Long-term, high-spatial resolution carbon balance monitoring of the Amazonian frontier: Predisturbance and postdisturbance carbon emissions and uptake


Corresponding author: M. Toomey, Department of Geography, University of California Santa Barbara, 1832 Ellison Hall, Santa Barbara, CA 93106, USA. (mtoomey@fas.harvard.edu)


[1] We performed high-spatial and high-temporal resolution modeling of carbon stocks and fluxes in the state of Rondônia, Brazil for the period 1985–2009, using annual Landsat-derived land cover classifications and a modified bookkeeping modeling approach. According to these results, Rondônia contributed 3.5–4% of pantropical humid forest deforestation emissions over this period. Similar to well-known figures reported by the Brazilian Space Agency, we found a decline in deforestation rates since 2006. However, we estimate a lesser decrease, with deforestation rates continuing at levels similar to the early 2000s. Forest carbon stocks declined at an annual rate of 1.51%; emissions from postdisturbance land use nearly equaled those of the initial deforestation events. Carbon uptake by secondary forest was negligible due to limited spatial extent and high turnover rates. Net carbon emissions represented 93% of initial forest carbon stocks, due in part to repeated slash and pasture burnings and secondary forest clearing. We analyzed potential error incurred when spatially aggregating land cover by comparing results based on coarser-resolution (250 m) and full-resolution land cover products. At the coarser resolution, more than 90% of deforestation and secondary forest would be unresolvable, assuming that a 50% change threshold is necessary for detection. Therefore, we strongly suggest the use of Landsat-scale (~30m) resolution carbon monitoring in tropical regions dominated by nonmechanized, smallholder land use change.

1 Introduction

[2] Tropical deforestation exerts considerable influence on the atmosphere, hydrosphere, and biosphere at a variety of scales [Bala et al., 2007; Wang et al., 2009]. Recent international policy discussions have focused on the potential to mitigate global climate change by reducing tropical deforestation, which is estimated to release nearly 1 Pg of carbon annually to the atmosphere or roughly 11% as much as that emitted by fossil fuel combustion [Harris et al., 2012; Baccini et al., 2012]. The proposed Reducing Emissions from Deforestation and Forest Degradation in Developing Countries (REDD+) UNFCCC, 2007; UN-REDD Programme, 2010] system would provide incentives for reducing emissions from deforestation and forest carbon, thus requiring accurate monitoring of those emissions. In this study, we consider the appropriate spatial scale for monitoring deforestation and subsequent land use change and their effects on carbon balance and emissions.

[3] Rates of tropical land cover change in recent decades were higher than those in any other part of the world [FAO, 2010], serving as the primary source of new agricultural lands during the 1980s and 1990s [Gibbs et al., 2010]. According to a pantropical study of deforestation “hotspots,” emissions from humid forests were 0.64 PgC yr−1, or 0.96 PgC yr−1 with dry tropical forests included [Achard et al., 2002]. Hansen et al. [2008a] performed a similar global study of deforestation hotspots for 2000–2005, reporting a loss of 27.2 Mha, similar to the rate in the 1990s, constituting a 1.39% decrease in humid forest cover. Deforestation in Brazil alone accounted for 47% of all clearing during the study period, with a rate of 26 × 103 km2 yr−1. Although Brazil's humid forest cover constitutes nearly 30% of the global total, the per-area forest loss, 3.6%, was the highest of any region during that period [Hansen et al., 2008a]. Indeed, as of 2006, 750,000 km2 of greater Amazonia have been deforested, 80% of which has taken place in Brazil [Soares-Filho et al., 2006].

[4] Although considerable advances in forest carbon monitoring have been made in the past decade, a number of logistical and methodological challenges limit our knowledge of the spatial and temporal dynamics of tropical deforestation and its consequences. Two primary challenges confront researchers: accurate, spatially explicit quantification of forest loss and proper characterization of carbon emissions resulting from land use and land cover change (LULCC). Satellite-based estimates of tropical forest cover are greatly inhibited by cloud cover [Asner, 2001], the cost of certain imagery products, and computational intensity of data processing at regional and global scales. Several compromises are available: (1) statistical sampling, (2) use of coarse-resolution satellite observations, and (3) decreased temporal sampling. Statistical sampling techniques [e.g., Achard et al., 2002; Hansen et al., 2008a] reduce sampling intensity by utilizing high-resolution imagery such as Landsat (30 m resolution) to focus on deforestation hotspots, producing biased estimates [Grainger, 2008]. By contrast, most large-scale studies of LULCC in tropical forests have relied on coarse-scale data (>250 m resolution) from satellites such as the Moderate Resolution Imaging Spectrometer (MODIS) or the Advanced Very High Resolution Radiometer (AVHRR) [e.g., Defries et al., 2002; Cardille and Foley, 2003; Hansen et al., 2008b]. Such coarse pixels may be too large to detect secondary growth [Alves et al., 2003; Ramankutty et al., 2007] and household-scale changes [Grainger, 2008], especially within already disturbed landscapes and forest fragments [Alves et al., 2003]. A third class of studies is characterized by extended periods between observations [Alves et al., 2003] masking interannual variability.

