The distribution of SOC mass by continents (Table 3) follows the pattern of terrestrial ecological zones. A large areal fraction of deserts obviously reduces the continental mean SOC stock, whereas a large fraction of frozen organic soil increases the continental mean SOC stock (Fig. 1).

Large SOC deposits exist in the frozen soils of the permafrost region and are vulnerable to the effects of global warming. The mass of these deposits, however, is not well known because the area and the stocks of the permafrost region are uncertain. The uncertainty in the area is characterized by the variation in the delineation and thus extent of the permafrost region among different maps and databases, which is due also to different definitions of “permafrost” and associated concepts.

One permafrost delineation is directly defined by the HWSD. The HWSD lists for each soil unit the presence of permafrost within the top 200 cm (a so-called “gelic phase”). SOC mass in the top 1 m of soils with a gelic phase is 164 Pg for a 13.1 Mm2 soil area (Table 4). A second delineation is given by the “Supplementary data to the HWSD” (Fischer et al., 2008). This database indicates on a 5′ grid the presence of continuous or discontinuous (i.e., excluding sporadic and isolated) permafrost that is based on the analysis of the snow-adjusted air frost number (H. van Velthuizen, IIASA, personal communication, 2011) as used for the Global Agro-ecological Zones Assessment v.3.0 (Fischer et al., 2008). This extent (19.5 Mm2 cell area, Fig. 2) encompasses the area of soils with a gelic phase and contains 185 Pg SOC on 16.7 Mm2 soil area according to the HWSD. A third permafrost delineation (24.9 Mm2 cell area) is described by the Circum-Arctic Map of Permafrost and Ground-Ice Conditions (CAMP; Heginbottom et al., 1993), which comprises 12 categories of permafrost and ground ice prevalence without a defined depth limit for the occurrence of permafrost. The CAMP permafrost region (including permafrost in the Alps and Central Asian ranges) represents 21.7 Mm2 soil area of the HWSD with 249 Pg SOC in the top 1 m.

SOC stocks in wetlands are considerable because water reduces the availability of oxygen and thus greatly reduces decomposition rates (Freeman et al., 2001). Draining of wetlands often greatly increases the decomposition of dead plant material, which results in the release of carbon dioxide into the atmosphere. This process can significantly affect the global C budget when it happens on a large scale. There is, however, no consensus of what constitutes a wetland at the global scale (Mitra et al., 2005). Therefore, the volume of wetland soil and its C mass are also uncertain (Joosten, 2010).

The most detailed and recent maps of global scope with detailed wetland classification (Köchy and Freibauer, 2009) are the Global Land Cover Characteristics database v.2.0 (GLCC; Loveland et al., 2000), which comprises up to 6 wetland types (“Wooded Wet Swamp”, “Rice Paddy and Field”, “Inland Water”, “Mangrove”, “Mire, Bog, Fen”, “Marsh Wetland”), and the Global Lakes and Wetland Database (GLWD; Lehner and Döll, 2004), which comprises 12 wetland categories. Both maps have a resolution of 0.5′. The GLCC originates from analysis of remote sensing data in the IGBP. Lehner and Döll compiled their database from existing maps, including the GLCC, and inventories. Some wetland types are restricted geographically due to the heterogeneous classification across the source materials. The categories “50–100 % wetland” and “25–50 % wetland”, for example, occur only in North America and “wetland complex” occurs only in Southeast Asia. One consequence is that the global extent of “bogs, fens, and mires” in the GLWD, 0.8 Mm2, is smaller than the Canadian area of peatlands, 1.1 Mm2 (Tarnocai et al., 2002), which is dominated by bogs and fens.

The spatial overlap of the GLWD and the GLCC categories is rather small (Table 6). Only the “Mire, Bog, Fen” category of the GLCC has been adopted completely by the GLWD (Lehner and Döll, 2004). Even categories with similar names like “Freshwater Marsh” vs. “Marsh Wetland” and “Swamp Forest, Flooded Forest” vs. “Wooded Wet Swamps” show little spatial overlap. Despite the GLWD's overall larger wetland area it does not include the areas identified as “rice paddies” in the GLCC.

