ArticlePDF Available

Quantifying small-scale deforestation and forest degradation in African woodlands using radar imagery


Abstract and Figures

Carbon emissions from tropical land‐use change are a major uncertainty in the global carbon cycle. In African woodlands, small‐scale farming and the need for fuel are thought to be reducing vegetation carbon stocks, but quantification of these processes is hindered by the limitations of optical remote sensing and a lack of ground data. Here, we present a method for mapping vegetation carbon stocks and their changes over a 3‐year period in a > 1000 km2 region in central Mozambique at 0.06 ha resolution. L‐band synthetic aperture radar imagery and an inventory of 96 plots are combined using regression and bootstrapping to generate biomass maps with known uncertainties. The resultant maps have sufficient accuracy to be capable of detecting changes in forest carbon stocks of as little as 12 MgC ha−1 over 3 years with 95% confidence. This allows characterization of biomass loss from deforestation and forest degradation at a new level of detail. Total aboveground biomass in the study area was reduced by 6.9 ± 4.6% over 3 years: from 2.13 ± 0.12 TgC in 2007 to 1.98 ± 0.11 TgC in 2010, a loss of 0.15 ± 0.10 TgC. Degradation probably contributed 67% (96.9 ± 91.0 GgC) of the net loss of biomass, but is associated with high uncertainty. The detailed mapping of carbon stock changes quantifies the nature of small‐scale farming. New clearances were on average small (median 0.2 ha) and were often additions to already cleared land. Deforestation events reduced biomass from 33.5 to 11.9 MgC ha−1 on average. Contrary to expectations, we did not find evidence that clearances were targeted towards areas of high biomass. Our method is scalable and suitable for monitoring land cover change and vegetation carbon stocks in woodland ecosystems, and can support policy approaches towards reducing emissions from deforestation and degradation (REDD).
Content may be subject to copyright.
Quantifying small-scale deforestation and forest
degradation in African woodlands using radar imagery
*School of Geosciences, University of Edinburgh, Edinburgh EH9 3JN, UK, The National Centre for Earth Observation, Natural
Environment Research Council, UK
Carbon emissions from tropical land-use change are a major uncertainty in the global carbon cycle. In African wood-
lands, small-scale farming and the need for fuel are thought to be reducing vegetation carbon stocks, but quantifica-
tion of these processes is hindered by the limitations of optical remote sensing and a lack of ground data. Here, we
present a method for mapping vegetation carbon stocks and their changes over a 3-year period in a >1000 km
region in central Mozambique at 0.06 ha resolution. L-band synthetic aperture radar imagery and an inventory of 96
plots are combined using regression and bootstrapping to generate biomass maps with known uncertainties. The
resultant maps have sufficient accuracy to be capable of detecting changes in forest carbon stocks of as little as
12 MgC ha
over 3 years with 95% confidence. This allows characterization of biomass loss from deforestation and
forest degradation at a new level of detail. Total aboveground biomass in the study area was reduced by 6.9 ±4.6%
over 3 years: from 2.13 ±0.12 TgC in 2007 to 1.98 ±0.11 TgC in 2010, a loss of 0.15 ±0.10 TgC. Degradation probably
contributed 67% (96.9 ±91.0 GgC) of the net loss of biomass, but is associated with high uncertainty. The detailed
mapping of carbon stock changes quantifies the nature of small-scale farming. New clearances were on average small
(median 0.2 ha) and were often additions to already cleared land. Deforestation events reduced biomass from 33.5 to
11.9 MgC ha
on average
Contrary to expectations, we did not find evidence that clearances were targeted towards
areas of high biomass. Our method is scalable and suitable for monitoring land cover change and vegetation carbon
stocks in woodland ecosystems, and can support policy approaches towards reducing emissions from deforestation
and degradation (REDD).
Keywords: agriculture, ALOS PALSAR, backscatter, carbon mapping, carbon stocks, emissions, land-use change, machamba,
Received 5 May 2011; revised version received 19 August 2011 and accepted 26 August 2011
Deforestation and other land-use change are major
components of the anthropogenic carbon (C) cycle,
transferring 0.92.2 PgC year
from the biota to the
atmosphere (Houghton, 2010). This number is highly
uncertain (Denman et al., 2007; Ramankutty et al., 2007;
Van Der Werf et al., 2009; Houghton, 2010) and esti-
mates often exclude many of the processes leading to
degradation of forest land (Houghton, 2010). Deforesta-
tion is primarily the result of the clearing of land for
agriculture (Geist & Lambin, 2002), both for the large-
scale production of global commodities (Defries et al.,
2010), and, particularly in Africa, for small-scale pro-
duction of food and cash crops (Burgess et al., 2002;
Fisher, 2010). In Africa, land-use change emissions are
thought to be in region of 0.3 ±0.2 PgC year
(Houghton & Hackler, 2006; Williams et al., 2007; Ciais
et al., 2011), but the data underlying these estimates
come from extrapolation of outdated, unreliable and
inconsistent national estimates (Grainger, 2008; Kinder-
mann et al., 2008; FAO, 2010).
Woodlands, characterized by an open tree canopy
and a continuous grass layer, cover 36% of vegetated
Africa (Mayaux et al., 2004), and as such represent a
low density, but large, stock of vegetation C (Dewees
et al., 2010). African countries dominated by woodlands
have high population densities, high population
growth rates and high, but uncertain, deforestation
rates (FAO, 2010). As a result, nations dominated by
woodlands contribute around half of Africa’s deforesta-
tion emissions [based on data from Mayaux et al. (2004)
and FAO (2010)]. Woodlands are particularly difficult
Present address: Claire Ghee, The James Hutton Institute,
Invergowrie, Dundee DD2 5DA, UK.
Correspondence: Casey M. Ryan, tel. + 44 131 650 7722,
fax + 44 131 662 0478, e-mail:
©2011 Blackwell Publishing Ltd 1
Global Change Biology (2011), doi: 10.1111/j.1365-2486.2011.02551.x
to monitor with optical/infrared remote sensing due to
inter- and intra-annual changes in tree leaf display
(Grainger, 1999) and the transient presence of a sub-
stantial grass layer which complicates the interpretation
of such satellite imagery (Archibald & Scholes, 2007).
Consequently, degradation emissions from woodlands
are highly uncertain, but are thought to be substantial
(Ahrends et al., 2010).
The standard approach to estimating C emissions
from deforestation is based on estimates of changes in
forest area, aided by increasingly robust estimates of
forest area change (Etter et al., 2006; Achard et al., 2007;
Hansen et al., 2008; Miettinen et al., 2011). Degradation
emission estimates, however, require repeat in situ mea-
surements of carbon density (GOFC-GOLD, 2010)
which are scarce (Ahrends et al., 2010), because they
require a large number of plots and strata to estimate
accurately any changes in mean C density. Recent air-
borne approaches (Asner et al., 2010) have not been
widely deployed to date. As a result, changes in forest
C density resulting from fire (Ryan & Williams, 2011)
or the selective extraction of biomass for fuel or timber
(Nepstad et al., 1999), are rarely assessed and have not
been fully included in emissions estimates (Houghton
& Hackler, 2006; Denman et al., 2007).
Even when combined, the methods outlined above
yield large uncertainties on estimates of changes in veg-
etation C stock, let alone on emissions (Houghton et al.,
2009). Uncertainty stems from three main issues: (i) The
use of arbitrary forest/nonforest thresholds whereas C
stocks are a continuous variable. Such thresholds are a
particular problem in woodlands where distinct edges
are rare, (ii) a lack of in situ measurements of C density,
and the potential artefacts associated with differential
land use on research plots and (iii) land that is defor-
ested or degraded may not be representative of the for-
est type in which it is found (Loarie et al., 2009;
Houghton, 2010). That is, farmers are likely to target
areas for deforestation based on careful consideration
of agricultural potential. Therefore, land with higher
than average soil fertility and/or water availability
might be expected to have a greater probability of con-
version to agriculture. The same areas might also be
expected to have higher than average soil and vegeta-
tion C stocks due to the increased productivity.
Small-scale, often shifting, cultivation exemplifies the
problem of estimating changes in C stocks in forests
and woodlands, producing a mosaic landscape that is
frequently misclassified (Mertz, 2009) and is rarely
well-represented by discrete land cover classes. The
landscape C dynamics (Williams et al., 2008) cannot be
adequately described by changes between categories
such as degraded land, forest or agriculture (Schmidt-
Vogt et al., 2009). Furthermore, the scale at which
small-scale farmers clear forest is constrained by a lack
of mechanical power and transport. This physical limit
results in small farms (scale ~ha) that may not be
recorded on land-use maps or detected with accuracy
by coarse-scale (~km
) satellite-based land cover change
Thus, there is a clear need for direct measurement at
sub-hectare resolution of changes in C stocks due to
land cover change (LCC, Houghton et al., 2009). Our
first aim is to demonstrate a method for such measure-
ments in African woodlands at a spatial resolution
sufficient to characterize small-scale farming and deg-
radation. This is achieved by the use of 25 m resolution
L-band radar imagery to map C stocks and their
changes through the years 20072010 in an area of cen-
tral Mozambique. Radar imagery has several advanta-
ges for this purpose. First, cloud and atmospheric
effects are largely irrelevant allowing observations at
cloudy sites. Second, L-band (23 cm wavelength) nor-
malized radar cross section (hereafter referred to as
backscatter) has been shown to have a reasonably direct
relationship to woody biomass up to a saturation at
around 50 MgC ha
(Le Toan et al., 1992; Rignot et al.,
1994; Magnusson et al., 2007; Karjalainen et al., 2009).
These advantages mean that spaceborne radar imagery
is increasingly used in support of land-use mapping in
the tropics (Van Der Sanden & Hoekman, 1999; Hoek-
man et al., 2010; Rahman & Sumantyo, 2010). Previous
work has shown that L-band radar backscatter is well
correlated to biomass across several African landscapes
(Mitchard et al., 2009).
The resultant C maps allow us to ask several ques-
tions about the nature of the processes of deforestation
and degradation:
How much carbon is lost to deforestation compared
to degradation?
• Are areas of high carbon density preferentially tar-
geted for LCC?
What is the carbon density of changed areas before
and after LCC?
• What size are LCC events, and how are they clus-
tered in space?
Site description and land-use history
Our study area covers 1160 km
in the Gorongosa and Nham-
atanda districts of Sofala province in central Mozambique. It is
primarily dominated by miombo woodland, the most wide-
spread vegetation type in Southern Africa (Frost, 1996). It has
a seasonal wet-dry climate with ~900 mm rain per year, with
82% falling in the 5 months between November and March
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
2C. M. RYAN et al.
(Ryan et al., 2011). The vegetation consists of miombo wood-
land on the well-drained flanks of the Rift valley, grading to
more scattered savanna on the poorly drained valley floor
(Tinley, 1977, 1982). The terrain is gently undulating in most
of the study area, but in the west there are steep slopes associ-
ated with the Pungue and Vunduzi rivers (Fig. S1). Ninety
seven percent of the study area has a slope of <10°(based
on 90 m resolution elevation data from the Shuttle Radar
Topography Mission (SRTM, Farr et al., 2007, http://srtm.
