This article appeared in a journal published by Elsevier.

The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/authorsrights

Author's personal copy

Applied Geography 41 (2013) 168e178

Contents lists available at SciVerse ScienceDirect

Applied Geography
journal homepage: www.elsevier.com/locate/apgeog

The tropical forest in south east Asia: Monitoring and scenario

modeling using synthetic aperture radar data
Rajesh Bahadur Thapa*, Masanobu Shimada, Manabu Watanabe, Takeshi Motohka,
Tomohiro Shiraishi
Earth Observation Research Center, Japan Aerospace Exploration Agency (JAXA), 2-1-1 Sengen, Tsukuba City, Ibaraki 305-8505, Japan

a b s t r a c t
Keywords:
PALSAR
Land cover change
Deforestation
Scenario analysis
LULUCF
Weights of evidence
Sumatra
REDD

Tropical forests play a major role in storing large carbon stocks and regulating energy, and water uxes,
but such forest cover is decreasing rapidly in spite of the policy attention on reducing deforestation.
High-resolution spatiotemporal maps are unavailable for the forests in majority of the tropical regions in
Asia because of the persistent cloud cover and haze cover. Recent advances in radar remote sensing have
provided weather-independent data of earth surface, thus offering an opportunity to monitor tropical
forest change processes with relatively high spatiotemporal resolutions. In this research, we aim to
examine the tropical deforestation process and develop a spatial model to simulate future forest patterns
under various scenarios. Riau Province from central Sumatra of Indonesia is selected as the study area;
this province has received much attention worldwide because the highest CO2 emission resulting from
tropical deforestation has been recorded. Annual time series PALSAR data from 2007 to 2010 were
analyzed for forest mapping and detecting land cover changes. A spatial model was calibrated using the
Bayesian method. Modeling parameters were customized for the local subregions that allocate deforestation on the basis of their empirical relationships to physical and socioeconomic drivers. The model
generated landscape spatial patterns mirrored the possible locations and extent of deforested areas by
2030 and provided time-series crucial information on forest landscape under various scenarios for future
landscape management projects. The results suggested that the current deforestation process is in a
critical stage where some subregions may face unprecedented stress on primary forest costing rivers and
forest ecosystems by the end of 2020. The perspective views of Riau Province generated by the model
highlighted the need for forest/environmental planning controls for the conservation of environmentally
sensitive areas.
2013 Elsevier Ltd. All rights reserved.

Introduction
Tropical forests play a major role in storing large amounts of
carbon stocks and in regulating energy and water uxes. These
forests are being diminished rapidly despite the attention to policies for reducing deforestation (FAO, 2010; Lambin & Meyfroidt,
2011). As deforestation proceeds, the living planet becomes more
vulnerable because of adverse impacts on the environment and
overall climate system related to the release of carbon, reduced
biodiversity, disturbed water regulation, and impacts on weather
patterns. Recently, tropical deforestation has been recognized as
the second largest source of greenhouse gas emissions and this
trend is expected to continue for the next several years (Harris

* Corresponding author. Tel.: 81 50 3362 7444; fax: 81 29 868 2961.

E-mail addresses: rajesh.thapa@jaxa.jp, thaparb@gmail.com (R.B. Thapa).
0143-6228/$ e see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.apgeog.2013.04.009

et al., 2012; IPCC, 2007; Saatchi et al., 2011). To reduce deforestation and forest degradation and to mitigate forest-related GHG
emissions, the Reducing Emissions from Deforestation and Forest
Degradation (REDD), an international agreement under the
United Nations Framework Convention on Climate Change, is being
advanced in several developing countries, including Indonesia
(Angelsen, 2008; NORAD, 2011). Successful implementation of this
agreement requires regularly updated spatial information on
changes in natural forest cover and the development of reference
scenarios for projecting deforestation and associated emissions.
Therefore, continuous monitoring of tropical deforestation and an
understanding of the causal effects will be essential in the future.
Examining these effects and envisioning future circumstances of
deforestation necessitate logically developed spatial models.
Spatially explicit data on changes in forest cover require investigating such changes, identifying the drivers of the changes, and
calibrating deforestation models. In the past, acquisition of spatial

Author's personal copy

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

data on forest cover was eld intensive, making it expensive and

difcult to acquire accurate, timely, and consistent data over a large
area. However, advances in remote sensing techniques have
reduced this limitation and offer vast opportunities to characterize
forest cover at reasonable spatiotemporal scales. Tropical regions in
Asia mostly covered by clouds, which creates a major barrier to
generating spatiotemporally consistent land use/cover maps using
optical remote sensing techniques. However, recent progress in
improving spatial resolution of synthetic aperture radar (SAR) has
signicantly reduced this barrier (Hoekman, Vissers, & Wielaard,
2010; Thapa et al., 2013). SAR techniques are particularly useful
for monitoring deforestation and forest degradation due to their
ability to monitor the earth surface in all weather and solar illumination conditions. Owing to longer wavelengths and shorter
frequencies, L-band SAR data enables mapping forest areas because
the microwave energy transmitted by the sensor penetrates forest
canopies and returns the backscatter representing forest structures.
Greater penetration of vegetation and weaker reection from
rough surfaces enable the acquisition of important information for
identifying and separating types of forests as well as other land-use
types (Shimada & Ohtaki, 2010; Walker, Stickler, Kellndorfer, Kirsch,
& Nepstad, 2010). However, the amount of backscattered energy
depends on the size and orientation of canopy structural elements
(i.e., leaves, branches, and stems), the moisture content of vegetation, and underlying soil conditions.
Remote sensing techniques have made it possible to map the
history of forest land cover and analyze dynamics of forest change
regardless of geographic size, but spatial modeling of deforestation
is necessary for understanding the complex process of forest cover
change; because the spatial modeling provides important abstract
information about the future, and enables testing the implications
of different forest policies. Various spatial models have been
developed from different perspectives to improve representation of
dynamic process of deforestation in the cellular automata (CA)
modeling environment by differentiating the probabilistic functions. For instance, logistic regression (Chowdhury, 2006;
Echeverria, Coomes, Hall, & Newton, 2008), genetic algorithm
(Soares-Filho, Rodrigues, & Follador, 2013; Venema, Calamai, &
Fieguth, 2005), weight of evidence (Maeda et al., 2011; SoaresFilho et al., 2010), and articial neural network (Khoi &
Murayama, 2011; Mas, Puig, Palacio, & Sosa-Lpez, 2004)
methods have been discussed in the contemporary literature. The
spatial models described in these works simulate changes in forest
cover based on previous and surrounding states and on biophysical
and socioeconomic driving factors. The focus of most of these
models is to investigate deforestation, understand the driving factors, simulate the dynamic processes of forest conversion, and
expedite understandings of the associated environmental impacts
so that measures to control deforestation can be formulated.
Understanding the dynamic process of deforestation in a region
depends greatly on timeespace relationships. This dynamic process
consists of a complex interaction between several components, such
as physiography, socioeconomic factors, and forest management
policies. Such complex interaction between changes in forest cover
and the various components, often called drivers, can be addressed
smoothly in CA with the weights of evidence (WofE) modeling
framework. CA is a simple grid-based system in which forest cover
changes can be represented in grid cells. CA-based models have
been demonstrated to be effective platforms for simulating dynamic
spatial interactions among biophysical and socioeconomic factors
associated with land-use and land-cover changes. The WofE method
is based on Bayes rule of probability with an assumption of conditional independence (Thapa, 2012). It has been used widely for
different applications, originally for modeling geologic phenomena
(Bonham-Carter, 1994) and later for modeling changes in land use

