen, J., Wu, Y.P., 2012. Advancing representation of hydrologic processes in the Soil and Li, T.J., Wang, G.Q., Huang, Y.F., et al., 2009. Modeling the process of hillslope soil
Water Assessment Tool (SWAT) through integration of the TOPographic MODEL erosion in the loess plateau. J. Environ. Inf. 14 (1), 1–10.
(TOPMODEL) features. J. Hydrol. 420, 319–328. Li, T.J., Wang, G.Q., Chen, J., 2010. A modified binary tree codification of drainage
Daly, C., Neilson, R.P., Phillips, D.L., 1994. A statistical topographic model for mapping networks to support complex hydrological models. Comput. Geosci. 36 (11),
climatological precipitation over mountainous terrain. J. Appl. Meteorol. 33, 1427–1435.
140–158. Li, T.J., Wang, G.Q., Chen, J., et al., 2011. Dynamic parallelization of hydrological model
Dirks, K.N., Hay, J.E., Stow, C.D., et al., 1998. High-resolution studies of rainfall on simulations. Environ. Modell. Software 26, 1736–1746.
Norfolk Island Part II: interpolation of rainfall data. J. Hydrol. 208 (3–4), 187–193. Li, X.H., Zhang, Q., Xu, C.Y., 2012. Suitability of the TRMM satellite rainfalls in driving a
Ehret, U., 2002. Rainfall and Flood Nowcasting in Small Catchments using Weather distributed hydrological model for water balance computations in Xinjiang catch-
Radar. Ph.D. Thesis. University of Stuttgart. ment, Poyang lake basin. J. Hydrol. 426, 28–38.
Goovaerts, P., 2000. Geostatistical approaches for incorporating elevation into the spatial Lin, Z.H., Mo, X.G., Li, H.X., et al., 2002. Comparison of three spatial interpolation
interpolation of rainfall. J. Hydrol. 288, 113–129. methods for climate variables in China. Acta Geog. Sin. 57 (1), 47–56 [In Chinese].
Huffman, G.J., Adler, R.F., Bolvin, D.T., et al., 2007. The TRMM Multi-satellite Lin, B.G., Zhang, P.C., Gu, S.S., 1997. Improvement of weather radar-measured regional
Precipitation Analysis (TMPA): Quasi-global, multi-year, combined-sensor pre- rainfall with comparison to other techniques. J. Nanjing Inst. Meteorol. 20 (3),
cipitation estimates at fine scales. J. Hydrometeorol. 8, 38–54. 334–340 [In Chinese].
Huffman, G.J., Bolvin, D.T., 2012. Real-Time TRMM Multi-Satellite Precipitation Analysis Lobligeois, F., Andreassian, V., Perrin, C., et al., 2014. When does higher spatial resolu-
Data Set Documentation. ftp://trmmopen.gsfc.nasa.gov/pub/merged/V7 tion rainfall information improve streamflow simulation? An evaluation using 3620
Documents/3B4XRT_doc_V7.pdf. flood events. Hydrol. Earth Syst. Sci. 18 (2), 575–594.
Jia, S.F., Zhu, W.B., Lv, A.F., et al., 2011. A statistical spatial downscaling algorithm of Mehrotra, R., Evans, J.P., Sharma, A., et al., 2014. Evaluation of downscaled daily rainfall
TRMM precipitation based on NDVI and DEM in the Qaidam Basin of China. Remote hindcasts over Sydney, Australia using statistical and dynamical downscaling ap-
Sens. Environ. 115 (12), 3069–3079. proaches. Hydrol. Res. 45 (2), 226–249.
Jiang, S.H., Ren, L.L., Yong, B., Hong, Y., Yang, X.L., Yuan, F., 2016a. Evaluation of latest Michaelides, S., Levizzani, V., Anagnostou, E.N., et al., 2009. Precipitation: measurement,
TMPA and CMORPH precipitation products with independent rain gauge observation remote sensing, climatology and modeling. Atmos. Res. 94, 512–533.
networks over high-latitude and low-latitude basins in China. Chin. Geog. Sci. 26 (4), Miller, S.W., Arkin, P.A., Joyce, R.J., 2001. A combined microwave/infrared rain rate
439–455. algorithm. Int. J. Remote Sens. 22 (17), 3285–3307.
