Available via license: CC BY-NC 4.0
Content may be subject to copyright.
ENVIRONMENTAL STUDIES Copyright © 2018
The Authors, some
rights reserved;
exclusive licensee
American Association
for the Advancement
of Science. No claim to
original U.S. Government
Works. Distributed
under a Creative
Commons Attribution
NonCommercial
License 4.0 (CC BY-NC).
Congo Basin forest loss dominated by increasing
smallholder clearing
Alexandra Tyukavina
1
*, Matthew C. Hansen
1
, Peter Potapov
1
, Diana Parker
1
, Chima Okpa
1
,
Stephen V. Stehman
2
, Indrani Kommareddy
1
, Svetlana Turubanova
1
A regional assessment of forest disturbance dynamics from 2000 to 2014 was performed for the Congo Basin
countries using time-series satellite data. Area of forest loss was estimated and disaggregated by predistur-
bance forest type and direct disturbance driver. An estimated 84% of forest disturbance area in the region is
due to small-scale, nonmechanized forest clearing for agriculture. Annual rates of small-scale clearing for agri-
culture in primary forests and woodlands doubled between 2000 and 2014, mirroring increasing population
growth. Smallholder clearing in the Democratic Republic of the Congo alone accounted for nearly two-thirds
of total forest loss in the basin. Selective logging is the second most significant disturbance driver, contributing
roughly 10% of regional gross forest disturbance area and more than 60% of disturbance area in Gabon. Forest
loss due to agro-industrial clearing along the Gulf of Guinea coast more than doubled in the last half of the
study period. Maintaining natural forest cover in the Congo Basin into the future will be challenged by an
expected fivefold population growth by 2100 and allocation of industrial timber harvesting and large-scale ag-
ricultural development inside remaining old-growth forests.
INTRODUCTION
The Congo Basin is home to the second largest massif of humid trop-
ical forests (HTFs) after the Amazon, performing globally important
ecosystem services and providing livelihood to the regional population
(1).ThecriticalroleoftheCongoBasin rainforests in climate regula-
tion and biodiversity conservation is recognized internationally and
has led to establishing collaborative sustainable forest resource man-
agement initiatives such as the Central Africa Regional Program for
the Environment and the Regional Programme for the Conservation
and Rational Use of Forestry Ecosystems in Central Africa. Understand-
ing forest disturbance dynamics in the region as a whole and on the
national scale is essential for policy-making and land use planning.
The presented study is focused on the six Congo Basin tropical
rainforest countries, namely, Cameroon (CAM), the Central African
Republic (CAR), the Democratic Republic of the Congo (DRC), Equa-
torial Guinea (EQG), Gabon (GAB), and the Republic of the Congo
(RoC). Differences in forest disturbance dynamics and drivers among
the Congo Basin countries vary owing to geographic, economic, and
demographic conditions (table S1); development history; and current
policy and institutional factors (2,3). Historically, forest loss in the
Congo Basin has been strongly linked to rural populations and sub-
sistence agriculture (4,5). However, per-capita food production and
food availability vary between Congo Basin countries (Table 1).
CAM stands out as a country with improving food production and,
in the regional context, relatively strong export and import sectors.
The oil-exporting countries GAB, EQG, and RoC form a group of
countries exhibiting decreasing food production. High food import
levels for RoC and especially GAB reflect the use of oil earnings to
support domestic food consumption. Oil exports account for 40 to
50% of the gross domestic product (GDP) in GAB and RoC and
80% of the GDP in EQG. Such high dependence on oil exports has
implications for economic and political stability in the face of price
shocks, such as those of 2014 to 2016 (6).
CAR has the lowest human development index of all countries
(table S1), reflected in Table 1 by marginal food exports and imports,
and the highest per-capita food aid shipments in the region. DRC is
unique in its declining food production, low food exports and imports,
and lack of food aid shipments. DRC isofparticularimportance,asit
is home to 60% of the remaining Congo Basin humid tropical rain-
forest (7). DRC is also unique because of its population pressure and
recent history of conflict and insecurity. The only country similar to
DRC in terms of persistent conflict, insecurity, and statelessness is the
CAR (table S1). However, DRC dwarfs CAR in terms of total popu-
lation and HTF resources. With more than 70 million people, DRC is
morethantwicethepopulationofCAM,CAR,EQG,GAB,andRoC
combined (table S1). For the citizens of DRC, which, along with CAR,
has a human development index in the bottom 10% of all countries,
there are few livelihood options. The vast majority of the population
consists of smallholder farmers, who feed not only themselves but also
nearby towns and cities (3,8).
Given the different economic, political, and social contexts within
Congo Basin rainforest countries, we can expect within-region variations
in land cover and land use change. For example, mineral and petroleum
exports tend to discourage deforestation, as oil wealth enables food
importation and reduced domestic agricultural output, with GAB a
clear example (9). Low populations also help to ensure low rates of
disturbance outside of commercial logging operations for countries
like GAB, RoC, and EQG (10). Recent investments in agro-industrial
development, mainly palm oil, are a relatively new threat to primary
forests in the Congo Basin (11). The Tropical Forest Alliance (12),
which seeks to implement sustainable palm oil development in Africa,
includes CAR, DRC, and RoC as signatories, but not CAM, EQG, or
GAB. The Gulf of Guinea countries have logistical advantages in this
sector over interior Congo Basin countries, mainly due to proximity to
ports. In GAB, the creation of special economic zones around ports is
part of new and ambitious development plans that include palm oil
expansion (13). Land use and land cover change in low-population,
forest resource–rich Congo Basin countries is likely attributable to ex-
tractive industries such as logging or agro-industrial development,
such as palm oil.
1
Department of Geographical Sciences, University of Maryland, College Park, MD
20740, USA.
2
College of Environmental Science and Forestry, State University of
New York, Syracuse, NY 13210, USA.
*Corresponding author. Email: atyukav@umd.edu
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 1of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
The present study uses a sample-based analysis to estimate the
rates of forest disturbance in Congo Basin countries between 2000
and 2014 and to attribute direct land use drivers to forest loss in differ-
entforesttypes.“Forest”is defined in the current study as any woody
vegetation exceeding 5 m in height and 25% in canopy cover at a 30-m
resolution (see the “Definitions”section). Global-scale (14,15) and
national-scale (16) wall-to-wall forest change maps provide diagnostic
information on the extent of land cover change in forests. However, all
maps contain errors and thus may underestimate or overestimate the
area of forest change. For example, pan-tropical sample-based studies
(17,18) report almost twice as much gross tree cover loss in Africa in
2000 to 2010 compared to the Hansen et al. map (14) for the same
time period, indicating significant map omission errors. Per good
practice recommendations, land cover change area estimates should
be derived from a probability sample of reference data (19,20) rather
than from counting map pixels, where reference data are defined as the
best practically available assessment of ground condition. Following
this guidance, the current study uses a stratified sampling approach
to estimating forest loss area, with the Hansen et al.(14) global forest
loss map used to construct strata to improve precision of the estimates
(see Materials and Methods). A probability sample of ground obser-
vations performed within months of detected forest disturbance
events and follow-up visits in the subsequent years would have been
ideal to determine the initial direct driver of forest loss and possible
future land cover and use transitions. This method was prototyped
by our team in a series of rapid ground surveys in the Mexican Yucatan
and Argentina. In the Congo Basin, however, such ground visits are less
feasible due to the lower quality and coverage of the road network and
safety concerns. The analysis of time series of all available Landsat
observations for the study period supplemented with detailed very
high resolution (1 m or better) imagery implemented in the current
study is a more cost-efficient alternative to ground surveys. Previous
pan-tropical studies of direct drivers of forest loss rely primarily on
national reports and literature reviews (21,22), which may be affected
by inconsistent definitions, national politics, and poor quality of
underlying data. Existing regional loss driver studies in the Congo
Basin have used case study reviews (5), statistical analysis of auxiliary
geospatial data sources (7,23), and interviews with experts (2), all of
which might also be affected by methodological inconsistencies and
data quality issues. Using remotely sensed imagery directly to derive
information on direct drivers of forest loss eliminates some of these
issues and allows estimating loss drivers across national borders
using the same data, method, and definitions. We first prototyped
this approach in the Brazilian Legal Amazon (24) by identifying di-
rect forest disturbance drivers and predisturbance forest types from a
sample of 10,000 Landsat pixels.
Direct drivers of forest disturbance in this study are defined as hu-
man activities or biophysical events that directly affect forest cover and
lead to canopy loss. Some direct drivers are distinguished using remote
sensing data relatively easily (e.g., road construction, settlement expan-
sion, mining, industrial selective logging, wildfires, and river meander-
ing). For other drivers, such as the clearing of forest for agricultural
activities, it is more difficult to identify the specific type of activity in
the absence of information on land tenure (smallholders versus in-
dustrial enterprises), type of crop or livestock (subsistence versus com-
mercial), and fallow cycle length. In these cases, we use a set of criteria
distinguishable in satellite imagery, such as size of individual clearing
(small-scale versus large-scale clearing of agriculture as a proxy to
smallholder versus industrial agriculture) and presence/absence of for-
est regrowth (to distinguish between semipermanent and rotational
agriculture). Less common clearing of forests for charcoal production
is included in the “small-scale clearing for rotational agriculture”class,
since the size of clearings and regrowth patterns of the two classes are
similar and often colocated. Of direct drivers associated with forest
degradation, we are able to detect only industrial selective logging
and stand-replacement fires but cannot quantify the area affected by
low-intensity artisanal logging, fuel collection, undercanopy livestock
grazing, and low-intensity fires not resulting in significant canopy loss.
