Content uploaded by Anupam Anand
Author content
All content in this area was uploaded by Anupam Anand on Jul 09, 2016
Content may be subject to copyright.
Earth science data records of global forest cover and change: Assessment
of accuracy in 1990, 2000, and 2005 epochs
Min Feng
a,
⁎, Joseph O. Sexton
a
, Chengquan Huang
a
, Anupam Anand
a,b
, Saurabh Channan
a
, Xiao-Peng Song
a
,
Dan-Xia Song
a
,Do-HyungKim
a
, Praveen Noojipady
a,c
,JohnR.Townshend
a
a
Global Land Cover Facility, Department of Geographical Sciences, University of Maryland, College Park, MD 20742, USA
b
Global Environment Facility, Washington, DC 20433, USA
c
Biospheric Sciences Laboratory, NASA/Goddard Space Flight Center, Greenbelt, MD 20771, USA
abstractarticle info
Article history:
Received 25 June 2015
Received in revised form 31 May 2016
Accepted 5 June 2016
Available online 16 June 2016
The Global Land Cover Facility (GLCF) global forest-cover and -change dataset is a multi-temporal depiction of
long-term (multi-decadal), global forest dynamics at high (30-m)resolution. Basedon per-pixel estimatesof per-
centage tree cover and their associated uncertainty, the dataset currently represents binary forest cover in nom-
inal 1990, 2000, and 2005epochs, as well as gains and losses over time. A comprehensive accuracy assessment of
the GLCF datasetwas performed using a global,design-based sampleof 27,988 independent, visually interpreted
referencepoints collectedthrough a two-stage,stratified samplingdesign wherein experts visuallyidentified for-
est cover and change in each of the 3 epochs based on Landsat and high-resolution satellite images, vegetation
index profiles, and field photos. Consistent across epochs, the overall accuracy of the static forest-cover layers
was 91%, and the overall accuracy of forest-cover change was N88% —among the highest accuracies reported
for recent global forest- and land-cover data products. Both commission error (CE) and omission error (OE)
were low for static forest cover in each epoch and for the stable classes between epochs (CE b3%, OE b22%),
but errors were larger for forest loss (45% ≤CE b62%, 47% bOE b55%) and gain (66% ≤CE b85%,
61% bOE b84%). Accuracy was lower in sparse forests and savannahs, i.e., where tree cover was at or near the
30% threshold used to discriminate forest from non-forest cover. Discrimination of forest had a low rate of com-
mission error and slight negative bias, especially in areas with low tree cover. After adjusting global area esti-
mates to reference data, 39.28 ± 1.34 million km
2
and 38.81 ± 1.34 million km
2
of forest were respectively
identified in2000 and 2005 globally, and 33.16 ± 1.36 million km
2
of forest were estimatedin the available cov-
erage of Landsat data circa-1990. Forest loss and gain were estimated to have been 0.73 ± 0.38 and 0.28 ± 0.26
million km
2
between 2000 and 2005, and 1.08 ± 0.53 and 0.53 ± 0.47 million km
2
between 1990 and 2000.
These estimates of accuracy are required for rigorous use of the data in the Earth sciences (e.g., ecology, econom-
ics, hydrology, climatology) aswell as for fusion withother records of global change.The GLCF forest -cover and -
change dataset is available for free public download at the GLCF website (http://www.landcover.org).
Published by Elsevier Inc. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
Keywords:
Accuracy assessment
Forest
Landsat
Global
Sampling
1. Introduction
Changes in Earth's forests impact hydrological, biogeochemical, and
energy fluxes, as well as ecosystems' capacity to support biodiversity
and human economies (Bonan, 2002; Nabuurs et al., 2007;
Schlesinger, 1997; Shvidenko et al., 2005; Townshend et al., 2012).
Long-term records of forest cover and change are needed across a
broad range of investigation, including climate and carbon-cycle model-
ing, hydrological studies, habitat analyses, biological conservation, and
land-use planning (Band, 1993; BenDor, Westervelt, Song, & Sexton,
2013; Conde et al., 2010; Haddad et al., 2015; Houghton, 1998; Lal,
1995; Smart, Swenson, Christensen, & Sexton, 2012; Song, Huang,
Saatchi, Hansen, & Townshend, 2015; Trainor, Walters, Morris, Sexton,
& Moody, 2013). Routine global monitoring of forest change has been
identified as a high priority in a number of national and international
programs, including the United Nations Framework Convention on Cli-
mate Change (UNFCCC) (UNFCCC, 2002), Food and Agriculture Organi-
zation of the United Nations (FAO) (FAO, 2010), Global Observation for
Forest and Land Cover Dynamics (GOFC-GOLD) (Townshend & Justice,
1988), Global Climate Observing System (Mason & Reading, 2004),
and the United States Global Change Research Program (Michalak,
Jackson, Marland, & Sabine, 2011).
Because a substantial proportion of forest cover and its changes
occur in small patches (Townshend & Justice, 1988), a requirement of
forest monitoring is repeated observation at resolutions b100 m—i.e.,
Remote Sensing of Environment 184 (2016) 73–85
⁎Corresponding author.
E-mail address: fengm@umd.edu (M. Feng).
http://dx.doi.org/10.1016/j.rse.2016.06.012
0034-4257/Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
by Landsat or “Landsat-class”satellites (Masek et al., 2006; Skole, Salas,
& Taylor, 1998; Townshend et al., 2004). One of the earliest efforts to
map forest change using Landsat data over large areas was NASA's
Landsat Pathfinder Humid Tropical Deforestation Project (Townshend
&Justice,1995), which provided the first assessments of deforestation
in the Amazon and several tropical countries (Skole & Tucker, 1993;
Steininger et al., 2001). Since then, Landsat-based forest-change assess-
ments have been conducted over a number of countries or regions, in-
cluding North America (Masek et al., 2008) , Paraguay (Huang et al.,
2007), the United States (Hansen et al., 2014), European Russia
(Potapov, Turubanova, & Hansen, 2011), the legal Amazon (http://
www.obt.inpe.br/prodes), the Democratic Republic of Congo (Potapov
et al., 2012), Bolivian Amazon (Steininger et al., 2001), and the humid
tropics (Kim, Sexton, & Townshend, 2015). Hansen and Loveland
(2012) provided a recent review of efforts to map land cover change
using Landsat data at regional to national scales.
Over the last few years, two parallel efforts have been devoted to
mapping global forest change using Landsat data. One was to harvest
Landsat-7 Enhanced Thematic Mapper Plus (ETM+) images to estimate
forest changes from 2000to 2012 (Hansen et al., 2013).The other, spon-
sored by NASA and led by the Global Land Cover Facility (GLCF), used
optimally selected Landsat ETM+, Thematic Mapper (TM), and Multi-
Spectral Scanner (MSS) images to produce a consistent, long-term re-
cord of global forest cover and change spanning the Landsat archive
from the 1970’s to the near-present (Feng, Huang, Channan, et al.,
2012a, Feng et al., 2013, Kim et al., 2014, Sexton, Song, et al., 2013a,
Sexton et al., 2015, Townshend et al., 2012). Development of the GLCF
1990, 2000, and 2005 forest-cover change products was recently com-
pleted, and preliminary accuracy estimates for these products have
been reported by Kim et al. (2014).Schepaschenko et al. (2015) validat-
ed these andother current globalforest-cover datasets against indepen-
dent, crowd-sourced reference data; and Sexton et al. (2016) map and
explain major differences between eight global forest-cover datasets.
Here we provide a comprehensive assessment of the GLCF Forest
Cover and Change data products for 1990, 2000, and 2005 epochs. We
include a brief overview of their development, a detailed description
of the assessment methods, and accuracy estimates at global and
biome levels. Insights and challenges in developing and validating glob-
al forest-cover change data products are also discussed.
2. The GLCF 30-m forest products
The GLCF global forest-cover and -change dataset is a multi-tempo-
ral depiction of long-term (multi-decadal), global forest dynamics at
high (i.e., 30-m) resolution, or pixel-size. Based on per-pixel estimates
of tree cover and their associated uncertainty (Sexton et al., 2013a,
2015, the dataset currently represents binary forest cover in nominal
1990, 2000, and 2005 years, or “epochs”, as well as gains and losses be-
tween epochs. Forest is defined as a minimum area of land of 0.27 ha
with ≥30% tree cover—i.e., as land cover, as opposed to land use
(Sexton et al., 2016; Townshend et al., 2012). The GLCF dataset com-
prises geographic layers representing forest cover in each nominal
year and change between years, as well as the uncertainty associated
with each. The forest-cover and change layers are based on 30-m reso-
lution estimates of surface reflectance (Feng et al., 2013) and are com-
patible with estimates of tree cover (Sexton et al., 2013a) and surface
water (Feng, Sexton, Channan, & Townshend, 2015). A moving box filter
of 3 x 3 pixels was applied to remove all (center) pixels which were not
neighbored by at least 2 (out of 8) pixels of the same type. This filter en-
sured that all patches (or holes) smaller than 3 pixels were removed
and thus ensured a minimum mapping unit of three pixels or 0.27 ha.
Per-pixel estimation of tree cover and its uncertainty is described by
Sexton et al. (2013a). Estimation of forest cover and change, as well as
the propagation of uncertainty from percent-tree to categorical forest
cover and change, are described by Sexton et al. (2015). Estimation
and validation of surface reflectance, used as covariates, are described
by Feng et al. (2013). A preliminary design-based validation of the
1990 forest-cover and 1990–2000 forest-cover change layers, as well
as model-based propagation of uncertainty from 2000 to 1990, are de-
scribed by (Kim et al., 2014). All data products are available for free pub-
lic download at http://www.landcover.org. Here we assess the accuracy
of forest-cover and –change layers in 1990–2000-2005 by design-based
accuracy assessment against independent reference data.
3. Assessment methods
3.1. Sampling design
Accuracy assessment employed a two-stage, stratified sampling de-
sign (Cochran, 1977; Sannier, Mcroberts, Fichet, Massard, & Makaga,
2014; Särndal, Swensson, & Wretman, 1992; Stehman, 1999; Stehman
& Czaplewski, 1998). To increase the representation of rare classes, ref-
erence data were sampled across the global land area in two stages, first
selecting Landsat WRS-2 tiles within predefined global strata and then
sampling pixels within each selected tile. The spatial location of sample
points was held constant for all time periods.
