Global-scale river routing models (RRMs) are commonly used in a variety of studies, including studies on the impact of climate change on extreme flows (floods and droughts), water resources monitoring or large-scale flood forecasting. Over the last two decades, the increasing number of observational datasets, mainly from satellite missions, and increasing computing capacities have allowed better performance by RRMs, namely by increasing their spatial resolution. The spatial resolution of a RRM corresponds to the spatial resolution of its river network, which provides the flow directions of all grid cells. River networks may be derived at various spatial resolutions by upscaling high-resolution hydrography data.
This paper presents a new global-scale river network at
Global-scale river routing models (RRMs) were primarily developed for climate studies. By simulating the flow routing processes through river networks, they allow climate models to close the water budget at the global scale. Therefore, several applications have been developed based on RRMs, including studies on the impact of climate change on extreme flows (floods and droughts, see e.g.
Over the last two decades, the increasing number of observational datasets, mainly from satellite missions, and increasing computing capacities have allowed better performance by RRMs, either by improving the representation of some processes (e.g.
The river network, which mainly provides the flow direction of each cell, is the main component of an RRM. Higher spatial resolution allows narrower rivers to be represented and confluences to be better localized, with potential positive impacts on streamflow simulations. The river networks of most RRMs are either grid based or vector based. These approaches differ in their definitions of unit catchments. In grid-based approaches, the river network is discretized on a regular Cartesian grid, so that unit catchments are rectangular pixels. On the other hand, vector-based river networks are based on irregular shapes of unit catchments extracted from high-resolution hydrography data.
For instance, TRIP
For grid-based models, the spatial resolution is defined by the size of the grid pixels, while for vector-based models, the spatial resolution relies on a threshold catchment area. For both approaches, the river network is generally derived from the upscaling of high-resolution hydrography data.
The HydroSHEDS dataset
Generally, grid-based approaches follow the D8 convention, meaning that each grid cell may flow into one of the eight neighbouring grid cells. Vector-based approaches are more flexible and may follow the D
Along with the river network at the appropriate spatial resolution, RRMs also require parameters that are consistent with the river network. Some parameters depend on the river network itself, such as the lengths and slopes of river stretches, and vary with the spatial resolution. Other parameters, including the roughness coefficient, river width or bankfull depth, may be calibrated or estimated using empirical relationships. In the latter case, these parameters may also indirectly depend on the spatial resolution. Finally, several RRMs use sub-grid approximations to represent fine-scale processes. For instance, some flooding schemes (e.g. in
Although some recent studies provide a new upscaled river network based on MERIT-Hydro (see e.g. the river network at the river geomorphology (length, slope, depth, roughness) floodplain roughness and sub-grid topography aquifer characteristics and sub-grid topography.
A direct quantitative assessment is not possible since there is, to our knowledge, no equivalent existing dataset at the same spatial resolution. As a consequence, to evaluate the new river network and the corresponding hydro-geomorphological parameters, we propose to use the CTRIP model
The main purpose of this paper is to present the global river network at
The paper is organized as follows. The derivation of the river network at
This section describes the methodology used to derive the river network at
River network datasets consist of flow direction maps that are generally derived from digital elevation models (DEMs) corrected for hydrology. With the increasing amount of satellite observations performed during recent decades, several methods have been proposed to derive river networks at various spatial resolutions using upscaling algorithms (for the D8 method, see e.g.
Recently, the Multi-Error-Removed Improved Terrain DEM (MERIT-DEM) was proposed by
Here, we applied the DRT algorithm using MERIT-Hydro as the basis for the high-resolution hydrography network to benefit from the most recently available dataset.
The following notations and definitions are used throughout the article:
“HR” for high resolution ( “12D” for “HD” for half-degree resolution “pixel” for a unit element at HR “cell” for a unit element at 12D.
