You are on page 1of 11

Journal of Hydrology 567 (2018) 626–636

Contents lists available at ScienceDirect

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

Research papers

Improving the use of ground-based radar rainfall data for monitoring and T
predicting floods in the Iguaçu river basin

A.S. Falcka,b, , V. Maggionia, J. Tomasellab, F.L.R. Dinizc, Y. Meia, C.A. Benetid, D.L. Herdiesc,
R. Neundorfd, R.O. Caramb, D.A. Rodrigueze
a
Sid and Reva Dewberry Dept. of Civil, Environmental and Infrastructure Engineering, George Mason University, USA
b
National Center for Monitoring and Early Warning of Natural Disasters (CEMADEN), Brazil
c
Center for Weather Forecast and Climate Studies (CPTEC/INPE), Brazil
d
Paraná Meteorological System (SIMEPAR), Brazil
e
Earth System Science Center (CCST/INPE), Brazil

A R T I C LE I N FO A B S T R A C T

This manuscript was handled by Marco Borga, This study investigates the efficiency of correcting radar rainfall estimates using a stochastic error model in the
Editor-in-Chief, with the assistance of upper Iguaçu river basin in Southern Brazil for improving streamflow simulations. The 2-Dimensional Satellite
Francesco Marra, Associate Editor Rainfall Error Model (SREM2D) is adopted here and modified to account for topographic complexity, season-
Keywords: ality, and distance from the radar. SREM2D was used to correct the radar rainfall estimates and produce an
Radar rainfall ensemble of equally probable rainfall fields, that were then used to force a distributed hydrological model.
Streamflow ensemble Systematic and random errors in simulated streamflow were evaluated for a cascade of sub-basins of the Iguaçu
Uncertainties precipitation catchment, with drainage area ranging from 1,808 to 21,536 km2). Results showed an improvement in the
Flood event
statistical metrics when the SREM2D ensemble was used as input to the hydrological model in place of the radar
rainfall estimates in most sub-basins. Specifically, SREM2D was able to remove the relative bias (up to 50%) in
the radar rainfall dataset regardless of the basin dimension, whereas the random error was reduced more pro-
minently in the larger basins (up to 100 m3 s−1). An event scale evaluation was also performed for nine selected
flood events in three sub-basins. SREM2D reduced the overestimation in the cumulative rainfall and streamflow
volumes during these events.

1. Introduction centers are located, amplified both frequency and magnitude of flash
floods (Pisani and Bruna, 2011). A recent survey of the Brazilian Water
Minimizing the loss of human lives and mitigating socio-economic National Agency has identified five areas that are highly vulnerable to
impacts associated with severe flooding depend on the ability to issue floods in the Iguaçu river basin (ANA, 2014; Fig. 1).
warnings with sufficient lead-times to enable preemptive mitigation The use of hydrological models in flood monitoring largely depends
actions. Flood prediction poses scientific and operational challenges to on the reliability and availability of real-time precipitation input data.
natural disasters centers mainly due to the difficulties in monitoring This is due the fact that the quality of hydrological forecasts is strongly
rainfall that directly impacts streamflow model simulations. These is- dependent on the initial moisture conditions in the basin, which is
sues are aggravated in basins with short response times (few hours), mostly determined by antecedent rainfall events. Due to their fine
where forecasting systems need to combine meteorological and hy- spatial and temporal resolution, ground radar precipitation estimates
drological input at fine temporal and spatial scales (Caseri et al., 2016). represent a viable option for monitoring and forecasting flood hazards.
Located in southern Brazil, the Iguaçu river basin has a long history However, these estimates are affected by errors due to unwanted echoes
of severe floods with significant socio-economic impacts (Garcia, 2016). from the local topography and the conversion of reflectivity into pre-
Land use (deforestation) and climate changes in recent decades have cipitation rate (i.e., Z-R relationship), among others (Anagnostou et al.,
contributed to worsen the effects of floods in the region. In addition, the 1999, 2010, 2017, 2018). Despite several corrections in the calibration
disordered development of riverine areas and the increased land surface of the Z-R relationship and attempts to assess these uncertainties, error
impermeability, mainly in the basin headwater areas where large urban residuals are still present in the final radar precipitation products. For


Corresponding author at: Sid and Reva Dewberry Dept. of Civil, Environmental and Infrastructure Engineering, George Mason University, USA.
E-mail addresses: asfalck@gmail.com, aline.falck@cemaden.gov.br (A.S. Falck).

https://doi.org/10.1016/j.jhydrol.2018.10.046
Received 27 May 2018; Received in revised form 3 September 2018; Accepted 18 October 2018
Available online 21 October 2018
0022-1694/ © 2018 Elsevier B.V. All rights reserved.
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

Fig. 1. Map of the Iguaçu river basin divided into 9 sub-basins with outlet at União da Vitória in Brazil. The precipitation radar location is shown with a black triangle
and its coverage of 150 km radius with a black line. (For interpretation to colours in this figure, the reader is referred to the web version of this paper.)

