Get 20M+ Full-Text Papers For Less Than $1.50/day. Subscribe now for You or Your Team.

Learn More →

National-scale estimation of gross forest aboveground carbon loss: a case study of the Democratic Republic of the Congo

National-scale estimation of gross forest aboveground carbon loss: a case study of the Democratic... Recent advances in remote sensing enable the mapping and monitoring of carbon stocks without relying on extensive in situ measurements. The Democratic Republic of the Congo (DRC) is among the countries where national forest inventories (NFI) are either non-existent or out of date. Here we demonstrate a method for estimating national-scale gross forest aboveground carbon (AGC) loss and associated uncertainties using remotely sensed-derived forest cover loss and biomass carbon density data. Lidar data were used as a surrogate for NFI plot measurements to estimate carbon stocks and AGC loss based on forest type and activity data derived using time-series multispectral imagery. Specifically, ˆ ´ ´ ´ ´ DRC forest type and loss from the FACET (Forets d’Afrique Centrale Evaluees par Teledetection) product, created using Landsat data, were related to carbon data derived from the Geoscience Laser Altimeter System (GLAS). Validation data for FACET forest area loss were created at a 30-m spatial resolution and compared to the 60-m spatial resolution FACET map. We produced two gross AGC loss estimates for the DRC for the last decade (2000–2010): a map-scale estimate .53:3 9:8 Tg C yr / accounting for whole-pixel classification errors in the 60-m resolution FACET forest cover change product, and a sub-grid estimate .72:1 12:7 Tg C yr / that took into account 60-m cells that experienced partial forest loss. Our sub-grid forest cover and AGC loss estimates, which included smaller-scale forest disturbances, exceed published assessments. Results raise the issue of scale in forest cover change mapping and validation, and subsequent impacts on remotely sensed carbon stock change estimation, particularly for smallholder dominated systems such as the DRC. Keywords: forest cover loss, carbon monitoring, REDD, remote sensing, uncertainty assessment, Congo 1. Introduction compensate developing countries for avoiding emissions due to likely future forest clearing and logging (Houghton 2012) through the emerging REDDC mechanism. The success of The United Nations Reducing Emissions from Deforestation REDDC will be defined by confirmed reductions in rates of and forest Degradation (UN-REDD) program seeks to deforestation and forest degradation. A program requirement is the capability to accurately map and monitor changes in Content from this work may be used under the terms of forest carbon by estimating gross emissions as a function of the Creative Commons Attribution 3.0 licence. Any further area of forest loss and density of carbon stocks within areas of distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. forest loss. 1748-9326/13/044039C14$33.00 1 2013 IOP Publishing Ltd Printed in the UK Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al National forest inventories (NFIs) could provide detailed global-scale carbon stock maps have been created recently and comprehensive information to produce national-scale using the synergy of field measurements, optical, lidar and carbon stock and change estimates. However, NFIs have radar remotely sensed data (Saatchi et al 2011, Baccini not been established in many developing countries that et al 2012). Another approach, presented here, is to calibrate participate in the UN-REDD program (Romijn et al 2012). lidar data using co-located field measurements (Baccini et al The United Nations Food and Agriculture Organization 2012). In this approach, a model is derived to convert lidar (FAO) and the UN-REDD are working on the general waveforms into biomass estimates. The derived model is guidelines for implementing multi-objective NFIs in these then extrapolated to a much larger population of lidar shots, countries (UN-REDD 2011). Meanwhile, alternative methods providing a biomass database for assigning carbon density of national-scale carbon stocks assessment independent of the values to mapped forest cover types. availability of systematically collected ground-based forest For REDDC countries, deforestation is likely to be the inventory data are being investigated and prototyped (GOFC- key category for greenhouse gas emissions estimates. A good GOLD 2010). Goetz et al (2009) provided an overview of the practice for these countries is to use at least IPCC Tier two or satellite-based methods of mapping and monitoring carbon three level assessments for this category of emissions, which stocks, and identified three general approaches: ‘stratify implies reporting uncertainties (Maniatis and Mollicone and multiply’ (SM), when a single carbon density value 2010). AGC stock and loss uncertainty estimates are also is assigned to each land cover type; ‘combine and assign’ crucial if these datasets are to be used as inputs to carbon cycle (CA), extending the SM approach by adding various ancillary and biosphere models. However, published land cover change spatial data layers; and ‘direct remote sensing’ (DR) approach, datasets that may be used as activity data often lack key aimed to derive the carbon stock estimates from machine accuracy assessment information (e.g. description of sampling learning algorithms based on satellite observations and other design, original error matrix, area of each map category, etc) detailed spatial data coupled with field measurements. The that would permit error-adjusted estimates of the change area last approach requires acquisition and processing of large (Olofsson et al 2013). The objectives of our analyses are: (i) to volumes of data to produce a national-scale carbon stock or illustrate the process of activity data accuracy assessment on loss estimate. The first approach, SM, also referred to as the the national level, applicable when using already published ‘biome-average approach’ (Gibbs et al 2007), is relatively land cover data or when creating a new data set, (ii) to easy to implement using a limited set of published data integrate uncertainties from activity and carbon data in a available at low or no cost. Although this approach is fairly national-level forest AGC loss estimate. generalized, in that it does not capture finer-scale spatial In this study, we implemented a SM (‘stratify and heterogeneity of carbon stocks, the accuracy of the estimates multiply’) approach for assessing gross forest AGC loss in can be increased via data refinements and overlays with other the Democratic Republic of the Congo (DRC), where forest data sets in a CA approach. cover change is dominated by smallholder land use and For a national-level aboveground carbon (AGC) loss industrial selective logging (Laporte et al 2007). Due to the assessment, SM approaches require a national-scale land aftermath of two civil wars, persistent political unrest and cover change dataset (activity data in the IPCC terminology lack of infrastructure, the DRC does not collect NFI data IPCC 2006) and mean AGC density estimates for each required for ground-based estimates of AGC stock and its land cover type (IPCC emission factors, here referred to as change. Our approach employs the best available activity and carbon data). Modifying the basic IPCC equation used to carbon data at the national scale—forest extent and loss maps calculate carbon emissions (IPCC 2006, vol 1, chapter 1.2), derived from Landsat imagery (Potapov et al 2012) and AGC the equation to estimate gross AGC loss within a study region estimates derived from GLAS-based canopy vertical structure or a country is the following: metrics (Baccini et al 2012). Results include new estimates of error-adjusted area of forest cover loss between 2000 and 2010, gross AGC loss, and associated uncertainties. AGC loss D 1AD CD (1) i i iD1 2. Data where 1AD (activity data) denotes the change in the extent of a given land cover type i, and CD (carbon data) represents 2.1. Activity data average vegetation carbon content per land cover type. Carbon data that are required for the national-scale AGC loss assessments in an SM approach could be derived from To estimate the area of forest loss, we used Landsat-based field inventory data (e.g. tree DBH and height measurements) year 2000 forest cover and 2000–2010 forest cover loss converted to aboveground biomass using allometric equations data from the Forets ˆ d’Afrique Centrale Evaluees ´ par ´ ´ ´ (e.g. for tropical forests—from Brown 1997 and Chave et al Teledetection (FACET) product, available online (ftp:// congo.iluci.org/FACET/DRC/). FACET data processing and 2005) or existing databases and maps of biomass carbon mapping methodology are described in Potapov et al (2012). density (e.g. Zheng et al 2013, Gibbs 2006, FAO 2010, The FACET dataset provides forest cover and gross forest Malhi et al 2006). Alternatively, biomass carbon content can be mapped using multi-source lidar and radar data that are cover loss for three forest types: primary humid tropical capable to capture vertical tree canopy structure (Goetz and forests, defined as mature humid tropical forest with canopy Dubayah 2011, Treuhaft et al 2009). Several regional and cover >60%; secondary forests, defined as regrowing forest 2 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 1. FACET forest cover and forest cover loss (Potapov et al 2012) combined with DRC wetland map (Bwangoy et al 2010): (a) forested area; (b) woodlands. with canopy cover >60%; and woodlands, defined as forested together with the combined FACET forest cover and DRC areas with canopy cover 30–60%. The spatial resolution of wetland maps to calculate mean AGC density values for the FACET data is 60 m per pixel. We further separated these six target forest classes. Only shots within the forested areas three forest types into terra firma (dryland) and wetland that did not experience forest cover loss between 2000 and sub-classes using the DRC wetland map of Bwangoy et al 2010 according to FACET were used for these calculations. (2010), resulting in six forest types in total. FACET forest The use of a large number of GLAS-estimated biomass cover loss was attributed to these new forest classes (figure 1). values to calculate biome-average AGC densities helps avoid In this manner, the different carbon content of the antecedent biases often inherent in estimates based on the compilation of forest cover could be directly related to disturbance dynamics point-based field measurements (i.e. paucity of sites over large in terra firma and wetland forested ecosystems. In this areas, inadequate stratification to capture variability, and other research, we conduct an explicit statistical validation of factors that limit their spatial representativeness). FACET forest cover loss for each of these forest types and derive the error-adjusted estimate of changed area based on 2.3. Validation data the validation sample. For the purposes of activity data validation, namely the 2.2. Carbon data uncertainty estimation for the FACET forest cover loss, we used all available original L1T Landsat images for years Mean AGC density values for each of the forest types 2000 and 2010 available at no charge from USGS archives were derived from GLAS-based biomass estimates. Baccini (http://glovis.usgs.gov/) and annual Landsat composites for et al (2012) developed a statistical model to predict AGC circa 2000, 2005 and 2010 (Potapov et al 2012). Year 2005 densities observed in the field using GLAS lidar energy composite images helped identify forest cover loss in the early metrics in order to estimate biomass per 65 m diameter 2000s that might be difficult to detect in 2010 Landsat images GLAS shot. The model was based on nearly 300 field sites due to rapid vegetation regeneration in the tropics. located in 12 countries across the tropics. GLAS-predicted In addition to the use of Landsat images for the AGC explained 83% of variance in the field-measured carbon validation (reference) classification, we also employed visual density at the GLAS-footprint scale with a standard error of interpretation of very high spatial resolution images available 1 TM 22:6 Mg C ha (Baccini et al 2012). For this study, we for the study region through Google Earth and through employed the GLAS-derived biomass data as if they were field a partnership between NASA and NGA that provides inventory data and did not incorporate this model uncertainty access to unclassified commercial high spatial resolution in downstream calculations. After screening GLAS data for satellite data from NGA archives for NASA Earth Science noise and filtering for slope (10 ), 371 458 AGC-estimated Investigators (http://cad4nasa.gsfc.nasa.gov/). A total of 1689 GLAS shots for the years 2004–2008 (figure 2) were analyzed high resolution images from multispectral and panchromatic 3 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 2. 