For each sample location the available Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper (ETM) data were sought from the Global Land Survey (GLS) database, freely available from the Earth Resources Observation Systems (EROS) Data Center (EDC) of the US Geological Survey (USGS). The GLS data have the advantage over the previous GeoCover dataset of higher geo-location accuracy with terrain correction being carried out with an improved digital elevation data at a 90 m resolution from the Shuttle Radar Topography Mission (SRTM) [20]. The data were downloaded at full resolution (30 × 30 m). The 20 km × 20 km imagettes (in UTM projection) covering the sample sites were extracted and visually screened (see Figure 2(a)). In the event of the data being unacceptable (due to cloud cover or artifacts) replacement data were sought from the Brazilian Space Agency, INPE (Instituto Nacional de Pesquisas Espaciais) with the help of the GEOSS (Global Earth Observing System of Systems) Land Surface Imaging Constellation (LSI). For the 1,230 samples sites covering the study area, 1,222 good image pairs were found for 1990–2000 and 1130 good pairs for 2000–2005. The missing sites were located in Central America, the Colombian Choco, the Central Amazon region and in the Guianas. The higher number of missing sites in the second period assessed (2000–2005) is due to the malfunctioning of the line scanner on the Landsat 7 ETM+ sensor after 1 June 2003, which results in the corruption of around two thirds of the image, thus severely reducing the amount of exploitable imagery. A full review of the available images and acquisition dates is given in Beuchlé et al. [21].

For each sample site, the satellite imagettes (from 1990, 2000 and 2005) were preprocessed for geometric control, radiometric calibration and normalization, segmentation and finally classification. The steps in satellite image preprocessing are described in detail in Bodart et al. [22]. An automatic check was made on the geometric matching between images using image correlation. The data were corrected to top of atmosphere reflectance and then the images were radiometrically normalized to each other using the dark object subtraction, in this case evergreen forest [23]. In this way differences in reflectance of a given surface due to haze or calibration issues between image dates were reduced [24]. Where no evergreen forest was present in a site a relative normalization was used, with an operator identifying a suitable radiometrically stable area in the site. The imagette from the year 2000 was taken as the reference for geometric and radiometric controls.

After pre-processing the imagettes were segmented so as to produce homogenous land units (segments) which could then be classified for each date [24]. The segmentation process was initially implemented on two-date imagery (1990 and 2000) in a single operation so as to ensure a coherence of segments in the two dates. The calibrated TM and/or ETM+ bands 3, 4 and 5 were used together with equal weights for all bands. An initial minimum mapping unit of 1 ha was set for the segments. This was a compromise between not having a large number of very small segments and avoiding segments with mixed land covers, which would create problems for a classifier. The segments from the individual imagettes were then classified using a supervised classification. In a second phase, these classified segments were aggregated into 5 ha segments so as to enable a feasible visual assessment of the classification by local experts.

For the first level classification at 1 ha, membership functions for the supervised classifier were generated from a spectral library of over one hundred homogeneous spectral classes. These were obtained by an interpreter collecting spectral signatures for existing land cover types from a systematic sub-sample of the imagettes [24]. A large number of spectral classes are required to cover the variability of spectral reflectance within any particular land cover class, e.g., the “tree cover” class consists of 15 spectral classes ranging from coniferous forests to mangrove forests to deciduous forests in different phenological phases. The 1 ha size classified segments, in theory composing of 100% pure cover, were automatically aggregated to a 5 ha level (see Figure 2(b)) into a few broad land cover classes based on the proportion of tree cover within segments: (i) ‘Tree Cover’ class with more that 70% 1 ha level tree cover within the 5 ha segment; (ii) ‘Tree Cover Mosaic’ class with 30 to 70% 1 ha level tree cover within the 5 ha segment, and three other non-forest classes with less than 30 % tree cover; (iii) ‘Other Wooded Land’ (iv) ‘Water’ and (v) ‘Other Land Cover’. The term ‘Other Wooded Land’ is used for describing any woody vegetation layer (shrubs) of less than 5 m height. For segments with a tree cover portion of less than 30%, a minimum proportion (70%) of a specific land cover (e.g., ≥70% ‘Other Wooded Land’) was required, otherwise the segment was assigned to the ‘Other Land Cover’. The ‘Other Land Cover’ class groups together all non-woody vegetation land cover classes, with the exception of ‘Inland Water’.