169

and land cover (Maeda et al., 2011; Soares-Filho, Cerqueira, &

Pennachin, 2002; Thapa, Shimada, Motohka, Watanabe, &
Shiraishi, 2011). The WofE method is robust in its handling of categorical and continuous spatial data, such as land cover maps and
distance maps, respectively. This method concerns the favorability
of detecting a certain event, for example, a land change event from
forest to non-forest, in relation to potential evidences that may be
topographic or socioeconomic types of drivers. The model is in loglinear form, which allows the weights from the evidential themes be
added while developing a probability map that can be easily
implemented in the CA modeling environment.
Riau Province in western Indonesia has recently received
worldwide attention owing to forest-related carbon emissions. The
province lost 583 thousand hectares (ha) of natural forest in dry
and peat-swamp areas and led in CO2 emissions (812 megatons)
the other provinces in the country during 2000e2005 (MoFor,
2008). Alarming deforestation has threatened forest carbon
stocks, peat drainage, and biodiversity in the province. The goal of
this research is to examine the deforestation process and develop a
spatial model to simulate spatial patterns of forest cover under
what-if scenarios to provide a basis for forest management in the
province. The CA model was calibrated using the WofE method and
four scenarios (business as usual, forest regeneration, forest conservation, and forest concession) were examined. Time-series
phased-array L-band synthetic aperture radar (PALSAR) data were
analyzed for time-series mapping of the forest cover.
Study area and database
Geography of Riau Province, Indonesia
Riau Province is located in central Sumatra, Indonesia (Fig. 1). It
is within the geographic region of 170 2400 S to 2 320 3600 N latitude
and 100103000 to 103 480 3900 E longitude. The province has varied
topography, including peat-swamp areas, basins, hillocks, mountains, rivers, and the coasts (eastesouth to north). It spans about 9
million ha with varying elevations up to 1200 m. The area has a
mostly cloudy and hazy atmosphere throughout the year. The
province consists of 12 administrative subregions (districts):
Bengkalis, Indragiri Hilir, Indragiri Hulu, Kampar, Dumai, Pekanbaru, Singingi, Pelalawan, Rokan Hilir, Rokan Hulu, Siak, and Meranti. Occupying 15% of the landscape, Indragiri Hilir is the largest
subregion, whereas Pekanbaru, headquarters of the province, is the
smallest with only 0.7% of the provincial territory. Currently, 5.54
million people live in the province (BPS, 2010). Being a major source
of pulp and paper, oilpalm, and petroleum products, the province is
of great economic importance for Indonesia. The forest landscape
has become the major source of land for industrial plantations in
recent years. The province has lost more than 35% of its forest since
1985 and is ranked to the most highly deforested province in
Sumatra (Uryu et al., 2010).
Spatial data preparation
Owing to the persistent cloudy and hazy conditions that exist
throughout the year in the study area, SAR data are an ultimate
choice for monitoring and assessing changes in forest cover.
PALSAR, a SAR sensor on the Advanced Land Observing Satellite
(ALOS), systematically acquired a large amount of data covering
Earths surface through the ve-year period of 2006e2011
(Rosenqvist, Shimada, Ito, & Watanabe, 2007). This L-band sensor,
designed with a long operating wavelength (23.6 cm), was appropriate for tropical forest monitoring because of its high sensitivity
to forest structure and moisture characteristics (Shimada, 2011).
We used PALSAR mosaic data products that were radiometrically

Author's personal copy

170

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

Fig. 1. The study area: subregions, urban centers, major roads, and water channels.

and geometrically calibrated at a 25 m spacing unit with the Japan

Aerospace Exploration Agencys (JAXA) mosaicking algorithm
(Shimada, Isoguchi, Tadono, & Isono, 2009; Shimada & Ohtaki,
2010). These products were developed using the ne beam dual
(FBD) mode of the ALOS-PALSAR sensor, which consists of horizontal transmit and horizontal receive (HH) and horizontal transmit and vertical receive (HV) polarizations at a 34.3 off-nadir
angle. This mode is highly applicable to the characterization of
forest cover in various tropical regions (Hoekman et al., 2010;
Thapa et al., 2013; Walker et al., 2010). The province required a
large set of mosaic data, data of eight 70 km wide ascending paths
of ALOS, which is equivalent to 45 standard PALSAR scenes
(70  70 km), for each year. In this study, we prepared data that
were acquired in dry season (JuneeOctober) of the years 2007,
2008, 2009, and 2010.
An automated mapping method was applied to prepare land
cover maps. The mapping method uses image segmentation and
threshold techniques to discriminate land cover types based on HH
and HV backscatter information in the PALSAR mosaic data. The
segmentation technique divides the HH and HV images into smaller
image objects of homogeneous regions. Among the various segmentation techniques (Trimble, 2012), we applied a multiresolution segmentation technique that minimizes heterogeneity
and maximizes homogeneity for a given number of image objects in
the mosaic data. The segmentation procedure begins with singleimage objects of one pixel and repeatedly merges them in pairs
in several loops to larger units as long as an upper threshold of

