The Surface Water Microwave Product Series v3.2 (SWAMPS) is a dataset of long-term, daily time series of the inundated area fraction derived from microwave remote sensing. The SWAMPS dataset provides estimates of terrestrial surface water dynamics, including for wetlands, rivers, lakes, ponds, reservoirs, rice paddies, and episodically inundated areas. SWAMPS provides estimates of the global inundated area fraction (fw) developed under the NASA Making Earth System Data Records for Use in Research Environments (MEaSUREs) program and is updated monthly. SWAMPS fw estimates are derived from a combination of passive microwave brightness temperature and active microwave radar backscatter from a variety of satellite sensors supplemented with a priori knowledge of land cover based on a static MODIS land cover product (Schroeder et al., 2015). The SWAMPS product includes daily gridded DMSP Special Sensor Microwave/Imager and Special Sensor Microwave Imager/Sounder (SSM/I-SSMIS) Pathfinder brightness temperature observations and active microwave backscatter from the NASA SeaWinds QuikSCAT Level-1B Sigma0 product and Advanced Scatterometer Level-1B (ASCAT) product, with ancillary snow water equivalent, land cover map, and normalized difference vegetation index (NDVI) from the Advanced Very High Resolution Radiometer (AVHRR) and MODIS for delineating snow cover and arid and semiarid areas. SWAMPS v3.2 is an update of v2.0 and includes a new cloud and snow mask, a quality control flag, a new land and ocean mask, freeze–thaw detection, and improved sensor intercalibration. For the purpose of this study, the SWAMPS v3.2 dataset, covering the years 2000 to 2018, was merged into a single monthly mean time series using samples flagged as “Valid Observations”. The sample sizes for the monthly mean range from 14–22 d globally with more measurements during the summer months and fewer measurements during winter months. For SWAMPS v3.2, the coastal zone was first filtered out using a Landsat-based 90 m mask of permanent ocean waters defined by the G3WBM global water body map dataset (Yamazaki et al., 2015) and then filtered using the MODIS MOD44W product to keep it consistent with the processing of the GSW dataset. The SWAMPS v3.2 data were remapped to WGS84 using bilinear interpolation at a 0.25∘ resolution with values aggregated from daily to monthly means.

The Global Surface Water Explorer (GSW) product is derived from 16 d Landsat thematic mapper imagery at a 30 m spatial resolution and identifies the presence or absence of water bodies over the period 1984–2016 (Pekel et al., 2016). We used this dataset to represent permanent water bodies which we define as those covered by open water for more than 50 % of the months during this time period. We used this as a permanent-water-body mask to avoid including temporary water bodies that are considered wetlands in our working definition. This distribution of long-term maximum permanent water was re-gridded to a 0.25∘ fractional area per grid cell and used for removing inland-water areas from SWAMPS v3.2. Because the coastal regions were masked out in the processing of SWAMPS, we used the MODIS product MOD44WC6 (Carroll et al., 2009) to generate an ocean mask in the processing of GSW to avoid over-deducting. The coastal region is defined as land areas along the coastline within 4 pixels (∼ 1 km) and was then intersected with the ocean-labeled pixels from MOD44WA1 to separate the ocean from inland water. The resulting ocean mask was then applied to remove coastal wetlands in GSW. The static long-term open-water area excluding coastal regions in GSW is 4.5 Mkm2, compared with the river and stream surface areas of 0.8 Mkm2 (Allen and Pavelsky, 2018).

We used static wetland maps to fill gaps left by wetland types that are under-represented or missed by the SWAMPS dataset. However, most static maps do not have global coverage or tend to have lower accuracy compared to the regional products, leaving us to take a separate merging approach for each of the three latitudinal bands.

