Land cover classification is a fundamental and vital activity that is helpful for understanding natural dynamics and the human impacts of land surface processes. Available multiple 1-km global land cover datasets have been compared to identify classification accuracy and uncertainties for vegetation land cover types, but they have not been adequately compared for water-related land cover types. Six 1-km global land cover datasets were comprehensively examined by focusing on three water-related land cover types (snow and ice, wetlands, and open water). The global mean per-pixel agreement measured by the class-specific consistency is high for snow and ice, medium for open water, and low for wetlands. The agreement is low for snow and ice in low latitudes and high for open water and snow and ice in high latitudes. Areas classified as wetlands in a pixel in one dataset are rarely classified as wetlands in the same pixel in the other five datasets. These areas are most often classified as forest, wetland, or shrub. Areas of snow and ice and open water in some regions are not always chronologically consistent among the datasets because nonsatellite data and different algorithms are used to determine the areas. Further research is necessary to reduce uncertainty in the water-related land cover classification and to develop an advanced classification algorithm that can detect water under a vegetation canopy for improvement in wetland classification. Chronological inconsistency between 1-km land cover datasets and satellite observation periods must also be addressed.
Land cover monitoring and classification is a fundamental and vital activity in land surface sciences, including ecology, hydrology, and climatology (e.g., Sanderson et al. 2002; Nakaegawa and Vergara 2010). It helps us to understand the dynamics of land surface processes and to simulate the processes in numerical models. As the amount of available satellite remote sensing data has increased, so has the demand for the detection of temporal changes in land cover types and the use of land cover monitoring in land surface sciences. As a result, several 1-km global land cover datasets have been produced by multiple research institutions, generally with moderate-resolution optical sensor data. These global land cover datasets have been used for both global and regional studies because they cover all continents and islands with sufficient spatial resolution (1 km). However, comparative analysis of different land cover datasets showed per-pixel agreement of only about 0.6 (e.g., Giri et al. 2005; Herold et al. 2008; McCallum et al. 2006). Although these studies made a comprehensive comparison of land cover types, they did not adequately compare water-related land cover types—that is, snow and ice, wetlands, and open water. The studies did find high per-pixel agreements and global total area for snow and ice among the 1-km land cover datasets but they did not examine accuracy and uncertainty for specific geographical features.
The per-pixel agreement for wetland areas is the lowest among the land cover types in the 1-km global land cover datasets. Darras et al. (1998) revealed that the Global Land Cover Characteristics (GLCC) for the International Geosphere–Biosphere Programme (IGBP) (GLCC.I; Loveland et al. 2000) and the atlas-based datasets produced by Matthews and Fung (1987) and Cogley (1994) do not identify large portions of wetland areas listed in the Ramsar dataset (Wetlands International 2002). In addition, the comprehensive 1-km global water-related dataset, Global Lake and Wetland Data level 3 (GLWD; Lehner and Döll 2004), has a much larger total global wetland area than GLCC.I and Moderate Resolution Imaging Spectroradiometer (MODIS; Strahler et al. 1999). There are also large differences in the latitudinal distribution of zonal total wetland areas north of 50°N (Lehner and Döll 2004). The areas of global total open water in several 1-km land cover datasets are in good agreement with one another (e.g., Nakaegawa 2011), and the latitudinal distribution of zonal total open water areas in GLWD coincides well with that of both satellite-based 1-km datasets (GLCC.I and MODIS). However, the zonal total open water area for MODIS is larger at 50°–70°N than that of GLWD and GLCC.I (Lehner and Döll 2004).
Foody (2008) insisted that the per-pixel agreement for all land covers mentioned above is pessimistically biased and that the common target accuracy of 0.85 is too high. Regional atmospheric model simulations have shown that an accuracy of less than 0.80 in a land cover dataset induces distinctive differences in precipitation through the interactions between the atmosphere and land surface (Ge et al. 2007) and that more accurate land cover type classifications induce better reproducibility of simulated surface air temperatures (Sertel et al. 2009), indicating present per-pixel agreements are not always sufficiently accurate for use in land surface sciences. In addition, uncertainty in global land cover datasets can produce uncertainty in estimations of a carbon cycle model (Jung et al. 2006). Nakaegawa (2011) demonstrated that the per-gridbox agreement between datasets can be slightly improved by upscaling the global land cover dataset from 1 km to a horizontal resolution of a numerical model (e.g., 1° and 3°) and that the accuracy for the upscaled global land cover dataset is directly dependent on that of the 1-km global datasets. Therefore, the classification accuracy of the land cover dataset is crucial in land surface sciences. As stated previously, the classification accuracy for water-related land cover types has not been sufficiently explored except for Lehner and Döll (2004), but they did not examine per-pixel agreement, examine snow and ice type, or use multiple satellite-based 1-km global land cover datasets. The present study focuses on three water-related land cover types—snow and ice, wetlands, and open water—and compares the classification results among six 1-km global land cover datasets.