instance, Cecinati et al. (2017) applied a geo-statistical technique to applied to a radar precipitation dataset to produce an ensemble of
improve radar rainfall estimates and Nerini et al. (2017) applied a non- rainfall estimates used to force the distributed hydrological model MHD
stationary stochastic ensemble generator that uses the short-space (Modelo Hidrológico Distribuído, Rodriguez and Tomasella, 2016). The
Fourier transformation to characterize radar rainfall uncertainty. Peleg model performance is assessed in nine sub-basins of the upper Iguaçu
et al. (2018) applied a rainfall ensemble generator to investigate the river basin in Southern Brazil. Characteristics of rainfall and streamflow
impact of precipitation variability at sub-pixel scales, preserving the during flooding events are investigated in terms of cumulative rain and
spatio-temporal structure and patterns to estimate rainfall extremes. flow volumes, time lag, and dispersion. The study region and the da-
Previous studies have demonstrated that rainfall error models have tasets are presented in Section 2. Section 3 describes the methodology,
the potential to improve the performance of hydrological models for including the hydrological model, SREM2D, and the hydrograph se-
monitoring and predicting extreme events (Hong et al., 2006; paration method used to identify the flooding events. Section 4 presents
Nikolopoulos et al., 2010; Maggioni et al., 2013; Falck et al., 2015). and discusses the results and Section 5 summarizes the conclusions.
These error models perturb the original rainfall estimates (usually from
satellite products; e.g., Derin et al., 2018) and statistically generate an
ensemble of equiprobable precipitation fields. Ensemble error correc- 2. Study area and dataset
tions of satellite rainfall estimates have been shown to improve model
streamflow simulations over the Tar river basin in North Carolina, U.S., The Iguaçu river basin is located in Southern Brazil (97%) and
(Maggioni et al., 2013), as well as across the Tocantins–Araguaia basin Argentina (3%), covering an area of 70,800 km2 (ÁGUASPARANÁ,
in the Brazilian Amazon region (Falck et al., 2015). 2017). The Iguaçu River is formed by the confluence of the Iraí and
Short-term precipitation forecasting can be based on the extra- Atuba rivers in the metropolitan area of Curitiba (3.5 million in-
polation of consecutive radar scans. To assess the uncertainty of radar habitants, IBGE, 2015). The springs are situated in the western edge of
rainfall estimates, probabilistic radar ensembles have been used in the Serra do Mar, running for 1320 km until they join the Paraná River
combination (or not) with numerical weather prediction models (Fig. 1). The predominant climatic regime in the basin is humid sub-
(Liguori and Rico-Ramirez, 2013). Codo and Rico-Ramirez (2018) tropical, oceanic with no dry season and hot summers (Alvares et al.,
showed that the extrapolation of radar rainfall fields could be improved 2013). Since the Iguaçu basin includes 10 multipurpose reservoirs
by a radar ensemble generator that accounts for the residual error be- (COPEL, 2017), the impact of flooding is reduced in the lower part of
tween radar rainfall estimates and rain gauge observations. the basin by reservoir operation. On the other hand, the upper Iguaçu
This study investigates the efficiency of correcting radar rainfall basin is mostly unregulated and encompasses a drainage area of
estimates using a stochastic rainfall error model to simulate streamflow 21,536 km2. Floods in the region are usually associated with diurnal
in a region characterized by complex topography and seasonal rainfall convection in small basins and stratiform and convective precipitation
variability. Specifically, the 2-Dimensional Satellite Rainfall Error associated with cold fronts in the larger basins.
Model (SREM2D), developed by Hossain and Anagnostou (2006), is Rainfall datasets used in this study were derived from a ground-
based radar and a rain gauge network. The Paraná Meteorological

627
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

System (SIMEPAR) S-band Doppler radar is located in Teixeira Soares in Table 2


Southern Brazil. Radar data were processed by SIMEPAR using the re- Calibration and validation periods for MHD-INPE, SREM2d and CPM.
flectivity of the Constant Altitude Plan Position Indicator (CAPPI) level MHD-INPE SREM2D CPM
at 3 km height above mean sea level. The spatial resolution is 1 km, the
temporal resolution is about 5 min for a full volume scans with 7 ele- Calibration January December December
2005–November 2007–November 2008–November
vations PPIs (Plan Position Indicator) and the first elevation angle of
2008 2008 2010
0.5 deg, whereas the data collection frequency is 10 min. The maximum Validation December 2008- December
range is 200 km, but the radar makes a scan in vigilance mode every November 2013 2008–November
30 min with the maximum range of 480 km around the radar. Radar 2010
data were aggregated to the 4 km/hourly spatio-temporal scale setting
the radar range to 150 km, as the radar performance commonly de-
grades at distances beyond ∼150 km (Ciach et al, 1997; Simpson and weighted distance between the pixel average altitude and the rain
Fox, 2018). Data are available for a period that ranges from December gauge altitude, as described by Lefèvre et al (2002). The Shuffled
2007 to November 2010. Complex Evolution Algorithm method (SCEA; Duan et al., 1992) com-
Hourly data from 40 rain gauges, 14 meteorological stations (where bines deterministic and probabilistic approaches through the systematic
air temperature, dew point temperature, atmospheric pressure, incident evolution of complex data points and the direction of the optimum. The
global radiation and wind speed are measured), and nine streamflow set of optimum parameters values was selected maximizing the mean
stations were available for an 8 year-long time series (January Nash–Sutcliffe efficiency (NSE) and the logarithm of the Nash–Sutcliffe
2005–December 2013). The hydrometeorological variables were efficiency (NSElog) of simulated vs. observed streamflow. The hydro-
quality-controlled by SIMEPAR. Additional information used in the logical model was calibrated for the period 2005–2008 using con-
hydrological model was a Digital Elevation Model (DEM) derived from tinuous series of interpolated meteorological data (Table 2). The model
the Shuttle Radar Topography Mission (SRTM) with 90 m spatial re- was spun up for a period of 720 hours to reduce the impacts of initial
solution (Farr et al., 2007), maps of land use and land cover from the conditions. For the validation run (2008–2013), the model used the
ProVeg (Vegetation Project; Sestini et al., 2002), and a soil map moisture states from the calibration run as initial conditions (cold
(Embrapa, 1981). start).

3. Methodology 3.2. The SREM2D error model

3.1. The MHD-INPE hydrological model The SREM2D model was developed by Hossain and Anagnostou
(2006) to correct satellite rainfall estimates based on a higher quality
The distributed hydrological model developed at the Brazilian reference dataset. SREM2D characterizes the multidimensional struc-
Institute for Space Research (INPE), namely MHD-INPE (Rodriguez and ture of satellite rainfall errors using stochastic space–time formulations.
Tomasella, 2016), is a regular grid-cell model that uses hydrological In this study, SREM2D was used to correct radar rainfall estimates with
response units in each cell as a function of land use and coverage. MHD- respect to the reference ground observations and to generate an en-
INPE is composed of four modules: (i) soil water balance; (ii) evapo- semble of equally probable reference-like rainfall fields. The multi-
transpiration; (iii) surface, sub-surface and groundwater fluxes; and (iv) dimensional structure of the error includes: (i) a temporal dimension
channel routing. This model has been successfully applied to climate (how does the error vary in time?); (ii) a spatial dimension (how does
change and land use/land cover change impact studies (Siqueira et al., the error vary in space?); and (iii) a retrieval dimension (what is the
2015; Mohor et al., 2015; Rodriguez and Tomasella, 2016), as a tool for magnitude of the precipitation error in rainy-areas?) (Hossain et al.,
hydrological monitoring and forecasting (Falck et al., 2015, Casagrande 2009). SREM2D considers both false alarms and missed rain through
et al., 2017, Tomasella et al., 2018), and by insurance companies for the joint spatial probability of successful delineation of rainy and non-
estimating economic indicators during droughts (Mohor and rainy areas. This is a critical element as real sensor data (including
Mendiondo, 2017). Additional details on the hydrological model can be radar rainfall estimates) exhibit spatial clusters of false rain and false
found in Rodriguez and Tomasella (2016). no-rain.
The MHD-INPE model was run in nine sub-basins of the Iguaçu river SREM2D has been successfully used to improve the performance of
basin (Table 1) over a 4 km regular grid at hourly temporal resolution. hydrological models for monitoring and predicting floods. Past studies
Model parameters were calibrated using hourly observations of me- have evaluated the propagation of satellite precipitation uncertainties
teorological data interpolated across the basin using the inverse square in streamflow simulations in mountainous basins and observed a de-
effective distance. The latter was calculated as the sum of the geodetic pendency of rainfall-runoff error propagation on the basin size
distance between the station and the center of the pixel and the (Nikolopoulos et al., 2010, Maggioni et al., 2013). Furthermore, Falck

Table 1
Iguaçu basin characteristics.
Sub-basin (index) Station Area (km2) Elevation (m) Distance radar (km)*

Mean Range Stand. Dev.