Many arctic wetlands, including peatlands, do not have surface inundation and thus are not captured by SWAMPS 3.2, but they still emit methane. We use the Northern Circumpolar Soil Carbon Database (NCSCD) to map permafrost and non-permafrost peatlands based on the Histel and Histosol soil orders (Hugelius et al., 2013). The NCSCD dataset is a digital polygon-based database compiled from harmonized regional soil classification maps in which data on soil order coverage have been linked to pedon data. In this study, the NCSCD wetland distribution is used as supplementary data for the latitudinal bands from 60–90∘ N. In this study we use a gridded version with a spatial resolution at 0.25∘. Permafrost and non-permafrost peatlands (Histels and Histosols, defined as > 40 cm surface peat) are mapped in the NCSCD from harmonized regional and national soil maps (Hugelius et al., 2013). However, these maps do not include occurrences of mineral soil tundra wetlands (with organic soil horizons of 0 to 40 cm) or smaller wetland complexes (Hugelius et al., 2020). To better include these types of wetlands, the NCSCD soil maps were combined with CircumArctic Wetlands based on Advanced Aperture Radar (CAWASAR) by Widhalm et al. (2015). The SAR data identify both organic and mineral wetland soils. They are based on Envisat advanced SAR data acquired in Global Monitoring mode (medium resolution) under frozen soil conditions; this represents surface roughness which can serve as a proxy for wetness levels in tundra. The wettest class was included as wetland. It corresponds to soils with > 25 kg C m2 in the top 100 cm (Bartsch et al., 2016). To avoid double counting of organic wetlands (peatlands) the datasets were overlaid so that any overlap between the datasets was removed, maintaining the NCSCD in the output data. The merged static map covers 2.3 Mkm2 for the high latitudes (> 60∘ N), including peatlands and mineral wetlands in the tundra biomes.

The distribution of tropical wetlands, including annually or seasonally waterlogged area and tropical peatlands, are derived from an expert-system mapping product (Gumbricht et al., 2017). We used the Center for International Forestry Research (CIFOR) wetland distribution for adjusting wetlands in the latitudinal bands from 60∘ S–40∘ N. This static map was generated by combining satellite images and topographic convergence indices (TCIs) by CIFOR. The TCI indices are calculated based on the Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) at a 250 m resolution with precipitation climatology from the WorldClim global dataset (Hijmans et al., 2005). A simplified hydrological model was used to estimate the local vertical water balance, runoff, and flood volumes. The topographic and hydrologic data are merged with MODIS (MCD43A4) images used for estimating the duration of wet and inundated soil conditions. The estimated areas of tropical peatlands and wetlands are ∼ 1.7 and ∼ 4.7 Mkm2 respectively. The estimated extent of CIFOR for the Cuvette Centrale tropical African peatland in the Congo Basin is 125 400 km2, which is in agreement with 145 500 km2 of a recent independent field investigation (Dargie et al., 2017).

The Global Lakes and Wetland Database (GLWD) (Lehner and Döll, 2004) is a global database of lakes, reservoirs, and wetlands based on the aggregation of aerial surveys, surveyor maps, and inventories at global and regional scales. While GLWD was generated from data sources now decades old, for some regions, it still represents the most complete wetland database available today. In this study, the GLWD wetland distribution is used to cover the temperate wetland only in the latitudinal band 40–60∘ N, outside the range of NCSCD and CIFOR. We used the Level-3 product, a global raster map that contains 12 classes of water bodies and wetlands at the 30 s resolution. We excluded the classes representing lakes, rivers, and reservoirs (1–3) and estimated the area of fractional wetland classes (9–12) as the midpoint from the range of each class. We then calculated the total fraction of wetland from all classes in 0.25∘ pixels. The estimated total wetland extent in GLWD is 8.7 Mkm2 for the globe and 2.7 Mkm2 for the 40–60∘ N bands.

The distribution of rice paddies is derived from the global dataset of monthly irrigated and rainfed crop areas for the year ca. 2000 (MIRCA2000) (Portmann et al., 2010). The datasets used to develop MIRCA2000 are based on compiling census-based land use datasets downscaled to a grid-cell level and thus are generally consistent with subnational statistics collected by national institutions and by the FAO (Food and Agriculture Organization of the United Nations). For this study, we extracted the annual maximum area of irrigated rice paddies from its original resolution at 5 arcmin and remapped to a 0.25∘ resolutions. We did not consider rainfed rice as we could not reliably separate lowland from upland cropping practices, with only the latter seasonally contributing to surface inundation. The estimated rice paddies in MIRCA2000 (irrigated 0.64 Mkm2, rainfed 1.13 Mkm2) are largely consistent with census-based national and subnational statistics from the FAO (1.54 Mkm2 for total area at ca. 2000) and slightly lower than a remote sensing estimate for irrigated rice paddies (0.66 Mkm2) (Salmon et al., 2015). We thus apply the monthly rice cover from 2000 across the entire 2000–2018 time series. This assumption ignoring year-on-year change in rice paddy area is reasonable given that its area increased by < 1.6 % over 2000–2017 according to IRRI world rice statistics (http://ricestat.irri.org:8080/wrsv3/entrypoint.htm, last access: 6 May 2021).