3.1.1. Biome definition
Biome-level stratification was based on the 16 major habitat types
delineated by the Nature Conservancy (TNC) Terrestrial Ecoregions of
the World dataset (TNC, 2012). Excluding inland water, deserts and
xeric shrublands, and rock and ice, we merged the major habitat types
into eight forest and non-forest biomes (Table 1). Among the 7277
WRS-2 tiles in the 8 biomes, the 5294 tiles completely contained within
any biome were assigned to their respective biomes, and tiles spanning
biome boundaries (including land/ocean boundaries) were excluded.
This reduced the land area for each of the 8 biomes available for sam-
plingby18.7–58.2% of each biome (Table 1).
3.1.2. Tile selection
Sampling within biomes focused on WRS-2 tiles exhibiting high
rates of vegetation change, which was detected using the Training
Data Automation and Support Vector Machines (TDA-SVM) change-de-
tection algorithm (Huang et al., 2008). The median vegetation-change
rate for each biome was then used as the threshold for discriminating
high- and low-change strata for that biome. Within each biome, eight
tiles were then randomly selected in the high-change stratum and
four tiles were randomly selected in the low-change stratum (Fig. 1).
The inclusion probability, p(T|G), of each WRS-2 tile, T,ineach
biome, G, was calculated as:
pTGjðÞ¼
nT
NT
;ð1Þ
Where n
T
is the desired number of sampled tiles within the popula-
tion of the stratum (N
T
); n
T
was set to 4 and 8 for low- and high-change
strata, respectively. A random number p
1
⁎was assigned to each tile, and
tiles with p
1
⁎bp(T|G) were selected asthe sample tiles. Globally, 89 tiles
were selected out of the intended 96 because only one tile met the cri-
terion for the “high-change”stratum in the boreal non-forest biome.
3.1.3. Point selection
Following biome-level sampling, each selected tile was divided into
8 strata representing forest/non-forest status in each of the two periods,
1990–2000 and 2000–2005. This preliminary forest/non-forest discrim-
ination was again performed by TDA-SVM. All pixels identified as cloud,
shadow, water, or no-data, as well as pixels located at the edge of two
classes, were excluded from the population. This exclusion reduced
the available land area for each of the 8 biomes by 3.8–13.2% (Table 1).
74 M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
The inclusion probability for each stratum was calculated as:
piSjðÞ¼
nS
NS
;ð2Þ
where the probability p(i|S) is the ratio of the desired number of
pixels (n
s
) to the total number of pixels in the stratum (N
S
). As recom-
mended by Congalton (1991) and Olofsson et al. (2014),n
s
was set to
50 for each stratum (S). A random number p
2
⁎was assigned to each
pixel, and pixels with p
2
⁎bp(i|S) were selected as the sample points. A
total of 27,988 points were thus collected across the globe. Fig. 2
shows the selected points in WRS-2 tile p224r078, located at the bound-
ary of Paraguay, Argentina, and Brazil.
3.2. Response design
Forest or non-forest cover in each pixel and each epoch was visually
identified by experienced image analysts using a web-based tool
presenting the GLS Landsat image(s)covering each location and auxilia-
ry information including; Normalized Difference Vegetation Index
(NDVI) phenology from MODIS, high-resolution satellite imagery and
maps from Google Maps, and geotagged ground photos (Fig. 3)(Feng
et al., 2012b). The Landsat images were presented in multiple 3-band
combinations—e.g., near infrared (NIR)-red (R)-green (G), R-G-blue
(B), and shortwave infrared (SWIR)-NIR-R. The extent of each selected
30-m Landsat pixel was extracted in the UniversalTransverse Mercator
(UTM) coordinate system and delineated in both the Landsat image and
in Google Maps to facilitate visual comparison. The NDVI profile was d e-
rived from the 8-day composited surface reflectance data (MOD09A1;
Vermote & Kotchenova, 2008; Vermote, Saleous, & Justice, 2002)with
nearest-neighbor interpolation, excluding data labeled ascloud or shad-
ow in the MOD09A1 Quality Assurance (QA) layer (Feng et al., 2012b).
The selected points were randomly distributed among 12 experts for
interpretation (Table 2). Experts visually checked the information provid-
ed by the tool and labeled each point either “forest”or “non-forest”for
each of the 3 epochs individually. Points with Landsat pixels
Table 1
Reclassification of TNCmajor habitat types(TNC, 2012) into biomestrata. The land area foreach biome is reportedin “Land area(km
2
)”column,and the percentageof that area reduced by
excluding tiles spanning boundaries is reported in “Spanning biome WRS-2tiles (%)”column. The percentage of the remained area after the “spanning biome”exclusion that further re-
duced by excluding edge pixels is reported in the “Edge pixels (%)”column.
Biome strata TNC biomes Land area (km
2
)
Percentage of area reduced
Spanning biome WRS-2 tiles (%) Edge pixels (%)
Tropical Evergreen Forests Tropical and Subtropical Moist Broadleaf Forests
Mangroves
Tropical and Subtropical Coniferous Forests
16,608,638 25.2 9.7
Tropical Deciduous Forests Tropical and Subtropical Dry Broadleaf Forests 6,780,454 18.7 8.4
Tropical Non-forest Tropical and Subtropical Grasslands, Savannas and Shrublands
Flooded Grasslands and Savannas (23°S - 23°N)
Montane Grasslands and Shrublands (23°S - 23°N)
15,296,731 28.0 5.5
Temperate Evergreen Forests Temperate Conifer Forests 3,843,538 50.9 13.2
Temperate Deciduous Forests Temperate Broadleaf and Mixed Forests
Mediterranean Forests, Woodlands, and Scrub
14,013,894 29.1 9.4
Temperate Non-forest Temperate Grasslands, Savannas and Shrublands
Flooded Grasslands and Savannas (23°S - 23°N)
Montane Grasslands and Shrublands (23°S - 23°N)
2,918,100 58.2 2.0
Boreal Forests Boreal Forests/Taiga 20,381,706 24.9 12.3
Boreal Non-forest Tundra 21,484,150 21.1 3.8
[Excluded] Deserts and Xeric Shrublands
Inland Water
Fig. 1. Biome strata and the 89 WRS-2 tiles selected within the sample.
75M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
contaminated with cloud or shadow were labeled as “cloud”and “shad-
ow”respectively. If an expert was unable to identify the cover of a pixel,
he or she was instructed to label it as “unknown”for further investigation.
Over 1000 points were collected in each decile of tree cover, with
nearly uniform sample size across the range of tree cover N10% cover
(Fig. 4). Of these points, N90% were labeled as forest or non-forest by vi-
sual interpretation in the 1990, 2000, and 2005 epochs, with only 6% of
the points remaining as “unknown”. Less than 1% of the points across all
epochs were interpreted as “cloud”or “shadow”. The distribution of the
unknown points in the 2000 epoch revealed that these difficult points
were rare (b4%) in areas of low or high tree-canopy cover but were
much more frequent in areas with 5–35% tree cover (Fig. 5).
3.3. Validation metrics
Based on the independent reference sample, the labeled points were
used to quantify the accuracy of the global forest-cover and -change
layers using validation metrics weighted by area (Card, 1982;
Congalton, 1991; Stehman, 2014; Stehman & Czaplewski, 1998). For
each reference datum, i, the agreement between estimated and refer-
ence cover or change, y, was defined:
yi¼1if
0if
^
ci¼
^
ci≠
ci
ci:
Weights were applied to the data to remove the effect of dispropor-
tional sampling, by standardizing the inclusion probability of each
observation proportional to the area of each stratum (Sexton, Urban,
Donohue, & Song, 2013b). Each point's weight, w
i
, was calculated as
the inverse of the joint standardized probability of its selection at the
tile- and pixel-sampling stages:
wi¼PiS
j
ðÞ
piS
j
ðÞ
PTG
j
ðÞ
pTG
j
ðÞ
¼nS
NS
ni
Ni
nG
NG
nT
NT
cos φi
ðÞ;ð4Þ
where P(i|S) is the inclusion probability of the desired number of
pixels (n
s
) to be randomly selected from the number of pixels in the
Landsat scene (N
S
), and P(T|G) is the probability of the desired
number of Landsat tiles (n
g
) selected from the total number of
Landsat scenes (N
g
) located inside the corresponding biome.
Adjusting the weight by the cosine of the pixel's latitude (φ)cor-
rects the sampling bias due to the increasing density of WRS-2
tiles with latitude.
Overall accuracy (OA) was calculated as the weighted number of
points showing agreement between the estimated and the reference
(i.e., human-interpreted) class—i.e., elements of the diagonal of the
confusion matrix—divided by the weighed total number of points
(n
a
):
OA ¼Xna
i¼1yiwi=Xna
i¼1wi:ð5Þ
The conditional probability of the estimate given the reference (i.e.,
human-interpreted) class, Pðcj^
cÞ(i.e., User's Accuracy, UA) and the
Fig. 2. Sampling of WRS-2 tile p224r078, located at the boundary of Paraguay, Argentina, and Brazil. The background image is a false-color (NIR-R-G) Landsat image of July 6, 2000.
76 M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
conditional probability of the reference class given the estimate Pð^
cjcÞ
(i.e., Producer's accuracy, PA) were calculated as:
CEc¼1−UAc¼1−∑n^
c
i¼1yiwi.∑n^
c
i¼1wi
ð6Þ
OEc¼1−PAc¼1−∑nc
i¼1yiwi.∑nc
i¼1wi
,(7)
where n^
cwere the points identified as type c(e.g., forest, non-forest,
forest gain, or forest loss) by the GLCF layers, and n
c
were the
points id entified as type cby the reference (Stehman, 2014).The inverse
of Pðcj^
cÞand Pð^
cjcÞwere interpreted as errors of commission and
omission respectively. The standard errors (SE) of the accuracy metrics
were calculated following the equations in Appendix A.1.
3.4. Area estimation
The reference points also provided a basis for sample-based estima-
tion of the areas of forest, non-forest, and of forest gain and loss
Table 2
Interpretation of the collected points for circa 1990, 2000, and 2005.