2004–2008 GLAS shots color-coded by the FACET forest type (Potapov et al 2012) combined with wetland map (Bwangoy et al 2010). sensors (Ikonos, WorldView-1, WorldView-2, Quickbird, by this objective, as well as by feasibility issues and time constraints. Orbview-5) for 2008–2011 time interval were used for the visual assessment of validation samples. In total, 503 out of a final 1061 validation samples had at least one matching 3.1.1. Sampling design and sample size. The target high resolution image available between 2000 and 2013, activity data class, forest cover loss, is relatively small TM either from Google Earth or from the NGA archive. These compared with the unchanged forest areas; the sampling images facilitated the forest cover loss validation, providing design should increase the sample representation of this rare information about forest cover type on date 1 (2000) or date 2 class in order to achieve a precise estimate of forest cover (2010). loss accuracy (Khorram 1999). Moreover, our objective is forest type-specific loss area estimation and its accuracy; stratified random sampling is an appropriate choice in this 3. Methods case (Stehman 2009). Initially, two strata within each forest type class were 3.1. Uncertainties from activity data considered: ‘no loss’ (forests, undisturbed between 2000 and 2010) and ‘loss’ (2000–2010 forest cover loss). However, The key objective of activity data validation is to estimate sufficient estimation of loss omission error within the large error-adjusted area of forest cover loss for each forest type and ‘no loss’ stratum requires special attention. Given a simple to quantify its uncertainty. Error-adjusted area estimation uses ‘loss’ and ‘no loss’ stratification, rates of false negatives validation sample data to adjust area of forest cover loss due to (change omission errors) could be poorly characterized classification errors (including omission errors and excluding (Khorram 1999). Furthermore, the FACET national-scale commission errors) present in the map product (Olofsson forest cover loss product is likely to be conservative, et al 2013). The choice of sampling design is determined i.e. omitting forest cover loss in comparison to committing 4 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 1. Distribution of samples among forest types using proportional and arbitrary sample allocation strategies for stratified random sampling. Proportional allocation (% samples) Forest type Based on forest area Based on loss area Arbitrary allocation (% samples) Primary forest 46 25 33 Secondary forest 11 55 17 Woodlands 21 13 25 Wetland primary forest 19 3 17 Wetland secondary forest 1 3 4 Wetland woodlands 2 1 4 forest loss. To address this issue we identified an additional Table 2. Allocation of sample size among validation strata. ‘probable loss’ stratum within each forest type class. This Forest type No loss Probable loss Loss Total stratum was constructed to target omitted forest cover loss Primary forest 200 70 63 333 in order to improve the loss area estimate for the AGC loss Secondary forest 30 87 50 167 calculation. We define the ‘probable loss’ stratum as a 1-km Woodlands 100 90 60 250 radius circular region around forest cover loss, assuming that Wetland primary 80 30 57 167 omission of loss is likely to occur in proximity to mapped forest loss. The choice of the 1-km wide ‘probable loss’ stratum Wetland 15 15 12 42 secondary forest is supported by the evidence that increased tree mortality in Wetland 15 15 12 42 temperate and tropical forests is generally observed up to 1 km woodlands from the forest edge (Broadbent et al 2008). A total of 18 strata were analyzed: ‘loss’, ‘probable loss’, and ‘no loss’ for each of the six forest types (terra firma and loss strata, we chose to have an allocation closer to equal, wetland primary forests, terra firma and wetland secondary which helped to target errors of commission (Stehman 2012) forests, terra firma and wetland woodlands). Allocation of among the ‘no loss’, ‘probable loss’ and ‘loss’ strata. A total samples among these strata should effectively address our sample size of 1000 was projected as feasible to be visually validation objective (see section 3.1) of minimizing standard interpreted by expert analysts. We imposed the condition that errors (SEs) of error-adjusted estimators of forest cover loss a sample size greater than 50 was required for the major area (Stehman 2012). forest types (primary, secondary forests, woodlands, wetland When considering allocation of samples among forest primary forests), the allocation of sample size per stratum (the types, we examined both the area of forest type and the sampling unit is one 60-m FACET pixel) was implemented as area of our target class (forest loss) within each forest type. shown in table 2. Proportional allocation of samples among forest types based For the chosen sample allocation we calculated SEs of on the forest type area would lead to small sample sizes from the estimated area of change using hypothetical omission secondary forest, woodlands and wetland forests: almost half and commission error rates in order to confirm that the of all samples in this case fall into the dense forest class chosen allocation would not lead to inflated standard errors. (table 1). Although forest cover loss in dense forests that have We compared our arbitrary allocation to proportional among high biodiversity and other high-value ecosystem services forest allocation with equal and proportional allocation among is important to estimate correctly, the majority of mapped loss strata and found that the arbitrary allocation performed forest cover loss occurred in secondary forests. However, as well or better than the other options. The equation allocation of samples based on the forest cover loss area used to calculate SEs of the estimated area of change for leads to the majority of samples being located in secondary each forest type is similar to equation (3) from Olofsson forests. In order to find a compromise between preserving et al (2013). However, after the assignment of reference a sufficient number of samples in the strategically important values to the samples during expert validation, we found dense forest class while adequately representing the relatively out that the ‘probable loss’ stratum contributed 35% of the small classes with high proportional forest cover change total variance in primary forest, 50% of the variance in (secondary forest, woodlands), we implemented an arbitrary secondary forest, and 20% of the variance in woodlands. allocation that was close to proportional by forest type area, Additional random samples were added to the ‘probable but adjusted for forest loss area (table 1). loss’ stratum of terra firma primary, secondary forests and The sample size allocation to the three strata within each woodlands (20, 30 and 10 samples respectively) in order to forest type was determined as follows. Because it is equally minimize the total SE of the loss area estimate. important for our primary validation objective (estimation of forest loss area for each forest type based on an error 3.1.2. Estimating area of forest loss and its uncertainty. matrix) to account for committed and omitted loss area, we addressed the need to account for omission errors by Visual interpretation of validation samples was performed at a creating the separate ‘probable loss’ strata within the original 30-m spatial resolution, enabling map-scale and sub-grid error ‘no loss’ class. Therefore, when allocating samples among assessments (FACET was made at a 60 m spatial resolution 5 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 3. Example of sample block visual interpretation; for the map-scale estimate, 0.5 loss is treated as no loss. The black stripe in the 2010 Landsat loss sample is a data gap due to the Landsat 7 scan-line corrector malfunction. Table 3. Error matrix of sample counts for map-scale and sub-grid area estimates. Reference strata Map-scale estimate Sub-grid estimate Forest type Map strata No loss Loss No loss Loss N of pixels in each stratum Primary forest No loss 200 0 200 0 147 647 298 No loss–probable loss 89 1 86.5 3.5 56 158 987 Loss 3 60 3 60 2 638 342 Secondary forest No loss 30 0 30 0 5 720 568 No loss–probable loss 107 10 98.5 18.5 35 535 337 Loss 00–10 3 47 3 47 5 619 034 Woodlands No loss 100 0 100 0 51 491 436 No loss–probable loss 98 2 97 3 39 725 284 Loss 00–10 7 53 7 53 1 374 079 Wetland primary forest No loss 80 0 80 0 67 675 696 No loss–probable loss 30 0 30 0 15 706 036 Loss 00–10 9 48 9 48 326 316 Wetland secondary forest No loss 15 0 15 0 1 506 946 No loss–probable loss 15 0 14.5 0.5 2 176 786 Loss 00–10 4 8 4 8 255 498 Wetland woodlands No loss 15 0 15 0 7 003 885 No loss–probable loss 15 0 15 0 2 477 979 Loss 00–10 2 10 2 10 97 176 using resampled 30-m Landsat time-series imagery). We a 60-m validation pixel as ‘loss’ only if the reference forest produced two forest loss area estimates for the DRC for the loss fraction detected using 30-m Landsat and/or high spatial last decade (2000–2010): a map-scale estimate accounting resolution was75% of pixel area. For the sub-grid estimate, for whole-pixel classification errors in the 60-m resolution three gradations of reference loss fraction per pixel were used: FACET forest cover change product, and a sub-grid estimate 1 (loss) with reference loss 75% of pixel area; 0.5 (mixed that took into account 60-m cells that experienced partial pixels) with reference loss between 75% and 25%; and 0 (no forest loss (table 3). For the map-scale estimate we treated loss) otherwise (figure 3). 6 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 4. Parameters for the calculation of error-adjusted area of forest cover loss within terra firma primary forests (map-scale estimate). Primary forest y n yN N Map area (ha) s u h h h u2h yh No loss 0 200 0=200 147 647 298 53 153 027 0.000 000 000 No loss–probable loss 1 90 1=90 56 158 987 20 217 235 0.011 111 111 Loss 60 63 60=63 2 638 342 949 803 0.046 082 949 Total 206 444 627 74 320 066 90 0:011 111 111 When the sampling strata and map classes being validated C 56 158 987 1 are the same, equations (2)–(4) from Olofsson et al (2013) 56 158 987 90 should be used to calculate error-adjusted area of forest cover 63 0:046 082 949 loss and its standard error based on a validation confusion C 2638 342 1 2638 342 63 matrix. In our case, there was a mismatch between sampling 1=2 strata (‘no loss’, ‘probable loss’, ‘loss’) and map classes 2 1 .206 444 627 / (‘loss’ and ‘no loss’) within each forest cover type arising from the attempt to target omitted forest cover loss by creating D 226 099 ha (6) the additional ‘probable loss’ stratum. Based on sampling theory (Cochran 1977), the following equation was employed A D 1129 210 443 156 ha: (7) to produce an unbiased estimator of the area of forest cover loss within each of the forest cover types when validation strata and map classes do not match (Stehman 2013, in review): 3.2. Uncertainties from carbon data N yN h h hD1 A D A  (2) tot N Table 5 presents the mean and population standard deviation (STD) derived from the number of GLAS shots per forest where A —total area of the forest cover type; tot type. Using the SM (‘stratify and multiply’) approach we y D 0:5 or 1 if pixel u (or it’s half) is in reference class ‘forest assigned a single mean AGC density value to each of the cover loss’, and y D 0 otherwise; u forest type classes to estimate gross AGC loss. To quantify u2h yN D , the sample-mean of the y values in stratum h; h u the uncertainty of this estimate, we employed the standard n —sample size in stratum h; deviation of the sample-mean’s estimate of a population mean, N —number of pixels in stratum h; the standard error of the mean (SEM). According to the central N—total number of pixels within the forest cover type. limit theorem, the distribution of sample estimates of the The standard error of the error-adjusted estimate of the mean is normally distributed, enabling us to calculate the 95% forest cover loss is: confidence interval (CI) of mean AGC density estimates as P s 1.96SEM. Table 5 shows mean AGC densities of our target H n yh 2 h N 1 hD1 h N n h h forests classes along with their 95% CIs. SE.A/ D A (3) tot .y yN / 2 h 3.3. Combination of the uncertainties u2h where s D , the sample variance for stratum h. yh n 1 A 95% confidence interval (assuming normal distribu- When calculating AGC loss for each forest type using tion) is: equation (1), uncertainty comes both from activity data (in O O A 1:96SE.A/: (4) our case—forest cover loss) and emission factors (carbon data). In order to combine uncertainties from these quantities, An example of the forest cover loss area estimation for terra the multiplication approach from the recent IPCC Guidelines firma primary forests (map-scale estimate) is presented in for National Greenhouse Gas Inventories (IPCC 2006, vol 1, table 4 and equations (5)–(7). chapter 3, p 28, equation (3.1)) was used: A D 74 320 065:72.0 147 647 298C  56 158 987 90 q 2 2 2 60 1 U D U C U CC U (8) total C  2 638 342/.206 444 627/ 1 2 where U is the percentage uncertainty in the product of total D 1129 210 ha (5) the quantities (half the 95% confidence interval divided by the SE.A/ D 74 320 065:72 total and expressed as a percentage). U is the percentage uncertainties associated with each of 200 0:0 147 647 298 1 the quantities. 147 647 298 200 7 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 5. GLAS-based AGC density estimates for the DRC forest types. Mean AGC densities are given with95% CI. Forest type Mean AGC density (Mg C ha ) Number of GLAS samples STD Primary forest 156.8 0.4 115 566 67.03 Secondary forest 94.8 0.7 31 443 67.45 Woodlands 71.2 0.2 121 671 44.24 Wetland primary forest 128.9 0.4 85 923 55.29 Wetland secondary forest 90.7 2.3 3 148 65.83 Wetland woodlands 66.5 0.8 13 707 45.81 Table 6. Original FACET and error-adjusted estimates of 2000–2010 forest cover loss within DRC forest types (95% CI). 2000–2010 forest cover loss (ha) Error-adjusted Forest type Map-scale estimate Sub-grid estimate FACET map Primary forest 1 129 210 443 156 1 690 800 645 694 949 803 Secondary forest 2 994 876 664 625 3 924 262 736 673 2022 852 Woodlands 722 979 396 475 865 990 439 210 494 668 Wetland primary forest 98 925 11 218 98 925 11 218 117 474 Wetland secondary forest 87 440 78 014 87 441 78 014 91 979 Wetland woodlands 29 153 7704 29 153 7704 34 983 For example, for the primary forest stratum, the 4. Results calculation of the U (using the map-scale 1AD estimate) total is the following: Applying the approach of adjustment for the classification errors described in section 3, we produced estimates of u ! SE.A/ AGC SEM forest cover loss within target DRC forest classes (table 6). U D  100 C  100 total Mean AGC Error-adjustment significantly increased estimated areas of forest loss in terra firma forest classes (primary, secondary 2 2 forests and woodlands); omission errors prevailed over 226 099:75 0:2 D  100 C  100 commission errors (figure 4). In the wetland forests and 1129 210 156:83 woodlands, on the contrary, more loss was committed in D 20:02%: (9) the map product; error-adjusted loss area estimates were smaller than those prior to adjustment. SE was highest in When calculating total gross AGC loss within the DRC wetland secondary forests and terra firma woodlands. High (summing AGC loss values for all forest types), the addition uncertainty in the wetland secondary forests is associated and subtraction approach from the IPCC Guidelines (IPCC with it being the smallest and spatially discontinuous class. 