Distinguishing forest loss by predisturbance forest type is impor-
tant because forests differ significantly in their carbon storage and bio-
diversity value. It is particularly significant to distinguish between high
conservation value primary forests and secondary forests, which are a
part of a shifting cultivation cycle. Most of the previous sample-based
studies do not differentiate primary and secondary forests (7,23).
FACET (Forêts d’AfriqueCentraleEvaluéespar Télédétection) atlases
distinguish primary and secondary forests and woodlands, but exist
Table 1. Food production and trade indicators. Data source: FAOSTAT Database (http://www.fao.org/faostat). Food production index 2014 (2004 to 2006 =
100) shows the relative level of the aggregate volume of agricultural production for the year 2014 in comparison with the base period 2004 to 2006.
Country
Food production index 2014
(2004–2006 = 100),
net per capita
Agricultural products export value
base price per capita, 2013
($ per person)
Agricultural products import value
base price per capita, 2013
($ per person)
Food aid shipments, 2014, per
capita (kg per person)
CAM 126 26 38 0.6
CAR 95 2 7 6.1
DRC 78 0.2 8 1.0
EQG 90 ———
GAB 73 28 215 —
RoC 82 2 58 1.8
Brazil 123 263 24 —
Indonesia 125 78 35 —
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 2of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
only for DRC and RoC (http://carpe.umd.edu/carpemaps/) and there-
fore can only be used as stratifiers in national-level, sample-based stu-
dies (25). The Hansen et al.(14) global forest loss map, which is a
stratifier in the current study, has been criticized for not distinguishing
between types of tree cover (26). We have addressed this concern in
past sample-based studies first by distinguishing between natural and
human-managed forests (17) and later by using a more detailed clas-
sification of forest types (24). The forest type classification in the cur-
rent study follows the more detailed classification approach and
includes the following five types of predisturbance forest cover (see
the “Definitions”section; fig. S2): (i) primary and mature secondary
dense HTFs, (ii) young secondary dense HTFs, (iii) primary wood-
lands and dry forests, (iv) secondary woodlands and sparse secondary
HTFs, and (v) tree and palm plantations.
To summarize, the objectives of the current study are the following:
(i) estimate 2000 to 2014 forest loss area in the six Congo Basin coun-
tries and temporal loss trends using a recommended good practice
probability sampling approach; (ii) identify direct drivers of forest dis-
turbance distinguishable in remote sensing imagery using the same
data, method, and definitions across national borders; and (iii) com-
pare annual rates of forest disturbance in different forest types across
the region.
RESULTS
The estimated 2000 to 2014 forest loss area in the study region is
16.6 ± 0.5 Mha (million hectares ± 1 SE) (table S2A). DRC alone
contributes a higher percentage of forest loss than the other five coun-
tries combined (69.1 ± 1.7%), followed by CAM (9.9 ± 1.2%), RoC (8.2 ±
1.2%), CAR (7.4 ± 0.8%), GAB (4.7 ± 0.9%), and EQG (0.7 ± 0.2%). The
average estimated annual area of forest clearing at the national level thus
ranges from almost 1 Mha in DRC (817 ± 28 thousands ha) to about ten
thousand hectares in EQG (8 ± 3 thousands ha).
Forest cover loss by direct driver
Small-scale forest clearing for agriculture is the largest direct driver
of forest disturbance in the region, contributing about 84% of the total
2000 to 2014 forest loss area (table S2A). This includes clearing for
rotational agriculture (82.1 ± 1.8%) and semipermanent conversion
of woody vegetation into cropland (2.1 ± 0.5%), both of which could
represent subsistence farming or production of commercial crops (27).
Small-scale forest clearing is likely nonmechanized, which is supported
by the size of individual clearings (median annual small-scale clearing
size is estimated at 1.8 ha) and the lack of access roads visible in very
high resolution imagery. At the national scale, small-scale clearing for
agriculture is the main direct disturbancedriverinallcountriesexcept
GAB (Fig. 1). In DRC and in CAR, more than 90% of all forest loss is
due to small-scale clearing for rotational agriculture. Semipermanent
conversion to cropland is much less common than rotational agricul-
ture, with CAM being the only country in which this conversion com-
prised more than 10% of the total forest disturbance area.
Large-scale clearing for agriculture (annual clearing size, >10 ha)
constitutes only about 1% (0.9 ± 0.2%) of the overall forest loss area
(table S2A and Fig. 1). This type of clearing, which is likely mechanized,
includes forest clearing for tree and palm plantations and industrial
pastures. CAM is the leading contributor to large-scale agro-industrial
clearing of the region (56.5 ± 9.4%), followed by DRC (21.3 ± 7.4%),
GAB (11.5 ± 5.4%), RoC (8.5 ± 7.9%), and EQG (2.2 ± 2.1%). Agro-
industry in the region has been experiencing a new wave of develop-
ment since 2004 (11). Therefore, large-scale agro-industrial clearing
is likely to become a more significant contributor to forest loss in
the future.
Construction accounts for about 1.5% of forest loss in the region,
which includes residential and commercial (1.0 ± 0.3%) and road (0.4 ±
0.1%) construction (table S2A and Fig. 1). The largest contribution of
construction to forest loss on the national level is observed in EQG
(18.7% of the national forest loss area), which is likely related to the
country’s large development projects of the last decade, such as con-
struction of the new capital city of Oyala (28).
Mining is a very rare forest disturbance driver, accounting for only
0.04±0.03%ofthetotalforestlossareaintheregion.Theestimateof
forest loss area due to mining has low relative precision (SE expressed
as percentage of driver area is 71%) because it is based only on two
sampled pixels: one in EQG and one in CAR. Quantifying forest loss
in the Congo Basin due to mining with high relative precision would
require a stratification specific to mining, for example, a combination
of existing mining concession boundaries and areas of semipermanent
bare ground gain, derived from remote sensing. In absolute terms, the
rarity of mining in our sample of 10,000 pixels gives us a good idea
regarding the magnitude of the forest loss due to mining (the 95%
confidence interval does not exceed 14,736 ha).
Industrial selective logging constitutes 9.5 ± 1.6% of forest loss area
in the region (table S2A and Fig. 1). The largest contributors are RoC
(39.5 ± 9.0%), GAB (30.7 ± 8.4%), and CAM (22.8 ± 7.9%), followed
by DRC (6.2 ± 4.6%) and CAR (<1%). Area affected by industrial se-
lective logging is defined using a 120-m buffer around logging damage
and roads visible in Landsat imagery (see Materials and Methods),
and therefore, the industrial selective logging area estimate is likely
conservative. Selective logging does not imply complete canopy loss
and hence does not result in the same carbon emissions as stand-
replacement forest disturbance drivers, which should be taken into
consideration when interpreting the carbon implications of the cur-
rent area estimates.
Fires, resulting in the loss of canopy, but not followed by agricul-
tural activities, account for 3.8 ± 1.0% of forest loss area in the region
(table S2A and Fig. 1). These are likely escaped agricultural fires or
fires set for hunting purposes: 78% of the sampled pixels identified
as “fire”were adjacent to the forest edge and human activities (roads,
settlements, and active fields). DRC contributes most of the region’s
fire disturbance (78.8 ± 7.4%), followed by RoC (8.4 ± 4.4%), CAR
(6.7 ± 3.5%), CAM (3.7 ± 2.4%), and GAB (2.4 ± 2.4%). Natural forest
disturbances, including windfalls and river meandering, contribute
only about 0.13 ± 0.07% to the total forest loss area.
Forest cover loss by predisturbance forest type
Forest loss in primary and mature secondary dense HTFs in 2000 to
2014 accounts for 43.7 ± 1.7% of forest loss area in the region (table S2A),
followed by clearing in young secondary dense HTFs (34.9 ± 1.6%),
primary woodlands and dry forests (16.8 ± 1.4%), secondary woodlands
and sparse secondary HTFs (4.2 ± 0.8%), and tree plantations,
established by the year 2000 (0.4 ± 0.1%). At the national scale, clearing
of primary and mature secondary dense HTFs is prevalent in GAB,
RoC, and CAM (Fig. 2). The extent of loss of primary and mature
secondary dense HTFs in DRC is comparable to the reclearing of young
secondary dense HTFs (Fig. 2), which indicates the presence of large
established shifting cultivation areas. Forest loss in CAR occurs mainly
within primary woodlands and dry forests. EQG has the lowest propor-
tion of forest loss within primary vegetation among all countries (Fig. 2).
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 3of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
Within primary and mature secondary dense HTFs, 70.3 ± 3.2%
of forest loss is due to small-scale clearing for rotational agriculture
(table S2A), which is an indication of shifting cultivation expanding
into previously undisturbed forest. This finding is consistent with
Molinario et al.(29), who found that the area under shifting culti-
vation and rural settlements in DRC grew by 10% between 2000 and
2010. Selective logging is also a significant contributor to primary and
mature secondary dense HTF loss (21.7 ± 3.2%), followed by fire (5.8 ±
1.8%), large-scale agro-industrial clearing (1.3 ± 0.3%), and road con-
struction (0.5 ± 0.2%).
