Mortonetal.CarbonBalanceandManagement2011,6:18 http://www.cbmjournal.com/content/6/1/18 REVIEW Open Access Historic emissions from deforestation and forest degradation in Mato Grosso, Brazil: 1) source data uncertainties Douglas C Morton1*, Marcio H Sales2, Carlos M Souza Jr2 and Bronson Griscom3 Abstract Background: Historic carbon emissions are an important foundation for proposed efforts to Reduce Emissions from Deforestation and forest Degradation and enhance forest carbon stocks through conservation and sustainable forest management (REDD+). The level of uncertainty in historic carbon emissions estimates is also critical for REDD +, since high uncertainties could limit climate benefits from credited mitigation actions. Here, we analyzed source data uncertainties based on the range of available deforestation, forest degradation, and forest carbon stock estimates for the Brazilian state of Mato Grosso during 1990-2008. Results: Deforestation estimates showed good agreement for multi-year periods of increasing and decreasing deforestation during the study period. However, annual deforestation rates differed by > 20% in more than half of the years between 1997-2008, even for products based on similar input data. Tier 2 estimates of average forest carbon stocks varied between 99-192 Mg C ha-1, with greatest differences in northwest Mato Grosso. Carbon stocks in deforested areas increased over the study period, yet this increasing trend in deforested biomass was smaller than the difference among carbon stock datasets for these areas. Conclusions: Estimates of source data uncertainties are essential for REDD+. Patterns of spatial and temporal disagreement among available data products provide a roadmap for future efforts to reduce source data uncertainties for estimates of historic forest carbon emissions. Specifically, regions with large discrepancies in available estimates of both deforestation and forest carbon stocks are priority areas for evaluating and improving existing estimates. Full carbon accounting for REDD+ will also require filling data gaps, including forest degradation and secondary forest, with annual data on all forest transitions. Keywords: Amazon, REDD+, IPCC, Tier, Approach, Landsat 1. Background agreement under the United Nations Framework Con- Tropical deforestation accounted for approximately 12% vention on Climate Change [7]. In 2010, the Cancun of anthropogenic CO emissions in 2008 [1]. Forest Agreements expanded the scope for climate mitigation 2 degradation from fire, logging, and fuel wood collection activities in forests to include the conservation and represents an additional source of carbon emissions enhancement of forest carbon stocks and sustainable from land use activities in tropical forest regions [1-6]. forest management, or REDD+ [8]. Recognition of the important contributions from defor- Proposed REDD+ mechanisms require a baseline or estation and forest degradation to anthropogenic green- reference emissions level against which future emissions house gas emissions led to proposals for Reduced can be compared [9,10]. Previous scientific studies have Emissions from Deforestation and forest Degradation estimated historic deforestation carbon emissions at (REDD) to be included in a post-2012 climate pan-tropical [1,3,11-13] or regional spatial scales, such as the Brazilian Amazon [14-19]. However, the spatial and temporal resolutions of previous deforestation emis- *Correspondence:[email protected] 1NASAGoddardSpaceFlightCenter,GreenbeltMDUSA sions estimates are likely too coarse for national REDD+ Fulllistofauthorinformationisavailableattheendofthearticle ©2011Mortonetal;licenseeBioMedCentralLtd.ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommons AttributionLicense(http://creativecommons.org/licenses/by/2.0),whichpermitsunrestricteduse,distribution,andreproductionin anymedium,providedtheoriginalworkisproperlycited. Mortonetal.CarbonBalanceandManagement2011,6:18 Page2of13 http://www.cbmjournal.com/content/6/1/18 baselines, given the potential inclusion of sub-national disagreement among available data products. In the sec- activities [8]. In addition, the input data and methods in ond manuscript, we describe a new model, the Carbon these studies were not necessarily consistent with gui- Emissions Simulator, to quantify the contribution from dance on national-scale reporting of emissions from for- source data evaluated in this study and model para- est lands from the Intergovernmental Panel on Climate meters to total uncertainties in forest carbon emissions. Change [20,21]. A range of forest carbon stock and The Carbon Emissions Simulator uses both Monte deforestation data products exist at national and sub- Carlo and error propagation techniques to quantify national scales that could be used to establish historic uncertainties in deforestation