Jiang, S.H., Zhou, M., Ren, L.L., Cheng, X.R., Zhang, P.J., 2016b. Evaluation of latest Mishra, A.K., Gairola, R.M., Varma, A.K., et al., 2011. Improved rainfall estimation over
TMPA and CMORPH satellite precipitation products over Yellow River Basin. Water the Indian region using satellite infrared technique. Adv. Space Res. 48 (1), 49–55.
Sci. Eng. 9 (2), 87–96. Mito, Y., Ismail, M.A.M., Yamamoto, T., 2011. Multidimensional scaling and inverse
Joyce, R.J., Janowiak, J.E., Arkin, P.A., et al., 2004. CMORPH: A method that produces

13
H. Shi et al. Journal of Hydro-environment Research xxx (xxxx) xxx–xxx

distance weighting transform for image processing of hydrogeological structure in Shi, H.Y., Li, T.J., Wang, G.Q., 2017a. Temporal and spatial variations of potential eva-
rock mass. J. Hydrol. 411, 25–36. poration and the driving mechanism over Tibet during 1961–2001. Hydrol. Sci. J.
Nair, S., Srinivasan, G., Nemani, R., 2009. Evaluation of multi-satellite TRMM derived http://dx.doi.org/10.1080/02626667.2017.1332416.
rainfall estimates over a Western State of India. J. Meteorol. Soc. Jpn 87 (6), Shi, H.Y., 2013. Computation of spatially distributed rainfall by merging raingauge
927–939. measurements, satellite observations and topographic information: A case study of
Nalder, I.A., Wein, R.W., 1998. Spatial interpolation of climatic normals: test of a new the 21 July 2012 rainstorm in Beijing, China. In: Proceedings of the 35th IAHR World
method in the Canadian boreal forest. Agric. For. Meteorol. 92, 211–225. Congress, vols I and II, 530-542.
Naoum, S., Tsanis, I.K., 2004. Orographic precipitation modeling with multiple linear Sinclair, S., Pegram, G., 2005. Combining radar and rain gauge rainfall estimates using
regression. J. Hydrol. Eng. 9 (2), 79–102. conditional merging. Atmos. Sci. Lett. 6, 19–22.
Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through conceptual models part 1 - Tabios, G.Q., Salas, J.D., 1985. A comparative analysis of techniques for spatial inter-
A discussion of principles. J. Hydrol. 10, 282–290. polation of precipitation. Water Resour. Bull. 21 (3), 365–380.
Nezlin, N.P., Stein, E.D., 2005. Spatial and temporal patterns of remotely-sensed and Tang, Q.H., Hu, H.P., Oki, T., et al., 2007. Water balance within intensively cultivated
field-measured rainfall in southern California. Remote Sens. Environ. 96 (2), alluvial plain in an arid environment. Water Resour. Manage 21, 1703–1715.
228–245. Todini, E., 2001. A Bayesian technique for conditioning radar precipitation estimates to
Osborn, H.B., 1984. Estimating precipitation in mountainous regions. J. Hydraul. Eng. rain-gauge measurements. Hydrol. Earth Syst. Sci. 5, 187–199.
110, 1859–1863. Vila, D.A., de Goncalves, L.G.G., Toll, D.L., et al., 2009. Statistical evaluation of combined
Pereira, F.A.J., Crawford, K.C., Hartzell, C., 1998. Improving WSR-88D hourly rainfall daily gauge observations and rainfall satellite estimates over continental South
estimates. Weather and Forecasting 13, 1016–1028. America. J. Hydrometeorol. 10, 533–543.
Qi, L., Wang, Y.Q., 2015. Discrepancies in different precipitation data products in the Bay Wang, H., Fu, X.D., Wang, G.Q., et al., 2011. A common parallel computing framework
of Bengal during summer monsoon season. Adv. Meteorol. 2015, 806845. for modeling hydrological processes of river basins. Parallel Comput. 37, 302–315.