Every single resulting classification from 1990–2000 was then inter-dependently visually controlled internally by the project staff. This was followed by an independent verification of these 1,222 change maps by national experts from forest and remote sensing institutes from the countries covered by the study. Where classification errors were found they were corrected. Once these two-date classifications were controlled and corrected the third date (2005) was segmented and classified using the year 2000 classification as a template for class labeling of year 2005. An inter-dependent visual control and correction was then made for the classifications of the 2000–2005 period.

It was not possible to find images corresponding to the exact reference dates for all sites. Images were acquired within a range from 1984 to 1992 for the first reference year (1990), 1997 to 2003 for the second reference year (2000) and 2004 to 2009 for the third reference year (2005) [21]. Each sample site’s area estimate was then adjusted to the baseline dates of 30 June 1990, 2000 and 2005: this was done by assuming that the land cover changes are constant during the given period (in the absence of some ancillary data at the regional level we cannot provide any other model). We therefore linearly adjusted the land cover change matrices to the reference dates/years. Cloudy areas within a sample site were considered as unbiased data loss, and were assumed to have the same proportions of land cover as the non-cloudy areas in the same site. This was achieved by converting the land cover change matrices for 1990–2000 and 2000–2005 to area proportions relative to the total land area of the sample site without clouds. It resulted in proportions matrices totaling 1 for the sum of their 5 × 5 elements.

For the missing sample sites (nine in the first period 1990–2000 and 110 in the second period 2000–2005) we used a local average from the surrounding 8 sample sites as surrogate results—with the hypothesis that surrounding sample sites are more representative of missing sample sites than the full available population. The following weights were applied for the local average of missing sites: (1) δ j j ′ = 1 d ( j , j ′ ) = 1 ( dif ( lat ) ) 4 + ( dif ( long ) ) 4where dif is the distance to the surrounding site involved. Small differences may appear between land cover proportions of year 2000 obtained from the two change matrices (1900–2000) and (2000–2005) due to the linear extrapolation to the reference dates. To correct for potential inconsistencies in estimates at regional level for the common year 2000, for each sample site the land cover proportion change matrix 2000–2005 was ‘calibrated’ to the land cover proportions of year 2000 from the 1990–2000 matrix by using a linear adjustment.

For the statistical estimation phase the sample sites were weighted in relation to their probability of selection. The sample frame, although systematic, does not have an equal probability of selection for each site due to the location of the confluence points. The circumference of the circle of latitude reduces in proportion to the cosine of the latitude; hence all sample units were given the weighting w_i, equal to 1 multiplied by the cosine of the latitude to account for the resulting higher spatial sampling frequency away from the equator. As the selected sample sites which contained a proportion of sea were considered as full sites (total of area proportions equal to 1), this compensated for those sample sites that contained a proportion of sea but were not selected as the center of the site—the confluence point—was located in the sea.

The proportions of land cover changes were then extrapolated to the study area using the Horvitz-Thompson Direct Expansion Estimator [25]. The estimator for each land cover class transition is the mean proportion of that change per sample unit, given by Equation (2): (2) y ¯ c = 1 m ∑ i = 1 n w i ⋅ y icwhere y_ic is the proportion of land cover change for a particular class transition ‘c’ in the i^th sample unit. The weight of the sample unit is w_i and m is the sum of the sample weights. The total area of change for this class transition Z_c is obtained from: (3) Z c = D ⋅ y ¯ cwhere D is the total area of the study region.

