Acquisition of Satellite imagery

While for the first two ‘reference years’ of the study (1990 and 2000), the coverage by TM imagery from Landsat 4, Landsat 5 (1990) and by ETM+ imagery from Landsat 7 (2000) is almost complete (Beuchle et al., 2011). Landsat imagery for the 2010 reference year suffers from the failure of Landsat 7's scan line correction in 2003. For example, for South America and for the 2010 epoch (mainly covered by Landsat 5 data), 13% of the sample sites are covered by alternative imagery from other fine or medium resolution optical sensors (in particular for regions with persistent cloud cover) with only 1% of the 1230 sample sites missing due to the lack of cloud-free images. Alternative satellite data for year 2010 were received for South America and Southeast Asia, namely from the RapidEye, AVNIR-2, Kompsat and Deimos-1 sensors. Alternative satellite data were also used for Africa, in particular over the Congo Basin from the Deimos-1, UK-DMC2, and SPOT-4 & 5 satellites. The intermediate epoch 2005 (which was used for the FAO FRA-2010 Remote Sensing Survey) is not reported here because (i) the analysis is comparing the two most recent decades and (ii) there is a significant number of missing sample units (due to missing imagery) for epoch 2005 (350 in total).

Preprocessing of satellite imagery

All satellite images were preprocessed, including geo-location correction, conversion into top-of-atmosphere reflectance, atmospheric correction (haze-correction and masking of cloud and cloud shadow) and normalization by dark-object subtraction on basis of dense evergreen humid forest areas (Bodart et al., 2011). The segmentation process targeted a Minimum Mapping Unit (MMU) of 5 ha, tolerating within each sample unit a maximum number of 5% of smaller polygons resulting from the segmentation procedure. Polygons smaller than 3 ha were merged with their neighbors (Raši et al., 2011). An automatic process of object-based change detection and classification was designed for Landsat images (Raši et al., 2013) and the process adapted for alternative imagery (Desclée et al., 2013).

Thematic legend for the analysis of satellite imagery

The FAO definition of ‘forest’ includes all areas of at least 0.5 ha size with tree cover (canopy) density greater than 10% and tree height greater than 5 m (FAO, 2010). However, these thresholds cannot be ‘measured’ from Landsat satellite imagery with high accuracy. A more feasible assessment is that the canopy density be greater than 30% (FAO, JRC, 2012). The 7th Conference of the Parties of the United Nations Framework Convention on Climate Change (UNFCCC) adopted (the ‘Marrakesh accords’) a forest definition with certain flexibility: ‘Forest’ is a minimum area of land of 0.05–1.0 ha with tree crown cover of more than 10–30%. To date for UNFCC reporting, most countries are defining forests with a minimum crown cover of 30%. Within our study, we used a MMU of 1 ha as an intermediate step in image processing and a 3 ha MMU for producing the final mapping results. We used five land-cover classes at 3 ha level: (i) areas with portion of tree cover over 70% within the unit (‘Tree Cover’); (ii) those with 30–70% portion (‘Tree Cover Mosaic’); (iii) all other woody vegetation (height <5 m), including shrubs and forest regrowths (‘Other Wooded Land’); (iv) Water; and (v) ‘Other Land Cover’. The satellite data from each reference year (1990, 2000, and 2010) are processed and classified into the five land-cover classes using a supervised classifier (Raši et al., 2011, 2013).

After visual quality control by national experts, we produce from the resulting three-date interpretations the matrices of the class transitions for the two periods 1990–2000 and 2000–2010. These matrices allow calculating gross and net rates of land-cover conversions. We may only miss the processes of reclearing of shrubland within a 10-year period. Forest areas are then calculated as 100% areas of the ‘Tree Cover’ class plus 50% areas of the ‘Tree Cover Mosaic’ class.

Our study does not fully assess forest degradation because selective logging and moderate forest fragmentation occurring below the defined MMU are usually not reflected by a change in the land-cover class. However, degradation is generally an initial phase in the conversion of forests (Numata et al., 2010).