The area is undergoing rapid land-use change: in 1992, with
the end of the Mozambican civil war, there were major popu-
lation movements in the area, including resettlement of semi-
abandoned rural areas such as Nhambita (Fig. 1). The only
surfaced road in the area, the EN1, was rebuilt, along with a
bridge over the River Pungue (rehabilitation of both took place
from 1999 to 2002), connecting the study area and Gorongosa
town to the Beira corridor, the major area of economic activity
in central Mozambique. Gorongosa town has grown rapidly
since 1992 (INE, 2010) with the district population increasing
by 50% from 1997 to 2007 to 117, 129. Currently, 69% of the
population is aged below 18 years (INE, 2010). Forest or
woodland loss in the area is primarily the result of: (i) clear-
ance for small-scale agriculture, notably for maize production
on farms of between ~1 and 2 ha and (ii) charcoal production,
involving the selective removal of medium size stems from an
area of ~0.2 ha surrounding temporary kilns. Charcoal is sold
along the EN1 highway for transport south to Inchope and Be-
ira, but there is also some demand for charcoal and fuel wood
in Gorongosa town. In addition, fire is extensively used to
manage the landscape and wild fires are common. Frequent
and intense fires can reduce biomass in these woodlands
(Furley et al., 2008; Ryan & Williams, 2011).
Carbon stock estimation
The basis of our approach is to produce a 3-year time series of
carbon maps of the area. The maps are produced using a com-
bination of satellite radar images and in situ carbon stock
inventories. We deal here only with carbon in the above-
ground woody vegetation pool (AGB, MgC ha
), although
more information on belowground biomass at this site can be
found in Ryan et al. (2011).
Radar imagery. Synthetic Aperture Radar (SAR) remote
sensing can provide information on vegetation biomass (Le
Toan et al., 1992), as well as many other characteristics of
the land surface (Woodhouse, 2006a). SAR utilizes an active
sensor aboard a satellite or plane to send out a beam of
energy and measures the intensity of the echoes that return
to the sensor. The parameter of interest is the backscatter
(technically the normalized radar cross section, a unitless
variable (m
)). Backscatter can loosely be thought of as
the ratio of the power that comes back from a patch of
ground to the power sent to that patch of ground (based on
the arbitrary assumption that the ground is an isotropic scat-
terer). The energy that returns to the sensor varies with the
proportion of the incident energy that is scattered by the
land surface, and the directionality of that scattering. When
energy of an appropriate wavelength is used, it interacts
with the structural elements of the tree canopy (branches
Goron gosa
Remo te Pa rk
Mbula wa
Sanc tuar y
0 8 16 24 324
Kilomete rs
Remo te Pa rk
Sanc tuar y
Goron gosa
Remote Park
Sanc tuar y
(b)(a) (c)
Fig. 1 Estimated carbon stocks in the study area in (a) 2007 and (b) 2010. Each image is derived from the mean of three ALOS PALSAR
scenes from that year’s dry season. (c) shows the areas detected as undergoing abrupt change (red) with a probability greater than 95%
and a reduction in biomass to less than 50% of original biomass. Areas that did not undergo change are indicated in grey, and white
indicates areas with <10 tC ha
. The maps are annotated with the local road network and the ‘sub-areas‘ used in Fig. 7. (GIS data cour-
tesy of ARA-CENTRO).
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
and trunks): typically, more woody biomass results in more
diffuse scattering and thus more energy being returned to
the sensor and a higher backscatter value is recorded. Back-
scatter is also affected by soil roughness (as it changes the
directionality of the scattering) and moisture (as it changes
the total proportion of scattering), as well as other environ-
mental factors. In this study L-band (~23 cm wavelength)
SAR imagery was used, as it is less affected by soil condi-
tions than shorter wavelengths, and is known to be able to
detect deforestation and to be sensitive to forest biomass
due to its ability to penetrate the forest canopy (Almeida-
Filho et al., 2005; Fransson et al., 2007; Karjalainen et al.,
2009; Mitchard et al., 2009). Ten images were acquired that
covered the study site spanning the dates 23 June 2007 to 1
October 2010 (Table 1). All images were for the dry season
and two or three were available for each year.
The technical details of the imagery and processing now fol-
low. Images were obtained from the phased array L-band syn-
thetic aperture radar sensor aboard the advanced land
observing satellite (ALOS PALSAR) in the Fine Beam Dual
mode (Shimada et al., 2010). All images were acquired on the
ascending pass, have an incidence angle centred on 34.3°, and
were provided at a pixel size of 12.5 m with 4 equivalent
looks, but then averaged to 25 m for 16 equivalent looks. Only
Horizontal-send Vertical-receive data are used here, as previ-
ous studies have shown this polarization to be more sensitive
to biomass than Horizontal-send Horizontal-receive (Le Toan
et al., 1992; Mitchard et al., 2009).
The images were processed using the Alaska Satellite Faci-
lity’s MapReady software v2.3.6 (ASF, 2010), combined with
~90 m elevation data from the SRTM. Each image was
converted from digital numbers to backscatter using the cali-
bration coefficients of Shimada et al. (2009) and a geometric
and radiometric terrain correction was applied. Visual com-
parison with optical imagery (IKONOS and Landsat) showed
that the resulting images were well geolocated (12 pixel
error, i.e. 2550 m) compared to prominent landscape features
such as roads, bridges and an airstrip. In particular, the
images were very closely geolocated to each other: no geoloca-
tion differences were visible between the ten images. Com-
monly, backscatter values are used on the logarithmic dB
scale, but here untransformed values of backscatter are used,
as the backscatter-biomass trend is usually linear in untrans-
formed space, but nonlinear in dB space.
Ground data. For each radar scene, backscatter values were
converted into carbon density using a regression equation
based on inventories of 96 plots in the forest, woodland and
cropland in the south of the study area. These data come from
a range of inventories conducted during 20062009, and
include plots of sizes ranging from 0.1 to 2.2 ha (mean ±SD
0.63 ±0.33 ha, Fig. 2). The inventories include all standing
trees >5 cm diameter at breast height (DBH). Many of the
woodland inventory data are described in Ryan et al. (2011),
as is the site-specific allometric equation for converting stem
diameter to carbon mass for each stem. Briefly, the data set
consist of: fifteen 1 ha square permanent sample plots in the
woodland and savanna (Ryan et al., 2011), eight 0.28 ha plots
used in the fire experiments of Ryan & Williams (2011), thirty
0.57 ha plots in a transect from open savanna to woodland
(Woollen et al. in prep; these had a nested sampling design,
see supplementary information), five 0.5 ha plots in relatively
remote woodland, 37 plots of 0.12.2 ha (mean 0.6 ha) on
cropland (Ghee, 2009) and one dense forest plot consisting of
three 0.04 ha subplots. Details of the plot data are given in the
supplementary information. Every plot is thought to have
avoided abrupt LCC (i.e. deforestation) during the study
period, assessed by repeat ground visits or satellite imagery;
however, we cannot rule out gradual changes due to
aggradation or degradation.
Table 1 Regression statistics for the relationship between radar backscatter (unitless) and aboveground woody biomass (MgC
RMA slope
(MgC ha
RMA intercept
(MgC ha
interval Intercept
interval adj R
(MgC ha
(MgC ha
(MgC ha
23-Jun-07 1562 13151810 20.9 26.6 to 15.1 0.39 10.4 10.9 1.7
08-Aug-07 1268 10821453 13.7 18.0 to 9.5 0.49 9.3 9.9 1.5
23-Sep-07 1329 11531505 11.0 14.5 to 7.5 0.58 8.4 8.7 1.4
10-May-08 1647 14171877 22.6 27.9 to 17.3 0.53 8.9 9.3 1.5
25-Jun-08 1632 13851880 19.9 25.2 to 14.5 0.44 9.8 10.4 1.6
28-Jun-09 1601 13631840 17.5 22.3 to 12.6 0.47 9.6 10.1 1.6
28-Sep-09 1303 11151490 11.2 15.0 to 7.4 0.50 9.2 9.6 1.5
16-May-10 1801 15182083 24.7 31.0 to 18.4 0.41 10.3 10.7 1.7
01-Jul-10 1679 14311926 25.6 31.6 to 19.6 0.48 9.5 9.9 1.6
01-Oct-10 1348 11691526 13.3 17.0 to 9.5 0.58 8.4 8.7 1.4
Mean 1517 12951739 18.0 22.9 to 13.1 0.49 9.4 9.8 1.6
RMA indicates reduced major axis regression; OLS, ordinary least squares regression; RMSE, root mean squared error.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
4C. M. RYAN et al.
Regression and error propagation. To convert backscatter
images into aboveground biomass (AGB in units of
MgC ha
), for each timeslice we regress plot mean backscat-
ter against plot AGB, assuming that the plot AGB is consistent
over the observational period. We use reduced major axis
(RMA) regression (Mitchard, 2011) implemented in MATLAB by
Trujillo-Ortiz & Hernandez-Walls (2010). RMA regression
minimizes the errors on both axes (rather than just on the
y-axis as in normal regression), which is appropriate because
there are errors in both data sets and the observer controls nei-
ther (Sokal & Rohlf, 1995).
To estimate error on the predictions from the regression, a
5000 92-fold cross-validation procedure was employed. Half
the dataset was withheld and used to estimate the root mean
squared validation error (RMSE), and bias (B) of predictions
from each regression. These statistics are defined as (Hui &
Jackson, 2007):
where Y
is a validation data point (not used in the regression)
Yiis the prediction from the regression. Band RMSE were
calculated for all 10 time slices 5000 times, each time with a
different random split of the data, and the mean values of B
and RMSE for the 5000 validations are reported. For indicative
purposes the adjusted R
of an ordinary least squares regres-
sion is also shown in Table 1.
The temporal covariance of the bias was quantified by look-
ing at the correlation of the bias between time steps for each of
the 5000 validations, i.e. by asking if the regression is biased
high in year x, is it likely to be biased high in subsequent years
Regression errors were propagated to the carbon maps with
a bootstrap procedure. Bootstrapping was used as it implicitly
includes any spatial or temporal covariance of the uncertainty
resulting from the AGB-backscatter regressions. For each set
of 10 carbon maps, 30 000 realizations were created each using
a different regression based on randomly selected regression
data (resampled with replacement). Derived quantities such
as carbon stock change, rates of loss, and the deforestation
and degradation totals, were also calculated for each of the
30 000 bootstraps, allowing the uncertainties introduced by
the regression to be estimated. All values are reported as the
mean ±the standard deviation of the 30 000 bootstraps. The
number of bootstraps was chosen after initial analysis showed
that repeat calculation with different random number seeds
using 10 000 bootstraps yielded identical results to three
significant figures.
Change detection algorithm
One of the aims of this article is to estimate parameters of pix-
els before and after LCC. This necessarily involves defining a
threshold of change. Here we focus on detecting abrupt losses
of AGB and define LCC of interest as being in pixels where:
(i) AGB is >20 MgC ha
at the start of the study, thus
increasing the likelihood that only forested pixels are exam-
ined, (ii) the AGB of the pixel after the change is reduced to a
minimum percentage (φ, default 50%) of its initial AGB, a
threshold which should exclude most natural changes in for-
est structure, and (iii) the probability that the change in AGB
occurred by chance (given the noise in the data) is <(1a)
(default a=0.95). The use of a ratio to define change is
appropriate with SAR imagery in power space, as ratioing
SAR images is generally preferred to differencing because of
its noise characteristics (Rignot & Van Zyl, 1993; Radke et al.,
The time series C(t), where Cis the AGB of a pixel over the
t=1,,10 images, provides a rich data source with which to
detect changes, and the 10 time slices allow more confidence
in using noisy data. For change detection, a simple iterative
algorithm locates the time-point of change, τ, at which the
mean of the preceding values of C(t) is most different to the
mean of the remaining part of the time series. The mean AGB
before (C1) and after (C2)τare then estimated. We use a
Monte Carlo procedure to estimate the probability, P,of
obtaining the change in AGB (C1C2) by chance, based on
sampling from the time series and including the RMSE valida-
tion errors from the regression (Table 1). A mathematical
description is given in the supplementary information.
Once changed pixels have been identified, each LCC event
(defined as adjacent changed pixels that have identical τ) was
automatically converted to a polygon in ARCGIS (ESRI, CA,
USA) and the area of each polygon calculated.