B1 Rio Negro 3040 892.22 790–1333 65.19 112


B2 São Bento 1808 879.91 777–1146 58.47 99
B3 Divisa 7168 830.53 760–945 39.18 75
B4 Porto Amazonas 3264 929.93 853–1119 49.09 96
B5 São Mateus do Sul 5440 849.68 761–968 50.75 35
B6 Pontilhão 1968 857.19 767–1210 75.00 41
B7 Fluviópolis 16,432 821.66 755–1246 77.88 89
B8 Sta Cruz do Timbó 2448 1058.41 748–1326 124.71 124
B9 União da Vitória 21,536 840.08 747–1226 110.64 85

* Distance from the radar to the centroid of each sub-basin.

628
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

et al. (2015, 2016) applied SREM2D to satellite precipitation products TLAG = CQ − CR (3)
over the Tocantins-Araguaia basin in Brazil (5280–764,000 km2) and
where R indicates rainfall; Q is streamflow; TR and TQ are the period of
showed that the systematic error in the streamflow ensemble decreases
rainfall and flow event, respectively; t is the time step. VR and VQ re-
as a function of catchment area, although the rainfall-to-streamflow
present the magnitude of cumulative rainfall and streamflow over the
error propagation did not show any dependency on the basin size.
event period, respectively. CR and CQ, are the temporal location of mass
In this study, SREM2D parameters were calibrated based on the
centers of hyetograph and hydrograph; their difference, TLAG, re-
radar rainfall estimates considering the rain gauge observations as
presents the mean time lag of flood response of a catchment. SR and SQ
“reference”. A regional calibration was chosen to account for the
represent the temporal degree of dispersion of rainfall and streamflow
complex topography of each sub-basin and the distance from the radar,
with respect to CR and CQ, where larger SR and SQ indicate less con-
which largely affect the precipitation error. Errors in the radar pre-
centrated patterns of the hyetograph and hydrograph, respectively. CR
cipitation estimates are indeed related to the spread of the radar beam,
and CQ are defined as:
as distance from the radar increases (usually > 150 km), causing dis-
tortion in the shapes and sizes of echoes, decreasing the radar detection ∫TR t·R (t ) dt ∫T t·Q (t ) dt
capability. Areas close to the radar (< 50 km) are affected by the blind CR = CQ = R
∫TR R (t ) dt ∫TR Q (t ) dt (4)
cone and secondary lobes generating ground clutter. Radar precipita-
tion estimates are also affected by topography due to ground clutter, VR, VQ, SR, SQ and TLAG were calculated for three precipitation datasets
which is reflected back to the radar and mistakenly interpreted by the (i.e., gauge-interpolated rainfall fields; radar precipitation estimates,
sensor as cloud water drops, resulting in incorrect high precipitation and SREM2D rainfall ensemble mean) and the corresponding stream-
rates. flow simulations (obtained by forcing the hydrological model with each
During the SREM2D calibration phase, only pixels where both radar precipitation dataset).
and rain gauge reference were available at the same time were con-
sidered. Moreover, a seasonal calibration was performed for the four 3.4. Performance analysis
following seasons: summer (December 2007–February 2008), fall
(March to May 2008), winter (June to August 2008), and spring The hydrological model performance was assessed using two me-
(September to November 2008). Then, in the running phase, the cali- trics: the Nash–Sutcliffe Efficiency (NSE) and the logarithmic
brated SREM2D was run to produce an ensemble of 25 members (that Nash–Sutcliffe efficiency (NSElog), defined as:
reproduce the reference precipitation dataset) for all pixels in each sub- N
basin during both calibration (December 2007–November 2008) and ∑t = 1 (QSt − QOt )
NSE = 1 − N
validation (December 2008–November 2010) periods at hourly time ¯ )2
∑t = 1 (QOt − QO (5)
step (Table 2).
N
∑t = 1 (log(QSt ) − log(QOt ))
NSElog = 1 − N
3.3. The characteristic point method ∑ (log(QO )−log(¯QO ))2 (6)
t=1 t

To evaluate the performance of radar rainfall ensemble in capturing where QOt represents observed hourly streamflow; QSt refers to
streamflow during a flooding event, we extracted flood events from the streamflow simulated by forcing the model either with rain gauge, or
long-term continuous streamflow time series using the Characteristic radar data, or the SREM2D ensemble; N indicates the number of time
Point Method (CPM), proposed by Mei and Anagnostou (2015). CPM is intervals; and overbars indicate the average value. NSE and NSElog can
a fully automated hydrograph separation method that uses time series range between −∞ and 1, where 1 represents the perfect match of
of rainfall and streamflow as input. The method defines a flood event simulated streamflow to the observed data; zero means that the model
based on the recession characteristics of streamflow rather than any simulation is as accurate as the mean of the observations; and negative
user-specific streamflow threshold. Moreover, it incorporates informa- values mean that the observation average is better than the model si-
tion of basin-average precipitation to identify the triggering rainfall of mulation.
flood events. CPM has been successfully applied to the development of a MHD-INPE was evaluated based on two additional metrics: the re-
comprehensive flood event database for the continental United States lative bias (rBIAS) and the unbiased root mean squared error
during 2002–2013 (Shen et al., 2017). (ubRMSE), defined as follows:
In this study, we used gauge-interpolated rainfall fields to derive 1 N
N
∑t = 1 (QSt − QOt )
basin-average rainfall for each sub-basin considering the total upslope rBIAS = 1 N
contributing area for the December 2008–November 2010 period. N
∑t = 1 (QOt ) (7)
These basin-average rainfall time series and corresponding streamflow
1 N
record were used as input to CPM to extract flood events. A total of 66 ubRMSE = ∑t=1 ¯ t ) − (QOt − QO
[(QSt − QS ¯ t )]2
N (8)
flood events was extracted for the nine sub-basins during the study
period (Table 2) and these flood events form the population for our For the SREM2D ensemble-forced simulation, rBIAS and ubRMSE
analysis. were calculated for each of the 25 members of the streamflow en-
A set of flood events properties was calculated for the flood events semble.
population: cumulative event rainfall and flow volumes (VR and VQ), To evaluate the events properties of rainfall and streamflow de-
dispersion of event rainfall and flow time series (SR and SQ), and time scribed in Section 3.3 we also used the rBIAS:
lag (TLAG). These parameters represent different aspects of the flood
(RE − RO) (QS − QO)
event and their accurate representation reflects the model performance rBIAS = rBIAS =
RO QO (9)
in capturing the magnitude, temporal variability, and timing of floods.
The flood event properties are defined as follows: where RO is the reference rainfall; RE refers to “Rainfall Estimate” (and
corresponds to either the radar observations or the SREM2D ensemble
VR = ∫T R
R (t ) dtVQ = ∫T
Q
Q (t ) dt
(1) mean); QO is the streamflow simulated by rain gauge observations; and
QS is the streamflow simulated by forcing with RE (i.e., either radar
∫TR (t − CR)2 ∫TQ (t − CQ)2 observations or the SREM2D ensemble mean). The optimal value of
SR = SQ = rBIAS is zero; the positive/negative values represent an overestimate/
∫TR R (t ) dt ∫TQ Q (t ) dt (2) underestimate in rainfall (or streamflow). ubRMSE can range between 0

