ArticlePDF Available

Global Percent Tree Cover at a Spatial Resolution of 500 Meters: First Results of the MODIS Vegetation Continuous Fields Algorithm


Abstract and Figures

The first results of the Moderate Resolution Imaging Spectroradiometer (MODIS) vegetation continuous field algorithm's global percent tree cover are presented. Percent tree cover per 500-m MODIS pixel is estimated using a supervised regression tree algorithm. Data derived from the MODIS visible bands contribute the most to discriminating tree cover. The results show that MODIS data yield greater spatial detail in the characterization of tree cover compared to past efforts using AVHRR data. This finer-scale depiction should allow for using successive tree cover maps in change detection studies at the global scale. Initial validation efforts show a reasonable relationship between the MODIS-estimated tree cover and tree cover from validation sites.
Content may be subject to copyright.
Copyright Ó2003, Paper 7-010; 4,977 words, 9 Figures, 0 Animations, 1 Table.
Global Percent Tree Cover at a
Spatial Resolution of 500 Meters: First
Results of the MODIS Vegetation
Continuous Fields Algorithm
M. C. Hansen*
Department of Geography, University of Maryland, College Park, Maryland
R. S. DeFries
Department of Geography, and Earth System Science Interdisciplinary Center, University
of Maryland, College Park, Maryland
J. R. G. Townshend
Department of Geography, and Institute for Advanced Computer Studies, University of
Maryland, College Park, Maryland
M. Carroll, C. Dimiceli, and R. A. Sohlberg
Department of Geography, University of Maryland, College Park, Maryland
Received 13 February 2003; accepted 5 May 2003
ABSTRACT: The first results of the Moderate Resolution Imaging
Spectroradiometer (MODIS) vegetation continuous field algorithm’s global
percent tree cover are presented. Percent tree cover per 500-m MODIS pixel is
* Corresponding author address: Dr. M. C. Hansen, Department of Geography, University of
Maryland, College Park, College Park, MD 20742.
E-mail address:
Earth Interactions Volume 7 (2003) Paper No. 10 Page 1
estimated using a supervised regression tree algorithm. Data derived from the
MODIS visible bands contribute the most to discriminating tree cover. The
results show that MODIS data yield greater spatial detail in the characterization
of tree cover compared to past efforts using AVHRR data. This finer-scale
depiction should allow for using successive tree cover maps in change detection
studies at the global scale. Initial validation efforts show a reasonable
relationship between the MODIS-estimated tree cover and tree cover from
validation sites.
KEYWORDS: Global land cover, Forest mapping, Regression tree
1. Introduction
Standardized maps of global forest cover serve many purposes, among them the
ability to estimate parameters for use in biogeochemical modeling procedures
(Bonan et al., 2002; DeFries et al., 2002), in delineating remaining intact forest and
woodland tracts for conservation and forestry concerns (Matthews, 2001), and in
monitoring ecological succession and natural processes in forests. Such maps also
reveal land use intensification when compared to potential vegetation conditions,
revealing the human impact on naturally forested ecosystems. Repeated efforts
over time can document change and aid in predicting future alterations to forest
ecosystems. The synoptic view of global satellite datasets affords the best
possibility of creating such maps. Initial efforts (DeFries et al. 1999; DeFries et al.
2000; Zhu and Waller 2001; Hansen et al. 2002) have demonstrated this capability.
This paper describes the creation of a new global percent tree cover map based
on 500-m data from the Moderate Resolution Imaging Spectroradiometer
(MODIS) instrument on board the National Aeronautics and Space Adminis-
tration’s (NASA’s) Terra spacecraft and represents the finest-scale global forest
information to date. The MODIS sensor represents a significant gain in spatial
detail due primarily to three facts. The first is the finer instantaneous field of view
of MODIS (250 and 500 m
) as compared to heritage Advanced Very High
Resolution Radiometer (AVHRR) instruments (1 km
). Second, due to the fact that
MODIS was built with seven bands specifically designed for land cover
monitoring, there is an improved spectral/spatial response compared to AVHRR.
This allows for greater accuracy in mapping due to more robust spectral signatures.
It also aids in reducing background scattering from adjacent pixels, as the MODIS
land bands were designed to limit the impact of atmospheric scattering. Third, 500-
m red and near-infrared data, two bands important for land cover mapping, are
created from averaged 250-m imagery. This resampling also reduces the percent
contribution of adjacency effects on 500-m pixels for these bands, allowing for
improved land cover estimates (Townshend et al., 2000). The result is a dataset that
reveals far more spatial detail than previous efforts. Maps such as the MODIS
global continuous field of percent tree cover map should be of use to more varied
scientific applications than previous coarse-scale maps.
Proportional per pixel tree cover estimates, or continuous fields of percent tree
cover, are an improved thematic representation over discrete classifications
(DeFries et al., 1995). Continuous field maps yield improved depictions of spatially
Earth Interactions Volume 7 (2003) Paper No. 10 Page 2
complex landscapes and the ability to use successive depictions to measure change
(Hansen and DeFries, 2003). Numerous methodologies exist to portray subpixel
vegetative cover. The techniques including fuzzy estimations of forest cover
(Foody and Cox, 1994), plant density isolines within multispectral scatterplots
(Jasinski, 1996; Zhu and Waller, 2001), empirically calibrated estimates using
multiresolution datasets (Zhu and Evans, 1994; Iverson, 1989; DeFries et al.,
1997), other multiresolution estimates that incorporate spatial arrangement
(Mayaux and Lambin, 1997), and endmember linear mixture modeling (DeFries
et al., 2000; Adams et al., 1995; Settle and Drake, 1993). This paper builds on prior
studies using AVHRR data to derive a global MODIS 500-m percent tree cover
map. The approach is an empirical, multiresolution calibration method that uses a
regression tree algorithm to estimate the percent tree canopy cover (Hansen et al.,
2002) The regression tree is a nonlinear, flexible model appropriate for handling
the variability present in global vegetation phenology. It also allows for the
calibration of the model along the entire continuum of tree cover, avoiding the
problems of using only endmembers for calibration.
2. Data
This initial attempt using MODIS imagery employed approximately 1 year of data.
The inputs consisted of 8-day minimum blue reflectance composites that were
made in order to reduce the presence of clouds in the datastream. However, this
procedure can lead to the inclusion of pixels within areas of cloud shadow. To
reduce the presence of cloud shadows, the data were converted to 40-day
composites using a second darkest albedo (sum of blue, green, and red bands)
algorithm. The inputs date from 31 October 2000 to 9 December 2001. An extra
40-day composite period was added to attempt to compensate for data gaps
resulting from temporary sensor outages.