Change detection validation
Two tests were used to produce accuracy statistics for the
change detection. To assess successful LCC detection we
delineated 92 agricultural fields formed by clearing woodland
0 10 20 30 40 50 60
round biomass, M
C ha–1
Number of plots
Fig. 2 Histogram of the aboveground biomass (AGB) of the
plots used to estimate and validate the AGB-backscatter rela-
tionship. Aboveground woody carbon stock is estimated from
diameter at breast height (DBH) measurements using a site-spe-
cific allometric equation. n=96.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
between March 2007 and June 2009. These conversions were
assessed by visual interpretation of multispectral SPOT 4
(20 m resolution) imagery from 7 March 2007 and IKONOS
(1 m) imagery from 24 June 2009 (an example is shown in
Fig. 6). The edges of the new farms were converted by hand to
polygons using ARCGIS. These areas of known LCC had a
mean area of 1.7 ±1.5 ha and were well distributed across the
study area. Although the AGB of these new farms is
unknown, visual inspection suggests that very few trees
remain and that they have crossed the change threshold out-
lined above. The miss rate is defined as the number of pixels
on these farms that are not detected as changed.
To assess false positives, an area of the Gorongosa National
Park, termed ‘The Sanctuary’ (Fig. 1) that is extremely
unlikely to have undergone human-induced LCC was analy-
sed. The Sanctuary is an electric-fenced area of 5761 ha
designed to retain animals imported to the Park (not including
elephants). Note that this definition of false positive is very
broad and includes both natural changes to the woodland as
well as random errors in detection as ‘false positives’.
The sensitivity of the hit rate and false positive rate to the
parameters used in the change detection algorithm was
assessed: φ,the fraction of AGB remaining on a changed
pixel was varied from 80% to 20%, and a, the threshold P-
value above which a LCC event is considered significant,
from 0.8 to 1.
Assessing preferential selection of high biomass areas
To test whether land with high AGB is preferentially selected
for LCC, compared with the null hypothesis that LCC is ran-
dom with respect to AGB, we compare the observed results to
a pseudodata set for which the null hypothesis is true. The
pseudodata time series of AGB for each pixel, denoted C*(t),is
constructed as follows: for each pixel, a time series is con-
structed with a constant mean equal to the observed AGB in
2007. Then a random 4% of pixels are pseudodeforested at a
time point randomly selected between t=2 and t=9. At this
time point, an abrupt change is imposed and the AGB reduced
to a mean of 10 MgC ha
. Noise was added to the time series
by drawing each value from a normal distribution with means
as described above and standard deviations N(t), the RMSE
validation errors from the regressions. The change detection
algorithm was run as described above to identify pixels that
were pseudochanged, estimating the parameters C1* and C2*,
the pseudo-equivalents of C1 and C2. The resulting distribu-
tions of C1* and C1 are compared using a two sample Kol-
mogorov-Smirnov test to test if the distribution of C1 is larger
than C1*.
Biomass loss due to deforestation and degradation
To compare the contributions of deforestation and degrada-
tion to the total carbon loss from the study area, the change
(ΔC) between the mean AGB of the three images in 2007
) and the mean AGB of the three images in 2010 (C
was examined. The image is classified into four categories that
follow the conventional definition of de/reforestation and de/
aggradation based on a binary forest/nonforest pixel classifi-
cation in 2007 and 2010. The forest/nonforest classification is
made using a default AGB threshold of 15 MgC ha
. How-
ever, the effects of varying the threshold from 5 to 20 MgC
were evaluated.
The deforestation AGB loss (Δ
) is defined as the net sum of
ΔC for pixels that shift from the forest to nonforest classes.
Reforestation gain (Δ
) is the net sum of ΔC for pixels moving
from nonforest to forest. Forest de/aggradation (Δ
) is the
net sum of ΔC for forest-pixels remaining forest, and nonforest
de/aggradation (Δ
) is the net sum of ΔC for nonforest pixels
remaining nonforest. Thus:
C2007 C2010 ¼DC¼DDþDRþDGF þDGN
and net deforestation is:
DD,net ¼DDþDR
Thus for all terms describing a change in biomass stocks,
positive numbers indicate loss of biomass and negative values
denote gains.
Ground data and regression
The 96 plots ranged in AGB from 0 to 56 MgC ha
(mean 15 ±12 MgC ha
; Fig. 2; ±indicates one stan-
dard deviation throughout). The ordinary least squares
regression of backscatter against AGB gave values of
from 0.40 to 0.58 (Table 1). For the ten images,
regression slopes ranged from 1272 to 1803 MgC ha
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055
Aboveground biomass (MgC ha–1)
−23.01 −20.00 −18.24 −16.99 −16.02 −15.23 −14.56 −13.98 −13.47 −13.01 −12.60
Radar backscatter, (dB)
Radar backscatter, (m2/m2)
Fig. 3 Regression of radar backscatter from ALOS PALSAR
and aboveground biomass. Ten regressions were performed,
one for each image in Table 1, but for clarity only the data and
regression lines from the first (June-2007, open circle, dashed
line), sixth (June-2009, closed circles, light line) and last (Octo-
ber-2010, crosses, heavy line) image are shown. Each line is fit to
minimize the errors on both axes (RMA regression).
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
6C. M. RYAN et al.
(Table 1; Fig. 3) and this translated into moderate vari-
ation in estimated carbon stocks through the study per-
iod. For instance, backscatter values for a pixel over a
patch of protected forest ranged from 89% to 107% of
the first image, which translated to a range of 47
61 MgC ha
(Fig. 4).
The validation procedure estimated RMSE validation
errors of 8.710.9 MgC ha
for the different time slices
(mean error 9.8 ±0.7 MgC ha
), and mean absolute
bias of 1.6 ±0.1 MgC ha
(Table 1). The worst-case
scenario for change detection is that these biases are
random between each scene, but the covariance of these
biases through time was high (Table S1), with, for
example, r>0.86 between the bias in the three scenes
from 2007, and r>0.44 for bias between years.
Carbon stocks and changes
The carbon stocks in the study area exhibit an east-west
gradient related to the topography, with AGB
~60 MgC ha
in the undisturbed areas to the west of
the Vunduzi river falling to ~30 MgC ha
in the centre
of the study area (Fig. 1). East, towards the floor of the
Rift valley, substantial biomass is restricted to the river
lines and high points. Imposed on this topographic
pattern, the effects of human disturbance are obvious,
with almost no large blocks of woodland remaining to
the south of the River Pungue and along the west side
of the highway. However LCC is less apparent to the
east of the highway, an area that is part of the buffer
zone of the Gorongosa National Park. The town of
Gorongosa marks the epicentre of a zone of reduced
The study area contained AGB of 2.13 ±0.12 TgC
in 2007 and 1.98 ±0.11 TgC in 2010, a loss of 0.15
±0.10 TgC, or 6.9 ±4.6% of the 2007 AGB over 3 years.
Pixels that lost >9tCha
contributed half the total loss
in AGB, with the remaining loss being in pixels that lost
<9 tC.
Spatial patterns of carbon stock change
Looking at the spatial patterns of change in C stocks
(Fig. 5), there were several areas of increasing AGB
mainly in the Park and Sanctuary (A; letters refer to
points marked on Fig. 5; names to the sub-areas delin-
eated with black lines in Fig. 1) as well as decreases (B).
Losses were observed all along the highway, but partic-
ularly to the west of the road; in comparison, to the east
of the road in the buffer zone of the Park there are
fewer areas of biomass loss (C). New farms were
opened in the area between the Vunduzi river and the
road, and a cluster of new farms can be seen in the inac-
cessible region to the west of the river (D). A new
power line built in the Park is visible along the bound-
ary of the Nhambita sub area (E). An area of private
land, in which charcoal production and agriculture are
not present (Eng. A. Serra, pers. comm. 2011) stands
out clearly from the surrounding decrease in AGB in
Mbulawa (F). In Nhambita a string of new farms along
a previously high AGB river line is visible (G).
The probability distribution functions (PDFs) of car-
bon stocks in the sub areas (Fig. 7) provide further
insight into the AGB stocks and changes. In the inhab-
ited areas (Gorongosa town, Nhambita, Mucombeze
and Mbulawa) the LCC is evident in the difference
between the 2007 and 2010 PDFs. In contrast, in the
Sanctuary, the 2010 PDF is shifted to higher biomass
values compared to 2007, indicating regrowth. The
Remote Park shows no change in AGB over the study
period. Vunduzi, an almost undisturbed, well wooded
area, is the only sub-area with a normal PDF. The bimo-
dal PDF of Gorongosa appears to consist of a defor-
ested PDF similar to Mucombeze combined with a
woodland PDF centred on 40 MgC ha
Deforestation and degradation
The loss of AGB across the study area over the 3 years
was 149 ±101 GgC. Using the default threshold of
forest being any pixel with AGB >15 MgC ha
, loss
01−Jan−2007 01−Jan−2008 01−Jan−2009 01−Jan−2010 01−Jan−2011
Aboveground biomass, MgC ha–1
Fig. 4 Example time series for a single pixel over (a) an undis-
turbed riverine forest (solid line, filled circles) and (b) an area of
woodland converted to a farm (dashed line, open circles). The
lines indicate the parameters as estimated by the change detec-
tion algorithm. Over the protected forest no change was
detected (mean aboveground biomass (AGB) of 55.0 ±4.5 MgC
). The new farm (illustrated in Fig. 6) was created between
June 2008 and June 2009. AGB before the change point was
42.7 ±1.8 and 5.3 ±5.3 MgC ha
afterwards. The probability
of this change being observed by chance is <0.001. The points
representing the forest pixel have been shifted forward by
10 days for clarity.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
of AGB due to deforestation was 92 ±11 GgC
(Table 2). This was offset by a reforestation gain in
AGB of 36 ±4.5 GgC, leading to a net deforestation
loss of 55 ±12 GgC. This can be compared to forest
degradation loss of 42 ±73 GgC and nonforest degra-
dation of 53 ±43 GgC, so that total degradation is
94 ±90 GgC. Thus most of the uncertainty in ΔC comes
from the degradation term.
Thus, the best guess of the percentage of net biomass
loss that can be attributed to degradation is 67% (med-
ian of the bootstraps). This result varied only slightly
with different thresholds of forest/nonforest: from 69%
with 10 MgC ha
as the threshold to 65% when using
20 MgC ha
as the threshold. This proportion of bio-
mass loss attributed to degradation is highly uncertain,
with an 80% confidence interval of 2180% (Fig. S2).
Fig. 5 Carbon stock change in the study area. The image shows the aboveground biomass (AGB) in 2010 as a percentage of AGB in
2007. Values greater than 100% indicate areas of biomass gain (blue) and below 100% areas of biomass loss (red). Features lettered A-F
are described in the main text. Areas with AGB <10 tC ha
are shown in white. The red box indicates the area shown in Fig 6. The red
line shows the EN1 highway. Rivers are marked in blue.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
8C. M. RYAN et al.
There is a 79% chance that degradation was >50% of
the net loss of AGB.
Characteristics of land cover change events
The change detection algorithm was able to detect per-
pixel abrupt changes of >12 MgC ha
, a threshold that
is a function of the noise associated with the measure-
ments of AGB, the number of observations and the
detection algorithm parameters aand φ. Using the
default parameters (a=0.95 and φ=50%) abrupt LCC
was detected in 2.61% of the study area.
In total, 3029 ha were detected as changed in 6761
events (mean size =0.45 ±0.80 ha, median =0.19 ha).
