Benchmark map of forest carbon stocks in tropical regions across three continents.

Abstract

Developing countries are required to produce robust estimates of forest carbon stocks for successful implementation of climate change mitigation policies related to reducing emissions from deforestation and degradation (REDD). Here we present a "benchmark" map of biomass carbon stocks over 2.5 billion ha of forests on three continents, encompassing all tropical forests, for the early 2000s, which will be invaluable for REDD assessments at both project and national scales. We mapped the total carbon stock in live biomass (above- and belowground), using a combination of data from 4,079 in situ inventory plots and satellite light detection and ranging (Lidar) samples of forest structure to estimate carbon storage, plus optical and microwave imagery (1-km resolution) to extrapolate over the landscape. The total biomass carbon stock of forests in the study region is estimated to be 247 Gt C, with 193 Gt C stored aboveground and 54 Gt C stored belowground in roots. Forests in Latin America, sub-Saharan Africa, and Southeast Asia accounted for 49%, 25%, and 26% of the total stock, respectively. By analyzing the errors propagated through the estimation process, uncertainty at the pixel level (100 ha) ranged from ± 6% to ± 53%, but was constrained at the typical project (10,000 ha) and national (>1,000,000 ha) scales at ca. ± 5% and ca. ± 1%, respectively. The benchmark map illustrates regional patterns and provides methodologically comparable estimates of carbon stocks for 75 developing countries where previous assessments were either poor or incomplete.