carbon emissions. By emission levels [21], but the suitability of existing data separating data and model-based uncertainties, the Car- for estimating historic carbon emissions and associated bon Emissions Simulator can be used to evaluate trade- uncertainties has not been thoroughly evaluated. offs for improving historic emissions estimates by year, The level of uncertainty in historic emissions baselines region, and source term. Together, these papers provide is critical for REDD+. Uncertainty in forest carbon emis- a comprehensive look at the data and research methods sions arises from estimated rates of deforestation and needed to quantify and reduce uncertainties in historic forest degradation, forest carbon stocks [22,23], and forest carbon emissions estimates for REDD+. emissions factors [16,24,25]. Large uncertainties could undermine the effectiveness of REDD+ by limiting the 2. Results ability to generate credits from mitigation actions, espe- 2.1 Deforestation cially if a conservative approach is used to estimate All five deforestation data products identified periods of REDD+ credits [26,27]. In the absence of a conservative increasing (2001-2004) and decreasing (2005-2007) approach, large uncertainties in historic emissions could deforestation rates in Mato Grosso (Figure 1, Table 1). lead to a situation in which mitigation actions fail to Annual deforestation rates were highly variable during generate climate benefits (i.e., “hot air”). the study period, ranging between 2,203 and 11,082 km2 Research to reduce uncertainties in REDD+ baselines yr-1 during 1990-2008. Three years with deforestation at national or sub-national scales may generate both rates greater than 10,000 km2 yr-1 accounted for more scientific and policy payoffs. Tropical deforestation than 25% of the total forest loss during this period remains the most uncertain term in the global carbon (1995, 2003, and 2004). budget [28]. Attention to the source and magnitude of On an annual basis, deforestation rates from different uncertainties in emissions estimates at the national level data products exhibited considerable variability (Figure can therefore help to constrain the global carbon bal- 1). In three consecutive years (2000-2002), deforestation ance. Reducing uncertainties at the national level may rates from SEMA were approximately half those from remove potential discounts from REDD+ carbon credits. PRODES-Digital, despite reliance on similar Landsat Recent studies suggest that uncertainties in rates of base data for both products. The range of annual defor- deforestation and forest degradation [26,29] and forest estation rates exceeded the expected performance of carbon stocks [30] can dramatically alter the cost-benefit satellite-based approaches (80-95% accuracy, [34]) in calculation for REDD+ from the country perspective. more than half of the years with multiple satellite-based Here, we use a structured approach to evaluate the deforestation products (1998-2008). Using a confidence source and magnitude of uncertainties in historic forest interval of ± 20%, low and high estimates of annual carbon emissions for the Brazilian State of Mato Grosso. deforestation did not overlap in these six years. The Mato Grosso is a hotspot of recent deforestation, inclusion of MODIS-based deforestation data increased accounting for more than 15% of humid tropical forest the range of annual deforestation estimates in 2005- losses worldwide during 2001-2005 [31-33]. Compared 2008, yet the INPE-DETER and Imazon-SAD estimates to other tropical forest regions, Mato Grosso also has a only represented the high and low values in 2008. wealth of data with which to evaluate historic emissions. The legacy of differences in satellite-based deforesta- In this, the first of two research articles, we review the tion datasets can also be seen in the spatial distribution available data for Mato Grosso on deforestation, forest of cumulative forest loss (Figure 2). In 1997, SEMA degradation, and forest carbon stocks to identify impor- deforestation estimates indicated greater cumulative for- tant data gaps and research needs to reduce source data est losses than PRODES-Digital data in northern Mato uncertainties in historic forest carbon emissions esti- Grosso (Figure 2a). By 2005, cumulative forest losses mates. We concentrate on five annual deforestation derived from PRODES-Digital data were higher across datasets and six estimates of forest carbon stocks in the state, with the greatest differences around the Xingu Mato Grosso (see Sections 5.3 and 5.4). We conclude River basin in eastern Mato Grosso (Figure 2b). These this study with a roadmap for research in support of areas of greatest uncertainty highlight the need for addi- REDD+ based on the spatial and temporal patterns of tional field and remote sensing research to