Ac

AFa

b

c

d

Le

a

ARA

KELSLDA

1

tacflwOoaueu

II

0h

International Journal of Applied Earth Observation and Geoinformation 28 (2014) 60–69

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

jo ur nal home p age: www.elsev ier .com/ locate / jag

nthropogenic pressure in East Africa—Monitoring 20 years of landover changes by means of medium resolution satellite data

ndreas B. Brinka,∗, Catherine Bodartb, Lukas Brodskyc, Pierre Defourneyd, Celine Ernstd,rancois Donneye, Andrea Lupie, Katerina Tuckovac

Joint Research Centre of the European Commission, Institute for Environment and Sustainability, Ispra, ItalyFood and Agriculture Organization of the United Nations, Rome, ItalyGISAT s.r.o. Praha, Czech RepublicEarth and Life Institute – Environmental Sciences, Faculté d’Ingénierie biologique, agronomique et environnementale – Université catholique de Louvain,ouvain-la-Neuve, BelgiumReggiani SpA, Joint Research Centre of the European Commission, Institute for Environment and Sustainability, Ispra, Italy

r t i c l e i n f o

rticle history:eceived 8 May 2013ccepted 13 November 2013

eywords:ast Africa IGAD regionand cover changeamplingandsat

a b s t r a c t

The East Africa IGAD (Intergovernmental Authority on Development in Eastern Africa) region with itsgreat variety of ecological regions experienced major changes during the last decades. This study assessesand quantifies the land cover dynamics in the region by applying a systematic sampling of mediumresolution Landsat and DMC Deimos imagery. 445 samples covering about 3% of the study area takenas a box of 20 km × 20 km around each 1 degree latitude and longitude intersects are processed andanalyzed. Statistical estimates of land cover change are produced by means of an automatic object-basedclassification in seven broad classes for the years 1990–2000 and 2000–2010. Figures of change for theEast Africa IGAD region are presented and land cover change processes such as loss of natural vegetation

MC Deimosnthropogenic impact

and increase of agriculture areas are analyzed. Results highlight the geographical distribution of landcover dynamics and show a 28% increase in agriculture area over the analyzed 20-year time frame. Theyearly agriculture area increase rate is around 1.4% for both assessed decades, however a strong increasein yearly deforestation rate – from 0.2% in the first period to 0.4% in the second period – has been observed.These figures are discussed within the context of the drivers of changes and the resulting impact to the

natural ecosystem.

. Introduction

East Africa is often described as the cradle of humanity. Forhousands of years people have been living in the area interactingnd transforming the natural environment. Pastoralism, shiftingultivation, permanent or semi-permanent agriculture and agro-orestry have altered the environment to a point that the presentandscape is the product of both natural variation in vegetation as

ell as human-induced changes (Bongers and Tennigkeit, 2010).ver the last decades anthropogenic impact and the competitionver land has become an issue of major concern and even conflictmong the rural and pastoral population in Africa. This is partic-

larly true for the East Africa IGAD region where in addition toxtensive agriculture expansion driven by a strong increasing pop-lation, recurring cycles of drought and famine are threatening

∗ Corresponding author at: Joint Research Centre of the European Commission,nstitute for Environment and Sustainability, Via E. Fermi 2749, I-21027 Ispra (VA),taly. Tel.: +39 0332785567; fax: +39 0332789960.

E-mail address: andreas.brink@jrc.ec.europa.eu (A.B. Brink).

303-2434/$ – see front matter © 2013 Elsevier B.V. All rights reserved.ttp://dx.doi.org/10.1016/j.jag.2013.11.006

© 2013 Elsevier B.V. All rights reserved.

agriculture production and hence lives of both people and live-stock (Meier et al., 2007; Molvaer, 1991). Hence, since the mid1990s the Intergovernmental Authority on Development in East-ern Africa (IGAD) aims to assist and complement the efforts of themember states in the fields of food security and environmentalprotection, promotion and maintenance of peace and security andhumanitarian affairs, and economic cooperation and integration.

Since the early 1970s Earth Observing satellites are monitoringour planet’s land surfaces at different time and spatial resolu-tions. From the beginning the African continent has been of majorinterest for this technology supporting the repeated assessmentof large and often inaccessible areas. Most studies applied eitherlow-resolution satellite data with high repeat cycles for continen-tal assessments of changes in vegetation status and phenologyproviding limited information on quantitative land cover changefigures (Bégué et al., 2011; Stroppiana et al., 2009; Diouf andLambin, 2001; Lambin and Ehrlich, 1997) or higher resolution

images with greater spatial accuracy but low repeat cycle appliedmainly for national or local case studies (Cabral et al., 2011; Tappanet al., 2004; Pellikka et al., 2009; Jansen et al., 2008; Serneelsand Lambin, 2001). An exception over the African continent is

arth O

ratw2bctCfisiHtdG1iU(so1mpgeBp

dtOueatar

fLt1pvpppd2oMaTt

2

2

Ao

A.B. Brink et al. / International Journal of Applied E

epresented by the humid tropics and in particular the Congo Basinrea which has been monitored accurately from low to high spa-ial and temporal resolution with both sample based methods asell as spatially continuous mapping approaches (Achard et al.,

002; Potapov et al., 2012; Mayaux et al., 2013). The reason for thiseing the greater interest of humid tropical forests in the globallimate processes and the climate change discussion and negotia-ions under the United Nations Framework Convention on Climatehange (UNFCCC). A recent study of Bodart et al. (2013) provides

or the first time regional estimates of forest cover and its changesn the dry African ecoregions between 1990 and 2000, using aystematic sample of medium-resolution Landsat type of satellitemagery which was processed consistently across the continent.owever also this study is mainly focusing on forests and deforesta-

ion rates. In addition, its geographical coverage is limited by thery forest areas within the Sudanian, Guinea-Congolia/Sudanian,uinea-Congolia/Zambezian and Zambezian ecoregions (White,983). These are covering only marginally the IGAD region, includ-

ng the southern part of Sudan and only minor areas in Ethiopia andganda. Finally, Brink and Eva (2011) demonstrate the capability

but also limitations) of using a regular grid of Landsat imageryamples to highlight areas of increase in agriculture and reductionf wood- and shrublands in the Somalia-Masai ecoregion (White,983) within the Greater Horn of Africa. Although different in theethodology – Brink and Eva (2011) is based on visual image inter-

retation and a higher sampling rate within the regular samplingrid – and slightly dissimilar in the study area – the Somalia-Masaicoregion is covering only partially the IGAD region – the study ofrink and Eva (2011) shows comparable results for the 1990–2000eriod over the matching areas of our study.