Based on the intersection of GLWD and the HWSD (Fig. 3), the global SOC mass in the top 1 m of soil of permanent and non-permanent wetlands (excluding lakes, reservoirs, and rivers) is 140 Pg (on 117 Mm2 soil area). Using the GLCC Global Ecosystems classification, the area covered by wetlands (excluding inland waters) is much smaller (3 vs. 12 Mm2) and contains only 34 Pg SOC (Table 7). The difference is partly due to the classification of large parts of North America (including the prairie) as temporary or patchy wetland in the GLWD, but even wetlands in a stricter sense cover twice the area and contain nearly twice the mass of SOC in the GLWD compared to the GLCC. Therefore, we combined both maps for the assessment of SOC stocks and masses (Table 7).

The differences in SOC mass estimates between the GLWD and the GLCC indicate that wetland types are defined heterogeneously and that especially the classification of swamp forests, marshes, mangroves, and rice paddies needs to be harmonized. The contrasting land cover classification could be overcome by using the more generic land cover classes developed within the UN Framework Convention on Climate Change (di Gregorio and Jansen, 2005). Remote sensing methods are being developed to improve the mapping of wetlands, e.g., the GlobWetland project (http://www.globwetland.org, and Journal of Environmental Management 90, special issue 7) or the Wetland Map of China (Niu et al., 2009).

Soils in the tropical land area (50 Mm2 within 23.5∘ N–23.5∘ S) contain 355 Pg SOC in the top 1 m (Table 8). The high intensity of rain in some parts of the tropics contributes to the presence of wetlands (union of GLWD and GLCC classes as in the previous section) in 9 % of the tropical land area (50 Mm2 within 23.5∘ N–23.5∘ S), containing 40 Pg SOC (Table 8, excluding lakes, reservoirs, and rivers). Most of the wetland SOC (27 Pg) is found in marshes and floodplains, as well as in swamp or flooded forests. The GLCC category with the highest SOC mass (10 Pg) is “Rice Paddy and Field” (1.2 Mm2 soil and cell area), but only 14 % of this area is recognized as wetland in the GLWD.

The permafrost region can be delineated according to different criteria (see previous section). Tarnocai et al. (2009) used the CAMP's permafrost classification (20.5 Mm2 grid cell area, excluding the Alps and Central Asian ranges) together with SOC and soil information from the Northern Circumpolar Soil Carbon Database (NCSCDB, Hugelius et al., 2013) to estimate SOC mass in the permafrost region. The NCSCDB includes soil profile data not incorporated into the HWSD. Data for calculating SOC stocks (C concentration, BD, depth, coarse fragments) in the upper 3 m were derived from 1038 pedons from northern Canada, 131 pedons from Alaska, 253 pedons from Russia, 90 peat cores from western Siberia, 266 mineral and organic soils from the Usa Basin database, and an unspecified number of profiles from the WISE database (v.1.1) for Eurasian soils. Extrapolations were used to estimate SOC mass in mineral soils and Eurasian peat soils > 1 m depth. The spatial extent of soil classes was obtained from existing digital and paper maps. Tarnocai et al.'s (2009) estimate of 496 Pg* for the 0–1 m depth is much higher than that of the HWSD's mass in the CAMP's permafrost region (233 Pg). The difference is partly due the limit of 2 m that the HWSD uses for distinguishing the “gelic phase”, whereas the CAMP does not refer to a depth limit (Heginbottom et al., 1993). The difference in mass is not only due to contrasting definitions and extent; even more so it is due to the greater SOC stock calculated from the NCSCDB (Table 5). In the NCSCDB the mean SOC areal density of soil in all permafrost classes is > 20 kg m-2, whereas the mean SOC areal density is 11.4 kg m-2 in the HWSD across all classes. The difference suggests that the BD of frozen organic soil is higher than assumed by us.