The seven MODIS land bands were used as inputs: blue (459–479 nm), green
(545–565 nm), red (620–670 nm), near infrared (841–876 nm), and midinfrared
(1230–1250, 1628–1652, 2105–2155 nm). The MODIS-composited data were
transformed into annual metrics that capture the salient points in the phenologic
cycle. A total of 68 metrics were derived from the composited data for bands 1–7
and the Normalized Difference Vegetation Index (NDVI). These were used as the
inputs for estimating percent tree cover. Metrics such as maximum annual NDVI or
mean growing season red reflectance represent generic signatures that can be used
to map global vegetation. The approach to deriving the metrics and training data is
fully described in Hansen et al. (Hansen et al., 2002).
Only the red and near-infrared MODIS bands are close to the bandwidths in the
long-term AVHRR sensor’s record. In addition to these two bands, AVHRR has
one midinfrared and two thermal bands that record brightness temperatures and
proved invaluable to mapping global land cover (Hansen et al., 2000). The
temperature bands act as surrogates for biome-level climatic variability. For
example, tropical drought deciduous woodlands can be stratified from tropical
humid forests using thermal brightness metrics. There is less evapotranspiration
during dry periods in seasonal woodlands, and this causes an increase in surface
temperature that is captured in the thermal bands.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 3
While the MODIS sensor has bands for measuring surface temperature [band 31
(10780–11280 nm) and band 32 (11770–12270 nm)], they are not currently
processed for use in land cover mapping. There is a surface temperature product
(Wan et al., 2002) that employs these bands, but its algorithms are land cover
dependent, precluding its use in mapping surface cover. Bands 31 and 32 of the
MODIS instrument are used to derive surface temperature. They mimic the
AVHRR thermal bands, and their inclusion in future reprocessing of the land
products is recommended. In place of the missing MODIS thermal data, which act
as a key regional stratification signal (Hansen et al., 2002), other features were
included. First, a three-region layer was included as a metric: extratropical north
(approximately 238N and above), tropical (approximately between plus and minus
238latitude), and extratropical south (approximately 238S and below). Second, an
archival 1-km AVHRR channel-4 brightness temperature (10300–11300 nm)
signal was used in metric form. These data are from 1995 to 1996 and represent the
most recent globally processed thermal images at 1 km for the AVHRR. The
thermal information is used within the algorithm to regionally stratify the globe, as
previously stated, and should not be significantly affected by land cover change
events since the time of the data’s acquisition. The thermal data exist at a 1-km
spatial resolution, and any land cover change in the interim would have to be very
extensive to impact the thermal signal. Even in these instances, the MODIS data at
a finer scale correspond to the detail in the training data, and these data should
drive most of the characterization. The AVHRR metrics and regional layers are
used alongside the MODIS data inputs.
The training data are derived by aggregating over 250 classified high-resolution
Landsat images to the MODIS grid. The Landsat images were classified into four
classes of tree cover, each class having a mean percent tree cover label. By
averaging the Landsat tree cover strata to MODIS cells, a 500-m continuous
training dataset was created. This training dataset contains over a million pixels,
which were systematically sampled to create a final training dataset of 271,149
pixels at the 500-m MODIS resolution. These training data have been used in a
number of global land cover mapping exercises and descriptions of their derivation
and distribution can be found in previous refereed studies (DeFries et al., 1998;
Hansen et al., 2000; Hansen et al., 2002).
3. Methods
The MODIS continuous fields of vegetation cover algorithm is described in
Hansen et al. (Hansen et al., 2002). It is an automated procedure that employs a
regression tree algorithm (Venables and Ripley, 1994). The regression tree is a
nonlinear, distribution-free algorithm that is highly suited for handling the
complexity of global spectral land cover signatures. The training data are used as
the dependent variable, predicted by the independent variables in the form of the
annual MODIS metrics. Outputs from the regression tree are further modified by a
stepwise regression and bias adjustment per Hansen et al. (Hansen et al., 2002).
The derivation of tree cover in this way creates the possibility of using subsequent
depictions to measure change. Hansen and DeFries (Hansen and DeFries, 2003)
Earth Interactions Volume 7 (2003) Paper No. 10 Page 4
Figure 1. Plot of validation field data from four test areas, each covering a wide
range of tree cover density, where y¼0.79xand R
¼0.95. See Hansen
et al. (Hansen et al., 2002a) for example from Zambia test area.
Figure 2. Final percent tree cover map in the Interrupted Goode Homolosine
Earth Interactions Volume 7 (2003) Paper No. 10 Page 5
used such an approach in detecting change based on the long-term AVHRR 8-km
Pathfinder dataset.
The output of the algorithm is the percent canopy cover per 500-m MODIS
pixel. Here percent canopy refers to the amount of skylight obstructed by tree
canopies equal to or greater than 5 m in height and is different than percent crown
cover (crown cover ¼canopy cover þwithin crown skylight). The canopy cover
definition is used in vegetation modeling exercises in which light availability is an
important parameter. Foresters, on the other hand, largely employ crown cover in
measuring forest density. Crown cover is a better measure when performing areal
inventories and is the variable used in many forest accounting procedures.
To better understand this relationship, ongoing field work is being performed
where both crown and canopy cover values are measured. Initial work suggests that
the mean forest label used in deriving canopy cover (80%) training data
corresponds to a 100% forested area in terms of crown cover. Figure 1 shows field
data gathered from four different sites to test this assumption and shows a
Table 1. Percent contribution to the overall reduction of sum of squares in the
regression tree structure aggregated in the following ways: (a) regression tree
splits aggregated by the band used in the metrics with total number of splits in
parentheses, (b) splits aggregated by individual metric with total number of
splits in parentheses and only the best 10 metrics shown, (c) metrics used in the
10 best individual splits in reducing overall sum of squares. These splits are
highlighted in the tree structure of Figure 4. For mean metrics the span of time is
listed by number of 40-day composites. For MODIS bands 1–7, metrics are
derived by looking at dark albedo values. For example, mean 1–3 band 6
represents the mean of the three darkest band 6 composite values. NDVI and
temperature means are based on finding the maximum ranked composites. For
example, mean 1–5 NDVI is the five highest NDVI composites averaged. The
only exceptions to this are the metrics binned using NDVI as a reference. For
example, in mean 1–3g band 1, the ‘‘g’’ indicates that the band 1 values are
found which correspond to the three greenest composites based on ranked
NDVI values. Thus, mean 1–3g band 1 is a mean red reflectance value, which
corresponds to the three greenest composites.
