Improving the Predictive Skill of a Distributed
Hydrological Model by Calibration on Spatial
Patterns With Multiple Satellite Data Sets
Moctar Dembélé
1,2
, Markus Hrachowitz
2
, Hubert H. G. Savenije
2
,
Grégoire Mariéthoz
1
, and Bettina Schaefli
1,3
1
Faculty of Geosciences and Environment, Institute of Earth Surface Dynamics, University of Lausanne, Lausanne,
Switzerland,
2
Faculty of Civil Engineering and Geosciences, Water Resources Section, Delft University of Technology,
Delft, Netherlands,
3
Now at Institute of Geography (GIUB), University of Bern, Bern, Switzerland
Abstract Hydrological model calibration combining Earth observations and in situ measurements is a
promising solution to overcome the limitations of the traditional streamflow‐only calibration. However,
combining multiple data sources in model calibration requires a meaningful integration of the data sets,
which should harness their most reliable contents to avoid accumulation of their uncertainties and mislead
the parameter estimation procedure. This study analyzes the improvement of model parameter selection by
using only the spatial patterns of satellite remote sensing data, thereby ignoring their absolute values.
Although satellite products are characterized by uncertainties, their most reliable key feature is the
representation of spatial patterns, which is a unique and relevant source of information for distributed
hydrological models. We propose a novel multivariate calibration framework exploiting spatial patterns and
simultaneously incorporating streamflow and three satellite products (i.e., Global Land Evaporation
Amsterdam Model [GLEAM] evaporation, European Space Agency Climate Change Initiative [ESA CCI]
soil moisture, and Gravity Recovery and Climate Experiment [GRACE] terrestrial water storage). The
Moderate Resolution Imaging Spectroradiometer (MODIS) land surface temperature data set is used for
model evaluation. A bias‐insensitive and multicomponent spatial pattern matching metric is developed to
formulate a multiobjective function. The proposed multivariate calibration framework is tested with the
mesoscale Hydrologic Model (mHM) and applied to the poorly gauged Volta River basin located in a
predominantly semiarid climate in West Africa. Results of the multivariate calibration show that the
decrease in performance for streamflow (−7%) and terrestrial water storage (−6%) is counterbalanced with
an increase in performance for soil moisture (+105%) and evaporation (+26%). These results demonstrate
that there are benefits in using satellite data sets, when suitably integrated in a robust model
parametrization scheme.
1. Introduction
One of the key challenges in hydrological modeling (Beven, 2019a; Singh, 2018) is the reliable representation
of the spatiotemporal variability of natural processes, to which the footprint of human activity is often
superimposed. In most places, available in situ observations are not sufficient to capture the spatiotemporal
heterogeneity of dominant hydrological processes (AghaKouchak et al., 2015; Hrachowitz & Clark, 2017).
With the upswing in development of distributed hydrological models (DHMs) that offer spatially explicit
predictions as an essential tool for decision making (Fatichi et al., 2016; Kampf & Burges, 2007; Paniconi
& Putti, 2015; Semenova & Beven, 2015), there is a growing interest in the plausibility of their spatial
patterns (Ko et al., 2019; Koch et al., 2018; Stisen et al., 2018; Wealands et al., 2005; Zink et al., 2018).
Most commonly, hydrological models are calibrated using streamflow data alone (Becker et al., 2019; Yassin
et al., 2017). The streamflow signal represents an integrated response of the hydrological system to a set of
natural drivers (e.g., climate and landscape) and anthropogenic influences (e.g., deforestation and reser-
voirs) occurring upstream of the measurement's location (Koch et al., 2015; Rientjes et al., 2013).
Although streamflow is key to understanding the temporal dynamics of a system, it does not disclose much
information on the system‐internal spatial heterogeneity of the hydrological processes (McDonnell et al.,
2007; Rajib et al., 2018). It therefore has little discriminatory power to constrain the feasible parameter space
©2020. American Geophysical Union.
All Rights Reserved.