2006, vol 1, chapter 3, p 28, equation (3.2)) was used to Woodland is a challenging forest type to map and monitor due estimate the uncertainty of the resulting quantity: to the gradients of tree canopy cover and seasonality as well total DRC as the comparatively uneven intensity of disturbance events, 2 2 2 all of which contributes to larger SEs. .U x / C .U x / CC .U x / 1 1 2 2 n n To compare AGC density estimates for our target forest jx C x CC x j 1 2 n classes with published estimates, we calculated average AGC (10) densities within the 6 DRC forest types using available where U is the percentage uncertainty in the sum of the spatially explicit vegetation carbon density products (Baccini total quantities (half the 95% confidence interval divided by the et al 2012, Saatchi et al 2011, Gibbs and Brown 2007, total and expressed as percentage); Kindermann et al 2008) and compared them with the x and U are the uncertain quantities and percentage i i GLAS-based estimates of the current study (figure 5). This uncertainties associated with them. comparison provides a general understanding of how well Thus, the overall uncertainty of gross AGC loss estimate our current estimates correspond to existing knowledge. for the entire DRC is: Examination of figure 5 shows that GLAS-based AGC density estimates are generally higher than those modeled using total DRC optical remotely sensed data (Baccini et al 2012, Saatchi et al 2 2 2 .U 1AGC / C.U 1AGC / CC.U 1AGC / total1 1 total2 2 totaln n D 2011, Gibbs and Brown 2007), probably because of spatial j1AGC C1AGC CC1AGC j 1 2 averaging (Goetz and Dubayah 2011, Zolkos et al 2013), but (11) don not exceed the estimates of Kindermann et al (2008) who where numbers (1–n) stand for the six forest cover types. employed FAO 2005 Forest Resources Assessment statistics. 8 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 4. Forest cover loss (2000–2010) within DRC forest types; error bars are the 95% CIs. Figure 5. Comparison of the AGC density estimates from the published datasets (error bars are the 95% CIs) and the current study. Table 7. Gross AGC loss estimates (2000–2010) with the uncertainty measures for DRC forest types ( is the 95% CI). Map-scale loss area estimate Sub-grid loss area estimate Forest type U (%) Gross AGC loss 2000–2010 (Pg C) U (%) Gross AGC loss 2000–2010 (Pg C) total total Primary forest 20.0 0.177 0.070 19.5 0.265 0.101 Secondary forest 11.3 0.284 0.063 9.6 0.372 0.070 Woodlands 28.0 0.051 0.028 25.9 0.062 0.031 Wetland primary forest 5.8 0.013 0.001 5.8 0.013 0.001 Wetland secondary forest 45.5 0.006 0.005 45.5 0.008 0.007 Wetland woodlands 13.5 0.002 0.001 13.5 0.002 0.001 DRC total 9.4 0.533 0.098 9.0 0.721 0.127 Sub-grid gross AGC loss estimates were 20–50% higher Differences between these estimates are mostly associated than map-scale ones for the major terra firma forests (primary, with the ‘loss’ and ‘probable loss’ strata, particularly in secondary forests and woodlands) and nearly equal for the regions where primary and secondary forest loss predominate. less widespread wetland forests (table 7, figures 6(b) and (c)). There are no significant differences in the forests and 9 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 6. Forest type and strata averages, aggregated to a 5-km grid: (a) year 2000 AGC; (b) map-scale estimate of 2000–2010 gross AGC loss; (c) sub-grid estimate of 2000–2010 AGC loss; (d) difference between sub-grid and map-scale estimates. Water bodies are shown in gray. Note that AGC values for both (b) and (c) are the same for the respective forest types. woodlands of the ‘no loss’ strata (figure 6(d)). For the whole confidence interval of the global sample-based estimate of the DRC, the sub-grid AGC loss estimate was 35% higher of Hansen et al (2010), but is significantly higher than than the map-scale estimate (table 7). the FACET map-based estimate without error-adjustment The comparison of gross forest cover loss and gross AGC (Potapov et al 2012). The sub-grid estimate, accounting for rates from this study with published estimates is presented in the finer-scale forest disturbance, is 30–40% higher than table 8. We report annual forest cover loss rates separately for published estimates for the DRC, and points to the difficulty primary and secondary forests, excluding woodlands (table 8) of mapping forest change in a landscape where smallholder to best match the definition of forests employed in the shifting cultivation predominates. For example, FACET forest most recent regional sample-based forest cover loss estimate cover loss has a mean patch area of 1.4 ha (Potapov et al by Ernst et al (2013) (all tropical moist forests, excluding 2012). While patch size is not the same as field size, it woodland savannahs and tropical dry forests). is worth noting that typical shifting cultivation practices in the tropics employ field sizes well under 1 ha (Aweto 5. Discussion 2013). The quantification of such change is challenging and represented by the comparatively large presence of The results reported in table 8 need to be considered in the mixed pixels in the FACET data. The difference of two context of inconsistencies in methodologies, definitions, and methodologically consistent loss area estimates based on areas of analysis (a direct consequence of the differences input data of different resolutions (60-m FACET and 30-m in the definitions of forest and woodlands). Our map-scale Hansen et al 2013, table 8) prior to error-adjustment illustrates 2000–2010 annual forest cover loss estimate within dense the issue: the 30-m product depicts 1.5 times more change forests (0.35%  0.03%) agrees well with the estimates than the 60-m one. Any binary (yes/no) change map will have of Ernst et al (2013) for the first half of the decade (0.32%  0.05%) and of Hansen et al (2013) for 2000–2012 scale-dependent omission errors. These ‘cryptic disturbances’ (0.34%). Our map-scale estimate also falls within the have been reported to add more than 50% of forest cover loss 10 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 8. Comparison of forest cover and carbon loss estimates for the DRC (95% CI). 2000–2005 2005–2010 Source Extent Annual gross forest cover loss (% of the forest area) Current study Map-scale ForestsC woodlands 0.32% 0.03% Sub-grid ForestsC woodlands 0.42% 0.03% Map-scale Forests 0.35% 0.03% Sub-grid Forests 0.47% 0.04% FACET map Potapov et al (2012)—60 m ForestsC woodlands 0.22% 0.25% Hansen et al (2013)—30 m ForestsC woodlands 0.34% Ernst et al (2013) Forests 0.32% 0.05% — Hansen et al (2010) ForestsC woodlands 0.12% 0.23% — Annual net forest cover loss (% of the forest area) FAO (2010) ForestsC woodlands 0.20% 0.20% Ernst et al (2013) Forests 0.22% 0.22% Annual gross AGC loss (Tg C yr ) Current study Map-scale ForestsC woodlands 53.3 9.8 Sub-grid ForestsC woodlands 72.1 12.7 Annual gross carbon loss (Tg C yr ) Harris et al (2012) ForestsC woodlands 23 — TM to existing Landsat-scale forest disturbance classifications for Google Earth high resolution imagery) allows defining the Amazon Basin (Asner et al 2005). reference values of validation samples without in situ measurements. Despite its advantages, the method is sensitive Table 8 reflects a second type of omission error related to sampling design and the associated decision of how to to algorithmic and/or data limitations. Estimates of forest loss allocate the sample size among validation strata. For the strata derived at a 30-m spatial resolution, particularly the Hansen and sample size allocation implemented in this study, the et al (2013) and Ernst et al (2013) products, have comparable decisions were advantageous; for the four largest forest types, gross forest cover loss rates, 0.34% and 0.32%  0.05%. However, the 30-m validation estimate is 0.47%  0.04%. the reduction in standard error attributable to the stratification Large area mapping algorithms are often conservatively was substantial. Specifically, the gain in precision due to implemented in attempting to avoid commission error. For stratification can be computed from the sample data (Cochran validation, the determination of loss/no loss is performed 1977, section 5A.11) as the ratio of the standard error that independently per sample and is free of this consideration. would have been obtained from simple random sampling Differences between the Hansen et al (2013) 30-m map and to the standard error obtained from the stratified design the Ernst et al (2013) 30-m sample estimates could be due implemented (same sample size for both designs). For the four to this fact. However, the estimate of Ernst et al (2013) was largest forest types, these ratios were 1.42 for primary forest, also sample based. The additional loss found in our validation 1.10 for secondary forest, 1.32 for woodlands, and 23.21 for effort compared to Ernst et al (2013), while partially due to wetland primary forest (the latter estimate is likely inflated by the use of very high spatial resolution data for a portion of the the fact that two of the three strata had 0% forest loss). The methodology is also highly dependent on the knowledge base reference samples, is not easily explained and may be more of the remote sensing experts performing visual interpretation related to definitional differences or other methodological factors. In summary, the difference between the 60-m FACET of validation samples. Finally, it is a function of the quality of loss rates of 0.22% and 0.25% and the 30-m loss rates of the reference imagery and the resulting clarity or conversely 0.34% and 0.32% is most likely related to the differing scales ambiguity in assigning change per validation sample. The of measurement. The difference between the 30-m loss rates map-scale and sub-grid estimates reflect the importance of this of 0.34% and 0.32% and the validation rate of 0.47% is most issue. likely related to limitations in mapping versus sampling or A further consideration in assessing the results concerns to other methodological factors. The discrepancy between the reference data and the potential volatility of the sample- map-scale and sub-grid estimates emphasizes the issue of based estimate itself. Table 4 illustrates this issue. The ‘loss’ scale in change area estimation for smallholder dominated stratum records 60 of 63 samples as having experienced terra landscapes like the DRC. firma primary forest cover loss, representing 905 574 ha of The approach for validating activity data employed in error-adjusted forest loss area. For the ‘probable loss’ stratum, this study is relatively straightforward and easy to implement. 1 of 90 samples was interpreted as having experienced forest The method allows for the generation of error-adjusted loss cover loss. Due to the much larger size of this stratum, this one area estimates from the existing land cover and vegetation sample accounts for an estimated 224 635 ha of error-adjusted maps. This approach does not require large volumes of data forest loss area, or fully 20% of terra firma primary forest processing and is therefore not limited by computational cover loss. Without the use of the ‘probable loss’ stratum facilities. The use of open access medium- and high resolution and the inclusion of this single sample of commission error, imagery for map product validation (USGS Landsat archive, results would indicate a slight underestimate of terra firma 11 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al forest cover loss. Validation studies should formally consider 2008). In total, GLAS-based AGC models explain from 73% likely regions of false negatives of forest change in developing (Lefsky et al 2005, Pflugmacher et al 2008) to 83% (current stratified sampling methods for error-adjusted area estimation. study; Baccini et al 2012) of the variance in field-estimated The validity of the sample-based estimate is a function of biomass. Regional forest inventory data are required to calibrate and validate the current forest type GLAS-based many factors, including the vagaries of any individual sample estimates. Additional field data collection could further refine data set used in creating the error-adjusted estimates. the estimates but, unfortunately, GLAS observations are Estimates of carbon density derived using different not available after 2009, posing a near-term challenge for methods can vary considerably within the same region improved AGC mapping and monitoring beyond the current (Houghton et al 2001), introducing uncertainty to the carbon models. As part of the process of establishing an NFI for loss estimation. However, recent published estimates of the DRC continues, other sources of remotely sensed data carbon loss from deforestation differ primarily due to major characterizing vegetation vertical structure, such as airborne disagreements in the quantification of the areal extent of forest lidar or spaceborne radar data, can bridge the gap until cover loss (Pan et al 2011, Harris et al 2012). The DRC gross systematic spaceborne lidar measurements become available AGC loss estimates from the current study (map-scale and to the scientific and REDDC implementation communities. sub-grid) are 2 to 3 times higher than the biomass carbon loss (total carbon, above- and belowground) estimate of Harris et al (2012) (table 8) due primarily to differences in the 6. Conclusion estimated area of forest cover loss. The Harris et al (2012) estimate is based on a global forest cover loss product by We applied a method of error-adjustment of forest cover Hansen et al (2010) that is highly uncertain in the DRC loss area to produce a national-scale gross forest AGC loss (SE D 100%, see table 8). Hansen et al (2010) employed estimate for the DRC based on a published forest cover loss a pan-tropical MODIS-based stratification to target sample dataset. We employed field-calibrated GLAS lidar-derived allocation with only 7 samples located in the DRC. The biomass carbon densities as a substitute for NFI data, which small sample size resulted in a high standard error (table 8). do not exist for the territory of the DRC. Two realizations of Harris et al (2012) reported a 90% carbon loss prediction the resulting DRC gross AGC loss estimate, map-scale and interval for the DRC, based on a Monte Carlo approach: sub-grid, were produced. The sub-grid AGC loss estimate 16–32 Tg C yr ; our current DRC gross AGC loss estimates, accounted for disturbances finer than the map grid scale of 1 1 map-scale (53:3 Tg C yr ) and sub-grid (72:1 Tg C yr ), 60 m and was higher than published estimates, highlighting are not within this interval. issues