Events larger than the median contributed 88% of the
total changed area; events >0.5 ha, 70%; and events
>1 ha contributed 49% of the total changed area. The
mean event size (assumed to be false positives) in
the Sanctuary (0.14 ha) was smaller than the events in
the inhabited sub-areas (means 0.340.52 ha). In the
inhabited areas, event size was smallest in Mucombeze
(mean 0.34 ha), where forested land is very scarce (26%
of land was forested in 2007) and largest (mean 0.52 ha)
in Vunduzi where all land was forested in 2007. The
other inhabited sub-areas have intermediate event sizes
and forest cover.
The mean change in AGB when an event occurred
was for the area to lose 21.3 ±5.5 MgC ha
, being
reduced from C1 =33.5 ±9.8 to C2 =11.9 ±18.4
MgC ha
. The reduction in AGB ranged from 12.2
56.9 MgC ha
. With the exception of the Vunduzi sub-
area, the mean pre-change AGB of changed pixels (C1)
was never more than 0.7 MgC ha
higher than if LCC
was random with respect to AGB (C1*), although the
differences were significant to P<0.001 (Fig. 7). This
suggests that the hypothesis that high AGB pixels are
preferentially subject to LCC is not true to a substantial
extent. However, in the Vunduzi sub-area, C1
(53.6 MgC ha
) was 4.7 MgC ha
higher (P<0.001)
than C1* (48.9 MgC ha
), suggesting a modest prefer-
ence towards high AGB land, but the number of chan-
ged pixels is relatively small (n=345).
Events were significantly clustered (P<0.001)
according to the nearest neighbour distance. The
observed mean distance from one event to the nearest
was 87 m compared to 206 m if the events were evenly
spaced. Many cleared areas are the result of small
expansions to the frontiers of already cleared land
(example in Fig. 6).
Change detection validation and sensitivity
Using our default parameters, LCC was detected in
1646 of the 2611 validation pixels, a miss rate of 35%.
Generally, the missed pixels were at the fringes of the
events only in 10 of the 90 validation events was no
LCC detected implying that the misses will affect
the size of detected events more than the number.
The false positive rate was 0.016% in the Sanctuary,
equal to 0.005% year
. The sensitivity analysis
(Fig. 8) showed that it is possible to suppress the false
positive rate further, but only at the expense of an
increase to the miss rate. A zero false positive rate
can be achieved in the Sanctuary by reducing φto
20% (for a>0.95), but this gives a miss rate of 74%.
Conversely, a lower miss rate can be achieved, say
20%, with φ=60% and a=0.9, but this gives a false
positive rate of 0.07%.
Can L-band radar accurately measure changes in
Our method was able to detect a loss (6.9 ±4.6%) of
biomass from the landscape over a relatively short
observation period of 3 years. This change in biomass
(ΔC) is equivalent to a reduction in C density from 18.4
to 17.1 MgC ha
across the whole study area, a change
that would be difficult to detect without very accurate
stratification and intensive ground-based sampling.
The relative error on the standing stock estimates is 6%,
but this increases to 67% for ΔC, the change in stocks
(Table 2). Without the high covariance of the regression
Table 2 Change in aboveground biomass (AGB) stocks classified by forest/nonforest transitions over the period 20072010
Transition type Area (km
) AGB loss over 3 years (GgC)
Deforestation (forest ?nonforest) 75 92 ±11
55 ±12 (33%)Reforestation (nonforest ?forest) 48 36 ±4.5
Forest degradation (forest ?forest) 530 42 ±73
94 ±90 (67%)Nonforest degradation (nonforest ?nonforest) 507 53 ±43
All 1160 149 ±101
The threshold for forest is AGB>15 MgC ha
. Negative changes indicate a gain in biomass stocks.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
errors through time (Table S1) the error on the change
in stocks would have been higher.
When the net changes in biomass are split between
deforestation and degradation, it is clear that most of
the uncertainty in ΔC comes from degradation. It is
much easier to detect change in the small area (75 km
that was deforested and which lost 92 ±11 GgC, com-
pared to the very large (1037 km
) area that was
degraded, which probably lost a comparable amount of
AGB (94 ±90 GgC) (Table 2). The nature of the chal-
lenge is illustrated by a best guess assessment of the
change in C density for the forest-remaining-forest area
from 30.3 to 29.4 MgC ha
However, at a pixel scale, abrupt change (including
degradation) is detectable at 95% confidence as long as
it exceeds 12 MgC ha
. Per pixel change detection
rates were 65%, and in 89% of change events some
change was detected. These rates are comparable to
another application of L-band SAR to detect >2 ha clear
cuts in Sweden (hit rate =76% (Pantze et al., 2009)).
However it is clear that we are at the limit of the resolu-
tion of this technique the ALOS imagery detects new
Radar-derived change
Optical/NIR imagery
10-May-08 28-Sep-09 01-Oct-10
28-Jun-09 01-Jul-10
25-Jun-08 16-May-10
0 0.5 1 2
SPOT-4 (20m)IKONOS (1m)LANDSAT-5 (30m)
Fig. 6 Detail of Fig. 5, showing the detected land cover change (LCC) events. The temporal evolution of forest clearance as detected by
the radar imagery is shown from 2007 to 2010 (left panels). The right hand panels show false-colour optical imagery of the same area
for comparison. The outlines of the change events up to July 2009 are overlaid on the middle right image.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
10 C. M. RYAN et al.
farms as having a bare centre and then a gradient of
biomass into the surrounding woodland, but on-the-
ground observations show that the farm/woodland
boundary is normally abrupt. Morphological detection
techniques have the potential to improve the very sim-
ple per-pixel change detection used here and to reduce
0 20 40 60 80
0 20 40 60 80
0 20 40 60 80
0 20 40 60 80
Remote park
0 20 40 60 80
0 20 40 60 80
0 20 40 60 80
C ha–1)
Pre−change AGB of deforested pixels (C1)
Pre−change AGB of pseudodeforested pixels (C1*)
(where probability of deforestation is random wrt to AGB)
(a) (b)
(c) (d)
Fig. 7 Probability density functions of aboveground biomass in 2007 (black solid line) and 2010 (grey solid line). Also shown is the
aboveground biomass (AGB) of changed pixels prior to change (dashed lines). The observed values (grey dashed lines), and
the expected value if deforestation was random with respect to biomass (black dashed line) are shown. The 2007 and 2010 data are the
mean of three images. The boundaries of each sub-area (ag) are shown in Fig. 1. No data are shown for the changed pixels in the
Sanctuary as the number of changed pixels (n=10) was to few to produce a meaningful probability distribution function.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
false positives. Such techniques may also be more accu-
rate in determining the size of each LCC event.
Backscatter-biomass relationships for spaceborne
L-band backscatter in ‘difficult’ conditions (i.e. mixed
age stands of multiple species) are reported with similar
statistics to that found here (e.g. R
=0.53 (Karjalainen
et al., 2009)) and slightly better results have been found
in even-aged stands (RMSE =30% of biomass (Magnus-
son et al., 2007)), or when considering a very wide range
of biomass (Mitchard et al., 2009). In comparison, air-
borne LiDAR has regression statistics between the
LiDAR metrics and AGB of RMSE of 23.5 MgC ha
=0.85 in an example from Peru (Asner et al.,2010).In
the present study the comparable figures were
RMSE =8.7 MgC ha
and R
=0.49, but the difference
in RMSE is probably due to the lower AGB of our study
Williams et al. (2008) found that regrowing woodland
at this site can accumulate between 0.4 and
0.9 MgC ha
AGB, and we expect that changes
in more mature woodland will be less than this. How-
ever, in two parts of the study area (Fig. 5A) rapid
increases in AGB beyond this rate are observed. Obser-
vations in this area show rapid woody encroachment,
probably the result of fire exclusion in the Sanctuary.
Stem numbers have approximately doubled in 4 years
(data not shown), but biomass has not increased by that
rate because the new stems are small saplings. This
unrealistic sensing of biomass accumulation raises the
issue of which (combination of) plot characteristics L-
band backscatter is most strongly related to stem
number, basal area or biomass (Woodhouse, 2006b)? In
mixed age, diverse sites this question has not yet been
resolved (Woodhouse, 2006b). If the biomass-backscat-
ter relationship is not direct, but is mediated by other
stand characteristics (Lucas et al., 2010), then the site-
specificity of the regression parameters needs to be
evaluated (Mitchard et al., 2009).
How much carbon is lost to deforestation compared
to degradation?
The large uncertainty on the degradation loss makes
it difficult to estimate the fraction of AGB losses attrib-
utable to degradation. However, a best guess is that
total degradation was 67% of total net losses of AGB.
This analysis thus provides tentative support for the
idea of a large and presently un-quantified loss of bio-
mass due to degradation in African woodlands and for-
ests. However, we caution against extrapolating this
change in AGB to emissions, because, firstly, it is likely
that many of the areas that are degraded are subse-
quently deforested (Ahrends et al., 2010), and secondly,
the time period over which a change in standing stocks
translates into a flux to the atmosphere is currently
unknown in this ecosystem.
Loss of AGB in our study area (0.43 MgC ha
was lower than has been reported from plot data in
the surrounds of Dar es Salaam (0.8 MgC ha
(Ahrends et al., 2010)). This is expected given that our
study site is ~100 km by road from the nearest city,
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1
Miss rate, %
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1
P value
False positive rate, %
Fig. 8 Miss rates and false positives for the land cover change (LCC) detection algorithm under a variety of definitions of change. The
P-value shown on the x-axis is the threshold probability above which a change event is considered real. The various lines indicate φ,
the fraction of biomass remaining after change, below which the pixel’s biomass needs to be reduced for the change to be detected. A
P-value of 0.95 and φ=50% were used as default values.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
12 C. M. RYAN et al.
Chimoio, which has a much smaller population and
thus pressure for resource extraction is likely to be
The distinction between deforestation and degrada-
tion is logical in the context of the area multiplied by C
density approach to C stock estimation. However, the
advent of high resolution maps of C stocks and their
change, where the resolution is similar to the area of a
single tree canopy, means that the distinction becomes
ambiguous in effect degradation is just very small-
scale deforestation. It may be more useful to character-
ize a LCC regime in terms of the frequency, area (and
shape), and intensity of biomass change, as these
parameters might be more easily related to the extent of
different land uses, which is normally the information
policy makers require.
• Are areas of high carbon density preferentially tar-
geted for LCC?
The difference between the AGB of land that under-
went LCC and the mean of comparable surrounding
woodland was <1 MgC ha
, suggesting there is no
substantial preference for LCC to be undertaken in
high biomass areas (Fig. 7). There are some indica-
tions that this may not be true in all areas (such as
Vunduzi), but for now, the hypothesis that emissions
might be substantially higher as a result of preferen-
tial deforestation of high AGB areas has not been sup-
ported. This suggests that the other criteria for the
location of new farms dominate (such as proximity to
existing farms, roads or dwellings) or that AGB is not
correlated to the agricultural suitability of the land.
The latter might be true even if potential AGB and soil
fertility are correlated because of past land use or
other disturbance to AGB.
What is the carbon density (MgC ha
) of changed
areas before and after LCC?
After a change event, land was not reduced to zero
AGB, but instead averaged 11.9 ±18.4 MgC ha
. This
result fits with observations on the ground, where large
trees are often left in newly cleared areas to provide
shade, or because of the disproportionate effort needed
to remove them (Ghee, 2009). In addition, trees on
deforested land are often ringbarked or otherwise
killed, but left standing, and as they dry out will scatter
radar waves less as their dielectric properties change.
However, there is evidence of non-linearity in the AGB-
backscatter regression at low values of backscatter, so
these measurements of very low AGB values may be
subject to additional uncertainties. This may be because
at very low biomass, backscatter will be strongly influ-
enced by the properties of the ground surface, includ-
ing surface roughness and soil moisture.
• What size are LCC events, and how are they clus-
tered in space?
