This dataset succeeds the GlobCover dataset based on the data from the MERIS sensor (onboard ENVISAT) collected at high resolution for surface water detection, together with the SPOT-VEGETATION time series (Herold et al., 2015) to aid in distinguishing wetlands from other vegetation covers. Global land cover maps at approximately 300 m (10 arcsec) resolution deliver data for three 5-year periods (1998–2002, 2003–2007 and 2008–2012). The extents of water bodies slightly changed between the first 5-year period and the third one (such as shrinking of the Aral Sea area by more than 55 %), but the extent of wetland classes (permanent wetlands and flooded vegetation classes) did not change significantly (the variation in wetland classes throughout these periods is less than 3 % of the total wetlands area). We acquired the last epoch data to represent the current state of wetlands (Fig. 1b). In ESA-CCI, wetlands are mixed classes of flooded areas with tree covers, shrubs or herbaceous covers plus inland water bodies, covering 3 % of the Earth land surface overall.

Prigent et al. (2007) used multi-sensor satellite data, including passive and active microwave measurements, together with visible and near-infrared reflectance to map the monthly mean inundated fractions at 0.25∘ resolution for a 12-year period (1993–2004). This dataset (GIEMS) gives the minimum and maximum extent of the inundated area (including wetlands, rivers, small lakes and irrigated rice). Fluet-Chouinard et al. (2015) used the GLC2000 land cover map (Bartholomé and Belward, 2005) to train a downscaling model for GIEMS at 15 arcsec resolution based on the HydroSHEDS digital elevation model (Lehner et al., 2008) and developed three static datasets for mean annual minimum, mean annual maximum and long-term maximum extent of the inundated areas (covering 3.9 %, 7.7 % and 10.3 % of the land surface area, respectively). In this study, we assumed that the mean annual maximum extent was the best representative measure for wetlands. In the following, GIEMS-D15 always indicates the mean annual maximum of GIEMS-D15 (Fig. 1c). Higher-resolution (3 arcsec) downscaling of GIEMS has been recently developed (Aires et al., 2017), but we overlooked this source because we focused our study on 15 arcsec resolution.

The JRC surface water products are a set of high-resolution maps (1 arcsec, ∼30 m) for permanent water and also for seasonal and ephemeral water bodies. These products are based on analysis of Landsat satellite images (Wulder et al., 2016) over a period of 32 years (1984–2015). Each pixel was classified as open water, land or a nonvalid observation. Open water is defined as any pixel with standing water, including fresh and saltwater. The study also quantifies the conversions, mostly referring to changes in state (lost or gained water extents, conversions from seasonal to permanent, etc.) during the observation period. In this study, we used the maximum surface water extent, which consists of all pixels that were under water at least once during the entire period, covering almost 1.5 % of the Earth land surface area (Fig. 1d).

Fan et al. (2013) performed global GW modeling to estimate the water table depth at 1 km resolution. This model assumes a steady flow, and lateral water fluxes are calculated using Darcy's law and the Dupuit–Forchheimer approximation for 2-D flow. Elevation is described at 30 arcsec resolution (by HydroSHEDS south of 60∘ N and otherwise by ASTER/NASA-JPL), and the recharge rates were modeled at 0.5∘ resolution using the WaterGAP model (Döll and Fiedler, 2008) based on contemporary meteorological forcing (1979–2007). To estimate subsurface transmissivity, the soil hydraulic conductivities were derived from the global Food and Agriculture Organization (FAO) digital soil maps (5 arcmin resolution) and US Department of Agriculture (USDA) soil maps over the United States (30 arcsec resolution) and subsequently assumed to decay exponentially with depth from the thin soil layer (2 m) down as a function of the local topographic slope. The decay factor is also adjusted for the permafrost region using an additional thermic factor (smaller transmissivity in permafrost areas). The modeled WTD was compared to observations available to the authors (more than 1 million observations, with 80 % of them located in North America). The resulting dataset suggests vast areas with a shallow water table over the tropics, along the coastal zones and in boreal areas of North America and Asia (almost 15 % of the land area for WTD ≤20 cm).

Flat downstream areas display a marked propensity to be saturated, which explains the wide use of topographic indices to delineate wetlands. Here, we use the global map of TI produced by Marthews et al. (2015) at 15 arcsec resolution. It relies on the original formulation of Beven and Kirkby (1979), as in Eq. (1), and on two global high-resolution digital elevation models (DEMs), viz. HydroSHEDS (Lehner et al., 2008) and Hydro1k (US Geological Survey, 2000) at 15 and 30 arcsec resolution, respectively. Hydro1k is used to fill the lack of information in HydroSHEDS north of 60∘ N, which is outside of the SRTM (Shuttle Radar Topography Mission) coverage. Because index values depend on pixel size, which varies with latitude, those researchers also applied the dimensionless topographic wetness index correction of Ducharne (2009) to transform the index values to equivalents for a 1 m resolution.