Biome-scale assessments are important in order to characterizeynamics and the impacts of these variations on similar vegetationypes – such as humid or dry forests – in distinct ecological zones.n the other hand, regional observations and evaluations, in partic-lar economic regions such as IGAD, are necessary to support theconomic region’s mission. Coordinated and harmonized policiest the regional scale in the fields of food security, habitat protec-ion and natural resource management require also standardizednd consistent monitoring and assessment of the environment ategional level.

The objective of this study is to provide land cover statisticsor the year 2000 based on a systematic sampling of high accurateandsat and DMC Deimos images over the entire IGAD region ando assess changes in a consistent manner over the time periods of990–2000 and 2000–2010 with a particular focus of the anthro-ogenic impact – defined here as agriculture expansion – on naturalegetation. The methodology is based upon the global TREES-3roject implemented by the Joint Research Centre of the Euro-ean Commission to monitor tropical forest cover changes for theeriods 1990–2000–2010 based on multi-date Landsat sample sitesistributed systematically over the global tropics (Achard et al.,009). The research of our study was conducted in the frameworkf the Seasonal and Annual Change Monitoring (SATChMo) Coreapping Service within the EU-FP7 funded Geoland-2 project. The

im and novelty of this particular work was to further develop theREES-3 approach- in order to assess land cover dynamics also inhe agricultural domain.

. Materials and methods

.1. Study area

Our study focuses on the countries within the Greater Horn offrica comprised in the IGAD region (Fig. 1), including the countriesf Djibouti, Eritrea, Ethiopia, Kenya, Somalia, Sudan and Uganda

bservation and Geoinformation 28 (2014) 60–69 61

(our study was done prior the official separation of Sudan into SouthSudan and Sudan, therefore in the text we refer only to the formerborders of Sudan).

The IGAD region covers a wide range of climatic and ecolog-ical regions, resulting in a wide variety of land cover types andland cover change dynamics. The geomorphology varies betweencoastal areas to flat and gentle sloping land to high mountain areassuch as Mount Kenya, Africa’s second highest peak. The ecoregionsmap of White (1983) provides a broad overview of the key vegeta-tion zones and their relation to climate and rainfall in the region.The natural land cover distribution is roughly mirrored in the cli-matic patters and the different vegetation zones of the region. TheIGAD region is characterized by nine ecoregions, although two ofthem – the Guineo-Congolian and the Giunea-Congolia/Sudanian –are of minor relevance due to its limited area in the IGAD region.The Somalia-Masai ecoregion is the most predominant one. It cov-ers about 33% of the entire IGAD region and is characterized byarid to semi-arid type of climate. The vegetation structure is pre-dominantly open to closed deciduous shrubs and woody vegetationincluding tree and shrub savanna, which become semi-evergreenand evergreen bushland and thicket on the lower slopes of themountain areas. The next ecoregion in importance of percentagearea is the Sudanian ecoregion which covers 23% of the total area.Most of the Sudanian region lies below 1000 m and is distinguishedby semi-arid climate in the North to equatorial savanna type of cli-mate in the southern part. The natural and semi-natural vegetationis characterized by woodland and in rare cases by dry forest. Butin most areas the natural vegetation has been modified by humanactivities. The practice of shifting cultivation, where woodland isin various stages of regrowth following a period of cultivation, istypical for this region (White, 1983). The Sahara and Sahel ecore-gions which describe the northern part of the IGAD region exhibitboth about 14% coverage respectively. The climate is arid and deserttype, with unreliable rainfall mostly below 500 mm per year, mak-ing any vegetation grow a challenge. Wooded grassland in theSouth and semi-desert grassland in the North are predominant withsome wood- and shrubland restricted to rocky outcrops (White,1983). About 10% of the IGAD region is distinguished by the uniqueAfromontane ecoregion. This zone is an archipelago-like centre ofendemism (White, 1983) which is common in tropical Africa onelevations above 2000 m. A considerable part of Ethiopia is hometo this ecoregion extending North into Eritrea and even Sudan.Furthermore, the highlands and mountain areas of Kenya are rep-resented by the Afromontane ecoregion and also to some extendthe most southern and western end of Uganda. In general meanannual rainfall is above 1000 mm in the forest zone, diminish-ing with increasing altitude outside the forest belt (White, 1983).Below the forest belt a natural transition zone towards the border-ing ecoregions is present. The remaining ecoregions have only alimited presence in the IGAD region. They include the Lake Victo-ria ecoregion with locally high and well distributed rainfall which issufficient to support rain forest, but in the majority semi-evergreenforest and wood- and shrubland represent the climax. Then theZanzibar-Inhambane ecoregion which is located along the Kenyancoast and stretches to some extend into southern Somalia. Herethe typical climax vegetation is represented by forest and alongthe coastline Mangrove forests. However, these has been exten-sively cleared to make place for tree plantations and aquaculture.Finally, the Guinea-Congolia/Sudania ecoregion is characterized byan equatorial savanna type of climate comparable to the southernend of the Sudanian ecoregion.

2.2. Method

A spatially continuous – often referred to as wall-to-wall map-ping – land cover change assessment may be the ultimate solution

62 A.B. Brink et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 60–69

ng fram

fbhactvane2

2

tDasdaJaAowcatlgT

registration, radiometric calibration and conversion to top ofatmosphere reflectance, cloud masking, haze correction and nor-malization or relative calibration. These steps are described in detail

Table 1Land cover classes and definition.

Class name Definition

Dense tree cover ≥70% tree cover portion in segmentOpen tree cover and tree cover mosaic 30–70% tree cover portionOther wooded land ≥70% shrubs, forest regrowthNatural grassland ≥70% grassland

Fig. 1. Study area consisting of the IGAD region and sampli

or analysing dynamics at the full spatial extent of a studied area,ut this method may not be realistic for large area mapping withigh-medium resolution images due to constraints on resources fornalysis and also because of technical difficulties in processing andlassifying complex type of landscapes. It has been demonstratedhat statistical sampling of high to medium resolution imagery pro-ides an alternative cost-effective and accurate approach to deriverea estimates of land cover and change at pan-tropical to conti-ental and regional level (Richards et al., 2000; FAO, 2001; Achardt al., 2002; Hansen et al., 2008; Gibbs et al., 2010; Brink and Eva,009, 2011; Bodart et al., 2013; Mayaux et al., 2013).