Statistical extrapolation to continental or tropical levels

It was not possible to acquire the satellite images at the exact reference date (Beuchle et al., 2011). Each sample site's area estimate was therefore linearly adjusted to the baseline dates of 30th June 1990, 2000, and 2010: this was done by assuming that the land-cover change rates are constant during the given period. Cloudy areas were considered as an unbiased loss of data, and assumed to have the same proportions of land cover as noncloudy areas within the same site. This is achieved by converting the land-cover change matrices 1990–2000 and 2000–2010 to area proportions relative to the total cloud-free land area of the sample units. For the missing sample units (4, 39, and 3 for 1990–2000 and 3, 39, and 3 for 2000–2010, for South America, Africa and Southeast Asia, respectively, from totals of 1230, 2045, and 741 sample units, respectively), we used a local average from surrounding sample sites as surrogate results. The following weights (δjj′) were applied for the local average of missing sites:

(1)

where d(j,j′) is the distance between two sites.

For the statistical estimation phase, the sample units are weighted in relation to their statistical probability of selection. Indeed, the sampling frame, although systematic, does not give equal probability to all sample units because the distance between sample units along a parallel is not the same as the distance along a meridian. All sample units are given a weight, equal to the cosine of the latitude to account for this unequal probability. The impact of these weights is moderate in tropical areas. The sample units that contain a proportion of sea compensate for unselected sample sites that contain a proportion of land (when the centre of the site is located in the sea) because they were considered as full sites.

The proportions of land-cover changes were then extrapolated to the study area using the Horvitz-Thompson Direct Expansion Estimator (Särndal et al., 1992). The estimator for each land-cover class transition is the mean proportion of that change per sample unit, given by Eqn 2:

(2)

where y_ic is the proportion of land-cover change for a particular class transition in the ith sample unit. The weight of the sample unit is w_i and m is the sum of the sample weights.

The usual variance estimation of the mean is known to have a positive bias (Stehman et al., 2011). Alternative estimators based on a local estimation of the variance have been shown to reduce the bias. We use an estimator of the standard error based on local variance estimation:

(3)

where f is the sampling rate, the weight w_jj′ is an average of the weights w_j and w_j′, and δ_jj′ is a decreasing function (1) of the distance between j and j′ (note that if we choose δ_jj′ = 0, we have the usual variance estimator). The standard error (SE) is then calculated as:

(4)

where n is the total number of available sample sites (i.e. not accounting for the missing sites even if they are replaced by a local average).

Accuracy assessment of the estimates of forest cover changes

The observations (source datasets) that are used to produce our results are derived from satellite interpretations. These surrogates to ground observations may be subject to the uncertainty (bias) (Foody, 2010), but we do not address such errors here. The use of such surrogate data for assessing area change is inevitable in many areas of the tropics where no ground observations exist and where large areas of inaccessible forests can only be monitored at affordable costs by the exploitation of satellite data. However, we performed an independent assessment over 1185, 1552, and 830 points (total: 3567 points) distributed systematically within a random subsample of 240, 338, and 166 sample units in South America, Africa, and Southeast Asia, respectively (a central point plus four points in the corners taken in each sample unit). In addition, we selected from a 9 × 9 systematic grid (81 points taken at 1 km distance in each sample unit) all points identified as change in land cover during the period 1990–2000, resulting in 1663, 1194, and 1425 points (total: 4282), respectively, for the three subregions. The corresponding polygons were carefully visually reinterpreted by independent experts using ancillary information when available (e.g., imagery from Google Earth© considering the date of imagery). This allows assessing the ‘consistency’ of the results of the interpretation.

To complement to this consistency assessment, we also compared our results to the INPE interpretations for period 1990–2000 (INPE, 2013) for a random selection of 34 sample units among the 411 sample units falling in the Brazilian Legal Amazon (Eva et al., 2012).

Combining forest cover change maps with pan-tropical biomass maps to estimate carbon fluxes

Biomass data for tree cover and other wooded land are spatially associated with each sample site, so that the per-site carbon emissions (and removals) can be calculated. We use three datasets of spatially explicit biomass information: the FAO ecozone map (FAO, 2012) combined with IPCC values (IPCC, 2006), and Baccini et al. (2012) and Saatchi et al. (2011) pan-tropical biomass maps.