Topography, however, is often not sufficient for wetland identification because climate and subsurface characteristics also control water availability and vertical drainage. Using the original TI formulation in Eq. (1), high index zones might coincide with flat arid areas, or inversely, low index values might occur at wetland zones with small upstream drainage areas over a shallow impervious layer. Several studies have focused on improving the topographic wetness index for wetland delineation by including other environmental factors or modifying the formulation of the wetness index (Rodhe and Seibert, 1999; Mérot et al., 2003; Manfreda et al., 2011). Therefore, we used the global TI dataset of Marthews et al. (2015) to supply the original TI, and also as a base map to derive two other variants of the index.

The first variant index is the TCI (topography–climate wetness index, inspired by Mérot et al., 2003): TCI=ln⁡a⋅Petan⁡(β)=TI+ln⁡(Pe), where Pe is the mean annual effective precipitation (in meters). The effective precipitation is first defined at the monthly time step as the monthly precipitation Pm,y (in meters) for month m and year y that is not evaporated or transpired using the monthly potential evapotranspiration EPm,y (in meters) as a proxy for total evapotranspiration: Pm,ye=max⁡(0,Pm,y-EPm,y). Pe is subsequently calculated as the sum of the 12 pluri-annual means of monthly effective precipitation. The required climatic variables are taken from the CRU monthly meteorological datasets (Sect. 2.2.3) for 1980–2016 to represent the contemporary period.

The second variant index (known as TCTrI, topography–climate–transmissivity index) is constructed by combining the effect of heterogeneous transmissivity (Rodhe and Seibert, 1999) with the above TCI: TCTrI=ln⁡a⋅PeTr⋅tan⁡(β)=TI+ln⁡(Pe)-ln⁡(Tr), where Tr (m2 s-1) is the transmissivity calculated by vertically integrating a constant Ks (saturated hydraulic conductivity in m s-1) from GLHYMPS over the first 100 m below the Earth's surface (Sect. 2.4.4).

To assess the impact of climate on wetlands, we used the Climatic Research Unit (CRU) monthly meteorological datasets. These datasets cover all land area from the beginning of the twentieth century (Harris et al., 2014). CRU climate time series are gridded to 0.5∘ resolution based on more than 4000 individual weather station records. To include a climate factor in the TI formulations, the time series of selected climate variables (i.e., precipitation and potential evapotranspiration based on the Penman–Monteith equation) are extracted for the contemporary period (1980–2016).

GLHYMPS is a global permeability and porosity map based on high-resolution lithology (Hartmann and Moosdorf, 2012). The permeability dataset and its derived hydraulic conductivity (Ks) estimates are given in vector format, with an average polygon size of approximately 100 km2. As noted by the developers of GLHYMPS (Gleeson et al., 2011, 2014), “lithology maps represent the shallow subsurface (on the order of 100 m)”, and thus hydraulic conductivity estimates are valid for the first 100 m of the subsurface layer. Thus, we estimated transmissivity as the integral of this constant Ks over these 100 m and used it to check whether use of the available transmissivity datasets in TI formulations can improve global wetland identification. It should be noted that the hydraulic conductivity dataset has two versions: with and without the permafrost effect. To consider the permafrost effect, Gleeson et al. (2014) used maps of the permafrost zonation index (PZI) from Gruber (2012) and homogenously assigned a rather low hydraulic conductivity (Ks=10-13 m s-1) for areas with PZI>0.99, i.e., in Siberian taiga forests and tundra, the Canadian Arctic Archipelago and Greenland. This choice leads to a very large contrast of Ks and transmissivity between permafrost and non-permafrost zones, which largely overrules the effects of lithology, so the high TI values (potential wetlands) become concentrated in permafrost areas. To preserve the influence of lithology, we rasterized the vector polygons of Ks without the permafrost effect to 15 arcsec resolution.

The GLWD, the Global Lakes And Wetlands Dataset, is based on the aggregation of regional and global land cover and wetland maps. This dataset contains three levels of information, and the most inclusive one is GLWD-3, which is in raster format. This dataset has an original 30 arcsec resolution and contains 12 classes for lakes and wetlands (maps and details are given in the Supplement, Sect. S1 and Fig. S1). For large zones prone to water accumulation but without solid information on existing wetlands, fractional wetland classes are defined (together they cover 4 % of the land surface area). This is particularly the case within the Prairie Pothole Region in North America and the Tibetan Plateau in Asia. Depending on the interpretation of fractional wetlands (by taking either the minimum, mean or maximum fraction of the ranges), wetlands cover between 5.8 % and 7.2 % of the land surface area. In this paper, we take the mean fraction in these areas, leading to a total wetland extent of 6.3 % of the land surface area.