In Central Mozambique, farmers have an arable area
of 1.4 ha on average (Simler et al., 2004). However,
most LCC events detected in this study were much
smaller than this (mean 0.45 ha), indicating that farms
are built up in size by repeated small clearances
(Fig. 6). The significant spatial clustering of LCC events
supports this. This ‘death by a thousand cuts’ type of
woodland clearance poses a stiff challenge for rapid
monitoring of LCC: a detection system that ignored
events <1 ha would miss the areal majority of LCC in
this study area, and so might need to operate over long
time scales to be effective.
Despite only using dry season imagery, there is consid-
erable variation through time in the backscatter values
of the ALOS PALSAR imagery, and the slope of the
backscatter-AGB regression. This is presumably due to
variations in environmental conditions such as soil
moisture, which is well known to influence L-band
backscatter (Rignot et al., 1994; Pulliainen et al., 1999;
Magnusson et al., 2007; Pantze et al., 2009; Lucas et al.,
2010), changes to understory vegetation, or sensor cali-
bration drift. This variation suggests that accurate bio-
mass change detection with this sensor will require
either: invariant features to be present in each scene;
correction for e.g. the effects of soil moisture with ancil-
lary data sets and models; or recalibration to ground
data at each time point. We adopted the latter approach
here, but with an inventory that was not repeated each
year, and thus an assumption that plot biomass has not
changed during the study period.
Furthermore, there are potentially significant uncer-
tainties associated with the stem diameter-biomass
allometry (Chave et al., 2004) even with the use of a
site-specific model. This can introduce bias of ~17%
(Ryan, 2009). Although the regression procedure used
here accounts for the random error, the bias will
remain and influence estimates of AGB stocks.
It should also be noted that the L-band biomass-back-
scatter response is known to saturate at the level of bio-
mass typically found in closed canopy forests. Studies
report saturation at between 30 and 100 MgC ha
(Mitchard et al., 2009; Lucas et al., 2010), although this
appears to depend on vegetation structure (Lucas et al.,
2010). Thus the method presented here which utilizes
L-band will be most useful in low C density woodlands
rather than higher density forests. The proposed
BIOMASS mission would employ P-band radar in an
effort to overcome this limitation (Le Toan et al., 2011).
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
• Multi-temporal L-band radar imagery can be effec-
tive in detecting small-scale deforestation, but may
fail to detect small levels of degradation over large
areas. Abrupt changes of more than 12 MgC ha
are detectable at 95% confidence on a 25 925 m
pixel scale.
In this area of Mozambique, LCC events are mostly
small (median = 0.2 ha) but range from 0.06 to 18 ha
and reduce aboveground carbon from 33.5 to
11.9 MgC ha
on average. They are strongly clus-
tered together, and many events are the expansion of
previously cleared land.
There was no evidence that the areas that were defor-
ested had a higher biomass than the average sur-
rounding woodland.
Degradation losses are likely to be substantial, with a
best guess that they represent 67% of the net biomass
loss. This number is extremely uncertain however.
The Nhambita community and staff of Envirotrade facilitated
fieldwork. Staff of the Gorongosa national park allowed
access to plots within the park. The GeoEye Foundation pro-
vided the IKONOS image. ESA and JAXA provided the
ALOS imagery (C1P.7493). CR was supported by the UK
NERC Carbon Fusion project, the Mpingo Conservation &
Development Initiative ( under
their REDD Pilot Project funded by the Royal Norwegian
Embassy in Tanzania, and the EU FP7 iREDD+project. EM is
funded by Gatsby Plants. We also thank Iain Cameron for
his advice and help. We thank the two anonymous reviewers
for their helpful comments.
Achard F, Defries R, Eva H et al. (2007) Pan-tropical monitoring of deforestation.
Environmental Research Letters,2, 045022.
Ahrends A, Burgess ND, Milledge SAH et al. (2010) Predictable waves of sequential
forest degradation and biodiversity loss spreading from an African city. Proceed-
ings of the National Academy of Sciences,107, 1455614561.
Almeida-Filho R, Rosenqvist A, Shimabukuro YE et al. (2005) Evaluation and per-
spectives of using multitemporal L-band SAR data to monitor deforestation in the
Brazilian Amazo
ˆnia. Geoscience and Remote Sensing Letters, IEEE,2, 409412.
Archibald S, Scholes RJ (2007) Leaf green-up in a semi-arid African savanna: separat-
ing tree and grass responses to environmental cues. Journal of Vegetation Science,
18, 583594.
ASF (2010) ASF MapReady User Manual. Alaska Satellite Facility Engineering Group,
Asner GP, Powell GVN, Mascaro J et al. (2010) High-resolution forest carbon stocks
and emissions in the Amazon. Proceedings of the National Academy of Sciences,107,
Burgess N, Doggart N, Lovett JC (2002) The Uluguru Mountains of eastern Tanzania:
the effect of forest loss on biodiversity. Oryx,36, 140152.
Chave J, Condit R, Aguilar S et al. (2004) Error propagation and scaling for tropical
forest biomass estimates. Philosophical Transactions of the Royal Society (B),359, 409
Ciais P, Bombelli A, Williams M et al. (2011) The carbon balance of Africa: synthesis
of recent research studies. Philosophical Transactions of the Royal Society (A),369,1
Defries RS, Rudel T, Uriarte M et al. (2010) Deforestation driven by urban population
growth and agricultural trade in the twenty-first century. Nature Geoscience,3, 178181.
Denman KL, Brasseur G, Chidthaisong A et al. (2007) Couplings between changes in
the climate system and biogeochemistry. In: Climate Change 2007: The Physical Sci-
ence Basis (eds Solomon S, Qin D, Manning M, Chen Z, Marquis M, Averyt KB,
Tignor M, Miller HL), pp. 499587. Cambridge University Press, Cambridge.
Dewees PA, Campbell BM, Katerere Y et al. (2010) Managing the miombo woodlands
of Southern Africa: policies, incentives and options for the rural poor. Journal of
Natural Resources Policy Research,2,5773.
Etter A, Mcalpine C, Phinn S et al. (2006) Characterizing a tropical deforestation
wave: a dynamic spatial analysis of a deforestation hotspot in the Colombian
Amazon. Global Change Biology,12, 14091420.
FAO (2010) Global Forest Resources Assessment 2010. FAO Forestry Papers, 163.
Food and Agriculture Organization of the United Nations, Rome.
Farr TG, Rosen PA, Caro E et al. (2007) The shuttle radar topography mission. Reviews
of Geophysics,45, RG2004.
Fisher B (2010) African exception to drivers of deforestation. Nature Geoscience,3, 375
Fransson JES, Magnusson M, Olsson H et al. (2007) Detection of forest changes using
ALOS PALSAR satellite images. In: Geoscience and Remote Sensing Symposium, 2007.
IGARSS 2007. IEEE International.
Frost P (1996) The ecology of miombo woodlands. In: The Miombo in Transition: Wood-
lands and Welfare in Africa (ed. Campbell BM), pp. 1155. Center for International
Forestry Research, Bogor, Indonesia.
Furley PA, Rees RM, Ryan CM et al. (2008) Savanna burning and the assessment of
long-term fire experiments with particular reference to Zimbabwe. Progress In
Physical Geography,32, 611634.
Geist HJ, Lambin EF (2002) Proximate causes and underlying driving forces of tropi-
cal deforestation. BioScience,52, 143150.
Ghee C (2009) Nitrogen and carbon dynamics in Mozambican smallholder agroforestry sys-
tems. MSc thesis, University of Edinburgh, Edinburgh.
GOFC-GOLD (2010) A Sourcebook of Methods and Procedures for Monitoring and Report-
ing Anthropogenic Greenhouse Gas Emissions and Removals Caused by Deforestation,
Gains and Losses of Carbon Stocks in Forests Remaining Forests, and Forestation. GOFC-
GOLD Project Office, Alberta.
Grainger A (1999) Constraints on modelling the deforestation and degradation of
tropical open woodlands. Global Ecology and Biogeography,8, 179190.
Grainger A (2008) Difficulties in tracking the long-term global trend in tropical forest
area. Proceedings of the National Academy of Sciences,105, 818823.
Hansen MC, Shimabukuro YE, Potapov P et al. (2008) Comparing annual MODIS and
PRODES forest cover change data for advancing monitoring of Brazilian forest
cover. Remote Sensing of Environment,112, 37843793.
Hoekman DH, Vissers MaM, Wielaard N (2010) PALSAR wide-area mapping of Bor-
neo: methodology and map validation. IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing,3, 605617.
Houghton RA (2010) How well do we know the flux of CO
from land-use change?
Tellus B,62, 337351.
Houghton RA, Hackler JL (2006) Emissions of carbon from land use change in
sub-Saharan Africa. Journal of Geophysical Research Atmospheres,111, G02003.
Houghton RA, Hall F, Goetz SJ (2009) Importance of biomass in the global carbon
cycle. Journal of Geophysical Research-Biogeosciences,114, G00E03.
Hui D, Jackson RB (2007) Uncertainty in allometric exponent estimation: a case study
in scaling metabolic rate with body mass. Journal of Theoretical Biology,249, 168177.
INE (2010) Estatísticas do Distrito de Gorongosa. Instituto Nacional de Estatı
Maputo Mozambique.
Karjalainen M, Pyysalo U, Karila K et al. (2009) Forest biomass estimation using
ALOS PALSAR images in challenging natural forest area in Finland. In: ALOS PI
2008 Symposium. ESA Special Publication SP-664, Island of Rhodes, Greece.
Kindermann GE, Mcallum I, Fritz S et al. (2008) A global forest growing stock, bio-
mass and carbon map based on FAO statistics. Silva Fennica,42, 387396.
Le Toan T, Beaudoin A, Riom J et al. (1992) Relating forest biomass to SAR data. Geo-
science and Remote Sensing, IEEE Transactions on,30, 403411.
Le Toan T, Quegan S, Davidson MWJ et al. (2011) The BIOMASS mission: mapping
global forest biomass to better understand the terrestrial carbon cycle. Remote Sens-
ing of Environment,115, 28502860.
Loarie SR, Asner GP, Field CB (2009) Boosted carbon emissions from Amazon defor-
estation. Geophysical Research Letters,36, L14810.
Lucas R, Armston J, Fairfax R et al. (2010) An Evaluation of the ALOS PALSAR
L-Band BackscatterAbove Ground Biomass Relationship Queensland, Australia:
impacts of Surface Moisture Condition and Vegetation Structure. Selected Topics in
Applied Earth Observations and Remote Sensing, IEEE Journal of,3, 576593.
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
14 C. M. RYAN et al.
Magnusson M, Fransson JES, Eriksson LEB et al. (2007) Estimation of forest stem vol-
ume using ALOS PALSAR satellite images. In: Geoscience and Remote Sensing Sym-
posium, 2007. IGARSS 2007. IEEE International.
Mayaux P, Bartholome
´E, Fritz S et al. (2004) A new land-cover map of Africa for the
year 2000. Journal of Biogeography,31, 861877.
Mertz O (2009) Trends in shifting cultivation and the REDD mechanism. Current
Opinion in Environmental Sustainability,1, 156160.
Miettinen J, Shi C, Liew SC (2011) Deforestation rates in insular Southeast Asia
between 2000 and 2010. Global Change Biology,17, 22612270.
Mitchard E (2011) Using satellite remote sensing to quantify woody cover & biomass across
Africa. PhD thesis. University of Edinburgh, Edinburgh.
Mitchard ETA, Saatchi SS, Woodhouse IH et al. (2009) Using satellite radar backscat-
ter to predict above-ground woody biomass: a consistent relationship across four
different African landscapes. Geophysical Research Letters,36, L23401.
Nepstad DC, Verssimo A, Alencar A et al. (1999) Large-scale impoverishment of
Amazonian forests by logging and fire. Nature,398, 505508.