The FAO ecozone map is combined with Above Ground Biomass (AGB) levels for the major tropical ecosystems (evergreen rainforests, rainforests with seasonal behavior, moist deciduous forests, dry forests, mountain ecosystems, shrubland, subtropical humid forests, grasslands, desert) derived from IPCC ‘Tier 1’ data, complemented by a regional dataset of aboveground biomass for Amazon basin (Malhi et al., 2006). We used the IPCC ratio of belowground biomass to aboveground biomass for conversion into total biomass (from 0.37 to 0.24 depending on the ecosystem).

The pan-tropical maps of forest carbon density at 500 m (Baccini et al., 2012) and 1 km resolution (Saatchi et al., 2011) are derived from a combination of field inventory plots, Geoscience Laser Altimeter System (GLAS) data points acquired from the ICESat Satellite and products of the moderate resolution imaging spectro-radiometer (MODIS). However, while Baccini et al. (2012) only estimated the AGB and aboveground carbon portion, Saatchi et al. (2011) also took into account the belowground biomass and carbon content. Both datasets show uncertainty information at continental level, and Saatchi's map provides uncertainty information for each pixel location. For the dataset of Baccini et al. (2012), we derive a map of the total carbon (above- and belowground) by applying the following equation used by Saatchi et al. (2011):

From these three pan-tropical biomass datasets, we derive five values of total biomass for the three wooded classes (‘Tree Cover’, ‘Tree Cover Mosaic’, and ‘Other Wooded Land’) for each sample site: (i) IPCC/Ecozone average; (ii) ‘Baccini average’; (iii) ‘Saatchi average’; (iv) ‘Saatchi maximum’; and (v) ‘Saatchi minimum’. The biomass values for (i), (ii), and (iii) are derived as follows: for each of the three available biomass maps, we consider the average biomass values of the pixels falling completely within the polygons labeled either as ‘Tree Cover’ or ‘Other Wooded Land’ in each single sample unit. As the polygons labeled as ‘Tree Cover Mosaic’ usually contain fragmented land cover and, moreover, are relative small leading to a limited number of 500 m or 1 km resolution pixels, we consider for this class the average biomass value of the ‘Tree Cover’ and ‘Other Wooded Land’ classes in each sample unit.

Baccini et al. (2012) provide relative standard deviation of errors on a continental basis at 6.6%, 3.6%, and 3.2% for tropical Africa, tropical America and tropical Asia, respectively, Saatchi et al. (2011) provide relative uncertainty ranges for different scales: from ±6% to ±53% at pixel (1 km resolution), ±5% at project (˜10 000 ha), and ±1% at national levels. For a sensitivity analysis of uncertainties, we use the uncertainty values reported in Saatchi's dataset at pixel level. For the ‘Tree Cover’ and ‘Other Wooded Land’ classes, we consider the average of the uncertainty values of the Saatchi pixels falling within the polygons labeled as one of these two classes for each single sample unit. These uncertainty averages are then used to produce ‘Saatchi maximum’ (‘Saatchi average’ plus uncertainty average) and ‘Saatchi minimum’ (‘Saatchi average’ minus uncertainty average) for each of these two classes. It should be noted that the average of the uncertainty values is around 60 t ha⁻¹ for all sample units with small variance between sample units.

We estimated total carbon as 50% of total biomass. We do not account for losses of carbon in soils.

In a previous study (Achard et al., 2004), the annual carbon gross emissions for the study area were calculated as the committed emissions over a 10 year period arising from 1 year of forest and woodland clearance. We aimed at taking into account the land-cover dynamics following deforestation, including the decay of product and slash pools (Ramankutty et al., 2007). These annual emissions accounted for 69% of the total initial carbon stocks. In this study, we consider only the maximum potential loss of carbon which can be emitted in the atmosphere over a long time period, corresponding to 100% of the total initial carbon stocks. In other words, we consider here that emissions would occur fully at time of clearing and referred to as committed emissions.