[5] The second challenge confronting researchers of tropical deforestation is the quantification of carbon stocks and emissions from LULCC. There is considerable uncertainty about the carbon stocks of tropical forests [Gibbs et al., 2007], even for the relatively well-studied Brazilian Amazon [Houghton et al., 2001; Kauffman et al., 2009]. There is also considerable uncertainty in carbon fluxes due to diverse land use trajectories and associated carbon-tracking schemes employed by various researchers [Ramankutty et al., 2007]. In particular, strategies differ greatly for tracking carbon after deforestation, including the decay of remaining slash and clearing of secondary forest.

[6] Despite the above challenges, the need for accurate, efficient, and repeatable carbon-monitoring schemes has become essential to the implementation of REDD+ [UNFCCC, 2007]. REDD+ would create financial incentives for forest-rich developing countries to protect existing forests from deforestation and degradation and to promote conservation, sustainable management, and enhancement of forest carbon stocks. While the architecture of an international REDD+ system is still being negotiated under the UNFCCC, Brazil already has numerous subnational projects and a national fund that channels international funding for REDD+ (the Amazon Fund). All levels of REDD+ require remote sensing-based monitoring, reporting, and verification (MRV) schemes by which recipients assess and openly disclose their own forest carbon stocks [Angelsen et al., 2009]. To assign credit and liability under a carbon market or national REDD+ policy, carbon stocks and emissions must be tracked at local and subnational levels. The integrity of this system will depend on having data of sufficient temporal and spatial resolution.

[7] In this study, we use high-spatial (28.5 m) and high-temporal (annual) resolution land cover data sets to analyze the long-term carbon balance and annual dynamics of an old Amazon frontier region in Rondônia, Brazil. Further, we examine the role of spatial scale in tracking LULCC and the implications for monitoring forest carbon, evaluating the effectiveness of coarse-scale imagery. The following questions guide this study: (1) What were the rates, trends, and interannual variability in the carbon stocks and fluxes of the Amazonian frontier in recent decades? (2) What was the spatial scale of LULCC change throughout that time and what are the implications for tropical forest carbon monitoring?

[8] To address these questions, we use the open-access Landsat satellite data archive along with a forest biomass map and carefully prescribed parameters of land use and the fate of carbon to derive accurate estimates of carbon emissions and remaining stocks. Additionally, we aim for a spatially comprehensive assessment by implementing the wall-to-wall approach suggested by Grainger [2008].

2 Site Description

[9] The Brazilian State of Rondônia (238,000 km2) serves as a microcosm of Amazonian development, with large tracts of established settlements along state and national highways juxtaposed with more recent and active deforestation activity near the edges [Roberts et al., 2002]. Mean rainfall in the state from 1987 to 2008 ranged from 1876 to 2159 mm, and mean temperatures were 24.3–25.7°C, with drier conditions in the south [World Data Center for Meteorology, 2010]. Accordingly, vegetation grades from upland and submontane moist forest in the north to semideciduous seasonal forest in the southeast. Lenses of savannah and forest-savannah ecotones are found throughout the state; savannah predominates in the extreme southeast (Figure 1). As of 2010, 39% of the state's mature forest had been cleared, constituting 13% of all deforestation in the Brazilian Amazon [INPE, 2011]. At present, the only substantive tracts of forests remaining are found in federally protected and indigenous areas.

Figure 1.

Maps of the modeled land cover classes for (left) 1984 and (right) 2010 with inset to highlight secondary forest. The black outline delineates the Rondônia state boundary.

[10] Due to its long settlement history, high deforestation rates, and accessible civil infrastructure, Rondônia has been the subject of a number of remote sensing-based LULCC studies using both coarse-scale [Malingreau and Tucker, 1988; Cardille and Foley, 2003] and fine-scale [Fearnside and Salati, 1985; Alves et al., 1999; Alves and Skole, 1996; Alves et al., 2003; Brown et al., 2005] imagery, all of which documented the rapid deforestation rate. Each of these studies was limited spatially and/or temporally, however, preventing the analysis of quickly changing and/or small-scale phenomena. Roberts et al. [2002] made a significant step towards temporal continuity, analyzing long-term (1975–1999) land cover changes for three Landsat scenes in central Rondônia, with an average interval of 2.18 years between images.

3 Methods

3.1 Landsat Time Series Data and Processing

[11] In this study, we used a continuous 27 year annual land cover history developed from seven Landsat scenes, covering 190,976 km2—approximately 5% of the Brazilian Legal Amazon (or 3% of greater Amazonia) at 28.5 m resolution. This region comprises 67.8% of the total state area and 83.9% of the deforested area [INPE, 2011]. This data set was comprised entirely of Landsat Thematic Mapper and Enhanced Thematic Mapper data with the exception of four images from 1986 which were acquired by the Multispectral Spectral Scanner. Images were initially selected based on minimal cloud cover (all selected images were <15% cloud cover) for a given year. Ten percent of the 189 images showed clear evidence of smoke and haze from biomass burning, necessitating corrections of the visible bands [Carlotto, 1999]. We implemented an eight-class land cover scheme: primary forest, secondary forest, pasture, green pasture, soil/urban, water, cloud, and a manually produced rock/savannah mask. Classified maps were produced via a standardized methodology established by Roberts et al. [2002], including georeferencing to Geocover base imagery, radiometric intercalibration using invariant targets, spectral mixture analysis, decision tree classification, and multitemporal correction filters.