identify the Mortonetal.CarbonBalanceandManagement2011,6:18 Page3of13 http://www.cbmjournal.com/content/6/1/18 Figure1AnnualdeforestationinMatoGrossoStateduring1990-2008.Bluenumbersindicatetheratiobetweenlowandhighannual deforestationestimates.IndividualdataproductsaredescribedinTable1. causes of consistent spatial discrepancies among satel- The spatial distribution of forest carbon stocks in lite-based estimates of forest area change. Mato Grosso differed markedly between Tier 2.m data products considered in this study. The Saatchi et al. and 2.2 ForestCarbon Stocks Imazon estimates of aboveground live biomass (AGLB) Average forest carbon stock estimates from Tier 1 and disagreed by ~50 Mg ha-1 in central and eastern por- Tier 2 data products for Mato Grosso varied by a factor tions of the state, and differences between the two pro- of two (Table 2). The Tier 1 estimate of forest carbon ducts exceeded 100 Mg ha-1 in northwest Mato Grosso stocks in Mato Grosso was the highest estimate of total (Figure 3). The conversion from AGLB to total biomass carbon (206 Mg C ha), based on the value for humid amplified the spatial discrepancies in Figure 3 because tropical forests in South America. The Tier 1 root-shoot expansion factors for BGB and aboveground dead bio- ratio (0.37) was much higher than for other products mass (AGDB) for the Imazon product were larger than (0.21-0.26), suggesting that below ground biomass in the Saatchi et al. data product (see Table 2). (BGB) accounts for part of the difference between Tier 1 and Tier 2 carbon stock estimates. Among Tier 2 data 2.3 Deforested Biomass products, the source of plot data, number of parameters Deforestation in Mato Grosso during 1993-2008 was in the biomass expansion factor, and methods to inter- concentrated in low biomass forest types. Average bio- polate between plot locations all contributed to the dif- mass in deforested regions increased during the study ference in carbon stock estimates. The wide range of period (Table 3) but remained below state-wide averages average forest carbon stocks for Mato Grosso suggests (Table 2). For the combination of SEMA deforestation that per-product uncertainty could be greater than ± data with Imazon biomass estimates, average AGLB in 50%, similar to an earlier assessment of biomass data deforested regions increased by 2.3 Mg ha-1 yr-1 during products by [22]. 1993-2005 (R2 = 0.85, Table 3). PRODES-Digital data Table 1Data sourcesforhistoric deforestation inMato Grosso, Brazil. Dataset Approach TemporalCoverage MMU1 Reference Sensor Method INPE-PRODES 2 1987-2008 6.25ha [58] Landsat Singleimage,visualinterpretation PRODES-Digital2 3 1997-2008 1ha [33] Landsat Singleimage,digitalprocessing,visualinterpretation SEMA2 3 1992-2005 1ha [32] Landsat Singleimage,digitalprocessing,visualinterpretation IMAZON-SAD3 3 2005-2008 12.5ha [59] MODIS2 Twoimages,digitalprocessing,automatedanalysis INPE-DETER3 3 2004-2008 25ha [60] MODIS2 Singleimage,digitalprocessing,visualinterpretation 1Minimummappingunit(MMU):thesmallestareaofnewdeforestationidentifiedinanyyear. 2AnnualdeforestationestimateswerenotavailablefromPRODES-Digitalduring1997-2000orSEMAfor1996,1998,and2000.Averagevaluesofforestloss betweenimagedateswereusedintheseyears(e.g.,forestareachangebetween1995-1997imageswasdividedequallybetween1996and1997). 3Alertdataproductsprovidenear-realtimemonitoringofdeforestationusingimageryfromtheMODISsensorsat250mresolution.Thesedataareprimarily intendedtoidentifythelocationofnewdeforestation,especiallyfordeforestationevents>25ha,ratherthanproviderobustestimatesofforestareachange. Mortonetal.CarbonBalanceandManagement2011,6:18 Page4of13 http://www.cbmjournal.com/content/6/1/18 Figure2SpatialdifferencesbetweenPRODES-DigitalandSEMAestimatesofcumulativedeforestationthrough1997(a)and2005(b), summarizedasthedifferenceinremainingforestarea(km2)withineach0.25°cell.AreasoutsideofthePRODESforestmaskappeargray. also suggest an increasing trend in average AGLB in deforestation during 2005-2008 (47.8 Mg ha-1) was lar- deforested areas using Imazon biomass estimates during ger than the total increase in average deforested biomass 2001-2008 (1.8 Mg ha-1yr-1, R2 = 0.43). Differences in from either product during the study period (< 30 Mg the location of recent deforestation between PRODES- ha-1, Table 3). Digital and SEMA had little impact on the average bio- mass in deforested areas during 2001-2005 from the 3. Discussion IMAZON product (< 2 Mg ha-1), as discrepancies The range of available deforestation and biomass data between these products were widely distributed across products provides a first estimate of source-data uncer- low and high biomass forests in Mato Grosso by 2005 tainties in historic deforestation carbon emissions. Find- (Figure 2b). ings in this study highlight how specific years, regions, Overall, the choice of