Type
Number of points
1990 2000 2005
Non-forest 10,657 11,244 11,929
Forest 15,221 15,194 14,448
Unknown 2025 1543 1494
Cloud 9 26 30
Shadow 30 28 28
Fig. 3. The web-based tool for visually identifyingforest cover at each sample point (Feng et al., 2012b).
77M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
(Olofsson et al., 2014).The proportion of forest-cover or -change class, c,
was estimated from the reference data and the mapped GLCF dataset
following:
^
ac¼1
XnG
k¼1AkXnG
k¼1Ak
1
XnT
j¼1cos φj
XnT
j¼1cos φj
1
Xnj
i¼1Nij Xnj
i¼1Nij
nic
ni
2
43
5
8
<
:9
=
;;
ð8Þ
where n
G
was the numberof biome-change strata and A
k
was the area
of stratum (k), which had n
T
selected WRS-2 tiles. The center point of
tile (j) was located at latitude (φ
j
), and the tile consisted of n
j
forest-sta-
tus strata. Forest-status stratum (i)consistedofN
ij
pixels, a ndnic
niwas the
number of points of class cover the total number of points in stratum
(i). Areal estimates with approximate 95% confidence were calculated
as ^
ac±1.96xAxffiffiffiffiffiffiffiffiffiffiffi
vð^
acÞ
p,whereAwas the total sampling area, equal to
104,460,279 km
2
for the eight biome strata, and vð^
acÞwas the variance
of the areal proportion (Appendix A.2). The difference between the
human-interpreted and the GLCF data at the points characterized the
“measurement bias”in the map (Stehman, 2013). The differences
were thenadded to themapped GLCF data areas to provide bias-adjust-
ed estimation of global forest-cover and forest-change areas.
4. Results
4.1. Accuracies of forest-cover layers
Accuracy of forest-cover detection was consistently high across all
biomes and epochs, with OA equaling 91% (SE ≈1%) in each of the
1990, 2000, and 2005 layers (Table 3). Commission errors (CE = 1-Pðcj
^
cÞ) and omission errors (OE = 1 -Pð^
cjcÞ)wereb10% for both forest and
non-forest classes in all epochs, for which SE b2.3%. The original, unad-
justed estimates showed a bias toward detection of non-forest, with the
forest class having a higher rate of omission errors (b21%) than com-
mission errors (b3%) and the non-forest class having a higher rate of
commission errors (b13%) than omission errors (b2%) in all epochs
and biomes (Table 4).
The largest overall accuracies (OA) were found in temperate forest
and non-forest, tropical evergreen, and boreal non-forest biomes—each
of which had OA N90% (SE b5%) (Table 4). OA were slightly lower in bo-
real forests (83% bOA b89%); OA of tropical deciduous forest ranged
from 80.7% to 84%; and OA of tropical non-forest ranged from 83.2% to
84.1%. Standard errors of OA were lowest (b1.6%) in evergreen forests
and temperate nonforest, slightly higher in deciduous and boreal forest
(b2.9%), and highest in boreal and tropical nonforest (b5%). Evergreen
and boreal forests had the lowest rate of omission error (OE b21%;
SE b3.5%) for the forest class, followed by deciduous forests
(24% bOE b55%; SE b9.6%) and non-forest biomes (59% bOE;
SE b7.6%). The non-forest class had low omission error (OE b10%;
SE b8.5%) in all biomes, and its commission error rate was larger
in the forest biomes (≤32.3%; SE b6.3%) than the non-forest biomes
(≤18.3%; SE b3.3%).
Although exclusion of biome boundaries could have artificially in-
creased the accuracies reported here, these estimates of accuracy are
likely conservative, given our exclusion of treeless biomes from the
sample and the uncertainty of identifying forest cover by visual inter-
pretation of satellite images (Montesano et al., 2009; Sexton et al.,
2015). Montesano et al. (2009) found that human experts achieved
18.7% RMSE in visual estimation of tree cover in high-resolution imag-
ery, and Sexton et al. (2015) found that visual confusion was greatest
near the threshold of tree cover used to define forests, especially
when interpretingchange. To investigate the relation between accuracy
and tree cover, OA of forest/non-forest cover in 2000 was plotted over
the range of coincident tree cover estimated by the Landsat tree-cover
dataset (Sexton et al., 2013a). A distinct concavity was evident in the re-
lation, which reached its minimum near the 30% tree-cover threshold
used to define forests (Fig. 6). The OA was large (N80%) where tree
cover was b0.1 or N0.35. Commission and omission errors were also in-
vestigated in relation to tree cover (Fig. 7). Commission error of the for-
est class was b10% except areas with tree cover b0.35, where the
commission error was b20%. Omission error of forest was b20% in
areas with N0.4 tree cover but increased in areas of sparse tree cover.
Fig. 4. Distribution of successfully interpreted points over the range of tree-canopy cover
estimated by the Landsat tree-cover (Sexton et al., 2013a).
Fig. 5. Percentage of “unknown”points interpreted for the 2000-epoch sample across the
range of tree-canopy cover estimated by the GLCF Landsat tree-cover layer (Sexton et al.,
2013a).
Table 3
Percentage accuracies of the 1990, 2000, and 2005 forest-cover layers relative to human-interpreted reference points. The standard error associated with each accuracy is reported in
parentheses.
Type
1990 2000 2005
Pðcj^
cÞPð^
cjcÞPðcj^
cÞPð^
cjcÞPðcj^
cÞPð^
cjcÞ
F97.2 (1.99) 79.8 (1.05) 98.2 (1.24) 79.9 (1.09) 97.9 (1.15) 79.8 (1.06)
N87.8 (1.93) 98.5 (1.10) 87.6 (2.28) 99.0 (1.19) 87.9 (2.20) 98.8 (1.44)
OA 90.9 (1.03) 91.1 (0.96) 91.2 (1.01)
78 M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
4.2. Accuracies of forest-change layers
Globally, overall accuracy of the 1990–2000 forest-change layer
equaled 88.1% (SE = 1.19%) and OA = 90.2% (SE = 1.1%) for the
2000–2005 forest-change layer (Table 5). In each period and biome,
OA ≥78.7% (SE b5%). The global accuraciesand standard errors of stable
forest (FF) and stable non-forest (NN) classes were similar respectively
to those ofthe stable forest and non-forest classes in the 1990, 2000,and
2005 layers, but the change classes—i.e., forest loss (FN) and forest gain
(NF)—had larger error rates than the static classes in the respective
epochs.
Commission and omission errors for forest loss were between 45%
and 62% globally, with SE between 1.72% and 23.48% (Table 5). Forest-
loss was detected most accurately, with errors dominated by commis-
sion, in temperate and tropical evergreen forest biomes (PA ≥71.7%;
UA ≥49.6%) (Table 6). This was likely due to relatively minimal impact
of vegetation phenology on canopy reflectance in evergreen forests.
Whether in temperate or tropical regions, detection of forest loss was
more accurate in evergreen forests than in their deciduous counterparts
(30% ≤PA b39%; 36.1% ≤UA ≤50.1%). In non-forest biomes, accuracy of
forest-loss detection was very low and dominated by omissions, but the
rarity of forests and their loss in these biomes made the impact of these
errors on overall accuracy small (Table 6).
Forest gain was consistently the most difficult dynamic to detect,
with OE and CE each N60% in all epochs (SE b17%) (Table 5). This was
likely due to the long traversal of intermediate tree cover during canopy
recovery from disturbance, compounded by the uncertainty of human
identification of change (Sexton et al., 2015). Producer's accuracies
tended to be largest in tropical evergreen forests (24.9% ≤PA ≤75.7%),
where canopy recovery following disturbance is fastest, and smallest
in non-forest biomes (PA b19%; UA b17%), where recovery is slower
and locations spend more time in intermediate ranges of canopy cover
(Table 6).
Table 4
Accuraciesof the global forest cover products estimated by biomes, expressed as percentages. The standard error associated with each accuracy is reported in parentheses.
Accuracy Type Boreal forest Boreal non-forest
Temperate
deciduous forest
Temperate
evergreen forest Temperate non-forest
Tropical
deciduous forest
Tropical
evergreen forest Tropical non-forest
OA 1990 88.2 (2.56) 98.1 (4.90) 93.0 (2.45) 93.9 (1.49) 98.4 (0.79) 80.7 (2.57) 93.7 (1.60) 83.2 (3.42)
2000 84.5 (2.81) 98.1 (1.95) 91.2 (2.54) 93.4 (1.41) 99.0 (0.56) 83.8 (2.46) 96.5 (1.10) 83.2 (3.43)
2005 83.7 (2.87) 98.2 (3.27) 90.1 (2.83) 93.0 (1.55) 99.2 (0.45) 84.0 (2.47) 96.7 (1.23) 84.1 (3.42)
Pð^
cjcÞF 1990 86.1 (1.66) 11.0 (2.35) 75.9 (9.57) 95.1 (3.31) 26.2 (5.76) 45.3 (2.68) 94.2 (2.48) 35.8 (2.09)
2000 80.1 (2.07) 12.1 (4.46) 72.3 (5.41) 92.0 (3.48) 38.6 (6.55) 47.5 (1.57) 96.6 (1.14) 37.2 (1.98)
2005 79.2 (2.54) 18.7 (7.60) 69.7 (1.97) 91.4 (3.00) 40.7 (3.33) 45.7 (1.51) 97.3 (1.63) 37.2 (1.64)
N 1990 92.9 (5.14) 100.0 (1.81) 98.7 (5.52) 92.3 (8.38) 100.0 (0.67) 98.8 (3.74) 90.6 (5.96) 99.5 (3.75)
2000 94.4 (6.82) 100.0 (1.81) 98.8 (4.39) 95.5 (6.83) 100.0 (0.67) 99.8 (3.09) 95.8 (5.86) 99.5 (6.85)
2005 93.2 (7.24) 100.0 (1.92) 98.9 (3.32) 95.6 (8.41) 100.0 (0.55) 99.6 (3.67) 93.8 (6.66) 99.8 (3.78)
Pðcj^
cÞF 1990 96.4 (3.17) 94.6 (0.00) 95.4 (2.88) 94.6 (2.59) 92.9 (3.54) 95.1 (2.20) 98.1 (1.31) 96.4 (2.25)
2000 97.0 (3.21) 87.6 (0.07) 96.2 (2.87) 97.1 (2.58) 94.4 (7.52) 98.9 (2.16) 99.2 (1.42) 96.5 (2.28)
2005 96.1 (3.22) 91.6 (0.04) 96.4 (2.88) 97.0 (3.12) 95.0 (3.45) 98.1 (2.16) 98.6 (1.33) 98.5 (2.22)
N 1990 75.0 (3.21) 98.1 (0.00) 92.4 (2.88) 92.9 (2.61) 98.4 (0.16) 78.0 (2.25) 74.9 (3.51) 81.8 (1.64)
2000 67.9 (2.90) 98.1 (0.02) 89.8 (2.86) 88.1 (2.59) 99.0 (0.17) 81.2 (2.17) 84.4 (4.74) 81.7 (1.02)
2005 67.7 (3.19) 98.2 (0.01) 88.4 (2.88) 87.7 (3.10) 99.2 (0.14) 81.8 (2.16) 88.3 (6.26) 82.6 (0.76)
Fig. 6. Overall accuracies of forest cover in relation to circa-2000 tree cover. Tree-cover
estimates were taken from Sexton et al. (2013a).