[12] Enhancements to the above classification products have been incorporated to address two common sources of inaccuracy in the classified product. First, to resolve confusion between early dry season pasture and secondary forest [Roberts et al., 2002], we used MODIS-derived pasture phenologies to identify growing season dates; these dates were applied in a multitemporal correction filter to reclassify early season pasture erroneously classified as secondary growth (Roberts et al., in preparation, 2013). Second, misclassification of primary forest on sunlit slopes as secondary forest [Roberts et al., 2002] was resolved using visual analysis and manual removal (Roberts et al., in preparation, 2013). Overall accuracy of the new classifications, assessed using recent Google Earth imagery (2008–2010) to identify ~400–500 reference pixels per class, was 90.5% with a Kappa coefficient of 0.887. The producer's and user's accuracy for the five principal land cover classes of interest (excluding clouds, the rock/savannah mask, and green pasture, which is an ephemeral form of pasture) was above 80%.

[13] In this study, the land uses under consideration were primary forest, secondary forest, and pasture. The primary and secondary forest land covers were treated solely as their respective land covers. The pasture land cover requires some reassignment: green pasture, soil, and burn classes are ephemeral forms of pasture and were reclassified as such. Several masks were applied. We masked out savannah that was not identified in the manual mask [Roberts et al., 2002] using RADAMBRASIL geographic layers [RADAMBRASIL, 1978]. Urban areas, as delineated by the Rondônia State Secretary for Environmental Development, were also masked out. Because our focus was on aboveground biomass and carbon, the water class, 0.6% of land cover on average, was masked out. With all masks applied, the area under study was 164,041 km2. Any remaining instances of the cloud class, which had not been reassigned by the multitemporal filters, were reassigned to the class from the preceding year. Our land cover classification scheme does not include a crop class due to the difficulties in separating secondary growth and crops and the limited extent in Rondônia [Caviglia-Harris and Harris, 2008]. An exception is the southeastern portion of the state where mechanized soy production dominates land use [Brown et al., 2005]; however, this land use differs little from pasture in terms of carbon dynamics [Kauffman et al., 2009]. Due to the central role that interannual corrections play in our classification protocol, the first (1984) and final (2010) years are not as accurate as intermediate years. Therefore, results will focus on the years 1985–2009.

3.2 Carbon Stocks and Flows

[14] We used a modified “bookkeeping” method of tracking the fate of carbon during and following deforestation [Houghton et al., 2000; Achard et al., 2002; DeFries et al., 2002; Ramankutty et al., 2007; Loarie et al., 2009; Numata et al., 2010]. The empirical relationships established for this methodology are based upon field data collected near Manaus [Fearnside et al., 1993; Carvalho et al., 1998] in the central Amazon and may differ substantively from land use practices, settlement history, soil fertility, and carbon balance of the southern Amazon. The modified bookkeeping method incorporates empirical models of primary forest biomass losses and secondary forest biomass accumulation based on field data from northern and central Rondônia [Alves et al., 1997; Guild et al., 1998]. Additionally, we account for pasture land use, which is not considered in the Houghton et al. [2000] model, based on data from Guild et al. [1998], Numata et al. [2003], and Kauffman et al. [2009]. All emission figures are reported in petagrams of carbon (PgC), assuming that 50% of biomass is carbon [Fearnside et al., 1993].

[15] In summary, the following assumptions were made in our bookkeeping model (please see sections 3.2.1 to 3.2.3 for details and sources):

  1. [16] Fifty-one percent of the original forest carbon is emitted during the initial burning, 39% is left as slash, 8% is used as timber products, and 2% is converted to elemental carbon. The slash pool has a combustion efficiency of 38% during two pasture burnings in the first 10 years, after which it decays at 10% yr−1; elemental carbon is assumed to decay at a rate of 0.1% yr−1, and wood products removed from site decay at 10% yr−1.

  2. [17] Pasture accumulates carbon as biomass with a density of 8.9 Mg ha−1 and is burnt, with 100% combustion efficiency, twice during the first 10 years: once 3 years after deforestation and a second time 6 years after forest cutting.

  3. [18] Secondary forest accumulates biomass at 7.41 Mg ha−1 yr−1. When secondary forest is burned and cleared, we assume a combustion efficiency of 59%. The remaining slash, 41% of the preburn biomass, decays at the rate of 10% yr−1.

3.2.1 Primary Forest Carbon Stocks

[19] We used a forest biomass map for the Brazilian Legal Amazon derived by enhanced kriging interpolation of field timber volume estimates [Sales et al., 2007; Sales, 2010]. Rondônia forest biomass ranged from 140 to 323 Mg/ha (2.5th and 97.5th quantiles) with a mean and median of 274 and 271 Mg ha−1, respectively. These figures match the mean biomass estimate of 271 Mg ha−1 derived from a survey of 44 Brazilian Amazonian sites [Houghton et al., 2001] and are comparable to more recent estimates for Neotropical evergreen broadleaf forest of 260 Mg ha−1 [Baccini et al., 2012]. Our estimates are higher, however, than the Brazilian mean biomass density estimate of 204 Mg ha−1 of Saatchi et al. [2011]. Caution should be taken when comparing with the latter two studies as they are based on syntheses of global-scale passive and active remote sensing data, calibrated by plot data, whereas ours are based solely on plot biomass estimates. Biomass was highest in the north, declining southward, paralleling the precipitation gradient. The lower biomass ranges are located in the semideciduous forest and woodlands in the far southeast. The Legal Amazon forest biomass map was a subset to the spatial domain of the land cover maps and resampled from its native 1 km resolution to 28.5 m using linear interpolation. All pixels classified as primary forest were assigned the value from the corresponding biomass map pixel to create the initial state.