The first step in the generation of the river network is to set up a land mask at the final resolution (
Particular attention has been paid to estuaries and the effective connections of estuaries to oceans and seas. For example, it may happen that a large river flows into a narrow estuary. In this case, the cell corresponding to the river outlet may be disconnected from the coast by cells considered to be land (i.e. with more than 50 % HR land pixels). To ensure an effective connection to the coast, closed seas (water cells surrounded by land cells) with less than 20 cells are first converted to land. Then the HR land mask is used to find the shortest path within the estuary from the river outlet to cells marked as ocean, and cells that follow this path are converted to ocean cells. In this process, only rivers with flow accumulations greater than 10 000 pixels (HR) are considered. An example of an estuary in Ireland is presented in Fig.
Example of an estuary opening: the red mask is the HR land mask, the blue mask is the 12D land mask, and the green mask represents the 12D cells converted from land to ocean to connect the river basin delineated in red to the ocean.
Using the land mask as a basis, the DRT algorithm is applied to upscale the MERIT-Hydro river network from 3 arcsec to
Rivers are first sorted by decreasing flow accumulation at the outlet. Rivers are treated in this hierarchical order to ensure that the representation of large rivers is as good as possible. The following steps are applied for each river. Given the river outlet, the HR river route is defined as the longest upstream river. The headwater cell is given by the first cell with a flow accumulation larger than a given threshold (10 % of the cell size, i.e. 1000 HR pixels). The flow direction of the river at For each cell, the downstream cell is found when the HR route exits the cell with a minimum length (60 % of the cell size if cardinal, 80 % if diagonal; see e.g. cell C7 for river #2 in Fig. If a downstream cell is already assigned (e.g. by a larger river), the river is diverted: a parallel route is created that is as close to the real river as possible. If the outlet is reached, the presence of loops is checked for and corrected if necessary, and the next largest river is treated (steps 2–6).
Example of network upscaling in the Hérault basin (France). Basin boundaries are drawn in red. Rivers are treated in descending order of drainage area and drawn with different colours; solid lines are used for HR and dashed lines for 12D.
River diversion (step 5) is an important feature of the algorithm as it allows the structure of the network to be conserved, but it simultaneously may raise problems with changes of river location (e.g. localization of gauge stations). To overcome this issue, it may be useful to keep a track of the relationship between HR and 12D rivers, which is done here by identifying each processed river with a unique id in both the HR and 12D networks.
Note that while river diversion is necessary with the D8 convention, it can be avoided with the D
An example of diversion in the Loire River basin (France) is shown in Fig.
Hypsometry (elevation with respect to the longitudinal distance along the river) is computed during the process so that values of river length, slope and elevation are assigned to each cell at the end of the process. Figure
Examples of hypsometry curves for the rivers in Fig.
The final river network at 12D at the global scale is represented in Fig.
The type of river network required by most river routing models (especially those working with the D8 convention) has to provide a flow direction for each cell of the model. This ensures the closure of the global-scale water budget in earth system models. The type of soil (nature, river, lake, cities, etc.) and other characteristics (such as climate zone) are therefore not considered when setting up the global-scale river network. As a consequence, flow directions are also derived over arid regions where no river exists or within cells where no headwater stream has developed. In that sense, the river network should be considered a drainage network.
Global-scale river network at
Regional-scale river network at
The DRT algorithm has been designed to conserve the river network structure as much as possible. The hierarchical river selection and river diversion have been set up for that purpose. The quality of the resulting river network has been assessed by
The qualitative assessment consists of visual comparisons of the river network from different sources, including the original MERIT-Hydro, the previous version of the CTRIP river network (CTRIP-HD) and Google Earth images. The boundary shapes of the basins have also been compared with those from CTRIP-HD, the original DRT network at 12D and the GRDC database
The quantitative assessment first relies on the relative differences between the basin area from the newly developed 12D river network and those from other datasets, including the original DRT, MERIT-Hydro and GRDC.
In addition, to assess the basin shape and coverage, an intersection-over-union (IoU) index is computed as
Over the 69 largest basins, the overall agreement between MERIT-Hydro and the 12D river network is very good, with a median relative area difference of 0.3 %, which demonstrates the robustness of the upscaling algorithm. A large part of this difference can be attributed to basins crossing arid regions. When neglecting such basins, the mean relative difference drops from 5.8 % to 3.7 %. This cause of differences is discussed below.