.2.1. Sampling strategy and classification schemeThe sampling design selected for this study consists of a sys-

ematic sampling of 20 km × 20 km subsets of Landsat TM/ETM andMC-Deimos imagery for the reference years 1990–2000–2010 tossess and quantify land cover dynamics in the IGAD region. Theampling framework is defined as a rectilinear grid based on integeregrees of geographical latitude and longitude as shown in Fig. 1nd is derived from the global TREES-3 project developed by theoint Research Centre of the European Commission in collaborationnd support to the Remote Sensing Survey of FAO’s Forest Resourcessessment 2010 (FAO, 2010) to monitor tree cover and its changesver the tropics (Achard et al., 2009). A total of 445 sample siteshere processed and analyzed for the 1990–2000–2010 period,

overing about 3% of the study area. The land cover classes detectedt each sample site and for all years were the following: dense

ree cover, open tree cover and tree cover mosaic, other woodedand, natural grassland, agriculture (non irrigated), agriculture irri-ated, bare and artificial areas and water (see class definition inable 1).

ework showing the 445 samples used for the assessment.

2.2.2. Satellite imagery and data selectionLandsat data were downloaded from the United States Geo-

logical Survey’s (USGS) National Centre for Earth ResourcesObservation and Science (http://glovis.usgs.gov) at full spatial andspectral resolution (30 m), while DMC-Deimos images at 22 m spa-tial resolution were provided by the European Space Agency (ESA)through the FP7 project Geoland2. Spatial resolution of both imagetypes were kept as provided. The integration of multi-sensors intothe processing chain is described in detail in Desclée et al. (2013).Image subsets of 20 km × 20 km were extracted and successivelyvisually screened in order to identify the best imagery in terms ofcloud cover and seasonal/radiometric characteristics (Beuchle et al.,2011).

2.2.3. Data pre-processingPre-processing includes assessment and correction of spatial

Agriculture (non irrigated) ≥70% agricultureAgriculture (irrigated) ≥70% agriculture (irrigated)bare and artificial ≥70% bare and artificialInland water Permanent water

A.B. Brink et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 60–69 63

Fig. 2. Flow-chart describing the Landsat multi-date pre-processing chain (modified from Bodart et al., 2011). Clouds and cloud shadows masking, haze correction andr h (1)

d age to

ib

LbtrtbbEiit

tc(trr

vauaocbtatitHp

adiometric normalization have been, applied on images previously identified witense forest was present in the sample site, a relative normalization of the 1990 im

n Bodart et al. (2011) and outlined succinctly in Fig. 2 and the textelow.

The majority of the Landsat data selected are part of the Globaland Survey (GLS) data set (Gutman et al., 2008). These data haveeen geometrically corrected and geo-referenced by the USGS tohe UTM (Universal Transverse Mercator) projection with a spatialesolution of 30 m and a remaining root mean square error of lesshan 50 m. The DMC Deimos imagery was intentionally designed toe comparable to Landsat imagery, in fact the reference system haseen processed by the image provider to match the GLS dataset.ffective precision of the spatial registration between multi-datemagery was controlled by visually assessing the relative image tomage geo-location. In most cases geo-location errors correspondedo a linear shift of 1 or 2 pixels.

In order to reduce differences arising from changing illumina-ion conditions and instrument errors and differences a radiometricalibration was performed by first converting raw digital numbersDN) into at-sensor spectral radiance for each band. Subsequently,he at-sensor radiance was converted into top-of-atmosphereeflectance. Finally, reflectance values were scaled into the byteange of 0–255 (Chander et al., 2009).

Remaining clouds and cloud shadows identified during theisual screening phase were masked by detecting all potential cloudnd cloud shadow pixels over the radiometric calibrated imagesing the automatic spectral rule-based preliminary mappingpproach proposed by Baraldi et al. (2006). Due to a high amountf commission and omission errors in this phase, a subsequent steponsisted of a sequential application of a post-processing algorithmased on morphological and topological methods designed to cleanhe results using the size, width and shape of clouds and shadowsnd the sun’s azimuth (Bodart et al., 2011). A part from clouds,ropical regions are often affected by haze which contaminate the

mages by semi-transparent clouds and aerosols. These influencehe spectral values in the images and hamper visual interpretation.aze affected image subsets were corrected by using the methodroposed by Lavreau (1991).

clouds, (2) haze or (3) evergreen dense forest. When no or not enough evergreen the 2000 image to the 2010 image has been applied.

Finally, radiometric normalization was applied to reduce thenoise introduced in the imagery by the variation in atmosphericconditions. For our purpose a relative normalization of the multi-temporal images to a common relative scale was applied to ensurespectral homogeneity of the data (Bodart et al., 2011).

Subsequently the processing method combined multi-dateimage segmentation and automatic classification steps followed byvisual control by image interpretation experts (Rasi et al., 2011) ini-tially focusing on the forest and other wooded land classes. Then,the classification model processed generic land cover maps for eachsite for the baseline year 2000 including also the remaining classesof natural grassland, agriculture, irrigated agriculture, bare andartificial and inland water. At this stage, an interactive approachwas applied, due to the highly diverse landscapes present in theregion. In the e-Cognition software a dedicated Architect GraphicalUser Interface (GUI) was designed for our specific purposes. The e-Cognition GUI allows fully automatic as well as interactive imageclassification, while users can easily configure, calibrate and exe-cute image analysis workflows (Benz et al., 2004). The approach isbased on a preliminary benchmarked classification model adoptinga two stage procedure: automated OBIA classification (Object-oriented Image Analysis) with the possibility to fine tune each classspecific variables and post-classification editing. Preliminary strat-ification of arid and non arid regions allows transferring rules froma sample to another sample. The need of manual post-editing isestimated to be about 20%, which is site specific. Main uncertain-ties in the classification remain in the low density shrubs and inagriculture, especially the irrigated cropland.

The land cover change (LCC) assessment is based on the LCC mapproduct processed according to the land accounting system (LAS)framework as explained in Fig. 3.

The LAS is a thematic and spatial simplification of the change

database. Land cover changes are aggregated into the thematicflows or indicators on a regular grid. This statistical assimilation dif-fers from cartographic generalization in the sense that the formerpreserves the original values while the latter incorporates the small

64 A.B. Brink et al. / International Journal of Applied Earth O

Fig. 3. Change analysis flow-chart describing the following steps: (1) land cover(LC) 2000 and 1990–2010 land cover change (LCC) database, (2) LC flow classifica-tion nomenclature, (3) change matrix derived from the image segmentation outputssr

otso–

twg

di

2

oaaavtaf

tored in shape file format, (4) database of flows, (5) pre-defined land accountingystem (LAS) analysis according to the specified indicators, (6) interactive tool foresults visualization and analysis by the end user.

bjects into the larger ones. The aggregation of land cover changeso the reference grid allows the integration with different dataources and types using the same grid. In this study we defined onlyne regular grid corresponding to the Area Frame Sampling Scheme

for each sample the following indicators were calculated:

Croplands extension – all changes TO agriculture or irrigated agri-cultureForest cover loss – forest/tree cover or mosaic forest TO any otherclassAnthropogenic impact – forest/tree cover, mosaic forest, otherwooded land and grassland TO agriculture or irrigated agricultureLoss of natural vegetation – forest/tree cover, mosaic forest, otherwooded land and grassland TO agriculture, irrigated agriculture,bare and artificial areas, water

Finally, according to the thematic indicators explained above,he results of the LAS are visualized on a dedicated interactiveeb-tool in form of maps, tables and charts (http://satchmo-africa.

isat.cz/).Since most significant, main focus of the following results and

iscussion section will concentrate on the anthropogenic impactndicator and on the forest cover loss index.