Carbon losses and removals from changes in forest and other wooded land cover

Our estimate of annual carbon gross loss from changes in forest and other wooded land cover are 887 MtC yr⁻¹ (range: 646–1238) and 880 MtC yr⁻¹ (602–1237) for the 1990s and 2000s, respectively (Table ​(Table2),2), the tropical humid regions contributing 66.7% and 62.2%, respectively (Fig. ​(Fig.3).3). Carbon losses are very similar over the two last decades when IPCC (2013) reports that ‘it is more likely than not that net CO₂ emissions from land-use change decreased during 2000–2011 compared to 1990–1999’. Carbon losses from deforestation of the last decade (2000–2010) represent circa 10% of Carbon emissions from fossil fuel combustion and cement production at 8.3 ± 0.7 PgC yr⁻¹ (IPCC, 2013).

Open in a separate window

Fig 3

Gross carbon losses from forest cover and Other Wooded Land (OWL) losses for all sample sites over the tree continents and the two decades: 1990–2000 and 2000–2010 (3 × 2 panels). Red color means that 100% of carbon losses are due to forest losses when blue color mean than 100% of carbon losses are due to OWL losses.

Taking our estimate of 10 years of regrowth as an estimate of accumulated forest regrowth areas at the beginning of each decade to provide an estimate of annual carbon gains in forest regrowth, we obtain potential annual removals (sinks) of carbon from forest regrowth at 115 MtC yr⁻¹ (range: 61–168) and 97 MtC yr⁻¹ (53–141) for the 1990s and 2000s, respectively (Table ​(Table33).

Comparison of results of carbon losses and removals with other studies

Our estimates of carbon losses and removals can be compared to published estimates of carbon emissions for Brazil. From the 707 sample units of our study which cover the Brazilian territory, our estimates of annual CO₂ emissions for the 1990s are 1091 MtCO₂ yr⁻¹ (±424 MtCO₂ yr⁻¹ when considering min–max errors from Saatchi map) as potential long-term emissions. For the 2000s, we estimate 1137 MtCO₂ yr⁻¹ (±405 MtCO₂ yr⁻¹). These estimates can be compared to the figures of annual net anthropogenic CO₂ emissions from the Land-use Change and Forestry sector (Second National Communication of Brazil to the UNFCCC: Brazil, 2010) which give 761, 821, 1249, and 1251 MtCO₂ yr⁻¹ for the accounting years 1990, 1994, 2000, and 2005, respectively. These net figures include four different components: sources from ‘Forest and Grassland conversions’ and from ‘Emissions and Removals from Soils’ and sinks from ‘Changes in Forest and Other Woody Biomass Stocks’ and from ‘Abandonment of Managed Land’. In the first Brazilian communication (Cerri et al., 2009), sources from soils represent an average of 7% of the sources from land conversions when the sinks compensate for around 25% of the sources. This translates into national figures of sources from forest conversions at around 1100 MtCO₂ yr⁻¹ for the 1990s and 1400 MtCO2 yr⁻¹ for the early 2000s during which rates of deforestation in Brazilian Amazon where much higher than for the late 2000s. The resulting official estimate of annual CO₂ emissions from forest conversions for the 1990s (1100 MtCO₂ yr⁻¹) is in very good agreement with our estimate of annual CO₂ emissions at 1091 MtCO₂ yr⁻¹.

Our estimate of carbon removals from forest regrowth for Brazil is 124 MtCO₂ yr⁻¹ for the 1990s and 149 MtCO₂ yr⁻¹ for the 2000s. By considering that in the official Brazilian figures of net emissions carbon sinks compensate for around 25% of the carbon sources, this would lead to a national figure of sinks from the Land-use Change and Forestry sector at 235 MtCO₂ yr⁻¹ (average for years 1990, 1994, 2000, and 2005). This figure is respectively 48% and 38% higher than our estimates, but it includes sinks from ‘Changes in Forest and Other Woody Biomass Stocks’ and from ‘Abandonment of Managed Land’ (the later corresponding to forest regrowths). Unfortunately, there is no information available from Cerri et al. (2009) nor from the second National Communication of Brazil to the UNFCCC (Brazil, 2010) on the share between these two components.