homogeneity is not exceeded locally. After the segmentation process was completed, the image objects were labeled by establishing
rules for different types of land cover. In the image-labeling procedure, different types of thresholds for HH and HV backscatters
were established to extract forest, water, and non-forest land cover
types. Land cover mapping rules were developed in eCognition
software (Trimble, 2012). Details of the mapping method can be
found in Thapa et al. (2013). Four time-series land cover maps
(Fig. 2) corresponding to the data-acquisition years were prepared
in order to model future spatial patterns of the forest. These maps
consist of natural forest, non-forest, and water categories and have
less than 10% overall mapping uncertainty.
Other supporting spatial data used for the spatial modeling
were altitude, slope, road, urban center, rural settlement,
waterway, district boundary, concession map, conservation map,
and eld-work data. The altitude and slope maps were derived
from the Shuttle Radar Topography Mission (SRTM) 90 m data,
whereas roads, urban centers, and rural settlements were digitized
from the provincial map (Riau Province map). The waterways
include major rivers channels, lakes, and coastal areas, which were
revised based on the PALSAR-based land cover map and SRTM map.
Concession and conservation maps were obtained from the Ministry of Forestry, Indonesia. Many eldwork trips were organized
for eld observation and verication in 2011 and 2012. The
Euclidean distance method was used to calculate proximity to
roads, urban centers, rural settlements, non-forest areas, and
waterways.

Author's personal copy

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

171

Fig. 2. Four land cover maps (A. 2007, B. 2008, C. 2009, and D. 2010) and locations of concession and conservation areas including deforested landscape observed in 2007e2010 (E).

Model calibration and validation

Landscape transition matrix calculation
In this paper, a subregion concept was used for the modeling
because the deforestation process is location specic and differs
with local landscape characteristics. In addition, coefcients localized at the subregion level are more robust for addressing local
relationships among deforestation factors than the averaged
province-level coefcient. The Markov chain (MC) approach was
used to calculate landscape transition matrices at the subregion
level. This approach analyzes two land cover maps from different
dates and produces a transition probability matrix. A landscape
transition matrix for each local subregion was calculated using the
subsequent map for calibrating the model. Similarly, a yearly
transition matrix for each subregion from the initial and nal maps
was prepared in order to predict the spatial patterns of forest
landscape under what-if scenarios.
Driving factors of deforestation
The selection of the best set of input variables to produce the
best t between the empirical data and the observable reality is a
very important aspect of deforestation modeling. After calculating
the land cover change, we identied a set of explanatory variables
governing the changes from forest area to non-forest area. The
variables available for the modeling analysis do not always represent the set needed to produce ideal simulation results. However,
there are no universal explanatory factors of forest cover change.
Although similar factors have been found by several studies, the
degree to which they contribute to landscape change differs

(Echeverria et al., 2008; Jaimes, Sendra, Delgado, & Plata, 2010;

Kamusoko et al., 2011; Khoi & Murayama, 2011; Maeda et al., 2011;
Soares-Filho et al., 2010; Thapa & Murayama, 2009). Individuals,
government plans and programs, landforms, landscape change
processes, and available resources often cause differences in the
importance of factors. We examined a set of driving factors, altitude, slope, forest conservation and concession areas, proximity to
non-forest areas, roads, waterways, urban centers, and rural settlements, and found that the factors substantially correlated to
landscape changes in the province. While evaluating the factors,
spatially explicit maps were prepared for the corresponding factors
and each factor map was overlain on the observed landscape
change during 2007e2010. The effects of the spatial factors on the
landscape change was calculated using the WofE method such that
weights for each factor were estimated from the measured association between the deforestation occurrences and the values on
the spatial patterns of the driving factors (Soares-Filho et al., 2010;
Thapa, 2012). In WofE, the posterior probability of land cover
change from forest to non-forest also called deforestation (D),
occurring given the presence of a spatial pattern of deforestation
factor (F) can be expressed as in Eq. (1).

PDjF PD*PFjD=PF

(1)

Equation (1) can be expressed in terms of odds by dening a

ratio of the probability that deforestation will occur to the probability that it will not occur as in the following (Eq. (2)).

ODjF OD*PFjD=PFjD0

(2)

where O(DjF) is the posterior odds event D, given a spatial pattern F,

O(D) is the prior odds event D, D0 represents the absence of event D,

Author's personal copy

172

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

and P(FjD)/P(FjD0 ) is the odds ratio. The natural logarithm of the

odds ratio is positive WofE denoted by W (Bonham-Carter, 1994;
Soares-Filho et al., 2010). The W values represent the attraction
between deforestation and the driving factors. The higher the W
value is the greater the probability that deforestation will occur.
The spatial independency of the input factor maps was tested
using Cramers V (Bonham-Carter, 1994). This test presents the
value in a range from 0 to 1, which shows the degree of association
between the maps being compared as independent to full,
respectively. Among the factors examined in this experiment, no
signicant spatial dependency was detected and the Cramers V
was less than 0.33 in all comparison cases. Furthermore, the
proximity to non-forest land was set to be dynamic considering the
forest cells expected to change into non-forest in a given time step
during the simulation. In this case, proximity to non-forest land
needs to be recalculated over the entire study area in each simulation run. The remaining factors were set to be static during the
model run.
Transition potential map calculation and simulation
The spatial distribution of deforestation across the province was
calibrated in a CA model integrating with WofE. WofE helps to
produce a transition potential map depicting the most favorable
areas for change (Soares-Filho et al., 2010). The spatial transition
potential (P) from forest to non-forest (D) for a given set of driving
factors can be expressed by the following equation (Eq. (3)):

X

X

w
P DjF1 XF2 X.Fn exp
w
n =1 exp
n

(3)

where, P(DjF1 X F2 X.Fn) is the conditional probability of deforestation given a set of spatial patterns of deforestation factors
(F1.Fn), and w
n represents the weights of evidence of deforestation occurring for spatial pattern of Fn.
After creating the transition potential map, the simulation
model was calibrated by internal parameters concerning the
average size, variance, and isometry of patches. These functions
enable the formation of a variety of sizes and shapes of patches of
land cover change during the simulation period. Patch isometry
may vary from 0 to 2, indicating a more isometric form of new
patches as the number increases. The sizes of change patches were
set according to a lognormal probability distribution; therefore, it is
necessary to specify the parameters of this distribution represented
by the mean and variances of the patch sizes to be formed (SoaresFilho, Rodrigues, & Costa, 2009). The mean patch size and variance
were determined from the source maps, whereas the isometry was
determined empirically. Using these internal parameters, the
quantity of land cover transition, and the reference map of 2007, we
simulated landscape patterns for 2010 to validate the simulated
map with the reference map of 2010. This simulation model was
designed in DINAMICA (Soares-Filho et al., 2009, 2013).