.2.4. Quality control and validationDeveloping an automated processing and classification scheme

ver vast areas with different landscapes and ecoregions, climatesnd seasons and human interactions is challenging. In order tossure the same classification quality standard for all the regionnd vegetation types we therefore followed a systematic three-step

isual quality control approach after the automated classifica-ion. This consists of a first visual screening of the segmentationnd classification results by internal experts to detect and corrector obvious errors, followed by an accurate quality control and

bservation and Geoinformation 28 (2014) 60–69

correction process by national and local experts from the vari-ous countries of the IGAD region. A last harmonization process byinternal experts was finally necessary in order to avoid differencesbetween countries.

In addition, the entire TREES-3 dataset over sub-Saharan Africa– on which our dataset is based upon – has undergone an inde-pendent validation assessment which has demonstrated an overallagreement of 87% for all classes and 94% for the tree cover classes.A detailed explanation of the validation exercise is described inBodart et al. (2013).

3. Results

3.1. Land cover year 2000

According to our results in the year 2000 the IGAD region wascovered by slightly over 331 million hectares (Mha) of natural veg-etation, which represents 62% of the total study area. The dominantland cover type is distinguished by the wood and shrub cover classwhich is distributed over the whole region – despite the mostnorthern area of Sudan – and accounts for more than half of theentire natural vegetation, covering about 186 Mha of the entirearea. Natural grasslands follow with nearly 117 Mha and also theseare well distributed in all countries. The forest zone on the otherhand is more localized, with dense tree cover restricted mainly tothe Afromontane areas in Ethiopia, Kenya and Uganda and opentree cover and tree cover mosaic more widely distributed, even iflimited to areas with sufficient rainfall. In total 28 Mha of forestsare covering the region, equalling to slightly over 5% of the totalarea. However from these less than 6 Mha are characterized byclosed tree cover and the remaining 22 Mha by open tree coverand tree cover mosaic. This means that according to our resultsonly 1% of the entire IGAD region is covered by closed forests.Bare areas are also predominant, covering about 112 Mha or 21%of the total area, but these are mostly marginalized to the northernpart of Sudan. Anthropogenic impact, distinguished by agricultureexpansion, has transformed the landscape and removed the nat-ural vegetation in vast areas. Almost 80 Mha of both rainfed andirrigated agriculture were covering the region in the year 2000, rep-resenting nearly 15% of all described land cover classes. Extensivecroplands are identified in the central and northern part of Ethiopia,in a central belt over Sudan, the fertile areas of South-West Kenyaand nearly the whole of Uganda. Minor agriculture areas can befound in the remaining countries of Somalia, Djibouti and Eritrea.The relative few irrigated croplands detectable by our method areconcentrated mostly along the Nile river in central-eastern Sudan,although minor areas can be found also in the other IGAD countries(Fig. 4).

3.2. Land cover change 1990–2000–2010

From the total IGAD area nearly 28 Mha have experienceda change in land cover for the analyzed twenty year timeperiod. This means that over 5% of the total IGAD area hasundergone a permanent modification in its land cover. Nearly13 Mha have been modified for the 1990–2000 period, while anincreasing 15 Mha of land has been altered for the 2000–2010epoch.

The main focus of this study is to assess the anthropogenicimpact on natural vegetation over the analyzed 20-year time frame.In the 1990–2000 time period about 7.6 Mha of natural vegeta-

tion has been converted to agriculture over the entire IGAD region.This value is just slightly reduced for the 2000–2010 time framewhere 7.4 Mha of natural vegetation has been lost towards agri-culture. The yearly rate of anthropogenic impact amounts to 1.44%

A.B. Brink et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 60–69 65

atistics for the year 2000 over the IGAD region on the right.

fga(

aaessfi

Table 2Yearly rates of land cover change over the IGAD region.

Indicator Total loss(1000 ha)

Yearly %rate

Fig. 4. (a) Land cover distribution on the left and (b) st

or the first decade and 1.39% for the second decade. If we disag-regate this figure, we note that 0.2% for the period 1990–2000nd 0.4% for the epoch 2000–2010 is related to deforestationTable 2).

Hence, the remaining annual loss of natural vegetation of 1.22%nd 1% for the first and second time period respectively arettributed to wood and shrubland and natural grasslands. How-

ver, while the loss of other wooded land and natural grasslandshow a slight decreasing trend over time – even if remaining sub-tantial – the deforestation rate has nearly doubled between therst and second analyzed decade. This is particularly true for the

Anthropogenic impact 1990–2000 7644 1.44Anthropogenic impact 2000–2010 7393 1.39Deforestation 1990–2000 1207 0.22Deforestation 2000–2010 2096 0.39

66 A.B. Brink et al. / International Journal of Applied Earth Observation and Geoinformation 28 (2014) 60–69

Fc

cdtltc

ahattaSitdc

Fig. 6. Anthropogenic impact indicator per sample site represented by the dark grey

ig. 5. Forest cover loss proportions by sample site represented by the dark greyolour: (a) 1990–2000 and (b) 2000–2010.

ountries of Kenya and the southern part of Sudan, where an evi-ent increase in deforestation is noticeable, as shown in Fig. 5. Onhe contrary, Uganda is experiencing a decreasing trend and to aesser extend also Ethiopia. Minor deforestation and with no par-icular trend can be found in Somalia and none is recorded for theountries of Eritrea and Djibouti.

The loss of natural vegetation towards agriculture is evident inll countries of the IGAD region for both considered time periods,ighlighting the apparent anthropogenic impact over the studyrea. Fig. 6 shows that the strongest dynamics are manifested inhe countries of Ethiopia – in particular for the first decade wherehe removal of wood and shrubland and grassland is strongest –nd the central belt and partly southern Sudan. Especially overudan the strong deforestation increase rate for the second decade

s responsible for the rise of the anthropogenic impact in thisime period. The country of Uganda even though experiencing aecreasing deforestation rate between the two decades (Fig. 5), isharacterized by a clear increase of loss of natural vegetation in the

colour: (a) 1990–2000 and (b) 2000–2010.