2. One-kilometer global land cover datasets
We used the six 1-km global land cover datasets listed in Table 1. Other datasets are available, such as that of the University of Maryland (UMD; Hansen and Reed 2000), but UMD is not appropriate for this study because snow and ice and desert are classified into a single land cover type (barren), ocean and open water are classified into a single land cover type (open water), and wetlands are not included as a land cover type. Some datasets are produced with the same satellite data and land cover types, whereas others are not (Table 1). These differences produce uncertainty in classification among the datasets. Water-related land cover types can be classified in many ways. For example, GLWD uses 12 land cover types. In this study we used three simple classes: snow and ice, wetlands, and open water. Table 1 also presents the corresponding land cover types in the six datasets for each of the three water-related land cover types. The category snow and ice is not defined in GLWD, and the definition of wetland slightly varies across the datasets (see Table 2).
GLCC for the Simple Biosphere model (SiB) (GLCC.S; Sellers et al. 1986) and GLCC.I (Loveland et al. 2000) were produced with the same satellite data, algorithm, and other factors, except for the land classification legend. Snow and ice were classified with a 12-month maximum value normalized difference vegetation index (NDVI) composite, which is less than the threshold values that depend on continental characteristics. Wetland was classified with an unsupervised cluster analysis that inputs the temporal NDVI composite. Open water was classified with near-infrared Advanced Very High Resolution Radiometer (AVHRR) data supplemented with open water information taken from the Drainage Network (DNNE) in the Digital Chart of the World (DCW; Danko 1992; Defense Mapping Agency 1992).
Snow and ice in MODIS were detected with the normalized difference snow index (NDSI) used in the MODIS snow and ice algorithm (Hall et al. 2001). Wetland was classified with a univariate decision tree—the same algorithm used for vegetation (C4.5; Quinlan 1993). Open water was extracted with land–ocean mask data produced from DNNE in MODIS (Strahler et al. 1999).
Global Land Cover 2000 (GLC2000) was produced by harmonizing the regionally optimized land cover legend for each continent to a less detailed global legend (Bartholomé and Belward 2005). Snow and ice were extracted with NDSI, and wetland was extracted with the same algorithm used for vegetation and with synthetic aperture radar (SAR) for water surfaces under vegetation canopies. Open water was detected with near-infrared data supplemented with DCW.
In Global Land Cover by National Mapping Organizations (GLCNMO; Tateishi et al. 2011), snow and ice were extracted from the MODIS–Terra snow cover global 0.05° Climate Modeling Grid (CMG) dataset (Hall et al. 2001) with threshold processing; for the Eurasian continent, snow and ice were extracted from the same dataset but with k-means cluster analysis. Wetlands were assigned as common wetland areas in both GLCC.I and GLC2000. Mangrove, one of the wetland areas, was extracted by visual interpretation of Landsat Enhanced Thematic Mapper images and from the mangrove world map atlas (World Conservation Monitoring Centre; WCMC 2010). Open water was extracted from seven MODIS bands of three periods in 2007 with unsupervised classification and visual labeling (Tateishi et al. 2011).
GLWD level 3 dataset was used in a raster format (Lehner and Döll 2004). Wetlands were extracted from a wetland map (WCMC 1993), and bog, fen, and mire areas as wetlands were extracted from GLCC for Ecosystem (GLCC.E), which has 100 land cover types. Although there are three different fractional wetland classes in the GLWD corresponding to class numbers 10–12 in Table 3, the other datasets do not have any fractional land cover classes; therefore, the fraction for the three wetland classes is assumed to be unity in this study. Open water was obtained from lakes in five existing datasets: the Mullard Space Science Laboratory Global Lakes Database (Birkett and Mason 1995), the Dataset of Large Reservoirs (Vörösmarty et al. 1997), ArcWorld (ESRI 1992), DCW DNNE, and GLCC.E. No snow and ice were included in the GLWD land cover legend.
