This is a multi-gas (CO2, CH4, N2O) spatially explicit (0.5∘) database on gross AFOLU emissions and associated uncertainties for the tropics and subtropics for the period 2000–2005, at Tier 2 and Tier 3 levels (see Supplement for the definition of tiers). This database locates the hotspots of tropical AFOLU emissions, which should help to estimate mitigation potentials, and prioritize the areas and the land activities that require most urgent mitigation action. It combines available published GHG datasets for the key sources of emissions in the AFOLU sector as identified by the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (AR5, Smith et al., 2014): deforestation, forest degradation (fire, wood harvesting), crop soils, paddy rice, and livestock (enteric fermentation and manure management). It also includes agricultural peatland decomposition using Tier 1 emission factors (see details in Roman-Cuesta et al., 2016). Forest emissions focus on aboveground biomass, with the exception of peat fires. More detailed methodological information is available in Roman-Cuesta et al. (2016).

FAOSTAT covers agriculture, forestry, and other land uses and their associated emissions of CO2, CH4 and N2O, following IPCC, 2006 guidelines at Tier 1 (Tubiello et al., 2013, 2014). Emissions are estimated for nearly 200 countries annually, for the reference period of 1961–2012 (agriculture) and 1990–2012 (FOLU), based on national activity data submitted by countries and further collated by FAO (Food and Agriculture Organization of the United Nations). Projected emission data are available for 2030 and 2050. FAOSTAT includes estimates of emissions from biomass fires, peatland drainage, and fires, based on geospatial information, as well as on forest carbon stock changes (both emissions and removals) based on national-level FAO Forest Resources Assessment data (FRA, 2010).

The Emissions Database for Global Atmospheric Research (EDGAR) provides global GHG emissions from multiple gases (CO2, CH4, N2O, HFCs, PFCs, and SF6) at 0.1∘ and country levels. The EDGAR database covers all IPCC sectors (energy, industry, waste management, and AFOLU), mostly applying IPCC 2006 guidelines for emission estimations (EDGAR, 2012). We downloaded the EDGAR 4.2 Fast Track 2010 (FT 2010). FT 2010 emissions cover the period 2000–2010 on an annual basis, at the country level.

Houghton's bookkeeping model calculates the net and gross fluxes of carbon (CO2 only) between land and atmosphere that result from land management (Houghton, 1999, 2012; Houghton and Hackler, 2001; Houghton et al., 2012). The net estimate includes emissions of CO2 from deforestation, shifting cultivation, wood harvesting, wood debris decay, biomass burning (for deforestation fires only, peatland fires were not included in our version of their data), and soil organic matter from cultivated soils. It also includes sinks of carbon in forests recovering from harvest and agricultural abandonment under shifting cultivation. The model, however, does not include forests that are not logged, cleared, or cultivated. Rates of growth and decomposition are ecosystem specific and do not vary in response to changes in climate, CO2 concentrations, or other elements of environmental change. Therefore, forests grow (and wood decays) at the same rates in 1850 and 2015. Unlike other databases, all carbon in a considered ecosystem is accounted for in live vegetation, soil, slash (woody debris produced during disturbance), and wood products. We downloaded regional annual emissions from the TRENDS (1850–2005) dataset for the tropics: Central and South (CS) America, tropical Africa, and South and South East Asia. Only net emissions were available. No spatially disaggregated data were offered (i.e. countries). Houghton data are, unlike all the other datasets, net aggregated FOLU estimates for CO2 only.

These are gross FOLU tropical emissions published by Baccini et al. (2012). Data are gross emissions for the period 2000–2010 disaggregated into deforestation (4.18 PgCO2 yr-1), wood harvesting (1.69 PgCO2 yr-1), biomass burning (2.86 PgCO2 yr-1), and wood debris decay (3.04 PgCO2 yr-1). We excluded this last variable to make it more comparable to the other datasets, where CWD is frequently excluded (Table 1). Baccini estimates refer to a tropical area slightly smaller than our study region and they are offered as an aggregated value (no continental or country data are available).

The EPA dataset contains global non-CO2 projected emissions for the period 1990–2030 for the agriculture, energy, industrial processes, and waste sectors, for more than twenty gases. EPA uses future net emissions projections of non-CO2 GHGs as a basis for understanding how future policy and short-term, cost-effective mitigation options can affect these emissions. EPA follows the Global Emissions Report, which uses a combination of country-prepared, publicly available reports consistent with IPCC guidelines and guidance (USEPA, 2013). When national emissions estimates were unavailable, EPA produced its own non-CO2 emissions using IPCC methodologies (i.e. international statistics for activity data, and the default IPCC Tier 1 emission factors). Deviations to this methodology are discussed in each of the source-specific methodology sections of USEPA (2012). No FOLU estimates are included in this dataset. We downloaded agricultural emissions offered at 5-year intervals at country level, disaggregated by gas (N2O and CH4) and by emission sources.

