Agricultural Systems 103 (2010) 609620

Contents lists available at ScienceDirect

Agricultural Systems
journal homepage: www.elsevier.com/locate/agsy

Modelling agricultural expansion in Kenyas Eastern Arc Mountains biodiversity hotspot

Eduardo Eiji Maeda *, Barnaby J.F. Clark, Petri Pellikka, Mika Siljander
University of Helsinki, Department of Geosciences and Geography, Gustaf Hllstrmin katu 2, 00014 Helsinki, Finland

a r t i c l e

i n f o

a b s t r a c t
The Taita Hills are the northernmost part of the Eastern Arc Mountains of Kenya and Tanzania, which is one of the most important regions for biological conservation in the world. The indigenous cloud forests in this area have suffered substantial degradation for several centuries due to agricultural expansion. In the Taita Hills, currently only 1% of the original forested area remains preserved. In order to create effective policies to preserve the natural resources and biodiversity of the Eastern Arc Mountains it is crucial to understand the causes and interactions involved in the landscape changes in the most degraded areas. The research presented here aimed to understand the role of landscape attributes and infrastructure components as driving forces of agricultural expansion in the Taita Hills. Geospatial technology tools and a landscape dynamic simulation model were integrated to identify and evaluate the driving forces of agricultural expansion and simulate future landscape scenarios. The results indicate that, if current trends persist, agricultural areas will occupy roughly 60% of the study area by 2030. Agricultural expansion will likely take place predominantly in lowlands and foothills throughout the next 20 years, increasing the spatial dependence on distance to rivers and other water bodies. The main factors driving the spatial distribution of new croplands were the distance to markets, proximity to already established agricultural areas and distance to roads. Other driving forces of the agricultural expansion, as well as their implications for natural resources conservation, are discussed. Further studies are necessary to integrate the effects of population pressure and climate change on the sustainability and characteristics of local agricultural systems. 2010 Elsevier Ltd. All rights reserved.

Article history: Received 18 January 2010 Received in revised form 7 July 2010 Accepted 16 July 2010 Available online 19 August 2010 Keywords: Land use change Simulation model Eastern Arc Mountains Taita Hills

1. Introduction The Eastern Arc Mountains (EAM) of Kenya and Tanzania maintain some of the richest concentrations of endemic animals and plants on Earth, and are thus considered one of the worlds top 25 biodiversity hotspots (Myers et al., 2000). The EAM comprise a chain of mountains located in southern Kenya and eastern Tanzania and are home for at least 96 endemic vertebrate species and 800 endemic vascular plant species (Burgess et al., 2007). Although this region is among the most important areas for biological conservation in the world, it has already lost approximately 80% of its original forest area (Hall et al., 2009). Forest loss in sub-Saharan Africa is proceeding at an alarming rate of 2.8 million ha per year; particularly in Afromontane areas the decrease is estimated to be 3.8% annually (Eva et al., 2006). Most of these losses are caused by agricultural expansion. Between the years 1975 and 2000 the agricultural areas increased 57% in sub-Saharan Africa (Brink and Eva, 2009). Although the

* Corresponding author. Tel.: +358 44 2082876. E-mail address: eduardo.maeda@helsinki. (E.E. Maeda). 0308-521X/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.agsy.2010.07.004

development of the agricultural sector is essential to improve food security in this region, the expansion of croplands without logistic and technological planning is a severe threat to the environment. Besides the imminent risk to biodiversity, indiscriminate agricultural activities may pose serious obstacles to water resources and soil conservation. One of the EAM sections most affected by agricultural expansion is the Taita Hills, which is the northernmost part of the EAM. Between 1955 and 2004, the indigenous forest areas in the Taita Hills decreased by 50% (Pellikka et al., 2009). Today, only 1% of the original forested area remains. Although only a small fraction of the indigenous cloud forests is preserved, the Taita Hills continue to have an outstanding diversity of ora and fauna and a high level of endemism. It is home for six endemic vertebrate species, three endemic bird species and at least 13 endemic plant species (Burgess et al., 2007; Brooks et al., 1998). Hence, a detailed study of this specic region is essential to preserve the remaining biodiversity and, most importantly, to expand the understanding of the interactions between human activities and landscape changes in the EAM. Such work can contribute to improving environmental protection policy in the regions of the EAM that are still intact but currently threatened by agricultural expansion.