Full text

Page 1

Benchmark map of forest carbon stocks in tropical
regions across three continents
Sassan S. Saatchia,b,1, Nancy L. Harrisc, Sandra Brownc, Michael Lefskyd, Edward T. A. Mitcharde, William Salasf,
Brian R. Zuttaa,b, Wolfgang Buermannb, Simon L. Lewisg, Stephen Hagenf, Silvia Petrovac, Lee Whiteh, Miles Silmani,
and Alexandra Morelj
aJet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109;bInstitute of Environment, University of California, Los Angeles, CA 90095;
cWinrock International, Ecosystem Services Unit, Arlington, VA 22202;dCollege of Natural Resources, Colorado State University, Fort Collins, CO 80523;
eInstitute of Geography, School of Geosciences, University of Edinburgh, Edinburgh EH8 9XP, United Kingdom;fApplied GeoSolutions, Durham, NH
03824;gEarth and Biosphere Institute, University of Leeds, Leeds LS2 9JT, United Kingdom;hAgence Nationale des Parcs Nationaux, République Gabonaise, BP
20379 Gabon;iDepartment of Biology, Wake Forest University, Winston-Salem, NC 27106; andjEnvironmental Change Institute, University of Oxford, Oxford
OX1 3QY, United Kingdom
Edited* by Susan E. Trumbore, University of California, Irvine, CA, and approved May 5, 2011 (received for review December 28, 2010)
Developing countries are required to produce robust estimates of
forest carbon stocks for successful implementation of climate
change mitigation policies related to reducing emissions from de-
forestationanddegradation(REDD).Herewepresenta“benchmark”
map of biomass carbon stocks over 2.5 billion ha of forests on three
continents, encompassing all tropical forests, for the early 2000s,
which will be invaluable for REDD assessments at both project
and national scales. We mapped the total carbon stock in live bio-
mass (above- and belowground), using a combination of data from
4,079 in situ inventory plots and satellite light detection and ranging
(Lidar) samples of forest structure to estimate carbon storage, plus
optical and microwave imagery (1-km resolution) to extrapolate
over the landscape. The total biomass carbon stock of forests in
the study region is estimated to be 247 Gt C, with 193 Gt C stored
aboveground and 54 Gt C stored belowground in roots. Forests in
Latin America, sub-Saharan Africa, and Southeast Asia accounted for
49%, 25%, and 26% ofthe total stock, respectively. By analyzing the
errors propagated through the estimation process, uncertainty at
the pixel level (100 ha) ranged from ±6% to ±53%, but was con-
strained at the typical project (10,000 ha) and national (>1,000,000
ha) scales at ca. ±5% and ca. ±1%, respectively. The benchmark
map illustrates regional patterns and provides methodologically
comparable estimates of carbon stocks for 75 developing countries
where previous assessments were either poor or incomplete.
forest biomass|forest height|microwave and optical imaging|error
propagation|carbon cycling
D
tropical regions, accounted for 12–20% of global anthro-
pogenic greenhouse gas (GHG) emissions in the 1990s and early
2000s (1–4) and these processes also impact the future potential
of forests to remove additional carbon from the atmosphere (5–
7). Estimates of GHG emissions from deforestation require in-
formation on both the area of forest loss and the corresponding
carbon stock of the land that is cleared (8, 9). Both are consid-
ered challenging to quantify accurately (10). Much of the em-
phasis to date has focused on improving spatially represented
estimates of forest area loss (11, 12). To improve confidence in
estimated emissions, equal emphasis is needed on improving
spatially explicit estimates of carbon stored in forests, which
remain uncertain in tropical regions (13). The largest proportion
of this uncertainty is in estimates of aboveground biomass (14,
15), which accounts for 70–90% of forest biomass carbon (16),
and its spatial variability that depends on factors such as climate,
human and natural disturbance and recovery, soil type, and to-
pographical variations (14, 17). Reducing the uncertainty in
emissions estimates requires temporally constrained estimates of
forest carbon content at a spatial scale that is fine enough to
capture the variability over the landscape and is quantified at the
scale of disturbance affecting the forest. Such information would
improve project- and national-level carbon stock estimates as
well as assist in the development of baseline information re-
eforestation and forest degradation, located primarily in
quired for reducing emissions from deforestation and degrada-
tion (REDD) activities designed to curb greenhouse gas emissions
from the land use sector (15, 18).
Efforts to estimate the distribution of biomass rely on remote
sensing techniques due to the wide geographical extent of forests,
difficult accessibility, and the limited utility of field inventories
owing to the natural spatial variability of forest biomass (8, 9, 14).
New remote sensing approaches using light detection and ranging
(Lidar) and radio detection and ranging (radar) from airborne
sensors have been successful in providing high-resolution esti-
mates of forest carbon density for small areas (19–21). However,
the spaceborne sensors needed to use these approaches for large-
scale mapping and monitoring efforts will not be available before
the end of this decade (22). Until then, cost-effective mapping of
carbon stocks for project- and national-scale assessments will rely
on a combination of satellite imagery and ground-based inven-
tory samples of forest carbon density (14, 21).
Here, we report on our use of global forest height data mea-
sured by the Geoscience Laser Altimeter System (GLAS),
onboard the Ice, Cloud, and land Elevation Satellite (ICESat)
(23), in combination with other remote sensing data bases and
ground data, to model the spatial distribution of aboveground
standing biomass density (AGB) (in megagrams of mass per unit
area) in forests across three continental regions for the early
2000s. Our approach includes >3 million Lidar shots collected
along the ICESat orbital tracks. For calibration of GLAS Lidar
height to AGB and for validation of AGB distribution, we use
AGB data from 4,079 available inventory and research plots
distributed over the study region. We estimate belowground
biomass carbon in roots from AGB using tree allometry (24).
Our approach results in a benchmark map of forest carbon
density at 1-km resolution. The accuracy of carbon estimates for
every pixel is evaluated by propagating individual components of
uncertainty through the spatial analysis.
Results
Relating Forest Height to Biomass. AGB estimates were compiled
for 4,079 geo-referenced in situ forest plots (>0.1 ha) distributed
over three continents and restricted to inventory dates between
1995 and 2005 (Fig. S1 and Table S1). Of these data, 493 cali-
bration plots (298 in Latin America, 75 in Africa, and 120 in
Southeast Asia) were located under the GLAS Lidar shots or
within the same forest stands (Table S2). Data for these plots
Author contributions: S.S.S., N.L.H., S.B., and W.S. designed research; S.S.S., N.L.H., S.B.,
M.L., and E.T.A.M. performed research; S.S.S., N.L.H., S.H., L.W., M.S., and A.M. contrib-
uted new reagents/analytic tools; S.S.S., N.L.H., S.B., M.L., E.T.A.M., W.S., B.R.Z., W.B.,
S.L.L., S.P., L.W., M.S., and A.M. analyzed data; and S.S.S., N.L.H., S.B., and S.L.L. wrote
the paper.
The authors declare no conflict of interest.
*This Direct Submission article had a prearranged editor.
1To whom correspondence should be addressed. E-mail: saatchi@jpl.nasa.gov.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1019576108/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1019576108PNAS
| June 14, 2011
| vol. 108
| no. 24
| 9899–9904
ENVIRONMENTAL
SCIENCES