RESEARCH ARTICLE
10.1029/2019WR026085
Special Section:
Advancing process representa-
tion in hydrologic models:
Integrating new concepts,
knowledge, and data
Key Points:
• A spatial calibration fram ework is
developed by combining four
noncommensurable variables
describing different hydrological
processes
• A new bias‐insensitive metric is
developed to incorporate spatial
heterogeneity in a multivariate
calibration scheme
• Only spatial patterns of satellite data
are used to improve the predictive
skill of a distributed hydrological
model
Supporting Information:
• Suppor ting Information S1
Correspondence to:
M. Dembélé,
moctar.dembele@unil.ch;
m.dembele@tudelft.nl
Citation:
Dembélé, M., Hrachowitz, M., Savenije,
H. H. G., Mariéthoz, G., & Schaefli, B.
(2020). Improving the predictive skill of
a distributed hydrological model by
calibration on spatial patterns with
multiple satellite data sets. Water
Resources Research, 56,
e2019WR026085. https://doi.org/
10.1029/2019WR026085
Received 1 AUG 2019
Accepted 13 JAN 2020
Accepted article online 15 JAN 2020
DEMBÉLÉ ET AL. 1of26
of a distributed model, i.e., the boundary flux or closure problem (Beven, 2006b). Consequently, a spatially
DHM calibrated only on streamflow is very unlikely able to reproduce a reliable spatiotemporal representa-
tion of other hydrological fluxes and states (Birhanu et al., 2019; Clark et al., 2016; Grayson & Bloschl, 2001;
Hrachowitz et al., 2014; Livneh & Lettenmaier, 2012; Minville et al., 2014), even if a multiscale parameter
regionalization (MPR) scheme is used (Rakovec et al., 2016). Mismatches between temporal and spatial pat-
terns should therefore be expected when comparing hydrological model outputs to other distributed obser-
vational data sets (Vereecken et al., 2008; Xu et al., 2014).
For a few decades, satellite remote sensing (SRS) has opened up new avenues for the development of spatial
hydrology (Cui et al., 2018; Engman & Gurney, 1991; Lettenmaier et al., 2015; McCabe et al., 2017; Mendoza
et al., 2002; Pasetto et al., 2018; Schmugge et al., 2002). The increasing and unprecedented availability of SRS
data at increasingly finer spatial and temporal resolutions has triggered the development of large‐domain
water management applications including flood and drought monitoring (Hapuarachchi et al., 2011;
Klemas, 2014; Revilla‐Romero et al., 2015; Senay et al., 2015; Sheffield et al., 2012; Su et al., 2017; Teng
et al., 2017; Wu et al., 2014). The use of SRS data in water resources monitoring is promising, and it has
led to an increasing number of studies on a variety of topics in hydrology, including precipitation, evapora-
tion, and soil moisture estimation (Cazenave et al., 2016; Chen & Wang, 2018; Cui et al., 2019; National
Academies of Sciences, Engineering, and Medicine, 2019; Schultz & Engman, 2012). SRS data complement
in situ hydrometeorological data (Balsamo et al., 2018), which are typically scarce and whose unavailability
hinders the understanding of environmental systems (Tang et al., 2009). This aspect is particularly relevant
for developing countries where research for development initiatives have been increasing in the recent years
(Montanari et al., 2015).
Besides direct use of SRS data for water resources monitoring and management (Cui et al., 2019; Sheffield
et al., 2018), an increasing body of literature addresses the question of how these data sets can be used to
improve hydrological modeling (Baroni et al., 2019; Clark et al., 2015; Guntner, 2008; Liu et al., 2012;
Nijzink et al., 2018; Paniconi & Putti, 2015; Parajka et al., 2009). The scienti fic community has, in fact, long
been advocating the use of spatial data for DHM evaluation (Beven & Feyen, 2002; Grayson & Bloschl, 2001;
Koch et al., 2015; Refsgaard, 2001; Wealands et al., 2005). SRS data sets have the potential to improve models
either via data assimilation (Leroux et al., 2016; Tangdamrongsub et al., 2017; Tian et al., 2017) or via cali-
bration (Bai et al., 2018; Li et al., 2018; Rientjes et al., 2013). In this context, data assimilation is used to
update the states of a given model, e.g., to compensate for model structural deficiencies (Spaaks &
Bouten, 2013). For parameter estimation (i.e., model calibration) with SRS data, the existing approaches con-
sist in using SRS data alone or in combination with in situ data, usually streamflow data (Immerzeel &
Droogers, 2008; Li et al., 2018; Rajib, Evenson, et al., 2018; Wambura et al., 2018). Calibration of hydrological
models without concomitant streamflow data remains challenging, and attempts to do so have only shown
limited success (Nijzink et al., 2018; Silvestro et al., 2015; Sutanudjaja et al., 2014; Wanders et al., 2014).
The simultaneous calibration of hydrological models with streamflow and different combinations of
complementary data from SRS is increasingly discussed in recent literature (Stisen et al., 2018).