Percent reduction in overall sum
of squares per metrics
aggregated by band
Percent reduction in overall sum
of squares per individual metric
Percent reduction in overall sum
of squares per regression tree split
Band 1 metrics (9) 68.1 Mean 1–3g band 1 (4) 67.7 Mean 1–3g band 1 60.1
Band 3 metrics (21) 9.6 Mean 1–3g band 3 (2) 6.7 Mean 1–3g band 1 7.3
Band 4 metrics (20) 5.4 Mean 1–3 temperature (6) 4.4 Mean 1–3g band 3 6.7
Temperature metrics (13) 5.1 Mean 1–3 band 6 (3) 3.6 Mean 1–3 band 6 3.3
NDVI metrics (14) 4.5 Rank 3 band 3 (16) 2.7 Mean 1–3 temperature 2.4
Band 6 metrics (5) 3.8 Mean 1–3 band 4 (5) 2.7 Mean 1–3 temperature 1.8
Band 5 metrics (10) 1.3 Mean 1–8 NDVI (3) 2.2 Mean 1–3 band 4 1.4
Regions (7) 1.0 Mean 1–8 band 4 (5) 1.6 Mean 1–8 NDVI 1.2
Band 2 metrics (5) 1.0 Mean 1–5 NDVI (6) 1.3 Mean 1–5 NDVI 1.1
Band 7 metrics (4) 0.2 Regions (7) 1.0 Mean 1–8 NDVI 0.9
Earth Interactions Volume 7 (2003) Paper No. 10 Page 6
reasonable relationship. Of course, different tree types have different relationships
between canopy and crown cover. Fir trees, for example, generally have little light
availability within the canopy. Four subalpine fir sites from Colorado reveal a 0.9
ratio of canopy to crown cover. Broadleaf Kalahari woodland trees in western
Zambia, on the other hand, have a greater presence within crown gaps and a 0.76
ratio. Information on stand species is not available at the global scale; however, the
0.8 slope in Figure 1 suggests a reasonable estimate for converting between canopy
and crown cover. It is suggested that users interested in deriving the crown cover
variable should divide the canopy cover layer by 0.8.
4. Results
The resulting regression tree yielded 109 terminal nodes. The largest node in terms
of surface area maps most of the tropical broadleaf evergreen forest and accounts
Figure 3. Plot of band-1 metric (mean red reflectance of the three greenest
compositing periods) vs estimated tree cover for a sample of 1000
Earth Interactions Volume 7 (2003) Paper No. 10 Page 7
for over 20% of dense (.40%) tree cover globally. This is a fairly homogeneous
cover type with a characteristic signature. The regression tree delineated two
subclasses of this forest type that represent more confused spectral signatures:
persistently cloudy areas and areas of regrowth/disturbance. The next largest node
in terms of tree cover maps dense needleleaf boreal forest. The final map is shown
in Figure 2.
The regression tree object can be studied to reveal which spectral information
drives the tree cover characterization. Table 1 shows which metrics add the most to
reducing the overall sum of squares in delineating tree cover strata. From Table 1,
note that the mean red reflectance corresponding to the three greenest composite
periods’ metric contributes most to mapping tree cover. Of the overall reduction in
the sum of squares, this metric alone contributes nearly 70% of the reduction. The
first split in the regression tree uses this metric, and this single split accounts for
60% of the reduction in the sum of squares. This metric is plotted against the
resulting estimated tree cover for a 1000-pixel sample in Figure 3. While this
metric alone cannot map global tree cover, it is clear that increasing canopy density
is correlated with lower red reflectance values due to the combined effects of
canopy shadowing and chlorophyll absorption.
Of interest is the fact that the MODIS visible bands (red, green, and blue) all
contribute significantly, while the near- and midinfrared bands largely do not. Only
band 6 performs comparably with the visible bands and is critical to mapping
inundated grasslands. This is a midinfrared band with strong water absorption
qualities that capture seasonal flooding events. Figure 4 shows the top levels of the
regression tree and how band 6 is used to map these grasslands. Although the
Figure 4. Regression tree object before stepwise regression and bias adjustment
steps. Estimated percent tree cover at the intermediate nodes is shown
within the ellipses. The metric used at each split is shown below each of
these nodes. Only one terminal node is shown as a rectangle; the tilde
(;) indicates more splits in the lower portion of the regression tree.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 8
infrared bands do not feature prominently, NDVI, derived using the near infrared,
is useful as seen in Table 1. It should also be noted that many of the most used
metrics of the visible bands are binned using NDVI to identify the greenest times
of the year.
The thermal signal of the AVHRR was used repeatedly, as seen in Table 1, and
underscores the need to include the MODIS thermal signal in the gridded land
products. The regional stratification was not as useful, only accounting for 1% of
Figure 5. Two subsets comparing the 2000–01 500-m MODIS tree cover map with
a 1995–96 1-km AVHRR tree cover map. (a) Area of Rondonia, Brazil,
from the AVHRR map, (b) same area from the 2000–01 MODIS map, (c)
area along the French–German border from the AVHRR map, and (d)
same area from the MODIS map.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 9
the overall reduction in the sum of squares. Two kinds of metrics were of little use:
amplitude metrics for measuring the absolute spectral change of cover through the
growing season and metrics associated with single-peak greenness dates, which
were largely unused by the regression tree.
The map has greatly increased spatial detail as compared to the AVHRR-derived
maps. Figure 5 shows two areas as examples. The human imprint on the landscape
is more readily seen as compared to the AVHRR example. There is the reasonable
expectation that consecutive comparisons of annual maps should reveal change.
Discrete breaks in tree cover due to administrative status, such as a national park,
government-owned lands, and transnational variations in land use intensity, are
clearly evident throughout the map. Figure 6 shows a region of southeastern Africa
where differential land use intensification is visible across national boundaries. The
rich detail present should be of use to land managers working at a regional scale
and in need of an internally consistent map. Fire history is present as well,
particularly in the boreal zone, as shown in Figure 7 by the number of quasi-
elliptical patterns that correspond to known fire scars. Further analysis of these data
should reveal if this kind of map can be used to determine likely succession
patterns, especially when other vegetation continuous field layers, such as leaf
type, are generated.
The map will be updated annually and used to monitor change in global tree
cover. Figure 8 shows the MODIS percent tree cover map with an overlay of a
change study using AVHRR data for 1982–99 (Hansen and DeFries, 2003).
Figure 6. Transnational boundary differences in percent tree cover. The highest
population density in this subset is found in Malawi, which is shown to
have greater clearing of tree cover than adjacent countries. Mozam-
bique is less disturbed as evidenced by the Tete Province jutting into
the more intensively used landscapes of Zimbabwe, Zambia, and
Malawi. The arm of the Congo extending into Zambia’s copper belt is
less disturbed than the heavily developed lands across the border.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 10
MODIS data from the 250- and 500-m bands should capture change in forest cover
more accurately. Optimum change study intervals, whether annual or 5–10-yr
epochs, will be sought. Improvements to the methodology, such as the inclusion of
MODIS thermal bands, will be implemented as soon as is feasible.