Our estimated total annual maximum area of global vegetated wetlands (excluding Greenland and Antarctica) is ∼ 13.0 Mkm2 (Fig. 3a). This estimate consists of 3.5 Mkm2 of mean annual minimum wetlands (Fig. 3b), 4.0 Mkm2 of seasonally inundated wetlands (MAmax minus MAmin) (Fig. 3c), and 5.5 Mkm2 of intermittently inundated wetlands (MALt minus MAmax) (Fig. 3d). Our estimated global total wetland area is slightly higher than that of GIEMS2 (Table 2) but is lower than that of a high-resolution version of the GIEMS initial version GIEMS-D15, which reports a long-term maximum of 17.3 Mkm2 (Fluet-Chouinard et al., 2015). Considering that WAD2M conservatively excludes rice paddies (0.59 Mkm2), rivers, streams, and lakes and ponds (2.52 Mkm2) while GIEMS-D15 includes these water bodies, one possible conclusion is that WAD2M applies the upward mergers of CIFOR and NCSCD, which has lower wetland estimates than those of GLWD, causing a lower long-term maximum than in GIEMS-D15. In addition, our estimated total area for intermittently inundated wetlands is close to the 5.2 Mkm2 reported for similar wetlands by GIEMS-D15, suggesting a good agreement for temporarily inundated areas between two independently developed products. Other recent studies (Hu et al., 2017; Tootchi et al., 2019), however, have proposed much higher global wetland areas of 27–29 Mkm2, which are likely overestimations due to their approaches based on topographic wetness indices that do not take into account the location of surface water tables. This leads to an overestimation of the inundated area with shallow groundwater tables and large inundated areas in, e.g., central Asia and South America that are not matched by other wetland maps.

Permanently inundated wetlands are located in well-documented wetland hotspots, including the Hudson Bay lowlands and West Siberian lowlands, where large extents of peat bogs and fens are not represented by SWAMPS v3.2. In tropical regions, key peatland areas along the Amazonian floodplain, in the Cuvette Centrale of the Congo (Dargie et al., 2017), and in the tropical peatlands in Indonesian Papua are all captured by WAD2M. The subtropical and boreal regions are the main contributors to the seasonally inundated wetlands. For the subtropical regions, the seasonal wetlands are largely located in Southeast Asia and the Sahel, where the variation in wetlands is mainly driven by the annual cycle of precipitation in monsoon regions. For the high latitudes, seasonal wetlands are primarily in the transition region from temperate to boreal across both North America and Eurasia. The high seasonality of inundation in these regions was also captured by the surface inundation retrievals of passive microwave observations from the Soil Moisture Active Passive (SMAP) mission (Du et al., 2018).

The latitudinal distribution of wetland area (Fig. 4) suggest that the Northern Hemisphere middle-to-high latitudes (> 45∘ N) have the highest coverage of wetland area with 45 ± 5 % of the total area of wetlands, followed by the equatorial region (10∘ S–10∘ N). A large portion of the intermittent wetlands are found in the northern middle–high latitudes, in regions that also have large areas of seasonal wetlands. The overall latitudinal pattern in WAD2M is similar to that of other estimates except for the Tootchi2019 one, which has the highest wetland area along the latitude gradient. The exception is over the mid-latitudes (20–40∘ N) where the wetland areas in GLWD are more extensive than that in WAD2M. The wetland areas in the arctic (> 60∘ N) in WAD2M have a lower wetland extent than in GLWD and NCSCD but higher than in GIEMS2. WAD2M shows a slightly higher wetland extent in the latitudinal band of 10–15∘ N compared to the other products, which we attribute to the higher intermittent wetlands in Southeast Asia detected by SWAMPS (Fig. 3d). The latitudinal gradient of the wetland area in WAD2M is similar to that of the previous version, SWAMPS-GLWD (Poulter et al., 2017), but with a reduced wetland area in the arctic (> 50∘ N) and at mid-latitudes (15–45∘ N), a consequence of the masking out the inland-water areas from GSW. Surface inundation products (GIEMS2 and SWAMPS) have limited observations in the high latitudes due to underestimates of wetland extent for unsaturated peatlands (Bohn et al., 2015) and due to the presence of snow and ice, and they are not reliable points of comparison in high latitudes.

