A multidisciplinary framework to derive global river reach classifications at high spatial resolution

The GloRiC framework draws from landscape ecology theory (Wiens 2002), the riverscape concept (Fausch et al 2002, Carbonneau et al 2012) and environmental flow principles (Poff et al 1997, Richter et al 1997, Arthington et al 2006) in that all three concepts emphasize the importance of a multi-scale approach to river characterization. This is particularly true for the landscape ecology and riverscape concepts, in which river habitat characteristics are perceived to be the result of processes interacting at various spatial and temporal scales (i.e., ranging from micro-habitat morphology to climatic patterns) which in turn lead to an associated diversity of species and habitat within a reach (Tonn et al 1990). The development of aquatic classifications (including lakes, watersheds and rivers) has also been viewed as being either 'top-down' (divisive) or 'bottom-up' (agglomerative) (Kingsford et al 2005) and the choice of classification approach is application specific (Frissell 1986, Naiman et al 1992, Fausch et al 2002).

The GloRiC framework was conceptualized based on a systematic review of 60 river classifications that used abiotic and biotic classifier variables, ranging from watershed to global scales (see SI-section 1, available online at stacks.iop.org/ERL/14/024003/mmedia). In total, these classifications contained 633 individual classifier variables, of which 193 were unique. Based on a qualitative analysis of the compiled variables, GloRiC proposes a consistent scheme for separating potential classifier variables into one of five thematic categories, allowing for a multidisciplinary characterization of the river reach (figure 1). Each category represents key river system characteristics as determined by hydrology, physiography and climate, fluvial geomorphology, water chemistry, and aquatic biology. The five categories are intended to offer two application pathways: they can either serve to develop five individual baseline sub-classifications; or they can offer a versatile mix-and-match framework with which to compose multiple, hierarchical levels of sub-classifications and combined classifications that broadly satisfy different goal-oriented needs, such as those created to support the assessment of environmental flow requirements (Bejarano et al 2010, Belmar et al 2011, Mackay et al 2014). The first four categories and associated sub-classifications represent various abiotic aspects of river reaches, and we propose that they can be combined to create physical 'river reach types' representing reaches with similar environmental potential and resources. Aquatic biology could then either be integrated with the abiotic classification to examine biotic responses to physical environmental potential, or be treated as a separate level of stratification.

Zoom In Zoom Out Reset image size

River ecosystems differ greatly from their terrestrial counterparts in their accumulating nature. Flow quantities, the annual flow regime, or water chemistry are examples of characteristics that are propagated and accumulated along the river network. For example, land cover in the headwater area may affect the chemical composition of downstream water (Tetzlaff and Soulsby 2008, Withers and Jarvie 2008). To create a river classification, it is thus essential to include an upstream/downstream perspective and to be able to trace connections and network topology. In general, hydrologic and chemical variables may be more appropriately treated with a cumulative upstream watershed perspective, while physio-climatic and biological variables may be better suited to a local perspective. Geomorphologic properties of river reaches may derive from both local and upstream sources.

To acknowledge this inherent feature of fluvial systems, GloRiC considers rivers as connected, networked, linear features, and emphasizes the river itself as the primary spatial entity, rather than the (sub-)basins through which rivers flow. GloRiC identifies river reaches as the smallest spatial unit. The advantage of using reaches is that they can represent conditions along the reach itself; the immediate surroundings of the contributing sub-catchment for each individual reach; or the entire upstream watershed that contributes to the specific reach (figure 2). In contrast, using sub-basins as the primary spatial entity would limit or preclude the use of hydrological or fluvial geomorphology characteristics relating to the stream channel itself. Moreover, many sub-basins contain multiple river reaches that differ substantially, ranging from small tributary reaches originating within the sub-basin to larger main stem reaches traversing the sub-basin. Lumping heterogeneous reaches within a single sub-basin convolutes the understanding of the effective spatial resolution of the analysis.

Zoom In Zoom Out Reset image size

Application of the GloRiC framework to develop a new, standardized, high-resolution global classification of river reaches required a spatially explicit and consistent database of both river geometry and descriptive hydro-environmental attributes. The HydroSHEDS database (Lehner et al 2008) is widely considered to be the most consistent river reach dataset currently available at the global scale (SI-section 2). HydroSHEDS was thus selected to provide the spatial foundation of GloRiC at 15 arc-second resolution (i.e. the river reaches were extracted from pixels measuring approximately 500 m × 500 m at the equator). Rivers were derived in HydroSHEDS for all locations where the long-term average discharge exceeds 100 liters per second or the upstream catchment area exceeds 10 km², resulting in a total global river length of 35.9 million kilometers. Smaller rivers were omitted to avoid increasing uncertainties due to errors in very small spatial units, either in HydroSHEDS or in auxiliary global data. Rivers are broken into reaches at all confluences, creating approximately 8.5 million individual river reaches with an average reach length of 4.2 km (std. dev. = 4.8 km).