Tier 2.m biomass data had a lar- and data products contribute to potential variability in ger impact on estimates of deforested biomass than the deforestation emissions estimates. Large (25-50%) dis- trend of increasing biomass in recently deforested areas crepancies remain between estimates of forest carbon from either product. The difference in average AGLB stocks and annual deforestation from different data pro- between Imazon and Saatchi et al. data products for ducts, even for estimates at the same Tier or Approach. Table 2Tier 1andTier 2data sourcesfortropical rainforest carbon stocksin Mato Grosso. Source TotalC:AGLB+AGDB+BGB AGDB,BGB(%AGLB) PlotData CarbonFraction(CF) Tier2 (MgC/ha)1 IPCC3 SA:206 9%,37% N/A 0.47 1 Houghtonetal.2001 BA:192 9%,21% LiteratureReview 0.5 2.a Brown&Lugo1992 BA:156 9%,21% RADAM4 0.5 2.a Nogueiraetal.2009 MT:159.7 13.91%,25.8%5 RADAM4 0.485 2.a Imazon;Salesetal.2007 MT:130.4±44.8 13.91%,25.8%5 RADAM4 0.485 2.m Saatchietal.2007 MT:99.0±58.0 9%,21% Houghtonetal.2001 0.5 2.m Tier1dataaretheIPCCdefaultvaluesforforestcarbonstocks,whereasTier2indicatescountry-specificdata(seeTable4).Totalforestcarbon(C)wasestimated fromabovegroundlivebiomass(AGLB)usingconversionfactorsfromeachsourceforabovegrounddeadbiomass(AGDB)andbelow-groundbiomass(BGB)asa percentageofAGLB. 1AveragetotalcarboninforestbiomassfortropicalrainforestSouthAmerica(SA),BrazilianAmazon(BA),orMatoGrosso(MT). 2Tier2biomassdataproductsweredividedbetweenregionalorstate-wideaveragevalues(Tier2.a)andspatially-explicitmapsofforestbiomass(Tier2.m). 3Asreportedby[80] 4TheRADAMBRASILfloristicinventory(DPNM,1973-1983). 5Nogueiraetal(2009)appliedadditionalcorrectionfactorsforAGLBindense(10.5%)andnon-dense(15.7%)foresttypes.Thesefactorswerealsoincludedin theImazonproduct. Mortonetal.CarbonBalanceandManagement2011,6:18 Page5of13 http://www.cbmjournal.com/content/6/1/18 Figure3MapofdifferencesbetweenSaatchietal.[75]andImazon[76]estimatesofAGLBinnorthernMatoGrossostate.Imazonestimates exceedthoseofSaatchietal.inred,orange,andyellowareas,whilegreenareasindicatehigherAGLBestimatesfromSaatchietal. Deforestationthrough2005isshowningray,andnon-forestareaswithinMatoGrossoappearwhite.Individualdataproductsaredescribedin Table2. Reconciling these differences is essential to reduce interested in revising estimates of forest area changes uncertainties in historic deforestation emissions esti- from 1972-present [35]. Landsat resolution (30 m) is mates and prevent the propagation of errors from subse- suitable for detailed estimates of forest area change [34], quent land-use transitions in disputed areas. provided that an accuracy assessment can be conducted Reducing source data uncertainties requires careful using very high resolution (< 5 m) imagery from air- methods to substitute space for time. The archive of borne or satellite data sources [36]. In the case of Mato Landsat satellite imagery is a rich resource for countries Grosso, where most deforestation occurs in large clear- ings (> 25 ha, [37]), forest area change estimates from moderate resolution (250 m) deforestation monitoring Table 3Mean aboveground live biomass ±1SD inareas systems do not differ much from estimates obtained ofrecentdeforestation (Mg ha-1). from Landsat-based deforestation maps (see Figure 1, Year SEMA/ PRODES-Digital/ PRODES-Digital/Saatchi Table 1). However, deforestation alert systems are inap- Imazon Imazon etal. propriate for monitoring small forest clearings [38] or 1993 158.9±46.9 forest degradation from selective logging [39] for esti- 1994 152.6±43.2 mates of historic carbon emissions. 1995 167.0±54.9 In contrast to the rich archive of historic satellite data, 1996 165.5±49.5 there is limited historic forest inventory data for Mato 1997 165.5±49.5 Grosso. Improving estimates of tropical forest carbon 1998 174.4±61.0 180.5±55.9 stocks will therefore require new data collection. A new 1999 174.4±61.0 180.5±55.9 National Forest Inventory is already underway in Brazil 2000 179.4±56.2 180.5±55.9 (http://ifn.florestal.gov.br), with field plots distributed on 2001 179.4±56.2 166.6±54.1 a regular grid (20 km × 20 km). New technologies offer 2002 181.1±53.4 172.7±57.0 the possibility to generate spatially explicit biomass 2003 179.6±56.9 183.4±58.8 maps using a more limited network of forest inventory 2004 185.7±59.0 181.0±56.5 plots and large-area sampling of forest heights with air- 2005 180.8±57.6 184.6±61.4 135.5±87.3 borne or spaceborne LiDAR [40-42]. However, contem- 2006 179.8±63.7 126.7±90.5 porary estimates of forest carbon stocks at the 2007 187.0±67.8 143.8±90.2 deforestation frontier must then be paired with data on 2008 179.1±56.0 136.5±82.3 historic deforestation and forest degradation to account Tables1and2provideadditionaldetailsregardingTier2.mbiomassdata for the impacts of historic land use on contemporary