629
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

Table 3 reproduce streamflow. Even in the case of B8, which is characterized by


Hydrological model calibration (2005–2009) and validation (2010–2013) re- extremely complex topography (Fig. 1), the NSE coefficient was above
sults. 0.8 and 0.7 during the calibration and validation periods, respectively.
Basin Calibration Validation According to Moriasi et al. (2007), such NSE values correspond to ex-
cellent model performance.
NSE NSElog NSE NSElog Moreover, due to the low density of the real time rain gauge net-
work, the interpolated rainfall field is affected by errors and un-
B1 0.409 0.658 0.559 0.692
B2 0.274 0.567 0.249 0.397 certainties, regardless of the chosen interpolation method, particularly
B3 0.623 0.763 0.628 0.393 in complex terrain areas such as the headwater catchments. This ex-
B4 0.555 0.552 0.664 0.817 plains (at least partially) the lower performance of the hydrological
B5 0.701 0.658 0.645 0.689
model for sub-basins B1 and B2 (Table 3). Although the radar is able to
B6 0.804 0.802 0.743 0.822
B7 0.823 0.845 0.776 0.869 capture the spatial structure of rainfall fields, further corrections may
B8 0.841 0.691 0.763 0.479 be necessary to remove biases and improve correlations with reference
B9 0.883 0.892 0.801 0.886 rain gauge observations. In this context, the use of ensemble techni-
ques, as proposed in this study, becomes an attractive tool, particularly
in early warning systems.
and +∞, where the optimal value is zero.

4.2. Rainfall error model


4. Results
A previous study by Falck et al. (2015) demonstrated that a local
4.1. Hydrological model calibration and validation SREM2D calibration is preferred to a single calibration for the entire
Tocantins-Araguaia basin, a large basin located in center of Brazil.
Table 3 shows calibration and validation results in terms of NSE and Specifically, they showed that parameters calibrated for the whole
NSElog for nine streamflow stations. Overall, the hydrological model basin (764,000 km2) could not fully represent rainfall errors in catch-
was able to reproduce the hydrological response of the Iguaçu basin in ments smaller than 25,000 km2. Even though the Iguaçu and Tocantins-
terms of frequency and magnitude of streamflow peaks and recession Araguaia basins are characterized by different precipitation regimes,
periods. For most sub-basins, NSE and NSElog were close to or higher size, and topography, the same approach was followed for the Iguaçu
than 0.7 during both calibration and validation periods. However, in basin, where complex terrain and distance from the radar location play
the headwater basins, like B1, B2 and B4, NSE showed values ranging a fundamental role in rainfall error characterization.
from 0.3 to 0.7. The poorer performance is likely due to the density of Fig. 2a shows the relative frequencies of elevation (hypsometric
the rain gauge network, which is crucial for representing the hydro- index) in the upper Iguaçu basin (black line) and in each sub-basin
logical response in smaller drainage areas with higher spatial rainfall (colored lines). The basin elevation ranges from 1333 m (in B1) to
variability influenced by topography. For instance, in sub-basin B2, 747 m at the basin outlet, located in B9, with an average of 889 m. The
where the model showed the worst performance, there exists only a average altitude of most sub-basins is relatively similar, ranging from
single rainfall station at its outlet, while five rain gauges are located in 821 m to 892 m, except for sub-basins B4 and B8, which have higher
sub-basin B4, where the performance of the hydrological model was mean altitude (929 m and 1058 m, respectively). Fig. 2b shows the
superior. In addition, hydrological models have limitations related to variance of elevation normalized by the total basin area (normalized
the complexity of headwater sub-basins, as important processes such as variance, NV), which is a measure of topographic complexity in each
hydrological partitioning are poorly characterized, even by physically sub-basin with respect to the whole catchment. Most sub-basins pre-
based models (Kelleher et al., 2015). sented NV values below 20%, whereas sub-basin 8 showed a NV of
Another source of uncertainties is the extrapolation of peak flows 48%, Thus, sub-basin B8 was analyzed separately because both the
from the rating curve, particularly in headwater basins where mea- hypsometric index and the normalized variance indicated contrasting
suring discharge during severe floods is logistically difficult. In the case differences when compared to the rest of the sub-basins.
of headwater basins B6 and B8, the model was able to properly The probability density function (PDF) and the cumulative density

1400 50
(a) (b)
Fraction of normalized variance (%)

Total
1300 B1
B2 40
B3
1200
B4
Elevation (m)

B5 30
1100 B6
B7
1000 B8
20
B9
900
10
800

700 0
0 20 40 60 80 100 1 2 3 4 5 6 7 8 9
Frequency (%) Sub-basin (index)
Fig. 2. (a) Hypsometric curves for the entire Iguacu basin (in black) and for each of its sub-basins (in color). The dashed line represents the mean elevation of the
whole basin (889 m). (b) Variance of the elevation in each sub-basin normalized by the whole basin area. (For interpretation to colours in this figure, the reader is
referred to the web version of this paper.)

630
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

0.25 1
B1 B1
B2 B2
0.2 B3 0.9 B3
B4 B4
B5 B5
0.15 B6 0.8 B6

Density
Density

B7 B7
B8 B8
0.1 B9 0.7 B9

0.05 0.6

(a) (b)
0 0.5
100 101 100 101
-1 -1
Rainfall Error (mm h ) Rainfall Error (mm h )
Fig. 3. Probability density function (a) and cumulative density function (b) of the error between radar rainfall estimates and rain gauge observations for each sub-
basin. (For interpretation to colours in this figure, the reader is referred to the web version of this paper.)