Inaccuracies associated with the mass estimates arise from incomplete knowledge of the spatial distribution of soil classes, soil depths, sparse distribution of soil profile data, and a lack of soil profiles with a full complement of measured data. Tarnocai et al. (2009) extensively discuss the uncertainty in their estimates. In terms of categories of confidence of the Intergovernmental Panel on Climate Change's Fourth Assessment Report (IPCC AR4), Tarnocai et al. have medium to high confidence (> 66 %) in the values for the North American stocks of the top 1 m, medium confidence (33–66 %) in the values for the Eurasian stocks of the top 1 m, and very low to low confidence (< 33 %) in the values for the other regional stocks and stocks of layers deeper than 1 m. Here we note only that major uncertainty is linked to the area covered by high-latitude peatlands (published estimates vary between 1.2 and 2.7 Mm2), which alone results in a range of 94–215 Pg SOC. In addition to the SOC mass in the top 1 m, Tarnocai et al. (2009) estimated that the permafrost region contains 528 Pg* in 1 to 3 m depth, and 648 Pg* in depths greater than 3 m. The C mass contained in > 3 m depth of river deltas is potentially great (241 Pg*; Tarnocai et al., 2009) but is based solely on extrapolation on the SOC stock and area of the Mackenzie River delta. Yedoma (Pleistocene loess deposits with high Corg) SOC mass (407 Pg*, > 3 m depth) is also associated with great uncertainty. The estimate (adopted from Zimov et al., 2006) is based on a sketched area of 1 Mm2 in Siberia (thus excluding smaller Yedoma deposits in North America) and mean literature values for depth (25 m), whose ranges extend > ± 50 % of the mean.

Wetlands with the highest Corg and highest SOC stocks are bogs, fens, mires, and marshes and the “25–50 %” and “50–100 %” wetlands in boreal North America. The latter two categories represent mostly bogs, fens, and small lakes. Due to their high Corg, these wetland types can also be classified as peatland.

The global area of peatland with a minimum peat depth of 30 cm is 3.8 Mm2 based on the International Mire Conservation Group Global Peatland Database (GPD; Joosten, 2010). Total SOC mass of peatlands in the GPD is 447 Pg* for their total depth. This estimate is considered conservative because mangroves, salt marshes, paddies, paludified forests, cloud forests, dambos, and Cryosols were omitted because of a lack of data. The information in the GPD is very heterogeneous. Missing data for calculating SOC mass had to be estimated. For some countries only the total area of peatland was known. When depth information was missing or not plausible, a depth of 2 m was assumed in the GPD, although most peatlands are deeper (Joosten, 2010). It is not clear which default values were used for Corg or BD in the assessment. C content (organic C fraction of ash-free mass) varies from 0.48–0.52 in Sphagnum peat to 0.52–0.59 in Scheuchzeria and woody peat (Chambers et al., 2010/2011). Values of BD show much stronger variation. Ash-free bulk density ranged from < 0.01 to 0.23 kg dm-3 in 4697 samples (Chambers et al., 2010/2011) with a median of 0.1 kg dm-3. The variation is due to water content, soil depth, plant material, and degree of decomposition (Boelter, 1968). The highest density is found in well-decomposed, deep peat of herbaceous or woody origin at low water content. When wet peatlands are drained, they may no longer qualify as wetlands, but they remain peatlands with high Corg and a large SOC mass. Drainage exposes the carbon to oxygen and thus accelerates peat decomposition and, depending on circumstances, an increase in BD. The great variation demands that BD of peatlands actually be measured at several depths and at ambient soil moisture at the same time as the C concentration. If this is not possible, PTFs of BD for peat ought to include water content, decomposition status, and plant material.

Peatlands with a certain thickness of organic layer qualify as Histosols. The HWSD adopted the FAO definition that “Soils having an H horizon of 40 cm or more of organic soil materials (60 cm or more if the organic material consists mainly of sphagnum or moss or has a bulk density of less than 0.1) either extending down from the surface or taken cumulatively within the upper 80 cm of the soil; the thickness of the H horizon may be less when it rests on rocks or on fragmental material of which the interstices are filled with organic matter” (FAO, 1997). The area covered by Histosols in the HWSD (Fig. 4) is 3.3 Mm2 (cell area multiplied by fraction of Histosol), slightly lower than the area given by the GPD, and contains 113 Pg SOC. The total area of cells with at least some fraction of Histosol, however, is 10 Mm2, containing 188 Pg SOC. The area of Histosol outside wetlands (1.7 Mm2) might indicate that a large portion of originally wet peatland has been drained and is exposed to decomposition.