Fig. 7. Accuracies of forest (A) and non-forest (B) in relation to circa-2000 tree cover (Sexton et al., 2013a).
79M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
The effect of tree cover on accuracy was investigated using the
2000–2005 forest-change layer (Fig. 8). Similar to that of the 2000 for-
est-cover layer, a distinct concavity was evident in the relationship be-
tween overall forest-change accuracy and tree cover, and accuracy
was lowest between 0.2–0.3 tree cover. Commission and omission er-
rors of stable forest and non-forest in relation to tree cover weresimilar
to those of forest and non-forestin the static layers(Fig. 9). The commis-
sion and omission error was high in areas with tree cover b0.35 and de-
creased to b60% in areas with tree cover N0.35. Commission and
omission errors of forest gain were both correlated to tree cover. The
omission error was b45% and commission error was b70% in areas
with 0.3–0.6 tree cover but N50% in high or low tree cover.
4.3. Global forest-area estimation
Table 7 reports estimates of the global areasof forest, non-forest, for-
est loss, and forest gain from the reference sample of human-
interpreted cover and the mapped GLCF estimates in 1990, 2000, and
2005. The sample of visually interpreted points yielded estimates of
40.18, 39.76, and 39.25 million km
2
of forest in circa 1990, 2000, and
2005 respectively. Sampling the GLCF estimates at the points and
adjusting for bias relative to the visual estimates yielded global esti-
mates of 39.28 ± 1.34 million km
2
in 2000 and 38.81 ± 1.34 million
km
2
in 2005, as well as a sub-global estimate of 33.16 ± 1.36 million
km
2
in 1990, for which the global coverage of Landsat images is incom-
plete (Channan et al., 2015; Gutman et al., 2008; Kim et al., 2014). Ad-
justed to the reference estimates, the GLCF layers reported 0.73 ±
0.38 million km
2
of forest loss and 0.28 ± 0.26 million km
2
of forest
gain between 2000 and2005, with average annual rates of forest loss es-
timated at0.15 ± 0.08 million km
2
/ year and forest gain at 0.06 ± 0.05
million km
2
/ year. The estimated forest loss and gain between 1990 and
2000 were 1.08 ± 0.53 and 0.53 ± 0.47 million km
2
respectively, with
forest loss and gain rates at roughly 0.11 ± 0.05 and 0.05 ± 0.05 million
km
2
/ year respectively.
5. Discussion
5.1. Global forest-area estimation
A growing community of research is developing around the goal of
detecting and estimating the area of forest cover and change globally
at high- (e.g., sub-hectare) resolution (Townshend et al., 2012). Al-
though variance remains due to differences in data, methods, and
even fundamental definitions of “forest”(Sexton et al., 2016), consensus
on the area and distribution of global forest cover is beginning to
emerge. The United Nations' 2010 Forest Resources Assessment (FRA)
(FAO, 2010) reported that the world's forests covered 41.68, 40.85,
40.61 million km
2
in 1990, 2000,2005—equaling about 31% of the global
land area. Hansen, Stehman, and Potapov (2010);Hansen et al. (2013)
calculated global forest areas of 32.7 and 41.5 million km
2
in successive
Landsat-based analyses, and Shimada et al. (2014) estimated 38.54,
38.22, 38.19, and 38.52 million km
2
of forest cover globally in 2007,
2008, 2009 and 2010 respectively, based on polarimetric L-band radar
measurements. Schepaschenko et al. (2015) estimated 33 million km
2
of global forest area by integrating eight prior forest data products, in-
cluding an early version of the global percent-tree canopy layer by
Sexton, Song, et al. (2013a), upon which our estimates of forest cover
and change here were based. Our adjusted estimates of 39.28 ± 1.34
million km
2
in 2000 and 38.81 ± 1.34 million km
2
in 2005 lie within
the range of these other global forest-area estimates; and although the
Hansen et al. (2013) and Shimada et al. (2014) estimates were adjusted
to match those of the FRA, they were both consistent with our estimates
that were adjusted based solely on independent visual interpretation.
The GLCF forest-cover and forest-change accuracies were among the
highest accuracies reported for recent global forest- and land-cover data
products. Gong et al. (2013) reported maximum UA of 80% and PA of
76% for a forest class mapped at 30-m resolution with four classifiers.
Chen et al. (2015) reported UA equaling 84% and PA equaling 92% for
the forest class of a 30-m resolution global land cover map produced
by pixel- and object-based classification and intensive human editing.
Shimada et al. (2014) reported overall accuracies equaling 85%–95% rel-
ative to independent reference datasets. Schepaschenko et al. (2015)re-
ported 93% overall, user's, and producer's accuracies for a 1-km
resolution global hybrid forest mask. Hansen et al. (2013) reported
87% user's accuracy and 88% producer's accuracy for global forest-loss
detection from 2000 to 2012, with which Kim et al. (2015) found strong
Table 5
Percentage accuracies of the global forest cover change layers for 1990–2000 and 2000–
2005 periods. The standard errorassociated witheach accuracyis reported in parentheses.
Type
1990–2000 2000–2005
Pðcj^
cÞPð^
cjcÞPðcj^
cÞPð^
cjcÞ
FF 97.5 (1.98) 78.5 (1.07) 98.2 (1.17) 79.4 (1.07)
FN 38.1 (3.60) 45.2 (4.63) 55.0 (5.89) 52.7 (2.16)
NF 15.3 (4.56) 16.8 (8.84) 34.0 (5.21) 39.3 (1.44)
NN 88.1 (2.75) 98.8 (1.72) 87.7 (2.43) 98.9 (1.67)
OA 88.1 (1.19) 90.2 (1.10)
Table 6
Percentage accuracies of the global forest cover change layers, estimated by biomes. The standard error associated with each accuracy is reported in parentheses.
Accuracy Type Boreal forest
Boreal
non-forest
Temperate
deciduous forest
Temperate
evergreen forest
Temperate
non-forest
Tropical
deciduous forest
Tropical
evergreen forest
Tropical
non-forest
OA 1990–2000 83.0 (3.30) 98.0 (4.99) 88.0 (3.07) 90.0 (1.81) 98.3 (0.85) 78.7 (2.50) 91.7 (2.06) 80.8 (3.49)
2000–2005 81.8 (3.04) 98.0 (3.83) 88.7 (2.99) 91.6 (1.44) 99.0 (0.58) 82.3 (2.49) 95.8 (1.92) 83.2 (3.44)
Pð^
cjcÞFF 1990–2000 81.5 (1.97) 9.8 (2.81) 76.0 (9.59) 93.5 (2.38) 35.4 (5.91) 43.6 (2.73) 93.2 (1.44) 33.6 (2.58)
2000–2005 77.7 (2.39) 12.7 (8.39) 71.9 (1.72) 91.3 (2.08) 39.8 (4.90) 45.6 (1.34) 96.8 (1.25) 36.5 (1.97)
FN 1990–2000 53.3 (10.12) 24.9 (14.29) 30.5 (7.10) 85.3 (11.76) 1.5 (7.35) 30.0 (14.12) 71.8 (7.01) 22.6 (3.59)
2000–2005 34.6 (8.42) –36.0 (15.18) 71.7 (11.53) 1.5 (7.93) 38.8 (19.04) 72.0 (11.52) 41.2 (23.48)
NF 1990–2000 35.9 (14.79) 5.2 (3.48) 10.6 (9.33) 29.3 (12.80) 2.2 (9.37) 12.9 (14.37) 24.9 (8.75) 4.9 (6.39)
2000–2005 45.6 (16.60) 0.2 (0.09) 18.9 (5.39) 35.2 (7.41) 18.6 (10.94) 18.9 (14.79) 75.7 (9.10) 0.1 (11.71)
NN 1990–2000 93.8 (10.18) 100.0 (1.82) 98.7 (5.57) 93.4 (7.92) 99.9 (0.74) 99.5 (3.40) 94.6 (6.07) 99.4 (3.89)
2000–2005 94.2 (8.31) 100.0 (2.24) 98.7 (3.43) 95.1 (8.38) 100.0 (0.68) 99.6 (3.11) 94.8 (7.04) 99.5 (3.83)
Pðcj^
cÞFF 1990–2000 95.9 (3.19) 93.7 (0.00) 95.6 (2.88) 95.8 (2.69) 96.3 (3.41) 96.7 (2.13) 98.5 (1.47) 97.2 (2.35)
2000–2005 96.3 (3.22) 87.0 (0.04) 96.6 (2.89) 97.1 (2.63) 94.4 (3.76) 99.1 (2.19) 99.0 (1.49) 98.5 (2.22)
FN 1990–2000 25.1 (3.22) 59.4 (1.72) 36.1 (3.37) 49.6 (2.84) 14.3 (2.68) 45.6 (2.44) 50.4 (17.86) 25.0 (9.31)
2000–2005 23.6 (3.85) 49.5 (10.59) 40.0 (5.72) 63.1 (14.64) 3.7 (12.93) 50.1 (2.16) 76.9 (4.02) 52.6 (3.56)
NF 1990–2000 33.1 (6.36) 99.6 (15.43) 18.7 (3.78) 47.9 (2.87) 1.6 (3.99) 13.8 (2.95) 11.1 (4.05) 5.0 (1.70)
2000–2005 15.6 (3.61) 0.5 (0.02) 37.2 (2.95) 32.8 (2.86) 16.7 (4.86) 27.4 (2.79) 49.2 (4.38) 18.7 (2.38)
NN 1990–2000 74.8 (6.65) 98.2 (0.02) 89.4 (2.87) 89.1 (2.60) 98.4 (0.21) 78.3 (3.06) 86.7 (3.92) 81.5 (1.90)
2000–2005 68.8 (4.00) 98.2 (0.02) 88.1 (2.87) 87.6 (2.67) 99.1 (0.16) 80.7 (2.23) 86.8 (7.19) 82.1 (1.03)
80 M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
correlation (R
2
= 0.96) of preliminary estimates from the GLCF 2000–
2010 forest-cover and –change layers across the humid tropics.