Only two other basins are significantly different in the HR and 12D networks: those of the Nelson River and the Churchill River (Canada).
Both river basins are connected via the South Indian Lake. The natural outlet of this lake flows into the Churchill River, but the lake is anthropized, and a part of the lake volume is diverted to the Nelson River basin for management purposes. The developer of MERIT-Hydro chose the Nelson River to be the major outlet of the South Indian Lake, considering the existing diversion project. We decided to disconnect this outlet, preferring to preserve the natural river network. Figure
Region surrounding the South India Lake in Canada where the river network has been corrected to follow the natural outlet of the lake to the Churchill River. Blue and red lines represent the river network at 12D and HR (MERIT-Hydro), respectively. The yellow line corresponds to the Nelson River and Churchill River delineation from GRDC. The yellow circles show the cells where the flow direction has been inverted to reconnect the lake to the Churchill River. The blue and red background masks correspond to the Nelson River and Churchill River basins extracted from MERIT-Hydro, respectively.
When compared to GRDC and DRT, the averaged relative area difference equals 5.6 % and 8.4 %, respectively. The median reaches 0.8 % in the comparison with DRT. This shows that, except for a few basins, the 12D river network and the original DRT are quite close. In Table S1, cells showing a relative area difference higher than 0.10 (10 %) are highlighted, and the potential cause of the difference is indicated by the background colour. Three main causes have been identified.
Most of the differences with GRDC and DRT come from the arid conditions characterizing parts of some basins (with a red background in Table S1). In such regions, the terrain is generally quite flat and often disconnected to the river network (endoreic). It is thus quite difficult to extract river networks, which explains the differences between the datasets (for example, within the basins of the Tigris–Euphrates and the Yellow, Senegal, Xi and Rufiji rivers). Nevertheless, the small amount of precipitation that can fall in such regions only partly infiltrates and is mostly evaporated. This volume of water never reaches the river network, so differences between river networks over arid regions can be neglected. This can be accounted for in the IoU index by removing arid regions from basin masks, with arid regions being defined as regions where the mean annual runoff is below a threshold fixed at 1 mm yr
Tigris–Euphrates river system. River networks from the new algorithm and from DRT are drawn in blue and in cyan, respectively. Basin boundaries from the new algorithm, from DRT and from GRDC are drawn in green, magenta and orange, respectively. The overlapping blue mask represents arid regions. The IoU for this basin equals 14 %, which decreases to 8 % when the arid regions are removed.
Another source of differences is related to some missing tributaries (green background in Table S1 in the Supplement). This is the case for many river deltas, including those in the Tocantins, the Xi, the Ural, the Dvina and the Chao Phrava basins. With the D8 convention, models cannot simulate river divergence (a cell can flow into only one other cell). Figure
The last noticeable difference is in the Neva River basin. It appears that in GRDC and DRT, Lake Saimaa (Finland) is disconnected from the Vuoksi River that flows into Lake Ladoga (Russia). As for the South Indian Lake, a significant part of the water is derived from Lake Saimaa to feed canals used for anthropogenic purposes (hydroelectricity, fluvial transport), which may reliably explain the disconnection of this sub-basin.
Lower Mississippi basin and Red River basin joining the Mississippi Delta. The Mississippi River network is drawn in blue and the Red River in black, while their boundaries are shown in green and grey, respectively. The orange line represents the basin boundary of the Mississippi River from GRDC.
Finally, the upscaling algorithm produced a reliable and consistent global river network at 12D that was very close to the GRDC database in terms of basin delineation for the 69 largest basins of the world. Since MERIT-DEM improved the HydroSHEDS high-resolution river network
Large-scale river routing models make use of a river network (flow direction) to propagate runoff within river basins to the oceans (in the case of exorheic basins). But the propagation dynamics also depends on geomorphological characteristics. These include river geometry (length, slope, width) and roughness (friction coefficient). For models that simulate floodplains, the topography (generally given as the relationships between the surface elevation, the area of the floodplain and the volume of water) is also needed, as well as the roughness in the floodplains. Similarly, when simulating the dynamics of groundwater and the exchanges with rivers, additional parameters are needed, such as soil porosity and transmissivity (or hydraulic conductivity). For large-scale models, floodplains and groundwater are usually simulated using a sub-grid approach, for example via a description of the distribution of the topography with respect to the elevation within each cell. This section describes the derivation of river parameters, as well as floodplain and groundwater sub-grid distributions, consistent with the river network derived in the previous section.