The geospatial river network data of HydroSHEDS is accompanied by a companion database, termed RiverATLAS, which provides a compendium of hydro-environmental attributes that were derived from existing global datasets and assigned to each river reach of HydroSHEDS (SI-section 2). RiverATLAS offers an open and readily available registry from which pre-calculated attributes were selected or new ones calculated to serve as the variables in developing GloRiC's sub-classifications and river reach types.

In order to create a first instance of global river reach types through supervised classification, we applied the GloRiC conceptual framework to the global river reach data and used statistical analyses to guide our decisions. First, we extracted a subset of river reach attributes that are commonly used in river classifications from the RiverATLAS database and derived an initial set of three sub-classifications (hydrologic, physio-climatic, and geomorphic). We could not create a chemical or biological sub-classification due to lack of reliable and consistent global data.

Our general classification approach is modeled after Olden et al (2012), and is based on (1) reducing the dimensionality of the data by screening potential input variables and selecting a subset of them to use as classifiers, and (2) subsequently implementing the classification. For the supervised classification, we reduced an initial set of classifier variables selected based on the systematic literature review using a principal component analysis (PCA) and a correlation matrix (SI-section 3).

After selecting the classifiers to be used, we derived the number of classes and class separation thresholds for each of the selected variables. The aim in this classification step was to identify a relatively low number of classes for each variable that reasonably represents natural groupings and major environmental features and avoids over-partitioning of the subsequent general global river reach types. A second goal was to use simple, replicable rules to increase the transparency of class separation. Class numbers and their value ranges were thus determined primarily by examining the statistical distribution of the values and separating them into quantile-based groups of either three, four, or five classes. The choice of using quantile-based groupings was informed by k-means clustering (SI-section 4).

The final step in the supervised classification was to combine the resulting classes from the three sub-classifications into one set of standardized, environmentally distinct global river reach types. The ultimate number of desired classes or types from any classification of geographic entities is a difficult consideration and there is almost always a trade-off between the representativeness (leading to a large number and a wider range of representation) and the manageability (prioritizing a lower number and reduced heterogeneity) of classes (Poff et al 2010). For the creation of global river reach types, we prioritized fewer, more generalized river reach types and aimed at keeping the number between 30 and 200. We believe that this range can depict both large-scale patterns, such as topography and climate, and finer scale patterns, such as flow regime magnitude and stream power, while still being manageable for global assessments or comparisons. To limit the number of resultant class-combinations from the three sub-classifications, we followed the approach of Sayre et al (2014) who proposed to first reduce the number of classes found in each sub-classification before combining them (SI-section 4).

In addition to the supervised classification exercise, we sought to reduce reliance on expert interpretation by implementing a second, more objective grouping of global river reach types. We therefore executed a multidimensional, unsupervised clustering of the data using an aspatial (not spatially constrained), all variables included, k-means clustering. The variables included in this multivariable clustering were the same set of variables used to develop the three sub-classifications described above.

K-means clustering was used because it can identify similarities across a large, multidimensional data space (Legendre and Legendre 2012). Variables were normalized and centered prior to their use in the k-means clustering algorithm. We used the simple structure index (Dolnicar et al 1999) to identify the most efficient partitioning solutions of k-means clustering between 30 and 200 final classes.

For the hydrologic sub-classification, two variables were selected through PCA (table 1): the long-term average discharge and a flow variability index, defined as the ratio of the highest average monthly discharge and the long-term average discharge. These two variables corresponded with the first two principal component axes, representing 83.1% and 6.4% of the variance, respectively. We divided the long-term average discharge into five logarithmic size ranges and the flow variability index into three ranges of variability (figure 3 caption for details). The combination of these ranges resulted in 15 hydrological classes (figure 3).

Zoom In Zoom Out Reset image size

The hydrologic class with the highest number of reaches is 'very low discharge, high variability' with more than 2.6 million river reaches. The resulting classes show a differentiation between more arid and more humid climatic zones. Large extents of arid and semi-arid areas such as the Sahara and Gobi deserts, Australia, and the American Prairies are dominated by the 'very low discharge, high variability' river class. River reaches located in large equatorial basins such as the Congo and the Amazon, as well as in large wetland complexes such as in Canada, Indonesia and Papua New Guinea are mostly classified as 'very low discharge, medium variability'. Note that some ephemeral reaches have no average flow (e.g., central Sahara, Atacama Desert) and are excluded from further analysis.

For the physio-climatic sub-classification, the first three main axes of the PCA, together representing 89.0% of variance, were associated with the long-term average of the minimum air temperature of the coldest month (52.3%); the climate moisture index (CMI) (23.9%); and elevation (12.8%). We divided the long-term average of the minimum air temperature into four ranges, the CMI into three ranges, and the elevation into two ranges (figure 4 caption for details). The combination of these ranges resulted in 24 physio-climatic classes (figure 4).

