Applied Geography 91 (2018) 99–110

Contents lists available at ScienceDirect

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

Environmental predictors of forest change: An analysis of natural T

predisposition to deforestation in the tropical Andes region, Peru
Vincent Baxa,∗, Wendy Francesconib
a
Universidad de Ciencias y Humanidades, Av. Universitaria 5175, Los Olivos, Lima39, Peru
b
International Center for Tropical Agriculture, Av. La Molina 1895, La Molina, Lima, Peru

A R T I C L E I N F O A B S T R A C T

Keywords: The spatial patterns of deforestation are usually non-randomly distributed across the landscape. While anthro-
Forest cover change pogenically driven processes are often addressed in land-use regulation policies and deforestation research, less
Conservation areas attention is given to the environmental factors that influence tropical deforestation. This study investigates to
Environmental deforestation drivers what extent climate conditions (temperature and precipitation) and biophysical landscape characteristics (ele-
Tropical montane forests
vation, slope, soil type, forest type, and distance to rivers) facilitate or mitigate deforestation processes in Peru's
Random Forest
tropical Andes. A Random Forest regression model was constructed for the entire Peruvian tropical Andes, and
separate models were developed for some of the known direct deforestation drivers in the region (coca pro-
duction, gold mining, and land-use by indigenous and non-indigenous communities). Soil type and precipitation
were identified as the most important deforestation predictors when the entire Peruvian tropical Andes was
considered, whereas distance to rivers was associated with deforestation by mining activities, and elevation and
temperature with coca cultivation areas. Using the regression results, a Random Forest classification model was
constructed to locate areas where the composition of environmental factors could either facilitate or mitigate
deforestation processes. It was found that almost 85% of the forests classified as having high to very high
probability to deforestation were located outside current protected areas. In order to increase conservation
impacts, the results suggest that greater consideration should be given to the distribution of environmental
factors when designing land-use regulation policies and establishing protected areas.

1. Introduction
Ongoing processes of forest destruction across tropical regions pose
a major threat to biodiversity, climate stability and the functioning of
biogeochemical and hydrological cycles (Bonan, 2008; Malhi et al.,
2008). In spite of the worldwide recognition of this environmental
problem and the implementation of manifold initiatives to halt further
reduction of tropical forest areas, the rates of deforestation in the tro-
pics have remained consistently at high levels (Achard et al., 2014;
Sloan & Sayer, 2015). Much attention goes to the tropical forest eco-
systems in Central and South America, which harbor some of the
world's greatest amounts of species diversity (Poveda, Álvarez, &
Rueda, 2011; Ribeiro et al., 2011, pp. 405–434) and constitute the
largest portion of the global terrestrial carbon sink (Pan et al., 2011).
Between 2000 and 2010, the extent of forest cover in Central and South
America shrunk by approximately 55,000 km2 per year, with important
deforestation hotspots located in northern Argentina, southeastern
Bolivia, western Paraguay and the Brazilian Amazon (Aide et al., 2013).
According to recent land-use change scenarios for the Neotropics
(Soares-Filho et al., 2013), in particular the lowland Amazon basin and
forests located at the Andean foothills will experience extensive forest
losses in the future, whereas forest recovery could be expected in the
highland forests of the Andean mountain range (Middendorp, Pérez,
Molina, & Lambin, 2016; Sanchez-Cuervo & Aide, 2013a). These dy-
namics of Neotropical forest cover change will largely shape the future
well-being of people relying directly on forest ecosystem services.
In the last decades, substantial efforts have been made by re-
searchers to determine why forest change happens and why the patterns
and rates of forest change vary across the landscape (Rudel, 2007). An
important contribution to the deforestation literature is the work by
Geist and Lambin (2002), which conceptualizes the dynamics between
fundamental social processes such as population changes, and human
activities or actions at the local level with direct impacts on forest cover
such as logging and agricultural expansion. Proximate causes and un-
derlying driving forces typically relate to anthropogenic systems,
however, environmental factors are also recognized to play a crucial
role in the process of forest cover change. Geist and Lambin (2002)
reviewed 152 subnational case studies from the tropical belt, out of

∗
Corresponding author.
E-mail addresses: vbax@uch.edu.pe (V. Bax), w.francesconi@cgiar.org (W. Francesconi).