3.2.2 Deforestation-Associated Carbon Losses

[20] Guild et al. [1998] reported a combustion efficiency of 47% and 54% for two fires conducted by private landowners in northern Rondônia—the mean of 51% is considerably higher than the 20% of Houghton et al. [2000]. Additionally, the Houghton et al. [2000] bookkeeping model assumes that dead wood left behind decays at a rate of 10% yr−1. In Rondônia, however, newly deforested areas are typically burned twice [Numata et al., 2003], and during each fire, 38% of the residual wood biomass is lost [Kauffman et al., 2009]. Therefore, we modeled forest biomass as follows: 51% of the original forest biomass is emitted during the initial burning, 39% is left as slash, 8% is used as timber products, and 2% is converted to elemental carbon. The slash pool has a combustion efficiency of 38% during two pasture burnings during the first 10 years, after which it decays at 10% yr−1; elemental carbon is assumed to decay at a rate of 0.1% yr−1, and wood products removed from site decay at 10% yr−1

[21] There is considerable disagreement on the effects of LULCC on belowground carbon in the tropics. Estimates differ in terms of the magnitude and direction based on management, climate, soil type, disturbance type, and sampling depth of the study [Holmes et al., 2006; Kauffman et al., 2009; Don et al., 2011]. Hence, following similar studies, we have not considered the effect of LULCC on soil carbon.

3.2.3 Pasture and Secondary Forest Carbon Dynamics

[22] Pasture management plays an important role in the carbon cycle of deforested land [Guild et al., 1998; Kauffman et al., 1998]. In central Rondonia, most landowners use two fires after the initial slash fire; smallholders do not use fire on pasture older than 10 years [Numata et al., 2003]. Combustion efficiency of pasture grass is essentially complete [Guild et al., 1998]. In this model, live pasture biomass is assumed to be 8.9 Mg ha−1 [Kauffman et al., 2009], similar to 10.7 Mg ha−1 assumed by Fearnside [1996]; pasture growth represents carbon uptake from the atmosphere (Figure 2). Pasture is assumed to undergo two burnings with 100% combustion efficiency for the first 10 years, once 3 years following deforestation and a second time 3 years later.

Figure 2.

Land use carbon model. See section 3.2 for rates of carbon emissions/uptake and frequency intervals. Circles represent pools, and boxes represent land cover changes. DF emis. = deforestation emissions and LU emis. = post-disturbance land use emissions.

[23] Abandoned pasture and shifting cultivation yield rapidly growing secondary forest. Secondary forests exhibit a rapid and linear rise in biomass during the first 15–20 years [Brown and Lugo, 1990; Alves et al., 1997; Zarin et al., 2001], after which there is considerable spread in the data [Brown and Lugo, 1990]. This study uses a chronosequence (n = 8) of secondary forest plots in central Rondônia (10.25°S, 62.87°W) from Alves et al. [1997] to establish rates appropriate for the state. Secondary forest biomass in Rondônia is very strongly correlated with stand age (R2 = 0.893, p < 0.001, n = 8). We calculated a linear regression model of biomass accumulation, forced through zero. The slope value (7.41 Mg ha−1 yr−1) shows good agreement with those of Brown and Lugo [1990] (6.25 Mg ha−1 yr−1), Zarin et al. [2001] (7.06 Mg ha−1 yr−1), and Helmer et al. [2009] (8.4 Mg ha−1 yr−1). As with pasture, secondary forest growth represents carbon uptake from the atmosphere (Figure 2).

[24] Secondary forests are frequently burned and cleared in Amazonia [Steininger, 2004]—a practice commonly unaccounted for in tropical carbon balance studies [Ramankutty et al., 2007]. Combustion efficiency of secondary forest is 59% [Hughes et al., 2000], somewhat higher than that of primary forest due to the prevalence of small-diameter, younger trees [Kauffman et al., 2009]. The remaining slash, 41%, decays at the rate of 10% yr−1 [Houghton et al., 2000].

[25] At the beginning of the study period, 1984, 1.5% of land cover was classified as secondary forest, and therefore, its age could not be determined. Hence, to assign ages to these secondary forest fragments, we tracked secondary forest persistence for the first 10 years of the time series (1984–1994) to construct a frequency histogram of new secondary forest (i.e., not secondary forest in 1984) ages. Secondary forest tracking was limited to this time period to yield secondary forest ages consistent with the settlement history [Roberts et al., 2002]. Pixels classified as secondary forest in 1984 were assigned ages by random sampling from this distribution, and their corresponding biomass was calculated accordingly using the biomass accumulation rate of 7.41 Mg ha−1 yr−1.