Table 4 function (CDF) of the error between radar rainfall estimates and rain
Summary of regions used for calibrating the SREM2D parameters. gauge observations were constructed based on a point-to-point com-
Region Basins Topography Distance of the radar parison for all pixels where both radar and rain gauge measurements
were available. PDFs and CDFs for each sub-basin are shown in Fig. 3.
R1 B8 Complex D > 50 km Error distributions (both in terms of PDFs and CDFs) show a similar
R2 B5; B6 Complex D < 50 km
behavior in all basins, except for B5 and B6. Specifically, the PDFs of
R3 B1:B4; B7; B9 Simple D > 50 km
these two basins are significantly skewed towards lower rainfall rates
(< 1 mm/h). These two sub-basins are located close to the radar, in-
Table 5 dicating that the distance from the radar affects the error in the rainfall
Calibrated SREM2D parameters. estimates. Consequently, SREM2D was calibrated for three different
regions (Table 4). The first region (R1) corresponds to basins that are
Metrics Summer Fall
far away from the weather radar location (distance > 50 km) and are
R1 R2 R3 R1 R2 R3 topographically complex, like B8. The second region (R2) corresponds
to basins that are closer to the radar (distance < 50 km), but topo-
POD parameters A 0.10 0.13 0.09 0.16 0.19 0.15 graphically simple, like B5 and B6. The third region (R3) includes ba-
POD parameters B 0.40 0.37 0.26 0.47 0.43 0.30
sins that are far from the radar (distance > 50 km) and topographically
False alarm (1/lambda) 0.24 0.23 0.24 0.25 0.24 0.24
Mean (mean gaus. log −0.08 −0.02 −0.04 −0.37 −0.07 −0.37 simple, like B1, B2, B3, B4, B7, and B9. Due to the seasonal differences
error) of rainfall events, SREM2D was calibrated for each season (summer,
Sigma (std. dev. gaus. log 1.18 1.00 1.13 0.96 0.84 1.00 fall, winter, and spring) separately. The calibrated parameters are
error)
summarized in Table 5.
Correlation length rain det. 25 55 12 1000 1000 1000
(km)
Fig. 4 shows the cumulative rainfall time series of the SREM2D
Correlation length no rain 16 8 8 16 16 8 ensemble median (gray solid line), the SREM2D ensemble minimum
det. (km) and maximum (light gray shaded area), radar rainfall estimates (black
Correlation length retrievel 16 20 8 16 20 16 solid line), and ground observations (black dashed line) for a small sub-
det. (km)
basin in R1 (B8, Santa Cruz do Timbó, Timbó river, 2448 km2), a
Lag-one correlation 0.15 0.27 0.21 0.33 0.31 0.28
PODnorain 0.97 0.98 0.97 0.97 0.98 0.98 medium-size sub-basin in R2 (B5, São Mateus do Sul, Iguaçu river,
5440 km2), and a large sub-basin in R3 (B9, União da Vitória, Iguaçu
Metrics Winter Spring river, 21,536 km2). In the medium sub-basin, the SREM2D ensemble
reduced significantly the bias in the radar data (by ∼2000 mm), re-
R1 R2 R3 R1 R2 R3
sulting in values closer to the rain gauge time series, considered to be
POD parameters A 0.16 0.20 0.19 0.15 0.18 0.14 the benchmark. However, in the small basin, the bias reduction caused
POD parameters B 0.46 0.43 0.39 0.29 0.41 0.36 an overall underestimation of ∼1000 mm, which is due to the fact that
False alarm (1/lambda) 0.23 0.24 0.23 0.18 0.24 0.24 the bias is not constant in time and may change from the calibration to
Mean (mean gaus. log −0.73 −0.40 −0.45 −0.45 −0.19 −0.14 the validation period. Over the entire basin, SREM2D produces the best
error)
Sigma (std. dev. gaus. log 0.67 0.64 0.68 0.68 0.80 0.86
correction (∼2000 mm) and the ensemble mean almost completely
error) overlaps the reference cumulative rainfall. The large basin showed
Correlation length rain det. 1000 1000 38 1000 1000 1000 smaller ensemble spread than the small and mid-size sub-basins, due to
(km) the fact that smaller areas are more affected by topographic hetero-
Correlation length no rain 16 8 8 20 8 8
geneity, which enhances the precipitation spatio-temporal variability.
det. (km)
Correlation length retrievel 16 20 16 16 55 16
det. (km) 4.3. Streamflow time series analysis
Lag-one correlation 0.27 0.15 0.17 0.17 0.16 0.46
PODnorain 0.97 0.98 0.97 0.97 0.99 0.98
The hydrological model ability of simulating streamflow when
forced with (i) rain gauge observations (triangles), (ii) radar rainfall
estimates (stars), and (iii) SREM2D rainfall ensemble (boxplots) was

631
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

Fig. 4. Time series of cumulative rainfall for: (a) a small sub-basin (B8, Region 1); (b) a medium sub-basin (B5, Region 2); and (c) a large sub-basin (B9, Region 3)
during the SREM2D validation period. The black solid line represents the radar rainfall estimates, black dashed line represents the ground observed rainfall, the gray
solid line represents the SREM2D ensemble median, and light gray shaded areas represents SREM2D ensemble minimum and maximum.

Fig. 5. (a) NSE and (b) NSElog of streamflow simulations forced with rain gauge observations (triangle), radar rainfall estimates (star) and SREM2D ensemble
(boxplot) during the SREM2D validation period (December 2008–November 2010).

assessed through the NSE and NSElog as a function of basin area during This can be explained by a combination of factors. Firstly, when com-
2008–2010 (Fig. 5). Results showed an improvement in the statistical pared to the rain gauge observations, both radar rainfall estimates and
metrics when the SREM2D ensemble was used as input to MHD-INPE the SREM2D ensemble had similar absolute error in B8, making
instead of the radar estimates in all sub-basins, except for B8 in the case streamflow simulations insensitive to the SREM2D bias correction
of the NSE, where the performance of the ensemble was slightly worse (Fig. 4a). In addition, radar rainfall estimates were consistently larger
when compared to the results generated using radar rainfall estimates. than observations, while the ensemble members underestimating rain

632
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

Fig. 6. (a) rBIAS and (b) ubRMSE of streamflow simulated by forcing the model with rain gauge observations (triangle), radar rainfall estimates (star), and SREM2D
ensemble (boxplot) for each sub-basin during December 2008–November 2010.