https://doi.org/10.1016/j.apgeog.2018.01.002
Received 16 October 2017; Received in revised form 29 December 2017; Accepted 9 January 2018
Available online 06 February 2018
0143-6228/ © 2018 Elsevier Ltd. All rights reserved.
V. Bax, W. Francesconi
which a third reported a link between deforestation and factors asso-
ciated with the biophysical environment, including a range of land-
scape attributes and climate variables. The effect of these environ-
mental factors is linked to human behavior at the local level, as they set
the necessary conditions for land-use change processes to occur, and
place physical thresholds on the types of land-use practices that are
feasible in a region (Aide et al., 2013). While this emphasizes the im-
portance of considering landscape elements and climate conditions at
local scales, the mainstream deforestation literature is particularly or-
iented towards analyzing the political, economic, and social context in
which forest change processes occur (Jusys, 2016; Lambin et al., 2001;
Robinson, Holland, & Naughton-Treves, 2014; Rudel, Defries, Asner, &
Laurance, 2009).
Nonetheless, a great variety of models have been developed
throughout the years to describe the role of the natural environment in
the deforestation process (Busch & Ferretti-Gallon, 2017; Kaimowitz &
Angelsen, 1998). The environmental variables commonly included in
models relate to land accessibility, land suitability and climate varia-
bility. Landscape characteristics that provide natural access routes to
forests include rivers and lakes (Salonen, Toivonen, Cohalan, &
Coomes, 2012), especially in areas where roads and other infrastructure
is scarce (Armenteras, Rudas, Rodriguez, Sua, & Romero, 2006). Fur-
thermore, forests located along the coastline that were better accessible
were found to be subjected to more deforestation compared to main-
land forests (Rudel & Roper, 1996). Elevation and slope gradients have
been associated with forest accessibility and deforestation as well (Bax,
Francesconi, & Quintero, 2016), although they particularly determine
the suitability of the land for productive activities (Pope et al., 2015).
Higher sloped terrain is less attractive for agriculture, given that har-
vests are generally lower (Barrowclough et al., 2016) and working the
land requires greater efforts and resources (Grau, Kuemmerle, &
Macchi, 2013). The relationship between deforestation and topography
is likely to become weaker through time when low-lying lands become
scarcer and exhausted, leaving people no other choice but to move to
steeper areas. Deforestation induced by the suitability of the land is also
determined by the quality of the soil, mainly within the context of
agricultural production (Laurance et al., 2002), and by forest type
(Chowdhury, 2006), given the potential preferences of loggers for tree
species with high economic value (Asner et al., 2005). Climate also
seems to affect deforestation through local variations in precipitation,
temperature, and dry season severity. Precipitation and dry season se-
verity can either have a hindering or an enabling effect: less rainfall
results in dryer forests which are easier to burn (Aragao et al., 2008)
while deficit or excessive rainfall tends to reduce crop yields (Grau,
Gasparri, & Aide, 2005). On the other hand, areas characterized by
moderate local temperatures provide desirable conditions for estab-
lishing human settlements, which transform the natural landscape
(Armenteras, Rodríguez, Retana, & Morales, 2011).
Given that the tropical Andes region is characterized by a great
variation in altitude, forest structure, temperature and rainfall patterns,
the way in which land cover transformations are being undertaken
could be related to these environmental attributes. A better under-
standing of nature's influence on deforestation decision making (here
defined as the decision of land-managers to conserve or convert forest)
is needed, as it is currently not adequately addressed in land-use reg-
ulation policies (Joppa & Pfaff, 2009; Miteva, Pattanayak, & Ferraro,
2012). However, studies specifically focusing on the environmental
dimensions of deforestation are scarce. Generally more attention is
given to human-related drivers and causes. To the best of our knowl-
edge, the work by Rolett and Diamond (2004) may be the best known
study that focuses on the effect of predisposing environmental factors
on forest transitions. Hence, understanding the biophysical and climate
context of forest cover change in the highly diverse landscape of the
Andean mountains could advance our understanding of montane forest
management. Furthermore, most studies on deforestation in the Neo-
tropical region focus on lowland Amazon ecosystems, while
100
Applied Geography 91 (2018) 99–110
information on Andean ecosystems remains limited (Armenteras et al.,
2011; Zuluaga & Rodewald, 2015). In particular, very few studies have
analyzed deforestation practices in the Peruvian Andes region. Current
rates of forest cover change and hotspot locations in the Peruvian Andes
are not provided in the scientific literature, and the drivers associated
with these changes are not clearly understood (Robiglio, Armas, Silva
Aguad, & White, 2014). Nonetheless, the tropical Andes have been
identified as the most critical biodiversity hotspot on the planet in
terms of plant and vertebrate species richness (Myers, Mittermeier,
Mittermeier, Da Fonseca, & Kent, 2000), which emphasizes the im-
portance of investigating deforestation dynamics in this region. More
specifically, the objectives of this study were to 1) identify and examine
the environmental factors that facilitate or mitigate deforestation in the
tropical Peruvian Andes; 2) analyze the influence of these environ-
mental factors on some of the known direct deforestation drivers in the
region; and 3) map the areas where the natural landscape facilitates or
mitigates deforestation.
2. Methods
2.1. Study area
Peru's tropical forest region can be disaggregated into the lowland
Amazon consisting of humid forests at low elevations, the northern
coast region consisting of dry forests, and the tropical Andes region
consisting of sub-tropical forests located along the eastern slopes and
valleys of the Andean mountain range. The Peruvian tropical Andes are
located between coordinates 3˚5′10 South, 79˚1′15 West, 14˚29′24
South and 68˚49′37 West, with most of the forests located at elevations
ranging from 500 up to 3000 m.a.s.l. (MINAM, 2015). The tropical
Andes extend over approximately 190,000 km2, which represents about
15% of the national territory (Fig. 1). Rainfall varies considerably, from
300 mm to 6000 mm per year, with the wettest areas found in the south
and the center at elevations between 500 and 800 m.a.s.l. Peru's Min-
istry of the Environment distinguishes twenty-three different types of
mature forest within the elevation range of 500–3000 m.a.s.l., including
tropical floodplain forests, high terraces, palm tree forests and cloud
forests (MINAM, 2015). Other natural ecosystems include Andean sa-
vanna vegetation and grasslands mainly located at higher elevations.
Soils with good biophysical properties for agricultural production, in-
cluding eutric and dystric cambisols (FAO-Unesco, 1990, p. 60) are
found at lower elevations along the eastern edge of the tropical Andes
and in central regions of the San Martín and Junín departments.
Peru currently comprises a total of 40 national parks, communal
reserves, national monuments and protected forests, out of which 22
are partly or completely located within the tropical Andes, covering
about 25% of the area. Road density in the region is relatively low
(8 km per 100 km2) owing to the highly sloped and arboreal landscape.
In consequence, the region is sparsely populated: an estimated 1.2
million people (INEI, 2015) are scattered over hundreds of small set-
tlements and a few bigger urban centers such as the cities of Rioja and
Moyobamba in the San Martín department, and the city of Tingo María
located in Huánuco. The rural population can be subdivided in in-
digenous inhabitants who originate from the region, and non-in-
digenous or colonist communities from varying regions of Peru. Rural
livelihoods are largely based on small scale and commercial agriculture,
as well as livestock farming activities. Some of the principal crops in-
clude coffee, which is produced at elevations ranging from 700 to
1200 m.a.s.l. primarily in the departments of Junín, Cusco, and San
Martín (Tulet, 2010); while rice and cassava are primarily cultivated in
San Martín, Loreto and Amazonas (INEI, 2014), and maize in San
Martín, Huánuco, Loreto and Junín (INEI, 2014). The tropical Andes are
marked by a long history of widespread cultivation of illicit coca leaves,
with massive impacts on forests and other biological diversity. The
production of cocaine has been accountable for large scale ecosystem
degradation in the Cusco region in the 1960s (Dourojeanni, 1992) and
V. Bax, W. Francesconi
in Huánuco and the Upper Huallaga Valley in San Martín since the
1980s (Young, 1996). Although coca cultivation takes place throughout
the entire Peruvian tropical Andes, current production sites are con-
centrated in the valleys of the Apurímac, Ene and Mantaro Rivers, lo-
cated in between the departments of Cusco, Huancavelica and Aya-
cucho. Other known deforestation drivers include fossil fuel extraction
(Cuba, Bebbington, Rogan, & Millones, 2014), with most of the active
oil and gas projects located in the center and the south of the tropical
Andes. In addition, informal artisanal gold mining activities take place
in the far north of the department of Amazonas (Alvarez-Berríos & Aide,
2015) and in the south of the Madre de Dios department (Swenson,
Carter, Domec, & Delgado, 2011).
2.2. Forest change 2017
A 2011 nationwide land cover layer produced by Peru's Ministry of
the Environment (MINAM, 2015) was used as a base-line map to
