This manuscript is a preprint currently under review in Journal of Hydrology.

Hence, its final

accepted version may be different from the current one. Once the manuscript will be fully
published the corresponding DOI link will be added on the right-hand side of this webpage. Please,
feel free to contact the corresponding author if you have any feedback

1
1 Evaluating precipitation datasets for large-scale distributed
2 hydrological modelling
3 M. Mazzoleni1,2, L. Brandimarte 3, A. Amaranto4

4
1
5 Department of Earth Sciences, Uppsala University, Uppsala, 75236, Sweden
6 2 Centre of Natural Hazards and Disaster Science (CNDS), Sweden
7 3 KTH Royal Institute of Technology, Department of Sustainable Development, Environmental
8 Sciences and Engineering, Stockholm, Sweden
9 4 Department of Electronics, Information, and Bioengineering, Politecnico di Milano, Piazza
10 Leonardo da Vinci, Milano, Italy
11

12 Corresponding author: Maurizio Mazzoleni ([Link]@[Link])

13 Keywords: Remote sensing, Distributed hydrological modelling, Precipitation datasets, large-

14 scale

15 Abstract

16 Over the past decades, a variety of valuable research studies has helped to advance our
17 understanding of the advantages and limitations of satellite derived precipitation datasets as a
18 forcing to hydrological models, in combination with or as an alternative to gauge data.
19 However, most studies have assessed the performance of only one single dataset (or a few),
20 have used global precipitation datasets to force lumped models on regional/large-scale basins or
21 have tested more complex distributed models only at small-scale basins. In addition, only few
22 studies have re-calibrated the model for each precipitation dataset or have investigated reanalysis-
23 based precipitation datasets.
24 We aimed at addressing these gaps in the literature: in particular, we compared the performance
25 of 18 different precipitation datasets when used as main forcing in a grid-based distributed
26 hydrological model to assess streamflow in medium to large-scale river basins. These datasets are
27 classified as Uncorrected Satellites (Class 1), Corrected Satellites (Class 2) and Reanalysis -

1
28 Gauges based datasets (Class 3). To provide a broad-based analysis, 8 large-scale river basins
29 (Amazon, Brahmaputra, Congo, Danube, Godavari, Mississippi, Rhine and Volga) having
30 different sizes, hydrometeorological characteristics, and human influence were selected. The
31 distributed hydrological model was recalibrated for each precipitation dataset individually.
32 We found that there is not a unique best performing precipitation dataset for all basins and that
33 results are very sensitive to the basin characteristics. However, a few datasets persistently
34 outperform the others: SM2RAIN-ASCAT for Class 1, CHIRPS V2.0, MSWEP V2.1, and
35 CMORPH-CRTV1.0 for Class 2, GPCC and WFEDEI GPCC for Class 3. Surprisingly,
36 precipitation datasets showing the highest model accuracy at basin outlets do not show the same
37 high performance in internal locations, supporting the use of distributed modelling approach rather
38 than lumped.

39 Introduction

40 In a recent study about worldwide information on precipitation ground measurements, Kidd et

41 al. (2017) estimated that “The total area measured globally by all currently available rain gauges
42 is surprisingly small, equivalent to less than half a football field or soccer pitch”. This limited
43 gauge representativeness, the scarce and unequal spatial distribution of rain gauges (Maggioni and
44 Massari, 2018) and the concern for the global decline of in-situ hydrologic measurements
45 (Stokstad, 1999; Shiklomanov et al., 2002) have motivated increasing attention on the
46 potentialities offered by the growing availability of satellite-retrieved precipitation products as an
47 alternative source of input data in hydrological modelling. In particular, the interest for satellite
48 products has grown over the past decade, with the increase in their temporal and spatial resolutions
49 (Stephens and Kummerow, 2007; Kidd and Huffman, 2011; Xie and Xiong, 2011; Brocca et al.,
50 2013; Funk et al., 2015; Duan et al., 2016; Beck et al., 2017a).
51 Since the first studies exploring the potentialities of incorporating satellite-based precipitation
52 data (e.g., Barrett and Martin, 1981; Schulz, 1996; Tsintikidis et al., 1999) to the latest global-scale
53 comprehensive evaluation of a number of precipitation datasets (Beck et al., 2017b), over the past
54 three decades the scientific literature has produced a variety of valuable research works that have
55 advanced our understanding of the advantages and limitations of satellite-derived precipitation
56 information in hydrologic modelling. Recently, Maggioni and Massarri (2018) proposed a

2
57 comprehensive review of previous studies on satellite-based precipitation input forcing
58 hydrological models.
59 The performance of precipitation datasets in hydrological applications has been assessed by
60 either comparing simulated and observed soil moisture (e.g., Brocca et al., 2013) or observed river
61 discharge. Here we focus on the latter and group previous studies based on common approaches
62 in performing their assessment.
63 Many studies focused their analysis on assessing the performance of a single precipitation
64 dataset in hydrological modelling (e.g., Artna et al., 2007; Collischonn et al., 2008; Yang et al.,
65 2017) or comparing the performance of few precipitation datasets in streamflow simulations (e.g.,
66 Yong et al., 2010; Bitew and Gebremichael, 2011; Behrangi et al., 2011; Falck et al., 2015, Camici
67 et al., 2018), thus limiting their analysis to specific products.
68 Often, reanalysis-based precipitation datasets were not taken into account (e.g., Moazami et
69 al., 2013; Zambrano-Bigiarini et al., 2017), or models were not re-calibrated for each precipitation
70 dataset (e.g., Voisin et al., 2008; Su et al., 2008; Li et al., 2013), thus missing to compare
71 uncorrected satellite products with the ones based on in-situ precipitation network.
72 Also, most of the previous studies have based the evaluation of precipitation datasets on
73 hydrological models that are lumped/conceptual (e.g., Behrangi et al., 2011; Essou et al., 2016),
74 thus not accounting for the inherent spatial variability of river basin characteristics that are
75 averaged over the watershed (e.g., Boyle et al., 2001; Carpenter and Georgakakos, 2006).
76 On the other hand, studies implementing fully or semi distributed hydrological modelling refer
77 only to a single specific case study (e.g., Stisen and Sandholt, 2010; Bitew et al., 2012; Pedinotti
78 et al., 2012; Casse et al., 2015; Falck et al., 2015; Wang et al., 2015; Tang et al., 2016; Camici et
79 al., 2018), or at sub-continental scale (e.g., Su et al. 2008; Li et al., 2013; Essou et al., 2016), thus
80 missing the opportunity to generalize the results.
81 Furthermore, only a limited number of applications have explored the suitability of satellite-
82 retrieved precipitation products in global/large-scale hydrological modelling (Fekete et al.; 2004
83 Voisin et al., 2008). A recent contribution from Beck et al. (2017b) performed a comprehensive
84 global evaluation of 22 precipitation datasets using gauge observations and hydrological modeling.
85 Nevertheless, these analyses did not explore the full range of available precipitation products
86 (i.e.,Voisin et al., 2008), whose offer over the past fifteen years has greatly increased, and are

3
 87 performed based on lumped models (Beck et al., 2017b) or did not compare the results of the
88 hydrological simulations with observed values (Fekete et al., 2004).
89 Our research questions emerged from the consideration that, notwithstanding the intrinsic
90 uncertainty and errors associated to satellite retrieved precipitation data (e.g., Hossain and
91 Anagnostou, 2004; Gottschalck et al., 2005; Hossain and Lettenmaier 2006; Hong et al., 2006;
92 Ebert et al., 2007; Tian and Peters-Lidard 2010; Stampoulis and Anagnostou, 2012; Libertino et
93 al., 2016; Maggioni and Massari, 2018), they represent a unique opportunity for hydrologic
94 applications and, in particular, they can potentially spark light on improved flow estimation in data
95 scarce or data poor basins (Serrat-Capdevilla et al., 2014).
96 We thus posed the following research questions, drawn from the highlighted gaps in the current
97 literature:

98 a) How does precipitation estimation from satellite data using different products compare when
99 forcing a distributed hydrologic model to estimate river discharge in basins with different
100 spatial scales, climatic zones and human influence (e.g. presence of reservoirs)?
101 b) How does the density of the precipitation monitoring network used to correct (some)
102 precipitation datasets affect river discharge estimation in different river basins?
103 c) Do spatial details gained by using distributed hydrological modelling and different
104 precipitation datasets lead to an improved representation of the river flow within the basin?
105 d) Is there a specific dataset that always outperforms the other ones at both outlet and internal
106 basin locations?

107 To reply to these research questions, we investigated the behaviour of 18 different alternative
108 sources of rainfall data on estimating river discharges over large areas. In particular, we tested
109 three classes of (quasi-) global precipitation dataset as main forcing of a grid-based distributed
110 hydrological model: 1) satellite-based rainfall products; 2) gauge-corrected satellite rainfall
111 products; and 3) reanalysis and gauge measured rainfall data. Eight large-scale river basins of
112 different size and hydrometereological characteristics are used as case studies: Amazon,
113 Brahmaputra, Congo, Danube, Godavari, Mississippi, Rhine and Volga river basins.

4
114 Material

115 Cases studies

116 The eight river basins (see Figure 1) were selected based on different basin size,
117 hydrometereological and climatic characteristics, presence of hydraulic structures and density of
118 precipitation gauge network. We thus selected mid to large-scale river basins belonging to different
119 continents and covering different climatic zones (Kottek et al. 2006). The Danube, Rhine and
120 Mississippi basins have a dense network of precipitation gauges and a significant presence of dams
121 and reservoirs which alter the natural hydrological response of the river basins. On the other hand,
122 the Amazon, Congo and Volga basins can be considered less affected by human interventions.
123 Table 1 (Molinier et al., 1993; Gaillardet et al, 1997; Wieriks and Schulte-Wulwer-Leidig, 1997;
124 Goolsby and Battaglin, 2001; Immerzeel, 2008; Jha et al., 2009; Csagoly et al., 2016; Harrison et
125 al., 2016) summarizes the main characteristics of each river, including climactic zones (Peel et al.,
126 2007), size of the drainage area, length of the main river, average river flow, number of flow
127 monitoring stations used to evaluate model performances (GRDC, 2018), number of reservoirs
128 from the GRanDv1 dataset (Lehner et al., 2011) and percentage of human footprint (i.e., human
129 population density, land transformation, electrical power infrastructure and access to land data, as
130 reported in Center for International Earth Science Information Network, CIESIN) (Kareiva et al.,
131 2007).