products(Imazon,Saatchietal.)andApproach3deforestationproducts (SEMA,PRODES-Digital),respectively. measurements (e.g., [43]). Routine sampling may be Mortonetal.CarbonBalanceandManagement2011,6:18 Page6of13 http://www.cbmjournal.com/content/6/1/18 needed to maintain updated field or LiDAR-based infor- full carbon accounting from deforestation and forest mation on forest carbon stocks for REDD+ [44] because degradation (Figure 5). static reference data are unable to account for increases At least two factors likely contributed to the observed in forest carbon stocks over time (e.g., [45]) or reduc- spatialandtemporaldiscrepanciesinannualdeforestation tions in biomass from forest disturbance (e.g., [46,47]). rates for Mato Grosso. First, none of the satellite-based What research is needed to reduce source data uncer- deforestation estimates were developed specifically for tainties in Mato Grosso and other Amazon regions? REDD+. Asaresult, forest degradationfromloggingand New measurements of forest carbon stocks and new firemayhavebeenincludedinhistoricdeforestationesti- estimates of forest area changes from remotely-sensed mates, especially in years with extensive damages from data are most critical in regions where existing products understory forest fires [48]. Incomplete information on disagree (Figure 4). Areas with high uncertainties in forest degradation and secondary forest dynamics also both forest biomass and deforestation rates provide an contributestosourcedatauncertaintiesforestimatingnet opportunity to collect complementary information on forest carbon emissions in Mato Grosso. Full carbon land use and carbon stocks to improve estimates of his- accountingfromdeforestationandforestdegradationwill toric carbon emissions. Improved estimates of forest requirecarefulconsiderationofsequentialland-usetransi- carbon stocks in areas with concentrated historic defor- tions(Figure5).Atime-seriesapproachtotrackdeforesta- estation are a specific priority for efforts to quantify his- tion, degradation, and secondary forest dynamics using toric emissions and establish REDD+ baselines. annual satellite imagery could improve emissions esti- Additional data collection and analysis in these areas are mates for Mato Grosso and other tropical forest regions needed to develop a consistent, validated approach for byreducingmisclassificationand“doublecounting”errors Figure4DataneedstoreduceuncertaintiesinhistoricdeforestationcarbonemissionsfromMatoGrosso,summarizedat0.25°spatial resolution.WhitecellsindicateareaswhereLandsat-basedestimatesofcumulativedeforestationthrough2005differby>40km2.Graycells indicateregionswhereaverageTier2.mestimatesofAGLBinremainingforestin2005differby>50Mgha-1.Cellswithdataneedsforboth deforestationandbiomassappearblack. Mortonetal.CarbonBalanceandManagement2011,6:18 Page7of13 http://www.cbmjournal.com/content/6/1/18 Figure5Landusetransitionsandrelateddataneedstoestimatecarbonemissionsfromdeforestationandforestdegradation.Full carbonaccountingrequiresdatafortherate(R)ofareachangeandassociatedchangesincarbonstocks(ΔC)fordeforestation(D),forest degradation(L),andregrowth(R).Allforestlandsmustmeetminimumheight(h),crowncover(CC),andarea(A)requirements,accordingto eachcountry’snationalforestdefinition.Solidarrowsrepresentprimarytransitionsfromforesttonon-forestordegradedforestlands;dashed arrowsrepresentsecondaryland-usetransitions. [48] that occur when degraded forests are deforestedfor The range of available data products provides an indi- agricultural use [49]. Second, time series methods may cation of the spatial and temporal variability associated also improve the consistency of deforestation estimates with estimates of deforestation and forest carbon stocks. over time. Deforestation estimates in this study were However, total uncertainties in historic emissions cannot based on interpretation of a single satellite image or a be estimated without validation efforts to characterize comparison between two successive images. Time series per-product uncertainties. Validation needs are greatest methods thatconsiderlonger periods of disturbance and in areas where existing products disagree (Figure 4), but recovery may improve the accuracy of change detection all deforestation and carbon stock data products should [50],especiallyforretrospectiveanalysestoestablish his- include a robust validation plan with routine field mea- toric baselines. Annual satellite data canbe usedto con- surements and airborne or spaceborne very high resolu- firm continued agricultural use of previously deforested tion imagery (< 5 m). areas,forestrecoveryfollowingdegradation,andtheageof Given the need for routine data collection on forest secondaryforestsfromlandabandonmenttoimprovecar- transitionsandassociatedcarbonlosses,developmentand bonstockestimatesinareasofactivelandusechange. maintenance of reporting information