GLC2000 does not include Antarctica; we used land cover types in Antarctica from GLCC.E to make GLC2000 global. GLCC.S, GLCC.I, and GLC2000 have the same land cover types in Antarctica. GLWD does not include snow and ice anywhere.
In addition to the six 1-km global land cover datasets, a snow analysis produced from surface observation data and satellite data in the Japanese 25-yr reanalysis (JRA-25; Onogi et al. 2007) was used to evaluate the interannual variability of seasonal snow cover.
Global total areas for each land cover type of each 1-km global land cover dataset and per-pixel agreements of each water-related land cover type were calculated, taking into account the effects of latitude on pixel area. The conventional per-pixel agreement indices, defined as Cohen’s κ or as the ratio of the trace of the contingency matrix to the total number of the pixels (Jensen 1996; Frey and Smith 2007), do not work for the three water-related land cover types because the global total area of a specific water-related land cover type is far smaller than the global total remaining area. A contingency matrix for two classes (Table 4)—for example, a specific water-related land cover type and the remaining types—can be constructed and class-specific consistencies (C) can be calculated as follows:
where a is the area-weighted number of pixels classified to the specific water-related land cover type in both the X and Y datasets, b is the area-weighted number of pixels classified to the specific water-related land cover type in the X dataset but to the remaining land cover types in the Y dataset, c is the area-weighted number of pixels classified to the remaining land cover types in the X dataset but to the specific water-related land cover type in the Y dataset, and d is the area-weighted number of pixels classified to the remaining land cover types in both datasets. This index is called the threat score in meteorology and is widely used for categorical weather forecast evaluation (Wilks 2006). The index correctly provides the degree of the per-pixel agreement even for the case where a ≪ d. In Eq. (1), d does not appear because the influence of the correct classification of the remaining land cover types corresponding to d can be excluded from the agreement evaluation.
The per-pixel agreement number for each of the water-related land cover types was calculated to identify the uncertainty in classification in the six 1-km global land cover datasets. The per-pixel agreement number for a specific land cover type is the number out of the six datasets that has the same land cover type in the same pixel. We used the aggregated legend of the 10 land cover types presented in Table 5 to determine how water-related land cover types in a given 1-km land cover dataset were classified into different land cover types in different datasets. This aggregated legend is based on the previously mentioned studies (e.g., Giri et al. 2005).
Maximum extent maps were developed for each of the three water-related land cover types to determine the largest potential coverage for each type. For each map, each pixel was classified as a particular water-related land cover type if any of the six 1-km global land cover datasets had that water-related land cover type in the pixel. For example, if only one 1-km land cover dataset had open water in a pixel, that pixel was classified as open water on the maximum extent map for open water. Each 1-km land cover dataset has its own systematic detection bias because of the satellite data and algorithm used. Accordingly, these maps can provide maximum extent information about each of the three water-related land cover types with less bias.
a. Global total area
Figure 1 depicts the global total area for each water-related land cover. Although both the classification algorithm and the satellite data for snow and ice in GLCC.S and GLCC.I are different from those of GLC2000 and MODIS, the global total snow and ice areas are very similar and the coefficient of variation is very small. The global total area for snow and ice in GLCNMO is slightly smaller than the others, despite the use of similar algorithms and datasets in MODIS and GLCNMO. This is partly a result of differing definitions of land and ocean, which will be discussed in more detail in section 4d.
GLCC.S and GLCC.I have almost the same global total wetland area, which is reasonable because their production process is the same, except for the land cover legends (Table 1). The global total wetland areas in the five datasets except for GLWD are smaller than the mean value, and those of MODIS and GLCNMO are much smaller. GLWD has a global total wetland area of 10.5 × 106 km2, which is about 10 times the size area of the other datasets. The value remains large, 8.0 × 106 km2, even after we accounted for the fractional ratio of wetland in a pixel. Atlas-based estimates of the global total wetland areas in Matthews and Fung (1987) and Cogley (1994) are about 5.0 × 106 km2, which is intermediate between the values for GLWD and the other five datasets. The GLWD uses other data sources, such as the WCMC wetland map, which may account for the large discrepancy in values; therefore, the coefficient of variation for wetland is quite large for the six datasets. However, it is also relatively large for the five datasets excluding GLWD because optical radiometers have difficulty in detecting seasonal wetlands (Darras et al. 1998).