Hu et al. (2017) proposed a potential wetland distribution using a precipitation topographic wetness index based on a new TI formulation in which the drainage area is multiplied by the mean annual precipitation. This formulation is based on the concept of the topography–climate wetness index (Mérot et al., 2003) in which the effective precipitation was introduced as the climate factor. The new index is calculated at 1 km resolution using GTOPO30 elevation data developed by the USGS. Wetlands are categorized into “water” and “non-water” wetlands based on regionally calibrated thresholds for each large basin of the world (level-1 drainage area of Hydro1k) using a sample trained adjustment model. The water classes of several land cover datasets are used to train the model for the water threshold, and the model for the non-water wetland threshold is trained on the regularly flooded tree cover and herbaceous cover categories (additional details are available in the Supplement, Sect. S1 and Fig. S2). The global coverage of the water and non-water wetland classes in Hu et al. (2017) is 22.6 % of the Earth land surface area (excluding lakes, Antarctica and the Greenland ice sheet), considering no loss due to human influence. This dataset gives the largest wetland extent within the accessible literature, with notably large water wetlands in South America and large non-water wetlands in central Asia and the North American continent. In this paper, we used the union of the water and non-water wetland classes of this dataset for further evaluations.

Hess et al. (2015) used the L-band synthetic aperture radar (SAR) data from the Japanese Earth resources satellite (JRES-1) imagery scenes at a 100 m resolution to map wetlands during the period 1995–1996 for high and low water seasons. The studied domain excludes zones with altitudes higher than 500 m and corresponds to a large fraction of the Amazon basin (87 %). Wetlands are defined as the sum of lakes and rivers (both covering 1 % of the basin area) and other flooded areas plus zones not flooded but adjacent to flooded areas and sharing wetland geomorphology. The flooded fraction of wetlands varies considerably (from 38 % to 75 %) between the low and high water season. The total maximum mapped wetland area extends over 0.8 million km2 and is used in the evaluation of CW maps in this study.

The map of potentially wet zones in France (les Milieux Potentiellement Humides de France Modélisée, MPHFM; Berthier et al., 2014) constructed at a 50 m resolution is based on the topo-climatic wetness index (Mérot et al., 2003) and the elevation difference to streams using the national high-resolution DEMs. Meteorological data for the calculation of the topo-climatic index (precipitation and potential evaporation rates; see further details in Sect. 3.2.2) are taken from the SAFRAN atmospheric reanalysis (Vidal et al., 2010) at 8 km resolution. Index thresholding for wetland delineation is performed independently in 22 hydro-ecoregion units and delimited based on lithology, drainage density, elevation, slope, precipitation rate and temperature. The wet fraction defining the threshold in each hydro-ecoregion is the fraction of hydromorphic soils (extrapolated from local soil maps to almost 18 % of the area of metropolitan France) taken from national soil maps at 1 : 250 000 (InfoSol, 2013). Additionally, the elevation difference between land pixels and natural streams was used to separate large streambeds and plain zones, which are difficult to model with indices based on topography. Based on MPHFM, potential wetlands extend over almost 130 000 km2 of France (23 % of the area of metropolitan France). The dataset was validated against available pedological point data (based on profiles or surveys) available over France. These point data are classified into wetlands and non-wetlands for the validation procedure. This procedure used statistical criteria such as spatial coincidence (SC; number of correctly diagnosed points over total number of points) and the kappa coefficient (modeling error compared with a random classification error).

The wetland definition behind the composite maps is focused on hydrological functioning, and we aim to include both seasonal and permanent wetlands as well as shallow surface water bodies (including rivers, both permanent and intermittent). Surface water bodies and wetlands are often hydrologically connected, and the transition between them is not sharp and varies seasonally. Moreover, these features are difficult to separate based on observations (either in situ or remote), and no dedicated exhaustive dataset is currently available (Raymond et al., 2013; Schneider et al., 2017). Inclusion of the shallow surface water bodies (in the RFW map) is compatible with the Ramsar classification, but we depart from this approach with respect to large permanent lakes, which are excluded from all input datasets to RFW and GDW maps (Sect. 2.2) because of their distinct hydrology and ecology compared with wetlands. In contrast, groundwater-driven wetlands can remain wet without inundation due to the presence of shallow water tables. As further discussed in Sect. 3.2, these areas are defined in this study as areas where the mean annual WTD is less than 20 cm, following similar assumptions in the literature (US Army Corps of Engineers, 1987; Constance et al., 2007; Tamea et al., 2010; Fan and Miguez-Macho, 2011).