Pantze A, Krantz AH, Fransson JES et al. (2009) Mapping and monitoring clear-cuts
in Swedish forest using ALOS PALSAR satellite images. Geoscience and Remote
Sensing Symposium, 2009 (IGARSS 2009) IEEE International, IGARSS 2009, pp. 589
Pulliainen JT, Kurvonen L, Hallikainen MT (1999) Multitemporal behavior of L- and
C-band SAR observations of boreal forests. IEEE Transactions on Geoscience and
Remote Sensing,37, 927937.
Radke RJ, Andra S, Al-Kofahi O et al. (2005) Image change detection algorithms: a
systematic survey. IEEE Transactions on Image Processing,14, 294307.
Rahman M, Sumantyo J (2010) Mapping tropical forest cover and deforestation using
synthetic aperture radar (SAR) images. Applied Geomatics,2, 113121.
Ramankutty N, Gibbs HK, Achard F et al. (2007) Challenges to estimating carbon
emissions from tropical deforestation. Global Change Biology,13,51
Rignot EJM, Van Zyl JJ (1993) Change detection techniques for ERS-1 SAR data. IEEE
Transactions on Geoscience and Remote Sensing,31, 896906.
Rignot E, Way J, Williams C et al. (1994) Radar estimates of aboveground biomass in
boreal forests of interior Alaska. IEEE Transactions on Geoscience and Remote Sensing,
32, 11171124.
Ryan CM (2009) Carbon cycling, fire and phenology in a tropical Savanna woodland in
Nhambita, Mozambique. PhD thesis, University of Edinburgh, Edinburgh, 257 pp.
Ryan CM, Williams M (2011) How does fire intensity and frequency affect miombo
woodland tree populations and biomass? Ecological Applications,21,4860.
Ryan CM, Williams M, Grace J (2011) Above- and Belowground Carbon Stocks in a
Miombo Woodland Landscape of Mozambique. Biotropica,43, 423432.
Schmidt-Vogt D, Leisz S, Mertz O et al. (2009) An assessment of trends in the extent
of Swidden in Southeast Asia. Human Ecology,37, 269280.
Shimada M, Isoguchi O, Tadono T et al. (2009) PALSAR Radiometric and Geometric
Calibration. Geoscience and Remote Sensing. IEEE Transactions on,47, 39153932.
Simler KR, Mukherjee S, Dava GL et al. (2004) Rebuilding After War: Micro-Level Deter-
minants of Poverty Reduction in Mozambique. International Food Policy Research
Institute, Washington, DC.
Sokal RR, Rohlf FJ (1995) Biometry: The Principles and Practice of Statistics in Biological
Research. W.H. Freeman, New York.
Tinley KL (1977) Framework of the Gorongosa Ecosystem. DSc thesis, University of Preto-
ria, Pretoria, 184 pp.
Tinley KL (1982) The influence of soil moisture balance on ecosystem patterns in
Southern Africa. In: Ecology of Tropical Savannas. (eds Huntley BJ, Walker BH).
Springer, Berlin.
Trujillo-Ortiz A, Hernandez-Walls R (2010) Geometric Mean Regression (Reduced Major
Axis Regression). Available at:
change/27918-gmregress (accessed 11 October 2011).
Van Der Sanden JJ, Hoekman DH (1999) Potential of airborne radar to support the
assessment of land cover in a tropical rain forest environment. Remote Sensing of
Van Der Werf GR, Morton DC, Defries RS et al. (2009) CO
emissions from forest loss.
Nature Geosciences,2, 737738.
Williams C, Hanan N, Neff J et al. (2007) Africa and the global carbon cycle. Carbon
Balance and Management,2,3.
Williams M, Ryan CM, Rees RM et al. (2008) Carbon sequestration and biodiversity
of re-growing miombo woodlands in Mozambique. Forest Ecology And Manage-
ment,254, 145155.
Woodhouse IH (2006a) Introduction to Microwave Remote Sensing. Taylor & Francis,
New York, London.
Woodhouse IH (2006b) Predicting backscatter-biomass and height-biomass trends
using a macroecology model. IEEE Transactions on Geoscience and Remote Sensing,
44, 871877.
Supporting Information
Additional Supporting Information may be found in the
online version of this article:
Appendix S1. Supplementary methods and results.
Please note: Wiley-Blackwell are not responsible for the con-
tent or functionality of any supporting materials supplied
by the authors. Any queries (other than missing material)
should be directed to the corresponding author for the
©2011 Blackwell Publishing Ltd, Global Change Biology, doi: 10.1111/j.1365-2486.2011.02551.x
... Mitchard et al. (2011) was able to detect deforestation and woody encroachment along forest--savanna ecotone of central Cameroon by comparing Advanced Land Observing Satellite (ALOS) Phased Arrayed L-band Synthetic Aperture Radar (PALSAR) data with L-band JERS-1 data acquired 11 years before. Ryan et al., 2012 detected land cover change and quantified the carbon stock losses of 11-33 MgC/ha in Mozambique woodlands using ALOS PALSAR data over a three-year period. Mitchard et al., 2013 detected increases in woodland carbon of 1.1 MgC/ha/year and decreases of 3 MgC/ha/year in Mozambique with ALOS PALSAR, but highlighted the need for more rigorous quantification of uncertainty. ...
... If we express these cover changes as a rate over, for example, 10 years, this would amount to ±0.025, ± 0.05 and > 0.05 fractional cover change per year, respectively. Several studies working in lower biomass woodlands in Africa have used a variety of methods to suggest that L-band SAR could detect relatively subtle biomass changes (<30 Mg/ha) (Mitchard et al., 2013;Odipo et al., 2016;Ryan et al., 2012;Wessels et al., 2019). Using a forest cover change error model, Cartus et al., 2018 found that L-band HV backscatter should be able to detect 50% cover change at hectare scale in mature forests, but would be unlikely to detect 50% cover change at lower 5 m heights of 5 m (similar to savannas), as error probability increases with decreasing height. ...
Global savannas are the third largest carbon sink with large human populations being highly dependent on their ecosystem services. However, savannas are changing rapidly due to climate change, fire, animal management, and intense fuelwood harvesting. In southern Africa, large trees (>5 m in height) are under threat while shrub cover (<3 m) is increasing. The collection of multi-date airborne LiDAR (ALS) data, initiated over a decade ago in the Lowveld of South Africa, provided a rare opportunity to quantify the ability of L-band SAR to track changes in savanna vegetation structure and this study is the first to do so, to our knowledge. The objective was to test the ability of ALOS PALSAR 1&2, dual-pol (HH, HV) data to quantify woody cover and volume change in savannas over 2-, 8- and 10-year periods through comparison to ALS. For each epoch (2008, 2010, 2018), multiple PALSAR images were processed to Gamma0 (γ⁰) at 15 m resolution with multi-temporal speckle filtering. ALS data were processed to fractional canopy cover and volume, and then compared to 5 × 5 aggregated (75 m) SAR mean γ⁰. The ALS cover change (∆CALS) and volume change between pairs of years were highly correlated, with (R² > 0.8), thus results for cover change applied equally to volume change. Cover change was predicted using (i) direct backscatter change or (ii) the difference between annual cover map product derived using the Bayesian Water Cloud Model (BWCM) and logarithmic models. The linear relationship between ∆γ⁰ and ∆CALS varied between year pairs but reached a maximum R² of 0.7 for 2018–2010 and a moderate R² of 0.4 for 2018–2008. Overall, 1 dB ∆γ⁰ corresponded to approximately 0.1 cover change. The three cover change models had very similar uncertainties with mean RMSE = 0.15, which is 13% of the observed cover change range (−0.6 to +0.6). The direct backscatter change approach had less underestimation of positive and negative cover change. The L-band backscatter had a higher sensitivity than suggested by previous studies, as it was able to reliably distinguish cover change at 0.25 increments. The SAR-derived cover change maps detected the loss of stands of big trees, and widespread increases in cover of 0.35–0.65 in communal rangelands due to shrub encroachment. In contrast, the maps suggest that cover generally decreased in conservation areas, forming distinct fence-line effects, potentially caused by significant increases in elephant numbers and frequent, intense wildfires in reserves.
... However, they do not provide information on changes in forest structure, for example related to biomass values or tree canopy heights. Previous research has related L-band backscatter signal to changes in AGB using a regional empirical regression model and bi-temporal ALOS PALSAR imagery (Ryan et al., 2012). Above ground biomass was estimated at the two time points, and then the difference in AGB was calculated by subtracting the two estimates. ...
... Previous work had successfully correlated decreases in Lband radar backscatter to structural changes in vegetation cover (Mitchard et al., 2011a;Ryan et al., 2012;Joshi et al., 2015). However, biomass estimation in tropical forests is inherently limited by the saturation of the radar signal (Mermoz et al., 2015). ...
Full-text available
In the last decades tropical forests have experienced increased fragmentation due to a global growing demand for agricultural and forest commodities. Satellite remote sensing offers a valuable tool for monitoring forest loss, thanks to the global coverage and the temporal consistency of the acquisitions. In tropical regions, C-band Synthetic Aperture Radar (SAR) data from the Sentinel-1 mission provides cloud-free and open imagery on a 6- or 12-day repeat cycle, offering the unique opportunity to monitor forest disturbances in a timely and continuous manner. Despite recent advances, mapping subtle forest losses, such as those due to small-scale and irregular selective logging, remains problematic. A Cumulative Sum (CuSum) approach has been recently proposed for forest monitoring applications, with preliminary studies showing promising results. Unfortunately, the lack of accurate in-situ measurements of tropical forest loss has prevented a full validation of this approach, especially in the case of low-intensity logging. In this study, we used high-quality field measurements from the tropical Forest Degradation Experiment (FODEX), combining unoccupied aerial vehicle (UAV) LiDAR, Terrestrial Laser Scanning (TLS), and field-inventoried data of forest structural change collected in two logging concessions in Gabon and Peru. The CuSum algorithm was applied to VV-polarized Sentinel-1 ground range detected (GRD) time series to monitor a range of canopy loss events, from individual tree extraction to forest clear cuts. We developed a single change metric using the maximum of the CuSum distribution, retrieving location, time, and magnitude of the disturbance events. A comparison of the CuSum algorithm with the LiDAR reference map resulted in a 78% success rate for the test site in Gabon and 65% success rate for the test site in Peru, for disturbances as small as 0.01 ha in size and for canopy height losses as fine as 10 m. A correlation between the change metric and above ground biomass (AGB) change was found with R ² = 0.95, and R ² = 0.83 for canopy height loss. From the regression model we directly estimated local AGB loss maps for the year 2020, at 1 ha scale and in percentages of AGB loss. Comparison with the Global Forest Watch (GFW) Tree Cover Loss (TCL) product showed a 61% overlap between the two maps when considering only deforested pixels, with 504 ha of deforestation detected by CuSum vs. 348 ha detected by GFW. Low intensity disturbances captured by the CuSum method were largely undetected by GFW and by the SAR-based Radar for Detecting Deforestation (RADD) Alert System. The results of this study confirm this approach as a simple and reproducible change detection method for monitoring and quantifying fine-scale to high intensity forest disturbances, even in the case of multi-storied and high biomass forests.
... These hold significance in understanding CBFM forests' contribution in leveraging forest management for broader national and international environmental goals. To assess forest biomass, we used imagery collected by the Phased Array Type L-band Synthetic Aperture RADAR (PALSAR 1 and 2) instruments carried by the Advanced Land Observing Satellites (ALOS and ALOS-2), which have often been used to assess tropical woody biomass changes (McNicol et al., 2018b;Mitchard et al., 2009;Ryan et al., 2012). They are particularly useful as they have the capacity to penetrate a forest canopy, providing information about forest structure. ...