gauge observations, particularly during more intense rainfall events. As using CPM. Table 6 shows the number of streamflow events in each sub-
the hydrological model tended to underestimate flow peaks, over- basin and the average properties of these events. Sixty-six events were
estimations in rainfall estimations tended to have a positive influence, identified by CPM in nine sub-basins during the period December
by correcting the model deficiency in simulating high flows. Finally, it 2008–November 2010 (Table 2).
should be noted that the extrapolation of rating curves for high flows Fig. 7 presents streamflow events from model simulations forced by
adds uncertainty, particularly in the case of small and fast response (i) radar rainfall estimates (black solid line), rain gauge observations
catchments, such as B8, where peak discharge is rarely measured. (black dashed line), the SREM2D ensemble median (gray solid line),
Systematic (rBIAS) and random (ubRMSE) errors in the simulated and minimum and maximum of the SREM2D ensemble envelope (light
streamflow time series were investigated for a cascade of basin scales gray shaded area). Three catchments of different drainage area were
(1808–21,536 km2) and are shown in Fig. 6. SREM2D was able to sig- then selected and performance statistics were computed for three dif-
nificantly remove the bias in the radar dataset regardless of the basin ferent rainfall events with contrasting behavior, resulting in nine re-
dimension, with an average bias removal of 0.5. The positive rBIAS in presentative streamflow events identified by CPM. The chosen
streamflow was also removed in B8, although it resulted in a slight streamflow events included six multi-peaks events and three single-
underestimation. The SREM2D ensemble-forced streamflow simulation peak events with recurrence period around 4.2 years in B8, 6.6 years in
improved the rBIAS in most sub-basins. On the other hand, the random B5, and 10 years in B9. These events were associated with stratiform
error, depicted by the ubRMSE, increased with the catchment area, as and/or convective precipitation. The hydrograph of each event was
larger flows are observed in larger sub-basins. In larger basins like B5, evaluated for all the simulation cases, i.e., using the three different
B7 and B9, SREM2D reduced the streamflow random error with respect rainfall inputs. Results showed that the use of SREM2D significantly
to simulations that used radar rainfall estimates as input to the hy- improved the streamflow simulations, removing the bias in the radar
drological model. rainfall product. Additionally, events in the small sub-basin (B8) in R1
and in the mid-sized sub-basin (B5) in R2 presented larger ensemble
spreads when compared to the large basin, which is due to higher
4.4. Event-based performance analysis precipitation uncertainties in regions characterized by complex terrain.
Fig. 8 shows the relative bias of rainfall and streamflow properties
In order to assess the ability of SREM2D to improve flood mon- for radar rainfall data and radar-forced simulations (stars) and for the
itoring and prediction, an event-based performance analysis was con- SREM2D ensemble mean and SREM2D ensemble mean-forced simula-
ducted. Firstly, flood events were identified during the SREM2D vali- tions (squares) for the events identified in Fig. 7. The radar dataset
dation period using rain and streamflow observations in each sub-basin significantly overestimated the cumulative rainfall volume (VR), which
led to an overestimation in the cumulative flow volume (VQ) of all the
Table 6 events. SREM2D was able to reduce the overestimation in the cumu-
Properties of observed rainfall events and observed streamflow for selected
lative rainfall volume (and corresponding streamflow volume) esti-
flood events.
mated by the radar. The largest overestimation of rainfall and stream-
Basin # Events Mean Mean Mean Mean Mean flow volumes was observed in mid-size and large sub-basins. The
rainfall streamflow time rainfall streamflow relative biases of TLAG for radar rainfall estimates and the SREM2D
duration duration (h) lag (h) volume volume
(h) (mm) (mm)
ensemble mean are close to zero, showing that both radar and SREM2D
well reproduced the time lag between mass centers of rainfall events
B1 12 495 627 147 176 124 and their corresponding hydrographs. The relative biases in the hye-
B2 5 1204 1638 329 366 212 tograph and hydrograph shape parameters (SR and SQ) are also very
B3 7 733 968 236 248 179
B4 6 1004 1360 253 253 225
close to zero. This demonstrated that both radar rainfall estimates and
B5 5 883 1170 249 267 223 the SREM2D ensemble mean were able to well represent the benchmark
B6 5 1194 1731 289 420 300 hyetograph and hydrograph shape parameters.
B7 4 1441 1955 415 439 347
B8 17 443 576 112 155 84
B9 5 1172 1553 339 375 272

633
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

Fig. 7. Streamflow events in B8 located in R1 (a, b, c), B5 located in R2 (d, e, f), and B9 located in R3 (g, h, i). Three events per sub-basin are shown. The black solid
line represents the streamflow forced by radar rainfall estimates, black dashed line represents the ground observed rainfall, the gray solid line represents the SREM2D
ensemble median, and light gray shaded areas represents SREM2D ensemble minimum and maximum.

5. Conclusions when the SREM2D ensemble was used as input to MHD-INPE in place of
radar estimates for all catchment sizes, except for a small basin in R1
This study investigated the efficiency of correcting radar rainfall where the model performance using radar estimates was already very
estimates using the SREM2D error model to simulate streamflow in the close to the one of the reference. Overall, SREM2D was able to sig-
Iguaçu basin, characterized by complex topography and seasonal nificantly remove the bias in the radar dataset regardless of the basin
rainfall variability. The MHD-INPE hydrological model was shown to size. Furthermore, the streamflow ensemble showed a reduction in the
properly simulate streamflow across the basin, when calibrated with systematic error when compared to streamflow simulated with rain
rain gauge observations. However, the model showed some limitations gauge for most sub-basins. The random error increased with the
in the headwater basins because of the high spatio-temporal pre- catchment area, but the proposed technique reduced the random error
cipitation variability in complex terrain. Since the calibration of the in larger basins.
hydrological model suffered from uncertainties related to low density Nine flood events were analyzed in the basin during the study
rain-gauge network, it is important to assess how rainfall-runoff model period and the use of SREM2D significantly improved streamflow si-
calibration results would be affected when bias-corrected radar rainfall mulations during these events. SREM2D also reduced the bias of cu-
estimates are used in place of interpolated rain-gauge measurements. mulative rainfall and streamflow volumes during the selected flood
Due to the region complex topography, the distance from the radar events. Moreover, both the radar dataset and the SREM2D ensemble
affecting the radar product quality, and precipitation seasonality, the were able to properly reproduce the flood event time lags and hyeto-
SREM2D error model was regionally calibrated during the four seasons. graph and hydrograph shape parameters.
Results showed that SREM2D reduced significantly the bias in the radar This study demonstrated the potential of using an error model –
rainfall product, particularly in medium and large size basins (5,440- traditionally applied to satellite precipitation products – to correct
21,536 km2). Improvements in the model performance were observed radar rainfall estimates for real-time streamflow monitoring/