132 Precipitation datasets

133 We assessed the performance of 18 gridded precipitation datasets in hydrological modelling

134 applications. In order to provide a fair comparison between datasets, we decided to classify them
135 according to their data source (see Table 2). Among datasets exclusively based on uncorrected
136 satellite data (Class 1) we analysed CHIRP V2.0 (Funk et al., 2015), CMORPH V1.0 (Joyce et al.,
137 2004), PERSIANN (Sorooshian et al., 2000), PERSIANN-CCS (Hong et al., 2004), SM2RAIN-
138 ASCAT (Brocca et al., 2013), and TMPA 3B42 RT V7 (Huffman et al., 2007). In most
139 applications, datasets of Class 1 were used as Near Real Time (NRT) products. We grouped in
140 Class 2 gauge-corrected datasets: CHIRPS V2.0 (Funk et al., 2015), CMORPH-CRT V1.0 (Joyce
141 et al., 2004), GPCP1DD V1.2 (Huffman et al., 2001), MSWEP V2.1 (Beck et al. 2017a),
142 PERSIANN-CDR (Ashouri et al., 2015), and TMPA 3B42 V7 (Huffman et al., 2007). Finally,

5
143 CPC Global Unified (Chen et al., 2008), GPCC (Schamm et al., 2014), GSMaP-RNL (Iguchi et
144 al., 2009), PFD (Sheffield et al., 2006), WFEDEI CRU (Weedon et al., 2014), and WFEDEI GPCC
145 (Weedon et al., 2014) belong to Class 3, reanalysis and gauge based datasets. It is worth noting
146 that these datasets have different spatial and temporal resolutions (see Table 2). The precipitation
147 datasets are resampled to a spatial resolution of 0.25 degrees and a daily temporal resolution (Liu
148 et al., 2017). In addition, because of the different temporal coverage of each dataset, the
149 overlapping period between January 1st, 2007 and December 31st, 2013 was selected for calibration
150 and validation analyses. A detailed review of different precipitation products is provided by Sun
151 et al. (2018).

152 Discharge dataset

153 We calibrated and validated the hydrological models using river discharge data provided by
154 the Global Runoff Data Centre (GRDC; [Link] The GRDC centre collects
155 river discharge data for more than 9500 stations from 161 countries. The collection period varies
156 spatially: the earliest data were from 1807, while the most recent were from 2018. In this study,
157 the daily discharge data (from 2007 to 2013) for 46 sensors across 8 different river basins (as
158 shown in Figure 1) were extracted from the whole GRDC database.

159 In particular, information from flow sensors located at the outlet of the 8 river basins were used
160 to calibrate and validate the hydrological model. Flow values at internal sensors were used in the
161 validation process of 5 of the 8 basins (Amazon, Danube, Godavari, Mississippi and Rhine) for
162 which GRDC flow data were available within the catchment for the simulated period.

163 Methodology

164 Distributed hydrological model

165 In this study, a grid-based hydrological model was developed to spatially estimate the flow
166 within river catchments with different spatial scales. Figure 2 shows a schematic representation of
167 the proposed modelling framework. For each grid (0.25º × 0.25º in this study), a conceptual HBV-
168 96 model (Bergström, 1992) was implemented to represent the rainfall-runoff processes and its
169 grid output was then routed downstream using the widely known Muskingum model (Cunge

6
170 1969). The choice of the HBV model was driven by its computational efficiency and its ability to
171 provide accurate forecast under a wide range of climatic conditions (e.g. Merz and Bloschl, 2004;
172 Bardossy, 2007; Jin et al., 2009; Demirel et al., 2015; Vetter et al., 2015). The connectivity between
173 grids was estimated using the flow direction and flow accumulation dataset developed by Wu et
174 al. (2011 and 2012).
175 The modelling steps implemented to calculate the discharge accumulation over the drainage
176 network at time step t are:

177 1- Calculate the flow contribute QR generated in each grid of the basin, as response of rainfall-
178 runoff processes using a conceptual lumped hydrological model.

179 2- Flows generated at the grids with lowest flow accumulation value (the most upstream part of
180 the catchment) were propagated downstream using the Muskingum routing model following
181 the connectivity characteristics (indicated as QP in Figure 2).

182 3- The total discharge in each grid was calculated: let’s consider cell N, showed in Figure 2,
183 located downstream of cells i and j. Once the upstream flows from grids i and j were routed
184 at the downstream cell N, the total discharge at cell N was calculated as:

(1)
QTN  Q RN  Q Pi  Q Pj

185 where QR is the generated flow from the conceptual hydrological model, QP is the propagated
186 flow from the upstream grids and QT is the total flow at the cell N.

187 4- Discharge QT at grid N is then propagated at the downstream grids.

188 5- Steps from 2 to 4 were sequentially repeated for each flow accumulation values up to the
189 highest one (i.e. the grid related to the basin outlet), in order to calculate the distributed
190 discharge accumulation QC within the catchment at the time step t.

191 A similar modelling framework is the widely used PCRaster software environment developed
192 by Karssenberg et al. (2009) for constructing iterative spatiotemporal environmental and
193 hydrological models (Bloschl et al, 2008; Cole and Moore, 2009; Thielen et al., 2009; PCraster,
194 2018).

195 It is worth noting that model states were initialized by running the model twice for the entire
196 record. Despite the effect of model’s initial condition has been widely discussed in the literature

7
197 (see for example Goodrich et al., 1994; Minet et al., 2011; Seck et al., 2015) there is still no clear
198 and automatic rule to determine the optimal spin-up time in hydrological models. However,
199 according to Rahman et al., (2016), up to date modelling exercises are currently performed by
200 running the simulation recursively through a specific period (which is typically a year) or by
201 running the model multiple times for different climatological conditions. For this reason, we
202 decided to initialize the states using two model runs. In addition, model states in each grid are
203 independent from the neighbouring cells, and interactions occur only in the propagation of the
204 generated total discharge.
205 For more information about the version of the HBV model implemented in this study, the
206 readers are referred to Seiber and Vis (2002) and Beck et al. (2016).

207 Performance measures

208 The performance of the distributed hydrological model forced by the different precipitation
209 input was calculated by means of the Nash Sutcliffe efficiency (NSE; Nash and Sutcliffe, 1970),
210 and Bias index, widely used among hydrologists (Moriasi et al., 2007). The reason for using NSE
211 instead of other indices as Root Mean Square Error, Pearson coefficient, or the Kling-Gupta
212 efficiency was that NSE is highly sensitive to peak flow values (Krause et al., 2005; Beck et al.,
213 2017b). In fact, the correct simulation of peak flows is highly dependent on different precipitation
214 forcing used in hydrological modelling, which is the main objective of this study.
215 NSE value of 1 represents a perfect model simulation, while NSE value equal to 0 indicates that
216 the model is as accurate as the mean of the observed flows. Ritter and Muñoz-Carpena (2013)
217 indicated that models scoring NSE≥0.65 provide acceptable results.
218 The Bias index was selected to assess the tendency of the model to overestimate (Bias index
219 greater than 1) or underestimate (Bias index smaller than 1) observed flow observations. The Bias
220 index is calculated as the ratio between the mean of the simulated and observed flow values.

221 Model calibration

222 One of the main challenges in large-scale hydrological modelling is the proper estimation of
223 model parameters (Anderton et al 2002). In the grid-based distributed model proposed in this study
224 14 model parameters need to be calibrated for each grid cell: 12 parameters from the HBV model
225 and 2 from the Muskingum routing model. Here, we used the global regionalized dataset developed

8
226 by Beck et al. (2016), spatially distributed and with a 0.25 degree resolution, to assign the initial
227 values of the 12 parameters of the HBV model. These datasets were then individually perturbed
228 by a correction coefficient, in order to estimate the optimal set of parameters for each different
229 precipitation product and river basin. We re-calibrated the hydrological model for each
230 precipitation dataset in order to get an unbiased and comprehensive comparison among the
231 datasets.
232 The 14 correction coefficients were calibrated by means of the least squares minimization
233 technique using the Broyden-Fletcher-Goldfarb-Shanno variant of the Davidon-Fletcher-Powell
234 minimization (DFPMIN) algorithm (Press et al., 1992). In particular, we aimed at maximizing the
235 NSE between the observed and simulated discharge values at the outlet, for all the considered
236 basins and precipitation datasets. This approach may be affected by “equifinality problem” (Beven
237 2001; 2002; Brooks et al., 2007) for large-scale basins in which calibration is performed using
238 only outlet discharge instead than flow measurements at internal points. However, due to the
239 limited flow information available at sensor locations within the catchments, we decided to use
240 these measurements for model validation rather than calibration purposes.
241 The length and characteristic of the flood events used for model calibration have great
242 influence on the estimation of the optimal model parameters. The calibration and validation
243 periods are shown in Table 3. It can be noticed that for some basins the calibration period is
244 posterior to the validation one, and vice versa: flood events included in the calibration period were
245 specifically selected in order to properly represent the hydrological variability of the basin, both
246 in terms of high and low flow. In fact, previous studies demonstrated that applying a hydrological
247 model in flow conditions different than the ones in calibration would result in unreliable model
248 performances (Seibert 2003; Singh and Bardossy, 2012; Brigode et al., 2013). The validation of
249 the distributed hydrological model was then performed on independent periods. The results of the
250 calibration and validation analyses are showed in the next sections.
251