We validated WAD2M against available independent fine-resolution datasets for the two methane-emitting hotspots, Amazon Basin lowlands (defined as the portion of the Amazon watershed below 500 m a.s.l.) and West Siberian lowlands. These two regions represent different wetland subtypes, vegetation compositions, and local hydrology, making them complementary for our validation.

The distribution of wetland area from WAD2M shows a similar spatial pattern for the Amazon Basin lowlands compared to the map based on JERS-1 SAR (Hess et al., 2015), which was used by Pangala et al. (2017) to estimate methane emissions. The distribution of WAD2M has a good similarity (κ= 0.54) to that of the independent, L-band synthetic aperture radar (SAR) map and is slightly lower than that of GIEMS2 (κ= 0.56; Fig. 6a) but higher than all other global products compared (range 0.1–0.2). WAD2M adequately captures the permanently inundated wetlands along with the Amazon Basin river channel network as well as temporarily flooded wetlands during the wet season (Fig. 5a). However, considerable spatial disagreements of the wetland location and extent were found among available datasets when compared with Hess et al. (2015). The disagreements captured in Fig. 6a are primarily related to the seasonally inundated Pantanal floodplains that have relatively flat terrains with dominant herbaceous/shrubland land cover (height < 5 m) and the Pastaza–Marañón wetland basin of the western Amazonia that have large areas of permanently inundated forested swamps with dense-canopy. The WAD2M-estimated wetland fraction exhibits a reasonably good agreement with independent SAR-based wetland products from Hess et al. (2015) for the Pantanal floodplains and the Ucayali–Marañón wetland basin. The WAD2M estimates for Pastaza–Marañón swamps are close to those of CIFOR and Hess et al. (2015) and are considerably lower than those of Tootchi2019. Over the Pantanal floodplains, WAD2M shows moderate-density wetlands, while Tootchi2019 suggest a widespread high-density area over the same extent. The CIFOR estimate is likely an underestimation given the limitations of its topographic-hydrology approach at estimating inundation over flat terrain like the Pantanal.

The comparison of multiple wetland mapping products for the West Siberian lowlands (Fig. 5b) shows that WAD2M permanent wetlands capture the general spatial distribution represented by most other products. WAD2M shows a good agreement with the independent dataset from Terentieva et al. (2016) that combines field survey and satellite images, in particular for inundation peatlands (e.g., fens and bogs) from 55–65∘ N in the taiga forests. The kappa coefficient of the WAD2M wetland map with Terentieva et al. (2016) for the West Siberian lowlands is 0.70 (Fig. 6b), higher than the value of GIEMS2, GLWD, SWAMPS-GLWD, and Tootchi et al. (2019) (kappa coefficient of 0.18, 0.57, 0.54, and 0.43 respectively). Wetlands in high latitudes above 65∘ N are more intermittent, caused by thawing of permafrost, and thus related to interannual climate variations.

Distinctive seasonal cycles in WAD2M can be observed across varying latitudinal bands (Fig. 7). The tropics (30∘ S–30∘ N) contribute 68 % of the global annual variation in wetland area, owing to the large wetting and drying cycles of tropical wetlands. Despite their large area of intermittent wetlands, the mid-latitudes have a less pronounced seasonal cycle with an average annual minimum of 0.9 Mkm2 and average annual maximum of 1.1 Mkm2 compared to the tropics and high latitudes. High-latitude wetlands again have a strong seasonal cycle with an average annual minimum of 0.24 Mkm2 and average annual maximum of 1.5 Mkm2. The seasonal cycle of WAD2M in mid-latitudes is small compared to that of GIEMS2 (Prigent et al., 2020), which is possibly due to different algorithms applied in SWAMPS and GIEMS2, especially in the way the vegetation contribution is accounted for. The seasonal cycle in the high latitudes is highest among the three regions, which is consistent with GIEMS2 and is mainly due to a significant annual freeze–thaw cycle.