634
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

Fig. 8. rBIAS of the rainfall and streamflow event for B8 (a, b, c), B5 (d, e, f), and B9 (g, h, i). VR is rainfall volume, VQ is streamflow volume, TLAG is time lag, SR is
dispersion of rainfall and SQ is dispersion of streamflow.

forecasting and flood risk mapping. As the methodology was tested only Vulnerabilidade_a_Inundaes.pdf&access=private.
in the Iguaçu basin, future work should verify the performance of the Anagnostou, E.N., Krajewski, W.F., Smith, J., 1999. Uncertainty quantification of mean-
areal radar-rainfall estimates. J. Atmos. Oceanic Technol. 16, 206–215. https://doi.
proposed methodology in other basins of the world to generalize these org/10.1175/1520-0426(1999) 016<0206:UQOMAR>2.0.CO;2.
results to different climatological regions. Anagnostou, M.N., Kalogiros, J., Anagnostou, E.N., Tarolli, M., Papadopoulos, A., Borga,
M., 2010. Performance evaluation of high-resolution rainfall estimation by X-band
dual-polarization radar for flash flood applications in mountainous basins. J. Hydrol.
Acknowledgments 394 (1–2), 4–16 Nov. 2010.
Anagnostou, M.N., Kalogiros, J., Nikolopoulos, E., Derin, Y., Anagnostou, E.N., Borga, M.,
This work was partially supported by the National Council for 2017. Satellite rainfall error analysis with the use of high-resolution X-band dual-
polarization radar observations over the Italian Alps. Springer, pp. 279–286. https://
Scientific and Technological Development (CNPq), Brazil and the doi.org/10.1007/978-3-319-35095-0_39.
Federal Agency for the Support and Evaluation of Graduate Education Anagnostou, M.N., Nikolopoulos, E.I., Kalogiros, J., Anagnostou, E.N., Marra, F., Borga,
(CAPES), Brazil. The authors would also like to thank the Paraná M., Mair, E., Bertoldi, G., Tappeiner, U., 2018. Advancing precipitation estimation
and streamflow simulations in complex terrain with X-band dual-polarization radar
Meteorological System (SIMEPAR), Brazil and the National Institute for
observations. Rem. Sens MDPI.
Space Research (CPTEC-INPE), Brazil, and the “Rede Sul Brasileira de Casagrande, L., Tomasella, J., dos Santos Alvalá, R.C., et al., 2017. Nat. Hazards 88, 741.
pesquisas sobre mudanças climáticas e prevenção aos desastres naturais https://doi.org/10.1007/s11069-017-2889-0.
(REDE CLIMASUL)”, Brazil and “Desenvolvimento de Sistema de Caseri, A., Javelle, P., Ramos, M.H., Leblois, E., 2016. Generating precipitation ensembles
for flood alert and risk management. J. Flood Risk Manage. 9, 402–415. https://doi.
Previsão de Enxurradas, Inundações e Movimentos de Massa em org/10.1111/jfr3.12203.
Encostas para Prevenção de Desastres Naturais”, Brazil projects. Cecinati, F., Rico-Ramirez, M.A., Heuvelink, G.B.M., Han, D., 2017. Representing radar
rainfall uncertainty with ensembles based on a time-variant geostatistical error
modelling approach. J. Hydrol. 548, 391–405. https://doi.org/10.1016/j.jhydrol.
Appendix 2017.02.053. ISSN 0022-1694.
Ciach, G.J., Krajewski, W.F., Anagnostou, E.N., Baeck, M.L., Smith, J.A., McCollum, J.R.,
See Tables 5 and 6. Kruger, A., 1997. Radar rainfall estimation for ground validation studies of the tro-
pical rainfall measuring mission. J. Appl. Meteor. 36, 735–747. https://doi.org/10.
1175/1520-0450-36.6.735.
References Codo, M., Rico-Ramirez, M., 2018. Ensemble radar-based rainfall forecasts for urban
hydrological applications. Geosciences 8 (8), 297. https://doi.org/10.3390/
geosciences8080297.
ÁGUASPARANÁ, Instituto das Águas do Paraná – Bacia Hidrográfica do Rio Iguaçu.
COPEL, 2017. Retrieved from https://www.copel.com/mhbweb/paginas/bacia-iguacu.
Retrieved from http://www.aguasparana.pr.gov.br/arquivos/File/BACIAS/iguacu.
jsf.
pdf.
Derin, Y., Anagnostou, E.N., Anagnostou, M.N., Kalogiros, J., Casella, D., Marra, A.C.,
Alvares, C.A., Stape, J.L., Sentelhas, P.C., de Gonçalves, J.L.M., Sparovek, G., 2013.
Panegrossi, G., Sanò, P., 2018. Passive microwave rainfall error analysis using high-
Koppen’s climate classification map for Brazil. Meteorologische Zeitschrift 22 (6),
resolution X-band dual-polarization radar observations in complex terrain. IEEE
711–728. https://doi.org/10.1127/0941-2948/2013/0507.
Trans. Geosci. Rem. Sens. https://doi.org/10.1109/TGRS.2017.2763622.
ANA, Atlas de Vulnerabilidade a inundações; 2014. Retrieved from http://metadados.
Duan, Q., Sorooshian, S., Gupta, H.V., 1992. Effective and efficient global optimization
ana.gov.br/geonetwork/srv/en/resources.get?id=243&fname=Atlas_de_
for conceptual rainfall-runoff models. Water Resour. Res. 28, 1015–1031. https://doi.

635
A.S. Falck et al. Journal of Hydrology 567 (2018) 626–636

org/10.1029/91WR02985. from rainfall and runoff records. J. Hydrol. 523, 636–649.