Six percent of the area of each of the two C-richest tropical wetland types are categorized as Histosols in the HWSD, totaling only 0.1 Mm2. Including non-wetlands, the total area of Histosols in the HWSD, 0.4 Mm2, agrees well with the most recent and detailed, independent estimate of tropical peatland area (Page et al., 2011, defining peatland as soil having > 65 % organic matter in a minimum thickness of 30 cm). The total mass of SOC in grid cells of the spatial layer with at least some fraction of Histosol is 24.2 Pg.

Page et al. (2011) used peatland area, thickness, BD, and Corg to calculate the SOC mass for each country within the tropics of Cancer and Capricorn. They tried to trace the original data and used best estimates where data were missing. Most data were available for area, but less data were available for thickness. Page et al. (2011) used 25 % of maximum thickness when only this information was reported instead of mean thickness and used 0.5 m when no thickness was reported. The percentiles of the frequency distribution of their best estimate of thickness weighted by their best estimate of area per country is 0–10 %: 0.5 m; 25 %: 1.75 m; 50–90 %: 5.5 m; 97.5 %: 7.0 m; mean: 4.0 m ± 2.2 m SD. This distribution can be used for estimates of SOC mass and associated uncertainty in other tropical peatlands. Data on BD and SOC concentration were rare. When they were provided they often referred only to the subsurface, although these parameters vary with depth. When these data were missing, Page et al. (2011) used 0.09 g cm-3 and 56 % as best estimates based on literature reviews. The best estimate of SOC mass for tropical peatlands of Page et al. (2011) is 88.6 Pg* for the whole soil depth, with a minimum of 81.7 and a maximum of 91.9 Pg*. If one assumes an average peat thickness of 4 m and uniform vertical mass distribution, the top 1 m contains 22 Pg* of SOC, close to our HWSD-based estimate for grid cells containing Histosol (24 Pg). Thus, peatlands may contain about 6 % of the tropical SOC mass within the first meter and approximately 21 % of the total tropical SOC mass (without depth limit). Obviously, the uncertainty in these estimates is great.

Joosten (2010) estimated SOC mass for individual tropical countries based on the Global Peatland Database. For some countries the difference between Joosten's and Page et al.'s estimates are large. For example, Joosten's estimate for Sudan is 1.98 Pg*, whereas Page et al. have 0.457 Pg*. These differences may be caused by different definitions of “peat” and variability in depth estimates, SOC concentration, and BD in the data sources.

The estimate of the global SOC mass within the top 1 m based on the HWSD (1062 Pg) can be improved if and where other sources provide better estimates. The HWSD estimate of SOC mass for tropical peatlands agreed well with other sources. The SOC mass in the permafrost region estimated by Tarnocai et al. (2009) appears to be more accurate than that of the HWSD. Therefore, for the permafrost region we substitute the HWSD-based estimate (-233 Pg, Table 5) by Tarnocai et al.'s estimate (+496 Pg). This calculation (1062–233 + 496 Pg) updates the global SOC mass within the top 1 m to 1325 Pg.

For including deeper soils in an estimate of the global SOC mass, we first consider estimates of deeper soil layers for the permafrost region and tropical peatlands. The best estimate of the SOC mass below 1 m for the permafrost region known to us is 1176 Pg (calculated from Tarnocai et al., 2009). In order to estimate the mass for 1–4 m depth of tropical peatlands, we use three-quarters of Page et al.'s best estimate for the top 4 m (66.5 Pg). An additional 389 Pg SOC is contained below 1 m outside the permafrost region and the tropics (Jobbágy and Jackson, 2000). In total, the mass of SOC in the soil is about 3000 Pg, but large uncertainties remain, especially for depths > 1 m.

Another source of uncertainty is the estimation of BD. The BD of peat varies between 0.05 and 0.26 kg dm-3 (Boelter, 1968). If the same range holds for Histosols (3.3 Mm2 Histosol area, 1 m depth, 34 % Corg), this variation alone introduces an uncertainty range of -56 to +180 Pg into the estimate of global SOC in the top meter, which is larger than the estimated annual global soil respiration (79.3–81.8 Pg C; Raich et al., 2002). The areal extent of organic soils, their depth, and the BD at different depths should therefore receive the greatest focus of future soil mapping activities.