We also compare our estimates of carbon losses during period 2000–2010 to two recent estimates of carbon emissions from tropical deforestation by Wood Hole Research Cente (WHRC) (Baccini et al., 2012) and WinRock International (Harris et al., 2012a,b) that, upon first review, seemed to differ widely. These studies cover only the first half of the 2000s, i.e. from year 2000 to year 2005 and use rates of deforestation which are different than ours (higher for South America and Africa and lower for Southeast Asia –Table 5). Our average continental estimates fall in between the estimates from the WHRC and WinRock studies for South America and Africa and at the level of the highest estimate (from WinRock) for Southeast Asia (Table ​(Table2).2). Our average estimates correspond well to FAO figures of forest biomass loss over the two decades (FAO, 2010). However, the FAO figures include also losses from forest degradation which are only partially considered in our study (only when deforestation is following degradation).

Compared to the most recent pan-tropical or global studies (Baccini et al., 2012; Harris et al., 2012a,b; Pan et al., 2011), we use spatial information on forest cover changes at a much higher spatial detail although limited to a sample of ca. 4000 sites (indeed only Harris et al., 2012a,b used a rigorous spatial approach combining deforestation and biomass maps at 1 km resolution followed by a detailed error propagation approach). Through our more detailed spatial approach, we provide a more direct and robust approach for producing estimates of carbon losses and removals from land-cover changes in tropics and we reconcile the significant regional differences in recent low estimates of carbon emissions (Table ​(Table55).

Recent estimates of removals from tropical regrowth from Pan et al. (2011) at 1.57 ± 0.50 and 1.72 ± 0.54 MtC yr⁻¹ for the 1990s and 2000s, respectively or gross uptake of carbon in the fallow cycle of shifting cultivation from Baccini et al. (2012) at 0.71 MtC yr⁻¹ for early 2000s fall outside our range of estimates at 0.12 ± 0.05 and 0.10 ± 0.04  MtC yr⁻¹ for the 1990s and 2000s, respectively (Table ​(Table3).3). Indeed, our global estimates of carbon removals from forest regrowth in the tropics are five to fifteen times lower than these earlier estimates. These discrepancies might be interpreted partly by removals from tropical forests that are recovering from logging which are not included in our study, but also partly by the use of a spatial approach with spatially more detailed values of biomass increment in our study compared to the use of a nonspatial book keeping model with national or continental values in these previous studies. Although our estimates of carbon removals from tropical regrowth are much lower than those of Pan et al., 2011, it may not imply large differences for the overall carbon fluxes budget. Indeed, the large potential overestimation of carbon removals by Pan et al. (2011) can be compensated in the net flux by a similar overestimation of carbon losses. In particular, the global estimate of tropical emissions from land-use changes for the 2000s from Pan et al. is much higher than ours (2.82 ± 0.45 and 0.88 ± 0.32 MtC yr⁻¹, respectively) and this difference cannot be only explained by the inclusion of emissions from forest degradation in Pan et al. estimate.

The vast majority of Landsat satellite data used in the study is coming from the GLS datasets (Global Land Survey) of the United States Geological Survey (USGS). Additional Landsat satellite imagery for years 1990 and 2000 were provided by ACRES (Australia), GISTDA (Thailand) and INPE. Complementary imagery (in particular for year 2010) was received from Astrium/French Development Agency (SPOT) and from the European Space Agency (AVNIR-2, Deimos-1, Kompsat, and RapidEye satellites) in the context of the TropForest project.

The authors would like to thank a large number of local forestry experts from the three continents for their support in reviewing the Landsat mapping results in the context of regional workshops (FAO, JRC, 2012).

The authors would also like to thank Javier Gallego of the JRC for his advices and contribution to the statistical aspects.

Finally, the authors would like to thank Scott Goetz of the Woods Hole Research Center for providing access to the full pan-tropical map of forest carbon density of Baccini et al. (2012) which is also available at http://www.whrc.org/mapping/pantropical/carbon_dataset.html. The carbon stock map of Saatchi et al. (2011) is available at http://carbon.jpl.nasa.gov.