A set of parameters related to rivers and describing the flow dynamics within the river network are derived in this subsection.
For each cell, the river slope and bed elevation parameters are directly derived from the original MERIT-Hydro adjusted elevation during the upscaling of the river network by considering the river reach at HR associated with each 12D cell. It should be noted that applying DRT is quite similar to the vector‐grid‐hybrid method
For the river length within each cell, the computation relies on the HR route within the cell, contrary to other methods that use the flow direction to compute the distance between the centre of the cell and the centre of the following cell and then multiply this distance by a constant meandering ratio (e.g. CTRIP-HD). Here, meanders are accounted for in the computation of distances in the HR river route.
The final river length within each cell is bounded between 1000 and 20 000 m. One may note that river reaches shorter than 1000 m correspond to headwaters, while reaches greater than 20 000 m correspond to highly meandering rivers.
The river slope is also bounded between 10
The river width
Distribution of river width from GRWL
Finally, smoothing is applied over each river (moving average with a 16-pixel width) to avoid unrealistic shrinkages (see Fig.
Examples of the combination of river widths from GRWL and Eq. (
The river depth
The last parameter related to the river hydro-geomorphology is the roughness coefficient. Here, we used the same methodology as in
Figure
River parameters for the Amazon (first row), the USA (second row) and Europe (third row): river slope
Floods may occur when the water height within the river exceeds the river depth, causing lateral flows over the river banks. A floodplain, described as an area surrounding a river that can be flooded during heavy rain events, provides water storage and directly impacts the water propagation within the river network.
High-accuracy representation of the flow dynamics within floodplains requires a highly accurate DEM and intensive computations to solve the 2D Saint-Venant equations. This can be done over small areas, such as urban areas, but not at regional scales.
A number of large-scale river models that account for floodplains and their impacts on the flow dynamics are based on sub-grid approximations
Here, in order to ensure consistency between the river network and the floodplain representation, the adjusted elevation from MERIT-Hydro is used as the baseline to compute the sub-grid distributions of elevation, cell fraction (related to the area) and volume of water within the floodplain. The method used to extract these distributions is described in
The floodplain roughness is used to estimate the flow velocity between the river and the floodplain, using the Manning–Strickler equation.
In addition, a floodplain roughness coefficient is estimated empirically as in
The 12 land types derived from the 1 km ECOCLIMAP-II database and their corresponding Manning roughness values.
Like floodplains, aquifers can significantly impact the propagation of water within rivers. Aquifers are usually recharged by the infiltration of water at the surface and can interact directly with rivers. The direction of the exchange between a river and an aquifer depends on the water elevation in the river and the water table depth. Just as for floodplains, some large-scale hydrology models (e.g.
To delineate the main aquifers that could be represented in large-scale hydrology models,
In
Aquifer numbering and parameters for the Amazon (first row), the USA (second row) and Europe (third row): aquifer number
Lastly, to simulate the exchanges between aquifers and rivers, the piezometric head has to be simulated and compared to the water level within the river. The piezometric head may be also used to represent upward capillary fluxes to the vegetation root layer
In this section, we set up a model configuration with the river network and the parameters described in Sects.
Both configurations are forced by runoff and drainage fields generated by the ISBA land surface model, as described in
Here we compare the performance of the new configuration (CTRIP-12D) to that of the previous one (CTRIP-HD). The performance mainly relies on comparisons between simulated and observed discharges at more than 10 000 in situ gauge stations over the globe.
A large number of in situ gauge stations have been considered for the comparison with simulated discharge. The data were extracted from various open-access databases described in Table
Description of the databases considered for the selection of in situ gauge stations with at least 3 years of discharge observations within the period 1979–2014. All websites were last accessed on 25 February 2021.