5. Validation
A multiresolution mapping approach in conjunction with field data is being used at
a number of sites to develop validation data for the percent tree cover map. The
exercise includes using field data along with IKONOS and Enhanced Thematic
Mapper Plus (ETMþ) data to create validation test areas the size of an ETMþ
image. Crown cover maps of IKONOS images are binned to ETMþcells and used
as continuous training data to map the percent crown cover for 30-m pixels. This
ETMþcrown cover map is then averaged to a 500-m resolution to validate the
MODIS map. Performing this exercise in a wide variety of biomes will help to
create a test bed against which successive iterations of the tree cover product can be
Figure 7. MODIS data with burn scar overlay. Black vectors represent burn scars
from 1990–2000, blue 1980–89, red 1970–79, magenta 1960–69, and
orange 1950–59. More recent scars are fairly well delineated in the tree
cover percentage map. Data are from Murphy et al. (Murphy et al.,
2000) and consist of a combination of ground and aerial surveys and
satellite image intepretations.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 11
Figure 8. A portion of South American tree cover with deforestation hot spot
overlay. The change areas are from a 19-yr study of 8-km AVHRR data
(Hansen and DeFries, 2003).
Earth Interactions Volume 7 (2003) Paper No. 10 Page 12
validated. The method has been initially tested for a Western Province, Zambia,
woodland site (Hansen et al., 2002) and is now being used in other areas.
Figure 9 shows results from a Colorado test area. Averaging the product
dramatically improves the validation measures. The greater scatter at 500-m spatial
resolution is probably an artifact of resampling in the MODIS data. As with all
global data processing, a nearest-neighbor scheme is used to reduce processing
time. This approach leads to a geometric degradation of the signal as the process is
repeated throughout the compositing process. Averaging the product to a 1-km
spatial resolution appears to ameliorate some of these effects.
6. Conclusions
The first layer of the MODIS Vegetation Continuous Field product, percent tree
canopy cover, has been generated and is available for use ( The MODIS product in
tile format for canopy cover is available online at the EROS data center http://
Figure 9. Validation data from CO WRS 035/032. (a) Plot of 500-m MODIS-
estimated percent tree crown cover vs 500-m crown cover validation
data derived from IKONOS/Enhanced Thematic Mapper Plus/field
data study, y¼0.99, R
¼0.81. (b) Data averaged to a 1-km resolution,
¼0.89. (c) Data averaged to a 2-km spatial resolution, y¼
¼0.94. (d) Validation data percent tree crown cover. (e)
MODIS-estimated percent tree crown cover at 500-m resolution.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 13 and per continent at glcf.umiacs.umd.
edu). The map reveals the improved spatial/spectral characteristics in the MODIS
data compared to heritage AVHRR data. This should lead to a wider variety of
applications that employ the MODIS-derived maps. Visible bands in the MODIS
data provided the most discrimination along with NDVI and AVHRR brightness
temperatures. This points out the need to add thermal information to the MODIS
land datastream. Currently in production are other vegetation layers, including the
percentages of herbaceous/shrub, bare ground, tree leaf type, and leaf longevity.
Upon completion, these maps should enhance the current understanding of global
land cover distributions and provide a basis for monitoring land cover change
Adams, J. B., D. E. Sabol, V. Kapos, R. A. Filho, D. A. Roberts, M. O. Smith, and A. R.
Gillespie, 1995: Classification of multispectral images based on fraction endmembers:
Application to land-cover change in the Brazilian Amazon. Remote Sens. Environ., 52,
Bonan, G. B., S. Levis, L. Kergoat, and K. W. Oleson, 2002: Landscapes as patches of plant
functional types: An intergrating concept for climate and ecosystem models. Global
Biogeochem. Cycles, 16, 1360–1384.
DeFries, R. S., et al., 1995: Mapping the land surface for global atmosphere–biosphere models:
Towards continuous distributions of vegetation’s functional properties. J. Geophys. Res.,
100, 20,867–20,882.
DeFries, R. S., M. C. Hansen, J. R. G. Townshend, and R. S. Sohlberg, 1998: Global land
cover classifications at 8km spatial resolution: The use of training data derived from
Landsat imagery in decision tree classifiers. Int. J. Remote Sens., 19, 3141–3168.
DeFries, R. S., J. R. G. Townshend, and M. C. Hansen, 1999: Continuous fields of vegetation
characteristics at the global scale at 1-km resolution. J. Geophys. Res., 104, 16,911–
DeFries, R. S., M. C. Hansen, J. R. G. Townshend, and A. C. Janetos, 2000: A new global 1-
km dataset for percentage tree cover derived from remote sensing. Global Change Biol.,
6, 247–254.
DeFries, R. S., R. A. Houghton, M. C. Hansen, C. B. Field, D. Skole, and J. Townshend, 2002:
Carbon emissions from tropical deforestation and regrowth based on satellite
observations for the 1980’s and 1990’s. Proc. Natl. Acad. Sci., 99, 14,256–14,261.
Foody, G., and D. Cox, 1994: Sub-pixel land cover composition estimation using a linear
mixture model and fuzzy membership functions. Int. J. Remote Sens., 15, 619–631.
Hansen, M. C., and R. S. DeFries, 2003: Detecting long term global forest change using
continuous fields of tree cover maps from 8km AVHRR data for the years 1982–1999.
Ecosystems, in press.
Hansen, M. C., R. S. DeFries, J. R. G. Townshend, and R. Sohlberg, 2000: Global land cover
classification at 1km spatial resolution using a classification tree approach. Int. J. Remote
Sens., 21, 1331–1364.
Hansen, M. C., R. S. DeFries, J. R. G. Townshend, L. Marufu, and R. Sohlberg, 2002a:
Development of a MODIS tree cover validation data set for Western Province, Zambia.
Remote Sens. Environ., 83, 320–335.
Hansen, M. C., R. S. DeFries, J. R. G. Townshend, R. A. Sohlberg, C. Dimiceli, and M.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 14
Carroll, 2002b: Towards an operational MODIS continuous field of percent tree cover
algorithm: Examples using AVHRR and MODIS data. Remote Sens. Environ., 83, 303–
Jasinski, M. F., 1996: Estimation of subpixel vegetation density of natural regions using
satellite multispectral imagery. IEEE Trans. Geosci. Remote Sens., 34, 804–813.
Matthews, E., 2001: Understanding the FRA. World Resources Institute Forest Briefing No. 1,
World Resources Institute, Washington, D.C., XX pp.
Mayaux, P., and E. F. Lambin, 1997: Tropical forest area measured from global land cover
classifications: Inverse calibration models based on spatial textures. Remote Sens.
Environ., 59, 29–43.
Murphy, P. J., J. P. Mudd, B. J. Stocks, E. S. Kasischke, D. Barry, M. E. Alexander, and N. H.
F. French, 2000: Historical fire records in the North American boreal forest, in Fire
Climate Change, and Carbon Cycling in the Boreal Forest, edited by E. Kasischke and
B. Stocks, Springer-Verlag, New York, 274–288.
Settle, J., and N. A. Drake, 1993: Linear mixing and the estimation of ground cover
proportions. Int. J. Remote Sens., 14, 1159–1177.
Townshend, J. R. G., C. Huang, S. N. V. Kalluri, R. S. Defries, S. Liang, and K. Yang, 2000:
Beware of per-pixel characterization of land cover. Int. J. Remote Sens., 21, 839–843.