Participatory forest management (PFM) has been applied to address declining tropical forest conditions. In the literature, there is a mixed evidence on PFM's role in improving forest conditions. However, most assessments ignore a relationship between household distance from PFM forests and impacts to non-PFM forests albeit being an essential aspect. Some PFM assessments show that distance matters in determining an individual's participation But, sparsely discussing the distance in relation to forest biophysical conditions in a landscape context. Drawing on the landscape approach and insights from Miombo forests of Kilwa in southern Tanzania, we illustrate the importance of studying PFM schemes in a landscape context to illuminate the relationship between household distance from PFM forests and impacts to non-PFM forests. Our study villages have forest abundances in areas between households and PFM forests. The average distance between households and PFM forests is 7.8 km. The long distances and forest abundances produce an ‘outbound effect’, whereby degrading and deforesting activities shift from PFM to non-PFM forests. Our analysis calls for landscape-level assessments that include forests under different governance regimes even those in unreserved landscapes – non-PFM forests. This is important for two reasons. First, for generating locally grounded contextual insights necessary for developing understandings of global forest conservation efforts from the ‘ground up’. Second, for revealing correct forest conditions in the entire landscape critical in the light of ongoing national and international interests to manage trees inside and outside designated forest reserves for both carbon sequestration and landscape restoration.
... Paradoxically, this is against the backdrop that Malawi started piloting REDD+ schemes as early as 2008, besides officially embracing the UN REDD+ partnership in 2014 [2,5,19]. Ryan et al. [26] observed that data on land use change emissions from the African Woodlands are generated from unreliable and inconsistent sources which have consequently provided difficulties in making comparisons or used under global REDD+ purposes [27]. ...
Full-text available
This study assessed forest inventory methods and standard operating procedures (SOPs) for estimating above-ground biomass (AGB) and carbon, as employed in the key REDD+ Miombo Forest Reserves of Malawi. Analysis of Variance statistical technique was applied to investigate the following methods and SOPs: (i) allometry, (ii) sample plot configuration, and (iii) dendrometric measurements. Results indicate that the allometric equations parameter significantly (P<0.001) affected the AGB estimations and was the highest contributor (97.95%) of the total variation. Malawi's specific allometry provided the highest AGB estimate (113.08±1.56 t/ha). In contrast, the Pan-Tropical/generalized allometric models substantially underestimated AGB within the range of 16.7-67.9%. Furthermore, the findings demonstrate that the use of varied sampling plot sizes significantly (P<0.001) affected the estimates of AGB. However, the plot size parameter contributed only 1.65% to the total variation. The 20m radius plot size registered the highest AGB (75.31±0.77 t/ha) compared to the 17.84m radius plot (66.12±1.61 t/ha). This signifies that the plot size of 17.84m radius underestimated the AGB by 12.2%. However, results on dendrometric measurements showed no significant (P>0.05) differences in the AGB estimates between the use of diameter tape (D-tape) and calliper in measuring dbh of individual trees despite the former yielding higher estimates of AGB (74.65±0.93 t/ha) than the latter (72.53±0.98 t/ha). This demonstrates that the use of calliper in measuring dbh underestimated AGB t/ha by only 2.8% compared to the use of D-tape. Therefore, the study recommends; employment of local allometry, adoption of a circular sampling design of 20m radius, and consistent use of D-tape in measuring dbh for AGB in Malawi's Miombo Woodlands. In conclusion, incorporation of these changes is envisaged to facilitate quick realisation of Malawi's REDD+ carbon payments, smooth running of the National Forest Inventory system, robust implementation and global recognition of the REDD+ efforts.
... Most of the studies on mapping ES have assessed the provision of regulating services such as carbon storage (Batjes, 2008;Egoh et al., 2011;Leh et al., 2013), water flow regulation (Egoh et al., 2008), soil accumulation and retention (Egoh et al., 2011;Leh et al., 2013). Mapping and quantifying woody biomass provided by forest ecosystems, when lacking primary data such as national forest inventories, relies on indicators derived from remote sensing data combined with ground observations (DeFries et al., 2007;Ryan et al., 2012). Therefore, to quantify and map the total biomass supplied by the forest ecosystem analysed in this article, we combine high resolution remote sensing images that measure changes in vegetation cover with biomass estimates from ground observations. ...
Full-text available
Community based management (CBM) is widely advocated as an effective method for governing and managing ecosystem services (ES). However, the distributional rules and maximum harvesting levels are likely to affect both the effectiveness of CBMs in maintaining ES and the fairness and equity of access to these ES. This article proposes a methodological approach for investigating normative trade-offs involved in CBM of forests, where forest conservation objectives need to be traded off against livelihoods objectives. The study uses remote sensing methods to quantify forest ES supply in Namizimu Forest Reserve in Malawi, and links this to demand for ES within the villages near the reserve. It then investigates how a plausible set of CBM rules can be developed to cap consumption of forest products to sustainable amount and quantifies, by using monetary valuation techniques, how these set of rules may affect the total well-being of local population. Our results demonstrate that, due to the spatial mismatches between demand and supply, the distribution of provisioning ES to the population across the harvesting area is unequal in biophysical terms. The current available stock of forest products is sufficient to cover the current demand, however, it is higher than the mean annual increment indicating that this level of consumption is ecologically unsustainable and will lead to forest degradation as shown under the business-as-usual scenario. We then examined the impact of governance and how CBM rules to allocate forest ES to different social groups (poor and rich) under a co-management regime will affect total societal welfare. We found that the distributional scenario that maximises total societal welfare expressed in monetary terms across the whole harvesting area is the scenario that distributes 40% of biomass to the rich group while the remaining 60% is allocated to the poor group. However, this scenario maximises Willingness to Pay (WTP) at total level but does not maximise WTP in each sub-area of forest but just for those that have a high availability for biomass. This indicates that the distributional rules that maximise total welfare at aggregate level may not maximise welfare at local level where constraints from biomass availability require to restrict further the distribution of forest products. When biomass availability is low, total societal welfare is maximised with distributional rules that distribute more trees to richer. Yet, a policymaker may choose a distributional rule that distribute more trees to the poor on normative grounds and forego the objective of maximising total welfare. In such cases the WTP analysis outlined in this paper can support the policymaker in choosing the distributional rule that minimise trade-offs between efficiency, i.e., maximising total welfare, and livelihoods objectives.
... However, incorporating C-band SAR only minimally improved the classification above what was possible when using optical imagery alone (Fig. 3). Utilizing L-band SAR, where freely available (for instance the PALSAR-2/PALSAR Mosaics 126 ) could potentially improve vegetation type separation capability further, by providing information about woody vegetation structure and biomass gradients [127][128][129] . Furthermore, the addition of texture imagery resulted in systematic misclassifications of some physiognomies which were not captured in our error metrics, despite appearing to improve classification OA 98 . ...
Full-text available
Native vegetation across the Brazilian Cerrado is highly heterogeneous and biodiverse and provides important ecosystem services, including carbon and water balance regulation, however, land-use changes have been extensive. Conservation and restoration of native vegetation is essential and could be facilitated by detailed landcover maps. Here, across a large case study region in Goiás State, Brazil (1.1 Mha), we produced physiognomy level maps of native vegetation (n = 8) and other landcover types (n = 5). Seven different classification schemes using different combinations of input satellite imagery were used, with a Random Forest classifier and 2-stage approach implemented within Google Earth Engine. Overall classification accuracies ranged from 88.6–92.6% for native and non-native vegetation at the formation level (stage-1), and 70.7–77.9% for native vegetation at the physiognomy level (stage-2), across the seven different classifications schemes. The differences in classification accuracy resulting from varying the input imagery combination and quality control procedures used were small. However, a combination of seasonal Sentinel-1 (C-band synthetic aperture radar) and Sentinel-2 (surface reflectance) imagery resulted in the most accurate classification at a spatial resolution of 20 m. Classification accuracies when using Landsat-8 imagery were marginally lower, but still reasonable. Quality control procedures that account for vegetation burning when selecting vegetation reference data may also improve classification accuracy for some native vegetation types. Detailed landcover maps, produced using freely available satellite imagery and upscalable techniques, will be important tools for understanding vegetation functioning at the landscape scale and for implementing restoration projects.
... For decades, empirical and semi-empirical relationships between radar cross section and forest canopies have been established for the retrieval of AGBC in varying ecosystems, in particular using long-wavelength radar, for example, from the Japanese Phased Array type L-band Synthetic Aperture Radar instruments aboard the Advanced Land Observing Satellite (ALOS-1/2 PALSAR-1/2) [8][9][10][11][12][13][14][15][16]. In African woodlands, previous work has also shown a strong biomass-backscatter relationship [17]. ...
Full-text available
Soil moisture effects limit radar-based aboveground biomass carbon (AGBC) prediction accuracy as well as lead to stripes between adjacent paths in regional mosaics due to varying soil moisture conditions on different acquisition dates. In this study, we utilised the semi-empirical water cloud model (WCM) to account for backscattering from soil moisture in AGBC retrieval from L-band radar imagery in central Mozambique, where woodland ecosystems dominate. Cross-validation results suggest that (1) the standard WCM effectively accounts for soil moisture effects, especially for areas with AGBC ≤ 20 tC/ha, and (2) the standard WCM significantly improved the quality of regional AGBC mosaics by reducing the stripes between adjacent paths caused by the difference in soil moisture conditions between different acquisition dates. By applying the standard WCM, the difference in mean predicted AGBC for the tested path with the largest soil moisture difference was reduced by 18.6%. The WCM is a valuable tool for AGBC mapping by reducing prediction uncertainties and striping effects in regional mosaics, especially in low-biomass areas including African woodlands and other woodland and savanna regions. It is repeatable for recent L-band data including ALOS-2 PALSAR-2, and upcoming SAOCOM and NISAR data.
... Tous ces différents aspects de changement d'affectation des terres influencent les facteurs d'émission (Diwediga et al., 2015 ;Polo-Akpisso et al., 2016). Parmi les cinq (5) (Luedeling & Neufeldt, 2012 ;Ryan et al., 2012). Les arbres qui existent dans les savanes présentent de petits diamètres et de petites tailles ; cause probable d'un potentiel de stockage de carbone moins important (Fousseni et al., 2019) par rapport aux autres formes d'utilisation des terres. ...