3.2.4 Analyzing the Spatial Scale of LULCC

[26] The spatial scale of LULCC throughout the time series was analyzed by recording and segmenting coded land cover transitions. As we were primarily concerned with the two dynamic carbon fluxes, deforestation and secondary forest regrowth, we analyzed the spatial scale of these two phenomena. Managed pasture, by contrast, neither gains nor loses significant amounts of carbon (unmanaged pasture, however, often transitions to carbon-accumulating secondary forest). For each pair of years (e.g., 1985–1986) in the land cover time series, deforestation “events” (discrete pixel clusters undergoing a primary forest-pasture or primary-secondary forest transition) were flagged and assigned a unique code in a separate binary image. Each output image was converted to a polygon layer which was used to produce a frequency histogram of sizes of land cover transitions for each year. The same process was followed for secondary forest fragments for each year in the classification time series.

[27] We also assessed the ability of coarse-resolution remote sensing instruments to detect deforestation and secondary forest by resampling the deforestation and secondary forest flags to 250 m resolution (the resolution of MODIS red and near-infrared bands used in Surface Reflectance and Vegetation Index data) and calculating their respective proportions of each grid cell. To analyze the spatial scale of LULCC and emissions, we flagged all deforestation events and secondary forest fragments observed throughout the time series. Despite the rigorous interannual corrections that were applied in the classification process, speckle and misregistration errors could still influence the accuracy of the remote sensing products and this analysis. Therefore, we discarded all deforestation events with four or fewer pixels—this threshold is applicable to most visually identified speckle and is equivalent to the number of pixels that would be masked out in a typical majority filter analysis using a 3 × 3 window. We also excluded the latter 3 years (2006–2009), which may have been inadequately corrected with our interannual correction filters (Figure 3).

Figure 3.

Land cover trends from 1985 to 2009 for primary forest, pasture, and secondary forest.

4 Results

4.1 Land Cover Trends and Deforestation Rates

[28] Forest was by far the most prominent class in 1985 with 93.0% cover; pasture and secondary forest covered only 5.1% and 0.8%, respectively (Figure 3). These proportions changed quickly, with pasture supplanting forest throughout the 25 year period, at the end of which primary forest covered 55.2% and pasture covered 38.4% of the study area. The coefficient of determination (R2) between annual primary forest and pasture cover (0.994) was much stronger than that between primary and secondary forest cover (0.549), illustrating the dominant role of pasture as the postdisturbance land use. There was a general increasing trend in secondary forest even though cover remained quite low, 0.8–2.5%, during the first 22 years (Figure 3). There is a slight upward inflection point in secondary forest during the final 3 years, which is likely an artifact of the interannual corrections. Of the original 154,526 km2, 63,976 km2 was deforested by 2009. Deforestation rates exhibited a slight increasing trend when evaluated as a function of total land cover and more so as a percentage of remaining forest cover (Figure 4). Mean rates were 1.50% and 2.03%, respectively; annual rates ranged from 0.92% to 2.4% and from 1.06% to 3.96%, respectively. By both measures, deforestation rates were highly variable on a year-to-year basis with some of the highest rates during the early 1990s (1.8–2.0% of total forest area) as well as the early 2000s (1.7–2.44%). Similar to the Brazilian Space Agency's deforestation monitoring scheme (Projeto de Monitoramento do Desflorestamento na Amazônia Legal: PRODES), we detected a decreasing trend after the mid-2000s peak [INPE, 2011], with rates more characteristic of the early 2000s. The decrease observed here is not as severe as that reported in PRODES, which estimates deforestation rates well below those of the past two decades.

Figure 4.

Annual deforestation rates as percentage of total area and remaining forest area and their trend lines.

[29] To develop a secondary forest age distribution, we recorded the age of new secondary forest at the end of the first 10 years (1985–1994). Based on this analysis 58.7%, 17.6%, and 9.3% of pixels were 1, 2, and 3 years old, respectively; only 1.1% of the secondary forest was continuously identified as secondary forest during the first 10 years (Figure 5).

Figure 5.

Probabilities of secondary forest age applied to initial conditions (1984) based on the frequency distribution of the first 10 years: 1985–1994.

4.2 Carbon Stocks and Fluxes

[30] In 1985, total aboveground forest carbon in the study region was 1.871 PgC, decreasing to 1.166 PgC by 2009, equal to an annual decrease of 0.028 PgC yr−1 or 1.51% yr−1 (Figure 6a; Table 1). Mean forest carbon increased slightly during this period, from 243.9 Mg ha−1 in 1984 to 255.0 Mg ha−1 in 2010 due to a slight deforestation bias in lower biomass forests in southern Rondônia [Sales et al., 2007]. Cumulative gross emissions were 0.761 PgC in 2009, with a mean annual rate of 0.030 PgC yr−1; this mean rate masks considerable interannual variability in both deforestation-related and postdisturbance land use emissions (Figure 7a). Carbon uptake by pasture and secondary forest regeneration partially offsets carbon emissions by 0.069 PgC, yielding net emissions of 0.692 PgC. Following the increasing deforestation trend, annual deforestation emissions increased nearly threefold from 1987 (0.010 PgC) to 2006 (0.027 PgC). Similarly, deforestation emissions experienced a 3 year dip from 2000 to 2002. Postdisturbance land use emissions increased more consistently throughout the time series (omitting the carbon pool buildup from 1985 to 1986). Intermittent fluctuations in deforestation were reflected as an increase in the land use emissions with a 2–3 year lag.