Multivariate (i.e., multiple variables) parameter estimation (Efstratiadis & Koutsoyiannis, 2010) can sub-
stantially reduce the feasible model and parameter space and lead to more realistic internal model
dynamics and related hydrological signatures (Clark et al., 2017; Shafii & Tolson, 2015), which can ulti-
mately enhance the overall representation of catchment functioning (Bergström et al., 2002; Rakovec
et al., 2016). Furthermore, and intimately linked to the above, multivariate calibration strategies can
considerably reduce equifinality (i.e., nonidentifiable model parameters in inverse modeling approaches;
Beven, 2006a; Savenije, 2001) and reduce prediction uncertainty (Fenicia et al., 2008; Fovet et al., 2015;
Gupta et al., 1998; Gupta et al., 2008; Hrachowitz et al., 2014; Schoups et al., 2005). However, important
open questions remain with respect to the combination of SRS data with streamflow data for model
parameter estimation. While some studies observed a significant improvement in the representation of
model outputs after SRS data incorporation (Chen et al., 2017; Leroux et al., 2016; Werth et al., 2009;
Yassin et al., 2017), others found minor changes or even major deteriorations (Stisen et al., 2018;
Tangdamrongsub et al., 2017; Tobin & Bennett, 2017). Such apparently contradictory conclusions are
case study specific and need to be understood as resulting from model structures, model parametriza-
tions, and trade‐offs between improving water balance estimates and related streamflow dynamics
and better representing other hydrological fluxes and states (Euser et al., 2013; Koppa et al., 2019;
10.1029/2019WR026085
Water Resources Research
DEMBÉLÉ ET AL. 2of26
Yassin et al., 2017). More generally, the key challenge results from the integration of several data
sources (SRS or in situ) in parameter estimation, which can be attributed to conflicting information
from different types of SRS data. Nonetheless, multivariate parameter estimation with SRS data remains
promising, especially when streamflow data availability is limited or the data quality is questionable.
Although SRS data are more accessible with higher spatiotemporal resolution compared to in situ observa-
tions, they are generally not direct measurements of hydrological processes, which adds a level of uncer-
tainty to any SRS‐based parameter estimation study (Ehlers et al., 2018; Knoche et al., 2014; Ma et al.,
2018). However, they provide spatial information on hydrological processes, which makes them a unique
and relevant information source for spatially distributed representations of the system in models (Stisen
et al., 2018; Wambura et al., 2018). For instance, many studies report different model performances when
using different satellite‐based products as input (e.g., precipiation; Pomeon et al., 2018; Thiemig et al.,
2013) or as calibration variables (e.g., evaporation, soil moisture, and terrestrial water storage; Bai, Liu, &
Liu, 2018; Nijzink et al., 2018). Nevertheless, for a given region, different products can give considerably dif-
ferent absolute values of a specific variable while they may exhibit plausible and similar spatial patterns
(Beck et al., 2017; Dembele & Zwart, 2016). Additionally, retaining only the spatial pattern information of
SRS data can substantially mitigate the uncertainty resulting from the fact that they are not direct observa-
tions, as long as their relative values are used rather than their absolute values (Mendiguren et al., 2017;
Wambura et al., 2018).
In the context of using SRS data for DHM calibration, the simultaneous use of more than one SRS product to
constrain several hydrological state or flux variables is uncommon (Clark et al., 2017; Lopez et al., 2017;
Nijzink et al., 2018), as is the incorporation of spatial pattern in the calibration scheme using bias‐insensitive
metrics (Demirel et al., 2018; Zink et al., 2018). Using different variables from SRS products simultaneously
in parameter estimation is in general not straightforward (Rajib et al., 2018; Silvestro et al., 2015; Tian et al.,
2017) because they all have limitations (e.g., spatiotemporal resolutions and accuracy), which can lead to sig-
nificant trade‐offs in multivariate calibration (Koppa et al., 2019).
In light of the above, we propose to test a novel multivariate calibration strategy in which a DHM will be
trained to simultaneously reproduce spatial patterns, i.e., relative spatial differences, of three variables from
different SRS products describing different components of the hydrological system (i.e., evaporation, soil
moisture, and terrestrial water storage), as well as in situ observations of streamflow. The proposed calibra-
tion framework combines simultaneously four noncommensurable variables and a new bias‐insensitive
metric for spatial pattern representation, which as a whole is different from previous studies (e.g., Demirel
et al., 2018; Koppa et al., 2019; Nijzink et al., 2018; Rakovec et al., 2016; Zink et al., 2018) and therefore makes
the novelty of this study. The following research hypotheses are tested:
1. Building upon previous work (e.g., Demirel et al., 2018; Rakovec et al., 2016; Zink et al., 2018), we assume
that simultaneously calibrating a DHM on four noncommensurable variables and spatial patterns of
satellite data considerably improves the predictive skill of the model, even for a DHM integrating a
MPR scheme.