The AR5 is a synthesis report, not a repository of global data. However, new AFOLU data are produced by merging peer-reviewed data such as Figs. 11.2, 11.4, 11.5, and 11.8 in chapter 11 of the AR5 (Smith et al., 2014). We will compare our six datasets with the data from these newly produced figures.

Table 1 shows a summary of key similarities and differences of the assessed AFOLU datasets and the data from the AR5. The exact variables used for each database are described in Table S1 in the Supplement.

Deforestation emissions were 2.9 (1.0–10.1), 3.7, 2.5, and 4.2 PgCO2 yr-1 (for Hotspots, FAOSTAT, EDGAR, and Baccini respectively), with Baccini and EDGAR showing the highest and the lowest values. However, their values represent very different scenarios: gross deforestation for the Hotspots and Baccini datasets (forest losses only), net deforestation for FAOSTAT (forest losses minus forest gains), and forest fire and post-burn decay for EDGAR (Table 3). The Hotspots dataset (Harris et al., 2012 and Baccini et al., 2012) offers gross deforestation estimates that rely on Hansen et al. (2010)'s forest cover loss areas. However, they report different tropical emissions (0.81 and 1.14 PgC.yr-1) because they use different carbon density maps: Harris et al. (2012) rely on Saatchi et al. (2011) and Baccini rely on Baccini et al. (2010). EDGAR does not provide a category for deforestation, and their Forest Fire and Decay category (5F; Table 3 and Table S1) is used as a proxy for deforestation (Tubiello et al., 2015). Such an approximation leads to underestimations since not all carbon losses from deforestation are necessarily associated with the use of fire (Tubiello et al., 2015). In spite of being net emissions, the deforestation estimates for FAOSTAT were higher than the gross estimates from Hotspots and Baccini. This is partly due to FAOSTAT's inclusion of fire emissions from humid tropical forests (see Sect. 3.2.3), which the other datasets did not have. Baccini's larger estimates of gross deforestation included more carbon pools than the other datasets (i.e. soil, coarse woody debris (CWD), litter). Baccini et al. (2012) reported that their estimated gross and net emissions from tropical deforestation were the same value (4.2 Pg CO2 yr-1). The difference with Houghton net emissions (5.4 PgCO2 yr-1; Fig. 2) corresponds, then, to non-offset carbon emissions from other land uses and activities included in the bookkeeping model: degradation by logging and shifting cultivation, decomposition and decay, and cultivated soils. Houghton tropical net emissions for 2000–2005 are high, but are lower than Houghton reported net estimates in the 1980s (7 PgCO2 yr-1; Houghton, 1999).

Forest degradation can be defined in many ways (Simula, 2009), but no single operational definition has been agreed upon by the international community (Herold et al., 2011a). It typically refers to a sustained human-induced loss of carbon stocks within forest that remains forest. In this study, similarly to Federici et al. (2015), we consider degradation to be any annual removal of carbon stocks that does not account for deforestation, without temporal-scale considerations (i.e. time needed for disturbance recovery or time to guarantee a sustained reduction of the biomass). We assessed two major degradation sources: wood harvesting and fire. Soil degradation is poorly captured in many datasets, and mainly focuses on fire in equatorial Asian peatland forests and drained peatlands (Hooijer et al., 2010). A better understanding of the processes and emissions behind forest degradation is key for climate mitigation efforts, not only because forest degradation is a widespread phenomenon (i.e. affects much larger areas than deforestation; Herold et al., 2011b), but also because the lack of knowledge of net carbon effects frequently results in assumptions of carbon neutrality of the affected standing forests, particularly for fire (Houghton et al., 2012; Le Quéré et al., 2014), which likely leads to an underestimation of forest and AFOLU emissions (Brando et al., 2014; Turetsky et al., 2015; Roman-Cuesta et al., 2016).