similar to Costanza (1989) but that accounts for fuzziness of location and category within a cell neighborhood. We used a two-way
reciprocal fuzzy similarity method (Almeida, Gleriani, Castejon, &
Soares-Filho, 2008; Soares-Filho et al., 2009) in which a representation of a cell is inuenced by the cell itself and also, to a lesser
extent, by the cells in its neighborhood. We compared the spatial
similarity between the reference map (2010) and the simulated
map (2010) at different scales. The degree of spatial similarity between the maps is presented in Fig. 3. The gure shows the predictive power of the model at spatial scales from 1  1 (pixel by
pixel) to 21  21. The model can produce 86% similar spatial patterns at the pixel level. Overall spatial similarity improves with
increasing window size, meaning that the model can predict spatial
patterns more accurately at coarser spatial resolution. Improving
similarity with coarser resolution will degrade spatial details,
which should be interpreted cautiously. For example, local level
planners and provincial level planners may need results at different
scales, and subregion level planners would anticipate much more
detailed results than those for the province level. Depending on the
scale requirements or information details of a particular policy
assessment, forest managers would have the option of selecting
different spatial resolutions to produce higher agreements. For
instance, more than 91% spatial similarity can be achieved with a
5  5 window. Because achieving higher accuracies will result in
coarser results, this option may be useful for provincial or national
planners but not for local planners.
Furthermore, the simulated change was quantitatively
compared to the observed change to clarify errors in the land cover
change simulation at the subregion level (Fig. 4). Contrasting patterns among the subregions between the observed change and the
simulated change are observed. Half of the subregions in the
province have less than 200 ha of difference. The Indragiri Hilir
subregion has the highest difference, whereas the Dumai subregion
has the lowest. The absolute difference between the observed
change and simulated change at the province level is nearly
1000 ha.
Forest process scenarios development
The successfully calibrated spatial model can be advanced to
what-if scenario development for the decision support process in
sustainable forest management. Such scenarios derive maps that
produce important information about what could happen if we
continue the current process further or make changes in preconditions (Thapa & Murayama, 2012). In many cases, models are
used to explore the consequences of changes by objectively

Model validation
Validation of a land change model is usually carried out by
comparing the predicted result to the reference map to determine
the prediction ability of the model. Spatial models require a comparison with a neighborhood context, because even maps that do
not exactly match cell-by-cell can present similar spatial patterns
and likewise spatial agreement within a certain cell vicinity. To
address this issue, several vicinity-based comparison methods have
been developed. For example, Costanza (1989) introduced a
multiple-resolution tting procedure that compares map t within
increasing window sizes. Hagen (2003) introduced a method

Fig. 3. Spatial similarity between the reference map (2010) and the simulated map
(2010) at different-scales.

Author's personal copy

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

173

concession areas for industrial plantations and selective logging,

are activated strictly from 2010 onward. The G-CPL scenario includes concession areas allocated for oil palm plantations, industrial plantations, and selective logging. This scenario assumes that
the future deforestation process will be conned to only the
concession areas. In the G-FC scenario, the deforestation process
will not continue the past trend and will occur only beyond the
conservation areas. The future spatial patterns of the deforestation
process for all the scenarios were extrapolated for the time period
of 2015e2030 at 5-year intervals.
Observed and simulated landscape spatial patterns from 2010
to 2030
Fig. 4. The absolute difference between the observed change and the simulated
change during the period 2007e2010.

modifying preconditions. Such explorative scenarios help to

delineate potential environmental consequences and develop
measures to alleviate deforestation in a particular region
(Kamusoko et al., 2011; Verburg, Eickhout, & van Meijl, 2008). To
demonstrate the usefulness of the modeling methods and to provide a comprehensible and alternative context for forest managers,
we dened four simple but plausible scenarios for the study area:
business as usual (BAU), forest regeneration (FR), governance e
forest conservation (G-FC), and governance e concession for industrial plantations and selective logging (G-CPL).
The BAU scenario assumed that the deforestation process would
continue the past trend everywhere in the province. The demand
and supply of forest and the forest policies impacting the past
deforestation trend would continue. In BAU, we used the 2010
reference map, calibrated weights of the driving factors, yearly
landscape transition matrices for each subregion, and the calibrated
transitional potential map to extrapolate future spatial patterns of
deforestation. In the FR scenario, a forest regrowth class was created
for the last three years of the time period by setting the condition
that if the non-forest land in a previous year was observed as forest
land in the latter year, the land would be assigned as regrowth class
in the land cover map of the latter year. In this case, three types of
land cover transitions may occur, i.e., forest to non-forest, nonforest to regrowth, and regrowth to non-forest. The driving factors
were the same as in the BAU scenario, but the weights and transition matrix were recalibrated to add the regrowth class.
The two governance scenarios (G-FC and G-CPL) assumed that
two broad forest policies, i.e., conservation areas for forest and

Table 1 shows the quantitative patterns of observed landscapes

in 2010. A subregion percent landscape is presented using the forest
cover maps shown in Fig. 2 to highlight the current condition of the
forest landscape observed from the PALSAR data analysis. At the
province level, 41% of the total land is covered by forest. Non-forest
and water lands account for 46% and 2% of the landscape, respectively. A large part of the forest (approximately 1 million ha, 11% of
the landscape) was deforested during the years 2007e2010.
Although deforestation occurred everywhere in the province, the
Rokan Hilir and Siak subregions suffered proportionately higher
deforestation in this period. Four of the twelve subregions, Indragiri
Hulu, Singingi, Pelalawan, and Meranti still have half of their
landscape in forest. The most urbanized subregion, provincial
headquarters (Pekanbaru), has the lowest forest cover (12%).
Some parts of the provinces forest lands are allocated for conservation areas and concession areas. Seventeen areas have been
allocated for forest conservation proposes in the province (Fig. 2-E).
Conservation lands account for 617,064 ha or 6.9% of the total
landscape. The Bukit Rimbang/Baling-baling Nature Reserve
located at the western border of the province that includes land of
the Kampar and Singingi subregions and Bukit Tigapuluh National
Park at the southern border of the province that includes land of the
Indragiri Hulu and Indragiri Hilir subregions are the largest conservation areas in the province. These two conservation areas
occupy 22% and 17%, respectively, of the total conservation landscape in the province. Kerumutan Wildlife Reserve in the Pelalawan
area and the Giam Siak Kecil-Bukit Batu UNESCO-MAB Biosphere
Reserve in the Siak area are the third and fourth largest conservation areas in the province. The spatial distribution of conservation
areas shows that all subregions except for Rokan Hulu in the
northwestern part of the province have some share of their landscape in conservation. Among the subregions, Indragiri Hulu in the