of scale and spatial averaging in AGC estimation. In our analysis, DRC gross forest AGC loss assessment Omitted disturbances were largely related to smallholder consists only of stand-replacement forest disturbance that agriculture land cover change, the detection of which is scale- can be observed at the mapping scale and in reference data. dependent. For the FACET product, the input Landsat imagery However, forest degradation processes that do not lead to the were averaged to 60 m and then classified, leading to the complete loss of tree canopy or cause small-scale canopy estimated scale-related omission error. Other processing steps openings, and can be detected only in the field or using can lead to change omission, either through the algorithm dense series of sub-meter remotely sensed data may result itself, for example image segmentation, post-processing of in significant AGC loss at the national scale (IPCC 2003, the output classification, or the application of a minimum Schoene et al 2007). One possible approach to assess the mapping unit. In Brazil, where agro-industrial land conversion loss of biomass from these disturbances could be based on results in large forest disturbances, the Brazilian Space monitoring changes in the area of intact forest landscapes Agency’s PRODES product 6.25 ha minimum mapping (Potapov et al 2008) and assigning an AGC loss value to the unit (the equivalent of approximately 69 Landsat pixels) forests that have undergone the transition from intact primary (INPE 2012), provides a viable deforestation monitoring to primary degraded and secondary forests (Margono et al approach. However, a 6.25 ha minimum mapping unit 2012, Zhuravleva et al 2013). For countries such as the DRC, for the DRC would omit the majority of change. For where large-scale agro-industrial forest disturbance is largely heterogeneous landscapes with change dynamics at or finer absent, the question of scale and its impact on AGC loss due than the resolution of Landsat data, higher spatial resolution to deforestation and degradation remains an important line of imagery to directly map such changes, or indirect methods scientific inquiry. to delimit degraded areas and subsequently relate to in situ We employed GLAS-based AGC estimates as a proxy for measurements, are required. the ground-based NFI data. There are some known issues and Our study also illustrates the importance of reference limitations concerning the estimation of biomass from GLAS forest state in assessing carbon dynamics, as with the metrics. For example, GLAS-estimated vegetation heights primary, secondary and woodland forest types presented here. often used in AGC models have on average 2–3-m error The Brazilian PRODES product, the current standard for when compared with USDA Forest Inventory and Analysis national-scale forest monitoring, quantifies only the loss of (FIA) and other field-measured heights (Pflugmacher et al primary forest in the Legal Amazon. While reducing primary 2008, Lefsky et al 2005, Sun et al 2008). GLAS-derived humid tropical forest loss is the main focus of climate biomass estimates are also known to be affected by the mitigation strategies such as REDDC, other forest types and season of data acquisition and terrain slope (Sun et al even trees outside of forests will be part of national carbon 12 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al accounting schemes. Our study underscored the importance Brown S 1997 Estimating biomass and biomass change of tropical forests: a primer FAO Forestry Paper 134 (Rome: Food and of monitoring other forest dynamics, as we found AGC loss Agriculture Organisation of the UN) in secondary forests to be 140% that of primary forests. The Bwangoy J B, Hansen M C, Roy D P, Grandi G De and Justice C O reuse of secondary forests remains a challenge to carbon 2010 Wetland mapping in the Congo Basin using optical and monitoring and the development of appropriate strategies radar remotely sensed data and derived topographical indices for reducing emissions, but monitoring all relevant forest Remote Sens. Environ. 114 73–86 Chave J et al 2005 Tree allometry and improved estimation of types and dynamics is required as national-scale programs are carbon stocks and balance in tropical forests Oecologia developed and implemented. 145 87–99 REDDC mechanisms will rely on accurate mapping Cochran W G 1977 Sampling Techniques (New York: Wiley) and monitoring of AGC (Houghton et al 2010). However, Ernst C, Mayaux P, Verhegghen A, Bodart C, Christophe M and scientific, technical and operational aspects of AGC mapping Defourny P 2013 National forest cover change in Congo Basin: deforestation, reforestation, degradation and regeneration for and monitoring are still in their infancy. Results from this the years 1990, 2000 and 2005 Glob. Change Biol. 19 1173–87 study have significant implications for policy initiatives FAO 2010 Global forest resources assessment 2010 FAO Forestry like REDDC. It is clear that the spatial scale of forest Paper 163 (Rome: Food and Agriculture Organization of the change characterization, reference information on forest type UN) and carbon stock, and sample representativeness, can all Gibbs H K 2006 Major World Ecosytem Complexes Ranked by Carbon in Live Vegetation: An Updated Database Using the dramatically impact AGC loss estimation. For example, GLC2000 Land Cover Product NDP-017b considering change at a 30 m validation scale, an extra 35% Gibbs H K and Brown S 2007 Geographical distribution of woody of AGC loss was estimated compared to the 60-m spatial biomass carbon in tropical africa: an updated database for 2000 scale; terra firma secondary forest cover loss accounted for NDP-055b 40% more AGC loss than that of terra firma primary forest Gibbs H K, Brown S, Niles J O and Foley J A 2007 Monitoring and estimating tropical forest carbon stocks: making REDD a loss; a single validation sample added 20% to map-scale reality Environ. Res. Lett. 2 045023 terra firma primary forest cover loss area. The volatility of Goetz S J, Baccini A, Laporte N T, Johns T, Walker W, results within this study indicates the DRC to be a challenging Kellndorfer J, Houghton R A and Sun M 2009 Mapping and environment for quantifying changes to forest carbon stocks, monitoring carbon stocks with satellite observations: a with implications for other countries as well. Eventual comparison of methods Carbon Balance Manag. 4 2 Goetz S J and Dubayah R 2011 Advances in remote sensing national monitoring systems will need to demonstrate technology and implications for measuring and monitoring spatio-temporal consistency given the various factors that forest carbon stocks and change Carbon Manag. 2 231–44 impact AGC loss estimation. While absolute accuracies GOFC-GOLD 2010 A sourcebook of methods and procedures for may differ due to some of the aforementioned factors, monitoring and reporting anthropogenic greenhouse gas relative consistency for any particular set of observations emissions and removals caused by deforestation, gains and losses of carbon stocks in forests remaining forests, and and spatial scale should be achievable and implementable. forestation Report Version COP16-1 (Alberta: GOFC-GOLD Demonstrating such consistency will be a proof of readiness Project Office, Natural Resources Canada) for REDDC monitoring. Hansen M C et al 2013 The first high-resolution global maps of 21st century forest cover change at press Hansen M C, Stehman S V and Potapov P V 2010 Quantification of Acknowledgments global gross forest cover loss Proc. Natl Acad. Sci. 107 8650–5 Harris N L, Brown S, Hagen S C, Saatchi S S, Petrova S, Salas W, Support for this study was provided by NASA’s Terrestrial Hansen M C, Potapov P V and Lotsch A 2012 Baseline map of Ecology program through grant number NNX12AB43G, carbon emissions from deforestation in tropical regions Science 336 1573–6 NASA Applied Sciences grant NNX12AL27G, and by the Houghton R A 2012 Carbon emissions and the drivers of United States Agency for International Development through deforestation and forest degradation in the tropics Curr. Opin. its CARPE program. Thanks to Brian Barker, Giuseppe Environ. Sustain. 4 597–603 Molinario and Alice Altstatt from the UMD CARPE team for Houghton R A, Greenglass N, Baccini A, Cattaneo A, Goetz S, the processing of high resolution remotely sensed data and Kellndorfer J, Laporte N and Walker W 2010 The role of valuable comments. science in reducing emissions from deforestation and forest degradation (REDD) Carbon Manag. 1 253–9 Houghton R A, Lawrence K T, Hackler J L and Brown S 2001 The References spatial distribution of forest biomass in the Brazilian Amazon: a comparison of estimates Glob. Change Biol. 7 731–46 Asner G P, Knapp D E, Broadbent E N, Oliveira P J C, Keller M Instituto Nacional de Pesquisas Especias 2012 Monitoring of the and Silva J N 2005 Selective logging in the Brazilian Amazon Brazilian Amazonian Forest by Satellite, 2000–2012 Science 310 480–2 IPCC 2003 Definitions and Methodological Options to Inventory Aweto A O 2013 Shifting Cultivation and Secondary Succession in Emissions from Direct Human-induced Degradation of Forests the Tropics (Wallingford, UK: CABI) and Devegetation of Other Vegetation Types (Hayama, Japan: Baccini A et al 2012 Estimated carbon dioxide emissions from The Institute for Global Environmental Strategies) tropical deforestation improved by carbon-density maps IPCC 2006 2006 IPCC Guidelines for National Greenhouse Gas Nature Clim. Change 2 182–5 Inventories ed H S Eggleston, L Buendia, K Miwa, Broadbent E, Asner G, Keller M, Knapp D, Oliveira P and Silva J T Ngara and K Tanabe (Hayama, Japan: IGES) 2008 Forest fragmentation and edge effects from deforestation Khorram S 1999 Accuracy Assessment of Remote Sensing-Derived and selective logging in the Brazilian Amazon Biol. Conserv. Change Detection ed S Khorram (Bethesda, MD: American 141 1745–57 Society for Photogrammetry and Remote Sensing) 13 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Kindermann G E, Mccallum I, Fritz S and Obersteiner M 2008 Saatchi S S, Harris N L, Brown S, Lefsky M, Mitchard E T A and A global forest growing stock, biomass and carbon map based Salas W 2011 Benchmark map of forest carbon stocks in on FAO statistics Silva Fenn. 42 387–96 tropical regions across three continents Proc. Natl Acad. Sci. Laporte N T, Stabach J A, Grosch R, Lin T S and Goetz S J 2007 108 9899–904 Expansion of industrial logging in central Africa Science Schoene D, Killmann W, Von Luepke H and LoycheWilkie M 2007 316 1451 Definitional issues related to reducing emissions from Lefsky M A, Harding D J, Keller M, Cohen W B, Carabajal C C, deforestation in developing countries Forest and Climate Espirito-Santo F D B, Hunter M O and de Olivera R Jr 2005 Change Working Paper 5 (Rome: Food and Agriculture Estimates of forest canopy height and aboveground biomass Organization of the UN) using ICESat Geog. Res. Lett. 32 L22S02 Stehman S V 2009 Sampling designs for accuracy assessment of Malhi Y et al 2006 The regional variation of aboveground live land cover Int. J. Remote Sens. 30 5243–72 biomass in old-growth Amazonian forests Glob. Change Biol. Stehman S V 2012 Impact of sample size allocation when using 12 1107–38 stratified random sampling to estimate accuracy and area of Maniatis D and Mollicone D 2010 Options for sampling and land-cover change Remote Sens. Lett. 3 111–20 stratification for national forest inventories to implement Stehman S V 2013 Estimating area and map accuracy for stratified REDDC under the UNFCCC Carbon Balance Manag. 5 9 random sampling when the strata are different from the map Margono B A, Turubanova S, Zhuravleva I, Potapov P, classes Int. J. Remote Sens. in review Tyukavina A, Baccini A, Goetz S and Hansen M C 2012 Sun G, Ranson K, Kimes D, Blair J and Kovacs K 2008 Forest Mapping and monitoring deforestation and forest degradation vertical structure from GLAS: an evaluation using LVIS and in Sumatra (Indonesia) using Landsat time series data sets SRTM data Remote Sens. Environ. 112 107–17 from 1990 to 2010 Environ. Res. Lett. 7 034010 Treuhaft R N, Chapman B D, Dos Santos J R, Gonc ¸alves F G, Olofsson P, Foody G M, Stehman S V and Woodcock C E 2013 Dutra L V, Grac ¸a P M L A and Drake J B 2009 Vegetation Making better use of accuracy data in land change studies: profiles in tropical forests from multibaseline interferometric estimating accuracy and area and quantifying uncertainty using synthetic aperture radar, field, and lidar measurements stratified estimation Remote Sens. Environ. 129 122–31 J. Geophys. Res. 114 D23110 Pan Y et al 2011 A large and persistent carbon sink in the world’s UN-REDD 2011 Expert meeting on assessment of forest inventory forests Science 333 988–93 approaches for REDDC Meeting Report (Rome: UN-REDD Pflugmacher D, Cohen W, Kennedy R and Lefsky M 2008 Regional Programme) applicability of forest height and aboveground biomass models Zheng D L, Prince S D and Wright R 2013 NPP Multi-Biome: for the Geoscience Laser Altimeter System Forest Sci. Gridded Estimates for Selected Regions Worldwide, 54 647–57 1954–1998, R3. Data set (Oak Ridge, TN: Oak Ridge National Potapov P V, Turubanova S A, Hansen M C, Adusei B, Broich M, Laboratory Distributed Active Archive Center) Altstatt A, Mane L and Justice C O 2012 Quantifying forest Zhuravleva I, Turubanova S, Potapov P, Hansen M, Tyukavina A, cover loss in democratic republic of the Congo, 2000–2010, Minnemeyer S, Laporte N, Goetz S, Verbelen F and Thies C with Landsat ETMC data Remote Sens. Environ. 122 106–16 2013 Satellite-based primary forest degradation assessment in Potapov P V et al 2008 Mapping the world’s intact forest the Democratic Republic of the Congo, 2000–2010 Environ. landscapes by remote sensing Ecol. Soc. 13 51 Res. Lett. 8 024034 Romijn E, Herold M, Kooistra L, Murdiyarso D and Verchot L 2012 Assessing capacities of non-Annex I countries for national Zolkos S G, Goetz S J and Dubayah R 2013 A meta-analysis of forest monitoring in the context of REDDC Environ. Sci. terrestrial aboveground biomass estimation using lidar remote Policy 19–20 33–48 sensing Remote Sens. Environ. 128 289–98 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Environmental Research Letters IOP Publishing

National-scale estimation of gross forest aboveground carbon loss: a case study of the Democratic Republic of the Congo

Loading next page...