Page 2

included tree height, basal area, and AGB over an area of 0.25–
1.0 ha (SI Materials and Methods). We used these plots to de-
velop a power-law functional relationship (AGB = aHk, R2=
0.85, P < 0.0001) between the ground-measured Lorey’s height
(basal area weighted height of all trees >10 cm in diameter) and
AGB derived from allometric equations based on harvested trees
(Fig. 1A). The relationship, including all 493 calibration plots
over three continents, was tested for consistency in AGB pre-
dictions and root mean square error (RMSE) values against
region-specific relationships (Fig. 1 B–D). Regional estimates
outperformed the combined estimate, with the AGB predictions
for a single region obtained by using a relationship developed for
a different region, increasing the RMSE by 5–15%. The largest
error resulted when AGB in tropical Asia was estimated from
the combined continental relationship (12% increase in RMSE)
or from the regional relationship developed for Latin America or
Africa (15% and 13%, respectively). The combined relationship
provided a consistent approach to convert Lorey’s height to AGB
over all forests in the study region with an overall uncertainty of
±23.8%. However, region-specific relations had higher accuracy
and were used to convert a similar height metric obtained from
GLAS data to AGB with an estimation uncertainty of ±15.8% in
Latin America, ±21.7% in Africa, and ±25.1% in Asia. To create
AGB samples at mapping units of 100 ha (1-km resolution), we
aggregatedfiveormoreGLASLidarshotsin1-kmresolutioncells
toproduce160,918pixelsamplesthatwerecombinedwithground
inventory data to model the spatial distribution of AGB (SI
Materials and Methods).
Spatial Modeling of Forest AGB. We used a data fusion model
based on the maximum entropy (MaxEnt) approach (SI Materials
and Methods) as well as spatial imagery from multiple sensors
[moderate resolution imaging spectroradiometer (MODIS),
shuttle radar topography mission (SRTM), and quick scatter-
ometer (QSCAT)] on earth-observing satellites to extrapolate
AGB measurements from inventory plots and GLAS data to the
entire landscape. The model produced a map of AGB along with
estimates of uncertainty at a spatial resolution of 1 km (Fig. 2A).
Forest area in each AGB range class [in millions (106) of hec-
tares] and total AGB [in millions of metric tons (Mt) (1 t = 1
Mg)] were computed for each continent by using the most in-
clusive fractional tree cover threshold from the MODIS vege-
tation continuous field product (SI Materials and Methods) to
define forests (10% tree cover). Latin America has 47% of its
forest area in high biomass classes (AGB > 100 Mg ha−1) (Fig.
2B) compared with 27% in Africa (Fig. 2C) and 56% in Asia
(Fig. 2D). Nearly 75% of Africa’s forests are distributed in
woodland savannas and dry forests that contain low biomass
(AGB < 100 Mg ha−1). In Latin America, similar forests cover
the cerrado of Brazil, woodlands of Argentina and Chile, and dry
forests and savanna woodlands of Central and South America. In
tropical Asia, low biomass density forests (44% of the area) also
include secondary and degraded forests that are widespread
throughout the region (24).
Although not as extensive in terms of area, forests with AGB >
100 Mg ha−1contain most of the biomass on each continent, with
83%, 59%, and 82% of all biomass found in high biomass classes
in Latin America (Fig. 2E), Africa (Fig. 2F), and Asia (Fig. 2G),
respectively. Forests with AGB > 250 Mg ha−1contain nearly
half of all forest biomass on each continent (41% in Latin
America, 38% in Africa, and 50% in Asia). Forests in the highest
biomass class (AGB > 350 Mg ha−1) alone comprise a significant
percentage of total biomass (7.4% in Latin America, 8.7% in
Africa, 7% in Asia), and they cover areas of Papua New Guinea
and central Borneo (Indonesia) in Asia, western Congo Basin in
Gabon and southern Cameroon, eastern Democratic Republic of
Congo in Africa, eastern and northern Amazonia along the
Guiana Shield, and southwestern Peru in Latin America. Similar
general patterns have been observed from analysis of ground
surveys (6, 25, 26). However, comparison with an earlier map of
AGB over the Amazon Basin (14) revealed differences in central
and western Amazonia, particularly in the Rio Negro Basin. The
earlier Amazon AGB map showed higher biomass values in the
central region but with large uncertainties due to a lack of ground
plots (14). The forests of the Rio Negro Basin are dominated by
swamp forests (varzea) and white sand vegetation (caatinga and
campinarana) with relatively low biomass density. In western
Amazonia, around the foothill of the Andes, the earlier AGB
Fig. 1.
tween Lorey’s height and above-
ground biomass of calibration
plots at spatial scales comparable
to GLAS footprints. (A) Combined
relation from three continents
(AGB = 0.3104H2.0608, R2= 0.85,
P < 0.001). (B–D) Allometric re-
lation (B) from 298 plots in Latin
America (AGB = 0.4574H1.9701,
R2= 0.89, P < 0.001), (C) from 75
plots in sub-Saharan Africa in-
cluding woodland savanna (AGB =
0.3542H2.0528, R2= 0.85, P < 0.001),
and(D)from120plotsinSoutheast
Asia including plots in secondary
and fragmented forests (AGB =
0.0508H2.5213, R2= 0.79, P < 0.001).
Allometric relations be-
9900
| www.pnas.org/cgi/doi/10.1073/pnas.1019576108 Saatchi et al.