Table 1
The observed landscape in 2010.
Sub-regions

Sub-regions area in ha

Landscape area in percentage

Forest

Non-forest

Water

Bengkalis
Indragiri Hilir
Indragiri Hulu
Kampar
Dumai
Pekanbaru
Singingi
Pelalawan
Rokan Hilir
Rokan Hulu
Siak
Meranti
Province total

829,988
1,352,990
770,840
1,146,800
209,757
61,138
549,805
1,266,960
892,142
701,428
840,598
366,904
8,989,350

44.66
21.35
57.56
37.80
48.23
12.31
55.95
53.63
31.00
31.08
44.05
52.24
41.04

43.99
64.89
29.57
51.70
41.13
75.54
32.00
33.33
51.90
57.47
42.21
35.57
46.12

1.96
4.64
1.37
0.77
0.00
0.76
1.39
1.95
1.37
0.91
1.67
0.78
1.85

Conservation area in percentage

Concession area in percentage

Deforested

Area

Forest

Deforested

Area

Forest

Deforested

9.28
9.06
11.43
9.66
10.64
11.39
10.51
11.06
15.57
10.41
12.07
10.98
10.89

11.11
1.32
20.04
9.16
1.68
1.27
8.05
9.50
0.45
0.00
8.25
1.41
6.86

52.37
96.34
96.87
85.93
53.66
88.26
94.75
83.42
0.00
0.00
86.73
87.32
83.45

4.70
3.29
1.30
6.12
9.39
1.94
1.93
8.16
5.74
0.00
1.73
3.56
4.21

38.26
30.23
17.65
33.53
52.15
12.17
34.98
51.03
37.75
18.70
50.86
23.90
35.44

53.63
47.12
37.84
23.91
62.12
4.13
43.90
57.45
43.56
18.70
50.17
73.67
46.48

11.31
11.71
18.69
9.83
10.97
6.64
12.26
13.11
13.88
15.78
15.95
11.44
12.98

Note: In the Conservation and Concession columns, the reported percentages for forest and deforested regions represent the proportion of land cover within the conservation
and concession areas of the corresponding sub-regions, respectively. The Deforested landscape indicates the forest land cleared during 2007e2010.

Author's personal copy

174

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

southwestern part of the province has the largest share (20%) of its
landscape in conservation (Table 1). Other subregions including
Bengkalis, Kampar, Pelalawan, Siak, and Singingi have 8e11% of
their landscapes allocated for conservation. Deforestation is
noticeably present in the conservation areas, which lost 4.2% of
their forest landscape during the period 2007e2010. However,
forest still covers the vast majority (83.5%) of the conservation
landscape.
Contrary to conservation, some parts of the provincial lands are
allocated to fullling demand for wood products and for other
economic activities. The forests in the concession areas are allowed
to be gradually cleared for oilpalm plantations, industrial plantations, and selective logging. All together, 3.1 million ha of land,
35.4% of the provincial landscape, are allocated for this purpose.
These concession areas are distributed sparsely in all subregions
(Fig. 2-D). The Dumai, Pelalawan, and Siak subregions have slightly
more than 50% of their land allocated for concession purposes
(Table 1). Pekanbaru, Indragiri Hulu, and Rokan Hulu have proportionately less concession area compared to other subregions.

The land cover map of 2010 shows that 13% of the concession
landscape was deforested in 2007e2010 but that 46.5% of the
concession landscape remains untouched. Concession land in forest
is largely available in the Meranti (73.7%), Dumai (62.1%), and
Pelalawan (57.4%) subregions.
Using the modeling conguration, calibrated parameters of the
2007e2010 period, and the input map of 2010, we performed
simulations under the four scenarios (BAU, FR, G-CF, and G-CPL)
with the goal of estimating alternative spatiotemporal patterns of
expected remaining forest and deforestation in the province by
2030. Simulation was run for each year. Fig. 5 shows the observed
and extrapolated spatial patterns of forests, non-forest, and deforested areas for the four scenarios for the years 2010, 2020, and
2030. Different spatial patterns can be seen in all the scenarios.
Difference in the spatial pattern of the deforestation areas between
the BAU and other three scenarios are apparent. If the historical
trend continues without any policy intervention as evidenced by
the BAU scenario (Fig. 5-A), the forest cover will consistently
disappear leaving only very few small patches of forest by 2030. The

Fig. 5. Observed and simulated landscape spatial patterns (2010e2030) by scenarios: A) BAU, B) FR, C) G-CF, and D) G-CPL. The deforested pattern by 2010 represents the observed
forest in 2007 changed to non-forest in 2010.

Author's personal copy

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

deforestation pattern extends to adjacent and environmentally

sensitive areas, such as conservation areas, forest in peat-swamp
areas in the northeastern part of the province, and dry-land areas
in the southwestern part in the province. Spatial patterns in the FR
scenario (Fig. 5-B) are very noisy due to the forest regrowth process.
The forest regrowth scenario seeds new regrowth areas in previously deforested areas but continues deforestation based on the
previous landscape history. In this map, the deforestation pattern
includes both change from forest to non-forest and change from
regrowth to non-forest. The majority of the regrowth spatial patterns are observed in the Pelalawan, Siak, and Bengkalis subregions.
Other subregions show regrowth at a lower level. The regrowth
scenario shows positive signals for the development of secondary
forest and for increasing future forest stocks.
The spatial patterns in the G-CF scenario (Fig. 5-C) are slightly
similar to those of the BAU scenario, except in the conservation
areas. The remaining forest patches are quite large in G-CF as the
result of forest conservation policy. All the remaining forests in the
conservation areas in 2010 remain untouched in future years.