Figure 6.

(a) Time series of the principal carbon pools, primary forest, atmosphere, secondary forest, and pasture (including elemental carbon and slash pools). (b) Smaller carbon pools, with elemental carbon, live pasture, primary slash pool, secondary slash pool, secondary forest and wood products.

Table 1. All Carbon Pools (PgC), 1985–2009, Including Primary Forest, Atmosphere, Elemental Carbon, Live Pasture, Primary Slash Pool, Secondary Slash Pool, Secondary Forest, and Wood Products
Figure 7.

(a) Annual and (b) cumulative carbon fluxes from deforestation, postdisturbance land use, and pasture/secondary forest uptake.

[31] The pasture and secondary aboveground forest carbon pools were minor compared to that of primary forest. Pasture carbon, including elemental carbon and slash pools, was 0.0152 PgC in 1985, increasing steadily until a maximum in 2006 of 0.0863 PgC, after which it decreased slightly to 0.0805 PgC (Figure 6a; Table 1). Among the constituent pasture pools, trends and quantities were quite variable. Primary forest slash was the largest pool, increasing from 0.011 to 0.044 PgC in 2006 and curtailing off to 0.033 PgC by 2009 due to the deforestation decrease (Figure 4). Live pasture carbon trends largely mirrored its expansion as the predominant land use, increasing from 0.0038 in 1985 to 0.0274 PgC in 2009.

[32] The live secondary forest carbon pool remained very low throughout the entire time series, owing to its limited spatial extent (Figure 3) and longevity (Figure 5). Mean secondary forest biomass throughout the time series was 9.72 MgC ha−1. Total secondary forest carbon increased from 0.001 to 0.004 PgC from 1985 to 1992, after which it oscillated between 0.0031 and 0.0047 PgC. There was a slight increase in the final 3 years of the time series to 0.011 PgC, most likely an artifact of the interannual corrections. Cumulative carbon uptake by secondary forest was 0.099 PgC, or 12.9% of cumulative total emissions (data not shown); recurrent clearing and burning of secondary growth kept the secondary forest carbon pool very small throughout the period, though. Secondary forest slash is the smallest of all pools, but due to consistent turnover of secondary forest, it nearly equaled that of live secondary forest in the late 1990s and early 2000s (Figure 6b; Table 1). As 8% of all harvested forest was assumed to be used as commercial wood products, the wood carbon pool is a relatively major one, increasing tenfold from 0.0022 to 0.0246 PgC over the 25 year period and roughly equaling the amount of carbon stored in live pasture throughout. The elemental carbon pool increased at a mean rate of 0.00063 PgC yr−1 to a pool of 0.0152 PgC in 2009. Elemental carbon, the product of primary and secondary forest burning, constituted a relatively significant carbon pool due to generous quantities created by primary forest conversion and very low decay rates.

4.3 Spatial Scale of LULCC

[33] As a function of their total area, larger deforestation events figured more prominently, yet 31.4% of deforestation occurred in areas smaller than 20 pixels, 63.1% occurred in areas smaller than 100 pixels, and 75% in areas smaller than 250 pixels (Figure 8a; Table 2). Assessed as a proportion of a 250 m MODIS pixel, one half of all deforested area was located in cells containing only 7% deforestation (Figure 9a). Alternatively, if one assumes that a proportional change of 50% is necessary to detect deforestation, only 8.0% of deforested area would be resolvable.

Figure 8.

Proportion of area (a) deforested and (b) in secondary forest. Area is grouped by 20-pixel wide bins (bars) and shown as a cumulative proportion (line).

Table 2. Proportion of Area Deforested and in Secondary Forest, Grouped by 20-Pixel Wide Bins, Up To 300 Pixelsa
Number of PixelsArea Deforestation (%)Area Deforestation Cumulative (%)Area Secondary (%)Area Secondary Cumulative (%)
  1. aSee Figure 8 for areas larger than 300 bins. Individual and cumulative proportions are shown.
Figure 9.

Proportion of (a) total deforested area and (b) secondary forest area occupying a 250 m pixel. Proportions are grouped by 0.04 wide bins (bars) and shown as a cumulative proportion (line).

[34] Despite its limited areal extent, secondary forest exhibited a similar size distribution as deforestation events. The median size of secondary forest was 10 pixels, (0.813 ha) and ranged from 5 to 118 pixels (2.5th and 97.5th quantiles). As a function of their total area, secondary forest fragments smaller than 100 pixels comprised 70.6% of the total area, while fragments smaller than 250 pixels accounted for 86.6% (Figure 8b; Table 2). When assessed as the proportion of a coarse MODIS pixel, secondary forest mirrored the deforestation results once again. One half of all secondary forest was found in grid cells containing 8% coverage (Figure 9b). Assuming, as above, that 50% coverage of a 250 m pixel would be required to identify secondary forest, only 5.5% of the secondary forest area would be detectable within a 250 m pixel.