Even while the various estimates of forest cover are converging, de-
tection of change remains comparatively challenging. The FRA2010 re-
ported −0.08 million km
2
and −0.05 million km
2
annual forest
change in 1990–2000 and 2000–2005, respectively. However,
confirming our previous estimates across the humid tropics (Kim et
al., 2015), the annual net forest-change rates estimated from the GLCF
data were 30.7% lower in 1990–2000 (−0.06 million km
2
) but 82.5%
higher in 2000–2005 (−0.09 million km
2
/year) thanthe FRA estimates.
Hansen et al. (2013) reported 2.29 million km
2
forest loss and 0.80 mil-
lion km
2
forest gain between 2000 and 2012, with annual rates of forest
loss and gain at 0.18 and 0.06 million km
2
, which were within the 95%
confidence level of our estimates of annual change rates. Corroborating
other efforts to detect change (Hansen et al., 2013; Potapov et al., 2011),
forest gainswere consistently the most difficult dynamicto detect. More
research is therefore still needed to increase the precision of satellite-
based forest-change detection to match the growing consensus among
global estimates of cover.
A major source of imprecision is semantic differences among
datasets (Sexton et al., 2016). Our definition of forest is based on criteria
consistent with those of the UNFCC and FAO, although the thresholds
were different. Our tree-cover threshold (30%) was more conservative,
which could lead to smaller estimates of forest areas (Sexton et al.,
2016). However, our minimum-mapping-unit (MMU) threshold of
0.27 ha was smaller, enabled by the pixel-size of Landsat data and in
turn enabling us to resolve smaller forest patches. Our spatial filtering
also removes the “tail”pixels at the end of linear features with exactly
one-pixel width; we neglected this effect in our area calculations due
to an assumed rarity in forests and it is potentially offset by the opposite
effect of also removing the tails of non-forest pixels. Neither actual nor
potential tree-height were considered due to their immeasurability in
currently available satellite imagery (Lefsky, 2010; Smart et al., 2012),
but an earlier study showed strong correlation (R
2
= 81%) between
the GLCF tree-cover estimates and percentage-cover of trees taller
than 5 m (Sexton et al., 2013a). The various thresholds likely also con-
tributed to the variance between estimations of global area by ourselves
and others.
5.2. Challenges and recommendations for global accuracy assessment
Independent assessment of accuracy is fundamental to improving
the reliability of maps of forest cover and change. Although a seemingly
Fig. 8. Overall accuracyof forest-cover change (2000–2005) in relation to circa-2000 tree
cover (Sexton et al., 2013a).
Fig. 9. Accuracy of the forest-cover change (2000–2005) layer in relation to circa-2000 tree cover (Sexton et al., 2013a).
81M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
simple conceptual exercise, accuracy assessment is complex and labori-
ous for spatio-temporally extensive land-cover datasets (Congalton,
1991; Foody, 2002; GFOI, 2013; McRoberts, 2011; McRoberts &
Walters, 2012; Olofsson et al., 2014; Stehman, 2000)—especially for
large regions or the globe (Olofsson, Foody, Stehman, & Woodcock,
2013; Tsendbazar, de Bruin, & Herold, 2014). A static, 30-m resolution
dataset covering Earth's terrestrial surface comprises roughly 166 bil-
lion pixels.
Validating datasets of such scope is challenging due to the difficulty
of collecting representative points across the range of forest types, nat-
ural and anthropogenic changes,and other environmental factors. In re-
cent years, efforts have been made to produce globally distributed
referencepoints by crowd-sourced, visual interpretation of high-resolu-
tion imagery made available by Google Earth ™(Fritz et al., 2011;
McCallum et al., 2015; Zhao et al., 2014). Corroborating these studies,
we have demonstrated that stratification of sampling across biomes
and independent, preliminary datasets provide an efficient sampling
framework for comprehensively assessing the accuracies and errors of
global data products at both global and biome scales. We also showed
that the adoption of a probabilistic approach provides insights into
how and where errors arise that provide a solid basis for where to
focus further efforts to improve global products.
Stehman and Czaplewski (1998) provide general guidelines for ac-
curacy assessment of land-cover datasets, acknowledging that “deci-
sions [among options for sampling, response, and estimation &
analysis protocols] should be based on the strengths and weaknesses
of each option to meet project objectives and practical constraints”.
Our exclusion of biomes, tiles, and pixels in the sampling design were
necessary practicalities of such a large effort. Biomes provided an inde-
pendent stratification based on climate and forest types (Olson et al.,
2001), and exclusion of the tiles spanning their boundaries was neces-
sary for meeting the assumptions of stratification. Excluding marginal
pixels and applying a 0.27-ha minimum mapping unit was necessary
to minimize the impact of the 50-m (1σ) geo-location accuracy of
Landsat images (Tucker, Grant, & Dykstra, 2004). Whereas these steps
might have resulted in slight over-estimation of accuracy, these effects
were likely offset by the exclusion of predominantly treeless biomes
(i.e., Deserts and Xeric Shrublands, Inland Water and Rock and Ice),
where the data products likely had higher accuracy.
The challenges of global accuracy assessment are multiplied when
considering multiple dates. Constrained by the availability of high-reso-
lution imagery, the reliability of reference datasets diminishes for
assessing land cover and change before the current era of high data
availability. Further, existing human-interpreted reference data (e.g.,
Fritz et al., 2011) may not precisely estimate the accuracy of satellite-
based data due to temporal mismatches between images used forrefer-
ence and for estimation.
Overcoming these challenges requires the use of human visual inter-
pretation supported by a diversity of information. Interpreting forest
cover at selected points in the classified images provides reference ob-
servations perfectly matching the targeted datasets, thus allowing accu-
racy assessment of and between each epoch using coincident
observations. Beyond that which is currently possible through automa-
tion, human cognition is capable of more reliable interpretation by in-
vestigating local reflectance in the context of surrounding patterns in
space and time, as well as expert knowledge on local ecology and land
use. However, human interpretation is also prone to error and uncer-
tainty (Montesano et al., 2009; Sexton et al., 2015). Our findings here
corroborate previous conclusions that errors and uncertainty in
human interpretation were associated with sparsely forested areas,
the tree-cover of which was near the decision threshold for discriminat-
ing forest from non-forest cover.
It is thus important toprovide analysts with a variety of information
to support the decision-making rules unique to each analyst. Web-en-
abled labeling tools (e.g. Feng et al., 2012b; Fritz et al., 2011; Zhao et
al., 2014) provide an efficient means of interpreting forest or other
cover types rapidly at a large number of locations by a distributed com-
munity of interpreters. High-resolution imagery is especially crucial in
sparsely forested regions (e.g., savanna, boreal forest). References de-
rived from high-resolution imagery can also be used for investigating
errors in the forest data products caused by remnant radiometric and
positional accuracies in Landsat data (Feng et al., 2013; Tucker et al.,
2004). Other geospatial and temporal data, including vector maps,
time-serial vegetation indices, vegetation height, and georeferenced
field photos provide complementary information to visually interpret
cover (Fritz et al., 2011; Lefsky, 2010; McCallum et al., 2015; Olofsson
et al., 2014). Spanning the range of spatial and temporal scales and a va-
riety of spectral and ecological characteristics with data relevant to each
analyst's expertise, these tools enable analysts together to span the
range of ecological and land-use conditions globally.
6. Conclusions
The Global Land Cover Facility (GLCF) global forest-cover and -
change dataset is a multi-temporal depiction of long-term, global forest
dynamics at 30-m resolution. Based on per-pixel estimates of tree cover
and their associated uncertainty, thedataset currently represents binary
forest cover in nominal 1990, 2000, and 2005 “epochs”, as well as gains
and losses between epochs. Understanding of errors and uncertaintiesis
crucial to use of the data, either for scientific application or for fusion
with other Earth-science datasets. Consistent across epochs, the overall
accuracy of the dataset is 91% for forest cover and N88% for forest-
change. Accuracy is lower in sparsely forested areas—i.e., with tree
cover near the 30%-cover threshold used to define forest from non-
forest—and for forest gain compared to static cover and forest loss. Dis-
crimination of forest had a low rate of commission relative to omission
error, especially in areas with low tree density. After adjusting global
area estimates to independent reference data, 39.28 ± 1.34 million
km
2
and 38.81 ± 1.34 million km
2
of forest were identified in 2000
and 2005 globally, and 33.16 ± 1.36million km
2
of forest were estimat-
ed for 1990 for the available coverage of Landsat data. The GLCF forest
datasets are available for free public download at the GLCF website
(http://www.landcover.org).
Table 7
Global forest-cover and-change areas estimated from the reference sample and mapped GLCF dataset.