Based on this definition, another feature of the proposed wetland maps is that they are static. As stated in Prigent et al. (2007), the maps represent the “climatological maximum extent of active wetlands and inundation” (for CWs and RFWs, respectively), i.e., the areas that happen to be saturated or near-saturated sufficiently frequently to develop specific features of wetlands (high soil moisture over a significant period of the year, potentially leading to reducing conditions in selected horizons and specific flora and fauna).

To project, resample, intersect/overlap and convert different datasets used in wetland mapping in this study, we relied on ArcMap software (Esri, ArcGIS Desktop: Release 10.3.1 Redlands, CA) and its different tools. All datasets were projected to a WGS84 equi-rectangular coordination system and subsequently resampled to a single resolution for facilitated fusion and comparison. The resulting raster datasets were processed with ArcMap tools available in almost any GIS software such as QGIS (Table 3).

The final resolution of the maps is targeted to 15 arcsec (∼500 m at the Equator) for consistency with the available water datasets. Therefore, all datasets were resampled to 15 arcsec resolution, which is within the resolution range of state-of-the-art wetland-related datasets. For datasets at coarser resolutions, each coarse pixel is disaggregated to 15 arcsec while retaining the same value. We used an “all-or-nothing” approach; i.e., the pixels are either fully recognized as wetland (or lake) or not at all, based on the dominant type if the input data are finer than 15 arcsec (ESA-CCI land cover and JRC surface water).

Eventually, each 15 arcsec global raster contains more than 80 000 pixels along a circle of 360∘ of longitude, and wetlands can exhibit notably small-scale patterns (e.g., patchy or river-like). To facilitate visual inspection, we calculated the mean wetland densities in 3 arcmin grids for most of the maps presented in this work. The same 3 arcmin resolution (∼6 km at the Equator) was used in calculating the spatial correlations. For zonal wetland area distributions, the area covered by wetlands in each 1∘ latitude band is displayed.

To identify the RFWs, we overlapped carefully selected datasets of surface water, land cover and wetlands, namely, the ESA-CCI land cover, GIEMS-D15 inundation surface and the maximum water extent in JRC surface water. These datasets were selected to include different types of data acquisition. The idea behind the fusion approach chosen in this work is that wetlands identified by the different datasets are all valid despite their uncertainties, although none of them are exhaustive. As a result, use of multiple inundation datasets fills the observational gap. Several other surface water datasets exist that were not used in this work, either because they mostly consist of lakes or because they rely on similar methodologies (Verpoorter et al., 2014; Yamazaki et al., 2015).

Overall, the RFW map covers 9.7 % of the land surface area (12.9 million km2) including river channels, deltas, coastal wetlands and flooded lake margins (Fig. 1e). Areal coverage of the RFWs is by definition larger than the area of wetlands in all three input datasets (Fig. 1b–d), which were selected to be representative of different types of data acquisition (sensors and wavelengths). Therefore, they correspond to different definitions of inundated areas, and their contribution to the RFW map is fairly different. In particular, the shared fraction of the three input maps is minuscule (5 % of the total RFW land surface area coverage), and is mostly composed of the large river corridors and ponds which are detectable by satellite visible range imaging techniques in the JRC dataset. The latter misses most understory inundations, which are better identified by the ESA-CCI dataset owing to specific vegetation classification. Finally, owing to the use of microwave sensors, GIEMS-D15 extends over larger areas since it captures both flooded areas and wet soils below most vegetation canopies except the densest ones (Prigent et al., 2007). In addition, the distribution of wetlands in GIEMS-D15 involves downscaling as a function of topography, and can be very different from the other datasets. Hence, 58 % of RFWs are solely sourced from GIEMS-D15, mostly in the Southeast Asian floodplains, Northeast Indian wet plains and rice paddies, and the Prairie Pothole Region (in the northern US and Canada). The ESA-CCI contribution is mainly found in the Ob River basin where wetland vegetation exists but wet soils are not easily detected by visible (JRC) or microwave (GIEMS-D15) observation. Due to its high resolution, JRC surface water adds small-scale wetlands such as patchy wetlands, small ponds and oases (0.4 % of the land surface area).