Zoom In Zoom Out Reset image size

The physio-climatic classes are generally similar to existing global climatic zonations, such as the Köppen–Geiger climatic regions (Peel et al 2007), yet regional topography exerts a strong influence in certain areas. This distinction from typical global terrestrial climate characterizations is important for a river-centric classification considering the importance of smaller-scale hillslopes on hydrology and sediment influx in river systems. The groupings show large, relatively homogenous areas of 'high or very high temperature, low moisture index, low elevation' in Africa, Australia, and the Middle East. Regions of 'very high temperature, medium or high moisture index, low elevation' are prevalent in Brazil, the Congo Basin and South-East Asia. Broad areas of the north are grouped into 'low temperature, low elevation' classes of varying moisture.

Of the three variables for the geomorphic sub-classification for which consistent global data at high spatial resolution were available (total stream power, lake-wetland influence, and stream gradient; table 1), we chose not to use the stream gradient because of its inclusion in the calculation of stream power. Given the limited number of remaining variables, no PCA was necessary for this sub-classification. The lake-wetland influence and the stream power were divided into two ranges each (figure 5 caption for details). This resulted in a combination of four geomorphic classes (figure 5).

Zoom In Zoom Out Reset image size

Due to the general global lack of river geomorphology data, the geomorphic sub-classification does not represent traditional classes of fluvial river morphology. Nevertheless, it successfully separated several large-scale distinctions, such as the vast extent of wetland-dominated areas in Canada and Alaska and low stream power areas in arid regions.

The direct combination of classes produced from the three sub-classifications (15, 24, and 4 possible classes from the hydrologic, physio-climatic, and geomorphic sub-classifications, respectively) yielded 1039 (out of 1440 possible) unique river reach types. As this number exceeded our targeted number of classes (between 30 and 200), we reduced the number of classes found in each sub-classification before combining them. The hydrology classes were reduced from 15 to five by using only the long-term average discharge ranges. The physio-climatic sub-classification was reduced from 24 classes to ten by merging all high elevation classes into two (cold and hot high elevation) and combining other similar classes. The geomorphic classes were reduced from four to three by merging the low and high stream power classes where lakes or wetlands are present. Also, after combining the sub-classifications, we refrained from distinguishing geomorphic classes for all rivers with 'very high discharge' according to the hydrologic sub-classification. This reduction resulted in a total of 150 possible (5 × 10 × 3) aggregated river reach combinations of which 127 were realized in the final typology (figure 6; SI-section 5).

Zoom In Zoom Out Reset image size

The most common river reach type globally is 'small river with low stream power and no presence of lakes or wetlands in very hot and high moisture climate' (20.8% of all reaches), followed by 'small river with low stream power and no presence of lakes or wetlands in warm and hot, high moisture climate' (11.5%). It should be noted, however, that arid and semi-arid river reaches may include many intermittent or ephemeral streams that do not flow for long periods of time, hence their actual extent at any one point in time is likely less dominant.

The k-means clustering resulting in 30 classes was chosen as the final clustering solution (figure 7). While at the low end of the range of investigated cluster number solutions, we chose this outcome to obtain a smaller number of natural groupings than we had identified in the supervised classification exercise (127). We felt that a smaller number of objectively-produced classes might be more relevant for freshwater management at a global scale. As the k-means clustering approach does not deliver explicit class thresholds, we refrain from translating the 30 numeric classes into more meaningful descriptions of dominant characteristics. However, in the clustering solution the influence of physio-climatic gradients is evident in the spatial pattern of rivers with smaller long-term average discharge. These small rivers are grouped into climatic regions based mostly on temperature and aridity. Unlike the supervised classification approach which yielded numerous types of large rivers, in the unsupervised clustering approach there was little separation of larger rivers into specific types. The most common class is number 11 representing 7.4% of all river reaches. This class correlates strongest with rivers with small and medium long-term average discharge in hot, humid areas.

Zoom In Zoom Out Reset image size

Fluvial and terrestrial ecosystems are different in terms of the processes that shape their respective habitats. While physio-climatic characteristics are often the basis for terrestrial or landscape classifications, hydrology and fluvial geomorphology are major driving factors for river systems in terms of habitat creation and disturbances. These latter two aspects of river reach types are crucial to the biota of fluvial environments and are explicitly covered in the GloRiC framework. As such, this framework represents a powerful communication tool that allows for a common vocabulary across river classifications. The flexibility of the GloRiC framework also supports and promotes the development of alternative, application-specific sub-classifications which could exploit the standardized and readily available attribute information offered in the RiverATLAS database.