2000–2010 period. This can therefore be attributed to the clearingof natural wood and shrubland and grasslands. On the contrary,the expanding anthropogenic impact for the 2000–2010 epoch inKenya is distinguished by an increasing deforestation rate concen-trated mainly in the southern and more fertile part of the country.Somalia is also showing a relatively high increase of agriculturearea per sample site in particular in the central and southern part ofthe country and around the Mogadishu area, predominantly asso-ciated with the increased clearing of wood and shrubland. Finally,Eritrea is exhibiting a relatively high anthropogenic impact rate forthe 1990–2000 period associated exclusively to the loss of otherwooded land and in particular natural grasslands in a few localizedareas. This rate of loss is diminished in the 2000–2010 time frame,however more widely distributed over the country. Djibouti, cov-

ered not even entirely by four sample sites is revealing only minorchanges.

arth O

4

4

stHa1imoc

ssmGdtiacu–ldooofmU(wcHdcvbnahRfcdp(oIvfiIdtbiTrbedm

A.B. Brink et al. / International Journal of Applied E

. Discussion

.1. Land cover year 2000 – comparison and analysis

In the year 2000 natural vegetation – such as forests, wood andhrublands and natural grasslands – was still representing 62% ofhe total land cover in the IGAD region according to our results.owever, anthropogenic impact has transformed the landscapend removed natural vegetation extensively, representing nearly5% of all described land cover classes. Bare areas are also predom-

nant, covering about 21% of the total area, but these are mostlyarginalized to the dry part of northern Sudan. The remaining 2%

f land cover is characterized by water bodies and other land coverlasses not reported within our study.

Comparing to other existing national or regional statistics for theame year is often problematic due to differences in data, methods,cale and definitions among others. However we note a nearly 100%atch of our closed tree cover figure when evaluated against theLC2000 dense forest class (Mayaux et al., 2004). In this case classefinitions are similar and the closed forest class is accurately iden-ified also by low resolution sensors such as the SPOT-Vegetationmages utilized for the GLC2000 map. Differences between the twossessments increase when comparing the open forest and treeover mosaic classes – we consider about 4% of the IGAD regionnder this class, whereas GLC2000 provides a value of nearly 3%

and diverge by over 50% in the shrub cover and other woodedand class. In addition to inconsistencies in class definitions, theifference in this case is also attributed to the limitation of low res-lution sensors in depicting the fragmented and open componentf this vegetation type showing the increased value and accuracyf our method. On the other hand when comparing the forest areaor the year 2000 of our study to the latest Forest Resource Assess-

ent undertaken by the Food and Agriculture Organization of thenited Nations (FAO, 2010) differences become significant. FAO

2010) reports about 100 Mha of forest area for the IGAD regionhereas our study assessment for this class is below 30 Mha when

ombining the tree cover and tree cover mosaic classes together.owever, the FAO approach is based on a much broader forest classefinition of 10% canopy cover – compared to our 70% for the treeover class and 30–70% for the tree cover mosaic class – and is pro-ided in form of statistics by the member countries in the best case,ut relies often on estimates and extrapolations with known weak-esses (Matthews, 2001). This results in often much higher forestrea statistics when related to remote sensing based methods asighlighted also by Bodart et al. (2013). Similarly to FAO’s Forestesource Assessment data also the FAO agriculture statistics suffer

rom a lack of consistency and completeness both in time and inoverage – relying as they do on the completeness of the nationalata sources. FAO statistics reveal that in 1997 only 29% of Africa’srimary production figures came from official country reportingNgendakumana, 2001). From the remaining figures 4% are semi-fficial and 67% of the crop production data are based on estimates.ndeed, from the seven IGAD countries only (former) Sudan has pro-ided official data according to FAO questionnaires. All remaininggures are based upon manual or FAO estimates (FAOSTAT, 2013).

t is therefore no surprise if also for the agriculture class our resultsiffer substantially from the FAO statistics. According to FAO statis-ics (FAOSTAT, 2013) only about 8% of the IGAD region is coveredy arable land and permanent crops in the year 2000. In compar-

son our agriculture area estimate for the study site is about 15%.his value is relatively close to the 18% figure provided for the IGADegion by the most recent cropland area map for Africa prepared

y Vancutsem et al. (2013). The authors harmonized and combinedxisting land cover/land use datasets to produce the most up toate and accurate cropland map for the continent. The GLC2000ap exhibit a value of 20% for the agriculture class, however it is

bservation and Geoinformation 28 (2014) 60–69 67

well known that the GLC2000 product overestimates this particularland cover class in this type of biome (Vancutsem et al., 2013).

4.2. Land cover change 1990–2010 – comparison and analysis

According to our study, 5% of the total IGAD area has undergonea permanent modification in its land cover over the analyzed 20-year time period. This figure includes both gain and loss of naturalvegetation. However, in total 15 Mha of natural vegetation has beenconverted to agriculture, the biggest portion coming from the otherwooded land class. Yet, we note that the strongest increase in lossof natural vegetation comes from the forest class. The loss of forestsnearly double from the first analyzed period to the second one, froma yearly deforestation rate of 0.2% for the 1990–2000 period to 0.4%in the 2000–2010 decade.

The 1990–2000 deforestation estimates from our study are closeto the figures provided by Bodart et al. (2013) using the samemethod and time interval for the African dry forests. In their studythe annual rate of deforestation (including both dense and opentree cover) amounts to 0.34%, however from this total value 0.37%is attributed to the Zambezian ecoregion and 0.25% to the Sudanianregion, the latter one being closer related to the IGAD environment.In contrast to our results showing an increasing deforestation trend,a recent study submitted by Mayaux et al. (2013) is reporting adecreasing deforestation tendency over the African rainforests –from 0.33% in the first epoch to 0.14% in the second time period. Thisparticular study is determining the evolution of the African humidforests for the same 1990–2000 and 2000–2010 time period andapplying again a comparable methodological approach, howeverthe forests in the IGAD region are not included in the assess-ment. Finally, comparing our deforestation estimates with the FAOstatistics from the latest Forest Resource Assessment (FAO, 2010)demonstrates again a significant higher deforestation value for theFAO report. For both the 1990–2000 and 2000–2010 periods theFAO statistics report a stable annual loss of forest cover of 0.77%over the IGAD region. However, several of the country data relies onoutdated information of various sources, dating back to the 1980sin the worst case for Somalia (FAO, 2010).