compare forest and non-forest areas and to identify the broad variety of
forest types found in the tropical Andes. This layer was produced on the
basis of Landsat 5 TM satellite imagery from 2011 at 30 m resolution,
combined with RapidEye and Google Earth imagery at approximately
5 m resolution. Vegetation cover change datasets based on satellite
detection, from Global Forest Watch (Hansen et al., 2013) and Terra-i
(Reymondin et al., 2012), were imported into ArcGIS version 10.1 to
identify deforestation sites between 2011 and 2017. These datasets
were projected to the WGS 1984 UTM Zone 18 S coordinate system,
clipped to the region of interest, and merged with the MINAM land
cover map. This resulted in a layer representing deforestation in the
101
Applied Geography 91 (2018) 99–110
Fig. 1. Distribution and extent of forest areas in the
Peruvian tropical Andes (approximately 500–3000 m.a.s.l.).
Peruvian tropical Andes up to the year 2017, at a 30 m resolution. To
globally estimate and confirm the reliability of the final deforestation
layer, extensive visual inspection of Landsat 7 ETM+ satellite imagery
with less than 10% cloud cover from the period 2012–2017 along with
Google Earth imagery from 2017 was carried out. All other natural land
cover types within the tropical Andes (< 2%) were removed from the
final layer as our analysis specifically focused on the transition from
forest to non-forest areas. In addition, all areas located at elevations
beyond 3000 m.a.s.l. were excluded from further analysis.
2.3. Environmental predictors of deforestation
Spatial datasets for several candidate biophysical and climate pre-
dictor variables were collected from different data sources (Table 1).
The relevance of taking these variables into account in the analysis was
assessed based on a preliminary literature review focused on defor-
estation in the Neotropics and tropical mountain regions (e.g.
Armenteras et al., 2011; Laurance et al., 2002). The datasets were im-
ported into ArcGIS and pre-processed in order to transform raw data
into suitable formats for later analyses (Table 1). A fishnet polygon grid
composed of 22,680 cells of 9 km2 was generated and used as an
overlay surface to extract the spatial data for the deforestation and
candidate predictor variables into a tabular database. At this spatial
scale, sufficient level of detail is provided to effectively relate the spa-
tial patterns of deforestation activity to the environmental variables
examined, while ensuring correct alignment of the different predictor
variable and deforestation datasets, which could be problematic in
fragmented landscapes (Kaimowitz, Mendez, Puntodewo, & Vanclay,
V. Bax, W. Francesconi
Table 1
Details of the spatial data used in this study.
Dataset Date Source
Response variable
Deforestation (%) 2011–2017 MINAMb, Terra-ic, GFWd
Biophysical predictor variables
Elevation (m.a.s.l.) 2011 ASTER GDEM V2e
Slope (degrees) 2011 ASTER GDEM V2e
Distance to rivers (meters) 2011 MINAMb
Soil type (classes) 1981 ONERNf
Mature forest type (classes) 2011 MINAMb
Climate predictor variables
Annual precipitation (mm/year) 1970–2000 WorldClim V2g
Average temperature (°C) 1970–2000 WorldClim V2g
a
Classes are provided in the supplementary material.
b
Ministerio del Ambiente (MINAM, 2015).
c
Terra-i (Reymondin et al., 2012).
d
Global Forest Watch (Hansen et al., 2013).
e
Advanced Spaceborne Thermal Emission and Reflection Radiometer Global DEM V2 (NASA METI, 2011).
f
Oficina Nacional de Evaluación de Recursos Naturales (ONERN, 1981).
g
WorldClim Global Climate Data (Fick & Hijmans, 2017).
2002, pp. 41–65). For cells without remaining forest cover, the value
for mature forest type could not be extracted, which resulted in missing
data (5% of the dataset). To account for this, the missing forest type
values were imputed using the bagImpute function implemented in the
caret package within the R statistical software (R Development Core
Team, 2016).
To better understand the potential relation between the selected
environmental variables and deforestation, a number of spatially ex-
plicit Random Forest (RF) (Breiman, 2001) regression models were
constructed using the randomForest package in R. This method was
selected owing to its outstanding predicting performance, its abilities to
deal with non-parametric and categorical data, its build-in measures to
rank the relative importance of predictor variables and its robustness to
spatial autocorrelation, overfitting and correlated predictor variables
(Evans, Murphy, Holden, & Cushman, 2011, pp. 139–159; Svetnik et al.,
2003). While particularly popular for land cover classification and
ecological applications (e.g. Nanni, Gasparri, & Grau, 2015), RF has
been used in a number of prior studies to assess the effect of predictor
variables on land cover change (e.g. Oliveira, Oehler, San-Miguel-
Ayanz, Camia, & Pereira, 2012; Sánchez-Cuervo & Aide, 2013b).
A RF model is composed of a combination of decision trees. The
individual trees are generated on the basis of bootstrapped training
samples consisting of approximately two thirds of the dataset (in our
case the values for deforestation and environmental predictors ex-
tracted from the 9 km2 grid cells). Within each tree, a random subset of
environmental predictor variables is selected to determine the split.
This generates a diverse set of decision trees which is needed to reduce
correlation between trees and the associated generalization error
(Breiman, 2001). To solve regression problems with a RF, the output
values of each individual tree are averaged to determine the final de-
forestation predictions. The remaining training data, approximately a
third of the original dataset called out-of-bag (OOB), is employed to
validate the prediction accuracy.
A model for the entire tropical Andes was developed using all
22,680 observations in the dataset. In addition, separate models were
developed to characterize some of the most important known direct
deforestation drivers in the region (Piu & Menton, 2014) (Table 2). We
developed deforestation models associated with indigenous and non-
indigenous communities as a proxy for land-use variables such as
agriculture and livestock. Indigenous and non-indigenous communities
were modeled separately, since indigenous communities are believed to
engage in less destructive forest management practices and livelihoods
compared to those adopted by non-indigenous communities (Porro,
Lopez-Feldman, & Vela-Alvarado, 2015). To verify this, the average
accumulative deforestation within a radius of 5 km from both
102
Applied Geography 91 (2018) 99–110
Scale Data processing procedure
30 m Merge of datasets into deforestation layer
30 m Set negative elevation values to null
30 m Identify rate of maximum change in z-value in elevation layer
30 m Euclidean distance to river
15 classesa Polygon to raster layer
23 classesa Polygon to raster layer
30 s Merge of datasets into average annual precipitation raster layer
30 s Merge of datasets into average annual temperature raster layer
community types was compared. Georeferenced point data of legally
entitled communities was used to locate the communities. To ensure a
fair land-use comparison, only 9 km2 grid cells with either an in-
digenous community or a non-indigenous community were selected for
the analysis. Other RF deforestation models developed looked at co-
caine production in the departments of Huánuco, Ucayali, Ayacucho
and Cusco, and artisanal gold mining in the south of Madre de Dios. For
both cocaine production and artisanal gold mining sites, data selection
was based on available cartographic material of the most affected re-
gions. Models for other known deforestation drivers including hydro-
carbon extraction could not be developed due to data limitations.
To run the models, two input parameters need to be specified: ntree
(the number of trees to grow) and mtry (the number of variables used to
split each node). To specify the models, we set the ntree parameter to
2000 to ensure stable results (500 trees is the default) and we de-
termined the optimal value for mtry by dividing the number of pre-
dictors (7) by 3 as suggested by Liaw and Wiener (2002). To measure
variable importance, the percentage of increase in Mean Square Error
(MSE) was computed by randomly permuting OOB data for a given
variable while leaving the other variables unchanged. Analysis of par-
tial dependence plots for each of the predictor variables was carried out
to estimate its direction and magnitude on the response variable.
2.4. Natural predisposition to deforestation
To locate the areas where natural characteristics embedded in the
landscape either facilitate or mitigate the deforestation process in the
Peruvian tropical Andes, a probabilistic Random Forest classification
model was developed using the R statistical program. To our knowl-
edge, only Di Lallo, Mundhenk, Zamora López, Marchetti, and Köhl
(2017) used RF for predicting deforestation probability, and hence,
the full potential of this approach is yet to be explored. RF classifi-
cation differs from RF regression, since the final prediction in the re-
sponse variable is determined by majority vote instead of averaging
the output values of each individual tree (Liaw & Wiener, 2002).
Raster datasets for the seven predictor variables (Table 1) were re-
sampled to a pixel size of 30 × 30 m and merged into a multi-band
raster image. Based on the deforestation raster dataset, a data sample
composed of 6000 forest and 6000 deforested sites was selected to
train the model, using a random points data layer generated in ArcGIS.
The same values for the input parameters ntree and mtry were spe-
cified as used for the RF regression models (2000 and 2 respectively).
The resulting model was applied to generate a deforestation prob-
ability map based on natural elements of the landscape. Four classes
were defined, ranging from very high to very low deforestation
V. Bax, W. Francesconi Applied Geography 91 (2018) 99–110