 
/lp/iop-publishing/national-scale-estimation-of-gross-forest-aboveground-carbon-loss-a-KOwOAItsVu

References (48)

Copyright
Copyright 2013 IOP Publishing Ltd
eISSN
1748-9326
DOI
10.1088/1748-9326/8/4/044039
Publisher site
See Article on Publisher Site

Abstract

Recent advances in remote sensing enable the mapping and monitoring of carbon stocks without relying on extensive in situ measurements. The Democratic Republic of the Congo (DRC) is among the countries where national forest inventories (NFI) are either non-existent or out of date. Here we demonstrate a method for estimating national-scale gross forest aboveground carbon (AGC) loss and associated uncertainties using remotely sensed-derived forest cover loss and biomass carbon density data. Lidar data were used as a surrogate for NFI plot measurements to estimate carbon stocks and AGC loss based on forest type and activity data derived using time-series multispectral imagery. Specifically, ˆ ´ ´ ´ ´ DRC forest type and loss from the FACET (Forets d’Afrique Centrale Evaluees par Teledetection) product, created using Landsat data, were related to carbon data derived from the Geoscience Laser Altimeter System (GLAS). Validation data for FACET forest area loss were created at a 30-m spatial resolution and compared to the 60-m spatial resolution FACET map. We produced two gross AGC loss estimates for the DRC for the last decade (2000–2010): a map-scale estimate .53:3 9:8 Tg C yr / accounting for whole-pixel classification errors in the 60-m resolution FACET forest cover change product, and a sub-grid estimate .72:1 12:7 Tg C yr / that took into account 60-m cells that experienced partial forest loss. Our sub-grid forest cover and AGC loss estimates, which included smaller-scale forest disturbances, exceed published assessments. Results raise the issue of scale in forest cover change mapping and validation, and subsequent impacts on remotely sensed carbon stock change estimation, particularly for smallholder dominated systems such as the DRC. Keywords: forest cover loss, carbon monitoring, REDD, remote sensing, uncertainty assessment, Congo 1. Introduction compensate developing countries for avoiding emissions due to likely future forest clearing and logging (Houghton 2012) through the emerging REDDC mechanism. The success of The United Nations Reducing Emissions from Deforestation REDDC will be defined by confirmed reductions in rates of and forest Degradation (UN-REDD) program seeks to deforestation and forest degradation. A program requirement is the capability to accurately map and monitor changes in Content from this work may be used under the terms of forest carbon by estimating gross emissions as a function of the Creative Commons Attribution 3.0 licence. Any further area of forest loss and density of carbon stocks within areas of distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. forest loss. 1748-9326/13/044039C14$33.00 1 2013 IOP Publishing Ltd Printed in the UK Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al National forest inventories (NFIs) could provide detailed global-scale carbon stock maps have been created recently and comprehensive information to produce national-scale using the synergy of field measurements, optical, lidar and carbon stock and change estimates. However, NFIs have radar remotely sensed data (Saatchi et al 2011, Baccini not been established in many developing countries that et al 2012). Another approach, presented here, is to calibrate participate in the UN-REDD program (Romijn et al 2012). lidar data using co-located field measurements (Baccini et al The United Nations Food and Agriculture Organization 2012). In this approach, a model is derived to convert lidar (FAO) and the UN-REDD are working on the general waveforms into biomass estimates. The derived model is guidelines for implementing multi-objective NFIs in these then extrapolated to a much larger population of lidar shots, countries (UN-REDD 2011). Meanwhile, alternative methods providing a biomass database for assigning carbon density of national-scale carbon stocks assessment independent of the values to mapped forest cover types. availability of systematically collected ground-based forest For REDDC countries, deforestation is likely to be the inventory data are being investigated and prototyped (GOFC- key category for greenhouse gas emissions estimates. A good GOLD 2010). Goetz et al (2009) provided an overview of the practice for these countries is to use at least IPCC Tier two or satellite-based methods of mapping and monitoring carbon three level assessments for this category of emissions, which stocks, and identified three general approaches: ‘stratify implies reporting uncertainties (Maniatis and Mollicone and multiply’ (SM), when a single carbon density value 2010). AGC stock and loss uncertainty estimates are also is assigned to each land cover type; ‘combine and assign’ crucial if these datasets are to be used as inputs to carbon cycle (CA), extending the SM approach by adding various ancillary and biosphere models. However, published land cover change spatial data layers; and ‘direct remote sensing’ (DR) approach, datasets that may be used as activity data often lack key aimed to derive the carbon stock estimates from machine accuracy assessment information (e.g. description of sampling learning algorithms based on satellite observations and other design, original error matrix, area of each map category, etc) detailed spatial data coupled with field measurements. The that would permit error-adjusted estimates of the change area last approach requires acquisition and processing of large (Olofsson et al 2013). The objectives of our analyses are: (i) to volumes of data to produce a national-scale carbon stock or illustrate the process of activity data accuracy assessment on loss estimate. The first approach, SM, also referred to as the the national level, applicable when using already published ‘biome-average approach’ (Gibbs et al 2007), is relatively land cover data or when creating a new data set, (ii) to easy to implement using a limited set of published data integrate uncertainties from activity and carbon data in a available at low or no cost. Although this approach is fairly national-level forest AGC loss estimate. generalized, in that it does not capture finer-scale spatial In this study, we implemented a SM (‘stratify and heterogeneity of carbon stocks, the accuracy of the estimates multiply’) approach for assessing gross forest AGC loss in can be increased via data refinements and overlays with other the Democratic Republic of the Congo (DRC), where forest data sets in a CA approach. cover change is dominated by smallholder land use and For a national-level aboveground carbon (AGC) loss industrial selective logging (Laporte et al 2007). Due to the assessment, SM approaches require a national-scale land aftermath of two civil wars, persistent political unrest and cover change dataset (activity data in the IPCC terminology lack of infrastructure, the DRC does not collect NFI data IPCC 2006) and mean AGC density estimates for each required for ground-based estimates of AGC stock and its land cover type (IPCC emission factors, here referred to as change. Our approach employs the best available activity and carbon data). Modifying the basic IPCC equation used to carbon data at the national scale—forest extent and loss maps calculate carbon emissions (IPCC 2006, vol 1, chapter 1.2), derived from Landsat imagery (Potapov et al 2012) and AGC the equation to estimate gross AGC loss within a study region estimates derived from GLAS-based canopy vertical structure or a country is the following: metrics (Baccini et al 2012). Results include new estimates of error-adjusted area of forest cover loss between 2000 and 2010, gross AGC loss, and associated uncertainties. AGC loss D 1AD CD (1) i i iD1 2. Data where 1AD (activity data) denotes the change in the extent of a given land cover type i, and CD (carbon data) represents 2.1. Activity data average vegetation carbon content per land cover type. Carbon data that are required for the national-scale AGC loss assessments in an SM approach could be derived from To estimate the area of forest loss, we used Landsat-based field inventory data (e.g. tree DBH and height measurements) year 2000 forest cover and 2000–2010 forest cover loss converted to aboveground biomass using allometric equations data from the Forets ˆ d’Afrique Centrale Evaluees ´ par ´ ´ ´ (e.g. for tropical forests—from Brown 1997 and Chave et al Teledetection (FACET) product, available online (ftp:// congo.iluci.org/FACET/DRC/). FACET data processing and 2005) or existing databases and maps of biomass carbon mapping methodology are described in Potapov et al (2012). density (e.g. Zheng et al 2013, Gibbs 2006, FAO 2010, The FACET dataset provides forest cover and gross forest Malhi et al 2006). Alternatively, biomass carbon content can be mapped using multi-source lidar and radar data that are cover loss for three forest types: primary humid tropical capable to capture vertical tree canopy structure (Goetz and forests, defined as mature humid tropical forest with canopy Dubayah 2011, Treuhaft et al 2009). Several regional and cover >60%; secondary forests, defined as regrowing forest 2 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 1. FACET forest cover and forest cover loss (Potapov et al 2012) combined with DRC wetland map (Bwangoy et al 2010): (a) forested area; (b) woodlands. with canopy cover >60%; and woodlands, defined as forested together with the combined FACET forest cover and DRC areas with canopy cover 30–60%. The spatial resolution of wetland maps to calculate mean AGC density values for the FACET data is 60 m per pixel. We further separated these six target forest classes. Only shots within the forested areas three forest types into terra firma (dryland) and wetland that did not experience forest cover loss between 2000 and sub-classes using the DRC wetland map of Bwangoy et al 2010 according to FACET were used for these calculations. (2010), resulting in six forest types in total. FACET forest The use of a large number of GLAS-estimated biomass cover loss was attributed to these new forest classes (figure 1). values to calculate biome-average AGC densities helps avoid In this manner, the different carbon content of the antecedent biases often inherent in estimates based on the compilation of forest cover could be directly related to disturbance dynamics point-based field measurements (i.e. paucity of sites over large in terra firma and wetland forested ecosystems. In this areas, inadequate stratification to capture variability, and other research, we conduct an explicit statistical validation of factors that limit their spatial representativeness). FACET forest cover loss for each of these forest types and derive the error-adjusted estimate of changed area based on 2.3. Validation data the validation sample. For the purposes of activity data validation, namely the 2.2. Carbon data uncertainty estimation for the FACET forest cover loss, we used all available original L1T Landsat images for years Mean AGC density values for each of the forest types 2000 and 2010 available at no charge from USGS archives were derived from GLAS-based biomass estimates. Baccini (http://glovis.usgs.gov/) and annual Landsat composites for et al (2012) developed a statistical model to predict AGC circa 2000, 2005 and 2010 (Potapov et al 2012). Year 2005 densities observed in the field using GLAS lidar energy composite images helped identify forest cover loss in the early metrics in order to estimate biomass per 65 m diameter 2000s that might be difficult to detect in 2010 Landsat images GLAS shot. The model was based on nearly 300 field sites due to rapid vegetation regeneration in the tropics. located in 12 countries across the tropics. GLAS-predicted In addition to the use of Landsat images for the AGC explained 83% of variance in the field-measured carbon validation (reference) classification, we also employed visual density at the GLAS-footprint scale with a standard error of interpretation of very high spatial resolution images available 1 TM 22:6 Mg C ha (Baccini et al 2012). For this study, we for the study region through Google Earth and through employed the GLAS-derived biomass data as if they were field a partnership between NASA and NGA that provides inventory data and did not incorporate this model uncertainty access to unclassified commercial high spatial resolution in downstream calculations. After screening GLAS data for satellite data from NGA archives for NASA Earth Science noise and filtering for slope (10 ), 371 458 AGC-estimated Investigators (http://cad4nasa.gsfc.nasa.gov/). A total of 1689 GLAS shots for the years 2004–2008 (figure 2) were analyzed high resolution images from multispectral and panchromatic 3 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 2. 