Page 3

map predicted much lower biomass compared with the new
benchmark map, but with higher uncertainty. These forests were
mapped with less uncertainty in the new benchmark map due to
the extensive GLAS Lidar sampling of forest structure.
National Assessment of Carbon Stocks. To estimate total biomass
carbon stocks and produce a forest carbon “benchmark” map
against which future changes can be assessed, we calculated
belowground biomass (BGB) as a function of AGB (BGB =
0.489 AGB0.89) (Fig. S2) and estimated total carbon as 50% of
total biomass (AGB + BGB) (27) (Fig. 3A). We defined forest
extent using three fractional cover thresholds (10%, 25%, and
30% tree cover) based on the range of thresholds used by in-
dividual Parties to the United Nations Framework Convention
on Climate Change (UNFCCC) proceedings (28) and estimated
total carbon stored in forests of each country in the study area
(Table S3). In Table 1, we present the five countries per conti-
nental region that contain the highest forest carbon stocks, along
with continental totals. From this analysis, we estimate the total
forest biomass carbon stocks at 10% tree cover as 247 Gt C, with
193 Gt C in AGB and 54 Gt C in BGB. Forests in Latin America
are the most extensive and contain ∼49% of the total biomass
carbon, followed by 26% in Asia and 25% in Africa. Applying
a higher tree cover threshold (30%) eliminates large areas of
savanna woodlands in Africa from the forest domain and reduces
the total carbon stock to 208 Gt C (16% reduction with 163 Gt C
for AGB and 45 Gt C for BGB). Among the countries analyzed,
Brazil, the Democratic Republic of Congo, and Indonesia have
the largest area of forest as well as the highest carbon stocks (62,
24, and 24 Gt C, respectively, at 10% tree cover). These esti-
1
2
3
4
5
6
7
8
9
10
11
A
BC
E
F
G
D
Fig. 2.
bounded at 30° north latitude and 40° south latitude to cover forests of Latin America and sub-Saharan Africa and from 60° to 155° east and west longitude.
The map was colored on the basis of 25–50 Mg ha−1AGB classes to better show the overall spatial patterns of forest biomass in tropical regions. Histogram
distributions of forest area (at 10% tree cover) for each biomass class were calculated by summing the pixels over Latin America in B, Africa in C, and Asia in D.
Similarly, total AGB for each class was computed by summing the values in each region with distributions provided for Latin America in E, Africa in F, and Asia
in G. All error bars were computed by using the prediction errors (SI Materials and Methods) from spatial modeling.
Distribution of forest aboveground biomass (circa 2000). (A) Forest aboveground biomass is mapped at 1-km spatial resolution. The study region was
Saatchi et al.PNAS
| June 14, 2011
| vol. 108
| no. 24
| 9901
ENVIRONMENTAL
SCIENCES