9
252 Results

253 Model calibration results

254 Following the approach described in the previous section, the distributed hydrological model
255 was calibrated for the 18 precipitation datasets, for each of the 8 river basins: 144 optimal sets of
256 perturbation model parameters were calculated. The results of the calibration phase are first shown
257 by class type of dataset (i.e., Class 1: satellite-based; Class 2: gauge-corrected and Class 3:
258 reanalysis/gauge measured) and then individually for each precipitation dataset.
259 Figure 3 shows the average simulated flow for each dataset class against the observed flow
260 hydrographs, at each basin’s outlet. A qualitative assessment of the simulated hydrographs
261 indicated that the proposed distributed model tends to properly represent the outlet flow when
262 compared to the observed one. The model was able to correctly represent peaks and flood timing
263 also in case of high flow variability, like on the Mississippi, Danube and Rhine basins (which are
264 highly regulated basins as showed in Table 1).
265 Table 4 shows the average NSE values obtained comparing simulated and observed outlet
266 flows. It can be observed that corrected satellites (Class 2) and reanalysis-gauges products (Class
267 3), provided the best model performance results. Class 1 outperformed the other two classes on
268 the Amazon basin, while on the Brahmaputra basin all classes showed similar NSE values. This
269 result may encourage the use of NRT products for river flow simulation on those large-scale basins
270 in which no rain gauges can be used to correct satellite products of precipitation. On average, Class
271 2 provided the best results on basins characterized by a similar Tropical and Temperate-Arid
272 climate like the Congo, Godavari and Mississippi basins (Peel et al., 2007). On the other hand,
273 Class 3 products gave better NSE values on basins with Temperate and Temperate-Cold climate
274 (e.g. Danube, Rhine and Volga) characterized by high values of human footprint.
275 The above results are related to the average NSE value over the datasets of each class type;
276 Figure 4 shows the NSE values for each individual precipitation datasets analysed. It can be
277 observed that no dataset outperforms the others, when comparing simulated and observed flows at
278 basins outlets.
279 By looking at individual products in each class, we can conclude that CHIRP V2.0 (Class 1),
280 CHIRPS V2.0 (Class 2) and CPC Global Unified (Class 3) are the datasets that provided the
281 highest NSE values, equal to 0.74, 0.77, and 0.78, respectively, when averaged over the 8 river

10
282 basins. In contrast, the lowest average NSE values of 0.21, 0.65 and 0.54 were obtained with
283 PERSIANN (Class 1), CMOPRH-CRT V1.0 (Class 2) and GSMaP-RNL (Class 3) respectively.
284 These values refer to average NSE performance and the conclusions can change if river basins are
285 considered individually. For example, the highest NSE values for the Brahmaputra basin were
286 equal to 0.87, 0.88 and 0.89 for SM2RAIN-ASCAT (Class 1), MSWEP V2.1 (Class 2) and
287 WFEDEI GPCC (Class 3). If we look at Figure 4 in terms of class behaviour, we can see that
288 models forced by Class 1 NRT products tend to provide higher variability in NSE values and an
289 overall mean NSE lower than the one obtained in case of Classes 2 and 3 products; while
290 comparable performance results can be found between Class 2 and 3 products.

291 Model validation results at basins outlets

292 We validated the distributed hydrological model at basin outlets for time periods and durations
293 different than in calibration. We present here our findings focusing on: a) model performance by
294 forcing dataset over the eight case studies and b) model performance by river basin, highlighting
295 the most and less performing forcing datasets.

296 a) Analysis of the model performance by forcing dataset

297 NSE values of each dataset obtained across the 8 basins are represented as boxplots in Figure
298 5. Overall, there is no a dataset outperforming the other ones. MSWEP V2.1 and GPCC were the
299 datasets that provided the highest average NSE values for Classes 2 and 3, SM2RAIN-ASCAT for
300 Class 1. Class 1 gave the highest variability of model results, which can be related to the difficulties
301 of these products to properly represent precipitation in areas with complex topography and
302 characterized by high spatiotemporal variability (Derin and Yilmaz, 2014). Among them,
303 PERSIANN was the dataset providing the highest variability and the lowest NSE when used as
304 input data. This can be due to the tendency of PERSIANN to overestimate observed precipitation
305 (Tian et al., 2009), in particular in North America as demonstrated by Sun et al. (2018). SM2RAIN-
306 ASCAT and CHIRP V2.0 were the datasets (among the ones in Class 1) which lead to the best
307 NSE (measured in terms of median value over the 8 basins) and lowest variability, respectively.
308 Similar results were achieved by the distributed model forced with CMORPH V1.0 and
309 PERSIANN-CCS. On average, datasets from Classes 2 and 3 gave the highest median NSE values
310 and lowest variability of the performance index. A part from CMORPH-CRT V1.0 and TMPA

11
311 3B42 V7, datasets of Class 2 provided similar NSE results, with MSWEP V2.1 outperforming all
312 the other datasets in the class. This might be due to the fact that MSWEP V2.1 is developed by
313 merging optimally daily gauge data, multiple satellite and reanalysis precipitation datasets.
314 Instead, the other products of Class 2 are only based on in-situ gauges with coarser temporal
315 resolution (Beck et al. 2017a). Regarding Class 3, the use of GSMaP-RNL as input in the
316 distributed model produced low and more variable model performances if compared to other
317 products in the same class. Comparable NSE values were obtained using CPC Global Unified,
318 GPCC, PFD and WFEDEI CRU. As expected, the corrected satellite products CHIRPS V2.0,
319 PERSIANN-CDR and TMPA 3B42 V7 gave higher model results and lower variability than their
320 corresponding raw and real-time products CHIRP V2.0, PERSIANN-CCS and TMPA 3B42 RT
321 V7 (Xie et al., 2017; Hussain et al., 2018). In general, we found that reanalysis products exhibit
322 better performance at latitudes dominated by intense, localized convective precipitation systems,
323 in agreement with Beck et al., 2017b.

324 b) Analysis of the model performance by river basin

325 Figure 6 summarizes the results obtained in validation, focusing on model performance for
326 each basin: datasets providing the highest and lowest NSE values per each class are highlighted.
327 The model was able to properly represent river flow at the Amazon basin outlet due to the
328 periodical trend of the flow hydrograph. High median values of NSE were obtained for medium
329 scale basins such as Rhine, Brahmaputra and Godavari. In addition, we can observe a higher
330 variability of NSE values in basins with strong human impact, both in terms of number of
331 reservoirs (e.g. Mississippi, Danube and Rhine) and human footprint (e.g. Volga). Low variability
332 of the NSE values was obtained for the Amazon, Brahmaputra and Godavari basins. This can be
333 related to the good ability of the distributed hydrological model to represent river flow in basins
334 with low human influence (e.g. Amazon) and smaller size basin (e.g. Brahmaputra and Godavari).
335 For what concerns the comparison among different classes, Figure 6 shows that, when using
336 uncorrected satellite products (Class 1, green bars in the figure) as input for the hydrological
337 model, the highest variability in modelling performances was obtained in most of the basins. For
338 example, high performance variability was obtained on the Mississippi basin using Class 1
339 datasets. This could be presumably due to the tendency of Class 1 products to provide results that
340 are more variable in winter rather than in summer (as recently showed by Beck et al., 2019). In

12
341 fact, one major challenge of Class 1 products is to properly represent snowfall and light rainfall
342 occurring in winter (Kongoli et al., 2003; Liu and Seo, 2013; Habib et al., 2009; Tian et al., 2009;
343 Beck et al., 2019). Classes 2 and 3 products gave NSE values higher than 0.5 for six out of the
344 eight basins. However, the class providing the highest NSE depends on the considered basin, as
345 shown also during calibration. In fact, Class 2 datasets gave highest median NSE values on the
346 Godavari, Mississippi and Congo, which are basins characterized by Tropical and Temperate-
347 Tropical climate. Similarly, reanalysis-gauges based products (Class 3), such as GPCC, WFEDEI-
348 GPCC and PFD, led to the best model results on the Danube, Rhine and Volga, respectively, in
349 line with the results obtained in calibration. This can be due to the dense network of in-situ sensors
350 used to derive Class 3 products. In addition, Class 3 datasets perform similarly in both summer
351 and winter seasons.
352 These results are key to understand how different (quasi-) global precipitation datasets classes
353 can be used to better force a distributed hydrological model for improving flood simulation, and
354 which product in each class has lower variability in the model results for a particular river basin,
355 with specific characteristics.
356 However, these conclusions are only valid when comparing simulated and observed discharge
357 values at the outlet of the eight basins. For this reason, we performed additional analyses to
358 compare simulated and observed flows also at sensors located at internal points. The results of this
359 further analysis are described in the next section.

360 Model validation results at internal locations

361 In this analysis, only those basins (5 out of 8: Amazon, Danube, Godavari, Mississippi and
362 Rhine) for which observed flow values in internal points were available from 2007 up to 2013 (the
363 overlapping period of the 18 precipitation datasets) were considered. It is worth noting that we did
364 not consider internal flow points located downstream of dam as flow variability may be strongly
365 affected by regulation rules which were not accounted in the distributed hydrological model.
366 Our findings are presented focusing on: a) model performance by forcing dataset; b) model
367 performance at different internal locations; c) model bias performance and d) best performing
368 datasets.

13
369 a) Analysis of the model performance by forcing datasets

370 Figure 7 represents the boxplots of the NSE values comparing simulated and observed flow in
371 both the internal locations and outlet of the 5 river basins. The variability of the NSE value
372 drastically increases when considering model performances at internal locations. This is due to the
373 fact that models were calibrated using only outlet flow information and no internal data.
374 As previously demonstrated during model calibration and validation at basin outlets, Class 1
375 datasets show higher variability in model performances, while Classes 2 and 3 products lead to
376 results that are more comparable. Furthermore, no dataset constantly outperforms the others. In
377 fact, model performances change according to the river basin and its characteristic. The highest
378 average NSE values were achieved using TMPA 3B42 RT V7, GPCC, CMORPH-CRT V.10 and
379 CHIRPS V2.0 for the Amazon, Danube, Godavari, Mississippi and Rhine, respectively. The good
380 performances obtained using CHIRPS as model forcing may be due to the use of sub-monthly
381 gauge observations to improve precipitation estimate and because CHIRPS is specifically designed
382 to provide the most temporally homogeneous record possible (Beck et al., 2017a). Similar results
383 for the TMPA 3B42 RT V7 datasets on the Amazon were obtained by Zubieta et al. (2015).
384 From the results reported in Figure 7, it can be observed that the distributed model was unable
385 to reliably represent flow within the Mississippi river as a high variation of the NSE values is
386 displayed for almost all the forcing datasets. In addition, the median values of the box-plot graph
387 for the majority of the datasets are lower than zero, indicating poor model performances. This can
388 be related to the inability of the model to properly represent hydrological processes in highly
389 regulated basins (e.g. with high number dams) rather than bad performances of the precipitation
390 datasets. Results that are more reliable were achieved in the other four basins. As previously
391 demonstrated, on the Amazon, datasets of Class 1 generally outperformed those in Classes 2 and
392 Class 3. This can be due to the systematic limitations of producuts from Class 3 in properly
393 detecting precipitation across South America (Sun et al., 2018).
394 On the Danube basin, both Classes 2 and Class 3 products showed NSE values higher than 0.5,
395 while on the Rhine basin there was higher variability of the median values of these two dataset
396 classes, which might be related to the high human influence on this basin. Moreover, corrected
397 satellites products (Class 2) gave, on the Godavari, higher median NSE values and lower
398 variability of model results than Class 3.