for REDD+ will In addition to reducing source data uncertainties, rea- likelyrequirededicatedcapacityforsatelliteandfielddata nalysis of historic changes in forest area can also facili- analysis.Consistentmethodsfordataanalysisarealsocri- tate sub-national allocation of deforestation baselines. ticalforREDD+[27].Eveninawell-characterizedregion Brazil recently selected the 1996-2005 period for defor- suchasMatoGrosso,multipledeforestationdataproducts estation baseline calculations [51]. However, annual were required to consider forest area changes during PRODES-Digital deforestation data are only available 1990-2008 because no single product provided annual beginning in 2000. Allocation of baseline deforestation estimatesduringtheentirestudyperiod.Thedevelopment information to Amazon states can be accomplished ofstandardsforREDD+monitoring,reporting,andverifi- using PRODES statistics (Approach 2), but below the cation(MRV)providesanopportunitytodesignasystem state scale, regional or project-scale activities may that can lower uncertainties in emissions estimates over require a new analysis of historic deforestation and for- time using the comparative approach described in this est degradation to provide Approach 3 data for all years paper.Ideally,theanalysisinthisstudywouldbethefirst during the baseline period. iterationofaroutineprocesstotargetnewdatacollection Mortonetal.CarbonBalanceandManagement2011,6:18 Page8of13 http://www.cbmjournal.com/content/6/1/18 in regions and years with largest uncertainties in carbon of forest cover and the criteria for deforestation. The stockanddeforestationestimates. Brazilian government defines their forest land as areas of at least one hectare in size with more than 30% 4. Conclusions crown cover of trees ≥ 5 m in height. This definition This study reviewed available data products for defores- selects the upper end of ranges for area (0.04-1.0 ha), tation, forest degradation, and forest carbon stocks in crown cover (10-30%), and tree height (2-5 m) in guide- Mato Grosso, Brazil to assess the level of uncertainty in lines established by the UNFCCC for the Clean Devel- source data for estimating historic forest carbon emis- opment Mechanism of the Kyoto Protocol [52]. sions for REDD+. Deforestation data showed consider- Deforestation occurs when any of these thresholds are able spatial and temporal variability, with Landsat-based crossed, typically during the conversion of forest for estimates of annual deforestation differing by > 20% in agricultural use (Figure 5). Within the scope of REDD+, most years. Forest carbon stock estimates exhibited even forest degradation is generally considered a reduction in greater variability, with more than a two-fold difference carbon stocks within forest land remaining as forest in carbon stock estimates in northwest Mato Grosso. [21], although a precise definition of forest degradation Limited information was available on forest degradation has not been adopted [53]. and secondary forest regeneration, suggesting that full Data on the rates of forest transitions and associated carbon accounting for REDD+ cannot be achieved with- changes in carbon stocks are classified according to the out additional satellite data analysis to quantify annual methods used for data collection (Table 4). We followed transitions involving degraded or regenerating forests. this guidance when reviewing and analyzing available The diversity of deforestation and carbon stock esti- forest area change data (Activity Data) and data on mates for Mato Grosso provides an initial indication of changes in carbon stocks (Emissions Factors) from tran- research needs to address source data uncertainties for sitions between forest land and other land uses [21]. For REDD+. Spatial and temporal patterns of disagreement estimates of deforestation, moving from Approach 1 to show priority areas for new data collection, and a coor- Approach 3 area change data involves a shift from glo- dinated strategy to estimate forest carbon stocks and bal or national survey methods (e.g., the Food and Agri- validate deforestation estimates in these areas could tar- cultural Organization’s periodic Forest Resource get the main source data uncertainties in Mato Grosso. Assessment surveys) to spatially-explicit estimates from Data needs for REDD+ differ from previous uses of satellite remote sensing data (Table 4). Approach 3 data deforestation information for enforcement of environ- are recommended as the basis for establishing REDD+ ment laws and private property rights. The additional baselines [21], since fine-scale spatial information (20-60 focus on source data uncertainties for REDD+ could m) is necessary to track sequential land-use transitions reduce large uncertainties in current emissions esti- at a given location through time (Figure 5). The specifi- mates, thereby