Rather than the usual variance estimation of the mean for systematic sampling [26] we used a local estimation of the variance as follows: (4) s 2 = ∑ j ≠ j ′ w j j ′ δ j j ′ ( y j − y j ′ ) 2 2 ∑ j ≠ j ′ w j j ′ δ j j ′where the weight w_jj′, is an average of the weights w_j and w_j′ and δ_jj′ is a decreasing function of the distance between j and j′ (Note that if we choose δ_jj′ = 1 j ≠ j′ we obtain the usual variance estimator). For w_jj′ we used a value of 1 (as a simplification because all the weights w_j are close to 1) and for δ_jj′ we applied Equation (1) also used for deriving local average of missing sites. The standard error (s.e.) is then calculated as: (5) se = s nwhere 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). The standard error represents the precision obtained with our sample scheme. The observations (source data sets) that are used to produce these results are derived from the satellite interpretations. These surrogates to ground observations may be subject to the uncertainty (bias) but we do not address such errors here [27]. Such errors are introduced mainly through misclassification where different land cover classes have similar spectral signatures. 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.

In our previous large-scale assessment using a similar visual inter-dependent interpretation method, we assessed the ‘consistency’ of the results of the interpretations [3]. The independent image-interpretation “consistency assessment” performed on a sub-sample of the data set resulted in some 93% and 96% agreement globally and for Latin America respectively for 10 aggregated land cover classes, and to some 91% and 96% agreement globally and for Latin America respectively for the forest cover change estimations. In addition to this earlier consistency assessment, we asked the relevant national institution, INPE, to independently visually interpret a random selection of 34 sample sites using our legend over the Brazilian legal Amazon. The comparison between INPE’s interpretations and our own shows a good correspondence both for the forest area of year 1990 and deforestation of period 1990–2000 with slopes close to 1 (1.017 and 1.008 respectively) and R² close to 1 (0.986 and 0.978 respectively) (Figure 3).

Due to the high investment in expert time required to visually control the interpretations we compared the land cover estimates and land cover change estimates derived from the 1990–2000 interpretations in 20 km by 20 km boxes with results using only the central 10 km by 10 km portion of the same interpretations. The close correspondence between these two estimates (Table 1) enabled us to process and analyse only the central 10 km by 10 km portion of the sample sites for the 2000–2005 period, resulting in a substantial savings in time and resources.

The selection of biomass levels and resulting carbon content for specific regions remains problematic due to lack of consistent data worldwide [28] and has been the source of some debate [29]. To estimate carbon fluxes from the changes in forest cover we created a spatially explicit data set by combining the FAO ecozone map with the Above Ground Biomass (AGB) map for tropical South America derived from maps of basal area and wood density extrapolated from data from over 200 forest plots [30]. The latter gives a gradation of AGB levels over the region, rather than a single figure as often used. The estimates from this map range from 200 to over 350 Mg·DW·ha⁻¹ AGB, and are on the whole lower than those estimates we used in our previous study. Spatially they show a fall in biomass from 350 t·ha⁻¹ in the Guinas and in the Central Amazonian basin to levels of 200 t·ha⁻¹ in the contact zones in the south and east and in the pre-montane forests of the Andes. The lower biomass in the ‘arc of deforestation’ is supported by recent direct measurements from [31,32]. Outside this area we associate AGB biomass levels for forests (‘Tree cover’) and shrubs (‘Other wooded land’) from the literature to each ecozone. We used IPCC [33] ratio of below-ground biomass to above-ground biomass for conversion to total biomass (ratio ranges from 0.37 for rainforest to 0.24 for sub-tropical humid forests) and we assumed 50% of biomass to be carbon. We did not account for losses of carbon in soils. The annual carbon emissions for the study area was calculated as the gross committed emissions over a 10 year period arising from one year of forest and woodland clearance [34]. This accounts for 69% of the total carbon that could be committed [35].