Soil monitoring is crucial for detecting changes in SOC stocks and as a reference for projecting changes in the global carbon pool using models (Wei et al., 2014; Wieder et al., 2014; Yan et al., 2014). The following conclusions from our study and a workshop of soil experts (Köchy and Freibauer, 2011) with respect to improved soil monitoring agree with more comprehensive recommendations by an international group of experts (Jandl et al., 2014). In situ measurements of soil Corg, soil depth, and BD must still be improved, collected, and made available for calculating global SOC mass. Extra care is necessary to reduce variability of data because variability reduces the potential of detecting change. Classification of soils as currently used in mapping produces uncertainty in the reported C stock when the characteristics of soil classes are aggregated and then used in further calculations. The use of pedotransfer rules and functions further increases the uncertainty in the real values. Since PTFs are entirely empirical in nature, it is preferable that they be derived from soils that are similar in nature to the soils to which the functions will be applied. For the purposes of detecting actual change in C stocks their uncertainty should be quantified. Of course it would be best if Corg, BD, and coarse fragments were measured at the same point or sample to reduce effects of spatial variability. Predictive mapping techniques, including geostatistics, modeling, and other quantitative methods (McBratney et al., 2003; Grunwald et al., 2011), especially in conjunction with proximal (radiometry, NIR spectroscopy) or hyperspectral remote sensing of soil properties (Gomez et al., 2008; Stockmann et al., 2015), can potentially reduce uncertainties in SOC mapping introduced by soil classification and help in interpreting spatiotemporal patterns. Regardless of whether soils are mapped in the classical way or by predictive methods, mapping of soils should be coordinated with the direct or indirect mapping of SOC input and its controlling factors (land use, land cover, crop type, land use history and land management) as well as extent and soil depth of wetlands, peatlands, and permafrost.

Uncertainty in SOC stocks in current maps could further be reduced if all soil types and regions were well represented by soil profile data with rich soil characteristics. Many soil profile data collected by governments and publicly funded projects remain unused because they are not available digitally; their use is restricted because of data protection issues, or because they are only known to a very limited number of soil scientists. Existing approaches such as the NCSCDB, the GlobalSoilMap.net project, and the Global Soil Partnership (coordinated by the FAO) are important steps to improve the situation. These activities would benefit further if all publicly funded, existing soil profile data were made publicly available to the greatest possible extent.

Another source of uncertainty is introduced because profile data and soil maps have been generated by a multitude of methods. Furthermore, if different methods are preferably used for particular soil types or regions, small differences multiplied by large areas can result in significant differences at the global level. Therefore, international activities to harmonize methods of sampling, calculation, and scaling should be supported. The harmonized methods should then actually be applied in soil sampling. Preferably, samples should be archived so that soils can be reanalyzed with improved or new methods or for checking data by more than one laboratory.

The strong effect of BD values on the calculation of SOC stocks and regional or global masses should guide the focus of global observation networks to improve not only the observation of SOC concentrations but also that of BD. Furthermore, our study describes for the first time the frequency distribution of SOC stocks within broad classes of land use/land cover and C-rich environments based on one of the most exhaustive, harmonized, and spatially explicit global databases available to date. The frequency distribution allows for a more focused spatial extrapolation and assessment of accuracy in studies where SOC is used as an independent variable (e.g., Pregitzer and Euskirchen, 2004). The frequency distributions also provide a foundation for targeting SOC conservation measures (Powlson et al., 2011) and for improving carbon accounting methods with associated uncertainties as used in the UNFCCC (García-Oliva and Masera, 2004).

CO2 emissions from soils are used in calculations of the global carbon cycle. Direct observations of CO2 emissions from soils (e.g., by eddy-flux towers), however, cannot be implemented in a spatially contiguous way. Indirect measurements by remote sensing can improve the spatial coverage but require ground observations for conversion from observed radiation to loss of CO2 from soils and distinction from other CO2 sources (Ciais et al., 2010). At the global scale, in situ measurements must be complemented by modeling activities, which are greatly improved if variation in key factors like SOC can be accounted for. Thus, more detailed information on the global distribution of SOC, both horizontally and vertically, including accounts of its accuracy and its variability, is necessary to improve estimates of the global carbon flow.