Gross emissions from forest degradation were larger than deforestation for the Hotspots, EDGAR, and Baccini datasets, with degradation-to-deforestation ratios of 108, 120, and 128 % respectively. FAOSTAT had degradation emissions of 60 % of the deforestation, partly due to its anomalously low fire contribution (see next section). Houghton et al. (2012) pointed out that global FOLU net fluxes were led by deforestation with a smaller fraction attributable to forest degradation, while the opposite was true for gross emissions (degradation being 267 % of deforestation emissions). This large ratio relates to their inclusion of shifting cultivation under degradation. This is a definition issue, which would not fit the definition of degradation chosen in this study, where a complete forest cover loss would represent deforestation and not degradation.

Fire led the gross forest degradation emissions in the tropics in 2000–2005 (Fig. 2): 2 (1.1–2.7), 0.2, 3.4, 2.9 Pg CO2 eq. yr-1 for the Hotspots, FAOSTAT, EDGAR, and Baccini datasets respectively; Fig. 2). The Hotspots estimates are conservative compared to Van der Werf et al. (2010)'s global emissions of 7.7 Pg CO2 eq. yr-1 for 2002–2007, due to the removal of CO2 from deforestation fires (to avoid double counting with deforestation emissions), the exclusion of fires in grasslands and agricultural residues, and Hotspots' smaller study area. FAOSTAT and EDGAR had the lowest and the highest fire values. The lowest values in FAOSTAT relate to omissions that are currently in the process of being corrected (S. Rossi, personal communication, 2016): (1) the complete exclusion of CO2 from fire in humid tropical forests and other forests (Table 3, Table S1), which FAOSTAT relocated as net forest conversion emissions, partly explains their larger deforestation values (FAOSTAT kept CH4 and N2O for fire in humid tropical forests and other forests), and (2) the use of default parameters for fuel in peats from the IPCC 2006 guidelines instead of the new IPCC Wetland Supplement, because they offer considerably higher values (Rossi et al., 2016). Moreover, FAOSTAT uses GFED3.0 burned area (Giglio et al., 2010) in their estimates while the other datasets use GFED3.0 emissions (Van der Werf et al., 2010). EDGAR fire emissions were the largest most likely because they included decay. Their dataset considers some undefined forest fires (5A) and wetland/peatland fires and decay (5D; Table 3; Table S1). Peatland decay probably explains EDGAR's larger emissions in Asia, while we assume that EDGAR's highest fire emissions for CS America might respond to deforestation fires which were not included in the Hotspots to avoid double counting with deforestation, and relocated in FAOSTAT to deforestation emissions (Fig. 3, Table 3). The Hotspots dataset showed higher gross fire emissions for Africa due to the inclusion of woodland fire, which EDGAR and FAOSTAT probably excluded. Baccini et al. (2012)'s fire emissions: 2.9 Pg CO2 eq. yr-1 (2000–2010) derived from Houghton's bookkeeping, but it is unclear how these emissions were estimated.

In spite of the importance of fire as a degradation source, this variable is frequently incompletely included, either through unaccounted gases (i.e. CH4 and N2O are excluded in the carbon community but their omission represent 17–34 % of the gross CO2 fire emissions; Valentini et al., 2014; Roman-Cuesta et al., 2016) or to unaccounted components (i.e. fires in tropical temperate forests such as conifers or dry forests such as woodlands are frequently excluded; Houghton et al., 2012). Unaccounted fire emissions are also derived from methodological choices (i.e. only interannual fire anomalies are considered; Le Quéré et al., 2014), from poor satellite observations such as understory fires in humid closed canopy forests; Alencar et al., 2006, 2012; Morton et al., 2013), or satellite fire omissions in certain regions (i.e. high Andean fires; Bradley and Millington, 2006; Oliveras et al., 2014). Other omissions relate to the current exclusion of non-Asian peatland fires (i.e. American tropical montane cloud forest peatland fires; Asbjornsen et al., 2005; Roman-Cuesta et al., 2011; Oliveras et al., 2013; Turetsky et al., 2015).

Fire suffers, moreover, from a series of assumptions that do not apply so easily to other types of degradation: (1) assuming a non-human nature of the fires (deforestation fire vs. wildfires), which in tropical areas contrasts with multiple citations referring to the 90 % human causality of fires (Cochrane et al., 1999; Roman-Cuesta et al., 2003; Alencar et al., 2006; Van der Werf et al., 2010); (2) assuming force majeure conditions that lead to non-controllable fires due to extreme climate conditions, which frequently result in incomplete assessment and reporting of emissions. This assumption contrasts with research on how human activities have seriously increased fire risk and spread in the tropics (Uhl and Kauffman, 1990; Laurance and Williamson, 2001; Roman-Cuesta et al., 2003; Hooijer et al., 2010), and clearly expose how most of the fires in the humid tropics would not occur in the absence of human influences over the landscape (Roman-Cuesta et al., 2003). (3) We assuming carbon neutrality and full biomass recovery after fire in standing forests. This is a generous assumption that contrasts with numerous studies on tropical forest die-back following fire events in non-fire adapted humid tropical forests (Cochrane et al., 1999; Barlow and Peres, 2008; Roman-Cuesta et al., 2011; Brando et al., 2012; Oliveras et al., 2013; Balch et al., 2015). All these phenomena cast doubts on the robustness of these assumptions and call for a much more comprehensive inclusion of fire emissions into forest degradation budgets.