We found that for the year 1990 there were 763 Mha (s.e. 10 Mha) of forest (Table 2(a): tree cover + 50% tree cover mosaic) in our study area and 191 Mha (s.e. 5.4 Mha) of other wooded land. By 2000 the forest area had fallen to 733 Mha (s.e. 10 Mha) with 189 Mha (s.e. 5.5 Mha) of other wooded land and by 2005 715 Mha (s.e. 10 Mha) of forests and 184 Mha (s.e. 5.5 Mha) of other wooded land. Comparisons with other estimates (Table 3) show a high variation between the different studies. Our forest area estimate is substantially less than the most recent FAO estimate for year 2000 of 877 Mha derived from the country survey of the Global Forest Resources Assessment 2010 (FRA 2010) [6], but FAO is using a land use definition of forests with 10% tree cover threshold. The limitations of the FAO country surveys are known, in particular due the lack of reliable figures for some countries and the difficulty to harmonize national definitions and estimates [1]. FAO reports forest cover for Brazil for year 2000 of 546 Mha which corresponds well to the sum of our year 2000 forest and other wooded land estimates for Brazil (derived from 808 sample units) of 542 Mha (406 Mha of forests and 136 Mha of other wooded land). The land cover map of South America for the year 2000 [36] gives a higher figure of 786 Mha for the forest area, but this map is also based on coarse resolution, 1 km, satellite imagery which are not considered appropriate to estimate with accuracy land cover areas and their changes [17]. Our previous study, TREES-II [3], reported an estimate of 647 Mha for tropical humid forests of Central and South America which would lead to a forest cover estimate of 727 Mha when corrected for Mexico and dry tropical forests of South America. This figure (727 Mha) is very close to our current study estimate (733 Mha). Our forest area estimate for year 2000 lies between the FAO estimate from the previous remote sensing survey of FRA 2000 (based on a lower intensity sample of Landsat satellite imagery) of 700 Mha [37] when adjusted for Mexico and the estimate from another remote sensing based study (based on coarse resolution—500 m-satellite imagery) at 814 Mha [4] when adjusted for the countries outside our study (Argentina, Chile and Uruguay). The circa 10% difference between our estimates (TREES-II and present study at 727 and 733 Mha) and the two other remote sensing based estimates for the same region (at 700 and 814 Mha) points to the remaining difficulties in comparing estimates from different data sources, methodologies and more importantly with different forest definitions. This comparison at regional scales indicates that a large proportion of our other wooded land class is considered as ‘forest’ by the other remote sensing based surveys.

To assess if our results are consistent with other assessments, we compared our forest area estimate for the Brazilian Legal Amazon region (derived from the 411 sample units covering this region) to an independent estimate from a wall to wall assessment using the same type of imagery. The Landsat Pathfinder forest estimate of Brazilian Amazonia of year 1988 [38] corrected to year 1990 leads to a rain forest area at 361.3 Mha which is 7.2% lower than the estimate of our study at 389.3 Mha (s.e. 5.6 Mha). However the Landsat Pathfinder estimate does not include the dry forests of Brazilian Amazonia (cerrado) which are considered to represent circa 10 percent of the total forest area [39]. Our previous TREES-II [3] led an estimate of tropical humid forests for Brazilian Amazonia at 381.9 Mha for 1990, a 1.9% difference with our new 1990 estimate. We consider therefore our present study estimate as a consistent estimate of total forest area of Brazilian Amazonia.