In terms of zonal distribution, 31 % of the RFWs are concentrated north of 50∘ N, with most of the wetlands formed in the Prairie Pothole Region and Siberian lowlands. Cold and humid climates and the poorly drained soils of the boreal forest regions in northern Canada on the Precambrian shield are the main hotspots of peat in the American continent. The same situation exists in the West Siberian Plain as well. The second zonal peak in RFWs lies between 20 and 33∘ N, where the major contributors are the vast floodplains surrounding the Mississippi, Brahmaputra, Ganges, Yangtze and Yellow rivers and Mesopotamian Marshes. A total of 30 % of the world's RFWs are found in tropical regions (20∘ N to 20∘ S), concentrated mainly in the Amazon, Orinoco and Congo river floodplains and in inundated portions of wetlands such as the Sudd swamp in South Sudan.

Due to a lack of integrated, standardized and globally distributed WTD observations, a sound approach to the location of groundwater-driven wetlands is the use of available global direct GW modeling results. In this study, we used the global WTD estimations of Fan et al. (2013), and the resulting wetland map is denoted as GDW-WTD. As explained in Sect. 3.1.1 we assumed the mean annual WTD in wetlands to be less than 20 cm, which results in a wetland area extending over 15 % of the land surface, with large wetlands in the northern areas and the Amazon basin (Fig. 2a). We also performed a sensitivity analysis on the areal fraction of wetlands with different WTD thresholds (Sect. S2, Figs. S3 and S4), revealing that the variation in total wetland fraction is quite weak (between 13.7 % and 16.7 %) for thresholds ranging from 0 to 40 cm. Therefore, a 20 cm threshold appears to be a credible representative value. However, the wetland fraction rapidly increases for deeper thresholds, showing that a clear distinction exists between shallow WTD areas (wetlands according to our definition) and the remainder of the land.

In line with many studies (Rodhe and Seibert, 1999; Curie et al., 2007; Hu et al., 2017), we define TI-based wetlands as the pixels with a TI above a certain threshold, defined to match a certain fraction of total land. In doing so, we prescribe the global GDW fraction as a chosen value, and the various TI formulations (Sect. 2.4.2) only change the geographic distribution of the corresponding wetlands. To apprehend the uncertainty related to the choice of the global GDW fraction, we tested two choices within the bounds derived from the global WTDs of Fan et al. (2013). In the first approach, we set the TI threshold such that the wet pixels (with high index values) cover 15 % of the land surface area, such as the fraction of WTD ≤ 20 cm according to Fan et al. (2013). The corresponding maps are noted as GDW-TI15, GDW-TCI15 and GDW-TCTrI15 in Table 2 and show fairly different patterns (Fig. 2b–d). The second approach assumes that the total wetland extent (this time including both GDWs and RFWs) covers 15 %. The TI thresholds are subsequently set such that the union of RFWs and GDW-TI (TCI/TCTrI), i.e., the composite wetlands, has the same extent as GDW-WTD. The resulting GDWs cover between 6 % and 6.6 % of the land area depending on the TI formulation and level of overlap with RFWs (Table 4) and are noted as GDW-TI6, GDW-TCI6.6 and GDW-TCTrI6. The patterns of these three maps are highly similar to those of GDW-TI15, GDW-TCI15 and GDW-TCTrI15, with diminished extents and densities (Fig. 2e–g).

As shown in Table 2, seven GDW maps are developed, consisting of GDW-WTD (Sect. 3.2.1) and six GDW-TIs (Sect. 3.2.2). The GDW-WTD map contains high wetland extents over the northern latitudes (Fig. 2a), in contrast to the other six GDW maps. The diagnosed wetlands of GDW-TI maps (Fig. 2b, e) are equally distributed over well-known arid areas such as the Sahara and the Kalahari Desert, the Australian Shield and the Arabian Peninsula as in wet regions such as the West Siberian Plain and northern Canada (Fig. 2b, e). As a result, for a given threshold (15 % in Fig. 3a), the distribution of wetlands derived from the simple TI is nearly uniform over different latitudes. Lower thresholds on TI variants (Figs. 2e–g and 3b) obviously result in a smaller wetland extent, with no major change in the zonal pattern when the wet fraction threshold changes from 15 % to 6 % (Figs. 2b–d and 3a, b).

Introducing a climate factor in the form of effective precipitation in GDW-TCI6.6 and GDW-TCI15 increases the value of the wetness index in wet areas and decreases it in dry climates (Figs. 2c, f and 3a, b). Therefore, previously diagnosed wetlands with a TI in dry climates disappear and transfer to regions with wet climates (such as the Amazon basin and South Asia). However, because transmissivity values sharply change by several orders of magnitude over regions with small permeability, the patterns of GDW-TCTrI maps are nearly replicas of the low hydraulic conductivity distribution in GLHYMPS (e.g., large diagnosed wetlands in North America and central Asia; Fig. 2d, g). Although at times GDW-TCTrI coincides with famous wetlands such as the Pampas in South America (Fig. 2d, g and near 25∘ S in Fig. 3a), diagnosed wetlands extend far beyond the actual wet regions into neighboring arid or semi-arid zones; e.g., vast diagnosed wetlands in the western Siberian lowlands extend southward towards the Kazakh upland arid zones. In the absence of precise and consistent information on subsurface characteristics (particularly for cold areas), GDW-TCTrI shows low wetland densities in zones with the known effect of transmissivity, such as the Hudson Bay Lowlands and the Prairie Pothole Region.