Young secondary dense HTFs and secondary woodlands and
sparse secondary HTFs are cleared almost exclusively in the context
of small-scale rotational agriculture (97.8 ± 0.5% and 95.9 ± 2.3%,
respectively; table S2A). Primary woodlands and dry forests are cleared
for both rotational (78.0 ± 4.3%) and semipermanent agriculture
(12.5 ± 2.9%). Old tree plantations are either cleared and replanted
again (68.6 ± 16.0%) or converted to small-scale shifting cultivation
(31.4 ± 16.0%).
Temporal trends of forest loss
Annual forest loss trends are analyzed at a regional scale by distur-
bance driver and predisturbance forest type (table S2B and Fig. 3).
Among the major disturbance categories (Fig. 3), small-scale clearing
for rotational agriculture is increasing both in primary and mature
secondary dense HTFs and in primary woodlands and dry forests. Ac-
celerating rates of small-scale clearing in these forest types are likely
linked to increasing population pressure (Fig. 4). However, at the na-
tional scale, not all countries display the same increasing trend of
small-scale clearing in primary forests and woodlands (Table 2). In
GAB, where industrial selective logging accounts for more forest loss
than small-scale clearing for agriculture (Fig. 1), encroachment of
small-scale agricultural activities into primary forests and woodlands
has slowed down by 2014 (Table 2). In CAR, small-scale clearing for
agriculture in primary forests and woodlands first accelerated and
then slowed down again (Table 2), possibly because of the civil war,
which started in 2012.
Small-scale forest clearing for rotational agriculture in secondary
forests displays a decreasing trend (Fig. 3), which is explained by
the way young secondary forests are defined in the current study
(see the “Definitions”section and Discussion). Industrial selective log-
ging in primary and mature secondary dense HTFs peaked at the
beginning and at the end of the study period (Fig. 3). Lower logging
rates in 2007 to 2008 may be linked to the decreased demand for tim-
ber during the global financial crisis (30).
DISCUSSION
Drivers of forest disturbance
Results of the current study provide a quantitative assessment of
regional socioeconomic drivers resulting in forest loss. Congo Basin
forests are being cleared primarilybymanualmeans:Nonmechanized
small-scale forest clearing for agriculture is responsible for 84% of the
total forest loss between 2000 and 2014. In the least-developed coun-
tries, DRC and CAR, small-scale clearing is even more dominant
(more than 90%). The dominance of local populations and subsistence
farming within the Congo Basin distinguishes it from deforestation
Fig. 1. Forest disturbance driver. (A) Reference disturbance driver for each sampled pixel. (B) National estimates of 2000 to 2014 forest loss area by disturbance
driver. Area estimates along with SEs are presented in table S2A. Forest clearing for small-scale rotational agriculture includes clearing for charcoal production, the
contribution of which does not exceed 10% of the class area (42).
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 4of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
dynamics in the Amazon Basin and Insular Southeast Asia. The Congo
Basin has historically lagged the Amazon Basin and Insular Southeast
Asia, the world’s other large remaining HTF regions, in the amount
and rate of tropical forest clearing. Table 1 includes data for Brazil
and Indonesia, the countries with the highest deforestation totals in
recent history, and illustrates dramaticallyincreasingfoodproduction
and agricultural export totals compared to the underdeveloped econo-
mies of the Congo Basin. Agro-industrial land use drivers of clearing in
Brazil are mainly pasture for cattle production and cropland for
soybean cultivation (31). In Indonesia, palm oil and forestry are the
principal land uses replacing primary forests (32). While agro-industrial
clearing has not been significant in the Congo Basin, there is nascent
large-scale clearing of forests for palm oil (11). For DRC, where small-
holder farming predominates, the main implement for clearing forests
remains the axe. From 2010 to 2014, the area of primary forest clearing
in DRC was equivalent to 54% of Indonesia’sand46%ofBrazil’sarea
of primary forest clearing (33). The fact that DRC’s clearing is largely by
hand and still equal to roughly one-half that of the two dominant de-
forestation countries is an indication of the scale of smallholder
cropland expansion in DRC.
The low level of development and political instability in the two
smallholder-dominated forest loss countries, DRC and CAR, is reflected
in forest clearing rates that are largely correlated with population
growth. Resulting population pressure on land resources can lead to
environmental degradation in efforts to produce sufficient food (34).
The increasing rate of forest loss due to smallholder agriculture reflects
the lack of agricultural intensification in the Congo Basin, which could
compensate for increasing population densities (35,36). Alternatives to
shifting cultivation practices are of particular importance given growing
populations, especially in DRC. Expected population growth is due to
persistently high fertility levels and increasing longevity. Recent United
Nations’population projections for DRC estimate 197 million people by
2050 and 379 million by 2100, when DRC is expected to be the fifth
most populous country in the world. Under the assumption that pop-
ulation growth continues to correlate with the increase in annual
primary forest loss area, all of DRC’s primary forests will have been
cleared by 2100. The strategy for survival in DRC is best reflected in
the concept of “Article 15,”the popular and imagined 15th article to
the 14-article constitution of the 32-year Mobutu dictatorship. Article
15 means “figure it out”and represents “an implicit social pact between
the state and its citizens since it allowed the former [the state] to retire
from public life and from its functions”(37).Thepracticaloutcomehas
been self-reliance in nearly every aspect of life for the citizens of DRC.
Since the end of the Mobutu regime, the ongoing conflict within DRC
has only exacerbated the challenges of DRC residents, with escalating
hunger and malnutrition due to prolonged conflict and displacement
(38). In terms of land cover and land use change, self-reliance in re-
sponse to statelessness is evident in elevated forest disturbance rates
compared to other Congo Basin countries, as the entire rural population
attempts to eke out a subsistence livelihood. Thefate of remaining HTFs
in DRC will be a function of alternative development strategies given a
daunting population growth trajectory.
The likely expansion of agro-industrial development will add to
the demographic challenge. Agro-industrial clearing was found primar-
ily in the Gulf of Guinea countries with 70% of the total 2000 to 2014
large-scale clearing for agriculture occurring after 2007 (table S2B).
Annual rates of industrial selective logging in primary and mature sec-
o
ndarydenseHTFshavealsobeensteadilyincreasingsince2007
(Fig. 3). This indicates that while subsistence agriculture is still the
leading driver of forest loss in the region, future industrial development
Fig. 2. Predisturbance forest type. (A) Reference predisturbance type for sampled pixels identified as forest loss. (B) National estimates of 2000 to 2014 forest loss
area by predisturbance forest type. Area estimates expressed in hectares along with SEs are presented in table S2A.
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 5of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
will bring new challenges to forest resources management. For example,
development of infrastructure for resource extraction and agro-industry
will likely facilitate encroachment of subsistence agriculture into previ-
ously undisturbed (intact) forest areas. Potapov et al.(39)illustratedsev-
eral dynamics associated with the expansion of certified sustainable
logging in the northern RoC, including expanding agriculture around
logging towns and increased infrastructure, such as paved roads and a
dam for hydroelectric power generation. Increasing infrastructure de-
velopment investment from China, India, and the Gulf states in the re-
cent years (40) may accelerate this process. Land use planning that
minimizes the conversion of natural forest cover for agro-industry will
serve to mitigate this nascent and growing threat to primary forests.
Temporal disaggregation of forest loss in different forest types
presented in the current study provides definitive information on
forest loss trends in the region. For example, the Hansen et al.(14)
global forest loss map underestimates forest loss area in the early 2000s
and overestimates forest loss in the 2010s, as demonstrated for the
Congo Basin (current study, fig. S7) and for the Brazilian Legal
Amazon (24). Disaggregation of forest loss area by disturbance driver
and predisturbance forest type provides context to previous forest loss
estimates. From previous studies (16,17,25), the total estimated area
of secondary forest loss in DRC for the 2000 to 2010 and 2000 to 2012
intervals exceeds the total area of primary forest loss to a greater de-
gree than the presented 2000 to 2014 study (Fig. 3), reflecting
increasing clearing of primary forests over time and decreasing pool
of year 2000 secondary forests available for reclearing. The annual
rates and general trend of primary forest loss in DRC agree with
Turubanova et al.(fig.S7).Ernstet al.(23) reported increasing rates
of deforestation between 1990 to 2000 and 2000 to 2005 in dense
forests of all countries except GAB (EQG not reported). While these
results are not directly comparable with the current estimates owing to
different definitions and study period, we observed similar national
trends of small-scale clearing for agriculture in primary forests and
woodlands between 2000 and 2014 (Table 2). Comparison of the
studies mentioned above and the FAO FRA 2015 (Global Forest
Resources Assessment 2015 of the Food and Agriculture Organization
of the United Nations) report (41) with the current results for DRC is
presented in table S3.