Classes
Sample-based estimation (km
2
) Pixel-based estimation (km
2
)
Interpreted Mapped Difference 95% confidence GLCF Adjusted
Forest 1990 40,181,147 33,303,728 6,877,420 1,359,789 26,280,999⁎33,158,419⁎
2000 39,763,403 32,692,869 7,070,533 1,334,874 32,204,035 39,274,568
2005 39,251,944 32,246,667 7,005,277 1,340,848 31,803,519 38,808,796
Forest loss 1990–2000 1,947,155 2,075,064 −127,910 533,205 1,207,117⁎1,079,207⁎
2000–2005 1,269,297 1,097,836 171,461 379,093 562,114 733,575
Forest gain 1990–2000 1,451,870 1,504,117 −52,246 468,094 576,759⁎524,513⁎
2000–2005 702,799 601,633 101,166 255,514 176,243 277,409
⁎Due to theunavailable Landsat datain eastern Russiaand western India(Channan et al., 2015;Gutman, Huang,Chander, Noojipady, & Masek, 2013;Kim et al., 2014), theareas for 1990
were only for the incomplete global coverage.
82 M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
Acknowledgements
Support for this effort was provided by the following National Aero-
nautics and Space Administration (NASA) programs: Making Earth Sci-
ence Data Records for Use in Research Environment (NNH06ZDA001N-
MEaSUREs), Land Cover and Land Use Change (NNH07ZDA001N-
LCLUC), NASA ACCESS (NH11ZDA001N-ACCESS) and NASA Indicators
(NNH12ZDA001N-INCA). We thank Linda Jonescheit Owen of LPDAAC
U.S. Geological Survey (USGS) for supporting our large Landsat data re-
quests, and thank our colleagues Katie Collins, Dr. Fu-Jiang Liu, and
Guang-Xiao Zhang for their efforts on interpreting the points. We would
also like to thank the four anonymous reviewers whose constructive com-
ments led to a better presentation of our research methods and results.
Appendix A
A.1. Variance of accuracy metrics
The variance of the accuracy metrics is described below. The points
in each forest/non-forest status stratum were randomly selected.
Hence, the variance of the OA for the stratum and the UA and PA of
class c(i.e., forest and non-forest for forest cover; FF, FN, NF, and NN
for forest-cover change) in the stratum were calculated following
Congalton & Green (2010: p116–119) and Olofsson et al. (2014):
vc
OA
¼1
∑n
i¼1n2
þi
∑
n
i¼1
n2
þic
UAi
1−c
UAi
niþ−1:
vc
UAc
¼c
UAc
1−c
UAc
ncþ−1
vc
PAc
¼1
∑n
k¼1
nþk
nkþ
nkc
n2
þc1−c
PAc
2c
UAc1−c
UAc
ncþ−1þc
PA2
c∑
n
i≠c
n2
þi
nic
niþ
1−nic
niþ
niþ−1ðÞ
2
6
6
43
7
7
5;
where n
ij
was the number of points in the error matrix at cell (i,j),
and n
i+
and n
+j
were respectively the summaries of row (i) and col-
umns (j) in the matrix.
The estimated variances (vð
^
θÞ) for the accuracy metrics (i.e., OA,UA,
and PA) of the globe and each biome were calculated following
(Cochran, 1977):
v
^
θ
¼∑
nG
k¼1
Ak
∑nG
l¼1Al
!
21
nG
∑
nT
j¼1
Wj
^
θj−b
θj
2
þ∑
nj
i¼1
W2
ij v
^
θij
"#
where, a biome (G) consisted of n
G
biome-change strata. Each
biome-change stratum (k) covered A
k
area and included n
T
selected
WRS-2 tiles. The weight for each tile (j) was calculated as:
Wj¼
cos φj
∑nT
i¼1cos φi
ðÞ
where, φ
i
is the central latitude of tile (j). A tile (j) consisted of n
j
for-
est status strata, and the accuracy for the tile (
^
θj) were estimated:
^
θj¼∑
nj
i¼1
Wij
^
θij
where, W
ij
was the weight for a forest status stratum (i) within tile
(j):
Wij ¼Nij
∑nj
i¼1Nij
where, N
ij
was the number of pixels in stratum (i)oftile(j). The
mean (b
θj) of accuracy (
^
θj) for tile (j) was calculated:
b
θj¼∑
nT
j¼1
Wj
^
θj
The standard error (SE) of the accuracy metrics was calculated as
square root of variance.
SE
^
θ
¼ffiffiffiffiffiffiffiffiffiffiffi
v
^
θ
r
A.2. Variance of area estimation
Similarly, the variance of the area estimates was calculated:
v^
ac
ðÞ¼
XnG
k¼1
Nk
XnG
l¼1Nl
0
@1
A
2
1
nsXnT
j¼1Wj
^
aj−XnT
i¼1Wj
^
aj
2þXnT
j¼1Xnj
i¼1WjWij
2
^
aij 1−^
aij
nij−1
:
The areal proportion of class c (^
aj) in tile (j) was calculated as the
weighted mean of the forest status strata in the tile:
^
aj¼∑
nj
i¼1
Wij
^
aij
where, the areal proportion (^
aij) in a forest status stratum (i) was cal-
culated by dividing the number of points of class c(m
ij
) by the total
number of points in the stratum (n
ij
).
References
Band, L.E.,1993. Effect of land surface representation on forest waterand carbon budgets.
Journal of Hydrology 150, 749–772. http://dx.doi.org/10.1016/0022-1694(93)90134-
U.
BenDor, T.,Westervelt, J., Song, Y., Sexton, J.O.,2013. Modeling parkdevelopment through
regional land use change simulation. Land Use Policy 30 (1), 1–12. http://dx.doi.org/
10.1016/j.landusepol.2012.01.012.
Bonan, G.B., 2002. Ecological Climatology: Concepts and Applications. Cambridge Univer-
sity Press, Cmbridge, UK http://dx.doi.org/10.1111/j.1745-5871.2009.00640.x.
Card, D.H., 1982. Using known map category marginal frequencies to improve estimates
of thematic map accuracy. Photogrammetric Engineering and Remote Sensing 48
(3), 431–439.
Channan, S., Feng, M., Kim, D., Sexton, J.O., Song, X., Song, D., Noojipady, P., Collins, K.,
Anand, A., Townshend, J.R., 2015. The GLS+: an Enhancement of the Global Land
Survey Datasets. Photogrammetric Engineering & Remote Sensing 81 (7),
521–525.
Chen, J., Chen, J., Liao, A., Cao, X., Chen, L., Chen, X., He, C., Han, G., Peng, S., Lu, M., Zhang,
W., Tong, X., Mills, J., 2015. Global land cover mapping at 30 m resolution: A POK-
based operational approach. ISPRS Journal of Photogrammetry and Remote Sensing
103, 7–27. http://dx.doi.org/10.1016/j.isprsjprs.2014.09.002.
Cochran, W.G., 1977. Sampling Techniques. third ed. John Wiley & Sons, New York.
Conde, D.a., Colchero, F., Zarza,H., Christensen, N.L., Sexton, J.O., Manterola, C., Chávez, C.,
Rivera, A., Azuara, D., Ceballos, G., 2010. Sex matters: Modeling male and female hab-
itat differences for jaguar conservation. Biological Conservation 143 (9), 1980–1988.
http://dx.doi.org/10.1016/j.biocon.2010.04.049.
Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely
sensed data. Remote Sensing of Environment 37 (1), 35–46. http://dx.doi.org/10.
1016/0034-4257(91)90048-B.
Congalton, R.G., Green, K., 2010. Assessing the Accuracy of Remotely Sensed Data: Princi-
ples and Practices. The Photogrammetric Record 25. http://dx. doi.org/10.11 11/j.
1477-9730.2010.00574_2.x.
FAO, 2010. Global Forest Resources Assessment 2010. Food and Agriculture Organization
of the United Nations. Rome: Food and Agriculture Organization of the United Na-
tions ISBN: 978–92–5-106654-6.
Feng, M., Huang, C., Channan, S., Vermote, E.F.,Masek, J.G., Townshend, J.R.,2012a. Quality
assessment of Landsat surface reflectance products using MODIS data. Computers &
Geosciences 38 (1), 9–22. http://dx.doi.org/10.1016/j.cageo.2011.04.011.
Feng, M., Huang, C., Sexton, J.O., Channan, S., Narasimhan, R., Townshend, J.R., 2012b. An
approachfor quickly labelingland cover types for multipleepochs at globally selected
locations. IEEE International Geoscience and Remote Sensing Symposium 2012,
6203–6206. http://dx.doi.org/10.1109/IGARSS.2012.6352674.
Feng, M., Sexton, J.O., Channan,S., Townshend, J.R., 2015. A global, high-resolution (30-m)
inland water body dataset for 2000: first results of a topographic-spectral classifica-
tion algorithm. International Journal of Digital Earth 1–21 http: //dx.doi.org/10.
1080/17538947.2015.1026420.
83M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
Feng, M., Sexton, J.O., Huang, C., Masek, J.G., Vermote, E.F., Gao, F., Naras imhan, R.,
Channan, S., Wolfe, R.E., Townshend, J.R., 2013. Global surface reflectance products
from Landsat: Assessment using coincident MODIS observations. Remote Sensing of
Environment 134, 276–293. http://dx.doi.org/10.1016/j.rse.2013.02.031.
Foody, G.M., 2002. Status of land cover classification accuracy assessment. Remote Sens-
ing of Environment 80 (1), 185–201. http://dx.doi.org/10.1016/S0034-
4257(01)00295-4.
Fritz, S., McCallum, I., Schill, C., Perger, C., See, L., Schepaschenko, D., van der Velde, M.,
Kraxner, F., Obersteiner, M., 2011. Geo-Wiki: An online platform forimproving global
land cover. Environmental Modelling & Software 1–14 http://dx.doi.org/10.1016/j.
envsoft.2011.11.015.
GFOI, 2013. Integrating remote-sensing and ground-based observations for estimation of
emissions and removals of greenhouse gases in forests. Methods and Guidance from
the Global Forest. Group on Earth Observations Observations Initiative, Geneva,
Switzerland.
Gong, P., Wang, J., Yu, L.L., Zhao, Y.Y.Y.Y., Liang, L., Niu, Z., Huang, X., Fu, H., Liu, S., Li, C., Li,
X., Fu, W., Liu, C., Xu, Y.,Wang, X., Cheng, Q., Hu,L., Yao, W., Zhang, H.H., Zhu, P., Zhao,
Z., Zheng, Y., Ji, L., Zhang, Y., Chen, H., Yan, A., Guo, J., Wang, L., Liu, X., Shi,T., Zhu, M.,
Chen, Y., Yang, G., Tang, P.,Xu, B., Giri, C., Clinton,N., Zhu, Z., Chen, J.J.,2013. Finer res-
olution observation and monitoring of global land cover: first mapping results with
Landsat TM and ETM + data. International Journal of Remote Sensing 34 (7),
2607–2654. http://dx.doi.org/10.1080/01431161.2012.748992.