The global total open water areas are similar for all six datasets because all six datasets use DCW; therefore, the coefficient of variation for open water is small. The slightly smaller area in GLCNMO stems from its use of a unique algorithm for open water.
The global total wetland and open water area in GLWD is the largest of the six datasets because of its much larger global total wetland area. The five global total water-related areas (excluding GLWD, which does not account for snow) are similar; therefore, the coefficient of variation is small for the five datasets. In addition, the total area of GLC2000 is about 7% larger than the mean total area. GLCNMO has the smallest water-related total area in the six datasets.
b. Global per-pixel agreement
Class-specific consistency was calculated to determine whether the same pixel for a pair of datasets has the same water-related land cover type. The class-specific consistency for snow and ice between GLCC.S and GLCC.I is almost unity, and the class-specific consistency for snow and ice among the other pairs of arbitrary datasets is also high—generally from 0.85 to 0.96. These values are reduced to 0.67–0.76 when Antarctica is excluded. Table 6 presents the class-specific consistency for wetlands and open water for all possible pairs of datasets. The consistency for wetlands is high for the GLCC.S and GLCC.I pair because the two datasets only differ in their land cover legend. The other pairs all have very low class-specific consistency (<0.10), and pairs including MODIS have low values (≤0.06).
In the 15-pair ensemble mean, a pixel of wetlands in a 1-km global land cover dataset was classified as a different land cover type in the same pixel of the other five datasets about 70% of the time. Wetlands were most often classified as forest (28%), wetland (27%), shrub (26%), cropland (8%), and grassland (7%; Fig. 2a). These results suggest that wetland classifications must be treated with caution. These classification differences are not unexpected because wetland areas are defined by whether water exists under a vegetation canopy (Table 2), and this type of distinction is very difficult to make using satellite-based classification schemes (Darras et al. 1998).
The class-specific consistency for open water between GLCC.S and GLCC.I is high, and the class-specific consistency between GLWD and GLCC.S or GLCC.I is about 0.65. The class-specific consistency values in most of the remaining pairs of datasets are about 0.5. The pairs including GLC2000 and GLCNMO have lower consistencies because they use a unique algorithm with satellite data. For all possible combinations of pairs, a pixel of open water of a 1-km global land cover dataset is classified identically in the other five datasets in 67% of cases (Fig. 2b). This value is high compared with the general per-pixel agreement of about 0.5 to 0.6 (Giri et al. 2005; Herold et al. 2008; McCallum et al. 2006). Alternate classifications include forest and shrub (both have probabilities of 9%) and wetland (0.01), as shown in Fig. 2b.
c. Latitudinal distribution
Figure 3 shows the latitudinal distribution of the zonal and global mean per-pixel agreement number for each water-related land cover among the six 1-km global land cover datasets. The global mean per-pixel agreement number for snow and ice among the five datasets that classify snow and ice is very high (4.62). The zonal mean per-pixel agreement numbers are high in high latitudes and low in mid- and low latitudes, indicating that the high global mean value is a result of good per-pixel agreement in the ice sheets in Antarctica and Greenland. The global zonal mean per-pixel agreement number decreases to 3.62 if Antarctica is excluded. High zonal mean per-pixel agreement numbers around 50°S latitude reflect the Patagonian Ice Field. The zonal mean per-pixel agreement numbers in mid- and low latitudes reach a maximum of 3, and the mean zonal mean per-pixel agreement number in these latitudes is 1.85. Therefore, uncertainty in the classification of snow and ice areas in these latitudes is high. The zonal mean coefficient of variation of the interannual variability of the seasonal snow cover ratio for June–August in the JRA-25 snow analysis for the 30-yr period from 1980 to 2009 exceeds 0.5 in latitudes lower than 45°N, whereas the zonal mean coefficients are smaller in higher latitudes. This large interannual variability is most likely responsible for the high level of uncertainty because the different datasets do not use the same satellite data over the same time period.