increasing the likelihood of generating city of source data on forest carbon stocks also increases benefits from REDD+ actions. from continental-scale averages for each forest type (Tier 1) to country-specific information (Tier 2). Few 5. Materials and methods countries have established Tier 3 efforts to repeatedly Below, we synthesize relevant IPCC guidance for source measure or model forest carbon stocks that could be data on changes in the area and carbon stocks in forest used to estimate historic emissions. lands (Section 5.1), describe the Mato Grosso study area (Section 5.2), and review available data on forest area 5.2 Study area changes (Section 5.3) and carbon stocks (Section 5.4). The state of Mato Grosso includes the southernmost extent of Amazon forests in Brazil (Figure 6). Data 5.1 IPCC Tiers and Approaches from the RADAMBRASIL floristic surveys (1973-1983) The definition of ‘forest’ forms the foundation of REDD indicate that Amazon forest and transition forest types + and related initiatives, establishing the spatial extent initially covered two-thirds of the state [54]. The Table4SummaryofIPCCdatacategoriesforActivityDataonforestareachangesandEmissionFactorsforchangesin carbon stocksfrom deforestation andforest degradation. ApproachesforActivityData:ForestAreaChanges TiersforEmissionFactors:ChangesinCarbonStocks 1.Non-spatialcountrystatistics 1.IPCCdefaultvaluesbycontinentandforesttype 2.Maps,surveys,andothernationalstatisticaldata 2.Countryspecificdataforkeyfactors 3.Spatiallyexplicitdatafrominterpretationofremote 3.Nationalinventoryofcarbonstocks,viarepeatedmeasurementsofkeystocksthrough sensingimagery timeormodeling Approach1andTier3dataproductswereunavailableforMatoGrossoduring1990-2008.Pleasesee[21]foramorecompletediscussionofIPCCGoodPractice Guidance. Mortonetal.CarbonBalanceandManagement2011,6:18 Page9of13 http://www.cbmjournal.com/content/6/1/18 Figure6Forestcoverextent(black)fromtheINPEPRODESprogramintheBrazilianStateofMatoGrosso(inset,white).Outsideofthe PRODESforestmask,areaswith>30%treecoverin2001appeardarkgray[55].ThePantanalbiomeinsouthernMatoGrossoisoutlinedin white. Brazilian Instituto Nacional de Pesquisas Espaciais 30 m pixel size), and the minimum mapping unit from (INPE) further refined the extent of Amazon forests these products is consistent with the one-hectare thresh- using Landsat satellite data under the PRODES (Moni- old for individual forest patches in Brazil’s national for- toramento da Floresta Amazônica Brasileira por Saté- est definition [34]. Most studies report gross rather than lite) program of annual deforestation assessments in net deforestation, as transitions involving secondary for- the Brazilian Amazon [33]. The PRODES forest mask est (e.g., agricultural abandonment and re-clearing) were is a common reference for many Amazon deforestation not routinely identified in historic assessments (see Fig- products. We adopted the PRODES forest mask for ure 5). our analysis to maintain consistency with results from We evaluated annual satellite-based estimates of other studies (Figure 2); however, the PRODES mask deforestation in Mato Grosso beginning in 1990, and we does not include all areas that could be classified as extended our evaluation through 2008 to include new forest in Mato Grosso based on the 30% crown cover deforestation products that were developed based on threshold [55]. moderate resolution (250 m) satellite imagery. We iden- tified five satellite-based estimates of Amazon deforesta- 5.3 Forestarea change data for Mato Grosso tion in Mato Grosso covering part or all of the 1990- 5.3.1 Deforestation 2008 timeframe (Table 1). We limited our review to Deforestation in the Brazilian Amazon has been moni- annual deforestation estimates, thereby excluding avail- tored for more than two decades using a variety of satel- able data from regional and global deforestation pro- lite sensors (e.g., [33,37,56]. Fine-scale mapping efforts ducts with periodic (5-10 year) evaluation periods have relied on Landsat or other high-resolution data (≤ [3,11,31,56,57]. Mortonetal.CarbonBalanceandManagement2011,6:18 Page10of13 http://www.cbmjournal.com/content/6/1/18 Deforestation data products were grouped according 5.4 Forestcarbon stocks to IPCC Approach. We categorized annual deforestation The amount and spatial distribution of forest carbon statistics from the PRODES-Analog product (INPE) as stocks in Amazonia are major sources of uncertainty in an Approach 2 dataset since these data were not spa- estimates of emissions from deforestation and forest tially explicit [58]. The PRODES-Digital and SEMA data degradation [22]. [70] estimated the total forest carbon products provided the longest time series of Approach 3 storage in the Brazilian Amazon as 39-93 Pg C, but the deforestation data. The length of the deforestation data seven