probability, using the natural break classification scheme to determine

Asner, Llactayo, Tupayachi, and Luna (2013)

the threshold pixel values (Pourtaghi, Pourghasemi, & Rossi, 2015).
To assess deforestation probability in protected areas, the extent of the
pixels belonging to each of the four classes within and outside pro-
tected areas was calculated in ArcGIS. Classification accuracy was
evaluated using the OOB estimate of error rate, where lower estimates
imply higher accuracies.

3. Results
Data source

UNODCb

The merge of the MINAM land cover map with the Global Forest
Watch and Terra-i datasets yielded a layer representing deforestation in
IBCa

IBCa
–

the Peruvian tropical Andes until the year 2017 (Fig. 2). On the basis of
this layer, the total extent of cleared forests equaled 44,200 km2, which
corresponds to 23% of the studied area. Between 2011 and 2017, more
environmental context of land-use change near non-indigenous

than 1150 km2 of forests were converted into other land-uses, yielding
environmental context of deforestation in the tropical Andes
environmental context of land-use change near indigenous

an annual deforestation rate of 19,300 ha per year (0.1%). This implies

that compared to the previously reported rate of 0.5–1.0% during the
1990–1997 period (Achard et al., 2002), deforestation in the Peruvian
tropical Andes has decreased. The most severely affected areas are lo-
environmental context of artisanal gold mining

cated in the departments of San Martín, Huánuco and Junín. In these

environmental context of coca production

areas, local deforestation rates of 2% were estimated by this study. In

contrast, the forests in the very north and the very south have been less
vulnerable to deforestation, which is reflected by lower deforestation
rates of about 0.2%. Protected areas show little change in land cover
compared to non-protected areas, with local deforestation rates ranging
from 0.02% to 0.04%.