Page 4

mates improve on forest carbon stock estimates reported pre-
viously (8, 13–15, 17, 18, 25) by providing a traceable and sys-
tematic approach to geographically locate the stock estimates for
further monitoring and verification. The forest definitions chosen
here using tree cover thresholds can readily change the estimates
of total carbon and area-weighted carbon densities at national
and regional scales.
Uncertainty Analysis. We assess the accuracy of the biomass car-
bon estimates by calculating the error as the difference between
the true mean biomass value (bootstrapped samples of ground
and Lidar-estimated AGB) and the predicted biomass value
(mapped at 1-km grid cell resolution) and propagating these
errors through the spatial modeling process (SI Materials and
Methods). Errors in the distribution of forest aboveground bio-
mass can be random or systematic in nature and can include the
following: (i) observation errors associated with the uncertainty in
estimates of Lorey’s height from GLAS Lidar, errors associated
with estimating AGB derived from GLAS Lidar height, and
errors in estimating BGB from AGB (27); (ii) sampling errors
associated with the spatial variability of AGB within a 1-km pixel
and the representativeness and size of inventory plots and GLAS
pixels over the landscape (29); and (iii) prediction errors associ-
ated with spatial analysis and mapping of AGB from significant
contributions from satellite imagery (Fig. S3) (14, 30). We
combined these three types of errors (SI Materials and Methods)
to quantify the uncertainty of total biomass carbon stock as the
95% bootstrapped confidence interval at the 1-km pixel level
(Fig. 3B). The overall uncertainty in mapping AGB at the pixel
scale averaged over all continental regions is estimated at ±30%,
but it is not uniform across regions or AGB ranges (±6% to
±53%) and depends on regional variations of forests, quality of
remote sensing imagery, and sampling size and distribution of
available ground and GLAS data. However, when averaged over
all AGB ranges, regional uncertainties were comparable: ±27%
over Latin America, ±32% over Africa, and ±33% over Asia
(Fig. S4). The uncertainty in total carbon stock at the pixel scale
averaged ±38% over all three continents after errors associated
with BGB estimation were included in the analysis.
We computed the uncertainty around carbon estimates at
national and regional scales by propagating errors associated
with observation, including the errors associated with BGB
estimates, sampling, and prediction. The uncertainty of carbon
stock estimates at the national level was calculated as the square
root of the sum of per-pixel errors for all pixels within the na-
tional boundary. This process reduced the relative errors as
sample area increased. The national estimates were found to be
constrained to within ±1% of the total carbon stock obtained
Fig. 3.
colored on the basis of a 12–25 Mg C ha−1range to show the spatial patterns. (B) The uncertainty of the benchmark map is estimated using error propagation
through a spatial modeling approach. The uncertainty is given in terms of plus or minus percent and it includes all errors associated with prediction from
spatial modeling, estimation of Lorey’s height from GLAS, estimation of AGB from Lorey’s height, errors from pixel level variations, and errors associated with
BGB estimation (SI Materials and Methods).
Benchmark map of carbon stock and uncertainty. (A) Forest carbon stock defined as 50% of AGB + BGB is mapped at 1-km pixel resolution and
9902
| www.pnas.org/cgi/doi/10.1073/pnas.1019576108 Saatchi et al.