Limitations
Direct assessment of disturbance drivers from remote sensing data is
advantageous in providing relative objectivity and consistency across
national borders. The legend is easily adaptable to include region-
specific drivers when applying the method in a different geographic
domain or globally. However, there are limitations in how much the-
matic detail can be interpreted. For example, we were not able to dis-
tinguish small-scale clearing for agriculture from forest clearing for
charcoal production, which may be colocated with the establishment
of new agricultural fields. Charcoal production is estimated to account
for up to 10% of forest loss in DRC and CAM (42). While fuel wood
collection in rural areas is largely offset through forest regeneration,
demand for energy from urban areas can lead to forest degradation
and deforestation. Charcoal is the fuel of choice in urban settings as
it is easier and cheaper to transport and store and produces more
energy per unit mass compared to wood. The Congo Basin’slargest
city, Kinshasa, sources charcoal within a 200-km radius with negligible
Fig. 3. Three-year moving average of annual forest loss area for the major disturbance categories in all countries. Each major disturbance category contributes
>0.5 Mha to the total 2000 to 2014 forest loss area. Forest clearing for small-scale rotational agriculture includes clearing for charcoal production, the contribution of
which does not exceed 10% of the class area (42). Error bands represent ±SE. Annual area estimates along with SEs are presented in table S2B. Prim., primary; sec.,
secondary; woodl., woodlands.
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 6of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
contributions from wood energy plantations (5). Improved spatial dis-
aggregation of charcoal as a driver of forest loss is needed.
Other drivers such as mining were virtually absent in the sample
population, reflecting their relative rarity compared to smallholder
agriculture. Given that the stratification is guided by observed forest
disturbance, rare land change drivers will not be well represented in
the analysis as SEs relative to the rare class area will often be large. For
rare classes, the 95% confidence interval establishes a useful upper
bound because it identifies that the change driver comprises “at most”
this percentage of the total loss. For example, at a 95% confidence level,
the true proportion of forest loss due to mining is at most 0.1% of
the total loss area of the region (<0.015 Mha), natural forest distur-
bance is at most 0.3% (<0.044 Mha), road construction is at most
0.7% (<0.11 Mha), large-scale clearing for agriculture is at most 1.3%
(<0.21 Mha), and commercial and residential construction is at most
1.6% (<0.26 Mha). One approach to targeting rare drivers more direct-
ly is by mapping forest disturbances by driver, which is often required
by countries for land use planning. Spatially explicit mapping of forest
disturbance drivers could be then supplemented with sample-based
analysis, providing map accuracy information and unbiased area esti-
mates. We have prototyped such approach to map forest conversion to
cropland in Brazil, but in the context of small-scale forest dynamics of
the Congo Basin, direct mapping of disturbance drivers may be more
challenging. Regardless, forest loss map information used in the current
study to target sample allocation through stratification significantly
increased precision of the estimates compared to simple random
sampling. For example, a simple random sample of 10,000 pixels would
have yielded a 4.8% SE of the total 2000 to 2014 forest loss area estimate
in the region, compared to the stratified SE of just 3.2%, resulting in a
33% reduction of uncertainty. For estimates of individual loss drivers
and forest types, we observed reductions in uncertainty from stratifica-
tion to be as much as 72% relative to the SE of simple random sampling.
Estimating temporal trends of small-scale forest clearing for rota-
tional agriculture in secondary forests is limited by the way the young
secondary forests are defined in the current study. We consider a
sampled pixel forested if it had forest cover in the year 2000; therefore,
fallows that reached a 5-m height threshold after 2000 and were later
cleared were not considered forest loss in this study. We therefore end
up with a limited pool of young secondary forests in the areas of sub-
sistence agriculture, which are recleared on average every 18 years (43).
By the end of our 14-year study period, most of the year 2000 young
secondary forests would have been cleared, resulting in decreasing rates
of clearing for this forest type. Including forest gain into the assessment
would have helped address this issue and track changes in young
secondary forests that regrewduringthestudyperiod.
Future potential advances
Forest type definitions are always a matter of debate and a source of
thematic uncertainty in forest loss area estimation. Using canopy
structure characteristics (% cover, height, and biomass density)
and disturbance history to define forest types may help reduce such
Table 2. Annual area of small-scale forest clearing for agriculture in
primary and mature secondary dense HTFs and primary woodlands
and dry forests (thousand hectares ± SE) by 5-year epochs. Forest
clearing for small-scale rotational agriculture includes clearing for charcoal
production, the contribution of which does not exceed 10% of the class
area (42). EQG had only 20 sampled pixels identified as forest loss, and this
small sample size did not yield adequately precise estimated annual loss
rates by 5-year epochs.
2000–2005 2005–2010 2010–2014
DRC 321 ± 26 403 ± 27 462 ± 33
CAR 64 ± 17 88 ± 20 80 ± 12
CAM 28 ± 7 37 ± 7 69 ± 16
RoC 9±3 24±8 35±9
GAB 17 ± 5 7 ± 3 4 ± 2
Fig. 4. Expansion of small-scale agriculture into recently undisturbed forests and woodlands (lines) and population growth in the region by country (bar
chart). Solid lines connect the annual forest loss area estimates and dashed lines represent the linear trend based on ordinary least squares regression. Forest clearing
for small-scale rotational agriculture includes clearing for charcoal production, the contribution of which does not exceed 10% of the class area (42). Error bars on the
area estimates represent 1 SE.
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 7of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
thematic uncertainty (17) and enable global applications of the method.
Forest type as defined in the current study is based on an objective cri-
terion of canopy density, which is modeled from optical remote sensing
data. Landsat-modeled canopy height was used as one of the quality
checks for forest type definition (table S4). In the future, improved
global light detection and ranging (LIDAR) data, for example, from
the proposed NASA Global Ecosystem Dynamics Investigation mis-
sion, will advance direct mapping of canopy height and other structural
metrics into forest definitions. Forest age and absence of disturbance
(e.g., “primary and mature secondary forests”versus “young primary
forests”) are somewhat harder to define and measure. Primary and in-
tact forest maps, derived using direct and indirect mapping methods
(33,39,44), could be used to supplement sample-based analysis.
Systematic acquisitions of high-frequency, very high spatial resolu-
tion images, a capability currently being developed by commercial
companiessuchasPlanet(45), may be an option for improving ref-
erence data. However, generic data access to such data sources by the
scientific and natural resource management communities is not a given.
Even with dense time series of very high resolution data readily avail-
able, visual assessment of gradual processes such as forest regrowth or
degradation will be challenging. Gradual forest changes can be modeled
on the basis of the spectral response from the canopy, but validation
data for these models will have to come from multiyear ground obser-
vationsortimeseriesofairborneLIDARdatameasuringsmallchanges
in canopy structure. Therefore, we currently only focus on quantifying
gross forest loss from disturbance events resulting in canopy damage
and do not aim to produce net forest change estimates or to quantify
the extent of forest degradation. We do, however, include industrial se-
lective logging, which is usually considered a type of forest degradation,
into the assessment of forest disturbance drivers. Because of the defini-
tion used (120-m buffer around visible logging damage), our estimates
of area affected by selective logging are likely to be considerably more
conservative compared with the estimates derived by outlining the
polygons of forests encompassing logging damages (46).
Attempts have been made to establish global reference data sampling
frames to validate global tree cover maps (47) and estimate the area of
different land cover types (48). These sampling frames that use stratified
random or systematic sampling with stratabeingbiomesorecoregions
are useful for the assessment of stable land cover classes. In the case of
small dynamic land cover change classes, such as forest loss (49)orbare
ground gain (50), stratification should be targeting these dynamics di-
rectly to ensure higher sampling efficiencies and lower uncertainty of
area estimates. Online tools for reference data collection, such as Collect
Earth (51), enable leveraging regional knowledge of image interpreta-
tion experts from around the world and transferring technical capacity
to the institutions in developing countries responsible for national land
coverchangereporting.Samplesizerequiredforglobalassessmentswill
depend on desired precision of the estimates and on the scale of report-
ing (e.g., biome-level estimates versus national versus subnational).
Forest loss area estimates, or activity data, enable carbon emissions
reporting (52). In the past, we have demonstrated two different
approaches to combining sample-based area estimates similar to those
derived in the current study with information on predisturbance forest
biomass (emissions factors). In the first approach (“stratify and mul-
tiply”), sampling domains with minimized within-domain carbon
density variance are created (17,25), and a single mean carbon density
value is assigned to the sample-based forest loss area estimate for each
domain. In the second approach, existing continuous forest biomass
maps are used to derive emissions factors per predisturbance forest
type class, identified from the sample (24). Although the current study
does not have the objective of estimating carbon losses associated with
forest loss, such estimates could easily be derived using emissions
factors from existing continuous maps or other sources of regional
or country-specific emissions factors.
MATERIALS AND METHODS
Definitions
Forest loss is defined in the current study as complete or partial removal
of woody vegetation, which reached a 5-m height threshold by the year
2000 and >25% tree canopy cover, within a sampled 30 m by 30 m pixel.
This includes “stand-replacement disturbance or the complete removal
of tree cover canopy at the Landsat pixel scale,”as defined by Hansen et al.
(14), and partial tree cover losses associated with boundary pixels and
selective logging. Forest loss was recorded in three gradations: 75 to
100% (counted as 100% of pixel area lost), 25 to 75% (50% of pixel area
lost), and <25% (0% of pixel area lost). These coarse gradations were
distinguishable in Landsat, which was the primary source of the
reference data (i.e., the observations used to produce the area estimates)
(fig. S1). The partial loss category includes pixels located on the edges of
2000 to 2014 forest disturbance patches (example: DRC sample 2615 at
http://glad.umd.edu/CAFR) and pixels located on the boundaries of for-
est patches in the year 2000 that have undergone complete clearing of
tree cover between 2000 and 2014 (example: Cameroon sample 118 at
http://glad.umd.edu/CAFR).