Venables, W. N., and B. D. Ripley, 1994: Modern Applied Statistics with S-Plus. Springer-
Verlag, New York, 462 pp.
Wan, Z., Y. Zhang, Q. Zhang, and Z. Li, 2002: Validation of the land-surface temperature
products retrieved from Terra Moderate Resolution Imaging Spectroradiometer data.
Remote Sens. Environ., 83, 163–180.
Zhu, Z., and D. L. Evans, 1994: U.S. forest typesand predicted percent forest cover from
AVHRR data. Photogramm. Eng. Remote Sens., 60, 525–531.
Zhu, Z., and E. Waller, 2001: Global forest cover mapping for the United Nations Food and
Agriculture Organization Forest Resources Assessment 2000 Program. UNFAO, Rome,
Italy, XX pp.
Earth Interactions is published jointly by the American Meteorological Society, the American Geophysical
Union, and the Association of American Geographers. Permission to use figures, tables, and brief excerpts
from this journal in scientific and educational works it hereby granted provided that the source is
acknowledged. Any use of material in this journal that is determined to be ‘‘ fair use’’ under Section 107 or that
satisfies the conditions specified in Section 108 of the U.S. Copyright Law (17 USC, as revised by P.IL. 94-
553) does not require the publishers’ permission. For permission for any other form of copying, contact one of
the copublishing societies.
Earth Interactions Volume 7 (2003) Paper No. 10 Page 15
... Table A1 in the Appendix summarizes the characteristics and sources of the datasets used. For more details about VCF and the spatial methods see Section A.1 in the Appendix, DiMiceli et al. (2021); Townshend et al. (2017), and Hansen et al. (2003). ...
... VCF-TC is better aligned to our research objectives because it allows assesing forest change as a continuous process at the pixel level (DiMiceli et al., 2021; Ryan et al., 2017). Our outcome variable of interest is the percent tree cover (VCF-TC), defined as the "amount of skylight obstructed by tree canopies equal to or greater than 5 m in height" (Hansen et al., 2003). We note that this differs from crown cover, "the amount of the ground which is encompassed by the tree's crown regardless of whether light penetrates." ...
Full-text available
More than one billion people worldwide receive cash or in-kind transfers from social protection programs. In low-income countries, these transfers are often conditioned on participation in labor-intensive public works to rehabilitate local infrastructure or natural resources. Despite their popularity, the environmental impacts of public works programs remain largely undocumented. We quantify the impact on tree cover of Ethiopia’s Productive Safety Net Program (PSNP), one of the world’s largest and longest-running public works programs, using satellite-based data of tree cover combined with difference-in-differences and inverse probability treatment weighting methodologies. We find that the PSNP increased tree cover by 3.8% between 2005 and 2019, with larger increases in less densely populated areas and on steep-sloped terrain. As increasing tree cover is considered an important strategy to mitigate global warming, our results suggest a win–win potential for social safety net programs with an environmental component.
... The MLCT Collection 6 was generated using a RF supervised classification algorithm within the R environment and a hierarchical classification model. The 500 m MODIS NBAR MC43A4 and MC43A2, the terrestrial ecozones of the Earth (Olson et al., 2001), MODIS water land mask (MCD44W) (Carroll et al., 2009), MODIS VCF (MCD44B) (Hansen et al., 2003), and a global cropland intensity map (Ramankutty et al., 2008), were the main inputs to generate the MLCT Collection 6 product. This product achieved an overall classification accuracy of 73.6%. ...
... The first global percent forest cover map (2000)(2001) was presented by Hansen et al. (2003) at a spatial resolution of 500 m and was derived from the MODIS sensor on board NASA's Terra spacecraft. The methodology was based on using a nonlinear supervised regression tree technique and multi-band MODIS data starting from October 2000 and December 2001. ...
Full-text available
Changes in global land cover (LC) have significant consequences for global environmental change, impacting the sustainability of biogeochemical cycles, ecosystem services, biodiversity, and food security. Different forms of LC change have taken place across the world in recent decades due to a combination of natural and anthropogenic drivers, however, the types of change and rates of change have traditionally been hard to quantify. This thesis exploits the properties of the recently released ESA-CCI-LC product – an internally consistent, high-resolution annual time-series of global LC extending from 1992 to 2018. Specifically, this thesis uses a combination of trajectories and transition maps to quantify LC changes over time at national, continental and global scales, in order to develop a deeper understanding of what, where and when significant changes in LC have taken place and relates these to natural and anthropogenic drivers. This thesis presents three analytical chapters that contribute to achieving the objectives and the overarching aim of the thesis. The first analytical chapter initially focuses on the Nile Delta region of Egypt, one of the most densely populated and rapidly urbanising regions globally, to quantify historic rates of urbanisation across the fertile agricultural land, before modelling a series of alternative futures in which these lands are largely protected from future urban expansion. The results show that 74,600 hectares of fertile agricultural land in the Nile Delta (Old Lands) was lost to urban expansion between 1992 and 2015. Furthermore, a scenario that encouraged urban expansion into the desert and adjacent to areas of existing high population density could be achieved, hence preserving large areas of fertile agricultural land within the Nile Delta. The second analytical chapter goes on to examine LC changes across sub-Saharan Africa (SSA), a complex and diverse environment, through the joint lenses of political regions and ecoregions, differentiating between natural and anthropogenic signals of change and relating to likely drivers. The results reveal key LC change processes at a range of spatial scales, and identify hotspots of LC change. The major five key LC change processes were: (i) “gain of dry forests” covered the largest extent and was distributed across the whole of SSA; (ii) “greening of deserts” found adjacent to desert areas (e.g., the Sahel belt); (iii) “loss of tree-dominated savanna” extending mainly across South-eastern Africa; (iv) “loss of shrub-dominated savanna” stretching across West Africa, and “loss of tropical rainforests” unexpectedly covering the smallest extent, mainly in the DRC, West Africa and Madagascar. The final analytical chapter considers LC change at the global scale, providing a comprehensive assessment of LC gains and losses, trajectories and transitions, including a complete assessment of associated uncertainties. This chapter highlights variability between continents and identifies locations of high LC dynamism, recognising global hotspots for sustainability challenges. At the national scale, the chapter identifies the top 10 countries with the largest percentages of forest loss and urban expansion globally. The results show that the majority of these countries have stabilised their forest losses, however, urban expansion was consistently on the rise in all countries. The thesis concludes with recommendations for future research as global LC products become more refined (spatially, temporally and thematically) allowing deeper insights into the causes and consequences of global LC change to be determined.
... The Bare-Earth SRTM Terrain DEM (BEST DEM) was computed by O'Loughlin et al., 2016 [2] from the SRTM4 DEM. VB was eliminated using canopy heights from the 30 Global Tree Height Map [6], canopy densities from the 250 m mesh size MODIS Vegetation Continuous Field (VCF) product [8], and climate and vegetation type classification maps and the ICESat ground elevation dataset. VB was defined as a function of canopy density multiplied by canopy height. ...