2004–2008 GLAS shots color-coded by the FACET forest type (Potapov et al 2012) combined with wetland map (Bwangoy et al 2010). sensors (Ikonos, WorldView-1, WorldView-2, Quickbird, by this objective, as well as by feasibility issues and time constraints. Orbview-5) for 2008–2011 time interval were used for the visual assessment of validation samples. In total, 503 out of a final 1061 validation samples had at least one matching 3.1.1. Sampling design and sample size. The target high resolution image available between 2000 and 2013, activity data class, forest cover loss, is relatively small TM either from Google Earth or from the NGA archive. These compared with the unchanged forest areas; the sampling images facilitated the forest cover loss validation, providing design should increase the sample representation of this rare information about forest cover type on date 1 (2000) or date 2 class in order to achieve a precise estimate of forest cover (2010). loss accuracy (Khorram 1999). Moreover, our objective is forest type-specific loss area estimation and its accuracy; stratified random sampling is an appropriate choice in this 3. Methods case (Stehman 2009). Initially, two strata within each forest type class were 3.1. Uncertainties from activity data considered: ‘no loss’ (forests, undisturbed between 2000 and 2010) and ‘loss’ (2000–2010 forest cover loss). However, The key objective of activity data validation is to estimate sufficient estimation of loss omission error within the large error-adjusted area of forest cover loss for each forest type and ‘no loss’ stratum requires special attention. Given a simple to quantify its uncertainty. Error-adjusted area estimation uses ‘loss’ and ‘no loss’ stratification, rates of false negatives validation sample data to adjust area of forest cover loss due to (change omission errors) could be poorly characterized classification errors (including omission errors and excluding (Khorram 1999). Furthermore, the FACET national-scale commission errors) present in the map product (Olofsson forest cover loss product is likely to be conservative, et al 2013). The choice of sampling design is determined i.e. omitting forest cover loss in comparison to committing 4 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 1. Distribution of samples among forest types using proportional and arbitrary sample allocation strategies for stratified random sampling. Proportional allocation (% samples) Forest type Based on forest area Based on loss area Arbitrary allocation (% samples) Primary forest 46 25 33 Secondary forest 11 55 17 Woodlands 21 13 25 Wetland primary forest 19 3 17 Wetland secondary forest 1 3 4 Wetland woodlands 2 1 4 forest loss. To address this issue we identified an additional Table 2. Allocation of sample size among validation strata. ‘probable loss’ stratum within each forest type class. This Forest type No loss Probable loss Loss Total stratum was constructed to target omitted forest cover loss Primary forest 200 70 63 333 in order to improve the loss area estimate for the AGC loss Secondary forest 30 87 50 167 calculation. We define the ‘probable loss’ stratum as a 1-km Woodlands 100 90 60 250 radius circular region around forest cover loss, assuming that Wetland primary 80 30 57 167 omission of loss is likely to occur in proximity to mapped forest loss. The choice of the 1-km wide ‘probable loss’ stratum Wetland 15 15 12 42 secondary forest is supported by the evidence that increased tree mortality in Wetland 15 15 12 42 temperate and tropical forests is generally observed up to 1 km woodlands from the forest edge (Broadbent et al 2008). A total of 18 strata were analyzed: ‘loss’, ‘probable loss’, and ‘no loss’ for each of the six forest types (terra firma and loss strata, we chose to have an allocation closer to equal, wetland primary forests, terra firma and wetland secondary which helped to target errors of commission (Stehman 2012) forests, terra firma and wetland woodlands). Allocation of among the ‘no loss’, ‘probable loss’ and ‘loss’ strata. A total samples among these strata should effectively address our sample size of 1000 was projected as feasible to be visually validation objective (see section 3.1) of minimizing standard interpreted by expert analysts. We imposed the condition that errors (SEs) of error-adjusted estimators of forest cover loss a sample size greater than 50 was required for the major area (Stehman 2012). forest types (primary, secondary forests, woodlands, wetland When considering allocation of samples among forest primary forests), the allocation of sample size per stratum (the types, we examined both the area of forest type and the sampling unit is one 60-m FACET pixel) was implemented as area of our target class (forest loss) within each forest type. shown in table 2. Proportional allocation of samples among forest types based For the chosen sample allocation we calculated SEs of on the forest type area would lead to small sample sizes from the estimated area of change using hypothetical omission secondary forest, woodlands and wetland forests: almost half and commission error rates in order to confirm that the of all samples in this case fall into the dense forest class chosen allocation would not lead to inflated standard errors. (table 1). Although forest cover loss in dense forests that have We compared our arbitrary allocation to proportional among high biodiversity and other high-value ecosystem services forest allocation with equal and proportional allocation among is important to estimate correctly, the majority of mapped loss strata and found that the arbitrary allocation performed forest cover loss occurred in secondary forests. However, as well or better than the other options. The equation allocation of samples based on the forest cover loss area used to calculate SEs of the estimated area of change for leads to the majority of samples being located in secondary each forest type is similar to equation (3) from Olofsson forests. In order to find a compromise between preserving et al (2013). However, after the assignment of reference a sufficient number of samples in the strategically important values to the samples during expert validation, we found dense forest class while adequately representing the relatively out that the ‘probable loss’ stratum contributed 35% of the small classes with high proportional forest cover change total variance in primary forest, 50% of the variance in (secondary forest, woodlands), we implemented an arbitrary secondary forest, and 20% of the variance in woodlands. allocation that was close to proportional by forest type area, Additional random samples were added to the ‘probable but adjusted for forest loss area (table 1). loss’ stratum of terra firma primary, secondary forests and The sample size allocation to the three strata within each woodlands (20, 30 and 10 samples respectively) in order to forest type was determined as follows. Because it is equally minimize the total SE of the loss area estimate. important for our primary validation objective (estimation of forest loss area for each forest type based on an error 3.1.2. Estimating area of forest loss and its uncertainty. matrix) to account for committed and omitted loss area, we addressed the need to account for omission errors by Visual interpretation of validation samples was performed at a creating the separate ‘probable loss’ strata within the original 30-m spatial resolution, enabling map-scale and sub-grid error ‘no loss’ class. Therefore, when allocating samples among assessments (FACET was made at a 60 m spatial resolution 5 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 3. Example of sample block visual interpretation; for the map-scale estimate, 0.5 loss is treated as no loss. The black stripe in the 2010 Landsat loss sample is a data gap due to the Landsat 7 scan-line corrector malfunction. Table 3. Error matrix of sample counts for map-scale and sub-grid area estimates. Reference strata Map-scale estimate Sub-grid estimate Forest type Map strata No loss Loss No loss Loss N of pixels in each stratum Primary forest No loss 200 0 200 0 147 647 298 No loss–probable loss 89 1 86.5 3.5 56 158 987 Loss 3 60 3 60 2 638 342 Secondary forest No loss 30 0 30 0 5 720 568 No loss–probable loss 107 10 98.5 18.5 35 535 337 Loss 00–10 3 47 3 47 5 619 034 Woodlands No loss 100 0 100 0 51 491 436 No loss–probable loss 98 2 97 3 39 725 284 Loss 00–10 7 53 7 53 1 374 079 Wetland primary forest No loss 80 0 80 0 67 675 696 No loss–probable loss 30 0 30 0 15 706 036 Loss 00–10 9 48 9 48 326 316 Wetland secondary forest No loss 15 0 15 0 1 506 946 No loss–probable loss 15 0 14.5 0.5 2 176 786 Loss 00–10 4 8 4 8 255 498 Wetland woodlands No loss 15 0 15 0 7 003 885 No loss–probable loss 15 0 15 0 2 477 979 Loss 00–10 2 10 2 10 97 176 using resampled 30-m Landsat time-series imagery). We a 60-m validation pixel as ‘loss’ only if the reference forest produced two forest loss area estimates for the DRC for the loss fraction detected using 30-m Landsat and/or high spatial last decade (2000–2010): a map-scale estimate accounting resolution was75% of pixel area. For the sub-grid estimate, for whole-pixel classification errors in the 60-m resolution three gradations of reference loss fraction per pixel were used: FACET forest cover change product, and a sub-grid estimate 1 (loss) with reference loss 75% of pixel area; 0.5 (mixed that took into account 60-m cells that experienced partial pixels) with reference loss between 75% and 25%; and 0 (no forest loss (table 3). For the map-scale estimate we treated loss) otherwise (figure 3). 6 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 4. Parameters for the calculation of error-adjusted area of forest cover loss within terra firma primary forests (map-scale estimate). Primary forest y n yN N Map area (ha) s u h h h u2h yh No loss 0 200 0=200 147 647 298 53 153 027 0.000 000 000 No loss–probable loss 1 90 1=90 56 158 987 20 217 235 0.011 111 111 Loss 60 63 60=63 2 638 342 949 803 0.046 082 949 Total 206 444 627 74 320 066 90 0:011 111 111 When the sampling strata and map classes being validated C 56 158 987 1 are the same, equations (2)–(4) from Olofsson et al (2013) 56 158 987 90 should be used to calculate error-adjusted area of forest cover 63 0:046 082 949 loss and its standard error based on a validation confusion C 2638 342 1 2638 342 63 matrix. In our case, there was a mismatch between sampling 1=2 strata (‘no loss’, ‘probable loss’, ‘loss’) and map classes 2 1 .206 444 627 / (‘loss’ and ‘no loss’) within each forest cover type arising from the attempt to target omitted forest cover loss by creating D 226 099 ha (6) the additional ‘probable loss’ stratum. Based on sampling theory (Cochran 1977), the following equation was employed A D 1129 210 443 156 ha: (7) to produce an unbiased estimator of the area of forest cover loss within each of the forest cover types when validation strata and map classes do not match (Stehman 2013, in review): 3.2. Uncertainties from carbon data N yN h h hD1 A D A  (2) tot N Table 5 presents the mean and population standard deviation (STD) derived from the number of GLAS shots per forest where A —total area of the forest cover type; tot type. Using the SM (‘stratify and multiply’) approach we y D 0:5 or 1 if pixel u (or it’s half) is in reference class ‘forest assigned a single mean AGC density value to each of the cover loss’, and y D 0 otherwise; u forest type classes to estimate gross AGC loss. To quantify u2h yN D , the sample-mean of the y values in stratum h; h u the uncertainty of this estimate, we employed the standard n —sample size in stratum h; deviation of the sample-mean’s estimate of a population mean, N —number of pixels in stratum h; the standard error of the mean (SEM). According to the central N—total number of pixels within the forest cover type. limit theorem, the distribution of sample estimates of the The standard error of the error-adjusted estimate of the mean is normally distributed, enabling us to calculate the 95% forest cover loss is: confidence interval (CI) of mean AGC density estimates as P s 1.96SEM. Table 5 shows mean AGC densities of our target H n yh 2 h N 1 hD1 h N n h h forests classes along with their 95% CIs. SE.A/ D A (3) tot .y yN / 2 h 3.3. Combination of the uncertainties u2h where s D , the sample variance for stratum h. yh n 1 A 95% confidence interval (assuming normal distribu- When calculating AGC loss for each forest type using tion) is: equation (1), uncertainty comes both from activity data (in O O A 1:96SE.A/: (4) our case—forest cover loss) and emission factors (carbon data). In order to combine uncertainties from these quantities, An example of the forest cover loss area estimation for terra the multiplication approach from the recent IPCC Guidelines firma primary forests (map-scale estimate) is presented in for National Greenhouse Gas Inventories (IPCC 2006, vol 1, table 4 and equations (5)–(7). chapter 3, p 28, equation (3.1)) was used: A D 74 320 065:72.0 147 647 298C  56 158 987 90 q 2 2 2 60 1 U D U C U CC U (8) total C  2 638 342/.206 444 627/ 1 2 where U is the percentage uncertainty in the product of total D 1129 210 ha (5) the quantities (half the 95% confidence interval divided by the SE.A/ D 74 320 065:72 total and expressed as a percentage). U is the percentage uncertainties associated with each of 200 0:0 147 647 298 1 the quantities. 147 647 298 200 7 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 5. GLAS-based AGC density estimates for the DRC forest types. Mean AGC densities are given with95% CI. Forest type Mean AGC density (Mg C ha ) Number of GLAS samples STD Primary forest 156.8 0.4 115 566 67.03 Secondary forest 94.8 0.7 31 443 67.45 Woodlands 71.2 0.2 121 671 44.24 Wetland primary forest 128.9 0.4 85 923 55.29 Wetland secondary forest 90.7 2.3 3 148 65.83 Wetland woodlands 66.5 0.8 13 707 45.81 Table 6. Original FACET and error-adjusted estimates of 2000–2010 forest cover loss within DRC forest types (95% CI). 2000–2010 forest cover loss (ha) Error-adjusted Forest type Map-scale estimate Sub-grid estimate FACET map Primary forest 1 129 210 443 156 1 690 800 645 694 949 803 Secondary forest 2 994 876 664 625 3 924 262 736 673 2022 852 Woodlands 722 979 396 475 865 990 439 210 494 668 Wetland primary forest 98 925 11 218 98 925 11 218 117 474 Wetland secondary forest 87 440 78 014 87 441 78 014 91 979 Wetland woodlands 29 153 7704 29 153 7704 34 983 For example, for the primary forest stratum, the 4. Results calculation of the U (using the map-scale 1AD estimate) total is the following: Applying the approach of adjustment for the classification errors described in section 3, we produced estimates of u ! SE.A/ AGC SEM forest cover loss within target DRC forest classes (table 6). U D  100 C  100 total Mean AGC Error-adjustment significantly increased estimated areas of forest loss in terra firma forest classes (primary, secondary 2 2 forests and woodlands); omission errors prevailed over 226 099:75 0:2 D  100 C  100 commission errors (figure 4). In the wetland forests and 1129 210 156:83 woodlands, on the contrary, more loss was committed in D 20:02%: (9) the map product; error-adjusted loss area estimates were smaller than those prior to adjustment. SE was highest in When calculating total gross AGC loss within the DRC wetland secondary forests and terra firma woodlands. High (summing AGC loss values for all forest types), the addition uncertainty in the wetland secondary forests is associated and subtraction approach from the IPCC Guidelines (IPCC with it being the smallest and spatially discontinuous class. 