Land cover change matrices were produced at three geographical levels: for the whole study area (Table 2), for Brazil and for the Brazilian Legal Amazon and again compare them to other studies (Table 3). Between 1990 and 2000 there was a net loss in the whole study area of 2.98 Mha·y⁻¹ (s.e. 0.26 Mha·y⁻¹) of forests rising to 3.74 Mha·y⁻¹ (s.e. 0.39 Mha·y⁻¹) in the 2000 to 2005 period (Table 2b). The gross loss of forests was rising from 3.74 Mha·y⁻¹ (s.e. 0.22 Mha·y⁻¹) in the 1990s (0.50% annually) to 4.40 Mha·y⁻¹ (s.e. 0.32 Mha·y⁻¹) in the 2000 to 2005 period (0.61% annually), with Brazil accounting for 69% of the total. The increase in the gross loss of other wooded land was more striking with a 45% increase between the two periods in the annual gross rate from 2.0 Mha·y⁻¹ (s.e. 0.11 Mha·y⁻¹) in the 1990s to 2.9 Mha·y⁻¹ (s.e. 0.37 Mha·y⁻¹) in the early 2000s. In both periods circa 60% of the cleared forests were converted to other land cover (this proportion is circa 68% for Brazilian Legal Amazon). The main area of forest loss in both periods was in the so called ‘arc of deforestation’ in the Brazilian Amazon (55% and 58% of the total for the 1990s and early 2000s respectively) (Figure 4), whereas changes in other wooded land were geographically more widespread, with high losses occurring in the Brazilian states of Mato Grosso and Tocantins. Large non-forest areas were also re-occupied by forests. Forest regrowth (defined as the gross gain of forests) decreased slightly from 0.65 Mha·y⁻¹ (s.e. 0.08 Mha·y⁻¹) in the 1990s to 0.46 Mha·y⁻¹ (s.e. 0.04 Mha·y⁻¹) in the early 2000s. During the first period the continent was losing more than 5.7 Mha·y⁻¹ of forests and other wooded lands a year (gross loss) rising to over 7.3 Mha·y⁻¹ in the second period, while re-colonization of these areas fell from 2.6 Mha·y⁻¹ to 2.5 Mha·y⁻¹ between the periods. Our gross annual deforestation rate for the period 2000–2005 corresponds closely to those of South Dakota State University (SDSU) for both Brazil and for the continent (Table 4). Both our study and that of SDSU give higher figures (from 12% to 18%) than the official ones provided by the Brazilian National Space Agency (INPE-Instituto Nacional De Pesquisas Espaciais) for the Legal Amazon [12]. These differences seem to be due to cerrado clearance which is not accounted for in INPE’s official figures [39].

The spatial distribution of forest change is highly concentrated in the study area, (Figure 4), highlighting the need for spatially explicit data on biomass.

We calculated the emissions resulting from changes in forest and other wooded land at three geographical levels, for the whole study area (Figure 5), for Brazil and for the Brazilian Legal Amazon and compare them to other estimates (Table 5). The annual carbon flux (10 year committed emissions) for the whole study area results in 482 MtC·y⁻¹ for the 1990s is close to our previous estimate at 519 MtC·y⁻¹, and rises to 583 MtC·y⁻¹ for the first half of the 2000s, reflecting the increase in deforestation in the latter period. For Brazil alone our estimate of the annual carbon flux for the 1990s is 351 MtC·y⁻¹ (888 MtCO₂·y⁻¹) as 10 year committed emissions or 508 MtC·y⁻¹ (1287 MtCO₂·y⁻¹) as potential long-term emissions (i.e., if all carbon from one year of deforestation and woodland clearance would go into the atmosphere). Similarly for the 2000–2005 period our estimates of emissions are 423 MtC·y⁻¹ (1,070 MtCO₂·y⁻¹) as 10 year committed emissions or 613 MtC·y⁻¹ (1,551 MtCO₂·y⁻¹) as potential long-term emissions. Our results can be compared to figures reported by Brazil to the UNFCC [40,41]: the annual net anthropogenic CO₂ emissions from the Land-use Change and Forestry sector are reported in Brazil’s second communication at 761, 821, 1,249 and 1,251 MtCO₂·y⁻¹ for the accounting years 1990, 1994, 2000 and 2005 respectively [40]. Brazil’s first national communication provided earlier estimates giving details of 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” [41]. For the year 1990 the sources were reported as 882 MtCO₂·y⁻¹ and 110 MtCO₂·y⁻¹ from Land Use conversions and from Soils respectively and the total sinks were reported at 234 MtCO₂·y⁻¹, when for the year 2005 the sources were 1026 MtCO₂·y⁻¹ and 65 MtCO₂·y⁻¹ and the total sinks were 230 MtCO₂·y⁻¹. If we consider our estimate of forest regrowth areas at the end of the period 1990–2000 (i.e., total of areas which have been reconverted to forests during the 10 year period) at 7.6 Mha (s.e. 0.67 Mha), and by using the rates of annual carbon gains in secondary vegetation used by Brazil in their national report (6.2 tC·ha⁻¹ for forest with biomass higher than 127 tC·ha⁻¹ and 5.1 tC·ha⁻¹ for less than 127 tC·ha⁻¹) we obtain a yearly sink of 168 MtCO₂·y⁻¹ for these regrowing forests accumulated during that 10-year period for all tropical South America. If we make the hypothesis that there have been similar rates of re-conversion to forests during the 1975–1990 period, a value of 336 MtCO₂·y⁻¹ corresponding to sinks from 20 years of accumulated regrowth can be considered as a conservative (i.e., maximum) estimate for annual removals during the study period of carbon from forest regrowth. The Brazilian estimate of 240 MtCO₂·y⁻¹ (average of years 1990, 1994 and 2005) which includes two components: sinks from changes in forest biomass stocks and sinks from abandonment of managed land, corresponds to 70% of this.