Forest loss year is defined as the year of the maximum percent
canopy cover removal. For example, if, initially, the sampled pixel
was cleared only partially and was later fully cleared, only the last
year was recorded as the loss year. When multiple complete or par-
tial vegetation clearing events occurred within the study period
(2001 to 2014), only the first complete clearing event was recorded.
Predisturbance forest categories include the following (fig. S2):
(i) primary and mature secondary dense (>60% tree canopy cover)
HTFs, (ii) young secondary dense HTFs, (iii) primary woodlands (25 to
60% tree canopy cover) and dry forests (>60% tree canopy cover, pres-
ence of dry season), (iv) secondary woodlands and sparse (25 to 60%
tree canopy cover) young secondary HTFs, and (v) plantations. We
did not use a minimal patch size in defining forest; instead, we defined
forest at the Landsat pixel scale as woody vegetation exceeding 5 m in
height and 25% in canopy cover. Quantitative thresholds of the vi-
sually interpreted forest type classes were verified using existing
Landsat-based tree cover (14) and height (53)models.Heightmodels
were reported to overestimate the height of tree cover <5 m (mean
absolute error of about 1.6 m), which may lead to inclusion of some
vegetation under 5 m into woodland classes in the current study. Mature
secondary dense HTFs are defined as disturbed in the past, but not dis-
tinguishable from the never disturbed primary dense HTFs in year 2000
Landsat imagery. Field evidence suggests that tropical secondary forests
restore structure and species richness similar to those of primary forests
after about 40 years (54).YoungsecondarydenseHTFsaremainlyasso-
ciated with the shifting cultivation areas, and these forests have a distinct
spectral signature (fig. S2B) and lower canopy heights compared to
primary and mature secondary HTFs. Primary woodlands and dry
forests represent natural vegetation outside HTF zones and are charac-
terized by distinct seasonality. Secondary woodlands and sparse young
secondary HTFs represent the sparse woody vegetation regrowth both
in HTF and in woodland zones. The plantation category represents palm
and tree plantations, established by the year 2000.
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 8of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
Forest disturbance drivers include four broad categories: human
forest clearing, industrial selective logging, fires, and natural distur-
bances (Fig. 1 and fig. S3). Human forest clearing includes clearing
for agriculture, clearing for construction, and clearing for mining.
Clearing for agriculture includes small- and large-scale clearing
categories. Large-scale clearings for agriculture have area of annual
clearing exceeding 10 ha; these are industrial mechanized clearings
for plantations and pastures (fig. S3C). Small-scale clearings have a
median annual clearing size of 1.8 ha (about 5 by 5 Landsat pixels),
which was estimated on the basis of manually digitized annual clearing
patches from Landsat composites for a random sample of 100 pixels
identified as “small-scale clearing.”Small-scale clearing for agriculture
was further distinguished into clearing for rotational (fig. S3A) and
semipermanent (fig. S3B) agriculture. Small-scale clearing for rota-
tional agriculture was characterized by forest regrowth starting 3 to
4 years after the clearing; for the clearings at the end of the study pe-
riod (after 2011), this driver was assigned on the basis of the regrowth
dynamics of the neighboring clearings. This disturbance category in-
cludes clearing for charcoal production, which cannot be reliably dis-
tinguished from slash-and-burn agriculture in the absence of very
high resolution imagery temporally colocated with charcoal burning.
Small-scale clearing for semipermanent agriculture was distinguished
from rotational agriculture by the absence of forest regrowth in the
years following the clearing.
Forest clearing for construction includes road (fig. S3D), residential
(fig. S3E), and commercial (fig. S3F) construction. The road construc-
tion category does not include roads that are the part of selective logging
infrastructure. Semipermanent roads, included in this category, were
distinguished from logging roads by the absence of forest regrowth
and presence of settlements and agricultural activities along the roads.
Residential construction and commercial construction were treated as
one class, since the absence of very high resolution imagery for some
sampled pixels would not allow consistently distinguishing these con-
struction types.
Forest clearing for mining is defined as removal of woody vegetation
in the process of mineral resources extraction (fig. S3G). This category is
characterized by the spectral signature of bare ground without significant
regrowth following forest disturbance, similar to that of construction.
Industrial selective logging is defined as canopy damage resulting
from logging infrastructure (logging roads, skid trails, and landings),
distinguishable in Landsat resolution, and a 120-m buffer around this
canopy damage to capture partial canopy loss associated with tree
felling and transportation. A 120-m buffer was initially selected for
a study in the Brazilian Legal Amazon (24) and was preserved in the
current study for consistency of the estimates. The sampled pixel was
labeled as “selective logging”if any visible logging damage was ob-
served in a 120-m buffer around it (white circle, fig. S3H). This forest
disturbance driver does not imply complete canopy loss within a sam-
pled pixel. Fragmentation effects of industrial selective logging were not
considered in the current study.
Forest loss from fire (fig. S3I) is defined as burning that was not
followed by agricultural activities, unlike small-scale clearing for
rotational agriculture. This includes areas affected by fires escaped
from slash-and-burn agricultural practices and fires set for hunting
purposes: 78% of the sampled pixels identified as “fire”were adjacent
to the forest edge and human activities (roads, settlements, and active
fields). We detected complete canopy loss in 72% of the sampled pixels
identified as fire, whereas the rest (28%) were non–stand-replacement
fires. Natural forest disturbances (fig. S3J) include river meandering,
windfalls, and other forest disturbance events (droughts and insect out-
breaks) that cannot be directly linked to human activities.
Sampling design
The study area consisted of the Congo Basin Forest Partnership
countries: CAM, CAR, DRC, EQG, GAB, and RoC (fig. S4). To estimate
forest loss area within this study region, we used a stratified sampling
design, which typically yields better precision compared to simple ran-
dom and systematic sampling designs (49,55). Strata were selected to
target forest cover loss (fig. S4) with the three strata defined as follows:
(i) “Loss,”any pixel that was mapped as forest loss during 2001 to 2014
where forest loss was determined from the global forest loss map
[Hansen et al.(14)]; (ii) “Probable loss,”60-m (two Landsat pixels)
buffer around mapped loss; (iii) “No loss,”all other areas outside of
mapped loss and the probable loss buffer (including both forested
and nonforested areas). The sampling unit was one Landsat pixel
(circa 30 m by 30 m); the mean pixel area within the study region in
geographic coordinates (latitude/longitude) was 766.13 m
2
.Thevaria-
tion of mean pixel area among the countries did not exceed 0.2% and
was therefore ignored. The total number of sample pixels was 10,000,
with 20% of the sample randomly allocated to the “Loss”stratum, 30%
to the “Probable loss”stratum, and 50% to the “No loss”stratum. Pixels
were allocated to the three sampling strata regardless of country bound-
aries. Countries were treated as poststrata in the area calculations. The
resulting distribution of sample pixels among the countries (poststrata)
and three sampling design strata is shown in table S5.
Estimation of area from the sample was performed using indicator
functions (56), since this approach works when the sampling strata are
different from the map classes and can also be used for the nonbinary
(proportional) reference sample labels (in our case, 0, 50, and 100%
forest loss). Forest loss area for each reference forest loss type (by dis-
turbance driver, by predisturbance forest type, and by year), reported
in table S2, was estimated using the following equation
^
A¼Atot ∑
H
h¼1
Nh
N
yhð1Þ
where A
tot
is the total study region area, Nis the total number of pixels
in the study region, His the number of poststrata (18, see table S5), n
h
is the sample size (number of sampled pixels) in poststratum h,N
h
is
the total number of pixels in poststratum h,y
u
is 1 or 0.5 if pixel u(or
its half) is classified as “forest cover loss”in the reference sample in-
terpretation and y
u
is 0 otherwise, and
yh¼∑u∈hyu
nhisthesamplemean
of the y
u
values in poststratum h.
To produce forest loss area estimates by disturbance driver, predis-
turbance forest type, and year, the definition of y
u
is modified so that
1 or 0.5 is recorded only if the loss area represents the specified dis-
turbance driver, predisturbance forest type, or year targeted by the
estimate, and y
u
= 0 if there is no forest cover loss or the sampled pixel
udoes not satisfy the definition of the target subset. The SE of the
sample-based loss area estimate is
SEð
^
AÞ¼Atot ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
∑
H
h¼1
N2
h1nh
Nh
s2
yh
nh
N2
v
u
u
u
tð2Þ
where s2
yh ¼∑u∈hðyu
yhÞ2
nh1is the sample variance for poststratum h.