175

However, during the simulation runs under the G-CF scenario,

deforestation pressure occurred beyond the conservation areas.
Most of the forests in the subregions located in the northern part of
the province are deforested by 2020. The governance scenario (GCPL) produced a balanced forest spatial pattern in the province
compared to other scenarios (Fig. 5-D). Deforestation occurred only
inside the lands allocated for concession. If this policy is implemented rigorously, all subregions would still have a considerable
amount of forest cover by the end of 2030.
Fig. 6 further illustrates the present and expected forest stock
patterns for the four scenarios observed in the maps by subregion
in 5-year intervals from 2010 to 2030. The decreasing patterns of
forest stock can be associated with deforestation patterns, which
provides an opportunity to understand the deforestation states at
the local level. The curves of the different scenarios show different
behavior during different time periods. The BAU curve remains in
the bottom line in all subregions, showing different states of
deforestation. If the deforestation trend continues as in the past,
the Pekanbaru region will have much stress on forest than the other

Fig. 6. The simulated deforested landscape according to subregions and scenarios.

Author's personal copy

176

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

subregions. Due to higher expected deforestation under the BAU

scenario, the Indragiri Hilir, Rokan Hilir, and Rokan Hulu subregions
are expected to have forest stock of less than 5% of the landscape by
2030. In this case, the FR and G-CPL scenarios may save the forest to
some extent. The expected forest stocks shown by the BAU and GCF curves in the Rokan Hilir and Rokan Hulu subregions are similar
either because they lacked conservation areas or because the conservation areas were completely deforested in the past. Throughout
the prediction period, the differences in expected deforestation
between the BAU and G-CF scenarios was widen in the Indragiri
Hulu, Kampar, and Pelalawan subregions. Deforestation will likely
be lower under the FR scenario than under the others in the
Bengkalis, Pekanbaru, Pelalawan, and Siak regions. It remains the
second alternative for reducing deforestation in all other regions,
where the G-CPL scenario is dominant. The G-CPL scenario shows a
high expectation of forest stock in the Indragiri Hulu, Kampar,
Singini, and Meranti subregions at even the end of 2030. Interestingly, the trend of expected deforestation under the G-CPL scenario
is very slow in Pekanbaru and Rokan Hulu compared to other regions. A rapid deforestation trend, compared to those of other
subregions, is observed for all scenarios in the Dumai subregion. As
compared to the other subregions, Indragiri Hulu will likely face the
largest deforestation under the BAU scenario by 2030 but the G-CPL
scenario provides a good tradeoff.
Fig. 7 presents the expected remaining forest in the province for
each scenario through the next two decades. A sharp decline of
forest is observed during the baseline period when forest land
decreased from 4.67 to 3.69 million ha, nearly 1 million ha of
deforestation, from 2007 to 2010. The impact of the various scenarios in estimated forest stocks is dynamic at different time periods. If the current trend continues, more than half of the forested
land will likely be cleared, leaving only 1.66 million ha, by the end
of current decade, as evidenced by the BAU. We may see less than
0.8 million ha of forest cover by 2030, which could result in adverse
environmental consequences including loss of biodiversity in the
province. The G-CF scenario slightly improves the forest preservation throughout the period as compared to BAU. The spatial distribution of forested land cover under this scenario is shown in
Fig. 5-C. Fig. 7 shows that forest stocks will likely be better such that
we may expect 1.8 million ha of forest land in 2020 and slightly
more than 1 million ha in 2030. Although the regrowth process
requires a long period of time to attain the quality of the natural
forest, it will likely expand the forest coverage in secondary forest.
Combining forest and regrowth, the FR scenario shows better forest
stocks than the BAU or G-CF. The FR scenario suggests that 2 million

Fig. 7. Time-series estimation of the remaining forest cover in the province.

ha of forest land can be expected by 2030. The remaining forest

trend of the G-CPL scenario is expected to be surprisingly different
from those of the other scenarios. By implementing the G-CPL
policy fairly, the province will likely have a good forest pattern to
2030, which is also evidenced by the spatial patterns in Fig. 5-D.
Unlike the other scenarios, G-CPL will likely follow the FR trend
until 2015 but is expected to slow the deforestation trend through
subsequent years. The estimated forest will likely be about 2.8
million ha in 2020 and 2.5 million ha in 2030, as shown by the GCPL curve in Fig. 7. If we consider the curve for the BAU scenario as a
reference line, the G-CF, FR, and G-CPL scenarios will likely save 0.3,
1.36, and 1.7 million ha of natural forest land, respectively, from
deforestation by the end of 2030.
Discussion and conclusion
This study used L-band SAR data for land cover mapping and
thus overcome the problems posed by persistent clouds and haze
effects, which affect the widely available optical remote sensing
data sets. Employing SAR remote sensing techniques, we traced the
spatiotemporal forest patterns for the past and examined the future
trends under four scenarios, which will be a reference asset for the
forest resources managers dealing with tropical forests. With regard the spatial patterns of land change in the past, the rate of
conversions from forested to non-forested areas has been rapid.
Biophysical characteristics, socioeconomic conditions, neighborhood interactions, and transportation accessibility in the province
have played important roles that led to such unexpected spatial
patterns. The forest lands in concession areas are mostly cleared for
the oil-palm and acacia plantations.
Modeling of deforestation and optimizing the forest conservation have been the objective of forest planning research for many
years. We developed a spatially explicit simulation model that
provides perspective landscape in what-if contexts and is responsive to policy intervention scenarios for forest preservation. As
Pontius et al. (2008) pointed out, a complex model does not
necessarily lead to higher predictive power. General models that
are exible, allowing increasing complexity, are often preferred in
planning process. Therefore, we established a simple CA deforestation modeling framework that requires less input data and serves
as learning tool for explaining the interrelations of land change
dynamics. It features the sub-region concept to integrate locationspecic deforestation process and provides scenario-based alternative mirrors of the future forest landscape in the province. The
Bayesian approach (weight of evidence) was adopted for calibrating
factors weights and land cover transition potential map. This
approach is robust in handling missing data and categorical data, as
shown by previous studies Bonham-Carter (1994) and Soares-Filho
et al. (2010). The modeling framework follows empirical calibration
and validation procedures. Comparison of the model-simulated
map with the reference map at different spatial resolutions provides more options for modeling accuracies. This is critical information for forest environmental planners to understand the
performance and predictability of a model at different spatial
scales.
In order to simulate the inuence of potential policy interventions on the land change processes that are depleting native
ecosystems, we examined four scenarios: business as usual, forest
regeneration, governance e forest conservation, and governance e
concession for industrial plantations and selective logging.
Although these scenarios are simple, they are based on the current
forest process, forest plan and policies of the province. These
scenarios connect the spatial modeling process and planning objectives such as land conservation for nature reserve and forest
land concession for other economic activities, which are the