2006, vol 1, chapter 3, p 28, equation (3.2)) was used to Woodland is a challenging forest type to map and monitor due estimate the uncertainty of the resulting quantity: to the gradients of tree canopy cover and seasonality as well total DRC as the comparatively uneven intensity of disturbance events, 2 2 2 all of which contributes to larger SEs. .U x / C .U x / CC .U x / 1 1 2 2 n n To compare AGC density estimates for our target forest jx C x CC x j 1 2 n classes with published estimates, we calculated average AGC (10) densities within the 6 DRC forest types using available where U is the percentage uncertainty in the sum of the spatially explicit vegetation carbon density products (Baccini total quantities (half the 95% confidence interval divided by the et al 2012, Saatchi et al 2011, Gibbs and Brown 2007, total and expressed as percentage); Kindermann et al 2008) and compared them with the x and U are the uncertain quantities and percentage i i GLAS-based estimates of the current study (figure 5). This uncertainties associated with them. comparison provides a general understanding of how well Thus, the overall uncertainty of gross AGC loss estimate our current estimates correspond to existing knowledge. for the entire DRC is: Examination of figure 5 shows that GLAS-based AGC density estimates are generally higher than those modeled using total DRC optical remotely sensed data (Baccini et al 2012, Saatchi et al 2 2 2 .U 1AGC / C.U 1AGC / CC.U 1AGC / total1 1 total2 2 totaln n D 2011, Gibbs and Brown 2007), probably because of spatial j1AGC C1AGC CC1AGC j 1 2 averaging (Goetz and Dubayah 2011, Zolkos et al 2013), but (11) don not exceed the estimates of Kindermann et al (2008) who where numbers (1–n) stand for the six forest cover types. employed FAO 2005 Forest Resources Assessment statistics. 8 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 4. Forest cover loss (2000–2010) within DRC forest types; error bars are the 95% CIs. Figure 5. Comparison of the AGC density estimates from the published datasets (error bars are the 95% CIs) and the current study. Table 7. Gross AGC loss estimates (2000–2010) with the uncertainty measures for DRC forest types ( is the 95% CI). Map-scale loss area estimate Sub-grid loss area estimate Forest type U (%) Gross AGC loss 2000–2010 (Pg C) U (%) Gross AGC loss 2000–2010 (Pg C) total total Primary forest 20.0 0.177 0.070 19.5 0.265 0.101 Secondary forest 11.3 0.284 0.063 9.6 0.372 0.070 Woodlands 28.0 0.051 0.028 25.9 0.062 0.031 Wetland primary forest 5.8 0.013 0.001 5.8 0.013 0.001 Wetland secondary forest 45.5 0.006 0.005 45.5 0.008 0.007 Wetland woodlands 13.5 0.002 0.001 13.5 0.002 0.001 DRC total 9.4 0.533 0.098 9.0 0.721 0.127 Sub-grid gross AGC loss estimates were 20–50% higher Differences between these estimates are mostly associated than map-scale ones for the major terra firma forests (primary, with the ‘loss’ and ‘probable loss’ strata, particularly in secondary forests and woodlands) and nearly equal for the regions where primary and secondary forest loss predominate. less widespread wetland forests (table 7, figures 6(b) and (c)). There are no significant differences in the forests and 9 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Figure 6. Forest type and strata averages, aggregated to a 5-km grid: (a) year 2000 AGC; (b) map-scale estimate of 2000–2010 gross AGC loss; (c) sub-grid estimate of 2000–2010 AGC loss; (d) difference between sub-grid and map-scale estimates. Water bodies are shown in gray. Note that AGC values for both (b) and (c) are the same for the respective forest types. woodlands of the ‘no loss’ strata (figure 6(d)). For the whole confidence interval of the global sample-based estimate of the DRC, the sub-grid AGC loss estimate was 35% higher of Hansen et al (2010), but is significantly higher than than the map-scale estimate (table 7). the FACET map-based estimate without error-adjustment The comparison of gross forest cover loss and gross AGC (Potapov et al 2012). The sub-grid estimate, accounting for rates from this study with published estimates is presented in the finer-scale forest disturbance, is 30–40% higher than table 8. We report annual forest cover loss rates separately for published estimates for the DRC, and points to the difficulty primary and secondary forests, excluding woodlands (table 8) of mapping forest change in a landscape where smallholder to best match the definition of forests employed in the shifting cultivation predominates. For example, FACET forest most recent regional sample-based forest cover loss estimate cover loss has a mean patch area of 1.4 ha (Potapov et al by Ernst et al (2013) (all tropical moist forests, excluding 2012). While patch size is not the same as field size, it woodland savannahs and tropical dry forests). is worth noting that typical shifting cultivation practices in the tropics employ field sizes well under 1 ha (Aweto 5. Discussion 2013). The quantification of such change is challenging and represented by the comparatively large presence of The results reported in table 8 need to be considered in the mixed pixels in the FACET data. The difference of two context of inconsistencies in methodologies, definitions, and methodologically consistent loss area estimates based on areas of analysis (a direct consequence of the differences input data of different resolutions (60-m FACET and 30-m in the definitions of forest and woodlands). Our map-scale Hansen et al 2013, table 8) prior to error-adjustment illustrates 2000–2010 annual forest cover loss estimate within dense the issue: the 30-m product depicts 1.5 times more change forests (0.35%  0.03%) agrees well with the estimates than the 60-m one. Any binary (yes/no) change map will have of Ernst et al (2013) for the first half of the decade (0.32%  0.05%) and of Hansen et al (2013) for 2000–2012 scale-dependent omission errors. These ‘cryptic disturbances’ (0.34%). Our map-scale estimate also falls within the have been reported to add more than 50% of forest cover loss 10 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Table 8. Comparison of forest cover and carbon loss estimates for the DRC (95% CI). 2000–2005 2005–2010 Source Extent Annual gross forest cover loss (% of the forest area) Current study Map-scale ForestsC woodlands 0.32% 0.03% Sub-grid ForestsC woodlands 0.42% 0.03% Map-scale Forests 0.35% 0.03% Sub-grid Forests 0.47% 0.04% FACET map Potapov et al (2012)—60 m ForestsC woodlands 0.22% 0.25% Hansen et al (2013)—30 m ForestsC woodlands 0.34% Ernst et al (2013) Forests 0.32% 0.05% — Hansen et al (2010) ForestsC woodlands 0.12% 0.23% — Annual net forest cover loss (% of the forest area) FAO (2010) ForestsC woodlands 0.20% 0.20% Ernst et al (2013) Forests 0.22% 0.22% Annual gross AGC loss (Tg C yr ) Current study Map-scale ForestsC woodlands 53.3 9.8 Sub-grid ForestsC woodlands 72.1 12.7 Annual gross carbon loss (Tg C yr ) Harris et al (2012) ForestsC woodlands 23 — TM to existing Landsat-scale forest disturbance classifications for Google Earth high resolution imagery) allows defining the Amazon Basin (Asner et al 2005). reference values of validation samples without in situ measurements. Despite its advantages, the method is sensitive Table 8 reflects a second type of omission error related to sampling design and the associated decision of how to to algorithmic and/or data limitations. Estimates of forest loss allocate the sample size among validation strata. For the strata derived at a 30-m spatial resolution, particularly the Hansen and sample size allocation implemented in this study, the et al (2013) and Ernst et al (2013) products, have comparable decisions were advantageous; for the four largest forest types, gross forest cover loss rates, 0.34% and 0.32%  0.05%. However, the 30-m validation estimate is 0.47%  0.04%. the reduction in standard error attributable to the stratification Large area mapping algorithms are often conservatively was substantial. Specifically, the gain in precision due to implemented in attempting to avoid commission error. For stratification can be computed from the sample data (Cochran validation, the determination of loss/no loss is performed 1977, section 5A.11) as the ratio of the standard error that independently per sample and is free of this consideration. would have been obtained from simple random sampling Differences between the Hansen et al (2013) 30-m map and to the standard error obtained from the stratified design the Ernst et al (2013) 30-m sample estimates could be due implemented (same sample size for both designs). For the four to this fact. However, the estimate of Ernst et al (2013) was largest forest types, these ratios were 1.42 for primary forest, also sample based. The additional loss found in our validation 1.10 for secondary forest, 1.32 for woodlands, and 23.21 for effort compared to Ernst et al (2013), while partially due to wetland primary forest (the latter estimate is likely inflated by the use of very high spatial resolution data for a portion of the the fact that two of the three strata had 0% forest loss). The methodology is also highly dependent on the knowledge base reference samples, is not easily explained and may be more of the remote sensing experts performing visual interpretation related to definitional differences or other methodological factors. In summary, the difference between the 60-m FACET of validation samples. Finally, it is a function of the quality of loss rates of 0.22% and 0.25% and the 30-m loss rates of the reference imagery and the resulting clarity or conversely 0.34% and 0.32% is most likely related to the differing scales ambiguity in assigning change per validation sample. The of measurement. The difference between the 30-m loss rates map-scale and sub-grid estimates reflect the importance of this of 0.34% and 0.32% and the validation rate of 0.47% is most issue. likely related to limitations in mapping versus sampling or A further consideration in assessing the results concerns to other methodological factors. The discrepancy between the reference data and the potential volatility of the sample- map-scale and sub-grid estimates emphasizes the issue of based estimate itself. Table 4 illustrates this issue. The ‘loss’ scale in change area estimation for smallholder dominated stratum records 60 of 63 samples as having experienced terra landscapes like the DRC. firma primary forest cover loss, representing 905 574 ha of The approach for validating activity data employed in error-adjusted forest loss area. For the ‘probable loss’ stratum, this study is relatively straightforward and easy to implement. 1 of 90 samples was interpreted as having experienced forest The method allows for the generation of error-adjusted loss cover loss. Due to the much larger size of this stratum, this one area estimates from the existing land cover and vegetation sample accounts for an estimated 224 635 ha of error-adjusted maps. This approach does not require large volumes of data forest loss area, or fully 20% of terra firma primary forest processing and is therefore not limited by computational cover loss. Without the use of the ‘probable loss’ stratum facilities. The use of open access medium- and high resolution and the inclusion of this single sample of commission error, imagery for map product validation (USGS Landsat archive, results would indicate a slight underestimate of terra firma 11 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al forest cover loss. Validation studies should formally consider 2008). In total, GLAS-based AGC models explain from 73% likely regions of false negatives of forest change in developing (Lefsky et al 2005, Pflugmacher et al 2008) to 83% (current stratified sampling methods for error-adjusted area estimation. study; Baccini et al 2012) of the variance in field-estimated The validity of the sample-based estimate is a function of biomass. Regional forest inventory data are required to calibrate and validate the current forest type GLAS-based many factors, including the vagaries of any individual sample estimates. Additional field data collection could further refine data set used in creating the error-adjusted estimates. the estimates but, unfortunately, GLAS observations are Estimates of carbon density derived using different not available after 2009, posing a near-term challenge for methods can vary considerably within the same region improved AGC mapping and monitoring beyond the current (Houghton et al 2001), introducing uncertainty to the carbon models. As part of the process of establishing an NFI for loss estimation. However, recent published estimates of the DRC continues, other sources of remotely sensed data carbon loss from deforestation differ primarily due to major characterizing vegetation vertical structure, such as airborne disagreements in the quantification of the areal extent of forest lidar or spaceborne radar data, can bridge the gap until cover loss (Pan et al 2011, Harris et al 2012). The DRC gross systematic spaceborne lidar measurements become available AGC loss estimates from the current study (map-scale and to the scientific and REDDC implementation communities. sub-grid) are 2 to 3 times higher than the biomass carbon loss (total carbon, above- and belowground) estimate of Harris et al (2012) (table 8) due primarily to differences in the 6. Conclusion estimated area of forest cover loss. The Harris et al (2012) estimate is based on a global forest cover loss product by We applied a method of error-adjustment of forest cover Hansen et al (2010) that is highly uncertain in the DRC loss area to produce a national-scale gross forest AGC loss (SE D 100%, see table 8). Hansen et al (2010) employed estimate for the DRC based on a published forest cover loss a pan-tropical MODIS-based stratification to target sample dataset. We employed field-calibrated GLAS lidar-derived allocation with only 7 samples located in the DRC. The biomass carbon densities as a substitute for NFI data, which small sample size resulted in a high standard error (table 8). do not exist for the territory of the DRC. Two realizations of Harris et al (2012) reported a 90% carbon loss prediction the resulting DRC gross AGC loss estimate, map-scale and interval for the DRC, based on a Monte Carlo approach: sub-grid, were produced. The sub-grid AGC loss estimate 16–32 Tg C yr ; our current DRC gross AGC loss estimates, accounted for disturbances finer than the map grid scale of 1 1 map-scale (53:3 Tg C yr ) and sub-grid (72:1 Tg C yr ), 60 m and was