3.1. Environmental predictors of deforestation

Based on the seven environmental variables included in the tropical

Model description

Andes RF model, 67% of the variance in deforestation could be ex-

plained. The model returned a mean of squared residuals equivalent to
Examines the
Examines the

Examines the

Examines the
Examines the
communities

communities

0.04. The OOB error rate estimates were used to measure and rank the
importance of the selected environmental variables in predicting de-
forestation in the tropical Andes (Fig. 3). Soil type and average annual
precipitation were identified as the most important predictors, re-
Dataset size (9 km2 grid cells)

flecting the extensive agricultural-based livelihood systems across most

of the region. The variable representing the type of forest ecosystem has
a marginal effect on deforestation as shown by the low percentage of
increase in MSE. Also the variables for slope, elevation and temperature
seem to have a limited impact on deforestation. Concurrently, ex-
cluding these variables from the analysis reduced the explanatory
22,680

power of the model only by 5%, 3% and 3% respectively, which con-

401

108

387
141

firms their relatively lower effect to the overall model performance.

Partial dependence plots were generated and inspected to interpret
Huánuco, Ucayali, Ayacucho and Cusco

the effect of single variables on the presence of deforestation. The

partial dependence plots of the continuous variables included in the
Throughout the tropical Andes

Throughout the tropical Andes

model (Fig. 4) show that most deforestation occurs at slopes lower than
United Nations Office on Drugs and Crime (UNODC, 2016).

20°, at altitudes lower than 2000 m.a.s.l., in areas with precipitation

Peruvian tropical Andes

less than 2000 mm/year and average temperatures greater than 20 °C.
Further, the impact of rivers was found to follow a U-shaped curve:
deforestation decreases as the proximal distance to rivers increases. Yet,
Madre de Dios

Instituto del Bien Común Peru (IBC, 2016).

there is a turning point at around 12 km, where deforestation increases

Location

as also the distance to rivers increases, reaching a plateau at around

35 km. The partial dependence plots for the categorical variables (soil
type and mature forest type; Fig. 4) are composed of simple bar charts.
Random Forest regression models.

In particular areas associated with eutric cambisols (CMe–VRe) seem to

Non-indigenous communities

be prone to deforestation. Eutric cambisols are mainly located in the

Indigenous communities

department of San Martín and are composed of fertile soils with high
Artisanal gold mining

agricultural potential (FAO-Unesco, 1990, p. 60). As far as forest types,

Coca production
Tropical Andes

the model identified subtropical low-hill forests with bamboo (bosque

de colina baja con paca; Bcb–pa) as most vulnerable, whereas sub-
tropical upland forests (bosque de colina alta; Bca) and high terraces
Model
Table 2

forests (bosque de terraza alta; Bta) are among some of the forest types
b
a

subjected to lower deforestation pressures.

103
V. Bax, W. Francesconi
Fig. 3. Random Forest regression results showing the relative importance of environ-
mental variables in estimating deforestation.
Separate deforestation models were developed to characterize some
of the known deforestation drivers in the tropical Andes (see Table 3).
The model results for indigenous and non-indigenous communities
presented similar environmental conditions in which deforestation oc-
curs. For both types of communities, precipitation and soil type were
among the most important predictors, corroborating the output pro-
vided by the tropical Andes model. The influence of rivers seems to be
more important for indigenous communities than for non-indigenous
ones, as reflected by the higher increase of MSE and the narrower
distance range to the nearest river (5 km and 14 km respectively).
104
Applied Geography 91 (2018) 99–110
Fig. 2. Deforestation status in 2017 in the Peruvian tropical
Andes.
Remarkably, forest type was not a prominent factor explaining defor-
estation activities in the tropical Andes, neither for indigenous nor for
non-indigenous communities. Average accumulative deforestation
within a radius of 5 km from the location of non-indigenous commu-
nities was substantially higher compared to deforestation near in-
digenous communities (70% and 44% respectively), indicating more
extensive pressures on the forests in the surroundings of non-indigenous
communities. At the same time, the environmental variables explained
a greater portion of variance in the deforestation near non-indigenous
than near indigenous communities (64% and 54% respectively).
The deforestation model for coca leaf production explained 82% of
the variance in deforestation. Elevation and average temperature were
the most important predictors, with the most susceptible regions lo-
cated at an elevation range between 500 and 1500 m.a.s.l. and average
temperature range of 18–24 °C. Further, in contrast to the other models,
the distance to rivers variable presents a strictly positive correlation,
indicating that more deforestation takes place further away from rivers.
The artisanal gold mining model output showed that only 33% of
the variance in deforestation could be explained on the basis of the
environmental variables studied. Hence, the extraction of metals seems
to be undertaken irrespective of landscape and climate features such as
slope and rainfall. In turn, the southern Madre de Dios region is char-
acterized by high rainfall patterns (5000–6000 mm per year) and poor
soil fertility, which mitigates the potential for agricultural production.
A strong negative correlation with increased distance to rivers was
determined, as the small scale mining operations are usually carried out
in rivers and alongside the riverbanks.
V. Bax, W. Francesconi Applied Geography 91 (2018) 99–110

Fig. 4. Partial dependence plots for continuous variables (slope, distance to rivers, average temperature, average annual precipitation and elevation) and categorical variables (mature
forest type and soil type), showing the marginal effect on deforestation. The values on the y-axis represent the average of all deforestation predictions (as a decimal fraction) for a given
value on the x-axis. Tick marks on the x-axis of continuous variable plots indicate the data distribution across the variable range in deciles.

3.2. Natural predisposition to deforestation
A RF classifier was applied to identify the areas in the tropical Andes
where natural elements in the landscape facilitate deforestation pro-
cesses and where they mitigate deforestation. The RF classification
model yielded a kappa coefficient of 0.63 and an overall classification
accuracy of 81.4, indicating that more than 80% of the forest and de-
forestation pixels was classified correctly. The deforestation probability
map generated is shown in Fig. 5. Red and orange shaded areas indicate
a combination of environmental factors that could positively influence
deforestation decision making, while the areas in blue represent less
attractive environmental conditions for forest conversion. Prone areas
are located in the northern half of San Martín and in the central/south
of Cusco. Both departments are characterized by fertile soils and low
precipitation rates. In contrast, in the far southeast of the tropical Andes
(Madre de Dios, Puno and the east of Cusco) where overall deforesta-
tion as well as current rates are relatively low, the natural landscape
seems to contribute to mitigating deforestation activities.
Based on the four deforestation probability classes, the extent of the
forests with low or very low probability equals about 105,000 km2. On
the other hand, an area of about 40,000 km2 was classified as having
high or very high deforestation probability (Table 4). Out of the forests

105
V. Bax, W. Francesconi Applied Geography 91 (2018) 99–110

Table 3
Random Forest output for the separate models for indigenous communities, non-indigenous communities, coca production and artisanal gold mining. The values for range/class reflect
where deforestation occurs for the respective variable. An unspecified class for soil type or forest type (–) indicates that deforestation occurs to a similar extent in areas associated with
any given type of soil or forest.

Indigenous communities Non-indigenous communities Coca production Artisanal gold mining

Range/class %Inc MSE Range/class %Inc MSE Range/class %Inc MSE Range/class %Inc MSE

Soil type Eutric cambisols 75.8 Eutric cambisols and regosols 43.5 Eutric regosols and leptosols 47.5 – 5.6
Precipitation (mm) < 2000 54.5 < 2000 35.2 1000–3500 41.8 4500–5500 10.3
Slope (degrees) < 25 29.5 < 25 13.8 < 30 54.2 < 10 12.3
Elevation (m.a.s.l.) > 700 43.0 500–1500 28.2 < 2000 58.8 < 700 14.6
Rivers (km) < 12 53.1 <5 10.1 >3 45.0 <2 23.5
Temperature (˚C) < 23 72.2 < 23 23.6 > 18 56.0 > 22 13.0
Forest type – 13.7 – 12.8 – 21.6 – 6.9
%Var explained 54.2 63.6 82.2 32.6

with high to very high probability to deforestation, owing to the

identified predisposing environmental conditions, only 20% and 11%
respectively are located within existing protected areas. This implies Table 4
Deforestation probability based on natural landscape characteristics within and outside
that more than 32,000 km2 of forest with high or very high deforesta-
protected areas.
tion probability are located outside protected areas. In contrast, 29%
and 44% of forest areas classified as having low and very low defor- Total area Protected forests Non-protected
estation probability respectively, given their soil suitability for agri- (km2) (%) forests (%)
culture, precipitation and river access characteristics, have been de-
Very low probability 62,944 44 56
signated as protected area. Low probability 41,771 29 71
High probability 25,321 20 80
4. Discussion Very high probability 13,815 11 89
Current non-forest area 45,709 2 98
(incl. water bodies)
It is well known that the geophysical variation in topography, cli-
mate and natural access routes such as rivers significantly influences

Fig. 5. Deforestation probability based on natural landscape fea-

tures.