With the exception of CW-WTD, which is always more similar to GDW-WTD because the latter is a component of the former, the validation criteria for the CW maps are rather small overall (between 0.2 and 0.6). However, the criteria are larger than the same values between the surface water and wetland datasets (less than 0.3 in Table 5 for the SPC of the globe and Table S1), showing their advantages. CW maps (especially CW-TCI maps) are more similar to GDW-WTD and Hu's map with respect to GLWD-3 because all but GLWD-3 share the GW modeling methodology. In contrast, the RFW map extends over a 60 % larger surface area than GLWD-3 and displays the highest similarity to GLWD-3, suggesting that wetlands in GLWD-3 are the regularly flooded ones. The inclusion of GDWs in the CW maps makes them depart from GLWD-3, but it markedly increases their similarity to the other two validation datasets for JI and SPC (e.g., SPC [RFW, GDW-WTD] = 0.3 versus SPC [CW-TCI15, GDW-WTD] = 0.6). As demonstrated in Fig. 4a (and also Table S1), increasing the GDW contribution from CW-TCI6.6 to CW-TCI15, as an example, also improves the similarity criteria (except the SC for GLWD-3 and GDW-WTD), justifying the need to account for the GDWs to provide a comprehensive description of wetlands. This is clearer for the global spatial correlation values which all increase when the contribution of GDW is increased from 6.6 % to 15 % (Table S1, first row block).

The following section breaks down the comparative wetland representation between our maps and that of the validation datasets at the regional scale. The selected regions encompass different climates, vegetation covers and ecosystems, both within and outside important wetland areas of the world, to assure the applicability of CW maps. These six regions are France in the temperate climate, the Amazon basin and Southeast Asia over the tropical zone, the cold boreal areas of the Hudson Bay Lowlands and the Ob River basin and the Sudd swamp in South Sudan with a semi-arid savannah climate.

Over France, wetland fractions from the validation datasets are highly inconsistent (Fig. 5). Visible range satellite imagery (JRC surface water) shows the smallest wet fraction (1 %). The GLWD-3 and ESA-CCI maps also produce low wetland coverage, whereas GIEMS-D15, which essentially forms the RFW map, gives 12 % coverage concentrated along the coastline and over the floodplains of the northern rivers. Wetlands from GW modeling-based datasets cover even larger areas (14 % and 18 % in GDW-WTD and Hu et al., 2017) and are scattered countrywide, except for the French Pyrenees and the Alps, with moderately denser wetlands along large rivers (such as the Rhine floodplain at the eastern border) and the Landes (southwestern shore). The MPHFM map (Berthier et al., 2014) can be considered as the most comprehensive validation dataset for the country because it relies on hydromorphic soil properties and was extensively validated. This map shows much larger wetland extents (23 % of France) than the above estimates because of its inclusion of both floodplains (along the Loire, Saône and Rhône rivers) and groundwater-driven wetlands, including those over the weakly permeable granites of Brittany (shown in green in Fig. 5g). These notorious wetlands are not considered in the global validation datasets but are captured to a good extent in CW maps (Fig. 5i, j).

By combining RFWs (which overlap with 20 % of MPHFM) and GDWs, our CW maps capture many features of the MPHFM map, including the total wetland extent (23 % for MPHFM versus 22 % and 25 % for CW-WTD and CW-TCI15) and correctly capturing most of the coastal and riparian wetlands (Fig. 5). The larger wetland fraction in MPHFM and CW maps is consistent with the work of Pison et al. (2018), who found that (wetland-driven) methane emissions over France deduced from atmospheric inversion were almost a third higher than direct estimates, from anthropogenic inventories and biogeochemical models driven by global wetland datasets (e.g., the overlap of GLWD and SWAMPS in Saunois et al., 2016). The added value of CW maps is demonstrated by the higher similarity criteria between CW maps such as CW-TCI15 and MPHFM (SPC = 0.52) than between surface water maps such as GIEMS-D15 and MPHFM (SPC = 0.43). However, it is difficult to identify the best CW map over France based on the similarity criteria against MPHFM because four of our CW maps (all shown in grey in Fig. 4b) display nearly the same values (Table S2).