The GloRiC framework provides an interdisciplinary approach to the classification of rivers and allows for rapid testing of the effects of parameter inclusion and omission. The inclusion of classifier variables that are derived both at the reach level (e.g. temperature) and with an upstream perspective (e.g. discharge) provides a landscape-integrated multi-scale perspective.

Our implementation of the GloRiC framework at the global scale using HydroSHEDS and RiverATLAS data delivered a spatially explicit set of 127 river reach types based on supervised expert classification, and 30 river reach types based on unsupervised statistical clustering. Both approaches represent characterizations and groupings of river reaches derived from hydrologic, physio-climatic and geomorphic properties. The availability of seamless georeferenced lines and attribute information for nearly 8.5 million river reaches represents an unprecedented spatial resolution for a river classification at the global scale. The resulting classifications allow for comparison across continents and provide a standardized baseline of different river reach types that can be used in a wide range of applications.

The geographical distribution of the resulting river reach types reflects general patterns at the global scale that are similar to coarser global eco-regionalizations, such as the 14 terrestrial biomes defined by Olson et al (2001), or the freshwater ecoregions of the world by Abell et al (2008). However, the inclusion of hydrologic variables, and in particular discharge magnitude, as well as a stronger focus on small-scale topography introduce a new geographic differentiation related to flow and local terrain heterogeneity.

River classifications to date have tended to focus on hydrologic variables (SI-section 1). Typically applied at local to regional scales, these classifications often rely on daily discharge data, which allows them to depict many specific characteristics of the river regime such as timing, rate of change, and recurrence (Melles et al 2014). In the absence of daily discharge data, the global river reach types presented here rely instead on long-term average monthly discharge. This limitation renders the results less specific and more limited to represent baseline conditions only. Nevertheless, the multidisciplinary GloRiC framework allows the inclusion of additional river characteristics based on physiography, climate and fluvial geomorphology to capture holistic patterns and elements of variability not covered by hydrology alone.

The presented first application of GloRiC does not include chemical or biological sub-classifications as proposed by the full framework, simply due to the lack of globally available and consistent data. Variables commonly used to represent the water chemistry of rivers, such as pH levels, salinity, or oxygen content are currently not available globally at high spatial resolution. Water temperature, however, may be partially reflected in the physio-climatic sub-classification. Similarly, the absence of biological data precludes the identification of regional differences in biodiversity patterns, such as distinct species assemblages—even for the same environmental river reach types—in different basins.

Any attempt to classify rivers at a global scale is dependent on available data. This first application of the GloRiC framework is based on globally comprehensive source datasets, which are typically less precise and accurate than locally available data. The datasets used to compile the global RiverATLAS compendium have certain limitations and uncertainties that need to be considered (SI-section 2). Particular caution is advised when using classes resulting from the hydrologic and geomorphic sub-classifications for smaller river reaches (0.1–1 m³ s⁻¹), as these are increasingly uncertain in the coarse-scale global model calculations. For projects undertaken at smaller scales, the GloRiC framework could be extended by incorporating data at finer spatial and temporal resolutions.

One of the main goals of this analysis was to derive a manageable number of final river reach types while capturing a wide variety of types across the globe. To achieve this goal, it was necessary to merge some similar classes. Hence, by design, the results of this analysis tend to highlight broad classes and omit rare or exceptional river reach types. For this reason, we caution against using the global river reach types in conservation planning efforts that emphasize the identification of rare habitat types. The GloRiC framework could, however, be applied to derive a greater number of river reach types if needed in other applications.

The GloRiC framework and underlying data have previously been used to create a classification of river reach types in the Greater Mekong Region (Lehner and Ouellet Dallaire 2013). We consulted with local experts during the process to select the best data for the region. In this data-poor area, the global datasets were well-suited given their consistency across political borders and high spatial resolution, demonstrating the advantage of using standardized global data when local data are scarce. During this Greater Mekong project, we found that the GloRiC framework provided a clear conceptual framework to easily identify the characteristics of the river systems that were deemed critical by the group of local, interdisciplinary experts.

Moreover, as a preliminary evaluation of our global results, we compared the output from the supervised classifications with classes from two regional river classifications located in Australia (Stein et al 2009) and in Spain (Belmar et al 2011). This preliminary evaluation indicated that the global application of GloRiC is capable of identifying patterns that are discernable at regional scales. Details of these comparisons are presented in SI-section 6.

Future work comparing the results of our global supervised classification with the results from our k-means or other unsupervised, aspatial, statistical clustering is encouraged. For example, it would be informative to implement clustering using statistical stopping rules (e.g. Metzger et al 2013) where the optimum number of clusters produced is a function of the intrinsic variability of the data, providing an instance of a completely unsupervised classification. As it stands, our unsupervised k-means clustering offers a novel, objective characterization of global river reaches for further evaluation by the river classification community.