There is not a unique way to estimate wood harvesting emissions as exposed in the guidelines for harvested wood products of the IPCC (IPCC, 2006). Assumptions regarding the final use of the wood products, decay times, substitution effects, international destination of the products, and time needed for forests to recover their lost wood can fully change the emission budgets. In our study, wood-harvesting emissions were 1.2 (0.7–1.6), 2.0, 1.7 PgCO2 yr-1 for the Hotspots, FAOSTAT, and Baccini data respectively (Tables 3, S1). Harvested wood products are derived from FAO country reports (i.e. FAOSTAT forest products). All datasets included fuel wood and industrial roundwood (Tables 3, S1). EDGAR excluded fuelwood from the AFOLU budget and placed it instead into the energy budget (EDGAR, 2012), which explains its absence in Fig. 2. Wood-harvesting emissions were larger in FAOSTAT than in the Hotspots data (Fig. 2), partly due to the inclusion of some extra categories of fuels (i.e. charcoal and residues) that were not included in the Hotspots database (Tables 3, S1). Charcoal represents 26 % of the total wood-harvesting emissions in FAOSTAT. Differences on wood harvesting affected Asia and CS America more (where the Hotspots data were half of the FAOSTAT data), while Africa presented almost identical values (Fig. 3). The reasons for these continental differences are unclear. Baccini's high emissions for wood harvesting could partly be related to their inclusion of extra biomass due to felling damages (i.e. 20–67 % of the AGB is damaged, and 20 % are left dead in BGB; Houghton, 1999).

Livestock emissions were the most homogeneous among the emissions sources (Fig. 2) with estimates of 1.2 (0.8–1.5), 1.1, 1.2, 1.1 Pg CO2 eq. yr-1 for the Hotspots, FAOSTAT, EDGAR, and EPA respectively, in range with the estimates in the AR5 (Fig 11.5 in Smith et al., 2014). Values were similar in spite of being derived from different tiers (i.e. Tier 3 for Herrero et al. (2013), Tier 1 for FAOSTAT and EDGAR. EPA used Tier 3 but for incomplete data series, otherwise Tier 1 was applied, USEPA, 2013). All datasets included enteric fermentation (CH4) and manure management (N2O, CH4). All of them relied on FAO data for livestock heads, although they used different years (i.e. 2000 for Herrero et al. (2013) data in the Hotspots, and 2007–2010 for EDGAR). From a continental perspective, FAOSTAT and EDGAR estimates were the closest while the Hotspots and EPA estimates were less similar. The Hotspots showed higher emissions for Africa and Asia and lower emissions for CS America compared to the other three datasets. Divergences likely relate to different tiers. CS America and Asia showed the highest values, with Africa following closely (Fig. 3), similar to what is reported in the AR5 (Smith et al., 2014). Globally, livestock farming is the largest source of CH4 emissions, with three-quarters of the emissions coming from developing countries, particularly Asia (USEPA, 2013, Tubiello et al., 2014). Three out of the top five emitting countries are in the tropics: Pakistan, India, and Brazil (USEPA, 2013) and, while Asia hosts the largest livestock emissions, the fastest growing trends in 2011 correspond to Africa (Tubiello et al., 2014).