The Amazon River basin is considered one of the richest tropical wetland ecosystems in the world (Mitsch and Gosselink, 2000). For ease of comparison, we limited our study to the domain of Hess et al. (2015), which covers 5 million km2 (Fig. 6). RFWs (mostly consisting of GIEMS-D15) show a pattern rather similar to that of GLWD-3 and Hess et al. (2015) (Figs. 4c and 6d, g, h), covering only the main drainage network of the Amazon and certain seasonally flooded wetlands and floodplains. However, certain spatial disagreements exist among these three datasets in seasonally flooded wetlands such as Llanos de Moxos (12∘30′–17∘30′ S, 63–68∘ W), the Roraima savannah and the Negro River basin (2∘ N–2∘ S, 60–65∘ W), which are larger in RFWs and Hess et al. (2015) than in GLWD-3.

The CW maps capture the wetland pattern of GDW-WTD and Hu et al. (2017) considerably better than RFWs (Fig. 6), highlighting the significance of groundwater wetlands over the Amazon. Wetland densities in CW maps, Hu et al. (2017) and GDW-WTD are more realistically high over the leached and swampy soils of the northern Amazon basin (e.g., Japurá–Solimões–Negro moist forests) and over the Purus–Madeira ecoregion, in line with recent estimates of wetlands and peatlands (Hess et al., 2015). This result suggests that the extended shallow peatlands of South America are the main causal contributor to the global tropical wetland extent (Gumbricht et al., 2017). The higher wetland densities of CW maps with respect to all satellite observations over these particular areas can be attributed to the coincidence of GDWs and dense rainforests (covering almost two-thirds of this domain), with large non-flooded wetlands over the Amazon. River channels and surrounding floodplains are better represented in CW maps, as compared with Hess et al. (2015), due to the inclusion of the RFW component. Similarly, CW compares well to other datasets because almost none of the river floodplains are delineated in Hu et al. (2017), and GDW-WTD misses the Tapajós River floodplain and portions of the downstream Amazon corridor. However, CW maps represent the wetland extent in lower density over grassland/savannahs and the Andes dry regions compared with the validation datasets.

The selected window over South and Southeast Asia stretches over notably wet regions, similar to the Amazon, but with severe human interference and deforestation (Miettinen et al., 2011; Stibig et al., 2013). In Southeast Asia, RFWs (mostly composed of GIEMS-D15) are larger than all validation datasets (Fig. 7d–g) because GIEMS-D15 also detects inundated areas associated with cultivation activities such as rice paddies (Fluet-Chouinard et al., 2015), which are not considered in inventories and GW modeling-based estimates. Over the window, RFWs and CWs coincide with the majority of wetlands in the validation sets, particularly over the Ganges–Brahmaputra floodplain, northern Indochina and Yunnan Plateau subtropical forests (Fig. 4d; SC between 0.59 and 0.91), showing the good agreement of our developed maps with respect to spatial patterns. As a general rule in Southeast Asia, floodplains and deltas (Ganges, Brahmaputra, Irrawaddy, Mekong and Red rivers) extend over larger areas in CW maps than in validation maps (Fig. 7), giving a more realistic extent than those in Fan et al. (2013) and Hu et al. (2017), considering the vast flood irrigated cultivation lands along floodplains. However, only a few small wetlands in the validation datasets are missed in RFWs (and CW maps), such as the upstream Mekong River corridor (near 20∘ N–102∘ E) and Irrawaddy River (near 25∘ N–97∘ E) in GLWD-3.

The CW-WTD and CW-TCI15 maps present patterns that are highly similar to each other (Fig. 7h, i) and to the validation datasets. However, high similarity criteria (especially SC) can be the result of the large extension of RFWs, itself overlapping almost all of the wetlands in the validation datasets. In addition, the similarity of CW-WTD and CW-TCI15, also derived from similarities between their GDW components, notes that groundwater wetland formation is almost completely explained by topography and climate (of the TCI formulation) in these areas and the negligible role of subsurface characteristics included in GDW-WTD.