106
V. Bax, W. Francesconi
the spatial patterns of human population and settlement distribution
(Lung, Lübker, Ngochoch, & Schaab, 2013; Small & Cohen, 2004). With
respect to the Andes, archeological evidence points to a very long his-
tory of human occupation of the subtropical Andean forests, especially
at elevations below 1500 m.a.s.l. where environmental conditions are
more conducive to agricultural production and permanent settlement
(Brush, 1982). With the arrival of the Spaniards in the early 1500s,
intensive land-use systems have been introduced (Turner & Butzer,
1992); traces of which are still apparent in today's Andean landscapes.
In agreement with the land-use history in neighboring Andean regions
(Erickson, 1999; Etter & van Wyngaarden, 2000), the environmental
circumstances for initial settlement and the Hispanic colonization of
Peru's tropical Andes have certainly influenced the way in which land-
use trajectories have changed over time.
The results provided by the Random Forest model for the entire
Peruvian tropical Andes show that more than two-thirds of the variance
in deforestation can be explained by the environmental factors ex-
amined. While soil type and rainfall were identified as strong predictor
variables, topographic characteristics such as slope, elevation and the
location of water bodies had a relatively lower influence in deforesta-
tion decision making. The importance of environmental factors may be
associated with the agricultural-based livelihood systems found within
the tropical Andes region. Fertile soils such as eutric cambisols increase
crop yields while rainfall determines the viability of the land for
farming. The best soils for agriculture are found in the department of
San Martín. Not surprisingly, San Martín has the highest deforestation
rates in Peru, driven by a large-scale production of rice, coffee, palm oil
and cacao (INEI, 2014). In terms of precipitation distribution in the
region, high rainfall values can be damaging to crop root systems,
leading to reduced agricultural yields. Furthermore, abundant rainfall
in sloped areas induces soil erosion which makes the land less suitable
for farming (Harden, 2001). At the same time, seasonal lower rainfall
facilitates forest clearing and burning in slash-burn agricultural prac-
tices. Considering the precipitation gradient found in the Peruvian
tropical Andes ranges from 300 to 6000 mm/year (with about 7% of the
area receiving less than 1000 mm/year), our results indicate that de-
forestation was more evident at the lower end of the precipitation
spectrum (i.e. less than 2000 mm/year). This is reflected by the largely
negative correlation with deforestation as shown in the partial depen-
dence plot generated in this study. The lower precipitation areas are
located on the western flank at elevations varying from 1500 to
3000 m.a.s.l. In addition to agriculture, other land-use practices could
be responsible for the large scale forest conversion in these areas. Young
and León (2000) point out that Peruvian montane forests in the ele-
vation range between 1500 and 2500 m.a.s.l. are attractive for timber
extraction and for the establishment of unimproved pastures, which are
activities that might be less dependent on rainfall conditions. While
socioeconomic, institutional and cultural factors may drive deforesta-
tion in the tropical Andes, the results of this study provide evidence that
also the favorable combination of environmental factors affects defor-
estation decision making in places such as San Martín.
In contrast, the lands of Madre de Dios and Puno were found to be
environmentally unattractive for farming, which is confirmed by low
deforestation rates and extensive areas of in-tact primary forest cover.
In Madre de Dios however, informal gold mining is an important ac-
tivity causing deforestation. This activity seems to be less dependent on
landscape or climate characteristics, given that rather the geologic lo-
cation of metals and possibly other mineral resources such as oil and
gas determines where deforestation occurs. Proximal river distance
seems to be the only environmental factor that influences where gold
mining is taking place, as shown by the Random Forest model results
presented in this study. The common sites for coca production, another
known direct deforestation driver in the Peruvian tropical Andes, pre-
sented stronger associations with the natural landscape. We found that
coca cultivation in Huánuco, Ucayali, Ayacucho and Cusco is bound to
specific environmental conditions in terms of temperature (> 18 °C),
107
Applied Geography 91 (2018) 99–110
slope (< 30°) and elevation (< 2000 m.a.s.l.). These habitat conditions
correspond to an optimal environment for coca cultivation in the tro-
pical Andes (Dillehay et al., 2010), pushing coca farming activities to
particular locations. Although coca cultivation also takes place in other
regions in Peru (i.e. the lowland Amazon basin), policy interventions in
the tropical Andes could be targeted based on these environmental
criteria. Proximal river distance had a positive correlation with coca
production sites. Given the illegal nature of coca cultivation, production
sites seem to be located in remote and isolated regions away from rivers
and perhaps other connecting networks such as roads.
Also examined, indigenous and non-indigenous communities across
the tropical Andes carry out deforestation activities in locations that
have similar environmental conditions. Nonetheless, the impact on
forest cover near non-indigenous communities was found to be much
more extensive than near indigenous communities (average accumu-
lative deforestation of 70% versus 44% respectively). Contrary to ex-
pectations, it was not possible to establish a strong link between the
type of forest and deforestation. This suggests that there is not a clear
preference among land-managers to convert a certain type of mature
forest into other land-uses. Nonetheless, particular non-indigenous
communities, mainly composed of migrants from other regions in Peru,
are known to engage in logging operations with dramatic effects on
forest cover (Young & León, 2000). Historically, widespread selective
logging has been an important economic activity in Peru's tropical
Andes and Amazon regions (White, 1978). These logging practices are
more likely to take place in forest types that include tree species with
high commercial value such as mahogany and cedar (Kometter et al.,
2004). Yet, valuable tree species may exist in different forest ecosys-
tems (Brown, Jennings, & Clements, 2003; Dauber, Fredericksen, &
Peña, 2005), which could explain why selective logging may not be
restricted to a particular forest type. Bamboo species for instance,
which have traditionally been used as construction material in Peru
(Escamilla & Habert, 2014), preponderate in five different forest types
as classified by Peru's Ministry of the Environment (MINAM, 2015). On
the other hand, attractive areas for agriculture, given their favorable
slope, precipitation and soil quality conditions, coincide with a broad
range of forest types. This could further explain why forest clearing for
agriculture occurs irrespective of the type of forest ecosystem. More-
over, other factors including the distance to settlements and logging
roads, land-tenure and the likelihood of government intervention may
influence agricultural expansion and the timber harvesting behavior of
loggers (Finley-Brook, 2007; Walker et al., 2013).
Considering the spatial and temporal dimensions of deforestation is
important when designing conservation measures (Apan, Suarez,
Maraseni, & Castillo, 2017). As relatively little attention is paid to en-
vironmental drivers in the deforestation literature, in turn their influ-
ence on deforestation decision making is not well addressed in existing
land-use management plans and regulation efforts, which are important
for the conservation of ecosystem services such as water provisioning
(Francesconi, Uribe, Valencia, & Quintero, 2017). Priority regions for
the establishment of protected areas are commonly justified on the
basis of two considerations: (1) the presence of vulnerable and rare
species, and (2) the vulnerability of the landscape to change (Rodrigues
et al., 2004). Meanwhile, Joppa and Pfaff (2009) have showed that
globally protected areas are usually located in higher, steeper, more
remote and agriculturally unsuitable areas. Also in the tropical Andes, a
large portion of the current protected area network is located in areas
where the natural landscape already deters deforestation probability. In
contrast, almost 85% of the remaining forest areas in the region that are
classified as having high to very high vulnerability to deforestation are
located outside protected areas (these classes correspond to a third of
all forests outside protected areas). Certainly, other considerations in-
cluding opportunity, practicality, cost and/or politics may influence the
selection of areas to designate for protection (Rodríguez & Young,
2000), which could be an explanation for the location bias of current
protected areas in the tropical Andes. Since the year 2000, the Peruvian
V. Bax, W. Francesconi
government expanded the protected area network within the tropical
Andes by 60% to increase conservation efforts of Peru's biological re-
sources. Regardless, better targeting of protected areas, taking into
account the potential and physical limits of lands for productive ac-
tivities would contribute to greater conservation impacts.
A localized assessment of forest areas that are vulnerable to defor-
estation and degradation is needed for conservation planning and for
regulating the sustainable use of forest resources (Etter, McAlpine,
Wilson, Phinn, & Possingham, 2006). Given that land-use change pro-
cesses occur within a specific spatial context, the development of local
models focusing on demarcated areas and distinct deforestation drivers
could be more relevant to policy interventions at local or regional
government levels, compared to models that assume homogeneity
across different regions. With the implementation of a reformed forest
and wildlife law in 2015 (law N˚ 29763), Peru's Ministry of Agriculture
initiated the process of forest-use zoning. Within this context, com-
prehensive information for rational forest management is obtained,
integrating ecological aspects in terms of soil and forest type, ecosystem
fragility and the distribution of biodiversity, with socioeconomic and
cultural aspects linked to the human occupation and activity in forest