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 9of12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
When estimating the relative contribution of each loss driver, forest
type, or country to the total area of forest loss, expressed as percentage,
both numerator and denominator are estimated from the sample, re-
sulting in a ratio estimator. Other examples for which a ratio estimator
is required include the estimates of contribution of each country to the
total area of forest loss of each driver, and the estimates of the contri-
bution of each loss driver to the total area of forest loss of each forest
type. The combined ratio estimator for stratified random sampling (57)
was therefore used to estimate these percentages, reported in Results
^
R¼
∑
H
h¼1
Nh
yh
∑
H
h¼1
Nh
xh
ð3Þ
where y
u
=1or0.5ifpixelu(or its half) is classified as belonging to a
specific driver, forest type, or country in the reference sample interpre-
tation, and y
u
= 0 otherwise; x
u
= 1 or 0.5 if pixel u(or its half) is
classified as forest cover loss in the reference sample interpretation,
and x
u
=0otherwise;
yh¼∑u∈hyu
nhis the sample mean of the y
u
values
in poststratum h;and
xh¼∑u∈hxu
nhisthesamplemeanofthex
u
values in
poststratum h.
The SE of the combined ratio estimator is
SEð
^
RÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
^
X2∑
H
h¼1
N2
h1nh
Nh
s2
yh þ
^
R2s2
xh 2
^
Rsxyh
=nh
sð4Þ
where
^
X¼∑H
h¼1Nh
xhis the estimated total area of tree cover loss
expressed in pixels, s2
yh and s2
xh are the sample variances in poststratum
h,andsxyh ¼∑u∈hxuyunh
xh
yh
nh1isthesamplecovarianceinpoststratumh.
Sources and availability of reference data
The primary source of reference data for the visual interpretation of
sampled pixels was Landsat data in the form of cloud-free annual
composites and 16-day observations (http://glad.umd.edu/CAFR).
Methods of Landsat data processing, cloud filtering, and compositing
are described in Potapov et al.(58). Sixteen-day observations were cloud
screened for the graphs of spectral indices (normalized difference veg-
etation index and normalized difference water index) and shortwave
infrared band reflectance. Non–cloud-screened 16-day composites were
used to identify the exact date of forest loss: Loss events are sometimes
visible through haze and translucent clouds, which would have been
removed in the automated process of cloud screening. To provide
landscape context to the visual interpretation of sampled pixels, annual
composites include a subset of 20 by 20 Landsat pixels (circa 36 ha)
around the sampled pixel and 16-day composites include a subset of
40 by 40 Landsat pixels (circa 144 ha).
Regionally, only 7% of the 10,000 sampled pixels had, on average,
less than one cloud-free Landsat observation per year (fig. S5A). These
pixels were clustered in the cloudiest areas along the coast and over
the HTFs in the core of the Congo Basin, introducing a spatial bias of
data availability. Consistent with this regional pattern, the country
mean of average number of cloud-free observations per year for each
sampled pixel (fig. S5A) was 0.9 in EQG, 1.1 in GAB, 2.0 in RoC, 3.0
inCAM,4.1inDRC,and4.7inCAR.InallcountriesexceptEQG,the
majority of sampled pixels had, on average, one or more cloud-free
observations per year, despite the large within-country variability of
available data. From all sampled pixels, 56% did not have any years
with zero cloud-free Landsat observations (fig. S5B), 20% had only one
gapyear,9%hadtwogapyears,4%hadthreegapyears,2%hadfour
gap years, and 9% had five or more gap years with zero cloud-free ob-
servations. Among the countries, EQG and GAB had the largest per-
centage of sampled pixels with at least one missing year (99 and 98%,
respectively), followed by RoC (82%), CAM (53%), DRC (40%), and
CAR (10%). This means that the error of forest loss occurrence and
date identification due to the reference data availability was probably
the highest in EQG and GAB and the lowest in CAR.
Landsat data availability also varied from year to year owing to the
characteristics of the Landsat satellite program (fig. S5C). Year 1999 had
the lowest data availability because Landsat 7 was launched in April
1999, and its predecessor Landsat 5 did not have a global data acquisi-
tion strategy. Year 1999 data were used only as a pre-2000 benchmark to
help identify year 2000 forest cover; thus, the low availability of 1999
data did not affect our results directly. Lower data availability occurred
in 2003 (fig. S5C) because of the malfunction of the Landsat 7 Scan Line
Corrector. This likely resulted in some underestimation of the year 2003
forest loss. The number of available cloud-free observations increased in
2013 and 2014 after the launch of Landsat 8 (fig. S5C), which might
have affected our interpretation results as well, leading to better detec-
tion of forest loss closer to the end of the study period.
The secondary source of reference data used primarily to help iden-
tify the initial forest cover and forest disturbance driver was very high
resolution data from Google Earth. The link opening Google Earth for
each sampled pixel is available from the interpretation interface (http://
glad.umd.edu/CAFR). From all sampled pixels, 74% had at least one
very high resolution (<1 m) image on Google Earth, 7% had image from
SPOT satellite (2.5 m resolution), and 19% had only Landsat.
Sample labeling protocol and confidence of
reference interpretations
All 10,000 sample pixels were initially screened by two independent
experts, who assigned forest loss (0, 50, or 100%) to each sample pixel.
Pixels identified as 50 or 100% loss were attributed with loss year, pre-
disturbance forest type, and forest disturbance driver (see the “Defini-
tions”section). Experts also recorded their confidence (high/low)
separately for the presence/absence of forest loss, forest loss year, pre-
disturbance forest type, and forest disturbance driver. After the initial
screening, sample pixels with disagreement between the two experts and
pixels marked as “low confidence”for any interpretation category were
additionally rechecked with the help of a third expert. Major sources of
uncertainty during sample interpretation and the ways they were ad-
dressed by the interpreters are listedintableS4.Additionalcheckswere
performed using auxiliarydatasourcesforthefollowingpixelsregard-
less of their initial confidence level:
(1) Primary and mature secondary HTF pixels with Landsat-
modeled year 2000 tree cover <90% (14);
(2) Primary and mature secondary HTF pixels with Landsat-modeled
year 2000 tree cover (14) >90% and year 2000 tree cover height (53)<15m;
(3) Young secondary HTF pixels with Landsat-modeled year
2000 tree cover height (53) >20 m;
(4)PrimaryandmaturesecondaryHTFsandprimarywoodlands
and dry forest pixels in DRC outside primary forest mask (33);
(5) Young secondary HTFs and secondary woodland pixels in
DRC within primary forest mask (33);
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 10 of 12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
(6) Pixels with forest loss year 3 or more years different from the
global forest loss map (14).
Sample pixels were iteratively rechecked by interpreters using aux-
iliary data until consensus on the final pixel labels and confidence
levels was reached.
Because we used the best available information for our reference
sample classification (visual interpretation of Landsat time series and
available very high resolution data), it is not possible to formally assess
the accuracy of our reference classification by comparing it to the
“truth.”In a sense, our current sample classification is the closest prac-
tical approximation to this truth in the absence of historic annual
(2000 to 2014) ground surveys or time series of very high resolution
data. We therefore can only indirectly assess the possible errors of
reference sample classification by analyzing certainty flags for each
sampled pixel. A total of 497 sampled pixels (5% of the total sample
size of 10,000; 274 “no loss”and 223 “loss”) were classified as low-
confidence presence/absence of forest loss during sample interpretation
(fig. S6). Sample pixels with low-confidence presence/absence of forest
loss were spread throughout the region but were somewhat clustered in
the cloudy coastal regions, particularly in the DRC province Bas Congo.
Confidence level was recorded separately for each forest loss
category (loss year, predisturbance forest type, and loss driver),
for both high- and low-confidence sample pixels identified as loss.
Years with the highest percentage of forest loss area coming from
low-confidence sample pixels (potential commission error) were
2007 (34%), 2003 (25%), and 2002 (25%); years with the lowest
percentage were 2011 (9%), 2006 (10%), and 2010 (10%). Annual
estimates for 2008 to 2014 had, on average, a smaller proportion of
area coming from low-confidence sampled pixels compared with
2001 to 2007 (14% versus 20%), which may be related to a better avail-
ability of cloud-free Landsat data (fig. S5) and the very high resolution
imagery from Google Earth in the later years. For the forest distur-
bance drivers, percentage of area coming from low-confidence sampled
pixels was the highest in the two smallest classes: mining (55%) and
natural disturbances (48%), followed by semipermanent small-scale
clearing for agriculture (24%), logging (14%), large-scale clearing for
agriculture (14%), fires (8%), rotational small-scale clearing for agri-
culture (5%), road construction (4%), and commercial and residential
construction (2%). These differences could be related to the higher
ambiguity of definitions for some of the classes. For example, con-
struction classes are the least ambiguous, since they usually occur in
the vicinity with already built-up areas and have a distinct postdistur-
bance spectral signature (concrete and dirt). Mining also has a distinct
postdisturbance bare ground signature, but artisanal mining typical for
the region is likely to be confused with natural disturbances (e.g., river
meandering). Among the predisturbance forest categories, young
secondary HTFs, secondary woodlands, and plantations had the high-
est potential commission error (22, 15, and 22%, respectively), due to
theirlikelyconfusionwithnon-woodyvegetation(youngfallows,tall
crops, and shrub). Primary and mature secondary HTFs and primary
woodlands and dry forests had lower potential commission error rates
(7 and 6%).
SUPPLEMENTARY MATERIALS
Supplementary material for this article is available at http://advances.sciencemag.org/cgi/
content/full/4/11/eaat2993/DC1
Fig. S1. Conceptual diagram of forest loss cases distinguishable via visual interpretation of a
single 30-m Landsat pixel.