Discrepancies are evident also in the agriculture class whencomparing to FAO statistics. We estimate a yearly rate of anthro-pogenic impact of 1.44% for the first decade and 1.39% for the seconddecade. FAOSTAT reports a yearly increase rate of arable land andpermanent crops for the 1990–2000 period of 0.90% and a strongincrease for the 2000–2010 epoch with a value of 1.87% for theIGAD region. Previous remote sensing based studies conducted byGibbs et al. (2010) and Brink et al. (2012) highlight the generaltrend of agriculture expansion on the one hand and the loss of nat-ural vegetation on the other hand for the entire sub-Saharan Africaregion. However, according to Mayaux et al. (2013) and Brink et al.(2012) the most important reduction of natural vegetation associ-ated with an increase of agriculture land, is occurring in the WestAfrican forests and in the Sahel belt. For the analyzed 20-year timeframe over the IGAD region we report a total anthropogenic impactrate – and hence increase in agriculture area – of 28%. Gibbs et al.(2010) and Brink et al. (2012) both describe an increase of agricul-ture land of almost 60% for the entire sub-Saharan Africa for earliertime periods – 1980–2000 for Gibbs et al. (2010) and 1975–2000for Brink et al. (2012). For the IGAD region this would mean that acertain saturation point has been reached, since our results reveala decreasing trend in agriculture expansion over the last ten to

twenty years. In fact, even if showing a strong increase in deforesta-tion rate, our total statistics reveal a slight negative anthropogenicimpact tendency – from a total of 14.4% for the 1990–2000 periodto 13.9% for the 2000–2010 epoch.

6 arth O

4

ocidcbpefaA(sUobo2peCpciid(sfataichoelIetffltU1sah

4

hcTtopitaa

8 A.B. Brink et al. / International Journal of Applied E

.3. Drivers and impact of land cover change in the IGAD region

Geist and Lambin (2002) associate the key underlying causef land cover change in the tropics to the expansion of agri-ultural land, followed and interlinked by wood extraction andnfrastructure development. Our study confirms this hypothesis byemonstrating the dominance of agriculture expansion in the landover change process. Wood extraction in terms of commercial tim-er but even more for fuelwood and charcoal production is oftenreceding the agriculture expansion process (land clearing), how-ver in particular wood extraction is not just a byproduct of biggerorces such as logging for timber and agriculture expansion, but

specific need to satisfy energy demand. Indeed, in sub-Saharanfrica the main use of extracted wood is for energy consumption

Kebede et al., 2010; Mugo and Ong, 2006; Mwampamba, 2007),upplying e.g. 97% of household energy needs in Ethiopia, 90% inganda and about 80% in Kenya (Mugo and Ong, 2006). Our figuresf change in the non-forest vegetation domain can be explainedy this process. FAOSTAT (2013) reports a population increase ratef 2.49% for the 1990–2000 period and a 2.29% increase for the000–2010 epoch. Even if a decreasing tendency is noticeable theopulation growth rates in the IGAD region are among the high-st in the world. The urban population growth rate is above 3%.onsequently, in the next future this will increase the demand andressure on natural resources in form of charcoal to meet householdooking demands and agriculture food products to feed the grow-ng population. Yet, in the context of land dynamics and changesn land use in the IGAD region, the growing importance of distantrivers in the land change process has to be taken into accountMeyfroidt et al., 2013). A recent study by Weinzettel et al. (2013)hows that international trade accounts for 24% of the global landootprint. The IGAD region is increasingly concerned by this glob-lization of land use process. International export opportunities inhe horticulture sector are growing fast, the global biofuels demandnd the ‘land grabbing’ problematic in general are key factors driv-ng new patterns of land investments and hence changes in landover and -use. Major international land deals in the IGAD regionave been reported in Sudan and Ethiopia, where 0.46% and 1.39%f the total land suitable for rainfed crops has been allocated to for-ign investors between 2004 and 2009 respectively, with the singleargest land allocation being 150,000 ha in size (Cotula et al., 2009).n Uganda since the 1980s horticulture is one of the fastest growingxport sectors, replacing traditional cash crops such as coffee andea. In fact, Uganda is currently the second largest producer of freshruits and vegetables in sub-Saharan Africa (Ogwal et al., 2010). Theoriculture industry in Kenya on the other hand has recorded con-inuous growth rates becoming the lead exporter to the Europeannion (EU) with a market share of over 35% (KFC, 2013; Barrett et al.,999). However, horticulture is practised principally in areas withufficient water and therefore in the forest zone of both Ugandand Kenya, putting strong pressure on the forests as shown by ourigh deforestation rates in these two countries.

.4. Data and method challenges and limitations

Our study has demonstrated the capability of using medium toigh resolution Landsat and DMC Deimos images to assess landover changes at a regional scale over a twenty year time period.he processing chain was optimized to handle big datasets simul-aneously over vast areas, however the automated classificationf images covering different ecoregions with distinctive seasonalatterns and landscapes as present in the IGAD region is challeng-

ng even to the most advanced approaches. Great care was giveno the selection of images with the same or as close as possiblecquisition date over the analyzed time period. But, this was notlways achieved and therefore changes in seasonal patterns were

bservation and Geoinformation 28 (2014) 60–69

attributed erroneously to changes in land cover in some cases. Thethree step visual quality control by internal as well as externalnational experts was therefore a mandatory and intense processto truthfully correct for these seasonal effects. In addition, ourclassification approach was able to accurately depict changes inthe tree cover and wood and shrubland domain, however boththe limited temporal coverage as well as spatial resolution of theimages limited the discrimination capabilities of the method toclassify natural grasslands and agriculture. Visual interpretationand the support of very high resolution images from Google Earthwere necessary to describe these vegetation classes and its changes.New sensors such as the recently launched Landsat 8 and Sentinel 2which will be launched in 2014 will allow to overcome most of thecurrent critical issues. Large swath coverage associated with hightemporal frequency as well as spectral and spatial resolution willallow better discrimination of vegetation dynamics over time andreduce classification uncertainties caused by seasonal patterns.

5. Conclusions

The assessment of natural resources at the scale of Regional Eco-nomic Communities in Africa is a fundamental information layersupporting the analysis of regional dynamics and hence the deci-sion making process. Yet, comparable and harmonised informationbetween countries is often missing. To the best of our knowledgefor the first time, a detailed and consistent analysis of changes inland cover – with a particular focus on the anthropogenic pressure- has been conducted over the whole IGAD region over a 20-yeartime frame using a global sampling of Landsat and DMC Deimosimagery and a region-specific harmonized approach. Between 1990and 2010 our study indicates that about 15 Mha of natural vege-tation has been lost towards agriculture with a yearly conversionrate of 1.44% and 1.39% for the 1990–2000 and 2000–2010 epochrespectively. Discrepancies between our results and national statis-tics provided by FAO are apparent, demonstrating once more theneed for comprehensive and systematic assessments and continu-ous monitoring over time of land cover dynamics with particularattention to the expansion of agriculture at the expenses of nat-ural vegetation. Furthermore, in order to analyze and understandthe current and future evolution of land cover dynamics and aboveall the anthropogenic impact in the IGAD region, national andregional, but in particular the global socio-economic dimensionhas to be further investigated. The systematic sampling schemeproposed in this study combined with current medium to highresolution images, but even more with the promising upcomingimages from new sensors such as Landsat 8 and Sentinel2, providea robust tool to assess land cover changes over time. This willenable improved formulation and effective monitoring of regionaland international environmental policies. In particular, such dataprovide crucial information for the IGAD region to accurately esti-mate the expansion process of agriculture areas at the expenses ofnatural vegetation and habitats.