The national figures reported by Brazil in the second communication to the UNFCCC correspond to net emissions of carbon emissions, i.e., with removals of carbon in managed areas and carbon gains in secondary vegetation areas, which explains why such figures are lower than our potential long-term emission estimates. However, the different contributions-sources from deforestation, removals in managed areas and in secondary vegetation—are not reported individually by Brazil. The Brazilian national figures include also carbon emissions from soils but the first national communication shows that emissions from soils represent a limited contribution, i.e., circa 7% of the emissions from land use conversions. On the other hand the Brazilian report does not seem to consider a decay model for distributing emissions over time, i.e., emissions factors are used as yearly direct emissions. This explains why the Brazilian estimates of net emissions are higher than our 10-year committed emission estimates.

The consequences of these land cover changes are multiple. The comparison of our results for emissions with the Brazilian national figures shows that the potential gross emissions from forest and woodland clearance are much higher than the net emissions reported in national reports. Our results show the importance of emissions from forest and woodland clearance in tropical America and the need to consider carefully what is accounted for when reporting emissions and removals from anthropogenic processes.

The more recent study on the role of the world’s forest in the carbon cycle [2] reports gross deforestation emission in South America at 1.51 and 1.37 Pg·C·y⁻¹ for the 1990s and 2000s (2000–2007 period) respectively which are around twice as high as our estimates of potential long-term emissions (i.e., full carbon loss) at 0.70 and 0.84 Pg·C·y⁻¹ for the 1990s and early 2000s (2000–2005 period) respectively, and circa three times higher than our estimates of 10 year committed emissions at 0.48 and 0.58 Pg·C·y⁻¹ for the 1990s and early 2000s (2000–2005 period) respectively. These high estimates for gross carbon emissions from tropical deforestation are based on FAO national statistics with their known limitations [1]. The same study also reports on sinks from tropical regrowth in South America giving 0.81 and 0.83 Pg·C·y⁻¹ for the 1990s and 2000s which compare to the Brazilian estimate from first communication at 0.065 Pg·C·y⁻¹ (average of years 1990, 1994, 2000 and 2005) and our own estimate for tropical South America at 0.092 Pg·C·y⁻¹ for the 1990s. If Brazil were to use the same rate of annual carbon gains in secondary vegetation as the global study (i.e., 3.8 tC·ha⁻¹) the Brazilian estimate of sinks would be even smaller. This indicates another high discrepancy between the results of [2] and the estimates from robust bottom-up approaches. Our consistent approach based on remote sensing for forest (and other woodland ecosystems) area and forest area change estimation combined with spatial information on the distribution of biomass in forests provide a more direct and robust approach for producing estimates of carbon emissions from land cover changes in tropical South America.