Fig. S2. Examples of predisturbance forest types.
Fig. S3. Examples of forest disturbance drivers.
Fig. S4. Study area and sampling strata.
Fig. S5. Availability of cloud-free 16-day Landsat observations for the sampled pixels.
Fig. S6. Sampled pixels with high and low confidence of presence/absence of forest loss.
Fig. S7. Comparison of annual forest loss estimates for DRC.
Table S1. Summary of selected socioeconomic indicators for the study countries.
Table S2A. Total 2001 to 2014 forest disturbance area by disturbance driver and
predisturbance forest type (million hectares ± SE).
Table S2B. Annual forest loss area by forest disturbance driver and predisturbance forest type
in all countries (million hectares ± SE).
Table S3. Comparison of forest loss estimates for DRC.
Table S4. Major sources of uncertainty during sample interpretation and measures to address
them.
Table S5. Distribution of sampled pixels (n
h
) among the country poststrata and three sampling
design strata (loss, probable loss, and no loss) and strata sizes (N
h
).
REFERENCES AND NOTES
1. A. Chhatre, A. Agrawal, Trade-offs and synergies between carbon storage and livelihood
benefits from forest commons. Proc. Natl. Acad. Sci. U.S.A. 106, 17667–17670 (2009).
2. Y. T. Tegegne, M. Lindner, K. Fobissie, M. Kanninen, Evolution of drivers of deforestation
and forest degradation in the Congo Basin forests: Exploring possible policy options to
address forest loss. Land Use Policy 51, 312–324 (2016).
3. T. K. Rudel, The national determinants of deforestation in sub-Saharan Africa.
Philos. Trans. R. Soc. Lond. B Biol. Sci. 368, 20120405 (2013).
4. Q. Zhang, D. Devers, A. Desch, C. O. Justice, J. Townshend, Mapping tropical deforestation
in Central Africa. Environ. Monit. Assess. 101,69–83 (2005).
5. C. Megevand, A. Mosnier, J. Hourticq, K. Sand ers, N. Doetinch em, C. Streck,
Deforestati on Trends in the Con go Basin: Reconc iling Economic Growt h and Forest
Protection (The World Bank, 2013).
6. F. Grigoli, A. Herman, A. Swiston, “A Crude Shock : Explaining the Impact of the 2014-16
Oil Price Decline Across Exporters,”IMF Working Paper 17/160, International Monetary
Fund, 2017.
7. P. Mayaux, J.-F. Pekel, B. Desclée, F. Donnay, A. Lupi, F. Achard, M. Clerici, C. Bodart,
A. Brink, R. Nasi, A. Belward, State and evolution of the African rainforests between
1990 and 2010. Philos. Trans. R. Soc. Lond. B Biol. Sci. 368, 20120300 (2013).
8. B. Fisher, African exception to drivers of deforestation. Nat. Geosci. 3, 375–376 (2010).
9. S. Wunder, Oil Wealth and the Fate of the Forest: A Comparative Study of Eight Countries,
Routledge Explorations in Environmental Economics (Routledge Press, 2005).
10. N. T. Laporte, J. A. Stabach, R. Grosch, T. S. Lin, S. J. Goetz, Expansion of industrial logging
in Central Africa. Science 316, 1451 (2007).
11. L. Feintrenie, Agro-industrial plantations in Central Africa, risks and opportunities.
Biodivers. Conserv. 23, 1577–1589 (2014).
12. TFA, “Tropical Forest Alliance 2020 Marrakesh Declaration for Sustainable Development
of the Oil Palm Sector in Africa, UNFCCC Twenty-Second Session of the Conference
of the Parties”(2012).
13. Government of Gabon, “Strategic Plan for an Emerging Gabon. Plan Strategique Gabon
Emergent: Vision 2025 et orientations stratégiques 2011-2016”(2012).
14. M. C. Hansen, P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina,
D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini,
C. O. Justice, J. R. G. Townshend, High-resolution global maps of 21st-century forest cover
change. Science 342, 850–853 (2013).
15. D. H. Kim, J. O. Sexton, P. Noojipady, C. Huang, A. Anand, S. Channan, M. Feng,
J. R. Townshend, Global, Landsat-based forest-cover change from 1990 to 2000.
Remote Sens. Environ. 155, 178–193 (2014).
16. P. V. Potapov, S. A. Turubanova, M. C. Hansen, B. Adusei, M. Broich, A. Altstatt, L. Mane,
C. O. Justice, Quantifying forest cover loss in Democratic Republic of the Congo,
2000–2010, with Landsat ETM + data. Remote Sens. Environ. 122,106–116 (2012).
17. A. Tyukavina, A. Baccini, M. C. Hansen, P. V. Potapov, S. V. Stehman, R. A. Houghton,
A. M. Krylov, S. Turubanova, S. J. Goetz, Aboveground carbon loss in natural and
managed tropical forests from 2000 to 2012. Environ. Res. Lett. 10,074002 (2015).
18. F. Achard, R. Beuchle, P. Mayaux, H. J. Stibig, C. Bodart, A. Brink, S. Carboni, B. Desclée,
F.Donnay,H.D.Eva,A.Lupi,R.Raši, R. Seliger, D. Simonetti, Determination of tropical
deforestation rates and related carbon losses from 1990 to 2010. Glob. Chang. Biol. 20,
2540–2554 (2014).
19. P. Olofsson, G. M. Foody, M. Herold, S. V. Stehman, C. E. Woodcock, M. A. Wuldere, Good
practices for estimating area and assessing accuracy of land change. Remote Sens.
Environ. 148,42–57 (2014).
20. GFO I, Integration of Remote-Sensing and Ground-Based Observations for Estimation of
Emissions and Removals of Greenhouse Gases in Forests: Methods and Guidance from
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 11 of 12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
the Global Forest Observations Initiative. Edition 2.0 (Food and Agriculture
Organization, 2016).
21. N. Hosonuma, M. Herold, V. De Sy, R. S. De Fries, M. Brockhaus, L. Verchot, A. Angelsen,
E. Romijn, An assessment of deforestation and forest degradation drivers in
developing countries. Environ. Res. Lett. 7, 044009 (2012).
22. R. A. Houghton, Carbon emissions and the drivers of deforestation and forest
degradation in the tropics. Curr. Opin. Environ. Sustain. 4, 597–603 (2012).
23. C. Ernst, P. Mayaux, A. Verhegghen, C. Bodart, M. Christophe, P. Defourny, National forest
cover change in Congo Basin: Deforestation, reforestation, degradation and regeneration
for the years 1990, 2000 and 2005. Glob. Chang. Biol. 19,1173–1187 (2013).
24. A. Tyukavina, M. C. Hansen, P. V. Potapov, S. V. Stehman, K. Smith-Rodriguez, C. Okpa,
R. Aguilar, Types and rates of forest disturbance in Brazilian Legal Amazon, 2000–2013.
Sci. Adv. 3, e1601047 (2017).
25. A. Tyukavina, S. V. Stehman, P. V. Potapov, S. A. Turubanova, A. Baccini, S. J. Goetz,
N. T. Laporte, R. A. Houghton, M. C. Hansen, National-scale estimation of gross forest
aboveground carbon loss: A case study of the Democratic Republic of the Congo.
Environ. Res. Lett. 8, 044039 (2013).
26. R. Tropek, O. Sedláček, J. Beck, P. Keil, Z. Musilová, I. Šímová, D. Storch, Comment on
“High-resolution global maps of 21st-century forest cover change.”Science 344, 981 (2014).
27. P. C. J. Moonen, B. Verbist, J. Schaepherders, M. B. Meyi, A. Van Rompaey, B. Muys,
Actor-based identification of deforestation drivers paves the road to effective REDD+ in
DR Congo. Land Use Policy 58, 123–132 (2016).
28. F. Zvomuya, On a whim: Equatorial Guinea building new capital city in the middle of the
rainforest. Mongabay Environ. News (2014).
29. G. Molinario, M. C. Hansen, P. V. Potapov, Forest cover dynamics of shifting cultivation in
the Democratic Republic of Congo: A remote sensing- based assessment for 2000 –2010.
Environ. Res. Lett. 10 (2015).
30. J. A. Sayer, D. Endamana, M. Ruiz-Perez, A. K. Boedhihartono, Z. Nzooh, A. Eyebe,
A. Awono, L. Usongo, Global financial crisis impacts forest conservation in Cameroon.
Int. For. Rev. 14,90–98 (2012).
31. D. Nepstad, D. McGrath, C. Stickler, A. Alencar, A. Azevedo, B. Swette, T. Bezerra,
M. DiGiano, J. Shimada, R. Seroa da Motta, E. Armijo, L. Castello, P. Brando, M. C. Hansen,
M. McGrath-Horn, O. Carvalho, L. Hess, Slowing Amazon deforestation through public
policy and interventions in beef and soy supply chains. Science 344, 1118–1123 (2014).
32. S. A. Abood, J. S. H. Lee, Z. Burivalova, J. Garcia-Ulloa, L. P. Koh, Relative contributions
of the logging, fiber, oil palm, and mining industries to forest loss in Indonesia.
Conserv. Lett. 8,58–67 (2015).