Acknowledgments

This research was supported by the GEOLAND-2 project (con-tract n. 218795), which is a Collaborative Project (2008–2012)funded by the European Union under the 7th FrameworkProgramme (http://www.gmes-geoland.info/). The research con-ducted for this study was done in the framework of the Seasonaland Annual Change Monitoring (SATChMo) Core Mapping Service

of the Geoland-2 project. The research team consisted of threeinstitutions, namely GISAT, the Université catholique de Louvainand the Joint Research Centre of the European Commisison. TheDMC Deimos images provided for the 2010 time period have been

arth O

fSrp

R

A

A

B

B

B

B

B

B

B

B

B

B

B

C

C

C

D

D

FFF

G

G

G

A.B. Brink et al. / International Journal of Applied E

unded by the project and were provided through the Europeanpace Agency. Finally, the authors would like to acknowledge theegional experts that took part in the quality control and validationrocess.

eferences

chard, F., Eva, H., Stibig, H.-J., Mayaux, P., Gallego, J., Richards, T., Malingreau, J.-P.,2002. Determination of deforestation rates of the world’s humid tropical forests.Science 297, 999–1002.

chard, F., Beuchle, R., Bodart, C., Brink, A.B., Carboni, S., Eva, H.D., Mayaux, P., Rasi,R., Simonetti, D., Stibig, H.-J., 2009. Monitoring forest cover at global scale: theJRC approach. In: Proceedings of the 33rd ISRSE Conference, 4 p.

araldi, A., Puzzolo, V., Blonda, P., Bruzzone, L., Tarantino, C., 2006. Automatic spec-tral rule-based preliminary mapping of calibrated Landsat TM and ETM+ images.IEEE Transactions on Geoscience and Remote Sensing 44 (9), 2563–2585.

arrett, H.R., Ilbery, B.W., Brown, A.W., Binns, T., 1999. Globalization and the chang-ing networks of food supply: the importation of fresh horticultural produce fromKenya into the UK. Transactions of the Institute of British Geographers 24 (2),159–174.

égué, A., Vintrou, E., Ruelland, D., Claden, M., Dessay, N., 2011. Can a 25-yeartrend in Soudano-Sahelian vegetation dynamics be interpreted in terms of landuse change? A remote sensing approach. Global Environmental Change 21,413–420.

enz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-readyinformation. ISPRS Journal of Photogrammetry & Remote Sensing 58, 239–258.

euchle, R., Eva, H.D., Stibig, H.-J., Bodart, C., Brink, A., Mayaux, P., Johansson, D.,Achard, F., Belward, A., 2011. A satellite data set for tropical forest area changeassessment. International Journal of Remote Sensing 32, 7009–7031.

odart, C., Eva, H., Beuchle, R., Rasi, R., Simonetti, D., Stibig, H.-J., Brink, A., Lindquist,E., Achard, F., 2011. Pre-processing of a sample of multi-scene and multi-dateLandsat imagery used to monitor forest cover changes over the tropics. ISPRSJournal of Photogrammetry and Remote Sensing 66, 555–563.

odart, C., Brink, A.B., Donnay, F., Lupi, A., Mayaux, P., Achard, F., 2013. Continentalestimates of forest cover and forest cover changes in the dry ecosystems of Africabetween 1990 and 2000. Journal of Biogeography 40 (6), 1036–1047.

ongers, F., Tennigkeit, T. (Eds.), 2010. Degraded forest in Eastern Africa, Manage-ment and Restoration. Earthscan, London.

rink, A.B., Eva, H.D., 2009. Monitoring 25 years of land cover change dynamicsin Africa: a sample based remote sensing approach. Applied Geography 29,501–512.

rink, A.B., Eva, H.D., 2011. The potential use of high-resolution Landsat satellitedata for detecting land-cover change in the Greater Horn of Africa. InternationalJournal of Remote Sensing 32 (21.).

rink, A.B., Eva, H.D., Bodart, C., 2012. Is Africa losing its naturalvegetation?—Monitoring trajectories of land cover change by means ofLandsat imagery. In: Giri, C.P. (Ed.), Remote Sensing of Land Use and Land Cover.CRC Press, Taylor & Francis Series in Remote Sensing Applications, Boca Raton,Florida, U.S.A., pp. 369–379.

abral, A.I.R., Vasconcelos, M.J., Oom, D., Sardinha, R., 2011. Spatial dynamics andquantification of deforestation in the central-plateau woodlands of Angola(1990–2009). Applied Geography 31, 1185–1193.

hander, G., Markham, B.L., Helder, D.L., 2009. Summary of current radiometric cal-ibration coefficients for Landsat MSS, TM, ETM+, and EO-1 ALI sensors. RemoteSensing of Environment 113 (5), 893–903.

otula, L., Vermeulen, S., Leonard, R., Keeley, J., 2009. Land grab or developmentopportunity? In: Agricultural Investment and International Land Deals in Africa.IIED/FAO/IFAD, London/Rome.

esclée, B., Simonetti, D., Mayaux, P., Achard, F., 2013. Multi-sensor monitoring sys-tem for forest cover change assessment in Central Africa. IEEE Journal of SelectedTopics in Applied Earth Observations and Remote Sensing 6 (1).

iouf, A., Lambin, E.F., 2001. Monitoring land-cover changes in semi-arid regions:remote-sensing data and field observations in the Ferlo, Senegal. Journal of AridEnvironments 48, 129–148.