Given that there is a surprising scarcity of independent wetland products with which to evaluate the seasonal patterns in middle and high latitudes, we only focus on the comparison of the seasonal cycle for the Amazon Basin, the largest regional contributor to the seasonal cycle of wetland extent. For the Amazon Basin lowlands, the estimates of wetland area exhibit a low seasonal pattern in WAD2M compared to the SAR-based high-resolution estimates from Hess et al. (2015). The flooded/inundated wetlands in WAD2M vary from 0.340 Mkm2 in the low-water season (October–November) to 0.401 Mkm2 in the high-water season (May–June). The satellite products based on passive microwave bands such as SWAMPS underestimate the seasonality and total wetland areas due to their limited ability to detect inundation outside of large wetlands and river floodplains. This indicates the need to improve the retrieval approach to account for the vegetation contribution in the processing of active and passive microwave signals.

The interannual variations in WAD2M suggest the effect of climate variations on global wetland extent across varying latitudinal bands (Fig. 7). Monthly anomalies, calculated by subtracting the 19-year mean monthly value from the monthly time series, reveal the changes in global wetlands in response to global climate variability such as El Niño–Southern Oscillation (ENSO) (Fig. 7b). For instance, a strong positive response in wetland areal anomalies was captured by WAD2M during the strong 2010–2011 La Niña event that temporarily increases the terrestrial water storage via affecting precipitation patterns globally (Boening et al., 2012). The signal for the recovery captured by WAD2M, i.e., the decline during the late stage of La Niña, is consistent with the estimated terrestrial water storage from GRACE and the ESA CCI Soil Moisture product (Fig. 8). The linear fit of the pan-tropical wetland anomalies for WAD2M over 2000–2018 shows no significant change (p > 0.1) in the wetland extent for the entire period. There are no trends of wetland extent for mid-latitudes and high latitudes (p > 0.1) as was also found with Landsat imagery (Wulder et al., 2018).

In general, variation in surface water in the tropics is primarily driven by precipitation, and the agreement in the patterns of the surface water extent and precipitation gives confidence in the interannual variability in wetland area estimation. At high latitudes, surface water runoff from snowmelt, not from direct precipitation, contributes towards the lower correlation between inundation extent and precipitation. A strong decline in wetland area during the early stage of El Niño in 2015–2016 was captured by all of the products. The GIEMS2 dataset shows similar patterns to those of WAD2M in two aspects: (1) tropical wetlands contribute to over 50 % of the global total wetland areas and the decadal change in wetland extent is mostly confined to the tropics and (2) the temporal variations in WAD2M are consistent with those of GIEMS2, where a sharp decrease in the tropics was found for 2010–2012, followed by an upward trend from 2013–2014. Note that despite the decline in 2000–2006, the WAD2M estimate is followed by a slight recovery during 2007–2014.

For the interannual variations at a river basin scale (Lehner and Grill, 2013), there is a generally good agreement in the interannual variation in wetland extent between WAD2M and four surface water products that are based on different methodologies (Fig. 8). All the products, including WAD2M, suggest a declining trend in wetland extent in the Amazon Basin since 2012, with strong negative anomalies during 2015–2016 when the strong El Niño event occurred. The temporary increase in wetland extent in WAD2M responding to the strong La Nina event of 2010–2011 is supported by the satellite-based observation of water storage and surface soil moisture, where good agreements of a strong positive phase in terrestrial water were found for the Orinoco Basin and Paraná Basin in South America. The temporal variations in wetland extent for the Ganges, Nile, Yangtze, and Mekong also suggest a good agreement between WAD2M and other products, except for the period since 2015 where GRACE observations suggest a strong decline in the Ganges Basin and Yangtze Basin while WAD2M observations remain constant or slightly decline. This discrepancy could be due to changes in irrigated rice extent, suggesting that the wetland extent in these regions so far is less influenced by groundwater depletion caused by human activity (Rodell et al., 2018). This can be supported by the wetland extent estimate from the TOPMODEL-based prognostic hydrological approach (Zhang et al., 2018), which explicitly excludes influence of human activity and attributes the change to the enhanced tropical precipitation since 2014.