2. Our new bias‐insensitive metric based on pixel‐by‐pixel locational matching can be used to improve the
calibration of a DHM on observed spatial patterns of hydrological processes even in the presence of
strong climatic gradients.
The overall goal of this study is to improve the spatial representation of dominant hydrological pro-
cesses of a DHM without significantly deteriorating the streamflow signal and reproducing plausible
dynamics of the hydrological system using spatial pattern information from SRS data sets. Such
improvement will be an asset for spatial hydrology and large‐domain water management applications
(e.g., water accounting, drought monitoring, and flood prediction) and might subsequently lead to
advances in prediction in ungauged basins (Blöschl et al., 2013; Hrachowitz et al., 2013; Sivapalan,
2003) with the use of readily accessible SRS data (Butler, 2014; Wulder & Coops, 2014). This work
embraces the fourth paradigm for hydrology (i.e., data‐intensive science, Peters‐Lidard et al., 2017)
and contributes to solving some of the issues (e.g., spatial variability and modeling methods) recently
identified as the 23 unsolved problems in hydrology in the 21st century (Blöschl et al., 2019). The pro-
posed multivariate calibration framework is tested with the mesoscale Hydrologic Model (mHM), with a
case study in the poorly gauged Volta River basin (VRB) in West Africa.
10.1029/2019WR026085
Water Resources Research
DEMBÉLÉ ET AL. 3of26
2. Study Area
The transboundary VRB is the study area. It covers approximately 415,600 km
2
across six countries of West
Africa. Figure 1 shows the physical and hydroclimatic characteristics of the VRB. The climate is character-
ized by a south‐north gradient of increasing aridity and varies from subhumid in the south to semiarid in the
north (Dembélé et al., 2019). Climate is driven by the Intertropical Convergence Zone, and four ecoclimatic
zones (i.e., Sahelian, Sudano‐Sahelian, Sudanian, and Guinean) can be identified (Figure 1a) based on the
average annual precipitation and agricultural features (Food and Agriculture Organization/Global
Information and Early Warning System, 1998; Mul et al., 2015). The characteristics of the four ecoclimatic
zones are given in Table 1. Actual evaporation exceeds 80% of annual rainfall in the basin (Andreini et al.,
2000; De Condappa & Lemoalle, 2009).
The topography is predominately flat as 95% of the relief is below 400 m above sea level (Figure 1b). The drai-
nage system is composed of four sub‐basins known as Black Volta (152,800 km
2
), White Volta (113,400 km
2
),
Oti (74,500 km
2
), and Lower Volta (74,900 km
2
). The Volta River flows over 1,850 km and transits in the
Lake Volta formed by the Akosombo dam before draining into the Atlantic Ocean at the Gulf of Guinea
(Williams et al., 2016). Land cover (Figure 1c) is dominated by savannah formed by grassland interspersed
with shrubs and trees covering about 88% of the basin area. Other land cover types include forest (9%), water
bodies (2%), and bare land and settlements (1%).
Figure 1. Physical and hydroclimatic characteristics of the Volta River basin.
Table 1
Characteristics of the Four Ecoclimatic Zones in the Volta River Basin
Ecoclimatic zones Climate class AI (−) P (mm/year) T
avg
(°C) T
min
(°C) T
max
(°C)
Sahel Savanna Arid 0.16 [0.12–0.20] 570 [470–610] 29 [29–30] 20 [20–21] 37 [36–38]
Sudano‐Sahelian Semiarid 0.29 [0.16–0.43] 790 [570–980] 29 [28–29] 20 [20–21] 36 [35–37]
Sudanian Savanna Semiarid/dry subhumid 0.47 [0.33–0.98] 1,010 [890–1290] 28 [26–29] 21 [19–23] 35 [32–36]
Guinean Savanna Dry subhumid/humid 0.70 [0.49–1.22] 1,190 [1030–1420] 28 [26–29] 21 [19–22] 34 [31–36]
Note. The annual mean value (with min‐max range in brackets) is given for each variable. The aridity index (United Nations Environment Programme, 1997) is
obtained from the global aridity index database (Trabucco & Zomer, 2018), and the WFDEI data (Weedon et al., 2014) are used for the long‐term (1979–2016)
estimation of annual precipitation and air temperature. AI = aridity index; P = precipitation, T
avg
= average air temperature; T
min
= minimum air temperature;
T
max
= maximum air temperature.