The global mean per-pixel agreement number for wetlands is very low (1.22; Fig. 3a). Relatively high zonal mean per-pixel agreement numbers are found in boreal high latitudes, but they are still less than 2. The large mean total wetland area is due to very large wetland areas only found in GLWD (Fig. 1). We therefore examined whether a wetland pixel in the other five datasets corresponded to a wetland pixel in GLWD and found that an average of 88% of them did correspond with GLWD wetland; conversely, there was only a 5% chance that a wetland pixel in GLWD corresponded with one in the other five datasets. The very large wetland areas found only in GLWD clearly reduce the overall zonal mean per-pixel agreement number. When GLWD is excluded, the global mean per-pixel agreement number is slightly larger (1.42), implying that the other five 1-km global land cover datasets have their own geographical distributions as inferred from Table 6.
The global mean per-pixel agreement number for open water for the six datasets is 3.00. The zonal mean per-pixel agreement numbers are low in high latitudes, which is in agreement with Lehner and Döll (2004), who compared open water among GLCC.E, MODIS, and GLWD. The low zonal mean per-pixel agreement numbers may be attributed to different interpretation of transition areas between wetlands and open water and of subpixel-scale small lakes (Lehner and Döll 2004). Zonal mean per-pixel agreement numbers are higher in several specific latitudes where large open waters are located: the Caspian and Aral Seas and the Great Lakes at about 45°N and Lake Victoria and the Amazon and Congo Rivers near the equator.
d. Geographical distribution
In this subsection, we show the geographical distribution of the per-pixel agreement numbers by focusing on selected areas for each water-related land cover type.
1) Snow and ice
As mentioned above, the ice sheets in Antarctica and Greenland are accurately classified as snow and ice, whereas ice fields and permanent snow in other areas are not. A local minimum of zonal mean per-pixel agreement number for snow and ice exists around 80°S (Fig. 3a) because the Ross and Filchner–Ronne Ice Shelves are classified as ocean in GLCNMO, whereas they are snow and ice in the four other datasets. These kinds of differences are also found in the coastal region on the North Sea side of Greenland and around the Svalbard Islands.
The largest snow and ice area in the low latitudes exists near the Tibetan Plateau corresponding to local maxima around 30°N (Fig. 3b). The per-pixel agreement number of snow and ice is high in the Himalayan and Tanggula Mountains, which are more than 5000 m above sea level (Fig. 4). GLC2000 has the largest snow and ice area in the Tibetan Plateau region of the five datasets. GLCC.S, GLCC.I, and GLCNMO all have similar smaller values, and MODIS has the smallest snow and ice area. MODIS and GLCNMO should show similar features because both use the MODIS snow cover datasets (Hall et al. 2001). We closely compared the MODIS snow products used in MODIS and GLCNMO in July, which is the month with the least snow cover, for 2001 and 2003 and found that snow and ice areas were larger in 2001 than in 2003. This fact contradicts the difference in snow and ice area between MODIS and GLCNMO in Fig. 4. The difference most likely stems from the algorithm used for snow and ice detection in GLCNMO, which employs an unsupervised k-means cluster analysis only for the Eurasian continent.
The global mean per-pixel agreement number for wetland is 1.22, suggesting that the geographical distribution of wetlands varies between the datasets. Higher per-pixel agreement numbers are confined to certain regions. In West Siberia, a per-pixel agreement number of 4 occurs along the main stream of the Ob River, Russia, and it ranges from 2 to 4 in regions between the Ob and Yenisei Rivers (Fig. 5a). A per-pixel agreement number of 2 is found in the Russian Far East and in the region surrounding Hudson Bay, Canada. A per-pixel agreement number of 4 or greater is found in the Everglades in Florida (Fig. 5b) and in the Sudd Swamp, Sudan (Fig. 5c). A per-pixel agreement number of 1 is also widely distributed in these regions (Fig. 5).
Frey and Smith (2007) compared ground-truth land cover data in West Siberia, which is very similar to the region depicted in Fig. 5a, with those of four 1-km global land cover datasets, including GLCC.I, MODIS, and GLWD. All of the datasets were found to significantly underestimate wetland areas as well as open water areas. The agreements with the ground-truth data were 0.02 for MODIS, 0.23 for GLCC.I, and 0.45 for GLWD. These low ground-truth agreement numbers correspond well to low per-pixel agreement numbers that are widely distributed in Fig. 5a, although the GLWD dataset does show a relatively high agreement with ground-truth data.