Gutman, G., Byrnes, R., Masek, J., Covington, S., Justice, C., Franks, S., Headley, R., 2008. To-
wards Monitoring Land-cover and Land-use Changes at a Global Scale: The Global
Land Survey 2005. Photogrammetric Engineering & Remote Sensing 74 (1), 6–10.
Gutman, G., Huang, C., Chander, G., Noojipady, P., Masek, J.G., 2013. Assessment of the
NASA-USGS Global La nd Survey (GLS) datasets. Remote Sensing of Environment
134, 249–265. http://dx.doi.org/10.1016/j.rse.2013.02.026.
Haddad, N.M., Brudvig, L.A., Clobert, J., Davies, K.F., Gonzalez, A., Holt, R.D., Lovejoy, T.E.,
Sexton, J.O., Austin, M.P., Collins, C.D., Cook, W. M., Damschen, E.I., Ewers, R.M.,
Foster, B.L., Jenkins, C.N., King, A.J., Laurance, W.F., Levey, D.J., Ma rgules, C.R.,
Melbourne, B.A., Nicholls, A.O., Orrock, J.L., Song, D., Townshend, J.R., 2015. Habitat
fragmentation and its lasting impact on Earth’s ecosystems. Science 1–9 (March).
Hansen, M.C., Loveland, T.R., 2012.A review of large areamonitoring of land cover change
using Landsat data. Remote Sensing of Environment 122, 66–74. http://dx.doi.org/10.
1016/j.rse.2011.08.024.
Hansen, M.C., Egorov, a., Potapov, P.V., Stehman, S.V., Tyukavina, a., Turubanova, S.a., Roy,
D.P., Goetz, S.J., Loveland, T.R.,Ju, J., Kommareddy, a., Kovalskyy, V., Forsyth, C., Bents,
T., 2014. Monitoring conterminous United States (CONUS) land cover change with
Web-Enabled Landsat Data (WELD). Remote Sensing of Environment 140, 466–484.
http://dx.doi.org/10.1016/j.rse.2013.08.014.
Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.a., Tyukavina, A., Thau,
D., Stehman, S.V., Goetz, S.J., Loveland, T.R., Kommareddy, A., Egorov , A., Chini, L.,
Justice, C.O., Townshend, J.R.G., 2013. High-Resolution Global Maps of 21st-Century
Forest Cover Cha nge. Science 342 (6160), 850–853. http://dx.doi.org/10.1126/
science.1244693.
Hansen, M.C., Stehman, S.V., Potapov, P.V., 2010. Quantification of global gross forest
cover loss. Proceedings of the National Academy of Sciences of the United States of
America 107 (19), 8650–8655. http://dx.doi.org/10.1073/pnas.0912668107.
Houghton, R.A., 1998. Historic role of forests in the global carbon cycle. Carbon dioxide
mitigation in forestry and wood industry. Springer, pp. 1–24.
Huang, C., Kim, S., Altstatt, A., Townshend, J.R.G., Davis, P., Song, K., Tucker, C.J., Rodas, O.,
Yanosky, A., Clay, R., Musinsky, J., 2007. Rapid loss of Paraguay's Atlantic forest and
the status of protected areas —A Landsat assessment. Remote Sensing of Environ-
ment 106 (4), 460–466. http://dx.doi.org/10.1016/j.rse.2006.09.016.
Huang, C., Song, K., Kim, S., Townshend, J.R.G., Davis, P., Masek, J.G., Goward, S.N., 2008.
Use of a dark object concept and support vector machines to automate forest cover
change analysis. Remote Sensing of Environment 112 (3), 970–985. http://dx.doi.
org/10.1016/j.rse.2007.07.023.
Kim, D., Sexton, J.O.,Townshend, J.R., 2015. Accelerated deforestation inthe humid tropics
from the 1990s to the 2000s. Geophysical Research Letters 42, 1–7. http://dx.doi.org/
10.1002/2014GL062777.Received.
Kim, D.-H., Sexton, J.O., Noojipad y, P., Huang, C., An and, A., Channan, S ., Feng, M.,
Townshend, J.R. , 2014. Global, L andsat-based fo rest-cover change from 1990 to
2000. Remote Sensing of Environment 155, 178–193. http://dx.doi.org/10.1016/j.
rse.2014.08.017.
Lal, R., 1995. Sustainable management of soil resources in the humid tropics Vol. 876.
United Nations University Press (Retrieved from) http://archive.unu.edu/unupress/
unupbooks/uu27se/uu27se00.htm.
Lefsky, M.A., 2010. A global forest canopy height map from the moderate resolution im-
aging spectroradiometer and the geoscience laser altimeter system. Geophysical Re-
search Letters 37 (15), 1–5. http://dx.doi.org/10.1029/2010GL043622.
Masek, J.G., Vermote, E.F., Saleous, N.E., Wolfe, R., Hall, F.G., Huemmric h, K.F., Gao, F.,
Kutler, J., Lim, T., 2006. A Landsat Sur face Reflectance Dataset for North America,
1990–2000. IEEE Geoscience and Remote Sensing Letters 3 (1), 68–72.
Masek, J., Huang, C., Wolfe, R., Cohen, W., Hall, F., Kutler, J., Nelson, P., 2008. North Amer-
ican forest disturbance mapped from a decadal Landsat record. Remote Sensing of
Environment 112 (6), 2914–2926. http://dx.doi.org/10.1016/j.rse.2008.02.010.
Mason, P.,Reading, B., 2004. Implementation plan for theglobal observing systems for cli-
mate in supportof the UNFCCC. 21st international conference on interactive informa-
tion processing systems for meteorology, oceanography, and hydrology (San Diego).
McCallum, I., Fritz, S., Schill, C., Perger, C., Achard, F., Grillmayer, R., Koch, B., Kraxner, F.,
Obersteiner, M., Quinten, M., 2015. Harnessing the power of volunteers, the internet
and Google Earth to collect and validate global spatial information using Geo-Wiki.
Technological Forecasting and Social Change http://dx.doi.org/10.1016/j.techfore.
2015.03.002.
McRoberts, R.E., 2011. Satellite image-based maps: Scientific inference or pretty pictures?
Remote Sensing of Environment 115 (2), 715–724. http://dx.doi.org/10.1016/j.rse.
2010.10.013.
McRoberts, R.E., Walters, B.F., 2012. Statistical inference for remote sensing-based esti-
mates of net deforestation. Remote Sensing of Environment 124 (0), 394 –401.
http://dx.doi.org/10.1016/j.rse.2012.05.011.
Michalak, A.M., Jackson, R., Marland, G., Sabine, C., 2011. A U.S. Carbon Cycle Science Plan.
(Retrieved from) http://www.carboncyclescience.gov/.
Montesano, P.M., Nelson, R., Sun,G., Margolis, H., Kerber, a., Ranson, K.J., 2009.MODIS tree
cover validation for the circumpolar taiga-tundra transition zone. Remote Sensing of
Environment 113 (10), 2130–2141. http://dx.doi.org/10.1016/j.rse.2009.05.021.
Nabuurs, G.J., Masera, O., Andrasko, K., Benitez-Ponce, P., Boer, R., Dutschke, M., Elsiddig,
E., Ford-Robertson, J., Frumhoff, P., Karjalainen, T., Krankina, O., Kurz, W.A.,
Matsumoto, M., Oyhantcabal, W., Ra vindranath, N.H., Sanchez, M.J.S., Zhang, X.,
2007. Forestry. In: Metz, B., Davidson, O.R., Bosch, P.R.,Dave, R., Meyer, L.A. (Eds.), Cli-
mate Change 2007: Mitigation. Contribution of Working Group III to the Fourth As-
sessment Report of the Intergove rnmental Panel on Climate Change. Cambridge
University Press, Cambridge, United Kingdom and New York, NY, USA, pp. 541–586
http://dx.doi.org/10.2753/JES1097-203X330403.
Olofsson, P., Foody, G.M., Herold, M., Stehman, S.V., Woodcock, C.E., Wulder, M.a., 2014.
Good practices for estimating area and assessing accuracy of land change. Remote
Sensing of Environment 148, 42–57. http://dx.doi.org/10.1016/j.rse.2014.02.015.
Olofsson, P., Foody, G.M., Stehman, S.V., Woodcock, C.E., 2013. Making better use of accu-
racy data in land change studies: Estimating accuracy and area and quantifying un-
certainty using stratified estimation. Remote Sensing of Environment 129, 122–131.
http://dx.doi.org/10.1016/j.rse.2012.10.031.
Olson, D.M., Dinerstein, E., Wikram anayake, E.D., Burgess, N.D., Powell, G.V.N.,
Underwood, E.C., D'Amico, J.A., Ito ua, I., Strand, H.E., Morrison, J.C., Loucks, C.J.,
Allnutt, T.F., Ricketts, T.H., Kura, Y., Lamoreux, J.F., Wettengel, W.W., Hedao, P.,
Kassem, K.R., 2001. Terrestrial ecoregions of the world: a new map of life on earth.
Bioscience 51 (11), 933–938 (Retrieved from http://www.bioone.org/doi/abs/
10.1641/0006-3568(2001)051[0933:TEOTWA]2.0.CO;2).
Potapov, P.V., Turubanova, S.a., Hansen, M.C., Adusei, B., Broich, M., Altstatt, A., Mane, L.,
Justice, C.O., 2012 . Quantifying forest cover loss in Democratic Republic of the
Congo, 2000-2010, with Landsat ETM+ data. Remote Sensing of Environment 122,
106–116. http://dx.doi.org/10.1016/j.rse.2011.08.027.
Potapov, P., Turubanova, S., Hansen, M.C., 2011. Regional-scale boreal forest cover and
change mapping usingLandsat data composites for European Russia. Remote Sensing
of Environment 115 (2), 548–561. http://dx.doi.org/10.1016/j.rse.2010.10.001.