... The estimated canopy penetration depth for the C-band Shuttle Radar Topography Mission (SRTM) signal was 50%. Carabajal and Harding 2006 [13] studied the penetration of C-band SRTM pulses into vegetation canopies using ICESat observations in several regions (Australia, Amazon, Africa, United States, and Asia) and the canopy density from Landsat VCF product [8]. Several relevant conclusions were drawn. ...
Full-text available
A large percentage of the Costa Rican territory is covered with high evergreen forests. In order to compute a 1″ Bare-Earth Digital Terrain Model (DTM) for Costa Rica CRDTM2020, stochastic Vegetation Bias (VB) was reduced from the 1″ NASADEM, Digital Elevation Model (DEM) based on the Shuttle Radar Topography Mission (SRTM) data. Several global models such as: canopy heights from the Global Forest Canopy Height 2019 model, canopy heights for the year 2000 from the Forest Canopy Height Map, and canopy density from the Global Forest Change model 2000 to 2019, were used to represent the vegetation in the year of SRTM data collection. Four analytical VB models based on canopy heights and canopy density were evaluated and validated using bare-earth observations and canopy heights from the Laser Vegetation Imaging Sensor (LVIS) surveys from 1998, 2005, and 2019 and a levelling dataset. The results show that differences between CRDTM2020 and bare-earth elevations from LVIS2019 in terms of the mean, median, standard deviation, and median absolute difference (0.9, 0.8, 7.9 and 3.7 m, respectively) are smaller than for any other of the nine evaluated global DEMs.
... The MOD44B MODIS VCF Coll. 3-Tree Cover (Hansen et al. 2003) data were used for mapping tree cover, whereas MOD44B MODIS VCF-Collection 5 data were used as an improvement (DiMiceli et al. 2011;Zomer et al. 2016). ...
Full-text available
This paper assesses the remote sensing datasets of biomass carbon on the agricultural landscape and their decadal change dynamics due to variation in tree cover dominance using geospatial technology in India. Remote sensing data showed that in the year 2000, 16.9% of all agricultural land (28.02 million hectares) in India had agroforestry land (at least 10% tree cover) which was further increased to 22.5% (37.30 million hectares) over 10 years (up to the year 2010). The total biomass carbon estimate in the year 2000 was found 1868.75 million tons of carbon (≈1.87 Pg C) over the Indian agriculture landscape (≈166 million hectares). Out of which approximately 1039 million tons (≈1.04 Pg C) of biomass carbon come from trees (with 55.7% contribution). Total biomass carbon loss between the periods of 2000 and 2010 was observed 31.19 million tons, whereas gain was 170.02 million tons. The decadal biomass carbon net gain was 138.83 million tons due to an increase in agroforestry land by 5.6% (9.27 million hectares). The mean biomass carbon in India increased from 11.29 to 12.13 t C ha −1 in 10 years, whereas the global mean increment is 20.4 to 21.4 t C ha −1 during the same base periods (Zomer et al in Sci Rep 6:29987, https:// doi. org/ 10. 1038/ srep2 9987, 2016). Our analysis critically addressed one of the past research gaps of the biomass carbon-related findings in the agriculture landscape due to tree cover variation. Such understanding will assist significantly agroforestry decision-makers of India in enhancing future harmonized blueprint for agroforestry.
... shrubs, crops and other herbaceous vegetation) were extracted from the MODIS Vegetation Continuous Fields (MOD44B.003; Hansen et al., 2003) to study the relationship of the vegetation stability metrics with vegetation cover. This dataset was subsequently resampled to the GIMMS grid by assigning to each GIMMS pixel the average fraction of all MODIS pixels situated within it. ...
... The Vegetation Continuous Fields (VCF) approach provides per-pixel continuous fractional cover of the major land cover types, namely woody vegetation, herbaceous vegetation, and nonvegetated land cover. The VCF products were created at global and national scales using Moderate Resolution Imaging Spectroradiometer (MODIS) (Hansen et al., 2003;DiMiceli et al., 2021) and Landsat data (Hansen et al., 2011). The VCF approach provides rich information on vegetation structure for various applications and can be transformed into a custom LCLU classification scheme (DiMiceli et al., 2021). ...
Full-text available
Recent advances in Landsat archive data processing and characterization enhanced our capacity to map land cover and land use globally with higher precision, temporal frequency, and thematic detail. Here, we present the first results from a project aimed at annual multidecadal land monitoring providing critical information for tracking global progress towards sustainable development. The global 30-m spatial resolution dataset quantifies changes in forest extent and height, cropland, built-up lands, surface water, and perennial snow and ice extent from the year 2000 to 2020. Landsat Analysis Ready Data served as an input for land cover and use mapping. Each thematic product was independently derived using locally and regionally calibrated machine learning tools. Thematic maps validation using a statistical sample of reference data confirmed their high accuracy (user’s and producer’s accuracies above 85% for all land cover and land use themes, except for built-up lands). Our results revealed dramatic changes in global land cover and land use over the past 20 years. The bitemporal dataset is publicly available and serves as a first input for the global land monitoring system.
Exposure to residential greenness might affect population health through increasing physical activity and social engagement, improving mental health, and reducing harmful environmental exposure. However, evidence on the association of greenness with risk of diabetes is still controversial. In this study, we recruited a total of 22,535 participants aged ≥18 years from Yinzhou District, Ningbo, Zhejiang Province, China to investigate the associations between residential greenness and risk of diabetes incidence. Residential greenness was estimated using Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Vegetation Continuous Field (VCF). We also calculated cumulative average NDVI, EVI and VCF values, and changes in NDVI, EVI and VCF during the follow-up period. We used Cox proportional hazards models controlling for demographic characteristics, lifestyles, individual socioeconomic status, history of diseases and particulate matter with an aerodynamic diameter ≤2.5 μm (PM2.5) to examine hazard ratios (HRs) and 95% confidence intervals (95% CIs) and assessed physical activity, body mass index (BMI) or PM2.5 as potential mediators. During 84,992.64 person-years of follow-up, a total of 1,154 incident cases of diabetes occurred. In multivariable models, living in the highest quartile of cumulative average NDVI, EVI and VCF within 250-m buffer was associated with 57% (HR = 0.43, 95% CI: 0.36, 0.52), 62% (HR = 0.38, 95% CI: 0.32, 0.45), and 55% (HR = 0.45, 95% CI: 0.38, 0.54) reduction in diabetes risk compared with the lowest quartile, respectively. Results remained similar for NDVI, EVI, and VCF within 500-m and 1000-m buffers. Stratified analyses showed stronger association for residential greenness and diabetes among older people. The association between greenness and diabetes did not appear to be mediated by physical activity, PM2.5 or BMI. Our findings suggested that higher residential greenness was significantly associated with lower risk of diabetes.