3) Open water
The area of Lake Chad declined by 95% since the early 1960s as a result of both decreased rainfall and increased irrigation (UNDP et al. 2000). In the area surrounding Lake Chad, the per-pixel agreement number for open water is 2–3, but in the central southeastern part of Lake Chad where open water remained in the 2000s, the per-pixel agreement number is 6 (Fig. 6a). Most of the Caspian Sea has a per-pixel agreement number of 6, but the northern coastal zones in Kazakhstan are lower (2–4). Garabogzaköl Bay, Turkmenistan, has a per-pixel agreement number of 6 in its central area and 4 in its surrounding areas (Fig. 6b). In 1980, the strait connecting the bay to the Caspian Sea was blocked and the bay decreased in size with salt deposition, leading to confined open water in the central area. In the 1990s, the strait was reopened and Garabogazköl Bay has been restored to its original state (Kosarev et al. 2009), showing that anthropogenic operations not only reduce but can also increase the amount of open water areas.
Lake Chad shows inconsistent chronological changes in the datasets. A satellite-based observation shows a monotonic increase in the open water area of Lake Chad since the late 1980s (Leblanc et al. 2011); however, the lake reaches a maximum area in 1980s (GLWD) and 2000/01 (MODIS) and a minimum area in 1992/93 (GLCC.S and GLCC.I), but a minimum also occurs in 2003 in GLCNMO. In addition, the Aral Sea also shows inconsistent chronological area changes (Fig. 7). The open water area reaches a maximum in the 1980s (GLWD), 1992/93 (GLCC.S and GLCC.I), and 2000/01 (MODIS), but it is at a minimum in 2000 (GLC2000) and 2003 (GLCNMO), although a ground-based observation shows a monotonic decrease in the open water area of the Aral Sea (Micklin 2007). On the other hand, Garabogzaköl Bay shows chronological changes (Fig. 7) incidentally consistent with a ground-based observation (Kosarev et al. 2009). These findings indicate that open water classifications in the six 1-km global land cover datasets do not always reflect the chronological dates of the satellite data used.
e. Maximum extent maps
The global total snow and ice area in the maximum extent map is 8% larger than the mean global total snow and ice area of the six 1-km global land cover datasets. When Antarctica is excluded, the maximum extent map is 38% larger than the mean global snow and ice map, which is consistent with the mean total and the global mean per-pixel agreement for snow and ice, excluding Antarctica.
The global total wetland area in the maximum extent map is about five times the size of the mean global total wetland area of the six datasets, or about 16% larger than the global total wetland area in GLWD. The very large wetland area in GLWD and low agreement among the wetland classifications are responsible for these differences. The global total wetland area (12.2 × 106 km2) in the maximum extent map developed here is about 4% larger than the global total wetland area (11.7 × 106 km2) in the maximum extent map developed by Lehner and Döll (2004) from Matthews and Fung (1987), Cogley (1994), and Stillwell-Soller et al. (1995) using GLCC.E and MODIS. This result reconfirms that uncertainty in wetland classification in the 1-km global land cover datasets is still large, and it raises the question about the actual size of the global total wetland area.
The global total open water area (7.1 × 106 km2) in the maximum extent map is about twice the size of the mean global total open water area of the six datasets. Previous estimates of global total open water area range from 3.2 × 106 km2 (Lehner and Döll 2004) to 4.6 × 106 km2 (Downing et al. 2006), which is about equivalent to the mean global total open water area in Fig. 1. The much larger area of the maximum extent map developed here may be the result of misclassification. We developed another maximum extent map for open water by classifying each pixel as open water if any two of the six datasets had open water in the pixel, and the global total open water area was 4.6 × 106 km2. The treatment of pixel fractions or the amount of open water in a pixel in the six datasets, which were not considered in this analysis, may also have impacted the results.
GLWD is the most widely used dataset in land surface sciences because it is comprehensive and extensive. We examined how GLWD classes of the water-related land cover types were classified in the other datasets.
Wetland areas in GLWD defined in Table 1 are classified as forest (32%), shrub (26%), and cropland (16%), where each value is the mean for the five possible pairs. The average ratio of GLWD wetland classified as wetland in the other five datasets to the total GLWD wetland area is 5%, which is much higher than the probability of an arbitrary pixel being classified as wetland in one of the five other datasets (0.7%). This can be explained by the much larger global total wetland area in GLWD as compared to the other five datasets (see Fig. 1). Since forest, shrub, and cropland with a water surface under the vegetation canopy can be classified as wetland, these classification differences in wetland area in GLWD are reasonable. In other words, the three land cover types in the other five datasets may include wetland areas, especially in areas classified as wetlands in GLWD.