33. S. Turubanova, P. V. Potapov, A. Tyukavina, M. C. Hansen, Ongoing primary forest loss in
Brazil, Democratic Republic of the Congo, and Indonesia. Environ. Res. Lett. 13,
074028 (2018).
34. P. S. Dasgupta, Population, poverty and the local environment. Sci. Am. 272,40–45 (1995).
35. E. Boserup, Population and Technological Change: A Study of Long-Term Trends
(The University of Chicago Press, 1965).
36. D. L. Carr, Proximate population factors and deforestation in tropical agricultural frontiers.
Popul. Environ. 25,585–612 (2004).
37. K. Vlassenroot, T. Raeymaekers, Conflict and Social Transformation in Eastern DR Congo
(Academia Press, 2004).
38. FAO, Living in and from the Forests of Central Africa (Rome, 2017).
39. P. Potapov, M. C. Hansen, L. Laestadius, S. Turubanova, A. Yaroshenko, C. Thies, W. Smith,
I. Zhuravleva, A. Komarova, S. Minnemeyer, E. Esipova, The last frontiers of wilderness:
Tracking loss of intact forest landscapes from 2000 to 2013. Sci. Adv. 3, e1600821 (2017).
40. V. Foster, W. Butterfield, C. Chen, N. Pushak, “Building Bridges: China’s Growing Role as
Infrastructure Financier for Sub-Saharan Africa. Trends and Policy Options; no. 5”
(The World Bank, 2009).
41. FAO, Global Forest Resources Assessment 2015 (2015); http://www.fao.org/
forest-resources-assessment/current-assessment/en/.
42. E. N. Chidumayo, D. J. Gumbo, The environmental impacts of charcoal production in
tropical ecosystems of the world: A synthesis. Energy Sustain. Dev. 17,86–94 (2013).
43. G. Molinario, M. C. Hansen, P. V. Potapov, A. Tyukavina, S. Stehman, B. Barker, M. Humber,
Quantification of land cover and land use within the rural complex of the Democratic
Republic of Congo. Environ. Res. Lett. 12, 104001 (2017).
44. A. Tyukavina, M. C. Hansen, P. V. Potapov, A. M. Krylov, S. J. Goetz, Pan-tropical hinterland
forests: Mapping minimally disturbed forests. Glob. Ecol. Biogeogr. 25, 151–163 (2016).
45. D. Butler, Many eyes on Earth: Swarms of small satellites set to deliver close to real-time
imagery of swathes of the planet. Nature 505, 143–144 (2014).
46. M. Hirschmugl, M. Steinegger, H. Gallaun, M. Schardt, Mapping forest degradation due to
selective logging by means of time series analysis: Case studies in Central Africa.
Remote Sens. 6, 756–775 (2014).
47. B. Pengra, J. Long, D. Dahal, S. V. Stehman, T. R. Loveland, A global reference database
from very high resolution commercial satellite data and methodology for application
to Landsat derived 30m continuous field tree cover data. Remote Sens. Environ. 165,
234–248 (2015).
48. J.-F. Bastin, N. Berrahmouni, A. Grainger, D. Maniatis, D. Mollicone, R. Moore, C. Patriarca,
N.Picard,B.Sparrow,E.M.Abraham,K.Aloui,A.Atesoglu,F.Attore,Ç.Bassüllü,A.Bey,
M. Garzuglia, L. G. García-Montero, N. Groot, G. Guerin, L. Laestadius, A. J. Lowe,
B. Mamane, G. Marchi, P. Patterson, M. Rezende, S. Ricci, I. Salcedo, A. Sanchez-Paus Diaz,
F. Stolle, V. Surappaeva, R. Castro, The extent of forest in dryland biomes. Science 358,
635–638 (2017).
49. M. Broich, S. V. Stehman, M. C. Hansen, P. Potapov, Y. E. Shimabukuro, A comparison of
sampling designs for estimating deforestation from Landsat imagery: A case study of
the Brazilian Legal Amazon. Remote Sens. Environ. 113, 2448–2454 (2009).
50. Q. Ying, M. C. Hansen, P. V. Potapov, A. Tyukavina, L. Wang, S. V. Stehman, R. Moore,
M. Hancher, Global bare ground gain from 2000 to 2012 using Landsat imagery.
Remote Sens. Environ. 194, 161–176 (2017).
51. A. Bey, A. Sánchez-Paus Díaz, D. Maniatis, G. Marchi, D. Mollicone, S. Ricci, J.-F. Bastin,
R. Moore, S. Federici, M. Rezende, C. Patriarca, R. Turia, G. Gamoga, H. Abe, E. Kaidong,
G. Miceli, Collect earth: Land use and land cover assessment through augmented visual
interpretation. Remote Sens. 8, 807 (2016).
52. IPCC, 2006 IPCC Guidelines for National Greenhouse Gas Inventories. Volume 1:
General Guidance and Reporting (IGES, 2006).
53. M. C. Hansen, P. V. Potapov, S. J. Goetz, S. Turubanova, A. Tyukavina, A. Krylov,
A. Kommareddy, A. Egorov, Mapping tree height distributions in Sub-Saharan Africa
using Landsat 7 and 8 data. Remote Sens. Environ. 185, 221–232 (2016).
54. T. Mitche, J. K. Zimmerman, J. B. Pascarella, L. Rivera, H. Marcano-Vega, Forest
regeneration in a chronosequence of tropical abandoned pastures: Implications for
restoration ecology. Restor. Ecol. 8, 328–338 (2000).
55. S. V. Stehman, Sampling designs for accuracy assessment of land cover. Int. J.
Remote Sens. 30, 5243–5272 (2009).
56. S. V. Stehman, Estimating area and map accuracy for stratified random sampling when
the strata are different from the map classes. Int. J. Remote Sens. 35,4923–4939 (2014).
57. W. G. Cochran, Sampling Techniques (John Wiley & Sons Inc., ed. 3, 1977).
58. P. V. Potapov, S. A. Turubanova, A. Tyukavina, A. M. Krylov, J. L. McCarty, V. C. Radeloff,
M. C. Hansen, Eastern Europe’s forest cover dynamics from 1985 to 2012 quantified
from the full Landsat archive. Remote Sens. Environ. 159,28–43 (2015).
Acknowledgments: Our work was facilitated by national-scale implementations of our
method in partnership with L. Mane and A. Mazinga of the Central African Satellite Forest
Observatory; L. Diackabana and C.-B. O. Diamansuka of the RoC’s National Center for Surveys
and Forest and Fauna Resource Management; I. Suspense and H. Makaya of the University
of Marien Ngouabi; and R. Siwe, T. Nana, and B. Socrates of Cameroon’s REDD+ Technical
Secretariat. Funding: Support for this study was provided by the United States Agency for
International Development through its Central Africa Regional Program for the Environment
and by the National Aeronautics and Space Administration. Author contributions: A.T.,
M.C.H., P.P., and S.V.S. designed the study. P.P. processed Landsat satellite data. I.K. and
P.P. designed and assembled sample interpretation web interface. A.T., D.P., C.O., S.T.,
and P.P. performed visual sample interpretation. A.T. performed statistical analysis. A.T. and
M.C.H. co-wrote the majority of the manuscript with all authors contributing to the final
version. Competing interests: The authors declare that they have no competing interests.
Data and materials availability: All data needed to evaluate the conclusions in the paper
are present in the paper and/or the Supplementary Materials. Final sample labels and
reference data for each sampled pixel are available from http://glad.umd.edu/CAFR.
Submitted 12 February 2018
Accepted 8 October 2018
Published 7 November 2018
10.1126/sciadv.aat2993
Citation: A. Tyukavina, M. C. Hansen, P. Potapov, D. Parker, C. Okpa, S. V. Stehman,
I. Kommareddy, S. Turubanova, Congo Basin forest loss dominated by increasing
smallholder clearing. Sci. Adv. 4, eaat2993 (2018).
SCIENCE ADVANCES |RESEARCH ARTICLE
Tyukavina et al., Sci. Adv. 2018; 4: eaat2993 7 November 2018 12 of 12
on November 7, 2018http://advances.sciencemag.org/Downloaded from
Congo Basin forest loss dominated by increasing smallholder clearing
Kommareddy and Svetlana Turubanova
Alexandra Tyukavina, Matthew C. Hansen, Peter Potapov, Diana Parker, Chima Okpa, Stephen V. Stehman, Indrani
DOI: 10.1126/sciadv.aat2993
(11), eaat2993.4Sci Adv
ARTICLE TOOLS http://advances.sciencemag.org/content/4/11/eaat2993
MATERIALS
SUPPLEMENTARY http://advances.sciencemag.org/content/suppl/2018/11/05/4.11.eaat2993.DC1
REFERENCES http://advances.sciencemag.org/content/4/11/eaat2993#BIBL
This article cites 43 articles, 9 of which you can access for free
PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions
Terms of ServiceUse of this article is subject to the
registered trademark of AAAS. is aScience Advances Association for the Advancement of Science. No claim to original U.S. Government Works. The title
York Avenue NW, Washington, DC 20005. 2017 © The Authors, some rights reserved; exclusive licensee American
(ISSN 2375-2548) is published by the American Association for the Advancement of Science, 1200 NewScience Advances
on November 7, 2018http://advances.sciencemag.org/Downloaded from