5 Discussion

5.1 LULCC Trends and Emissions

[35] There was a temporally dynamic but generally increasing deforestation rate throughout the study period (Figure 4). The mean rate 1.50% yr−1 was somewhat lower than the rate of 1.9% yr−1 (1985–2008) estimated by Numata et al. [2010] but close to the mean annualized rate of 1.64% yr−1 (1975–1999) of Roberts et al. [2002]. Our estimates of the decrease in deforestation rates since 2006 are similar to those reported in PRODES [2011] but document a less steep decline. This may be due in part to the focus in PRODES on corte raso (clear cuts) [INPE, 2006], whereas our method documents small-scale, incremental annual clearing. As expected, the wall-to-wall method produced more conservative estimates than those employing the hotspot method, estimated as 3.2% yr−1 (1990–1997) [Achard et al., 2002]. Despite differing spatial domains, masks, and versions of the Roberts et al. [2002] classification products, our estimates of proportional forest cover loss match those of Numata et al. [2010] for the same period (1985–2008): 34% versus 33%, respectively. Numata et al. [2010] incorporated an earlier version of the Sales et al. [2007] biomass map, which featured significantly higher forest biomass values than the Sales [2010] map that we used. Hence, our observed forest carbon losses of 0.028 PgC yr−1 are lower than those of Numata et al. [2010] who reported forest biomass losses of 0.0384 PgC yr−1. Changes in mean forest biomass, from 243.9 Mg ha−1 in 1984 to 255.0 Mg ha−1 in 2010, follow those of Loarie et al. [2009] who found that from 2001 to 2007 the mean forest biomass of the larger Brazilian Amazon increased by 9.8%, from 366 to 402 Mg ha−1. Our increase was less dramatic, owing to the dearth of high-biomass forests in the drier climate of Rondônia.

[36] There was a limited amount of secondary forest in Rondônia throughout the time series, never rising above 5% and hovering between 1–3% during most years. Because of the low proportional cover in our study area, cumulative carbon uptake by secondary forest, 0.099 PgC, represents only 12.9% of cumulative emissions from all pools. This finding contrasts with that of Alves et al. [2003], who reported that 28% of deforested area had regenerated in both 1985 and 1995; our findings indicate secondary forest replacement rates of 10% and 12%, respectively, in agreement with Roberts et al. [2002] and Neeff et al. [2006]. The latter reported very similar (12–13%) and consistent replacement rates during the period 1978–2002. Also similar to our findings, Neeff et al. [2006] encountered only a 1% increase in secondary forest area between 1997 and 2002, despite rapid expansion during the 1980s. Our findings are in agreement with several studies [Roberts et al., 2002; Alves et al., 2003; Neeff et al., 2006] documenting the highly dynamic and short-lived nature of secondary forest; 85% of secondary forest was 3 years or younger (Figure 5). We also note that in household surveys conducted in 2005 and 2009 in Rondônia, households reported rather limited secondary forest cover on their properties—4.1% (n = 295) and 3.6% (n = 556), respectively (J. L. Caviglia-Harris et al., unpublished data, 2013)—corroborating our remote sensing-based findings. This rate is slightly lower than the Amazon-wide rate reported by Houghton et al. [2000]; state-specific rates are not included, barring direct comparison.

[37] Live pasture carbon accumulation played a very minor role in the Rondônia carbon budget; however, slash pools and land use-associated emissions from pasture lands were quite prominent. Guild et al. [1998] estimated that annual pasture fire emissions in Rondônia were 67–94% of those from deforestation fires, a range which brackets the ratio that we calculated from cumulative emissions—76.5% (Figure 7). Among years, the ratio varied based on interannual variability in deforestation and the lagged combustion of slash pools, allowing land use emissions to exceed deforestation emissions occasionally. Kauffman et al. [2009] estimated that the combined emissions from deforestation and subsequent, frequent fires of slash and regenerating vegetation are 123% of the original aboveground live biomass of primary forest. This estimate is notable considering their relatively low combustion efficiency estimate (39%) of slash and burn techniques. Our figures suggest that total cumulative emissions do exceed initial carbon stocks over short periods. Gross cumulative emissions in Rondônia over 25 years (Figure 7b), equal to 0.719 PgC, were 102% those of forest biomass losses, 0.705 PgC. On a decadal basis, that ratio was 101%, 104%, and 119% during the 1980s, 1990s, and 2000s, respectively, in close agreement with Kauffman et al. [2009]. Carbon uptake by pasture and secondary forest regrowth partially offsets those emissions (Figure 7b), and consequently, net carbon fluxes to the atmosphere, 0.655 PgC, were 92.9% those of forest biomass losses. As the combustion efficiency of slash and burn was only 51%, these results illustrate the central role of postdisturbance land use in regulating carbon fluxes.

[38] The remote sensing methodology applied in this study did not account for selective logging, which has been identified as a source of carbon losses [Asner et al., 2005] in the Brazilian Legal Amazon. Thirty-eight percent of selectively logged areas are clear cut within 5 years [Asner et al., 2006], however, and so this fraction would eventually be subsumed by the deforestation category. Our model of pasture carbon dynamics is also relatively simple, ignoring variability in pasture management and intermediate stages between pasture and secondary forest. Distinct successional stages are difficult to distinguish, even from airborne videography [Powell et al., 2004], and we lack the methodology and ancillary data to ascertain pasture management from the Landsat imagery. It is likely that some of the high secondary forest turnover rates encountered here represent poorly managed pasture parcels whose spectral reflectance lies near the boundary between the pasture and secondary forest endmembers used.