Open water in GLWD is primarily classified in the five other datasets as open water (66%), shrub (12%), and forest (11%). Latitudinal distributions of zonal total open water areas in GLWD are generally similar among the five datasets; however, discrepancies in open water areas between 50° and 70°N latitude were noted, as was also pointed out by Lehner and Döll (2004). The results may be affected by a very large number of small lakes in glaciated or permafrosted terrain or peatland (Smith et al. 2007) that are irresolvable in GLWD as well as by the interpretation of transition areas between open water and other land cover types.
Figure 8 shows scatterplots between open water lake areas and the agreement ratio of total open water area in GLWD and each dataset. The shoreline pixels of open water lake areas generally have lower agreement ratios than the inner pixels. The ratio of shoreline pixels to inner pixels is smaller in large open water areas than it is in small ones. Therefore, large lake open water areas tend to have higher agreement ratios than small ones, which is clearly seen for large open water lakes (>1000 km2) in Fig. 8. However, the agreement ratios vary from 0 to 1 for relatively small lake open water areas (<100 km2). For full-resolution products such as MODIS, the location of small open water areas may move around or change in size and shape with subsequent passes as a function of viewing angle and conditions, such as cloud and aerosol interventions. We examined eight of the largest 32 lakes (Table 7), all of which are larger than 5000 km2 and have an agreement ratio of lower than 0.8. GLCNMO has low agreement ratios for all eight lakes, which may be a result of temporal variations in lake open water area at each lake. For example, large-scale irrigation decreased the open water area of the Aral Sea and Lake Chad (e.g., Revenga et al. 1998; UNDP et al. 2000). Both artificial and natural variations are responsible for the low agreement ratio in the human-controlled lakes (Great Salt, United States; Itaparica, Brazil; Volta, Ghana; and Smallwood Lakes, Canada), and differences in the complicated shoreline and many small islands characterizing a glacial lake are responsible for the low agreement ratio in Reindeer and Nettiling Lakes, Canada.
The definitions used for wetland in each 1-km global land cover dataset (Table 1) are widely used but they represent only some possible definitions. To quantify this uncertainty, we conducted a similar class-specific consistency analysis but added potential wetland cover types to the wetland definition for GLC2000 and GLWD, as indicated in Table 1 with italicized letters. In comparison with the results presented in Table 6, the class-specific consistency for wetlands slightly decreased by about 0.005 in pairs including GLWD, and the class-specific consistency for wetland in pairs including GLC2000 increased by about 0.01. The GLWD and GLC2000 pair had a relatively larger increase in class-specific consistency (0.025), which indicates that the results obtained in this study are probably robust for different definitions of wetland.
A close examination shows that a pixel of GLWD’s swamp forest and flooded forest was classified as forest 62.3% of the time in the other five datasets and as shrub another 20.6% of the time in the five except for GLC2000, where it was classified as shrub only 7.3% of the time. Although Lehner and Döll (2004) classified forest as wetland when they calculated the global and zonal total wetland areas, forest in GLWD is rarely classified as wetland (1.4%) in the five other datasets. Therefore, the category of forest in the five other datasets must include some flooded forest and shrub areas because of differing interpretations of wetlands and the difficulty in using satellite-based classification for wetlands, as previously mentioned.
These results demonstrate that the five satellite-based 1-km global land cover datasets excluding GLWD show similar features in quality when compared to GLWD and generate a cluster. The GLWD was primarily developed by compiling multiple datasets in addition to the satellite-based observations as mentioned above, and it should be considered as a comprehensive dataset. The GLWD has much more accurate agreement with ground-truth data than satellite-based land cover datasets (Frey and Smith 2007). Therefore, the GLWD may be considered to be the best 1-km global water-related land cover dataset currently available and as a benchmark. Even so, GLWD does not always work well for the evaluation of other 1-km global land cover datasets. GLWD shows relatively good class-specific consistency for wetlands and open water with GLCC.S and GLCC.I (Table 6) because wetland and open water areas are not always independent in these datasets. Therefore, a global ground-truth dataset for water-related land cover types is fundamentally essential to address the accuracy of each dataset as Frey and Smith (2007) demonstrated.