areas (MINAGRI, 2016). Considering biophysical factors and climate
conditions in the designation of forest zones should also be part of a
comprehensive approach. Biophysical factors and climate conditions
define many of the environmental limitations and production potentials
of land, while on the other hand, they influence the decision of land-
managers to convert forests into other uses. The results of this study
support the use of soil quality characteristics and the variation in pre-
cipitation rates as dominant factors explaining deforestation in the
tropical Andes. In addition, for localized deforestation drivers such as
cocaine production and gold mining, temperature, elevation and river
distance were also important environmental predictors. Taking these
factors into account through a tailor-made approach would contribute
to the identification of areas prone to deforestation, which can help in
the designation of forest-use regimes and in turn better formulate forest
conservation and management activities.
5. Conclusions
Using the tropical Peruvian Andes as a case study, we provided an
example of how spatially explicit models can be used to characterize
localized deforestation processes at the landscape level. This study
applied Random Forest to examine the spatial patterns of deforestation
and explore which types of deforestation activities are feasible
throughout the region. In contrast to prior deforestation modeling
studies, we specifically examined to what degree the structure of the
natural landscape along with local climate conditions explain defor-
estation patterns, rather than focusing on socioeconomic drivers such as
road infrastructure and human population measures.
The outcomes of this study indicate that the environmental con-
text of the tropical Peruvian Andes region strongly influences the
land-use behavior of land-managers. A generalized Random Forest
regression model shows that throughout the region, precipitation and
the quality of the soil largely determine whether agriculture and
other land-use practices are feasible or not. Consequently, these en-
vironmental factors facilitate the deforestation process in one place,
while posing a natural barrier that hinders the possibilities for pro-
ductive activities in another. In addition to this generalized model,
we developed localized Random Forest models to better understand
to what degree some of the most prominent deforestation activities in
the region are controlled by environmental conditions. Coca culti-
vation was effectively linked to temperature, elevation and slope
characteristics, showing that this activity is undertaken in locations
defined by specific environmental criteria. In contrast, gold mining
practices were found to be less dependent on landscape and climate
characteristics. These varying relationships between predisposing
environmental factors and specific deforestation activities across the
108
Applied Geography 91 (2018) 99–110
region emphasize the importance of assessing deforestation drivers at
local scales.
This study suggests that a greater consideration of environmental
factors as predictors of land-use patterns in regions of highly diversified
natural landscapes such as Peru's tropical Andes, could improve land-
use regulation policies such as forest zone designation and the estab-
lishment of protected areas. Yet, meeting the needs of forest users
without jeopardizing the functioning of ecosystems along with the
provisioning of ecosystem services constitutes a major challenge in
current policy and decision making. Rational and well-informed forest
management supported by empirical research will be critical for the
safeguarding of the remaining tropical forest resources and the nu-
merous livelihoods relying on them.
Conflicts of interest
None.
Acknowledgement
This work was supported by funding provided by Universidad de
Ciencias y Humanidades.
Appendix A. Supplementary data
Supplementary data related to this article can be found at http://dx.
doi.org/10.1016/j.apgeog.2018.01.002.
References
Achard, F., Beuchle, R., Mayaux, P., Stibig, H. J., Bodart, C., Brink, A., et al. (2014).
Determination of tropical deforestation rates and related carbon losses from 1990 to
2010. Global Change Biology, 20(8), 2540–2554.
Achard, F., Eva, H. D., Stibig, H.-J., Mayaux, P., Gallego, J., Richards, T., et al. (2002).
Determination of deforestation rates of the world's humid tropical forests. Science,
297(5583), 999–1002.
Aide, T. M., Clark, M. L., Grau, H. R., López-Carr, D., Levy, M. A., Redo, D., et al. (2013).
Deforestation and reforestation of Latin America and the Caribbean (2001–2010).
Biotropica, 45(2), 262–271.
Alvarez-Berríos, N. L., & Aide, T. M. (2015). Global demand for gold is another threat for
tropical forests. Environmental Research Letters, 10(1), 014006.
Apan, A., Suarez, L. A., Maraseni, T., & Castillo, J. A. (2017). The rate, extent and spatial
predictors of forest loss (2000–2012) in the terrestrial protected areas of the
Philippines. Applied Geography, 81, 32–42.
Aragao, L. E. O., Malhi, Y., Barbier, N., Lima, A., Shimabukuro, Y., Anderson, L., et al.
(2008). Interactions between rainfall, deforestation and fires during recent years in
the Brazilian Amazonia. Philosophical Transactions of the Royal Society of London B
Biological Sciences, 363(1498), 1779–1785.
Armenteras, D., Rodríguez, N., Retana, J., & Morales, M. (2011). Understanding defor-
estation in montane and lowland forests of the Colombian Andes. Regional
Environmental Change, 11(3), 693–705.
Armenteras, D., Rudas, G., Rodriguez, N., Sua, S., & Romero, M. (2006). Patterns and
causes of deforestation in the Colombian Amazon. Ecological Indicators, 6(2),
353–368.
Asner, G. P., Knapp, D. E., Broadbent, E. N., Oliveira, P. J., Keller, M., & Silva, J. N.
(2005). Selective logging in the Brazilian Amazon. Science, 310(5747), 480–482.
Asner, G. P., Llactayo, W., Tupayachi, R., & Luna, E. R. (2013). Elevated rates of gold
mining in the Amazon revealed through high-resolution monitoring. Proceedings of
the National Academy of Sciences, 110(46), 18454–18459.
Barrowclough, M., Stehouwer, R., Alwang, J., Gallagher, R., Mosquera, V. B., &
Dominguez, J. (2016). Conservation agriculture on steep slopes in the Andes: Promise
and obstacles. Journal of Soil and Water Conservation, 71(2), 91–102.
Bax, V., Francesconi, W., & Quintero, M. (2016). Spatial modeling of deforestation pro-
cesses in the central Peruvian Amazon. Journal for Nature Conservation, 29, 79–88.
Bonan, G. B. (2008). Forests and climate change: Forcings, feedbacks, and the climate
benefits of forests. Science, 320(5882), 1444–1449.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
Brown, N., Jennings, S., & Clements, T. (2003). The ecology, silviculture and biogeo-
graphy of mahogany (Swietenia macrophylla): A critical review of the evidence.
Perspectives in Plant Ecology, Evolution and Systematics, 6(1–2), 37–49.
Brush, S. B. (1982). The natural and human environment of the central Andes. Mountain
Research and Development, 2(1), 19–38.
Busch, J., & Ferretti-Gallon, K. (2017). What drives deforestation and what Stops It? A
meta-analysis. Review of Environmental Economics and Policy, 11(1), 3–23.
Chowdhury, R. R. (2006). Driving forces of tropical deforestation: The role of remote
sensing and spatial models. Singapore Journal of Tropical Geography, 27(1), 82–101.
V. Bax, W. Francesconi
Cuba, N., Bebbington, A., Rogan, J., & Millones, M. (2014). Extractive industries, liveli-
hoods and natural resource competition: Mapping overlapping claims in Peru and
Ghana. Applied Geography, 54, 250–261.
Dauber, E., Fredericksen, T. S., & Peña, M. (2005). Sustainability of timber harvesting in
Bolivian tropical forests. Forest Ecology and Management, 214(1), 294–304.
Di Lallo, G., Mundhenk, P., Zamora López, S. E., Marchetti, M., & Köhl, M. (2017). REDD
+: Quick assessment of deforestation risk based on available data. Forests, 8(1), 29.
Dillehay, T. D., Rossen, J., Ugent, D., Karathanasis, A., Vásquez, V., & Netherly, P. J.
(2010). Early Holocene coca chewing in northern Peru. Antiquity, 84(326), 939–953.
Dourojeanni, M. (1992). Environmental impact of coca cultivation and cocaine produc-
tion in the Amazon region of Peru. Bulletin on Narcotics, 44(2), 37–53.
Erickson, C. L. (1999). Neo-environmental determinism and agrarian ‘collapse’in Andean
prehistory. Antiquity, 73(281), 634–642.
Escamilla, E. Z., & Habert, G. (2014). Environmental impacts of bamboo-based con-
struction materials representing global production diversity. Journal of Cleaner
Production, 69, 117–127.
Etter, A., McAlpine, C., Wilson, K., Phinn, S., & Possingham, H. (2006). Regional patterns
of agricultural land use and deforestation in Colombia. Agriculture, Ecosystems &
Environment, 114(2), 369–386.
Etter, A., & van Wyngaarden, W. (2000). Patterns of landscape transformation in
Colombia, with emphasis in the Andean region. AMBIO: A Journal of the Human
Environment, 29(7), 432–439.
Evans, J. S., Murphy, M. A., Holden, Z. A., & Cushman, S. A. (2011). Modeling species
distribution and change using random forest. Predictive species and habitat modeling in
landscape ecology. Springer.
FAO-Unesco (1990). Soil map of the world (Revised legend. Reprinted with corrections.
World Soil Resources Report (1990)).
Fick, S. E., & Hijmans, R. J. (2017). WorldClim 2: New 1-km spatial resolution climate
surfaces for global land areas. International Journal of Climatology, 37, 4302–4315.
Finley-Brook, M. (2007). Indigenous land tenure insecurity fosters illegal logging in
Nicaragua. International Forestry Review, 9(4), 850–864.
Francesconi, W., Uribe, N., Valencia, J., & Quintero, M. (2018). Modeling for manage-
ment: A case study of the cañete watershed, Peru. In D. Rivera, A. Godoy-Faúndez, &
M. Lillo-Saavedra (Eds.). Andean Hydrology (pp. 84–101). Boca Ratón, FL, USA: CRC
Press.
Geist, H. J., & Lambin, E. F. (2002). Proximate causes and underlying driving forces of
tropical deforestation: Tropical forests are disappearing as the result of many pres-
sures, both local and regional, acting in various combinations in different geo-
graphical locations. BioScience, 52(2), 143–150.
Grau, H., Gasparri, N., & Aide, T. (2005). Agriculture expansion and deforestation in
seasonally dry forests of north-west Argentina. Environmental Conservation, 32(2),
140.
Grau, R., Kuemmerle, T., & Macchi, L. (2013). Beyond ‘land sparing versus land sharing’:
Environmental heterogeneity, globalization and the balance between agricultural
production and nature conservation. Current Opinion in Environmental Sustainability,
5(5), 477–483.
Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S., Tyukavina, A.,
et al. (2013). High-resolution global maps of 21st-century forest cover change.
Science, 342(6160), 850–853.
Harden, C. P. (2001). Soil erosion and sustainable mountain development: Experiments,
observations, and recommendations from the Ecuadorian Andes. Mountain Research
and Development, 21(1), 77–83.
IBC. Base de Datos del Sistema de Información sobre Comunidades Nativas de la Amazonía
Peruana - SICNA. (2016). http://191.98.188.187/ibcmap retrieved 28-08-2017.
INEI (2014). Compendio estadístico Perú 2014. Lima, Peru: Instituto Nacional de Estadística
e Informática.
INEI (2015). Estimaciones y Proyecciones de Población Total por Sexo de las Principales
Ciudades. Lima: Peru, Instituto Nacional de Estadística e Informática.
Joppa, L. N., & Pfaff, A. (2009). High and far: Biases in the location of protected areas.
PLoS One, 4(12), e8273.
Jusys, T. (2016). Fundamental causes and spatial heterogeneity of deforestation in Legal
Amazon. Applied Geography, 75, 188–199.
Kaimowitz, D., & Angelsen, A. (1998). Economic models of tropical deforestation: A review.
Bogor, Indonesia: Center for International Forestry Research.
Kaimowitz, D., Mendez, P., Puntodewo, A., & Vanclay, J. K. (2002). Spatial regression
analysis of deforestation in Santa Cruz, Bolivia. Land Use and Deforestation in the
Amazon, University of Florida Press.
Kometter, R., Martinez, M., Blundell, A., Gullison, R., Steininger, M., & Rice, R. (2004).
Impacts of unsustainable mahogany logging in Bolivia and Peru. Ecology and Society,
9(1), 12.
Lambin, E. F., Turner, B. L., Geist, H. J., Agbola, S. B., Angelsen, A., Bruce, J. W., et al.
(2001). The causes of land-use and land-cover change: Moving beyond the myths.
Global Environmental Change, 11(4), 261–269.
Laurance, W. F., Albernaz, A. K., Schroth, G., Fearnside, P. M., Bergen, S., Venticinque, E.
M., et al. (2002). Predictors of deforestation in the Brazilian Amazon. Journal of
Biogeography, 29(5–6), 737–748.
Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R News,
2(3), 18–22.
Lung, T., Lübker, T., Ngochoch, J. K., & Schaab, G. (2013). Human population distribu-
tion modelling at regional level using very high resolution satellite imagery. Applied
Geography, 41, 36–45.
Malhi, Y., Roberts, J. T., Betts, R. A., Killeen, T. J., Li, W., & Nobre, C. A. (2008). Climate
change, deforestation, and the fate of the Amazon. Science, 319(5860), 169–172.
Middendorp, R. S., Pérez, A. J., Molina, A., & Lambin, E. F. (2016). The potential to
restore native woody plant richness and composition in a reforesting landscape: A
modeling approach in the Ecuadorian Andes. Landscape Ecology, 31(7), 1581–1599.
109
Applied Geography 91 (2018) 99–110
MINAGRI (2016). Guía Metodológica para la Zonificación Forestal. Lima, Peru: Ministerio
de Agricultura y Riego.
MINAM (2015). Mapa Nacional de Cobertura Vegetal - memoria descriptiva. Lima, Peru:
Ministerio del Ambiente.
Miteva, D. A., Pattanayak, S. K., & Ferraro, P. J. (2012). Evaluation of biodiversity policy
instruments: What works and what doesn't? Oxford Review of Economic Policy, 28(1),
69–92.
Myers, N., Mittermeier, R. A., Mittermeier, C. G., Da Fonseca, G. A., & Kent, J. (2000).
Biodiversity hotspots for conservation priorities. Nature, 403(6772), 853–858.
Nanni, A. S., Gasparri, N. I., & Grau, H. R. (2015). Redistribution of forest biomass in an
heterogeneous environment of subtropical Andes undergoing agriculture adjustment.
Applied Geography, 62, 107–114.
NASA METI (2011). Advanced Spaceborne thermal emission and reflection radiometer
(ASTER) global digital elevation model version 2 (GDEM V2) ." NASA EOSDIS land
processes DAAC, USGS Earth resources observation and science (EROS) center. Sioux
Falls, South Dakota: Japan Space Systems, an implementing agency for ASTER
Science Project and ASTER GDS Project.
Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., & Pereira, J. M. (2012).
Modeling spatial patterns of fire occurrence in Mediterranean Europe using Multiple
regression and random forest. Forest Ecology and Management, 275, 117–129.
ONERN (1981). Mapa de Suelos del Perú. Lima, Peru: Oficina Nacional de Evaluación de
Recursos Naturales.
Pan, Y., Birdsey, R. A., Fang, J., Houghton, R., Kauppi, P. E., Kurz, W. A., et al. (2011). A
large and persistent carbon sink in the world's forests. Science, 333(6045), 988–993.
Piu, H. C., & Menton, M. (2014). The context of REDD+ in Peru: Drivers, agents and in-
stitutions. Bogor, Indonesia: Center for International Forestry Research.
Pope, I., Bowen, D., Harbor, J., Shao, G., Zanotti, L., & Burniske, G. (2015). Deforestation
of montane cloud forest in the central highlands of Guatemala: Contributing factors
and implications for sustainability in Q’eqchi’communities. The International Journal
of Sustainable Development and World Ecology, 22(3), 201–212.
Porro, R., Lopez-Feldman, A., & Vela-Alvarado, J. W. (2015). Forest use and agriculture in
Ucayali, Peru: Livelihood strategies, poverty and wealth in an Amazon frontier. Forest
Policy and Economics, 51, 47–56.
Pourtaghi, Z. S., Pourghasemi, H. R., & Rossi, M. (2015). Forest fire susceptibility map-
ping in the Minudasht forests, Golestan province, Iran. Environmental Earth Sciences,
73(4), 1515–1533.
Poveda, G., Álvarez, D. M., & Rueda, Ó. A. (2011). Hydro-climatic variability over the
Andes of Colombia associated with ENSO: A review of climatic processes and their
impact on one of the Earth's most important biodiversity hotspots. Climate Dynamics,
36(11–12), 2233–2249.
R Development Core Team (2016). R: A language and environment for statistical computing.
Vienna, Austria: R Foundation for Statistical Computing. www.R-project.org.
Reymondin, L., Jarvis, A., Pérez-Uribe, A., Touval, J., Argote, K., Rebetez, J., et al. (2012).
A methodology for near real-time monitoring of habitat change at continental scales
using MODIS-NDVI and TRMM. Remote Sensing of Environment (submitted for pub-
lication).
Ribeiro, M. C., Martensen, A. C., Metzger, J. P., Tabarelli, M., Scarano, F., & Fortin, M.-J.
(2011). The Brazilian Atlantic forest: A shrinking biodiversity hotspot. Biodiversity hot-
spots, Springer.
Robiglio, V., Armas, A. D., Silva Aguad, C., & White, D. (2014). Beyond REDD+ readi-
ness: Land-use governance to reduce deforestation in Peru. Climate Policy, 14(6),
734–747.
Robinson, B. E., Holland, M. B., & Naughton-Treves, L. (2014). Does secure land tenure
save forests? A meta-analysis of the relationship between land tenure and tropical
deforestation. Global Environmental Change, 29, 281–293.
Rodrigues, A. S., Akcakaya, H. R., Andelman, S. J., Bakarr, M. I., Boitani, L., Brooks, T. M.,
et al. (2004). Global gap analysis: Priority regions for expanding the global protected-
area network. AIBS Bulletin, 54(12), 1092–1100.
Rodríguez, L. O., & Young, K. R. (2000). Biological diversity of Peru: Determining priority
areas for conservation. AMBIO: A Journal of the Human Environment, 29(6), 329–337.
Rolett, B., & Diamond, J. (2004). Environmental predictors of pre-European deforestation
on Pacific islands. Nature, 431(7007), 443–446.
Rudel, T. K. (2007). Changing agents of deforestation: From state-initiated to enterprise
driven processes, 1970–2000. Land Use Policy, 24(1), 35–41.
Rudel, T. K., Defries, R., Asner, G. P., & Laurance, W. F. (2009). Changing drivers of
deforestation and new opportunities for conservation. Conservation Biology, 23(6),
1396–1405.
Rudel, T., & Roper, J. (1996). Regional patterns and historical trends in tropical defor-
estation, 1976-1990: A qualitative comparative analysis. Ambio. Stockholm, 25(3),
160–166.
Salonen, M., Toivonen, T., Cohalan, J.-M., & Coomes, O. T. (2012). Critical distances:
Comparing measures of spatial accessibility in the riverine landscapes of Peruvian
Amazonia. Applied Geography, 32(2), 501–513.
Sanchez-Cuervo, A. M., & Aide, T. M. (2013a). Identifying hotspots of deforestation and
reforestation in Colombia (2001–2010): Implications for protected areas. Ecosphere,
4(11), 1–21.
Sánchez-Cuervo, A. M., & Aide, T. M. (2013b). Consequences of the armed conflict, forced
human displacement, and land abandonment on forest cover change in Colombia: A
multi-scaled analysis. Ecosystems, 16(6), 1052–1070.
Sloan, S., & Sayer, J. A. (2015). Forest Resources Assessment of 2015 shows positive
global trends but forest loss and degradation persist in poor tropical countries. Forest
Ecology and Management, 352, 134–145.
Small, C., & Cohen, J. (2004). Continental physiography, climate, and the global dis-
tribution of human population. Current Anthropology, 45(2), 269–277.
Soares-Filho, B., Nepstad, D., Curran, L., Voll, E., Cerqueira, G., Garcia, R., et al. (2013).
LBA-ECO LC-14 modeled deforestation scenarios, Amazon basin: 2002-2050. Data set.
V. Bax, W. Francesconi
Available on-lineOak Ridge, Tennessee, U.S.A: Oak Ridge National Laboratory
Distributed Active Archive Centerhttp://dx.doi.org/10.3334/ORNLDAAC/
1153http://daac.ornl.gov.
Svetnik, V., Liaw, A., Tong, C., Culberson, J. C., Sheridan, R. P., & Feuston, B. P. (2003).
Random forest: A classification and regression tool for compound classification and
QSAR modeling. Journal of Chemical Information and Computer Sciences, 43(6),
1947–1958.
Swenson, J. J., Carter, C. E., Domec, J.-C., & Delgado, C. I. (2011). Gold mining in the
Peruvian Amazon: Global prices, deforestation, and mercury imports. PLoS One, 6(4),
e18875.
Tulet, J.-C. (2010). Peru as a new major actor in Latin American coffee production. Latin
American Perspectives, 37(2), 133–141.
Turner, B. L., & Butzer, K. W. (1992). The Columbian encounter and land-use change.
Environment: Science and Policy for Sustainable Development, 34(8), 16–44.
110
Applied Geography 91 (2018) 99–110
UNODC (2016). Perú Monitoreo de Cultivos de Coca 2015. Lima, Peru: United Nations
Office on Drugs and Crime.
Walker, R., Arima, E., Messina, J., Soares-Filho, B., Perz, S., Vergara, D., et al. (2013).
Modeling spatial decisions with graph theory: Logging roads and forest fragmentation
in the Brazilian Amazon. Ecological Applications, 23(1), 239–254.
White, S. (1978). Cedar and mahogany logging in eastern Peru. Geographical Review,
68(4), 394–416.
Young, K. R. (1996). Threats to biological diversity caused by coca/cocaine deforestation
in Peru. Environmental Conservation, 23(01), 7–15.
Young, K. R., & León, B. (2000). Biodiversity conservation in Peru's eastern montane
forests. Mountain Research and Development, 20(3), 208–211.
Zuluaga, G. J. C., & Rodewald, A. D. (2015). Response of mixed-species flocks to habitat
alteration and deforestation in the Andes. Biological Conservation, 188, 72–81.