14
399 b) Analysis of the model performance at different basins

400 For what concerns model results with respect to the spatial distribution of the sensors, Figure
401 8 represents the boxplots of the NSE values obtained at different sensor locations across all
402 precipitation datasets. In addition, the human footprint at each internal gauge was represented
403 using the dataset provided by Kareiva et al. (2007). Model performance at each location varied
404 with the size of the basin. In fact, the variability of the median values of NSE at internal sensors
405 of the basin was found higher when the size of the basin increases. This was clear for the Amazon
406 and Mississippi rivers. Here, it was difficult to find a consistent spatial pattern of NSE values with
407 respect to the order of the river reaches. For example, sensor 6 of the Amazon basin had higher
408 median NSE of the boxplot graph than sensor 2 (downstream of sensor 6) and sensors 10 and 11
409 (upstream of sensor 6). On the contrary, sensor 12 had higher median NSE than both sensor 5 and
410 3 that are located downstream along that particular reach.
411 An additional remark is that model performances showed erratic behaviour in upstream sensors
412 located in reaches converging into the same downstream reach. However, two main mechanisms
413 can be found: (1) it can happen that both upstream reaches had lower NSE median values than the
414 downstream sensor (e.g. downstream sensor 9 and upstream sensors 14 and 11 in the Mississippi
415 basin; or downstream sensor 6 and upstream sensors 10 and 11 in the Amazon basin) or; (2), model
416 performances at upstream reaches compensate each other, showing as an average resulting NSE
417 value at the downstream sensor (e.g. downstream sensor 12 and upstream sensors 6 and 4 in the
418 Mississippi basin).
419 It is interesting noting that no correlation was found between human footprint at each gauge
420 location and model performance. As mentioned in the method section, the overall flow at a
421 particular location is given by the sum of the flow generated by the model in that particular grid
422 and the upstream contribute routed at the downstream location. For this reason, the flow value at
423 a specific location is not affected by the high or low value of the human footprint in that location
424 as it may not be representative of the average value of human footprint within the basin (e.g.
425 upstream part of the basin), which has higher influence on the model performances. This is clear
426 in the Amazon basin, in which the majority of the basin has low human footprint (see Table 1),
427 but flow gauges are located in urbanized populated areas with higher human footprint close to the
428 river.

15
429 Furthermore, the results of this analysis showed that model results at the basin outlets were
430 higher than the ones at internal points of the basins. This was not surprising, since the model was
431 calibrated at the outlet of each basin and no information from internal points was used during the
432 calibration process. However, there were cases in which the model at internal locations performed
433 better than at the outlet of the basin. In order to further investigate this aspect, Figure 9 represents
434 the percentage of sensors in which NSE was higher at internal points (NSEI) than at the outlet of
435 the basin (NSEO) for the different precipitation datasets.
436 The first visible result is that on the Amazon, Danube and Rhine basins only few precipitation
437 datasets have a percentage of NSEI>NSEO higher than zero. As already mentioned in the previous
438 analyses, datasets belonging to Class 1 showed higher variability in model results (especially
439 PERSIANN, PERSIANN-CCS, and CMORPH V1.0) and higher percentage of NSEI>NSEO if
440 compared to the other two datasets classes. This is evident in Figure 9 for the Mississippi and
441 Godavari basins. The latter shows a persistent percentage value of NSEI>NSEO higher than 0: this
442 is because simulated flow at sensor 4 (see Figure 8) always outperformed the model performances
443 at the basin outlet.

444 c) Analysis of the model bias performance

445 To further assess model performance at internal points, Figure 10 illustrates the percentage of
446 precipitation datasets, for each sensor location, that overestimates (Bias index greater than 1) the
447 observed flow: this is called %Bias in the following. The red circle indicates that 100% of datasets
448 overestimates the observed flow, while 0% shows that all the precipitation datasets led to
449 underestimation of the river flow at a particular sensor location.
450 A very interesting result is that, despite the fact that the NSE values are highly dependent on
451 sensor location, the percentage of Bias index higher than 1 (called %Bias in the following) was
452 not influenced by the different dataset classes. In particular, the distributed model generally
453 provided uniform Bias values on medium scale basins (Danube, Godavari and Rhine), while
454 heterogeneous values were obtained on large-scale basins (Amazon and Mississippi). The model
455 underestimated flow observations within the Rhine (4th row) and Danube (5th row) basins,
456 regardless of precipitation product class; this behaviour could be attributed to high human
457 influence on both basins. On the contrary, the model overestimated observed flow on the Godavari
458 (3rd row). Different results were achieved for large-scale basins, Amazon (1st row) and Mississippi

16
459 (2nd row). The former showed a flow underestimation of the distributed model on the Northern
460 part of the basin, while the opposite occurred on the Southern reaches. Instead, for the Mississippi
461 river basin the model seemed to overestimate observed flow. However, underestimation was also
462 visible for the Eastern reaches of the basin.
463 An interesting aspect visible on the Amazon, Mississippi and Rhine basins is that sensors
464 located at two upstream reaches converging into the basin outlet usually overestimated and
465 underestimated the observed flow in the upstream reaches, and provided accurate results at the
466 outlet. This phenomenon might be due to the model calibration, which focuses on optimised model
467 parameters using flow outlet with the consequent compensation of flow at upstream opposite river
468 reaches. This is clearly illustrated in Figure 11, where the ensemble of simulated flows using all
469 the precipitation datasets is represented in red for overestimation (red arrow in the map, Bias index
470 greater than 1.1), blue for underestimation (blue arrow, Bias index smaller than 0.9), and green for
471 Bias index values close to 1 (green circle).

472 d) Best performing datasets

473 To conclude the assessment of model results at internal locations, in Figure 12 we show the
474 datasets that provided the highest NSE value for each sensor location in each river basin. It can be
475 observed that there was not a clear “winner” dataset, as it is showed in the validation analysis at
476 the basin outlet. However, datasets of Class 2 often provide the highest NSE values. This can be
477 due to the explicit use of daily gauge data for correcting satellite products. In particular, out of the
478 47 locations within the 5 river basins, Class 2 outperformed the other two classes 26 times (see
479 Table 5). CHIRPS V2.0 (12 times), MSWEP V2.1 (7 times), CMORPH-CRT V2.0 (4 times) and
480 PERSIANN-CDR (3 times) are the datasets in Class 2 that gave the highest NSE values. The higher
481 model performances achieved with CHIRPS V2.0, MSWEP V2.1, and CMORPH-CRT V2.0 over
482 PERSIANN-CDR can be related to the fact that PERSIANN-CDR indirectly incorporate gauge
483 data.
484 Moreover, the model forced with CHIRPS V2.0 gave the highest NSE values in the Rhine and
485 Mississippi basins, MSWEP V2.1 on both Mississippi and Danube basins, CMORPH-CRT V2.0
486 mainly in the Godavari basin, while PERSIANN-CDR on the Amazon basin. This suggests that
487 using sub-monthly gauge observations improves precipitation products, and consequently flow
488 simulation. In addition, the good performances of CHIRPS and MSWEPS V2.1 may be due to the

17
489 use of high-resolution gauge-based climatic datasets to determine the most accurate long-term
490 precipitation mean (Beck et al. 2017b).
491 Regarding Class 1 datasets, SM2RAIN-ASCAT provided the best results on the Amazon (at 4
492 sensors) and Danube (2 sensors). This result obtained with SM2RAIN-ASCAT on the Amazon
493 was quite surprising as Massari et al. (2017) stated that soil moisture based rainfall estimates
494 perform reasonably well in semi-arid climates rather than wet ones. We speculate that the good
495 performance of SM2RAIN-ASCAT on the Amazon can be mainly driven by the model calibration
496 and the optimal set of parameters used. TMPA 3B42 RT V7 and PERSIANN-CCS showed also
497 high NSE values on the Amazon (3 times) as they provide a good precipitation estimate in wet
498 tropical regions. Class 3 datasets gave high NSE values for 11 sensors of the five river basins. In
499 particular, besides PFD, all datasets of Class 3 provided the highest NSE at least for one sensor in
500 the basins. The results of this last analysis are collected in Table 5.
501 Another interesting result was that datasets producing the best model results at the basin outlet
502 did not provide the highest model performances at internal sensor locations. For instance, on the
503 Rhine basin the GPCC dataset allowed to achieve the highest model performance at the basin
504 outlet, while at internal points the CHIRPS V2.0 dataset was the one showing the best model results
505 in most of the cases. A similar pattern can be observed on the Mississippi river. This can be due to
506 the better spatial representation of the precipitation field by the CHIRPS V2.0 dataset and to the
507 fact that NSE index obtained at the basin outlet had comparable value as the one achieved with the
508 best performing dataset. On the other hand, SM2RAIN-ASCAT and CMORPH-CRT V1.0
509 outperformed the other datasets in both basin outlets and on most of the internal locations of the
510 Amazon and Godavari basins, respectively.