The six 1-km global land cover datasets in this study do not always apply the state-of-the-art algorithms to extract water-related land cover areas except for snow and ice areas. In addition, atlas-based digital data developed in the early 1980s have been used. As mentioned above, open water areas temporally change in size on a decadal time scale (e.g., Lake Chad). Some users may be better served with fully satellite-based open water classification so that a 1-km global land cover dataset has temporal features consistent with the satellite observation period used for the development; other users may prefer climatological or temporal mean open water classification, depending on the user’s purpose. The normalized difference water index computed from near-infrared (Knight et al. 2009), visible green radiometer data (McFeeters 1996), and other methods could prove useful as detection techniques for chronological changes.
As mentioned earlier, optical radiometer data have difficulty detecting wetland areas if the water surface is under a vegetation canopy (Darras et al. 1998). Some new methods for wetland classification have been developed, including methods using optical radiometer data and topographical variables (Landmann et al. 2010) and using both optical radiometer and SAR data (Henderson et al. 2002), both of which might provide a successful solution (Tateishi et al. 2011). In addition, SAR data are being applied to detect wetland areas in forested wetlands (e.g., Lang et al. 2008), Andean wetland forages (Moreau and Letoan 2003), and other areas (Henderson and Lewis 2008). These new algorithms may work well at an experimental site or a regional scale, but they have not yet been examined on a continental scale. Nevertheless, the use of new algorithms is expected to provide less uncertainty in estimates of water-related land cover areas in 1-km global land cover datasets.
We comprehensively examined six 1-km global land cover datasets (GLCC.S, GLCC.I, GLC2000, MODIS, GLCNMO, and GLWD) by focusing on three water-related land cover types (snow and ice, wetlands, and open water). The global mean per-pixel agreement numbers are high for snow and ice, medium for open water, and low for wetlands; however, the latitudinal distributions of zonal mean per-pixel agreement numbers show distinctive features. The uncertainty is high for snow and ice in low latitudes and low for open water and snow and ice in high latitudes or regions. Areas classified as wetlands in a pixel in one dataset are rarely classified as wetlands in the same pixel in the other five datasets. These areas are most often classified as forest, wetland, or shrub. Areas of snow and ice and open water in some regions are not always chronologically consistent among the datasets because nonsatellite data and different algorithms are used to determine the areas.
These datasets play an important role in estimations and predictions generated by both statistical and dynamical models in land surface science, and the class-specific consistencies for the three water-related land cover types are not high enough. When one of the existing 1-km global land cover datasets is used, it is crucial to understand the features of the specific dataset. In this study, we focused on the per-pixel agreement number on a regional to global scale. The differences between the datasets may be enhanced at a smaller scale and should be examined.
Although we quantified the agreement (uncertainty) in the water-related land cover types among the six 1-km global land cover datasets, we have not yet explored which dataset should be used for a given purpose or how to reduce the quantified uncertainty. Intensive evaluation of the land cover datasets against ground-truth data may be a straightforward methodology, but none of the 1-km global land cover datasets is likely to show high agreement, as was demonstrated for West Siberia by Frey and Smith (2007). The choice of a dataset depends on purpose and region. The agreement number and the maximum extent map used in this study may provide fundamental information on making such choices.
These datasets are produced using various combinations of satellite data, algorithms, land cover types, and other factors, all of which contribute to the uncertainty in classification. In the meteorological community, historical global daily atmospheric analyses have been performed using the latest data assimilation system—a process known as reanalysis (e.g., Kanamitsu et al. 2002; Onogi et al. 2007). The reanalysis of different satellite datasets with a consistent classification algorithm may provide a consistent time series of temporal 1-km global land cover datasets and reveal detectable interannual variability in land cover types.
I am grateful to all the institutions and developers who kindly provided the six 1-km global land cover datasets. This study was funded, in part, by KAKENHI (Grant 19106008) of Ministry of Education, Culture, Sports, Science and Technology (MEXT). Additional support was provided by the KAKUSHIN Program of MEXT. This study was also conducted as part of an integrated Meteorological Research Institute project on developing a seasonal prediction system using a global atmosphere–ocean coupled general circulation model.
This article is included in the Hydrology in Earth System Science and Society (HESSS) special collection.