Sannier, C., Mcroberts, R.E., Fichet, L.-V., Massard, E., Makaga, K., 2014. Using the regres-
sion estimator with Landsat data to estimate proportion forest coverand net propor-
tion deforestation in Gabon. Remote Sensing of Environment 151, 138–148. http://
dx.doi.org/10.1016/j.rse.2013.09.015.
Särndal, C.-E., Swensson, B., Wretman, J.H.,1992. Model assisted survey sampling.Spring-
er-Verlag New York, New York.
Schepaschenko, D., See, L., Lesiv, M., McCallum, I., Fritz, S., Salk, C.,Moltchanova, E., Perger,
C., Shchepashchenko, M., Shvidenko, A., Kovalevskyi, S., Gilitukha, D., Albrecht, F.,
Kraxner, F., Bun, A., Maksyutov, S., Sokolov, A., Dürauer, M., Obersteiner, M.,
Karminov, V., Ontikov, P., 2015. Development ofa global hybrid forest mask through
the synergy of remote sensing, crowdsourcing and FAO statistics. Remote Sensing of
Environment 162, 208–220. http://dx.doi.org/10.1016/j.rse.2015.02.011.
Schlesinger, W.H., 1997. Biogeochemistry: an analysis of global change. second ed. Aca-
demic Press, San Diego, California, USA.
Sexton, J.O., Song, X.-P., Feng, M., Noojipady, P., Anand, A., Huang, C., Kim, D.-H., Collins,
K.M., Channan, S., Dimiceli, C., Townshend, J.R., 2013a. Global, 30-m resolution con-
tinuous fields of tree cover:Landsat-based rescaling of MODIS Vegetation Continuous
Fields with lidar-based estimates of error. International Journal of Digital Earth 6 (5),
427–448. http://dx.doi.org/10.1080/17538947.2013.786146.
Sexton,J.O., Urban, D.L., Donohue, M.J.,Song, C., 2013b. Long-term landcover dynamics by
multi-temporal classification across the Landsat-5 record. Remote Sensing of Envi-
ronment 128, 246–258. http://dx.doi.org/10.1016/j.rse.2012.10.010.
Sexton, J.O., Noojipady, P., Anand, A. , Song, X.-P., McMahon, S., Huang, C., Feng, M.,
Channan, S., Townshend, J.R., 2015. A mo del for the propagation of uncertainty
from continuous estimates of tree cover to categorical forest cover and change. Re-
mote Sensing of Environment 156, 418–425. http://dx.doi.org/10.1016/j.rse.2014.
08.038.
Sexton, J.O., Noojipady, P., Song, X.-P., Feng, M., Song, D.-X., Kim, D.-H., Anand, A., Huang,
C., Channan, S., Pimm, S.L., Townshend, J.R., 2016. Conservation policy and the mea-
surement of forests. Nature Climate Change 6, 192–196. http://dx.doi.org/10.1038/
nclimate2816.
Shimada, M., Itoh, T., Motooka, T., Watanabe, M., Shiraishi, T., Thapa, R., Lucas, R., 2014.
New global forest/non-forest maps from ALOS PALSAR data (2007-2010). Remote
Sensing of Environment 155, 13–31. http://dx.doi.org/10.1016/j.rse.2014.04.014.
Shvidenko, a., Barber, C.V., Persson, R., Gonzalez, P., Hassan, R.M.,Lakyda, P., McCallum, I.,
Nilsson, S., Pulhin, J., van Rosenburg, B., Scholes, B., 2005. Forest and Woodloand Sys-
tems. Millennium ecosystem assessment series. Ecosystems and human well-being:
Current State and Trends. Findings of the Condition and Trends Working Group. (Mil-
lennium Ecosystem Assessment).
Skole, D., Tucker, C., 1993. Tropical Deforestation and Habitat Fragmentation in the Ama-
zon: Satellite Data from 1978 to 1988. Science 260 (5116), 1905–1910. http://dx.doi.
org/10.1126/science.260.5116.1905.
Skole, D.L., Salas, W.A., Taylor, V., 1998. Global Observations of Forest Cover:Coarse Resolu-
tion Products Design Strategy,Report of a Workshop EROS Data Center. Paris, France.
(Retrieved from) http://www.fao.org/gtos/gofc-gold/docs/GOLD_4.pdf.
Smart, L.S., Swenson, J.J., Christensen, N.L., Sexton, J.O., 2012. Three-dimensional charac-
terization of pine forest type and red-cockaded woodpecker habitat by small-
84 M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
footprint, discrete-return lidar. Forest Ecology and Management 28 1, 100–110.
http://dx.doi.org/10.1016/j.foreco.2012.06.020.
Song, X.-P., Huang, C., Saatchi, S.S., Hansen, M.C., Townshend, J.R., 2015. Annual Carbon
Emissions from Deforestation in the Amazon Basin between 2000 and 2010. PloS
One 10 (5), e0126754. http://dx.doi.org/10.1371/journal.pone.0126754.
Stehman,S.V., Czaplewski, R.L.,1998, Januaryy. Design and Analysisfor Thematic Map Ac-
curacy Assessment - an application of satellite imagery. Remote Sensing of Environ-
ment 64, 331–344. http://dx.doi.org/10.1016/S0034-4257(98)00010-8.
Stehman,S.V., 1999. Basic probability sampling designsfor thematic map accuracyassess-
ment. International Journal of Remote Sensing 20 (12), 2423–2441. http://dx.doi.org/
10.1016/S0034-4257(99)00090-5.
Stehman, S.V., 2000. Practical implications of design-based sampling inference for the-
matic map accuracy ass essment. Remote Sensi ng of Environment 72 (1), 35–45.
http://dx.doi.org/10.1016/S0034-4257(99)00090-5.
Stehman, S.V., 2013. Estimating area from an accuracy assessment error matrix. Remote
Sensing of Environment 132, 202–211. http://dx.doi.org/10.1016/j.rse.2013.01.016.
Stehman, S.V., 2014. Estimating area and map accuracy for stratified random sampling
when the strata are different from the map classes. International Journal of Remote
Sensing 35 (13), 37–41. http://dx.doi.org/10.1080/01431161.2014.930207.
Steininger, M.K., Tucker, C.J., Townshend, J.R.G., Killeen, T.J., Desch, A., Bell, V., Ersts, P.,
2001. Tropical deforestation in the Bolivian Amazon. Environmental Conservation
28 (2), 127–134. http://dx.doi.org/10.1017/S0376892901000133.
TNC, 2012. Terrestrial Ecoregions of the World dataset. (Retrieved February 6, 2015,
from) http://maps.tnc.org/files/metadata/TerrEcos.xml.
Townshend, J.R., Justice, C.O., 1995. Spatial variability of images and the monitoring of
changes in the Normalized Difference Vegetation Index. International Journal of Re-
mote Sensing 16, 2187–2195. http://dx.doi.org/10.1080/01431169508954550.
Townshend, J.R.G., Justice, C.O., 1988. Selecting the spatial resolution of satellite sensors
required for global monitoring of land transformations. International Journal of Re-
mote Sensing 9 (2), 187–236. http://dx.doi.org/10.1080/01431168808954847.
Townshend, J.R., Justice, C.O., Skole, D.L., Belward, A., Janetos, A., Gunawan, I., Goldammer,
J.G., Lee, B., 2004. Meeting the Goals of GOFC. Land Change Science. Vol. 6, pp. 31–55.
http://dx.doi.org/10.1007/978-1-4020-2562-4.
Townshend, J.R., Masek, J.G., Huang, C., Vermote, E.F., Gao, F., Channan, S., Sexton, J.O.,
Feng, M., Narasimhan, R., Kim, D., Song, K., Song, D., Song, X.-P., Noojipady, P., Tan,
B., Hansen, M.C., Li, M., Wolfe, R.E., 2012. Global characterization and monitoring of
forest cover using Landsat data: opportunities and challenges. International Journal
of Digital Earth 5 (5), 373–397. http://dx.doi.org/10.1080/17538947.2012.713190.
Trainor, A.M., Walters, J.R., Morris, W.F., Sexton, J., Moody, A., 2013. Empirical estimation
of dispersal resistance surfaces: A case study with red-cockaded woodpeckers. Land-
scape Ecology 28 (4), 755–767. http://dx.doi.org/10.1007/s10980-013-9861-5.
Tsendbazar, N.E., de Bruin, S., Herold, M., 2014. Assessing global land cover reference
datasets for different user communities. ISPRS Journal of Photogrammetry and Re-
mote Sensing 103, 9 3–114. http://dx.doi.org/10.1016/j.isprsjprs.2014.02.008.
Tucker, C.J., Grant, D.M., Dykstra, J.D., 2004. NASA's global orthorectified Landsat data set.
Photogrammetric Engineeringand Remote Sensing 70 (3), 313–322 (Retr ieved from)
http://cat.inist.fr/?aModele=afficheN&cpsidt=15517253.
UNFCCC, 2002. Report of the Conference of the Parties on the second part of its seventh
session, held at Marrakesh from 29 October to 10 November, addendum, part two:
action taken by the conference of parties. FCCC/CP/2001/13/Add.1. (Retrieved
from) http://unfccc.int/documentation/documents/advanced_search/items/6911.
php?priref=600001855.
Vermote, E.F., Kotchenova, S.Y., 2008. MOD09 (Surface Reflectance) User's Guide. Produc-
tion 09 (Retrieved from http://modis-sr.ltdri.org/products/MOD09_UserGuide_v1_2.
pdf).
Vermote, E.F., Saleous, N.Z.E., Justice, C.O., 2002. Atmosphericcorrection of MODIS data in
the visible to middle infrared: first results. RemoteSensing of Environment 83 (1–2),
97–111. http://dx.doi.org/10.1016/S0034-4257(02)00089-5.
Zhao, Y., Gong, P., Yu, L., Hu, L., Li, X., Li, C., Zhang, H., Zheng, Y., Wang, J., Zhao, Y., Cheng,
Q., Liu, C., Liu, S., Wang, X., 2014. Towards a common validation sample set for global
land-cover mapping. International Journal of Remote Sensing 35 (13), 4795–4814.
http://dx.doi.org/10.1080/01431161.2014.930202.
85M. Feng et al. / Remote Sensing of Environment 184 (2016) 73–85