511 Conclusions

512 In this study, we compared the performance of 18 different precipitation datasets when used
513 as main forcing in a grid-based distributed hydrological model to simulate river flow in large river
514 basins. The 18 datasets were classified as Uncorrected Satellites (Class 1), Corrected Satellites
515 (Class 2) and Reanalysis - Gauges based datasets (Class 3). To provide a broad-based analysis, 8
516 river basins (Amazon, Brahmaputra, Congo, Danube, Godavari, Mississippi, Rhine and Volga)
517 having different sizes (medium to large), hydrometeorological characteristics, and human

18
518 influences were considered. The distributed hydrological model was re-calibrated for each
519 precipitation dataset individually.
520 The findings of this study underlined the importance of the proper selection of precipitation
521 products for large-scale distributed hydrological modelling purposes. We showed that there is not
522 a unique best performing precipitation dataset for all basins and results are very sensitive to the
523 basin characteristics. However, few datasets persistently outperform the others, i.e. SM2RAIN-
524 ASCAT for Class 1, CHIRPS V2.0, MSWEP V2.1, and CMORPH-CRTV1.0 for Class 2, GPCC
525 and WFEDEI GPCC for Class 3. The main findings of our study, related to the research questions
526 that we originally posed, are:

527 1) The distributed hydrological model was able to properly represent river flow even if highly
528 sensitive to differences in precipitation forcing (in agreement with Voisin et al. 2008). In
529 particular, the model was mainly affected by the scale of the basin and by the human
530 influence expressed as presence of reservoirs (see number of reservoirs in Table 1) and
531 human footprint in the basin; different climatic zones had an indirect impact on the
532 hydrological model through their influence on precipitation datasets;
533 2) Uncorrected Satellite datasets (Class 1), which can be considered as NRT products,
534 provided the lowest and most variable model results when compared to the other two
535 classes of products. SM2RAIN-ASCAT is the dataset that provided the best results within
536 Class 1 products.
537 3) Precipitation datasets belonging to Class 2 outperformed the other datasets in basins with
538 Tropical and Temperate-Arid climate (e.g. Congo, Mississippi and Godavari), while Class
539 3 datasets showed the highest NSE values in Temperate and Temperate-Cold basins (e.g.
540 Danube, Rhine and Volga). In addition, datasets from Class 3 gave the best performances
541 at basin outlets in case of dense precipitation monitoring networks, as in the Danube and
542 Rhine basins.
543 4) Performance of model results decreased when comparing simulated distributed flows at
544 internal basin locations with the observed ones. Higher variability of model results could
545 be observed for large-scale basins like the Amazon and Mississippi. This can be due to the
546 complex hydrometeorological characteristics and highly non–linear rainfall–runoff
547 response (e.g. Mississippi basin, as described in Beck et al., 2017b). As expected, the
548 highest NSE values were obtained at the basin outlets.

19
549 5) Despite the different spatial results provided by the distributed model, the bias index value
550 did not (significantly) change when using one class of dataset or the others.
551 6) Surprisingly, precipitation datasets showing the highest model result at the basin outlet did
552 not provide a corresponding highest result at internal locations, supporting the use of
553 distributed modelling approach rather than lumped. Overall, datasets from Class 2 provided
554 the highest model results at internal basin locations.

555 Our conclusions are based on specific river basins (although large-scale and representing
556 different parts of the globe), and a specific hydrological model (although distributed). Thus, to
557 refine our findings, we suggest future studies on this field to: a) test other physically based
558 distributed hydrological models; b) calibrate the hydrological model using also internal sensors;
559 c) assess influence of different model spatial resolutions, d) better understanding on how the
560 physical characteristics of the datasets might affect model results . Furthermore, in this study the
561 values of the parameters are calibrated for each precipitation dataset. Consequently, we suggest
562 for a future research direction to perform a global sensitivity analysis of distributed hydrological
563 models to both input data and model parameters, in order to understand the spatial distribution of
564 the governing principle of the hydrological processes and to guide the modeler towards an
565 informed selection of the precipitation dataset.
566 The outcomes of this study are valuable to support the selection of precipitation dataset to
567 achieve reliable model results for global and large-scale applications. This is the first research that
568 attempts to model large-scale basins using multiple global datasets of precipitations within a
569 distributed hydrological model. Our research offers promising results that might be key in
570 assessing flow values in data scarce river basins.

571 Acknowledgements

572 The Authors gratefully thank the GRDC for providing the observed Q data. This research was
573 partly supported by the European Research Council (ERC) within the project
574 “HydroSocialExtremes: Uncovering the Mutual Shaping of Hydrological Extremes and Society”,
575 ERC Consolidator Grant no. 761678. Part of this research was supported by the Swedish Strategic
576 research programme StandUP for Energy.
577

20
578 Conflicts of Interest

579 The authors declare no conflict of interest.

580 Author contributions

581 Conceptualization: M.M.; Data curation: M.M. and A.A.; Methodology. M.M. and L.B.;
582 Distributed Hydrological modelling: M.M.; Investigation: M.M., L.B. and A.A.; Original draft:
583 M.M.; Writing - review & editing. M.M., L.B. and A.A
584
585 .

21
586 References

587 Anderton S, Latron J, Gallart F. 2002. Sensitivity analysis and multi-response, multi-criteria
588 evaluation of a physically based distributed model. Hydrological Processes 16(2):333–353.
589 [Link] [Link]
590 Artan, G., Gadain, H., Smith, J.L., Asante, K., Bandaragoda, C.J., Verdin, J.P., 2007. Adequacy
591 of satellite derived rainfall data for streamflow modeling. [Link] 43, 167–185.
592 Ashouri, H., Hsu, K., Sorooshian, S., Braithwaite, D. K., Knapp, K. R., Cecil, L. D., Nelson, B.
593 R., and Pratt, O. P.: PERSIANNCDR: daily precipitation climate data record from
594 multisatellite observations for hydrological and climate studies, B. Am. Meteorol. Soc., 96,
595 69–83, 2015.
596 Bardossy, A. (2007), Calibration of hydrological model parameters for ungauged catchments,
597 Hydrol. Earth Syst. Sci., 11, 703–710.
598 Barrett, E.C., Martin D.W., 1981. The Use of Satellite Data in Rainfall Monitoring. Academic
599 Press (London), 340 pp.
600 Beck, H. E., A. I. J. M. van Dijk, A. de Roo, D. G. Miralles, T. R. McVicar, J. Schellekens, and L.
601 A. Bruijnzeel (2016), Global‐scale regionalization of hydrologic model parameters, Water
602 Resour. Res., 52, 3599–3622, doi: 10.1002/2015WR018247.
603 Beck, H.E., van Dijk, A.I., Levizzani, V., Schellekens, J., Miralles, D.G., Martens, B., de Roo, A.,
604 2017a. MSWEP: 3-hourly 0.25 global gridded precipitation (1979–2015) by merging gauge,
605 satellite, and reanalysis data. Hydrol. Earth Syst. Sci.21 (1), 589.
606 Beck, H. E., Vergopolan, N., Pan, M., Levizzani, V., van Dijk, A. I. J. M., Weedon, G. P., Brocca,
607 L., Pappenberger, F., Huffman, G. J., and Wood, E. F.: Global-scale evaluation of 22
608 precipitation datasets using gauge observations and hydrological modeling, Hydrol. Earth Syst.
609 Sci., 21, 6201-6217, [Link] 2017b.
610 Beck, H. E., Pan, M., Roy, T., Weedon, G. P., Pappenberger, F., van Dijk, A. I. J. M., Huffman,
611 G. J., Adler, R. F., and Wood, E. F.: Daily evaluation of 26 precipitation datasets using Stage-
612 IV gauge-radar data for the CONUS, Hydrol. Earth Syst. Sci., 23, 207-224,
613 [Link] 2019.
614 Behrangi, A., Khakbaz, B., Jaw, T.C., AghaKouchak, A., Hsu, K., Sorooshian, S., 2011.
615 Hydrologic evaluation of satellite precipitation products over a mid-size basin. J. Hydrol. 397
616 (3–4), 225–237.

22
617 Bergström, S.: The HBV model – its structure and applications, SMHI Reports RH 4, Swedish
618 Meteorological and Hydrological Institute (SMHI), Norrköping, Sweden, 1992.
619 Beven K. 2001. How far can we go in distributed hydrological modelling? Hydrology and Earth
620 System Sciences 5(1):1–12.
621 Beven K. 2002. Towards a coherent philosophy for modelling the environment. Proceedings of
622 the Royal Society A–Mathematical Physical and Engineering Sciences 458(2026):2465–2484.
623 Bitew, M.M., Gebremichael, M., 2011. Assessment of satellite rainfall products for streamflow
624 simulation in medium watersheds of the Ethiopian highlands. Hydrol. Earth Syst. Sci. 15,
625 1147–1155.
626 Bitew, M.M., Gebremichael, M., Ghebremichael, L.T., Bayissa, Y.A., 2012. Evaluation of high-
627 resolution satellite rainfall products through streamflow simulation in a hydrological modeling
628 of a small mountainous watershed in Ethiopia. J. Hydrometeorol. 13, 338–350.
629 Bloschl, G., Reszler, C. R., and Komma, J.: A spatially distributed ¨ flash flood forecasting model,
630 Environ. Mod. & Soft., 23, 464– 478, 2008.
631 Boyle, D.B., Gupta, H.V., Sorooshian, S., Koren, V., Zhang, Z., Smith, M., 2001. Toward
632 improved streamflow forecasts, value of semi-distributed modeling. Water Resources
633 Research, 37 (11), 2749-2759.
634 Brigode, P., Oudin, L. Perrin, C. (2013) Hydrological model parameter instability: A source of
635 additional uncertainty in estimating the hydrological impacts of climate change, Journal of
636 Hydrology, 476, 410-425.
637 Brocca, L., Melone, F., Moramarco, T., Wagner, W., 2013. A new method for rainfall estimation
638 through soil moisture observations. Geophys. Res. Lett. 40 (5), 853–858.
639 Brooks ES, Boll J, McDaniel PA. 2007. Distributed and integrated response of a geographic
640 information system–based hydrologic model in the eastern Palouse region, Idaho.
641 Hydrological Processes 21(1):110–122.
642 Camici, S., Ciabatta, L., Massari, C., and Brocca, L., How reliable are satellite precipitation
643 estimates for driving hydrological models: A verification study over the Mediterranean area,
644 Journal of Hydrology, 563, 950-961, 2018.
645 Carpenter, T.M., Georgakakos, K.P., 2006. Intercomparison of lumped versus distributed
646 hydrologic model ensemble simulations on operational forecast scales. Journal of Hydrology,
647 329(1–2), 174-185.