Vegetation covers were one of the first land covers to receive special attention when thematic Land Use Cover (LUC) maps first appeared. Interest in this subject has remained strong since then because of the valuable information that these datasets provide for monitoring forests, deforestation and climate change, among other issues. A wide variety of thematic LUC datasets characterizing vegetation covers are currently available. In this chapter, we review eleven of these datasets, most of which provide long series of LUC maps, so permitting the study of LUC change. In thematic terms, most of the maps provide information on the vegetation or tree cover fraction per pixel, so characterizing the vegetation covers on Earth in great detail. A specific dataset has been found that maps mangrove distribution across the globe at 30 m for one date (1997/00). It is not included in this review because of its high specificity, which means it is only of interest to certain communities of users. Of all the products reviewed here, the World’s Forests 2000 is probably the most basic, providing information about three wooded cover categories for the year 1995/96 at a spatial resolution of 1 km. SYNMAP is a very specific thematic map designed to meet the needs of the carbon cycle and vegetation modelling community, which was produced at a spatial resolution of 1 km and with a legend of 48 categories. Among the maps providing information on the fraction of vegetation cover per pixel, the Hybrid Forest Mask 2000 (1 km) and the PTC Global Version (500 m–1 km) offer relatively coarse resolutions and few points in time: just one date in the former (2000) and two in the latter (2003, 2008). The Forests of the World 2010 is also available for just one year (2010), albeit at a more detailed spatial resolution (250 m). Various datasets provide information on the cover fraction for long periods of time at medium and high spatial resolutions. FCover provides the longest time series (1999-present) at 1 km, although since 2014 this dataset is also available at 300 m. Modis VCF also offers a long data series (2000–2019) at a spatial resolution of 250 m. MEaSUREs Vegetation Continuous Fields (VCF) is another thematic LUC dataset providing information on the tree cover fraction of the earth surface for a very long time period: 1982–2016. However, it is not reviewed here because of its coarse spatial resolution (around 5.6 km at the Equator). At very detailed spatial resolutions, GFCC30TC Landsat VCF (30 m) provides data on the cover fraction for four different points in time, between 2000 and 2015. It also gives information on forest change for two periods (1990–2000/2000–2005) through the associated GFCC30FCC dataset. The Hansen forest map (30 m) also provides one of the longest time series, from 2000 to 2019. Global FNF is the dataset with the highest resolution (25 m) of all those reviewed. It is available for two periods of time: 2007–2010 and 2015–2017. In thematic terms, however, this dataset is less detailed, in that it only differentiates between forest and non-forest covers. TanDEM-X Forest/Non-Forest also provides information on the forest extent at high spatial resolution (50 m). However, the map is only available for one point in time. Like Global FNF, it was also obtained from the classification of radar data.
Full-text available
Context Forest restoration plays an important role in global efforts to slow biodiversity loss and mitigate climate change. Vegetation in remnant forests can form striking patterns that relate to ecological processes, but restoration targets tend to overlook spatial pattern. While observations of intact reference ecosystems can help to inform restoration targets, field surveys are ill-equipped to map and quantify spatial pattern at a range of scales, and new approaches are needed. Objective This review sought to explore practical options for creating landscape-scale forest restoration targets that embrace spatial pattern. Methods We assessed how hierarchy theory, satellite remote sensing, landscape pattern analysis, drone-based remote sensing and spatial point pattern analysis could be applied to assess the spatial pattern of reference landscapes and inform forest restoration targets. Results Hierarchy theory provides an intuitive framework for stratifying landscapes as nested hierarchies of sub-catchments, forest patches and stands of trees. Several publicly available tools can map patches within landscapes, and landscape pattern analysis can be applied to quantify the spatial pattern of these patches. Drones can collect point clouds and orthomosaics at the stand scale, a plethora of software can create maps of individual trees, and spatial point pattern analysis can be applied to quantify the spatial pattern of mapped trees. Conclusions This review explored several practical options for producing landscape scale forest restoration targets that embrace spatial pattern. With the decade on ecosystem restoration underway, there is a pressing need to refine and operationalise these ideas.
Full-text available
In this paper we consider how we may determine the relative proportions of ground cover components in a mixed pixel. We assume the usual linear model for signal mixing and examine a number of methods, closely related, for estimating the proportions. We also show how the precision of our estimates can be defined. We introduce a new estimator which is based on regularisation principles and which produces a smoother set of images than other methods, and gives more accurate estimates. The methods are compared on a simulated data set.
Full-text available
This paper on reports the production of a 1 km spatial resolution land cover classification using data for 1992-1993 from the Advanced Very High Resolution Radiometer (AVHRR). This map will be included as an at-launch product of the Moderate Resolution Imaging Spectroradiometer (MODIS) to serve as an input for several algorithms requiring knowledge of land cover type. The methodology was derived from a similar effort to create a product at 8 km spatial resolution, where high resolution data sets were interpreted in order to derive a coarse-resolution training data set. A set of 37 294 x 1 km pixels was used within a hierarchical tree structure to classify the AVHRR data into 12 classes. The approach taken involved a hierarchy of pair-wise class trees where a logic based on vegetation form was applied until all classes were depicted. Multitemporal AVHRR metrics were used to predict class memberships. Minimum annual red reflectance, peak annual Normalized Difference Vegetation Index (NDVI), and minimum channel three brightness temperature were among the most used metrics. Depictions of forests and woodlands, and areas of mechanized agriculture are in general agreement with other sources of information, while classes such as low biomass agriculture and high-latitude broadleaf forest are not. Comparisons of the final product with regional digital land cover maps derived from high-resolution remotely sensed data reveal general agreement, except for apparently poor depictions of temperate pastures within areas of agriculture. Distinguishing between forest and non-forest was achieved with agreements ranging from 81 to 92% for these regional subsets. The agreements for all classes varied from an average of 65% when viewing all pixels to an average of 82% when viewing only those 1 km pixels consisting of greater than 90% one class within the high-resolution data sets.
Full-text available
[1] While most land models developed for use with climate models represent vegetation as discrete biomes, this is, at least for mixed life-form biomes, inconsistent with the leaf-level and whole-plant physiological parameterizations needed to couple these biogeophysical models with biogeochemical and ecosystem dynamics models. In this paper, we present simulations with the National Center for Atmospheric Research land surface model (NCAR LSM) that examined the effect of representing vegetation as patches of plant functional types (PFTs) that coexist within a model grid cell. This approach is consistent with ecological theory and models and allows for unified treatment of vegetation in climate and ecosystem models. In the standard NCAR LSM the PFT composition and leaf area for each grid cell are obtained by classifying grid cells as 1 of 28 possible biomes. Here, we develop a data set from 1-km satellite data that provides each model grid cell a unique PFT composition and leaf area for each PFT. Global simulations at 3degrees x 3degrees spatial resolution showed that ground temperature, ground evaporation, and northern high-latitude winter albedo exhibited direct responses to these landscape changes, which led to indirect effects such as in soil moisture and sensible and latent heat fluxes. Additional simulations at 2degrees x 2degrees and 1degrees x 1degrees spatial resolution showed that low-resolution simulations masked landscape heterogeneity in both approaches but the satellite-based, continuous representation of vegetation reduced model sensitivity to resolution. It is argued that the use of spatially continuous distributions of coexisting PFTs is a necessary step to link climate and ecosystem models.