higher than published estimates, highlighting are not within this interval. issues of scale and spatial averaging in AGC estimation. In our analysis, DRC gross forest AGC loss assessment Omitted disturbances were largely related to smallholder consists only of stand-replacement forest disturbance that agriculture land cover change, the detection of which is scale- can be observed at the mapping scale and in reference data. dependent. For the FACET product, the input Landsat imagery However, forest degradation processes that do not lead to the were averaged to 60 m and then classified, leading to the complete loss of tree canopy or cause small-scale canopy estimated scale-related omission error. Other processing steps openings, and can be detected only in the field or using can lead to change omission, either through the algorithm dense series of sub-meter remotely sensed data may result itself, for example image segmentation, post-processing of in significant AGC loss at the national scale (IPCC 2003, the output classification, or the application of a minimum Schoene et al 2007). One possible approach to assess the mapping unit. In Brazil, where agro-industrial land conversion loss of biomass from these disturbances could be based on results in large forest disturbances, the Brazilian Space monitoring changes in the area of intact forest landscapes Agency’s PRODES product 6.25 ha minimum mapping (Potapov et al 2008) and assigning an AGC loss value to the unit (the equivalent of approximately 69 Landsat pixels) forests that have undergone the transition from intact primary (INPE 2012), provides a viable deforestation monitoring to primary degraded and secondary forests (Margono et al approach. However, a 6.25 ha minimum mapping unit 2012, Zhuravleva et al 2013). For countries such as the DRC, for the DRC would omit the majority of change. For where large-scale agro-industrial forest disturbance is largely heterogeneous landscapes with change dynamics at or finer absent, the question of scale and its impact on AGC loss due than the resolution of Landsat data, higher spatial resolution to deforestation and degradation remains an important line of imagery to directly map such changes, or indirect methods scientific inquiry. to delimit degraded areas and subsequently relate to in situ We employed GLAS-based AGC estimates as a proxy for measurements, are required. the ground-based NFI data. There are some known issues and Our study also illustrates the importance of reference limitations concerning the estimation of biomass from GLAS forest state in assessing carbon dynamics, as with the metrics. For example, GLAS-estimated vegetation heights primary, secondary and woodland forest types presented here. often used in AGC models have on average 2–3-m error The Brazilian PRODES product, the current standard for when compared with USDA Forest Inventory and Analysis national-scale forest monitoring, quantifies only the loss of (FIA) and other field-measured heights (Pflugmacher et al primary forest in the Legal Amazon. While reducing primary 2008, Lefsky et al 2005, Sun et al 2008). GLAS-derived humid tropical forest loss is the main focus of climate biomass estimates are also known to be affected by the mitigation strategies such as REDDC, other forest types and season of data acquisition and terrain slope (Sun et al even trees outside of forests will be part of national carbon 12 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al accounting schemes. Our study underscored the importance Brown S 1997 Estimating biomass and biomass change of tropical forests: a primer FAO Forestry Paper 134 (Rome: Food and of monitoring other forest dynamics, as we found AGC loss Agriculture Organisation of the UN) in secondary forests to be 140% that of primary forests. The Bwangoy J B, Hansen M C, Roy D P, Grandi G De and Justice C O reuse of secondary forests remains a challenge to carbon 2010 Wetland mapping in the Congo Basin using optical and monitoring and the development of appropriate strategies radar remotely sensed data and derived topographical indices for reducing emissions, but monitoring all relevant forest Remote Sens. Environ. 114 73–86 Chave J et al 2005 Tree allometry and improved estimation of types and dynamics is required as national-scale programs are carbon stocks and balance in tropical forests Oecologia developed and implemented. 145 87–99 REDDC mechanisms will rely on accurate mapping Cochran W G 1977 Sampling Techniques (New York: Wiley) and monitoring of AGC (Houghton et al 2010). However, Ernst C, Mayaux P, Verhegghen A, Bodart C, Christophe M and scientific, technical and operational aspects of AGC mapping Defourny P 2013 National forest cover change in Congo Basin: deforestation, reforestation, degradation and regeneration for and monitoring are still in their infancy. Results from this the years 1990, 2000 and 2005 Glob. Change Biol. 19 1173–87 study have significant implications for policy initiatives FAO 2010 Global forest resources assessment 2010 FAO Forestry like REDDC. It is clear that the spatial scale of forest Paper 163 (Rome: Food and Agriculture Organization of the change characterization, reference information on forest type UN) and carbon stock, and sample representativeness, can all Gibbs H K 2006 Major World Ecosytem Complexes Ranked by Carbon in Live Vegetation: An Updated Database Using the dramatically impact AGC loss estimation. For example, GLC2000 Land Cover Product NDP-017b considering change at a 30 m validation scale, an extra 35% Gibbs H K and Brown S 2007 Geographical distribution of woody of AGC loss was estimated compared to the 60-m spatial biomass carbon in tropical africa: an updated database for 2000 scale; terra firma secondary forest cover loss accounted for NDP-055b 40% more AGC loss than that of terra firma primary forest Gibbs H K, Brown S, Niles J O and Foley J A 2007 Monitoring and estimating tropical forest carbon stocks: making REDD a loss; a single validation sample added 20% to map-scale reality Environ. Res. Lett. 2 045023 terra firma primary forest cover loss area. The volatility of Goetz S J, Baccini A, Laporte N T, Johns T, Walker W, results within this study indicates the DRC to be a challenging Kellndorfer J, Houghton R A and Sun M 2009 Mapping and environment for quantifying changes to forest carbon stocks, monitoring carbon stocks with satellite observations: a with implications for other countries as well. Eventual comparison of methods Carbon Balance Manag. 4 2 Goetz S J and Dubayah R 2011 Advances in remote sensing national monitoring systems will need to demonstrate technology and implications for measuring and monitoring spatio-temporal consistency given the various factors that forest carbon stocks and change Carbon Manag. 2 231–44 impact AGC loss estimation. While absolute accuracies GOFC-GOLD 2010 A sourcebook of methods and procedures for may differ due to some of the aforementioned factors, monitoring and reporting anthropogenic greenhouse gas relative consistency for any particular set of observations emissions and removals caused by deforestation, gains and losses of carbon stocks in forests remaining forests, and and spatial scale should be achievable and implementable. forestation Report Version COP16-1 (Alberta: GOFC-GOLD Demonstrating such consistency will be a proof of readiness Project Office, Natural Resources Canada) for REDDC monitoring. Hansen M C et al 2013 The first high-resolution global maps of 21st century forest cover change at press Hansen M C, Stehman S V and Potapov P V 2010 Quantification of Acknowledgments global gross forest cover loss Proc. Natl Acad. Sci. 107 8650–5 Harris N L, Brown S, Hagen S C, Saatchi S S, Petrova S, Salas W, Support for this study was provided by NASA’s Terrestrial Hansen M C, Potapov P V and Lotsch A 2012 Baseline map of Ecology program through grant number NNX12AB43G, carbon emissions from deforestation in tropical regions Science 336 1573–6 NASA Applied Sciences grant NNX12AL27G, and by the Houghton R A 2012 Carbon emissions and the drivers of United States Agency for International Development through deforestation and forest degradation in the tropics Curr. Opin. its CARPE program. Thanks to Brian Barker, Giuseppe Environ. Sustain. 4 597–603 Molinario and Alice Altstatt from the UMD CARPE team for Houghton R A, Greenglass N, Baccini A, Cattaneo A, Goetz S, the processing of high resolution remotely sensed data and Kellndorfer J, Laporte N and Walker W 2010 The role of valuable comments. science in reducing emissions from deforestation and forest degradation (REDD) Carbon Manag. 1 253–9 Houghton R A, Lawrence K T, Hackler J L and Brown S 2001 The References spatial distribution of forest biomass in the Brazilian Amazon: a comparison of estimates Glob. Change Biol. 7 731–46 Asner G P, Knapp D E, Broadbent E N, Oliveira P J C, Keller M Instituto Nacional de Pesquisas Especias 2012 Monitoring of the and Silva J N 2005 Selective logging in the Brazilian Amazon Brazilian Amazonian Forest by Satellite, 2000–2012 Science 310 480–2 IPCC 2003 Definitions and Methodological Options to Inventory Aweto A O 2013 Shifting Cultivation and Secondary Succession in Emissions from Direct Human-induced Degradation of Forests the Tropics (Wallingford, UK: CABI) and Devegetation of Other Vegetation Types (Hayama, Japan: Baccini A et al 2012 Estimated carbon dioxide emissions from The Institute for Global Environmental Strategies) tropical deforestation improved by carbon-density maps IPCC 2006 2006 IPCC Guidelines for National Greenhouse Gas Nature Clim. Change 2 182–5 Inventories ed H S Eggleston, L Buendia, K Miwa, Broadbent E, Asner G, Keller M, Knapp D, Oliveira P and Silva J T Ngara and K Tanabe (Hayama, Japan: IGES) 2008 Forest fragmentation and edge effects from deforestation Khorram S 1999 Accuracy Assessment of Remote Sensing-Derived and selective logging in the Brazilian Amazon Biol. Conserv. Change Detection ed S Khorram (Bethesda, MD: American 141 1745–57 Society for Photogrammetry and Remote Sensing) 13 Environ. Res. Lett. 8 (2013) 044039 A Tyukavina et al Kindermann G E, Mccallum I, Fritz S and Obersteiner M 2008 Saatchi S S, Harris N L, Brown S, Lefsky M, Mitchard E T A and A global forest growing stock, biomass and carbon map based Salas W 2011 Benchmark map of forest carbon stocks in on FAO statistics Silva Fenn. 42 387–96 tropical regions across three continents Proc. Natl Acad. Sci. Laporte N T, Stabach J A, Grosch R, Lin T S and Goetz S J 2007 108 9899–904 Expansion of industrial logging in central Africa Science Schoene D, Killmann W, Von Luepke H and LoycheWilkie M 2007 316 1451 Definitional issues related to reducing emissions from Lefsky M A, Harding D J, Keller M, Cohen W B, Carabajal C C, deforestation in developing countries Forest and Climate Espirito-Santo F D B, Hunter M O and de Olivera R Jr 2005 Change Working Paper 5 (Rome: Food and Agriculture Estimates of forest canopy height and aboveground biomass Organization of the UN) using ICESat Geog. Res. Lett. 32 L22S02 Stehman S V 2009 Sampling designs for accuracy assessment of Malhi Y et al 2006 The regional variation of aboveground live land cover Int. J. Remote Sens. 30 5243–72 biomass in old-growth Amazonian forests Glob. Change Biol. Stehman S V 2012 Impact of sample size allocation when using 12 1107–38 stratified random sampling to estimate accuracy and area of Maniatis D and Mollicone D 2010 Options for sampling and land-cover change Remote Sens. Lett. 3 111–20 stratification for national forest inventories to implement Stehman S V 2013 Estimating area and map accuracy for stratified REDDC under the UNFCCC Carbon Balance Manag. 5 9 random sampling when the strata are different from the map Margono B A, Turubanova S, Zhuravleva I, Potapov P, classes Int. J. Remote Sens. in review Tyukavina A, Baccini A, Goetz S and Hansen M C 2012 Sun G, Ranson K, Kimes D, Blair J and Kovacs K 2008 Forest Mapping and monitoring deforestation and forest degradation vertical structure from GLAS: an evaluation using LVIS and in Sumatra (Indonesia) using Landsat time series data sets SRTM data Remote Sens. Environ. 112 107–17 from 1990 to 2010 Environ. Res. Lett. 7 034010 Treuhaft R N, Chapman B D, Dos Santos J R, Gonc ¸alves F G, Olofsson P, Foody G M, Stehman S V and Woodcock C E 2013 Dutra L V, Grac ¸a P M L A and Drake J B 2009 Vegetation Making better use of accuracy data in land change studies: profiles in tropical forests from multibaseline interferometric estimating accuracy and area and quantifying uncertainty using synthetic aperture radar, field, and lidar measurements stratified estimation Remote Sens. Environ. 129 122–31 J. Geophys. Res. 114 D23110 Pan Y et al 2011 A large and persistent carbon sink in the world’s UN-REDD 2011 Expert meeting on assessment of forest inventory forests Science 333 988–93 approaches for REDDC Meeting Report (Rome: UN-REDD Pflugmacher D, Cohen W, Kennedy R and Lefsky M 2008 Regional Programme) applicability of forest height and aboveground biomass models Zheng D L, Prince S D and Wright R 2013 NPP Multi-Biome: for the Geoscience Laser Altimeter System Forest Sci. Gridded Estimates for Selected Regions Worldwide, 54 647–57 1954–1998, R3. Data set (Oak Ridge, TN: Oak Ridge National Potapov P V, Turubanova S A, Hansen M C, Adusei B, Broich M, Laboratory Distributed Active Archive Center) Altstatt A, Mane L and Justice C O 2012 Quantifying forest Zhuravleva I, Turubanova S, Potapov P, Hansen M, Tyukavina A, cover loss in democratic republic of the Congo, 2000–2010, Minnemeyer S, Laporte N, Goetz S, Verbelen F and Thies C with Landsat ETMC data Remote Sens. Environ. 122 106–16 2013 Satellite-based primary forest degradation assessment in Potapov P V et al 2008 Mapping the world’s intact forest the Democratic Republic of the Congo, 2000–2010 Environ. landscapes by remote sensing Ecol. Soc. 13 51 Res. Lett. 8 024034 Romijn E, Herold M, Kooistra L, Murdiyarso D and Verchot L 2012 Assessing capacities of non-Annex I countries for national Zolkos S G, Goetz S J and Dubayah R 2013 A meta-analysis of forest monitoring in the context of REDDC Environ. Sci. terrestrial aboveground biomass estimation using lidar remote Policy 19–20 33–48 sensing Remote Sens. Environ. 128 289–98

Journal

Environmental Research LettersIOP Publishing

Published: Dec 1, 2013

There are no references for this article.