23
648 Casse, C., Gosset, M., Peugeot, C., Pedinotti, V., Boone, A., Tanimoun, B.A., Decharme, B., 2015.
649 Potential of satellite rainfall products to predict Niger River flood events in Niamey. Atmos.
650 Res. 163, 162–176.
651 Chen, M., Shi, W., Xie, P., Silva, V. B. S., Kousky, V. E., Higgins, R. W., and Janowiak, J. E.:
652 Assessing objective techniques for gauge-based analyses of global daily precipitation, J.
653 Geophys. Res., 113, D04110, [Link] 2008.
654 Cole, S. J. and Moore, R. J.: Distributed hydrological modelling using weather radar in gauged
655 and ungauged basins, Adv. Water Resour., 32, 1107–1120, 2009.
656 Collischonn, B., Collischonn, W., Tucci, C.E.M., 2008. Daily hydrological modeling in the
657 Amazon basin using TRMM rainfall estimates. J. Hydrol. 360, 207–216.
658 Csagoly, P., Magnin, G., & Hulea, O. (2016). Danube River Basin The Wetland Book (pp. 1-12):
659 Springer.
660 Cunge, J. A. (1969). “On the subject of a flood propagation computation method (Muskingum
661 method).” J. Hydraul. Res., 7(2), 205–230
662 Demirel, M. C., M. J. Booij, and A. Y. Hoekstra (2015), The skill of seasonal ensemble low-flow
663 forecasts in the Moselle River for three different hydrological models, Hydrol. Earth Syst. Sci.,
664 19(1), 275–291.
665 Derin, Y., & Yilmaz, K. K. (2014). Evaluation of multiple satellite-based precipitation products
666 over complex topography. Journal of Hydrometeorology, 15(4), 1498–1516.
667 Duan, Z., Liu, J., Tuo, Y., Chiogna, G., Disse, M., Evaluation of eight high spatial resolution
668 gridded precipitation products in Adige Basin (Italy) at multiple temporal and spatial scales,
669 Science of The Total Environment, 573, 1536-1553, 2016.
670 Ebert, E.E., Janowiak, J.E., Kidd, C., 2007. Comparison of near-real-time precipitation estimates
671 from satellite observations and numerical models. Bull. Am. Meteorol. Soc. 88, 47–64.
672 [Link]
673 Essou, G. R. C., Arsenault, R., Brissette, F. P., 2016. Comparison of climate datasets for lumped
674 hydrological modeling over the continental United States. J. Hydrol., 537, 334–345.
675 [Link] 2016.
676 Falck, Aline S. et al., 2015. Propagation of satellite precipitation uncertainties through a distributed
677 hydrologic model: a case study in the Tocantins-Araguaia basin in Brazil. J. Hydrol. 527, 943–
678 957.

24
679 Fekete, B. M., Vörösmarty, C. J., Roads, J. O., and Willmott, C. J.: Uncertainties in precipitation
680 and their impacts on runoff estimates, J. Climate, 17, 294–304, 2004.
681 Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Rowland, J.,
682 Harrison, L., Hoell, A., and Michaelsen, J.: The climate hazards infrared precipitation with
683 stations – a new environmental record for monitoring extremes, Scientific Data, 2, 150066,
684 [Link] 2015.
685 Gaillardet, J., Dupre, B., Allegre, C. J., & Négrel, P. (1997). Chemical and physical denudation in
686 the Amazon River Basin. Chemical geology, 142(3-4), 141-173.
687 Goodrich, D., Schmugge, T., Jackson, T., Unkrich, C., Keefer, T., Parry, R., Bach, L., and Amer,
688 S.: Runoff simulation sensitivity to remotely sensed initial soil water content, Water Resour.
689 Res., 30, 1393–1405, 1994.
690 Goolsby, D. A., & Battaglin, W. A. (2001). Long term changes in concentrations and flux of
691 nitrogen in the Mississippi River Basin, USA. Hydrological processes, 15(7), 1209-1226.
692 Gottschalck, J., Meng, J., Rodell, M., Houser, P., 2005. Analysis of multiple precipitation products
693 and preliminary assessment of their impact on global land data assimilation system land surface
694 states. J. Hydrometeorol. 6, 573–598. [Link] JHM437.1.
695 Habib, E., Henschke, A., and Adler, R. F.: Evaluation of TMPA satellite-based research and real-
696 time rainfall estimates during six tropical-related heavy rainfall events over Louisiana, USA,
697 Atmos. Res., 94, 373–388, 2009.
698 Harrison, I. J., Brummett, R., & Stiassny, M. L. (2016). Congo River Basin. The Wetland Book:
699 II: Distribution, Description and Conservation, 1-18.
700 Hong, Y., Hsu, K.-L., Sorooshian, S., and Gao, X.: Precipitation Estimation from Remotely Sensed
701 Imagery Using an Artificial Neural Network Cloud Classification System, J. Appl. Meteorol.,
702 43, 1834–1853, 2004.
703 Hong, Y., Hsu, K. L., Moradkhani, H. and S. Sorooshian (2006). Uncertainty quantification of
704 satellite precipitation estimation and Monte Carlo assessment of the error propagation into
705 hydrologic response. Water resources research, 42(8).
706 Hossain, F., Anagnostou, E.N., 2004. Assessment of current passive-microwave- and infrared-
707 based satellite rainfall remote sensing for flood prediction. J. Geophys. Res. 109, D07102.
708 [Link] Corrigendum, 110, D06115,
709 doi:10.1029/2005JD005831.

25
710 Hossain, F., Lettenmaier, D.P., 2006. Flood prediction in the future: Recognizing hydrologic issues
711 in anticipation of the Global Precipitation Measurement mission. Water Resour. Res.42,
712 W11301.
713 Huffman, G. J., Adler, R. F., Morrissey, M. M., Bolvin, D. T., Curtis, S., Joyce, R., McGavock,
714 B., and Susskind, J.: Global precipitation at one-degree daily resolution from multi-satellite
715 observations, J. Hydrometeorol., 2, 36–50, 2001.
716 Huffman, G. J., Bolvin, D. T., Nelkin, E. J., Wolff, D. B., Adler, R. F., Gu, G., Hong, Y., Bowman,
717 K. P., and Stocker, E. F.: The TRMM Multisatellite Precipitation Analysis (TMPA):
718 quasiglobal, multiyear, combined-sensor precipitation estimates at fine scales, J.
719 Hydrometeorol., 8, 38–55, 2007.
720 Hussain, Y., Satgé, F., Hussain, M.B. et al. Theor Appl Climatol (2018) 131: 1119.
721 [Link]
722 Iguchi, T., Kozu, T., Kwiatkowski, J., Meneghini, R., Awaka, J., and Okamoto, K.: A Kalman
723 filter approach to the Global Satellite Mapping of Precipitation (GSMaP) from combined
724 passive microwave and infrared radiometric data, J. Meteorol. Soc. Jpn., 87A, 137–151, 2009.
725 Immerzeel, W. (2008). Historical trends and future predictions of climate variability in the
726 Brahmaputra basin. International Journal of Climatology, 28(2), 243-254.
727 Jha, P. K., Tiwari, J., Singh, U. K., Kumar, M., & Subramanian, V. (2009). Chemical weathering
728 and associated CO2 consumption in the Godavari river basin, India. Chemical geology, 264(1-
729 4), 364-374.
730 Jin, X., C. Xu, Q. Zhang, and Y. D. Chen (2009), Regionalization study of a conceptual
731 hydrological model in Dongjiang basin, south China, Quat. Int., 208 (1–2), 129–137.
732 Joyce, R. J., Janowiak, J. E., Arkin, P. A., and Xi, P.: CMORPH: A method that produces global
733 precipitation estimates from passive microwave and infrared data at high spatial and temporal
734 resolution, J. Hydrometeorol., 5, 487–503, 2004.
735 Kareiva, P., Watts, S., McDonald, R. and Boucher, T. (2007) Domesticated Nature: Shaping
736 Landscapes and Ecosystems for Human Welfare, Science, 316(5833), 1866-1869,
737 DOI:10.1126/science.1140170.
738 Karssenberg, D., Schmitz, O., Salamon, P., De Jong, C., and Bierkens, M. F. P.: A software
739 framework for construction of process-based stochastic spatio-temporal models and data
740 assimilation, Environ. Mod. & Soft., 25, 1–14, doi:10.1016/[Link].2009.10.004, 2009.

26
741 Kidd, C. and Huffman, G. (2011), Global precipitation measurement. Met. Apps, 18: 334-353.
742 doi:10.1002/met.284.
743 Kidd, C., Becker, A., Huffman, G.J., Muller, C.L., Joe, P., Skofronick-Jackson, G., Kirschbaum,
744 D.B., 2017. So, how much of the Earth’s surface is covered by rain gauges? Bull. Am.
745 Meteorol. Soc. 98, 69–78. https:// [Link]/10.1175/BAMS-D-14-00283.1.
746 Kongoli, C., Pellegrino, P., Ferraro, R. R., Grody, N. C., and Meng, H.: A new snowfall detection
747 algorithm over land using measurements from the Advanced Microwave Sounding Unit
748 (AMSU), Geophys. Res. Lett., 30, [Link] 2003.
749 Kottek, M., J. Grieser, C. Beck, B. Rudolf, and F. Rubel, 2006: World Map of the Köppen-Geiger
750 climate classification updated. Meteorol. Z., 15, 259-263. DOI: 10.1127/0941-
751 2948/2006/0130.
752 Krause, P., Boyle, D. P., and Bäse, F.: Comparison of different efficiency criteria for hydrological
753 model assessment, Adv. Geosci., 5, 89–97, [Link] 2005.
754 Lehner, B., C. Reidy Liermann, C. Revenga, C. Vörösmarty, B. Fekete, P. Crouzet, P. Döll, M.
755 Endejan, K. Frenken, J. Magome, C. Nilsson, J.C. Robertson, R. Rodel, N. Sindorf, and D.
756 Wisser. 2011. High-Resolution Mapping of the World's Reservoirs and Dams for Sustainable
757 River-Flow Management. Frontiers in Ecology and the Environment 9 (9): 494-
758 [Link]://[Link]/10.1890/100125.
759 Li, L., Ngongondo, C.S., Xu, C.Y., Gong, L., 2013. Comparison of the global TRMM and WFD
760 precipitation datasets in driving a large-scale hydrological model in southern Africa.
761 Hydrology Research, 44(5), 770-788.
762 Libertino, A., Sharma, A., Lakshmi, V., Claps, P., 2016. A global assessment of the timing of
763 extreme rainfall from TRMM and GPM for improving hydrologic design. Environ. Res. Lett.
764 11, 1.
765 Liu, G. and Seo, E.-K.: Detecting snowfall over land by satellite high-frequency microwave
766 observations: The lack of scattering signature and a statistical approach, J. Geophys. Res.-
767 Atmos., 118, 1376–1387, 2013.
768 Liu, X., Yang, T., Hsu, K., Liu, C., and Sorooshian, S.: Evaluating the streamflow simulation
769 capability of PERSIANN-CDR daily rainfall products in two river basins on the Tibetan
770 Plateau, Hydrol. Earth Syst. Sci., 21, 169-181, [Link]
771 2017.