Author's personal copy

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

necessary aspects for preparing land development alternatives in

the future. The model generated the perspective remaining forest
and deforestation landscapes under the four scenarios for the next
two decades. The patterns mirror where and how the deforestation in the province is heading by 2030 and provides time-series
crucial information on forest landscape under the different scenarios for future landscape management projects. Results presented dynamic patterns and composition of deforestation
alternatives for different time periods. The perspective views of
the province generated by the model highlighted the need for
forest/environmental planning controls for the conservation of
environmentally sensitive areas.
In addition, the analytical approaches presented in this research
may help to identify the locations and options for REDD programs
to achieve multiple benets, based on conditions at a given time. It
is important in REDD framework to understand how future
deforestation process may affect forest patterns in long run while
implementing different plans, policies, and strategies. Scenario
analyses enabled exploring a range of potential issues and outcomes, in order to help visualize possible patterns of forest land
cover change and their impacts in time and space. Each scenario in
this study derives a set of spatially explicit deforestation storylines
with different possibilities incorporating simulations that help
people to grasp the possible risks and possibilities involved in
particular courses of action by testing forest plans and strategies.
For example, the BAU scenario can be combined with forest carbon
stock in order to set reference emission scenario. Consequently, the
simulated future forest cover changes under different simulation
scenarios provide important insights to decision makers for REDD
preparedness activities. For instance, the different simulation scenarios have provided spatial as well quantitative time-series
regional-scale forest cover changes, which would be useful for
the target areas that require immediate intervention. However,
scenarios incorporating more local policies and public perceptions
may be required for developing a more credible storyline for
REDD in action.
The WofE approach used in this modeling assumes that the
changes in forest lands follow linear patterns, while forest change
process may attain a non-linear trend in some regions (Carmona &
Nahuelhual, 2012; Khoi & Murayama, 2011). Therefore, any land
cover change model should be designed with caution. Although the
performance of the model shows good agreement between the
spatial patterns of modeled and reference maps, the resulting
patterns are inuenced by the underlying driving factors where
some of them may change over time. For example, road accessibility is a major driving factor of forest change in this study. Owing
to the lack of spatial information on future roads plan, this research
assumed that the existing roads will be constant and inuence
equally overtime. However, in reality, new roads or water channels
may be built in future, providing transportation access and
impacting forest pattern changes. Furthermore, we considered the
deforestation process as a two-state process except in the FR scenario, i.e., the conversion of forest land to non-forest. The regrowth
class in land cover maps used in FR scenario is framed with assumptions. Inclusion of more forest types change process will
enable to better decision making process as well as improve the
accuracy of forest carbon stocks assessment. Despite these limitations, this study has presented the usefulness of SAR data and the
optimization technique for alternative forest preservation in the
future, which are valuable for environmental planning and natural
resources management. The modeling process presented in this
study is simple and easy to replicate when necessary, can be used
for analyzing forest land cover changes in other tropical regions
where the amount and quality of spatial information and other
ancillary data are very limited.

177

Acknowledgments
We are thankful to Prof. I Nengah Surati Jaya and Edwin Setia
Purnama from the Faculty of Forestry, Bogor Agricultural University, Indonesia and Yumiko Uryu from WWF-Indonesia for sharing
their thoughts and eld experiences in this area of research. We
wish to thank to the anonymous reviewers for their creative comments and suggestions, which helped us to improve this
manuscript.
References
Almeida, C., Gleriani, J., Castejon, E., & Soares-Filho, B. (2008). Neural networks and
cellular automata for modeling intra-urban land use dynamics. International
Journal of Geographical Information Science, 22(9), 943e963.
Angelsen, A. (2008). REDD models and baselines. International Forestry Review,
10(3), 465e475.
Bonham-Carter, G. (1994). Geographic information systems for geoscientists: Modeling
with GIS. New York: Pergamon.
BPS. (2010). Hasil Sensus Pnduduk 2010: Data Agregat per Provinsi (in Indonesian
language).
http://dds.bps.go.id/eng/download_le/SP2010_agregat_data_
perProvinsi.pdf. Accessed 07.03.12.
Carmona, A., & Nahuelhual, L. (2012). Combining land transitions and trajectories in
assessing forest cover change. Applied Geography, 32(2), 904e915.
Chowdhury, R. R. (2006). Driving forces of tropical deforestation: the role of
remote sensing and spatial models. Singapore Journal of Tropical Geography,
27(1), 82e101.
Costanza, R. (1989). Model goodness of t: a multiple resolution procedure.
Ecological Modelling, 47, 199e215.
Echeverria, C., Coomes, D. A., Hall, M., & Newton, A. C. (2008). Spatially explicit
models to analyze forest loss and fragmentation between 1976 and 2020 in
southern Chile. Ecological Modelling, 212(3e4), 439e449.
FAO. (2010). Global forest resources assessment. In FAO Forestry paper No. 163: . Rome:
Food and Agriculture Organization of the United Nations, 378 pp.
Hagen, A. (2003). Fuzzy set approach to assessing similarity of categorical maps.
International Journal of Geographical Information Science, 17(3), 235e249.
Harris, N. L., Brown, S., Hagen, S. C., Saatchi, S. S., Petrova, S., Salas, W., et al. (2012).
Baseline map of carbon emissions from deforestation in tropical regions. Science, 336(6088), 1573e1576.
Hoekman, D. H., Vissers, M. A. M., & Wielaard, N. (2010). PALSAR wide-area mapping of Borneo: methodology and map validation. IEEE Journal of Selected Topics
in Applied Earth Observations and Remote Sensing, 3(4), 605e617.
IPCC. (2007). Climate change 2007: mitigation of climate change. In B. Metz,
O. R. Davidson, P. R. Bosch, R. Dave, & L. A. Meyer (Eds.), Contribution of working
Group III to the fourth assessment report of the intergovernmental panel on climate
change). Cambridge: Cambridge University Press.
Jaimes, N. B. P., Sendra, J. B., Delgado, M. G., & Plata, R. F. (2010). Exploring the
driving forces behind deforestation in the state of Mexico (Mexico) using
geographically weighted regression. Applied Geography, 30(4), 576e591.
Kamusoko, C., Oono, K., Nakazawa, A., Wada, Y., Nakada, R., Hosokawa, T., et al.
(2011). Spatial simulation modelling of future forest cover change scenarios in
Luangprabang Province, Lao PDR. Forests, 2(3), 707e729.
Khoi, D. D., & Murayama, Y. (2011). Modeling deforestation using a neural networkMarkov model. In Y. Murayama, & R. B. Thapa (Eds.), Spatial analysis and
modeling in geographical transformation process: GIS-based applications (pp.
169e190). Dordrecht: Springer.
Lambin, E. F., & Meyfroidt, P. (2011). Global land use change, economic globalization,
and the looming land scarcity. Proceedings of the National Academy of Sciences,
108(9), 3465e3472.
Maeda, E. E., Almeida, C. M., Ximenes, A. C., Formaggio, A. R., Shimabukuro, Y. E., &
Pellikka, P. (2011). Dynamic modeling of forest conversion: simulation of past
and future scenarios of rural activities expansion in the fringes of the Xingu
National Park, Brazilian Amazon. International Journal of Applied Earth Observation and Geoinformation, 13(3), 435e446.
Mas, J. F., Puig, H., Palacio, J. L., & Sosa-Lpez, A. (2004). Modelling deforestation
using GIS and articial neural networks. Environmental Modelling & Software,
19(5), 461e471.
MoFor. (2008). IFCA consolidation report: Reducing emissions from deforestation and
forest degradation in Indonesia. Jakarta, Indonesia: Forestry Research and
Development Agency.
Norad. (2011). Real-time evaluation of Norways international climate and forest
initiative: Contributions to national REDD processes 2007e2010. Country
Report: Indonesia, Oslo, Norway.
Pontius, R. G., Boersma, W., Castella, J., Clarke, K. C., de Nijs, T., Dietzel, C., et al.
(2008). Comparing the input, output, and validation maps for several models of
land change. Annals of Regional Science, 42, 11e47.
Rosenqvist, A., Shimada, M., Ito, N., & Watanabe, M. (2007). ALOS PALSAR: a
pathnder mission for global-scale monitoring of the environment. IEEE
Transactions on Geoscience and Remote Sensing, 45(11), 3307e3316.
Saatchi, S. S., Harris, N. L., Brown, S., Lefsky, M., Mitchard, E. T. A., Salas, W., et al.
(2011). Benchmark map of forest carbon stocks in tropical regions across three