Qian, W.H., 2004. Synoptic meteorology. Peking University Press, Beijing [In Chinese]. Wang, G.Q., Wu, B.S., Li, T.J., 2007. Digital Yellow River model. J. Hydro-Environ. Res. 1
Qin, Y.X., Chen, Z.Q., Shen, Y., Zhang, S.P., Shi, R.H., 2014. Evaluation of satellite rainfall (1), 1–11.
estimates over the Chinese Mainland. Remote Sensing 6, 11649–11672. Wang, G.Q., Fu, X.D., Shi, H.Y., et al., 2015. Watershed Sediment Dynamics and
Sangati, M., Borga, M., 2009. Influence of rainfall spatial resolution on flash flood Modeling: A Watershed Modeling System for Yellow River. In: In: Yang, C.T., Wang,
modelling. Nat. Hazards Earth Syst. Sci. 9 (2), 575–584. L.K. (Eds.), Advances in Water Resources Engineering, Handbook of Environmental
Schermerhorn, V.P., 1967. Relations between topography and annual precipitation in Engineering Volume 14 Springer International Publishing.
western Oregon and Washington. Water Resour. Res. 3 (3), 707–711. Wang, G.S., Xia, J., Chen, J., 2009. Quantification of effects of climate variations and
Shen, Y., Xiong, A.Y., Wang, Y., Xie, P.P., 2010. Performance of high-resolution satellite human activities on runoff by a monthly water balance model: a case study of the
precipitation products over China. J. Geophys. Res. Atmos. 115, D02114. Chaobai River basin in northern China. Water Resour. Res. 45 (7), W00A11.
Shi, H.Y., Fu, X.D., Chen, J., et al., 2014. Spatial distribution of monthly potential eva- Wang, H., Zhou, Y., Fu, X.D., et al., 2012. Maximum speedup ratio curve (MSC) in parallel
poration over mountainous regions: case of the Lhasa River basin, China. Hydrol. Sci. computing of the binary-tree-based drainage network. Comput. Geosci. 38 (1),
J. 59 (10), 1856–1871. 127–135.
Shi, H.Y., Li, T.J., Liu, R.H., Chen, J., Li, J.Y., Zhang, A., Wang, G.Q., 2015. A service- Woldemeskel, F.M., Sivakumar, B., Sharma, A., 2013. Merging gauge and satellite rainfall
oriented architecture for ensemble flood forecast from numerical weather prediction. with specification of associated uncertainty across Australia. J. Hydrol. 499,
J. Hydrol. 527, 933–942. 167–176.
Shi, H.Y., Li, T.J., Wei, J.H., Fu, W., Wang, G.Q., 2016b. Spatial and temporal char- Wotling, G., Bouvier, C., Danloux, J., et al., 2000. Regionalization of extreme precipita-
acteristics of precipitation over the Three-River Headwaters region during tion distribution using the principal components of the topographical environment. J.
1961–2014. J. Hydrol.: Reg. Stud. 6, 52–65. Hydrol. 233 (1–4), 86–101.
Shi, H.Y., Li, T.J., Wang, K., Zhang, A., Wang, G.Q., Fu, X.D., 2016a. Physically based Zhang, A., Li, T.J., Si, Y., Liu, R.H., Shi, H.Y., Li, X., Li, J.Y., Wu, X., 2016. Double-layer
simulation of the streamflow decrease caused by sediment-trapping dams in the parallelization for hydrological model calibration on HPC systems. J. Hydrol. 535,
middle Yellow River. Hydrol. Process. 30 (5), 783–794. 737–747.
Shi, H.Y., Li, T.J., Wei, J.H., 2017b. Evaluation of the gridded CRU TS precipitation da- Zope, P.E., Eldho, T.I., Jothiprakash, V., 2012. Study of spatio-temporal variations of
taset with the point raingauge records over the Three-River Headwaters region. J. rainfall pattern in Mumbai city, India. J. Environ. Res. Dev. 6 (3), 545–553.
Hydrol. 548, 322–332.

14