27
772 Maggioni, M. and Massari, C., 2018. On the performance of satellite precipitation products in
773 riverine flood modeling: A review. J. Hydrol. 558, 214–224.
774 Massari, C., Crow, W., and Brocca, L.: An assessment of the performance of global rainfall
775 estimates without ground-based observations, Hydrol. Earth Syst. Sci., 21, 4347-4361,
776 [Link] 2017.
777 Merz, R., and G. Bloeschl (2004), Regionalisation of catchment model parameters, J. Hydrol., 287
778 (1–4), 95–123.
779 Minet, J., Laloy, E., Lambot, S., andVanclooster, M.: Effect of high-resolution spatial soil moisture
780 variability on simulated 5 runoff response using a distributed hydrolologic model, Hydrol.
781 Earth Syst. Sci. 15, 1323-1338, 2011.
782 Moazami, S., Golian, S., Kavianpour, M. R., Hong, Y., 2013. Comparison of PERSIANN and V7
783 TRMM Multi-satellite Precipitation Analysis (TMPA) products with rain gauge data over Iran.
784 Int. J. Remote Sens., 34, 8156–8171.
785 Molinier, M., Guyot, J.-L., De Oliveira, E., Guimarães, V., & Chaves, A. (1995). Hydrologie du
786 bassin de l’Amazone. Proc. Grands Bassins Fluviaux Péri-atlantiques, 1, 335-344.
787 Moriasi, D. N., Arnold, J. G., Van Liew, M. W., Bingner, R. L., Harmel, R. D., and Veith, T. L.:
788 Model evaluation guidelines for systematic quantification of accuracy in watershed
789 simulations, Transactions of the American Society of Agricultural and Biological Engineers,
790 50, 885–900, 2007.
791 Nash, J. E. and Sutcliffe, J. V.: River flow forecasting through conceptual models part I – a
792 discussion of principles, J. Hydrol., 10, 282–290, 1970.
793 PCraster: Environmental Modelling language, available at: http:// [Link] (last access:
794 01 December 2018), 2018.
795 Pedinotti, V., Boone, A., Decharme, B., Cr´etaux, J. F., Mognard, N., Panthou, G., Papa, F.,
796 Tanimoun, B. A., [Link] of the ISBA-TRIP continental hydrologic system over the
797 Niger basin using in situ and satellite derived datasets. Hydrol. Earth Syst. Sci., 16, 1745–1773.
798 Peel, M. C., Finlayson, B. L., and McMahon, T. A.: Updated world map of the Köppen-Geiger
799 climate classification, Hydrol. Earth Syst. Sci., 11, 1633-1644, [Link]
800 11-1633-2007, 2007.
801 Press, W.H., et al., 1992. Numerical recipes in FORTRAN. 2nd ed. Cambridge, UK: Cambridge
802 University Press.

28
803 Rahman, M. M., Lu, M., and Kyi, K. H.: Seasonality of hydrological model spin-up time: a case
804 study using the Xinanjiang model, Hydrol. Earth Syst. Sci. Discuss.,
805 [Link] in review, 2016.
806 Ritter A., Muñoz-Carpena, R. Performance evaluation of hydrological models: Statistical
807 significance for reducing subjectivity in goodness-of-fit assessments, Journal of Hydrology,
808 480, 33-45, 2013.
809 Schamm, K., Ziese, M., Becker, A., Finger, P., Meyer-Christoffer, A., Schneider, U., Schröder,
810 M., and Stender, P.: Global gridded precipitation over land: a description of the new GPCC
811 First Guess Daily product, Earth Syst. Sci. Data, 6, 49-60, [Link]
812 2014, 2014.
813 Schultz, G.A., 1996. Remote sensing applications to hydrology: runoff. Hydrological Sciences
814 Journal 41, 453–475.
815 Seck, A., Welty, C., and Maxwell, R. M.: Spin-up behavior and effects of initial conditions for an
816 integrated hydrologic model, Water Resour. Res., 51, 2188–2210,
817 [Link] 2015.
818 Seibert, J. and Vis, M. J. P.: Teaching hydrological modeling with a user-friendly catchment-
819 runoff-model software package, [Link] Syst. Sci., 16, 3315–3325,
820 [Link] 2012.
821 Seibert, J., 2003. Reliability of model predictions outside calibration conditions. Nordic Hydrology
822 34, 477–492. doi : 10.2166/nh.2003.028.
823 Serrat-Capdevila, A., Valdes, J.B., Stakhiv, E., 2014. Water management applications for satellite
824 precipitation products: synthesis and recommendations. J. Am. Water Resour. Assoc. 50, 509–
825 525. [Link]
826 Sheffield, J., Goteti, G., and Wood, E. F.: Development of a 50-year high-resolution global dataset
827 of meteorological forcings for land surface modeling, J. Climate, 19, 3088–3111, 2006.
828 Shiklomanov, A.I., Lammers, R.B., Vörösmarty, C.J., 2002. Widespread decline in hydrological
829 monitoring threatens pan-Arctic research. Eos, Transactions American Geophysical Union 83,
830 13.
831 Singh, K.S. and Bárdossy, A. (2012) Calibration of hydrological models on hydrologically unusual
832 events, Advances in Water Resources, 38, 81-91.

29
833 Sorooshian, S., Hsu, K.-L., Gao, X., Gupta, H. V., Imam, B., and Braithwaite, D.: Evaluation of
834 PERSIANN system satellitebased estimates of tropical rainfall, B. Am. Meteorol. Soc., 81,
835 2035–2046, 2000.
836 Stampoulis, D., Anagnostou, E.N., 2012. Evaluation of global satellite rainfall products over
837 continental Europe. J. Hydrometeorol. 13, 588–603. [Link] org/10.1175/JHM-D-11-
838 086.1.
839 Stephens, G.L. and C.D. Kummerow, 2007: The Remote Sensing of Clouds and Precipitation from
840 Space: A Review. J. Atmos. Sci., 64, 3742–3765,[Link]
841 Stisen, S., Sandholt, I., 2010. Evaluation of remote-sensing-based rainfall products through
842 predictive capability in hydrological runoff modelling. Hydrol. Process. 24, 879–891.
843 Stokstad, E., 1999. Scarcity of rain, stream gages threatens forecasts. Science 285, 1199–1200.
844 Su, F., Hong, Y., Lettenmaier, D.P., 2008. Evaluation of TRMM Multisatellite Precipitation
845 Analysis (TMPA) and its utility in hydrologic prediction in the La Plata basin. J.
846 Hydrometeorol. 9, 622–640.
847 Sun, Q., Miao, C., Duan, Q., Ashouri, H., Sorooshian, S., & Hsu, K.-L. (2018). A review of global
848 precipitation data sets: Data sources, estimation, and intercomparisons. Reviews of
849 Geophysics, 56, 79–107. [Link]
850 Tang, G., Zeng, Z., Long, D., Guo, X., Yong, B., Zhang, W., Hong, Y., 2016. Statistical and
851 hydrological comparisons between TRMM and GPM level-3 products over a midlatitude
852 basin: Is day-1 IMERG a good successor for TMPA 3B42V7? J. Hydrometeorol. 17 (1), 121–
853 137.
854 Thielen, J., Bartholmes, J., Ramos, M.-H., and De Roo, A.: The European Flood Alert System –
855 Part 1: Concept and development, Hydrol. Earth Syst. Sci., 13, 125–140, doi:10.5194/hess-13-
856 125- 2009, 2009.
857 Tian, Y., C. D. Peters-Lidard, J. B. Eylander, R. J. Joyce, G. J. Huffman, R. F. Adler, K. Hsu, F.
858 J. Turk, M. Garcia,and J. Zeng (2009), Component analysis of errors in satellite-based
859 precipitation estimates,J. Geophys. Res.,114, D24101,doi:10.1029/2009JD011949.
860 Tian, Y., Peters-Lidard, C.D., 2010. A global map of uncertainties in satellite-based precipitation
861 measurement. Geophys. Res. Lett. 37 (24).