data products reviewed in that study disagreed record is critical for estimating annual emissions; a about the spatial distribution of low and high-biomass minimum of 10 years of historic deforestation data are forest types within the region. Recent efforts to refine recommended to estimate the contribution from pre- maps of forest carbon stocks in Amazonia have focused vious clearing activity to emissions in any given year on new plot measurements of forest biomass [71], [16]. We also included two “alert” data products (INPE- improved allometric relationships relating wood volume DETER and Imazon-SAD) from efforts to monitor to biomass [72,73], and extrapolation of plot-based data deforestation in near-real time based on moderate reso- using climate metrics [74], satellite-based estimates of lution (250 m) satellite data [59,60]. Inclusion of alert forest canopy reflectance [75], and geostatistical meth- data products allowed us to characterize the additional ods [76]. Revised estimates of the total forest carbon uncertainty in annual deforestation estimates that could stocks in Amazonia fall within the original range arise if only alert-type data on area change were described by [70], albeit with lower forest biomass in available. areas of active deforestation in southern and eastern 5.3.2 Forest Degradation and Secondary Forests Amazonia than previously estimated [19,73,75]. Remain- Few satellite-based estimates of forest degradation exist ing uncertainties in the spatial distribution of forest bio- for Mato Grosso. New algorithms to detect selective log- mass arise from the small number of forest plots [77] ging [4,61,62] and understory forest fires [48,61,63,64] and the influence of historic land use on forest carbon using Landsat data were developed in Mato Grosso. stocks, especially along the deforestation frontier However, only one estimate of selective logging was [16,49,70,78]. Direct estimates of aboveground biomass available with statewide coverage over multiple years [4]. from LiDAR or Radar remote sensing instruments have [4] estimated that selective logging in Mato Grosso aver- the potential to address these concerns [40-42,79], but aged 9,367 km2 yr-1 during 1999-2002. Excluding logged no direct satellite-based measurements of forest biomass areas that were deforested by 2004, the average annual were available for this study. logged area during 1999-2002 was 6,923 km2 yr-1 [49]. We compared one Tier 1 and five Tier 2 datasets of No satellite-based estimates of understory forest fires or forest carbon stocks in Mato Grosso (Table 2). Tier 1 fuel wood collection were available for Mato Grosso, dataforcarbonstocksintropicalforestsrepresentconti- even for a single year. nental-scale averagesfor each forest type [53,80]. Tier 2 Knowledge of the extent and frequency of land aban- biomass datasets were derived from Amazon forest donment to secondary forest is critical for estimating inventory plots, either from the Brazilian government’s net carbon emissions from deforestation, since carbon RADAMBRASIL survey [54] or a compilation of forest accumulation in secondary forests may partially offset biomass plots from the scientific literature [70,75]. The deforestation carbon losses [16,65]. As in the case of RADAMBRASIL inventory is the most intensive survey forest degradation, few satellite-based estimates of sec- oftimbervolumesinBrazilianforestsconductedtodate, ondary forest extent were available for Mato Grosso. with 440 one-hectare plots in Mato Grosso [54]. Con- Most studies estimated secondary forest area for only verting timber volume into AGLB, including all plants one period in time rather than following the dynamics regardlessoftimberutility,requirestheuseofabiomass of land abandonment and re-clearing of secondary for- conversion and expansion factor [18,73,81]. Similarly, est. Previous estimates of the amount of historic defor- aboveground dead biomass (AGDB) and belowground estation in some stage of forest regrowth varied from biomass(BGB)aretypicallyestimatedusingrelationships 12-17% in Mato Grosso in three studies conducted with among field-measured AGLB, woody debris, and root- satellite data from 2000-2008 [66-68]. Across the entire shoot ratios [70,72,73,82] (Table 2). Data products from Brazilian Amazon, the amount of historic deforestation Houghton and Saatchi et al. were based on forest bio- in some stage of secondary forest regrowth ranged from mass plots from the scientific literature, adjusting for 20-36% over different epochs [56,68,69]. However, a rig- AGDBandBGBinasimilarmanner when thesequanti- orous comparison of secondary forest data products was tieswerenotdirectlymeasured[70,75]. not possible due to differences in the timing of recent Tier 2 biomass maps for Amazonia rely on statistical studies. methods to extrapolate plot-based measurements across