Page 5

from averaging the pixel mean carbon values (Table S2). Spatial
aggregation using different size windows shows that the error
stays bounded to within ±5%, even at 10,000 ha (100 1-km
pixels), due to the increased sample area. These errors do not
include any systematic bias that may exist in biomass carbon
allometry for AGB and BGB, in the in situ data inputs into the
allometric equations (e.g., accurate botanical identifications for
assigning wood density information, the wood density data itself,
measurements of structure, etc.) or in the spatial sampling of
forest biomass (SI Materials and Methods).
Discussion
The benchmark map provides a spatially refined and methodo-
logically comparable carbon stock estimate for forests across 75
developing countries and improves upon previous assessments
based on often old and incomplete national forest inventory data
(27) and earlier spatial products (Fig. S5) (8, 14). Systematic
quantification of the errors improves and constrains the pan-
tropical estimate of total tropical forest biomass carbon (247 Gt C
at 10% tree cover) and similarly national-scale carbon stocks
(Table S3). Given the large uncertainty of forest carbon estimates
for individual 1-km pixels (>30%), the map should be used pri-
marily for national- and project-scale assessments (>10,000 ha).
Reducing the uncertainty at the pixel level would require (i)
higher-resolution data from future spaceborne radar and Lidar
measurements to capture the spatial variability of forest structure
and improve the estimation of aboveground biomass at spatial
scales of <1 ha (21, 22) and (ii) in situ data with appropriate
sampling schemes to match the scale of the remotely sensed
measurements. Existing airborne Lidar and radar data can provide
high-resolution estimates of project-scale forest carbon stocks (19,
21), but are likely to be unsuitable for large-scale wall-to-wall
national- or continental-scale forest carbon monitoring systems.
Indeed, one key issue with any attempt to quantify forest
carbon stocks is assessing the uncertainties inherent in estimating
AGB from ground plots (29). Throughout this paper data asso-
ciated with ground-based inventory plots have been assumed to
be error-free. Although the methods used here do represent best
estimates, estimating biomass from stem diameter, using conti-
nental or pan-tropical allometric equations, will introduce non-
random errors, which may be significant and systematic (6, 29).
Given the paucity of information on plot-based errors, we have
Table 1. Estimates of forest carbon stocks in the five largest national pools for each continent in the study region
Canopy cover threshold
10%25% 30%
CountryArea (Mha)
Total C
(Gt C)
C density
(Mg C/ha)Area (Mha)
Total C
(Gt C)
C density
(Mg C/ha) Area (Mha)
Total C
(Gt C)
C density
(Mg C/ha)
Democratic Republic of
Congo
Cameroon
Republic of Congo
Gabon
Angola*
Total sub-Saharan Africa
Brazil
Peru
Colombia
Venezuela
Bolivia
Total Latin America
Indonesia
Myanmar
Papua New Guinea
India†
Malaysia
Total Asia and Oceania
Total study region
20524 118 177 23128 16422134
36
28
24
73
775
596
80
84
61
74
1,209
165
5
4
4
3
129
144
160
44
80
102
153
122
118
84
99
142
146
147
89
172
137
100
30
24
22
42
539
481
75
67
50
65
977
127
42
37
43
25
359
1,875
4
4
4
3
142
160
164
66
93
116
158
138
134
90
112
155
155
152
104
179
155
115
27
23
21
34
447
442
73
64
47
61
893
121
40
36
36
25
336
1,677
4
4
4
2
151
162
165
70
106
123
160
141
139
94
119
158
157
153
112
180
159
124
62
61
12
10
7
6
50
56
12
9
7
6
48
54
12
9
7
6
120
23
110
20
107
19
49
43
63
30
7
6
6
5
6
6
4
5
6
6
4
4
474
2,458
6556 54
208 246 215
All carbon values were calculated by using the pixel-based AGB value to compute BGB and total carbon (AGB + BGB). Carbon density (Mg C ha−1) values
were calculated from the ratio of total carbon to total forest area at national or regional levels.
*Central African Republic replaces Angola as no. 5 in Africa using 25% and 30% canopy cover thresholds.
†India switches from rank 4 to rank 5 using 25% and 30% canopy cover thresholds.
Fig. 4.
canopy cover. Distribution of total (above- and belowground) biomass carbon
stocks is shown. Carbon was computed by using the pixel-based AGB values to
estimate root biomass and summing values across thresholds of percentage of
tree cover (10%, 25%, and 30%) obtained from intersecting the 2001 MODIS
vegetation continuous field (VCF) product (SI Materials and Methods) with the
benchmark map. Similar distributions for each continent were produced sep-
arately. The error bars are based on the difference between the upper and
lower limits of predicted AGB for computing total carbon (Table S2).
Total carbon stock across the study region at three thresholds of
Saatchi et al. PNAS
| June 14, 2011
| vol. 108
| no. 24
| 9903
ENVIRONMENTAL
SCIENCES