AO, 2001. Global Forest Resources Assessment 2000. Forestry Paper 140. Rome.AO, 2010. Global Forest Resources Assessment 2010. Forestry Paper 163. Rome.AO, 2013. FAOSTAT – 2013 Database. Food and Agriculture Organization, Rome,

Available online at: http://faostat.fao.org/ (accessed 29.04.14).eist, H.J., Lambin, E., 2002. Proximate causes and underlying driving forces of trop-

ical deforestation. Bioscience 52, 143–150.ibbs, H.K., Ruesch, A.S., Achard, F., Clayton, M.K., Holmgren, P., Ramankutty, N.,

Foley, J.A., 2010. Tropical forests were the primary sources of new agriculturalland in the 1980s and 1990s. Proceedings of the National Academy of Sciencesof the United States of America 107, 16732–16737.

utman, G., Byrnes, R., Masek, J., Covington, S., Justice, C., Franks, S., Headley, R.,2008. Towards monitoring land-cover and land-use changes at a global scale:

bservation and Geoinformation 28 (2014) 60–69 69

the global land survey 2005. Photogrammetric Engineering & Remote Sensing74 (1), 6–10.

Hansen, M., Stehman, S., Potapov, P.V., Loveland, T.R., Townshend, J.R.G., DeFries, R.S.,Pittman, K.W., Arunarwati, B., Stolle, F., Steininger, M.K., Carroll, M., Di Miceli,C., 2008. Humid tropical forest clearing from 2000 to 2005 quantified by usingmulti-temporal and multi-resolution remotely sensed data. Proceedings of theNational Academy of Sciences of the United States of America 105, 9439–9444.

Jansen, L.J.M., Bagnoli, M., Focacci, M., 2008. Analysis of land-cover/use changedynamics in Manica Province in Mozambique in a period of transition(1990–2004). Forest Ecology and Management 254, 308–326.

Kebede, E., Kagochi, J., Jolly, C.M., 2010. Energy consumption and economic devel-opment in Sub-Sahara Africa. Energy Economics 32, 532–537.

Kenya Flower Council, 2013. http://www.kenyaflowercouncil.org web site accessedon 05.08.2013.

Lambin, E.F., Ehrlich, D., 1997. Land-cover changes in sub-saharan Africa(1982–1991): application of a change index based on remotely sensed surfacetemperature and vegetation indices at a continental scale. Remote Sensing ofEnvironment 61 (2), 181–200.

Lavreau, J., 1991. De-hazing landsat thematic mapper images. PhotogrammetricEngineering & Remote Sensing 57 (10), 1297–1302.

Matthews, E., 2001. Understanding the FRA 2000, Forest Briefing 1.Mayaux, P., Bartholomé, E., Fritz, S., Belward, A., 2004. A new land-cover map of

Africa for the year 2000. Journal of Biogeography 31, 861–877.Mayaux, P., Pekel, J.-F., Desclée, B., Donnay, F., Lupi, A., Achard, F., Clerici, M., Bodart,

C., Brink, A.B., Nasi, R., Belward, A., 2013. State and evolution of the African rain-forests between 1990 and 2010. Philosophical Transactions of the Royal SocietyB 368, 20120300.

Meier, P., Bond, D., Bond, J., 2007. Environmental influences on pastoral conflict inthe Horn of Africa. Political Geography 26, 716–735.

Meyfroidt, P., Lambin, E.F., Erb, K.-H., Hertel, T.W., 2013. Globalization ofland use: distant drivers of land change and geographic displacementof land use. Current Opinion in Environmental Sustainability 5, 1–7,http://dx.doi.org/10.1016/j.cosust.2013.04.003.

Molvaer, R.K., 1991. Environmentally induced conflicts? A discussion based on stud-ies from the Horn of Africa. Security Dialogue 22, 175–188.

Mugo, F., Ong, C., 2006. Lessons of eastern Africa’s unsustainable charcoal trade,ICRAF Working Paper no. 20. World Agroforestry Centre, Nairobi, Kenya.

Mwampamba, T.H., 2007. Has the woodfuel crisis returned? Urban charcoalconsumption in Tanzania and its implications to present and future forest avail-ability. Energy Policy 35, 4221–4234.

Ngendakumana, V., 2001. Data quality as limiting factor in the measuring and anal-ysis of food supplies – FAO’s Africa experience. Joint ECE/EUROSTAT/FAO/OECDmeeting on food and agricultural statistics in Europe. Geneva, 17–19 October.Food and Agricultural Organization, Rome.

Ogwal, F., Guloba, M., Weick, V., 2010. Horticulture production and biodiversity inUganda: benefits and risks associated with export growth strategies. BridgesTrade BioRes Review 4 (2).

Pellikka, P.K.E., Loetjoenen, M., Siljander, M., Lens, L., 2009. Airborne remote sensingof spatiotemporal change (1955–2004) in indigenous and exotic forest cover inthe Taita Hills, Kenya. International Journal of Applied Earth Observation andGeoinformation 11, 221–232.

Potapov, P.V., Turubanova, S.A., Hansen, M.C., Adusei, B., Broich, M., Altstatt, A., Mane,L., Justice, C.O., 2012. Quantifying forest cover loss in Democratic Republic of theCongo, 2000–2010, with Landsat ETM+ data. Remote Sensing of Environment122, 106–116.

Rasi, R., Bodart, C., Stibig, H.-J., Eva, H., Beuchle, R., Carboni, S., Simonetti, D., Achard,F., 2011. An automated approach for segmenting and classifying a large sample ofmulti-date Landsat imagery for pan-tropical forest monitoring. Remote Sensingof Environment 115, 3659–3669.

Richards, T.S., Gallego, J., Achard, F., 2000. Sampling for forest cover change assess-ment at the pan-tropical scale. International Journal of Remote Sensing 21,1473–1490.

Serneels, S., Lambin, E.F., 2001. Proximate causes of land-use change in NarokDistrict, Kenya: a spatial statistical model. Agriculture, Ecosystems and Envi-ronment 85, 65–81.

Stroppiana, D., Boschetti, M., Brivio, P.A., Carrara, P., Bordogna, G., 2009. A fuzzyanomaly indicator for environmental monitoring at continental scale. EcologicalIndicators 9, 92–106.

Tappan, G.G., Sall, M., Wood, E.C., Cushing, M., 2004. Ecoregions and land cover trendsin Senegal. Journal of Arid Environments 59, 427–462.

Vancutsem, C., Marinho, E., Kayitakire, F., Linda, S., Fritz, S., 2013. Harmonizing andcombining existing land cover/use datasets for cropland areas monitoring at theAfrican continental scale. Remote Sensing.

Weinzettel, J., Hertwich, E.G., Peters, G.P., Steen-Olsen, K., Galli, A., 2013. Affluence

drives the global displacement of land use. Global Environmental Change 23 (2),433–438, http://dx.doi.org/10.1016/j.gloenvcha.2012.12.010.

White, F., 1983. The Vegetation of Africa: A Descriptive Memoir to Accompany theUnesco/AETFAT/UNSO Vegetation map of Africa. Natural Resources Research,Unesco, Paris, 20.