610

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

The improvement in models and computer capacity during the past decades has allowed an increasing number of studies aiming at the sustainable use of natural resources and land use planning. For instance, land use and land cover change (LUCC) simulation models provide robust frameworks to cope with the complexity of land use systems (Veldkamp and Lambin, 2001; Washington et al., 2010). Such models are considered efcient tools to project alternative scenarios into the future and to test the stability of interrelated ecological systems (Koomen et al., 2008). Understanding the circumstances and driving forces of change is an essential step for elaborating public policies that can effectively lead to the conservation of natural resources. Among dynamic spatial models, frameworks operating on a cellular automata (CA) basis have arisen as a feasible alternative for the analysis of land use dynamics and in the exploration of future landscape scenarios. CA models consist of a simulation framework in which space is represented as a grid of cells, and a set of transition rules determine the attribute of each given cell considering the attributes of its neighboring cells (Almeida et al., 2003). For instance, CLUE-S is a spatially explicit, multi-scale model that can be applied to describe land change dynamics through the determination and quantication of the bio-geophysical and human drivers of land use (Verburg et al., 2002). Another widely used CA simulation model is Dinamica-EGO (Soares-Filho et al., 2002, 2009). Dinamica-EGO is an environmental modelling platform developed by the Centre for Remote Sensing of the Federal University of Minas Gerais, Brazil (CSR-UFMG). This platform allows the design of static or dynamic simulations involving nested iterations, dynamic feedbacks and multi-scale approaches. Moreover, geospatial technologies, such as remote sensing and geographical information systems (GIS), have made available an unprecedented opportunity for new studies in terms of data collection, availability and processing capacity. Nevertheless, scientists currently face the challenge of integrating these technologies to better understand the coupled relations between human activities and environmental changes. In this context, land change science has emerged as an interdisciplinary eld that aims to understand the dynamics of LUCC as a coupled human environment system (Turner et al., 2007). This study addresses this exact issue, aiming to understand the role of landscape attributes and infrastructure components as driving forces of agricultural expansion in the Taita Hills. In order to achieve this objective, remote sensing, GIS techniques and a LUCC simulation model were integrated to identify and evaluate the driving forces of LUCC and simulate future landscape scenarios. Thus, it is hoped that the results of this study may represent an important instrument for households, researchers and policy makers to better cope with future changes in local agricultural systems.

maize species and sorghum and millet species are cultivated. In the lower midland zones with average rainfall between 500 and 700 mm, dryland maize types and onions are cultivated, among others. Moreover, the area is considered to have high scientic interest and there is a high potential for succeeding in connecting economic development and community-based natural resource management (Himberg et al., 2009).

3. Material and methods This research integrated remote sensing, GIS techniques and a spatially explicit simulation model of landscape dynamics, Dinamica-EGO (Soares-Filho et al., 2007), to assess the driving forces of agricultural expansion in the study area and simulate future scenarios of land use. A general description of the applied method is illustrated in Fig. 2. The model receives as inputs land use transition rates, landscape variables and landscape parameters. The landscape parameters are intrinsic spatially distributed features, such as soil type and slope, which are kept constant during the simulation process. The landscape variables are spatialtemporal dynamic features that are subjected to changes by decision makers, for instance roads and protected areas. The land use transition rates were also considered to be decision variables, given that this modelling exercise was based on the assumption that agricultural expansion rates can be modied by public policies or other external forces. The model is driven by land use and land cover maps (LULCM) from two selected dates: 1987 (initial landscape) and 2003 (nal landscape), which are used as inputs to represent the historical land use transitions in the study area. The dates of the LULCM were chosen based on two criteria. The rst criterion was that the landscape changes between the initial and nal landscape should accurately represent the ongoing land change activities in the study area. That is to say, the agricultural expansion rates between 1987 and 2003 were assumed to be representative of current trends. The second criterion relied on the availability of cloud free satellite images to assemble the LULCM.

3.1. Landscape variables and parameters In total, ten landscape attributes (variables/parameters) were used as inputs for the model, nine of which were static and one of which was dynamic. Static inputs are those that are kept constant throughout the model run, while dynamic inputs refer to those that undergo changes during the model run. All landscape attributes were represented by raster images with a 20 m spatial resolution. The description of each layer is as follows: Distance to roads (DRo): Euclidian distance in meters to main and secondary roads. Distance to Markets (DM): The markets were represented by main villages in the region; the distance to markets raster was created by calculating the Euclidian distance in kilometers to centre of each village. Digital Elevation Model (DEM): The 20 m spatial resolution DEM was interpolated from 50-feet interval contours captured from 1:50,000 scale topographic maps, deriving an estimated altimetric accuracy of 8 m and an estimated planimetric accuracy of 50 m. Distance to Rivers (DRi): Represented by the Euclidian distance in meters to main rivers. Two sources were used to extract the river network in the study area. Firstly, GIS tools were used to automatically identify the rivers based on a ow accumulation grid obtained using the DEM. Subsequently, eventual errors in the automatic classication were corrected using a 1:50,000 scale topographic map.