Full-text available
En Afrique de l'Ouest et particulièrement au Togo les écosystèmes aussi bien naturels qu'anthropisés sont pourvoyeurs de nombreux services et biens. Cette étude est une contribution à une meilleure gestion des ressources naturelles des paysages sous forte emprise des activités socioéconomique et très sensible aux effets des changements climatiques. Les données des formes d'utilisations des terres des années 1977, 1987, 1997, 2007, et 2017 ont été obtenues suite à une classification des images satellites de type Landsat. Ces données ont servi à évaluer l'empileur d'activités dans le socle grâce aux analyses géostatistique de type intersection et différence sous QGIS 3.81. Les facteurs d'émission des différentes formes d'utilisation des terres identifiées dans la zone d'étude ont été estimées grâce à l'équation allométrique de Chave après un inventaire forestier réalisé dans 99 placettes de 1 ha distribués suivant une maille de 5 km. Enfin les émissions liées à la déforestation et la dégradation des terres a été évaluées en multipliant les données d'activités par les différents facteurs d'émissions. Les résultats montrent que les facteurs d'émissions des savanes (Sa), parcs agroforestiers (PA), de la végétation marécageuse saisonnière (VMS), des jardins de case (JC) sont respectivement 0,66 t/ha, 1,69 t/ha, 1,35 t/ha et 0,58 t/ha. Il est constaté une forte conversion des savanes au profit des parcs agroforestiers évaluée à période de 1977 à 1987 les savanes qui ont demeuré ont stocké 22879,24 t.C-1. La conversion des savanes en parcs agroforestiers a occasionné une émission de 37185,07 t.C-1. Les émissions liées à la conversion des savanes en d'autres formes d'utilisations des terres entre 1977 et 2017 sont respectivement de 60626,15 t.C-1 (Pa), de 9743,09 t.C-1 (VMS), et de 289,58 t.C-1 (Jc). La préservation des écosystèmes s'avère indispensable pour toute politique d'optimisation du potentiel d'atténuation des émissions de gaz à effet de serre en l'occurrence CO2. L'étude bien qu'ayant un caractère local pourrait contribuer à la déclinaison des politiques liées aux mécanismes REDD+, NAMA, et MDP auxquels le Togo a souscrit ; tout en continuant son approfondissement. Mots Clés : Séquestration du carbone, affectation des terres, Emission de carbone, Togo. Abstract The accelerated landscape change in Togo is the source of various emissions linked to the allocation of forms of land use. This study is a contribution for a better management of natural resources in landscapes under strong influence of socioeconomic activities and very sensitive to the effects of climate change. Data on land use forms for the years 2007, and 2017 obtained following a classification of Landsat-type satellite images were used to evaluate the activity data in the base thanks to geostatistical analyzes of the intersection and difference type under QGIS 3.81. The emission factors of the different forms of land use identified in the study area were estimated using the allometric equation of Chave after a forest inventory carried out in 99 plots of Folega et al. 2021 Rev Écosystèmes et Paysages (Togo), 2021, N o 01, vol 01 ; 58-72pp 59 1 ha distributed according to a 5 km grid. Finally, the emissions linked to deforestation and land degradation were evaluated. The results show that the emission factors of savannas (Sa), agroforestry parks (PA), seasonal swamp vegetation (VMS), and home gardens (HG) are respectively 0.66 t.ha-1 , 1.69 t.ha-1 , 1.35 t.ha-1 and 0.58 t.ha-1. It is noted a strong conversion of savannas especially for the benefit of agroforestry parks over the entire time series. The loss of savannahs for agroforestry parks is There is also a gradual increase in the home gardens over the same period, confirming the highly anthropized nature of the landscape. The conversion of savannas into agroforestry parks resulted in an emission of 37 185.07 tEqC. This emission trend linked to the conversion of savannas to other form of land use is observed over the same time series. These emissions are respectively between 1977 and 2017 60,626.15 tEqC (Pa), 9743.09 tEqC (VMS), and 289.58 tEqC (HG). The study although having a local character could contribute to the declination of policies related to REDD+, NAMA, and MDP mechanisms to which Togo has subscribed; while enhancing substantial improvement.
Understanding changes to aboveground biomass (AGB) in forests undergoing degradation is crucial for accurately and completely quantifying carbon emissions from forest loss and for environmental monitoring in the context of climate change. Monitoring forest degradation as compared to deforestation presents technical challenges because degradation involves widespread, low-intensity AGB removal under varying temporal dynamics. Charcoal production is a key driver for forest degradation in Africa and is projected to increase in the future years. In Sub-Saharan Africa (SSA), where charcoal production drives widespread ABG removal, the utility of optical remote sensing for degradation quantification is challenged by the large inter-seasonal variation and high complexities in ecosystem structure. Limited field measurements on tree structure and aboveground biomass density (AGBD) in many parts of the SSA also impose constraints. In this study, we present a novel data fusion approach combining 3D forest structure from NASA's GEDI Lidar with optical time-series data from Landsat to quantify biomass losses associated with charcoal-related forest degradation over a 10-year time period. We used machine learning models with Landsat spectral indices from the time period of limited hydric stress (LHS) as predictor variables. By applying the best performing Random Forest (RF) model to LandTrendr-stabilized annual LHS Landsat composites, we produced annual forest AGBD maps from 2007 to 2019 over the Mabalane district in southern Mozambique where the dry forest ecosystem was under active charcoal-related degradation since 2008. The RF model achieved an RMSE value of 7.05 Mg/ha (RMSE% = 42%) and R² value of 0.64 using a 10-fold cross-validation dataset. We quantified a total AGB loss of 2.12 ± 0.06 Megatons (Mt) over the 10-year period, which is only 6.35 ± 2.56% less than the total loss estimated using field-based data as previously published for the same area and time. In addition to quantifying biomass loss, we constructed annual AGBD maps that enabled the characterization of disturbance and recovery. Our framework demonstrates that fusing GEDI and Landsat data through predictive modeling can be used to quantify past forest AGBD dynamics in low biomass forests. This approach provides a satellite-based method to support REDD+ monitoring and evaluation activities in areas where field data is limited and has the potential to be extended to investigate a variety of different disturbance events.
Full-text available
O mineral de ouro no distrito de Manica é extraído ao longo do rio Revué. As intensas actividades de extração e processamento deste mineral tem criado alterações nas classes de uso e cobertura de terras das áreas em que são praticadas as actividades garimpo. Vários estudos estão sendo desenvolvidos sobre os impactos negativos da mineração artesanal de ouro neste distrito sobre meio ambiente, não abordando variação das classes de uso de cobertura. Este trabalho visa mapear áreas degradadas pela mineração entre os anos de 2000 a 2019 na microbacia do Revue, estimando a quantidade de vegetação nativa suprimida. Para alcançar os objetivos propostos, foi realizado o mapeamento do uso e cobertura da terra através da classificação supervisionada e pelo cálculo do Índice de Vegetação por Diferença Normalizada de imagens de satélite Landsat 05 e Landsat 08 (OLI). As imagens foram pré-processadas e classificadas pelos Sistemas de Informação Geográfica (SPRING 5.5.2, ArcGIS 10.2.1 e QGIS 3.2.3). Os resultados mostraram que o atual padrão de uso e ocupação da terra é marcadamente caracterizado pela presença de áreas de mineração artesanal, industrial e pelo plantio de eucaliptos e pinheiros que se expandem em direção a áreas de vegetação nativa e de solo exposto.
Full-text available
Previous estimates of the flux of carbon from land use change in sub-Saharan Africa have been based on highly aggregated data and have ignored important categories of land use. To improve these estimates, we divided the region into four subregions (east, west, central, and southern Africa), each with six types of natural vegetation and five types of land use (permanent crops, pastures, shifting cultivation, industrial wood harvest, and tree plantations). We reconstructed rates of land use change and rates of wood harvest from country-level statistics reported by the Food and Agriculture Organization (FAO) (1961-2000) and extrapolated the rates from 1961 to 1850 on the basis of qualitative histories of demography, economy, and land use. We used a bookkeeping model to calculate the annual flux of carbon associated with these changes in land use. Country-level estimates of average forest biomass from the FAO, together with changes in biomass calculated from the reconstructed rates of land use change, constrained the average biomass of forests in 1850. Comparison of potential (predisturbance) forest areas with the areas present in 1850 and 2000 suggests that 60% of Africa's forests were lost before 1850 and an additional 10% lost in the last 150 years. The annual net flux of carbon from changes in land use was probably small and variable before the early 1900s but increased to a source of 0.3 +/- 0.2 PgC/yr by the end of the century. In the 1990s the source was equivalent to about 15% of the global net flux of carbon from land use change.
In response to the urgent need for improved mapping of global biomass and the lack of any current space systems capable of addressing this need, the BIOMASS mission was proposed to the European Space Agency for the third cycle of Earth Explorer Core missions and was selected for Feasibility Study (Phase A) in March 2009. The objectives of the mission are 1) to quantify the magnitude and distribution of forest biomass globally to improve resource assessment, carbon accounting and carbon models, and 2) to monitor and quantify changes in terrestrial forest biomass globally, on an annual basis or better, leading to improved estimates of terrestrial carbon sources (primarily from deforestation); and terrestrial carbon sinks due to forest regrowth and afforestation. These science objectives require the mission to measure above-ground forest biomass from 70 degrees N to 56 degrees Sat spatial scale of 100-200 m, with error not exceeding +/- 20% or +/- 10 t ha(-1) and forest height with error of +/- 4 m. To meet the measurement requirements, the mission will carry a P-Band polarimetric SAR (centre frequency 435 MHz with 6 MHz bandwidth) with interferometric capability, operating in a dawn-dusk orbit with a constant incidence angle (in the range of 25 degrees-35 degrees) and a 25-45 day repeat cycle. During its 5-year lifetime, the mission will be capable of providing both direct measurements of biomass derived from intensity data and measurements of forest height derived from polarimetric interferometry. The design of the BIOMASS mission spins together two main observational strands: (1) the long heritage of airborne observations in tropical, temperate and boreal forest that have demonstrated the capabilities of P-band SAR for measuring forest biomass; (2) new developments in recovery of forest structure including forest height from Pol-InSAR, and, crucially, the resistance of P-band to temporal decorrelation, which makes this frequency uniquely suitable for biomass measurements with a single repeat-pass satellite. These two complementary measurement approaches are combined in the single BIOMASS sensor, and have the satisfying property that increasing biomass reduces the sensitivity of the former approach while increasing the sensitivity of the latter. This paper surveys the body of evidence built up over the last decade, from a wide range of airborne experiments, which illustrates the ability of such a sensor to provide the required measurements.
The goal of quantifying the woody cover and biomass of tropical savannas, woodlands and forests using satellite data is becoming increasingly important, but limitations in current scientific understanding reduce the utility of the considerable quantity of satellite data currently being collected. The work contained in this thesis reduces this knowledgegap, using new field data and analysis methods to quantify changes using optical, radar and LiDAR data. The first paper shows that high-resolution optical data (Landsat & ASTER) can be used to track changes in woody vegetation in the Mbam Djerem National Park in Cameroon. The method correlates a satellite-derived vegetation index with field-measured canopy cover, and the paper concludes that forest encroached rapidly into savanna in the region from 1986-2006. Using the same study area, but with radar remote sensing data from 1996 and 2007 (ALOS PALSAR & JERS-1), the second paper shows that radar backscatter correlates well with field-measured aboveground biomass (AGB). This dataset confirms the woody encroachment within the park; however, in a larger area around the park, deforestation dominates. The AGB-radar relationships described above are expanded in the next paper to include field plots from Budongo Forest (Uganda), the Niassa Reserve (north Mozambique), and the Nhambita Community Project (central Mozambique). A consistent AGB-radar relationship is found in the combined dataset, with the RMSE for predicted AGB values for a site increasing by
Although the plant species composition in the various vegetation types of southern Africa is influenced by soil properties such as nutrient status, pH, salinity and texture, the overwhelmingly important factor determining the spatial distribution of forest, savanna and grasslands is soil moisture balance. This is a function of a single feature, or several in combination, ie texture and consistence, presence or absence of a pan horizon, distance of this horizon from the surface, macro- and microrelief, and salinity.
The Uluguru Mountains in eastern Tanzania contain at least 16 endemic vertebrate and 135 endemic plant taxa, with hundreds of more taxa shared only with forests in eastern Tanzania and Kenya. This degree of endemism is exceptional in tropical Africa, and the Uluguru Mountains are one of the 10 most important tropical forest sites for conservation on the continent. Surveys carried out during 19991,600 m, and concentrated in submontane forest. During the recent surveys most of the endemic and near-endemic vertebrate species known from the Uluguru Mountains were re-recorded, but three endemic snake species and two near-endemic bird species were not found. These species were previously known from the elevations where deforestation has been greatest. More than 50 plant species are also known only from the altitude range that has been heavily deforested. The primary cause of forest loss has been clearance for new farmland. The forest that does remain is largely confined to Catchment Forest Reserves managed for water by the Tanzanian Government. Without these reserves the loss of forest, and hence the loss of biodiversity, in the Uluguru Mountains would most likely have been much greater.