10.1029/2019WR026085
Water Resources Research
DEMBÉLÉ ET AL. 4of26
3. Data Sets
The data sets used to set up and run the distributed model for the 2000–2012 period include the basin mor-
phological data (elevation, slope, land cover, etc.) and meteorological data (i.e., rainfall and air temperature).
In situ streamflow data and complementary data from SRS are used to calibrate and to evaluate the model
performance. A description of the data sets with their characteristics and their sources is given in Table 2.
The streamflow data were obtained from different organizations (see the Acknowledgements section).
Previous work by Dembélé et al. (2019) describes the preprocessing of the streamflow data set, which was
quality checked and whose missing data portions were gap filled.
Concerning the SRS products, the terrestrial water storage (S
t
) anomaly data derived from changes in surface
mass, which is related to the Earth's gravity field, is obtained from the Gravity Recovery and Climate
Experiment (GRACE; Landerer & Swenson, 2012; Tapley et al., 2004). Over land, S
t
is the sum of snow,
ice, surface water, soil moisture, and groundwater. The data release RL05 (Swenson, 2012) is used in this
study. It is a simple arithmetic mean of different solutions from three processing centers: Jet Propulsion
Laboratory, Center for Space Research at University of Texas, and Geoforschungs Zentrum Potsdam.
Table 2
Overview of the Modeling Data Sets
Variables Product Spatial resolution Temporal resolution Reference
Model setup
Meteorological data
Rainfall CHIRPS v2.0 0.05° Daily Funk et al. (2015)
http://chg.geog.ucsb.edu/data/chirps/
Temperature (average, minimum,
and maximum)
WFDEI 0.5° Hourly Weedon et al. (2014)
http://www.eu‐watch.org/data_
availability
Morphological data
Terrain characteristics (elevation, slope, aspect,
flow direction, and flow accumulation)
GMTED 2010 225 m (0.0021°) Static Danielson and Gesch (2011)
https://topotools.cr.usgs.gov/
Soil properties (horizon depth, bulk density,
and sand and clay content)
SoilGrids 250 m (0.0023°) Static Hengl et al. (2017)
https://www.isric.org/explore/soilgrids
Geology GLiM v1.0 0.5° Static Hartmann and Moosdorf (2012)
https://doi.pangaea.de/10.1594/
PANGAEA.788537
Land use land cover Globcover 2009 300 m (0.0028°) Static Bontemps et al. (2011)
http://due.esrin.esa.int/page_globcover.
php
Phenology (leaf area index) GIMMS 8 km (0.0833°) Bimonthly Tucker et al. (2005), Zhu et al. (2013)
http://cliveg.bu.edu/modismisr/lai3g‐
fpar3g.html
Model calibration/evaluation
In situ data
Streamflow ‐ Point Daily Multiple organizations
(see the Acknowledgements section)
Complementary satellite products
Terrestrial water storage anomaly GRACE TellUS v5.0 1° Monthly Tapley et al. (2004), Landerer
and Swenson (2012)
https://grace.jpl.nasa.gov/
Surface soil moisture ESA CCI SM v4.2 0.25° Daily Dorigo et al. (2017)
https://www.esa‐soilmoisture‐cci.org/
Actual evaporation GLEAM v3.2a 0.25° Daily Martens et al. (2017),
Miralles et al. (2011)
https://www.gleam.eu/
Land surface temperature
(only for model evaluation)
MYD11A2 v6 1 km (0.0083°) 8‐day Wan et al. (2015)
https://lpdaac.usgs.gov/products/
myd11a2v006/
Note. CHIRPS = Climate Hazards Group InfraRed Precipitation with Station data; ESA CCI SM = European Space Agency Climate Change Initiative soil moist-
ure; GIMMS = Global Inventory Modelling and Mapping Studies; GLEAM = Global Land Evaporation Amsterdam Model; GLiM = Global Lithological Map;
GMTED = Global Multi‐resolution Terrain Elevation Data; GRACE = Gravity Recovery and Climate Experiment; WFDEI = WATCH Forcing Data methodology
applied to ERA‐Interim data.
10.1029/2019WR026085
Water Resources Research
DEMBÉLÉ ET AL. 5of26