The estimates of cropland emissions reached values of 0.18 (0.16–0.19), 0.56, 0.6, and 0.64 Pg CO2 eq. yr-1 for the Hotspots, FAO, EDGAR, and EPA datasets respectively, for N2O and CO2 from changes in soil organic carbon content. Cropland soil emissions (N2O and soil organic carbon stocks (CO2) heavily depend on land management practices (i.e. tillage, fertilization, and irrigation practices) and climate (Crowther et al., 2015). We chose exactly the same land practices in all datasets to allow comparisons (Tables 3, S1). For this reason, we excluded N2O emissions from grassland soils, drainage of organic soils, and restoration of degraded lands (Table 3). This restrictions resulted in lower emissions than those estimated for cropland soils in the AR5 (Fig. 11.5 in Smith et al., 2014). The Hotspots and EPA showed the lowest and the highest estimates (Figs. 2, 3). With the exception of the Hotspots, the other datasets agreed well at the tropical scale, with FAOSTAT and EDGAR being almost identical, also at continental scales. EPA disagreed more than the other datasets at the continental scale, with underestimations for Asia that were probably related to the parameterization of its emission model. All three datasets used FAO activity data, and for EDGAR and FAOSTAT the same emission factors must have been used. The Hotspots showed anomalously low emissions partly because it only included six major crop types (maize, soya, sorghum, wheat, barley, and millet) for which the emission model (DAYCENT) counted on reliable parametrization (S. Ogle, personal communication, 2016). Emissions from other important crops in the tropics (e.g. sugar cane, tobacco, tea) were excluded, as well as emissions from croplands in organic soils, due to model constraints.

Estimates of drained peatlands (mainly for agricultural purposes) suggest large omissions in the Hotspots database with emissions 1 order of magnitude lower (28 TgCO2 eq. yr-1) than FAOSTAT (ca. 500 TgCO2 eq. yr-1) and 1 order of magnitude lower than the values reported for peatland drainage in Asia alone (Hooijer et al., 2010; 355–855 TgCO2 eq. yr-1). The lower values in the Hotspots dataset relate to much smaller agricultural areas with histosols (0.4 mill ha) than those reported by FAOSTAT for the same countries (7 mill ha). This area difference is partly due to the methodological approach used by Ogle et al. (2013) in which only six major crop covers are considered: maize, wheat, sorghum, soya beans, millet, and barley, and partly to the unmatching spatial scales of histosols and croplands (i.e. 1 km for histosols and 50 km for croplands) which result in underestimations of the final area.

When paddy fields are flooded, decomposition of organic material gradually depletes the oxygen present in the soil and floodwater, causing anaerobic conditions in the soil that favour methanogenic bacteria that produce CH4. Some of this CH4 is dissolved in the floodwater, but the remainder is released to the atmosphere, primarily through the rice plants themselves. Net emission estimates for paddy rice were 0.55 (0.4–0.833), 0.33, 0.37, 0.30 Pg CO2 eq. yr-1 for the Hotspots, FAOSTAT, EDGAR, and EPA datasets respectively. The Hotspots showed the highest emissions (Fig. 2), but only in Asia (Fig. 3). Part of the reason behind these differences refers to the final gases estimated in Li et al. (2013) which included CH4, N2O and decomposition of soil organic carbon (SOC; CO2; Table 3, S1), while the others only focused on CH4. In Li et al. (2013)'s estimates, N2O were 48 % of the CH4 emissions, explaining the doubled emissions in the Hotspots database. SOC was a sink, with -0.076 PgCO2 yr-1.

Based on the explanations above, Table 4 points out the likely least reliable emission sources for each dataset considering disagreements among emission estimates due to biased/divergent/incomplete definitions and methods. Houghton's sinks are suggested as least reliable since they suffer from compatibility issued with IPCC guidance and exclude sinks from non-disturbed areas and forests undergoing disturbances other than wood harvesting or recovery from shifting cultivation (Grassi and Dentener, 2015; Federici et al., 2016).

Different datasets were developed for different purposes that have influenced the methods and approaches chosen to estimate their land use GHGs. Thus, EDGAR was created with an air pollution focus making its land emissions weaker. In contrast, FAOSTAT carries FAO's focus on land, particularly agriculture (data available since the 60s), with forest data added later through the FRA assessments (1990, 2005, 2010, 2015). The “Hotspot” database was created to identify the areas with the largest land use emissions in the tropics (emissions hotspots), while Houghton's accent is on historical LULUCF emission trends (since 1850). EPA concentrates on industrial, energy, and agricultural emissions. Forests are excluded with an interest on human health and mitigation. Moreover, due to its long existence, several datasets rely on FAOSTAT long-term agricultural data, which is probably the reason behind the higher homogeneity of agricultural emission estimates (i.e. crops, rice, and livestock among datasets). FAOSTAT forest emissions use FRA data, which get updated every 5 years. Different FRA versions strongly influence forest emission estimates which makes it important to acknowledge the FRA version used when contrasting FAOSTAT emissions and when comparing estimates (i.e. differences up to 22 % between the forest sink estimates using FRA2015 and FRA2010 have been reported by Federici et al., 2015). Similarly, official updates of Houghton's bookkeeping TRENDS data, as well as researchers' self-tuned versions of his model, result in emission differences that are difficult to track.