Author's personal copy

178

R.B. Thapa et al. / Applied Geography 41 (2013) 168e178

continents. Proceedings of the National Academy of Sciences, 108(24), 9899e

9904.
Shimada, M. (2011). Model-based polarimetric SAR calibration method using forest
and surface-scattering targets. IEEE Transactions on Geoscience and Remote
Sensing, 49(3), 1712e1733.
Shimada, M., Isoguchi, O., Tadono, T., & Isono, K. (2009). PALSAR radiometric and
geometric calibration. IEEE Transactions on Geoscience and Remote Sensing,
47(12), 3915e3932.
Shimada, M., & Ohtaki, T. (2010). Generating large-scale high-quality SAR mosaic
datasets: application to PALSAR data for global monitoring. IEEE Journal of
Selected Topics in Applied Earth Observations and Remote Sensing, 3(4), 637e656.
Soares-Filho, B. S., Cerqueira, G. C., & Pennachin, C. L. (2002). DINAMICA-a stochastic
cellular automata model designed to simulate the landscape dynamics in an
Amazonian colonization frontier. Ecological Modelling, 154(3), 217e235.
Soares-Filho, B., Moutinho, P., Nepstad, D., Anderson, A., Rodrigues, H., Garcia, R.,
et al. (2010). Role of Brazilian Amazon protected areas in climate change
mitigation. Proceedings of the National Academy of Sciences, 107(24), 10821e
10826.
Soares-Filho, B. S., Rodrigues, H., & Costa, W. (2009). Modeling environmental dynamics with Dinamica EGOAvailable at www.csr.ufmg.br/dinamica.
Soares-Filho, B. S., Rodrigues, H., & Follador, M. (2013). A hybrid analytical-heuristic
method for calibrating land-use change models. Environmental Modelling &
Software, 43, 80e87. http://dx.doi.org/10.1016/j.envsoft.2013.01.010.
Thapa, R. B. (2012). Weight of evidence in geospatial analysis. In Y. Murayama (Ed.),
Progress in geospatial analysis (pp. 97e106). Tokyo: Springer.
Thapa, R. B., Itoh, T., Shimada, M., Watanabe, M., Takeshi, M., & Shiraishi, T.
(2013). Evaluation of ALOS PALSAR sensitivity for characterizing natural

forest cover in wider tropical areas. Remote Sensing of Environment. submitted for publication.
Thapa, R. B., & Murayama, Y. (2009). Land use change factors in Kathmandu Valley:
a GWR approach. In B. G. Lees, & S. W. Laffan (Eds.), Proceedings of the 10th
international conference on GeoComputation (pp. 255e260). Sydney: The University of New South Wales.
Thapa, R. B., & Murayama, Y. (2012). Scenario based urban growth allocation in
Kathmandu Valley, Nepal. Landscape and Urban Planning, 105(1e2), 140e148.
Thapa, R. B., Shimada, M., Motohka, T., Watanabe, M., & Shiraishi, T. (2011). Spatial
modeling of tropical deforestation using PALSAR data. In 32nd Asian conference
on remote sensing 2011 (pp. 1468e1471). ACRS, 2011(2).
Trimble. (2012). Reference book: eCognition developer 8.7.1. Munchen: Trimble Germany GmbH.
Uryu, Y., Purastuti, E., Laumonier, Y., Sunarto, S., Budiman, A., Yulianto, K., et al.
(2010). Sumatras forests, their wildlife and the climate. Jakarta, Indonesia: WWF
Indonesia Technical Report.
Venema, H. D., Calamai, P. H., & Fieguth, P. (2005). Forest structure optimization
using evolutionary programming and landscape ecology metrics. European
Journal of Operational Research, 164(2), 423e439.
Verburg, P. H., Eickhout, B., & van Meijl, H. (2008). A multi-scale, multi-model
approach for analyzing the future dynamics of European land use. Annals of
Regional Science, 42, 57e77.
Walker, W. S., Stickler, C. M., Kellndorfer, J. M., Kirsch, M. K., & Nepstad, D. C. (2010).
Large area classication and mapping of forest land cover in the Brazilian
Amazon: a comparative analysis of ALOS/PALSAR and Landsat data sources. IEEE
Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 3(4),
594e604.