For the comparison between observed and simulated discharges, one must first localize the gauge station within the river network of the model. A very common method of achieving this consists of looking for the grid pixel around the station for which the drainage area is the closest to the one reported in the station metadata. However, in some cases, this can lead to the erroneous selection of the CTRIP pixel corresponding to a certain station. Such problems can happen unwittingly (see the example in Fig.
Since the coordinates and the drainage area of each station are known, it is possible to delineate the catchment related to the station from the MERIT-Hydro database. First, the pixel in the HR grid corresponding to the station is designated by selecting the pixel that minimizes a criterion that combines the distance to the station and the drainage area. At such a high resolution, the method can be considered robust enough to avoid mislocalization.
The second step consists of sorting the CTRIP pixels around the station (as in Fig.
Consequently, each station is assigned a CTRIP pixel more consistently than when using classical approaches. This process is applied for CTRIP-HD and CTRIP-12D. It also ensures that basins smaller than one grid pixel are excluded from the selection, since
Example of necessary relocalization using the mask overlapping method. The station (red dot) is the Oberndorf station (GRDC id 6342910) on the Danube River (48.947
The main metric used to quantify the performance of each simulation is the modified Kling–Gupta efficiency (KGE,
We also use the normalized information contribution (NIC), which is particularly suited to quantifying the improvement between two simulations, as in
In this section, the modelling results are evaluated by comparing simulated and observed river discharges at the 13 516 gauge stations selected from various open-access databases, as described in Sect.
Figure
To verify that poor performance is mainly due to these reasons and not to the new parametrization at 12D, the next section compares the performance of CTRIP-12D with that of CTRIP-HD when both are run in the same configuration.
Kling–Gupta efficiency for CTRIP-12D over 11 238 gauge stations (with KGE
NIC of the Kling–Gupta efficiency between CTRIP-12D and CTRIP-HD over 2164 gauge stations (with KGE
Considering that the CTRIP-HD model in its current version has been extensively validated (e.g.
Panels
By applying the methodology to localize gauge stations within the river network (see Sect.
To get a closer look at the difference in performance between CTRIP-HD and CTRIP-12D, panels in Fig.
Better performance could be expected for smaller basins, since these basins are represented by just a few cells at HD, and the difference between the basin delineation at HD and 12D could be relatively high, leading to different contributing areas. The better performance of CTRIP-12D for larger basins is less expected. Indeed, the processes and forcing are the same for both configurations, and the parameters are derived using similar strategies and relationships. The improvement in the correlation and variability demonstrates that a better-defined river network improves the dynamics of river propagation within the basin and interactions with floodplains and aquifers. Other potential sources of differences between the models include (1) the reference HR dataset (HydroSHEDS for CTRIP-HD, MERIT-Hydro for CTRIP-12D), which impacts the generation of floodplain and aquifer sub-grid parametrization, and (2) the use of observation-based river widths for CTRIP-12D.
The river network and hydro-geomorphology datasets (including the floodplain and aquifer parametrizations) are freely available for download from Zenodo (
This article has presented a new global-scale river network at
The new river network and hydro-geomorphological parameters have been implemented in a new version of the CTRIP model
For perspective, it should be mentioned that the derivation of some parameters for some regions could be improved by using existing local or national data. For example, aquifers could be better described by the Référentiel Hydrogéologique Français (BDRHF) database available for France, or by hydrogeological maps from USGS for the United States.
In grid-based approaches, the river network is discretized on a regular Cartesian grid, so that unit catchments are rectangular pixels with their own hydrogeomorphological characteristics. The complete dataset described here is particularly well suited to a number of large-scale RRMs that use a gridded structure for global hydrological studies (see Table 2 in
The supplement related to this article is available online at:
SM and BD both designed the study and contributed to the paper.
The contact author has declared that neither they nor their co-author has any competing interests.
Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We are very grateful to the four reviewers and the editor for their relevant remarks that helped us improving the clarity of the paper.
This paper was edited by Hanqin Tian and reviewed by Dai Yamazaki, Lishan Ran, and two anonymous referees.