There is currently a great deal of interest in understanding and quantifying the extent of natural and human-caused fire in the different biomes throughout the world. In Chapter 8, Shvidenko and Nilsson examined the fire statistics for the Russian boreal forest region and showed that there is much uncertainty in these data because of the management and monitoring practices in this country. Although a much more accurate record for fire activity exists for the North American boreal forest region, this data set is still incomplete. In this chapter, we discuss the pedigree of the data sets that comprise the North American boreal forest record so that their limitations can be understood. Finally, the recent advances in geographic information systems (GIS) and their adoption by natural resource management agencies have led to the creation of databases that can be used to examine the spatial extent of fires in more detail. In this chapter, we also discuss the use of these spatial data sets for examining fire in the North American boreal forest.
Two forest-cover maps of national scale have been produced under the 1993 Forest and Rangeland Renewable Resources Planning Act (RPA) Assessment Update program. Both maps are based on 1-kilometre resolution Advanced Very High Resolution Radiometer (AVHRR) data, and use of multitemporal and multisource remote sensing data analyses. The forest-type groups map depicts the distribution patterns of 25 forest-cover types over the United States. Complementary to the forest-type groups map is the predicted percent forest-cover map of the conterminous United States which portrays the relative amount of forest cover per square kilometre. The two maps, in both digital and paper forms, provide current forest information for the U.S. Techniques and results of the 1993 RPA forest mapping project are presented, and apparent changes in forest cover since 1967 are also described. 24 refs.
Mixed pixels occur commonly in remotely-sensed imagery, especially those with a coarse spatial resolution. They are a problem in land-cover mapping applications since image classification routines assume ‘pure’ or homogeneous pixels. By unmixing a pixel into its component parts it is possible to enableinter alia more accurate estimation of the areal extent of different land cover classes. In this paper two approaches to estimating sub-pixel land cover composition are investigated. One is a linear mixture model the other is a regression model based on fuzzy membership functions. For both approaches significant correlation coefficients, all >0·7, between the actual and predicted proportion of a land cover type within a pixel were obtained. Additionally a case study is presented in which the accuracy of the estimation of tropical forest extent is increased significantly through the use of sub-pixel estimates of land-cover composition rather than a conventional image classification.
The geographic distribution of vegetation over the Earth's land surface is traditionally described using classification schemes with discrete numbers of vegetation types. When such land cover data sets are used as boundary conditions in Earth system models, abrupt boundaries and unrealistic homogeneity are introduced into parameter estimates. This paper proposes an alternative approach to describe global land cover with continuous fields of vegetation characteristics. A linear mixture model is applied to 1-km advanced very high resolution radiometer data to estimate proportional cover for three important vegetation characteristics: life form (percent woody vegetation, percent herbaceous vegetation, and percent bare ground), leaf type (percent needleleaf and percent broadleaf), and leaf duration (percent evergreen and percent deciduous). Linear discriminants for input into the mixture model are derived from 30 metrics representing the annual phenological cycle. Through comparison with training data derived from a global network of Landsat multispectral scanner scenes, we conclude that the linear assumption implicit in the linear mixture model is not severely violated. The linear relationships between percent cover as determined from the training data and the linear discriminants are used to estimate end-member values, and the mixture model is applied to derive the seven layers of global continuous fields. The availability of Moderate Resolution Imaging Spectroradiometer data in the future holds promise for refining the simple technique used in this paper to derive improved global continuous fields.
This paper reports a study which aims to (i) develop methodologies for global land cover classi® cations that are objective, reproducible and feasible to implement as new satellite data become available in the future and (ii) provide a global land cover classi® cation product based on the National Aeronautics and Space Administration /National Oceanic and Atmospheric Administration Path® nder Land (PAL) data that can be used in global change research. The spatial resolution for the land cover classi® cation is 8km, intermediate between our previously published coarse one degree by one degree spatial resolution and the 1km global land cover product being developed under the auspices of the International Geosphere Biosphere Program. We ® rst derive a global network of training sites from Landsat imagery, using 156 Landsat scenes mostly from the Multispectral Scanner System, to identify over 9000 pixels in the PAL data where we have high con® dence that the labelled cover type occurs. We then use the training data to test a number of metrics that describe the temporal dynamics of vegetation over an annual cycle for potential use as input variables to a global land cover classi® cation. The tested metrics are based on: (i) the ratio between surface temperature and Normalized Di Ä erence Vegetation Index (NDVI); (ii) seasonal metrics derived from the NDVI temporal pro® le, such as length of growing season; (iii) a rule-based approach that determines cover type through a series of hierarchical trees based on surface temperature and NDVI values; and (iv) annual mean, maximum, minimum and amplitude values for all optical and thermal channels in the Advanced Very High Resolution Radiometer (AVHRR) (PAL) data. Highest mean class accuracies from a decision tree classi® er were obtained using the annual mean, maximum, minimum, and amplitude values for all AVHRR bands. Finally, we apply these metrics to 1984 PAL data at 8km resolution to derive a global land cover classi® cation product using a decision tree classi® er. The classi® cation has an overall accuracy between 81.4 and 90.3%. The Landsat images used for deriving the training data and the methodology for classi® cation of AVHRR data at 8km resolution can also be applied to 1km AVHRR data and, in the future, Moderate Resolution Imaging Spectroradiometer (MODIS) data at 250 and 500m resolution. Digital versions of the land cover dataset and detailed documentation can be found on the World Wide Web at /landcover/8km-map.html.
Many countries periodically produce national reports on the status and changes of forest resources, using statistical surveys and spatial mapping of remotely sensed data. At the global level, the Food and Agriculture Organization (FAO) of the United Nations has conducted a Forest Resources Assessment (FRA) program every 10 yr since 1980, producing statistics and analysis that give a global synopsis of forest resources in the world. For the year 2000 of the FRA program (FRA2000), a global forest cover map was produced to provide spatial context to the extensive survey. The forest cover map, produced at the U.S. Geological Survey (USGS) EROS Data Center (EDC), has five classes: closed forest, open or fragmented forest, other wooded land, other land cover, and water. The first two forested classes at the global scale were delineated using combinations of temporal compositing, modified mixture analysis, geographic stratification, and other classification techniques. The remaining three FAO classes were derived primarily from the USGS global land cover characteristics database (Loveland et al. 1999). Validated on the basis of existing reference data sets, the map is estimated to be 77% accurate for the first four classes (no reference data were available for water), and 86% accurate for the forest and nonforest classification. The final map will be published as an insert to the FAO FRA2000 report. FOR. SCI. 49(3):369–380.