5.2 The Importance of Spatial Resolution in LULCC Monitoring

[39] In this study we have encountered a trend of small-scale and incremental deforestation occurring in Rondônia. Our results indicate that medium-resolution imagery, such as Landsat, is necessary to conduct annual monitoring of LULCC and carbon balance in tropical regions dominated by smallholder agriculture and settlements. Similarly, Hansen et al. [2008c] found that in the Congo, MODIS is too coarse for such small-scale deforestation. The same may not necessarily be true of areas dominated by mechanized clearing and agriculture, such as the Brazilian State of Mato Grosso, the site of significant large-scale clearing for industrial farming [Morton et al., 2006; 2008; Hansen et al., 2008b; van Der Werf et al., 2009]. Additionally, it is apparent that imposing large, stringent size thresholds, beyond those necessary to ensure quality control and/or classification accuracy, can result in substantial omissions of carbon fluxes. Recognition of this fact introduces methodological challenges—specifically data gaps due to cloud cover in areas lacking a distinct dry season, such as the northwestern Amazon or western equatorial Congo.

[40] Shifting forest policy may also underscore the importance of higher spatial resolution monitoring. On 25 May 2012, the president of Brazil signed into effect portions of major revisions to the Brazilian forestry code. The new laws would create exemptions for smallholders to the 80% forest cover on land parcels required since 1996; additionally, it would allow states to determine their own limits [Tollefson, 2011]. At the time of writing, the exact interpretation of the final law was not clear. Yet, in light of this approved legislation, small-scale forest clearing by landowners could play an ever increasing contribution to carbon emissions, and MRV schemes will increasingly rely on higher spatial resolution monitoring.

5.3 Rondônia: From Carbon Source to Sink?

[41] How does Rondônia figure into the larger context of pantropical LULCC and what are the implications for REDD+? The amount of forest carbon lost in the study area over 25 years, 0.705 PgC (Figure 7b), is roughly equal to the annual pantropical humid forest carbon losses during the 1990s (0.64 PgC yr−1) [Achard et al., 2002] and early 2000s (0.81 PgC yr−1) [Harris et al., 2012]. While this may seem unrealistic for the relatively small state, Rondônia accounts for 13.1% of all deforested area in the Brazilian Legal Amazon [INPE, 2011], which in turn accounts for 47% of global humid forest clearing [Hansen et al., 2008a]. Hence, Rondônia accounts for 6.2% of the annual pantropical deforestation. Further, according to our estimates, the study area contributed 3.5–4% of annual pantropical deforestation emissions. If baseline reference levels were based on historical deforestation, as is the case with the Amazon Fund, landowners in Rondônia would stand to benefit significantly from the REDD+ compensation. In agreement with official figures [INPE, 2011], we found that deforestation declined in the latter half of the past decade. Should this trend continue, REDD+ incentives will create an attractive and effective alternative to the punitive measures which are generally credited with reducing deforestation in recent years [Nepstad et al., 2009]. The dearth of secondary forest in Rondônia also highlights an important opportunity for the state in a REDD+ regime that compensates for regeneration and restoration of forest carbon stocks, as well as avoids deforestation and forest degradation. Promoting reforestation and regeneration would have further benefits in terms of local ecosystem services, like water quality, and potentially reduce leakage by providing new sources of wood products, thus reducing the profitability of expanding access into mature forest for timber extraction [Brown and Lugo, 1990].

6 Conclusions

[42] This study modeled the amount of carbon release and uptake within the Amazonian frontier state of Rondonia, Brazil using a seamless 27 year (1984–2010) Landsat-based land cover classification time series. Over this time period, we note a relatively high level (1.50% yr−1) of deforestation and resultant emissions (1.51% yr−1). We estimated that the amount of forest carbon lost in the study area over 25 years, 0.705 PgC (Figure 7b), is roughly equal to the annual pantropical humid forest carbon losses during the 1990s (0.64 PgC yr−1) [Achard et al., 2002]. Forest was replaced predominantly by pasture throughout the time period. Pasture harbored a significant amount of biomass, predominantly in the form of primary and secondary forest slash, while live pasture biomass was a very minor carbon pool. Secondary forest remained a relatively minor land use, comprising less than 5% of the landscape and exhibiting frequent turnover to pasture. Consequently, secondary forest regrowth carbon uptake did not substantially compensate for deforestation emissions. Hence, net carbon fluxes to the atmosphere during the study period, 0.655 PgC, were 92.9% those of original forest biomass.

[43] This study hinged on high-spatial resolution monitoring of land cover change. Analysis of the spatial scale of deforestation and secondary forest regrowth underscored the importance of this approach. Assuming that a proportional change of 50% is necessary to detect deforestation or secondary forest regrowth, only 8.0% and 5.5% of annual change in these two classes would be detectable within a 250 m MODIS pixel. The insensitivity of coarser remote sensing instruments will bear importance on the implementation of REDD+ financial mechanisms and recent Brazilian legislation [Stokstad, 2011; Tollefson, 2011] which will likely pivot deforestation more towards smallholders along the Amazonian frontier.


[44] Many thanks are given to the Rondônia survey and field team we worked with during the 2009 field season. This work was funded by NSF grant BCS-0751292 and NASA Headquarters under the NASA Earth and Space Science Fellowship Program, grant NNX10AP15H, with support from NASA's Biological Diversity Program (NNX07AF16G).