30
862 Tsintikidis, D., Georgakakos, K.P., Artan, G.A., Tsonis, A.A., 1999. A feasibility study on mean
863 areal rainfall estimation and hydrologic response in the Blue Nile region using METEOSAT
864 images. Journal of Hydrology 221, 97–116.
865 Vetter, T., Huang, S., Aich, V., Yang, T., Wang, X., Krysanova, V., and Hattermann, F.: Multi-
866 model climate impact assessment and intercomparison for three large-scale river basins on
867 three continents, Earth Syst. Dynam., 6, 17–43, [Link] 2015.
868 Voisin, N., Wood, A. W., and Lettenmaier, D. P.: Evaluation of precipitation products for global
869 hydrological prediction, J. Hydrometeorol., 9, 388–407, 2008
870 Wang, S., S. Liu, X. Mo, B. Peng, J. Qiu, M. Li, C. Liu, Z. Wang, and P. Bauer-
871 Gottwein, 2015: Evaluation of Remotely Sensed Precipitation and Its Performance for
872 Streamflow Simulations in Basins of the Southeast Tibetan Plateau. J.
873 Hydrometeor., 16, 2577–2594, [Link]
874 Weedon, G. P., Balsamo, G., Bellouin, N., Gomes, S., Best, M. J., and Viterbo, P.: The WFDEI
875 meteorological forcing data set: WATCH Forcing Data methodology applied to ERA-Interim
876 reanalysis data, Water Resour. Res., 50, 7505–7514, 2014.
877 Wieriks, K., & Schulte‐Wülwer‐Leidig, A. (1997). Integrated water management for the Rhine
878 river basin, from pollution prevention to ecosystem improvement. Paper presented at the
879 Natural Resources Forum.
880 Wu, H., J. S. Kimball, N. Mantua, and J. Stanford (2011), Automated upscaling of river networks
881 for macroscale hydrological modeling, Water Resour. Res., 47, W03517,
882 doi: 10.1029/2009WR008871.
883 Wu, H., J. [Link], [Link], [Link], L. [Link], and R. [Link] (2012), A new global river
884 network database for macroscale hydrologic modeling, Water Resour. Res., 48, W09701,
885 doi:10.1029/2012WR012313.
886 Xie, P., and A.‐Y. Xiong (2011), A conceptual model for constructing high‐resolution gauge‐
887 satellite merged precipitation analyses, J. Geophys. Res., 116, D21106,
888 doi: 10.1029/2011JD016118.
889 Xie, P., R. Joyce, S. Wu, S. Yoo, Y. Yarosh, F. Sun, and R. Lin, 2017: Reprocessed, Bias-
890 Corrected CMORPH Global High-Resolution Precipitation Estimates from 1998. J.
891 Hydrometeor., 18, 1617–1641, [Link]

31
892 Yang, Y., Du, J., Cheng, L., Xu, W., 2017. Applicability of TRMM satellite precipitation in driving
893 hydrological model for identifying flood events: a case study in the Xiangjiang River Basin,
894 China. Nat Hazards, 87, 1489–1505. DOI 10.1007/s11069-017-2836-0
895 Yong, B., Ren, L. L., Hong, Y., Wang, J.H., Gourley, J.J., Jiang, S.H., Chen, X., Wang, W. 2010.
896 Hydrologic evaluation of Multisatellite Precipitation Analysis standard precipitation products
897 in basins beyond its inclined latitude band: A case study in Laohahe basin, China. Water
898 Resources Research, 46(7), w07542.
899 Zambrano-Bigiarini, M., Nauditt, A., Birkel, C., Verbist, K., Ribbe, L., 2017. Temporal and spatial
900 evaluation of satellite-based rainfall estimates across the complex topographical and climatic
901 gradients of Chile. Hydrol. Earth Syst. Sci., 21, 1295–1320. [Link]
902 1295-2017, 2017.
903 Zubieta, R.,Geritana, A., Espinoza, J. C., and Lavado W.: Impacts of Satellite-based Precipitation
904 Datasets on Rainfall-Runoff Modeling of the Western Amazon Basin of Peru and Ecuador, J.
905 Hydrol., 528, 599–612, [Link] 2015.
906
907

32
908 Tables

909 Table 1. Overview of the characteristic of the 8 river basins selected as case studies.
Climatic Basin Area River length Average Q # flow # % of human
Basin
zone (km2).103 (km) (m3/s).103 sensors reservoirs footprint
Amazon Tropical 6150 6992 209 12 8 8.6
Temperate-
Brahmaputra 525 2900 20 1 5 17.7
Cold
Congo Tropical 4000 4300 41 1 4 18.3
Temperate-
Danube 817 2857 7.1 7 175 39.9
Cold
Godavari Tropical 315 1465 3.4 4 58 34.6
Temperate-
Mississippi 3000 3800 17 16 698 27.9
Arid
Rhine Temperate 200 1223 2.9 8 24 43.7
Temperate-
Volga 1300 3700 8.1 1 16 28.6
Subpolar

910

911

1
912 Table 2. Overview of the 18 (quasi-) global precipitation datasets used as main forcing of the distributed hydrological model; the
913 acronym NRT indicates Near Real Time.

Spatial Spatial Temporal Temporal

# Name Data access
resolution coverage resolution coverage
Class 1: Uncorrected satellite datasets (Satellite)
Land,
1 CHIRP V2.0 0.05° Daily 1981–NRT [Link]
<50°
2 CMORPH V1.0 0.25° <60° 30 min 1998–NRT [Link]
3 PERSIANN 0.25° <60° Hourly 2000–NRT [Link]
4 PERSIANN–CCS 0.04° <60° Hourly 2003–NRT [Link]
5 SM2RAIN–ASCAT 0.50° Land Daily 2007–2015 [Link]
6 TMPA 3B42 RT V7 0.25° <50° 3–Hourly 2000–NRT [Link]
Class 2: Gauge corrected satellite datasets (Corrected Satellite)
Land,
7 CHIRPS V2.0 0.05° Daily 1981–NRT [Link]
<50°
CMORPH–CRT
8 0.25° <60° 30 min 1998–2015 [Link]
V1.0
9 GPCP1DD V1.2 1.00° Global Daily 1996–2015 [Link]
10 MSWEP V2.1 0.10° Global 3–hourly 1979–NRT [Link]
11 PERSIANN–CDR 0.25° <60° 6–Hourly 1983–2016 [Link]
12 TMPA 3B42 V7 0.25° <50° 3–hourly 2000–2017 [Link]
Class 3: Reanalysis and gauge based datasets (Reanalysis-Gauges)
[Link]/psd/data/gridded/[Link]
13 CPC Global Unified 0.50° Land Daily 1979–NRT
recip
[Link]
14 GPCC 1.00° <60° Daily 1988–2013
fulldata-daily_v1_doi_download
15 GSMaP–RNL 0.10° <60° Daily 2001–2013 [Link]
16 PFD 0.25° Global 3–hourly 1948–2012 [Link]
17 WFEDEI CRU 0.25° Global 3–hourly 1979–2015 [Link]
18 WFEDEI GPCC 0.25° Global 3–hourly 1979–2013 [Link]
914

1
915 Table 3. Calibration and validation periods for the different basins.

Calibration Validation
Basin Start End Start End
Amazon 01-01-2007 31-12-2009 01-01-2010 31-12-2013
Brahmaputra 01-01-2007 31-12-2009 01-01-2010 31-12-2012
Congo 01-01-2007 31-08-2010 01-01-2010 31-12-2013
Danube 01-01-2009 31-12-2010 01-01-2007 31-12-2008
Godavari 01-01-2010 31-12-2013 01-01-2007 31-12-2009
Mississippi 13-04-2010 31-12-2013 01-01-2008 31-12-2010
Rhine 01-01-2008 19-05-2011 20-05-2011 31-12-2013
Volga 01-01-2007 31-12-2008 01-01-2009 31-12-2010
916

917

1
918 Table 4. Average NSE calibration values obtained comparing the observed flow at the basin
919 outlets with the average simulated flow using different datasets classes.

Basin Class 1 Class 2 Class 3

Amazon 0.92 0.89 0.89
Brahmaputra 0.84 0.83 0.83
Congo 0.46 0.72 0.63
Danube 0.50 0.83 0.85
Godavari 0.85 0.89 0.88
Mississippi 0.50 0.85 0.84
Rhine 0.56 0.77 0.84
Volga 0.55 0.49 0.78
920

921

2
922 Table 5. Number of times a dataset provided the highest value of NSE at a given sensor
923 location.

Dataset # of highest NSE Class

CHIRPS V2.0 12 2
MSWEP V2.1 7 2
SM2RAIN-ASCAT 6 1
CMORPH-CRT V2.0 4 2
GPCC 3 3
PERSIANN-CDR 3 2
WFEDEI GPCC 3 3
TMPA 3B42 RT V7 2 1
PERSIANN-CCS 2 1
WFEDEI CRU 2 3
GSMaP-RNL 2 3
CPC Global Unified 1 3
924

925

926

3
927 Figures

928
929 Figure 1. Case studies and in-situ streamflow sensors locations.

930

931

4
932
933 Figure 2. Distributed modelling framework: the flow accumulation map is used to assess the
934 cell connectivity and propagate downstream the grid flow generated by rainfall-runoff processes.
935

5
936

937
938 Figure 3. Comparison between observed and average simulated flow using precipitation
939 datasets of classes 1, 2 and 3 at different basin outlets (in brackets the name of the stations).

940

6
941
942 Figure 4. NSE values from calibration, calculated comparing the simulated flow obtained
943 using different datasets as input in the distributed hydrological model and observed flow at the
944 different basin outlets.

945

7
946
947 Figure 5. Boxplots of the NSE values (from model validation) at basin outlets for all
948 precipitation datasets over the 8 river basins. Outliers are represented by red crosses.
949

8
950
951 Figure 6. Boxplots of the NSE values (from model validation) at basin outlets divided in dataset
952 classes (Class 1, 2 and 3 visualized in green, blue and red respectively). In addition, datasets
953 providing the maximum and minimum NSE for each class on each basin are represented as the
954 boxplot extremes
955
956

9
957
958 Figure 7 Boxplots of the NSE values (from model validation) for all precipitation datasets
959 (numbers from 1 to 18 refer to Table 2) over all sensors, both at the outlet and internal locations.
960 Outliers are represented by red crosses.

961

10
962
963 Figure 8. Boxplots of the NSE values (from model validation) for all sensor locations, over all
964 precipitation datasets. Outliers are represented by red crosses.
965
966

11
967
968 Figure 9. Percentage of internal sensors with NSE values higher than the one at basin outlet
969 for all precipitations datasets on the 8 river basins.

970

12
971
972 Figure 10. Percentage of precipitation datasets (divided by classes) with Bias index higher
973 than 1 (overestimation) for each sensor location. Percentage of 100% indicates that all datasets
974 overestimate observed flow at a given sensor location.
975

13
976
977 Figure 11. On the left side: average values of the Bias index for each sensor location on three
978 selected river basins. Red and blue arrows indicate average Bias index value higher than 1.1
979 (overestimation) and lower than 0.9 (underestimation). The green circles indicate average Bias
980 index between 0.9 and 1.1. On the right side: ensemble of simulated flow for all precipitation
981 datasets for three specific locations represented in the maps on the left side. Red and blue
982 ensembles indicate overestimation and underestimation, respectively.
983

14
984
985 Figure 12. Representation of the precipitation datasets that provided the best NSE values for
986 each river basin.

987

15