The Hudson Bay Lowlands (HBL) are a vast flat wetland area in the low subarctic regions of North America dominated by extensive peatlands, swamps and marshes (Mitsch and Gosselink, 2000; Packalen et al., 2014), where below-freezing temperatures for most of the year reduce drainage in the soil layer (Hamilton et al., 1994). A systematic contrast is noted between inundation maps (Fig. 8a–c; maximum wet fraction: 21 %) and validation datasets (Fig. 8d–f; minimum wet fraction: 49 %), underlining the inability of satellite imagery to capture wetlands in this area (e.g., Landsat images used in JRC surface water; Fig. 8c). Surprisingly, GLWD-3 has a pattern notably similar to that of the other two validation maps due to the comprehensive wetland maps in Canada available to its developers. Moreover, HBL is one of the few regions where similarity indices sharply increase with increased GDW contribution (Table S1). The Jaccard index rises from 0.46 to 0.53 when increasing the total GDW extent from 6.6 % to 15 % between CW-TCI6.6 and CW-TCI15. CW maps perform fairly well, particularly CW-WTD, which predominantly obtains the highest validation criteria (Fig. 4e). Due to an explicit parameterization of the permafrost (adjusted to reproduce the “observed wetland areas” in North America; Fan and Miguez-Macho, 2011), dense wetlands are extended south of 50∘N in the GW model by Fan et al. (2013), which are less dense in CW-TCI15 in the absence of a soil-freezing mechanism. Comparing wetlands detected through satellite imagery and validation datasets, GW modeling appears to be the best wetland delineation method over boreal zones due to nonpermanent surface inundation, shallow WTD or snow/ice cover.

The Ob River basin in western Siberia extends over ∼3×106 km2. The annual variability of the inundated area is large (e.g., Mialon et al., 2005), making this basin one of the largest wetland complexes in the world, which contributes to buffer peak discharge during the flooding period (Grippa et al., 2005). Wetland fractions in different datasets compare similarly to HBL, except for GLWD-3, which appears to underestimate the total wetland extent although the climatic and geomorphologic properties are nearly alike. Datasets recognizing the contribution of GW to wetland formation (Fig. 9e, f, h, i) indicate consistently higher wet fractions than others. However, CW-TCI15 appears to miss wetlands south of 60∘ N that are extended to the upstream Ob River basin near 50∘ N in both GDW-WTD and Hu et al. (2017), most likely due to the permafrost effect on wetland formation. With respect to the evaluation criteria, CW-WTD often displays better performances, although CW-TCI15 shows the highest SPC. We also find that CW-TCI15 outperforms CW-TCI6.6 for all criteria and validation dataset combinations (Table S1). TCTrI-based CW maps fail to surpass others in the validation process, considering that we used the transmissivity map without the permafrost adjustment due to its imprecise representation of hydraulic conductivity in these zones (Sect. 2.4.4). CW-WTD tends to better capture the wetland extent and spatial pattern, with more concentrated wetlands in the downstream lowlands and northwestern regions (65∘ N–65∘ E) of the basin due to RFWs. Overall, considering the wetland fraction solely attributed to GDWs in CW-TCI15 and CW-WTD (13 % and 29 % of the basin area) and the difference found between the inundation and validation dataset (Fig. 9, first and second row), it becomes clear that the uncertainty of the wetland extent and spatial pattern is rather high over boreal zones.

This large wetland is located in eastern South Sudan, nearly 300 m above mean sea level, and is the largest freshwater wetland in the Nile basin (Sutcliffe et al., 2016). The Sudd swamp extent estimations are highly uncertain in the literature, ranging from 7.2 to 48×103 km2 (Mohamed et al., 2004, and references therein). Over the selected window, the wetlands and surface water distribution is also highly disparate and varies from 1 % to 27 % for different datasets (Fig. 10). Additionally, wetlands in Hu et al. (2017) are rather patchy and show sharp density changes with what seems to be periods of 0.5∘. Because GLWD-3 appears to represent only flooded wetlands (with the same wetland fraction of RFWs and overlapping with one-third of them), and the study of Hu et al. (2017) contains technical issues, GDW-WTD can be considered as the only comprehensive validation dataset over the Sudd swamp.

The CW datasets in Fig. 10 show high wetland densities in the central floodplain, in rather good agreement with GLWD-3, GDW-WTD and regional estimates of saturated soil (compared with visuals in Mohamed et al., 2004, and Mohamed and Savenije, 2014). CW-WTD compares more similarly to validation datasets, closely followed by CW-TCI15 (Fig. 4g), but the main difference between these two CW maps is that the groundwater wetlands in CW-TCI15 are extended southwest into the southern national park (over local flat valley bottoms) but are more concentrated over the main floodplain in the SE–NW direction for CW-WTD. The total wetland fraction is nearly equal in CW-TCI15 and CW-WTD (25 % and 27 % of the selected window area), underlining a primary role of topography and climate in wetland formation. Considering the wetland fraction in the RFW map (mostly consisting of ESA-CCI wetlands) and GDW-WTD, groundwater wetlands appear to be the dominant feature in the Sudd swamp, as is the case for CW-WTD and CW-TCI15. The added value of CW maps with respect to GDW-WTD is not substantial, but they additionally contain the seasonally flooded plains downstream of the White Nile (top right of the selected window in Fig. 10g, 12–14∘ N, 32–34∘ E), which are not completely captured by validation datasets due to the inclusion of RFWs.