Empresa Brasileira de Pesquisa Agropecuária (EMBRAPA), 1981. Soil Map of Brazil at 1:5, Mohor, G.S., Rodriguez, D.A., Tomasella, J., Siqueira Júnior, J.L., 2015. Exploratory
000,000 Scale. Serviço Nacional de Levantamento e Conservação de Solos/Embrapa, analyses for the assessment of climate change impacts on the energy production in an
Rio de Janeiro, RJ. Brazil (in Portuguese). Amazon run-of-river hydropower plant. J. Hydrol.: Reg. Stud. 4, 41–59. https://doi.
Falck, A.S., Maggioni, V., Tomasella, J., Vila, D.A., Diniz, F.L.R., 2015. Propagation of org/10.1016/j.ejrh.2015.04.003.
satellite precipitation uncertainties through a distributed hydrologic model: a case Mohor, G.S., Mendiondo, E.M., 2017. Economic indicators of hydrologic drought in-
study in the Tocantins-Araguaia basin in Brazil. J. Hydrol. 527, 943–957. https://doi. surance under water demand and climate change scenarios in a Brazilian context.
org/10.1016/j.jhydrol.2015.05.042. ISSN 0022-1694. Ecol. Econ. 140, 66–78. https://doi.org/10.1016/j.ecolecon.2017.04.014.
Falck, A.S., Vila, D.A., Tomasella, J., Maggioni, V., Diniz, F.L.R., 2016. Avaliação de um Moriasi, D.N., et al., 2007. Model evaluation guidelines for systematic quantification of
Modelo Estocástico de Erro Multidimensional Aplicado a Estimativas de Precipitação accuracy in watershed simulations. Trans. ASABE 50, 885–900. https://doi.org/10.
por Satélite. Revista Brasileira de Meteorologia 31 (52–63), 2016. https://doi.org/10. 13031/2013.23153.
1590/0102-778620140042. Nerini, D., Besic, N., Sideris, I., Germann, U., Foresti, L., 2017. A non-stationary stochastic
Farr, T.G., Rosen, P.A., Caro, E., Crippen, R., Duren, R., Hens-Ley, S., Kobrick, M., Paller, ensemble generator for radar rainfall fields based on the short-space Fourier trans-
M., Rodriguez, E., Roth, L., Seal, D., Shaffer, S., Shimada, J., Umland, J., Werner, M., form. Hydrol. Earth Syst. Sci. 21, 2777–2797. https://doi.org/10.5194/hess-21-
Oskin, M., Burbank, D., Alsdorf, D., 2007. The shuttle radar topography mission. Rev. 2777-2017.
Geophys. 45 (2). https://doi.org/10.1029/2005RG000183. ISSN 1944-9208n/a–n/a. Nikolopoulos, E.I., Anagnostou, E.N., Hossain, F., Gebremichael, M., Borga, M., 2010.
Garcia, J.R., 2016. Avaliação do agravamento de inundações na bacia hidrográfica do Understanding the scale relationships of uncertainty propagation of satellite rainfall
Alto Iguaçu e afluentes do Alto Ribeira, Paraná. Retrieved from http://journals. through a distributed hydrologic model. J. Hydrometeorol. 11, 520–532. https://doi.
openedition.org/confins/11025. org/10.1175/2009JHM1169.1.
Hong, Y., Hsu, K.-L., Moradkhani, H., Sorooshian, S., 2006. Uncertainty quantification of Peleg, N., Marra, F., Fatichi, S., Paschalis, A., Molnar, P., Burlando, P., 2018. Spatial
satellite precipitation estimation and Monte Carlo assessment of the error propaga- variability of extreme rainfall at radar subpixel scale. J. Hydrol. 556, 922–933.
tion into hydrologic response. Water Resour. Res. 42, W08421. https://doi.org/10. https://doi.org/10.1016/j.jhydrol.2016.05.033. ISSN 0022-1694.
1029/2005WR004398. Pisani, M.A.J., Bruna, G.C., 2011. How to minimize urban flooding: permissible archi-
Hossain, F., Anagnostou, E.N., 2006. Assessment of a multi-dimensional satellite rainfall tecture and urban planning. Engevista 13 (1), 40–45.
error model for ensemble generation of satellite rainfall data. Geosci. Remote Sens. Rodriguez, D.A., Tomasella, J., 2016. On the ability of large-scale hydrological models to
Lett. (GRSL) 3 (3), 419–423. simulate land use and land cover change impacts in Amazonian basins. Hydrol. Sci. J.
Hossain, F., Tang, L., Anagnostou, E.N., Nikopoulos, E.I., 2009. A Pratical guide to a 61 (10), 1831–1846. https://doi.org/10.1080/02626667.2015.1051979.
space-time stochastic error model for simulation of high resolution satellite data, Sestini, M.F., Alvalá, R.C.S., Mello, E.M.K. et al., 2002. Elaboração de mapas de vegetação
2009, http://iweb.tntech.edu/fhossain/papers/SREM2DGuide.pdf. para utilização em modelos meteorológicos e hidrológicos. São José dos Campos,
IBGE (Instituto Brasileiro de Geografia e Estatística), 2015. Retrieved from https:// INPE. Retrieved from http://urlib.net/rep/sid.inpe.br/marciana/2003/03.05.15.05?
cidades.ibge.gov.br/brasil/pr/curitiba/panorama. ibiurl.language=pt-BR.
Kelleher, C., Wagener, T., McGlynn, B., 2015. Model-based analysis of the influence of Shen, X., Mei, Y., Anagnostou, E.N., 2017. A comprehensive database of flood events in
catchment properties on hydrologic partitioning across five mountain headwater the contiguous United States from 2002 to 2013. Bull. Amer. Meteor. Soc. 98,
subcatchments. Water Resour. Res. 51, 4109–4136. https://doi.org/10.1002/ 1493–1502. https://doi.org/10.1175/BAMS-D-16-0125.1.
2014WR016147. Simpson, M.J., Fox, N.I., 2018. Dual-polarized quantitative precipitation estimation as a
Lefèvre, M., Remund, J., Albuisson, M., Wald, L., 2002. Study of effective distances for function of range. Hydrol. Earth Syst. Sci. 22, 3375–3389. https://doi.org/10.5194/
interpolation schemes in meteorology Annual Assembly, European Geophysical hess-22-3375-2018.
Society, Nice, April 2002. Geophysical Research Abstracts 4: EGS02-A-03429. Siqueira Junior, L., Tomasella, J., Rodriguez, D.A., 2015. Impacts of future climatic and
Liguori, S., Rico-Ramirez, M.A., 2013. A review of current approaches to radar-based land cover changes on the hydrological regime of the Madeira river basin. Clim.
quantitative precipitation forecasts. Int. J. River Basin Manage. 12 (4), 391–402. Change 129, 117–129. https://doi.org/10.1007/s10584-015-1338-x.
https://doi.org/10.1080/15715124.2013.848872. Tomasella, J., Gonçalves, A.S., Falck, A.S., Caram, R.O., Diniz, F.L.R., Rodriguez, D.A.,
Maggioni, V., Vergara, H.J., Anagnostou, E.N., Gourley, J.J., Hong, Y., Stampoulis, D., Prado, M.C.R., Negrão, A.C., Medeiros, G.S., Siqueira, G.C., 2018.. Probabilistic flood
2013. Investigating the applicability of error correction ensembles of satellite rainfall forecasting in the Doce Basin in Brazil: effects of the basin scale and orientation and
products in river flow simulations. J. Hydrometeorol. https://doi.org/10.1175/JHM- the spatial distribution of rainfall. J. Flood Risk Manage. e12452. https://doi.org/10.
D-12-074.1. 1111/jfr3.12452.
Mei, Y., Anagnostou, E.N., 2015. A hydrograph separation method based on information

636

You might also like