2. Study area Taita Hills is the northernmost part of the EAM biodiversity hotspot, situated in the middle of the Tsavo plains of the Taita-Taveta District in the Coastal Province, Kenya (Fig. 1). Taita Hills cover an area of 1000 km2. The population of the whole Taita-Taveta district has grown from 90,146 (1962) persons to over 300,000 (Republic of Kenya, 2001). The indigenous cloud forests have suffered substantial loss and degradation for several centuries as abundant rainfall (annual 1100 mm) and rich soils (cambisols and humic nitosols) have created good conditions for agriculture. The agriculture in the hills is intensive small-scale subsistence farming. In the lower highland zone and upper midland zone, the typical crops are maize, beans, peas, potatoes, cabbages, tomatoes, cassava and banana (Soini, 2005). In the slopes and lower parts of the hills with average annual rainfall between 600 and 900 mm, early maturing

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

611

Fig. 1. Geographic location of the study area.

Fig. 2. General description of the method, in which landscape attributes obtained using remote sensing and GIS techniques are used as inputs for a LUCC model. The model evaluates the role of each attribute in the land changes and simulates future landscape scenarios.

Protected Areas (PA): Characterized by the national parks and conservation areas close to the Taita Hills. A segment of Tsavo East National Park is located in the northeastern part of the study area and a small section of Taita Hills Game Sanctuary in the southwestern. Soil Type (ST): The soil map was obtained from the Soil and Terrain Database for Kenya (KENSOTER), at scale 1:1 M, compiled by the Kenya Soil Survey (Batjes and Gicheru, 2004). Slope (S): The slope (%) was extracted from the DEM. Insolation (I): Annual average solar radiation in watt hours per square meter (W h/m2) for the whole year was created from the DEM using ArcGIS 9.3. Mean annual precipitation (P) was obtained by the compilation of long term mean precipitation grids interpolated from available

meteorological data in surrounding areas using ANUSPLINE software (Hutchinson 1995; Maeda et al., 2010a). Distance to croplands (DC): Represented by the Euclidian distance to already established croplands. This layer was the only dynamic landscape attribute used as an input for the model, which means that this variable undergoes changes during the model run as new cropland patches are created. 3.2. Global transition rates and local transition probabilities Global transition rates refer to the total amount of changes for each type of land cover given in the simulation period, without taking into account the spatial distribution of such changes. The transition rates were calculated by cross-tabulation, which produced as

612

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

output a transition matrix between the LULCMs from 1987 and 2003. The transition matrix describes a system that changes over discrete time increments, in which the value of any variable in a given time period is the sum of xed percentages of the value of the variables in the previous time period (Soares-Filho et al., 2009). The land use/land cover for the Taita Hills was mapped from SPOT 4 HRVIR satellite images (path & row 143357), with 20 m spatial resolution and green, red and Near-infrared (NIR) spectral bands. Because of the rugged terrain in the area, the images were orthorectied using a 20 m planimetric resolution DEM. Atmospheric correction was implemented utilizing the historical empirical line method (HELM) (Clark and Pellikka, 2009). The SPOT images were classied according to a nomenclature derived using the land cover classication system (LCCS) of the Food and Agriculture Organization of the United Nations (FAO) and the United Nations Environment Programme (UNEP) (Di Gregorio, 2005). The land cover classication methodology itself utilized a multi-scale segmentation/object relationship modelling (MSS/ORM) approach implemented with the Deniens software tool (Burnett and Blaschke, 2003). Central to this methodology is the generation of meaningful image objects, relating to homogeneous land cover patches, by multi-scale segmentation based on both spectral and textural characteristics of the imagery. More detailed description of the method is presented in Clark and Pellikka (2009). An accuracy assessment for the 2003 classication was undertaken based on ground reference test data, independent of the training data, collected during eld visits in 2005/2006 using stratied random road sampling. Additional reference points for the test data were collected from 0.5 m resolution airborne true-colour digital camera imagery acquired in January 2004. The local transition probabilities, different from the global transition rates, are calculated for each grid cell considering the natural and anthropogenic characteristics of the site. Different methods can be used to estimate local probabilities in Dinamica-EGO, such as logistic regression and neural networks (Soares-Filho et al., 2002; Almeida et al., 2008). In the particular case of this work, the transition probability of each cell was calculated in Dinamica-EGO using the weights of evidence (WoE) method. The WoE is a Bayesian method, in which the effect of each landscape variable on a transition is calculated independently of a combined solution (Soares-Filho et al., 2002). The spatial probability of a transition is given by the following equation (Bonham-Carter, 1994):

where P{Vi/T} is the probability of variable Vi occurring in the presence of transition T, given by the number of cells where both Vi and T are found divided by the total number of cells where T is found; and P{Vi/T } is the probability of variable Vi occurring in the absence of transition T, given by the number of cells where both Vi and T are found divided by the total number of cells where T is not found. Hence, the W+ values represent the attraction between a determined landscape transition and a certain variable. The higher the W+ value is, the greater is the probability that a certain transition takes place. On the other hand, negative W+ values indicate lower probability of a determined transition occurring in the presence of the respective variable range. Based on the W+ values of each range for every variable considered, Dinamica-EGO generates a spatially explicit probability map, in which each cell is assigned the probability for a determined transition. 3.3. Calibration of Dinamica-EGO internal parameters The Dinamica-EGO platform has the advantage of using stochastic algorithms for land use change allocation. Two transition algorithms are responsible for the allocation of the land use/land cover changes: expander and patcher. The expander function performs the expansion of previously existing patches of a certain class. The patcher function, in turn, is designed to generate new patches through a seed formation mechanism (Soares-Filho et al., 2002). The applied algorithms consist of scanning the initial land cover/land use map to sort out the cells with the highest probabilities and then arrange them in a data array. Following this procedure, cells are selected randomly from top to bottom of the data array (the internal stochastic selection mechanism can be loosened or tightened depending on the degree of randomisation desired). In a nal step, the land cover/land use map is again scanned to perform the selected transitions. Hence, the rst parameter to be calibrated in the simulation is the percentage of changes that will be addressed by each of these two algorithms. For instance, in regions where landscape changes happen exclusively by the expansion of existing patches, all changes should be arranged by the expander function. The next parameters to be adjusted are the mean and variance of new patches sizes. These parameters can be independently adjusted for the expander and patcher functions. The model also includes another heuristic parameter denoted the patch isometry index. A high isometry index results in compact patches, while low values are reected in more fragmented formations. The calibration of all parameters was conducted simultaneously. Firstly, initial values are set based on empirical knowledge obtained in previous eld work in the region. Next, the model is run to simulate the landscape scenario in 2003 using the map from 1987 as the initial landscape. The simulated landscape for 2003 is compared with the reference LULCM by visual analysis, and the parameters adjusted through heuristic procedures in order to improve the results. The model is iteratively executed and evaluated, and the results compared with the previous run. These procedures are repeated until an optimal result is found in the model evaluation. The model tting was evaluated using an adaptation of the method proposed by Hagen (2003), in which multiple resolution windows are used to compare the simulated and the reference maps within a neighbourhood context. Each type of change is analysed separately using pairwise comparisons involving maps of differences: (i) between the initial land use/land cover map and a simulated one, and (ii) between the same initial land use/land cover map and the reference one. This modication is able to tackle two matters. First, as it deals with only one type of change at a time, the overall two-way similarity measure can be applied to the entire map, regardless of the different number of cells per

Pn OfT g e i1 W x;y Pn ; Px;y fT =V 1 \ V 2 \ \ V n g P W i 1 x; y 1 OfT g t j 1 e

where Px,y is the probability of transition in a cell with coordinates x, y; T is land use/land cover transition; Vi is all possible landscape variables i selected to explain transition T; and O{T} is the odds of a transition, represented by the ratio between a determined transition probability and the complementary probability of non-occurrence, described by following equation:

OfT g

PfT g PfT g

where P{T} is the probability of transition T occurring, given by the number of cells where the relevant land use/land cover transition occurred divided by the total number of cells in the study area; P{T } is probability of transition T not occurring, given by the number of cells where the relevant land use/land cover transition is absent divided by the total number of cells in the study area; and W x;y is the weight of evidence for a determined landscape variable range, dened by the following equation:

W loge

PfVi=T g ; PfVi=T g

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

613

category. Second, the inherited similarity between the initial and simulated maps can be eliminated from this comparison by simply ignoring the null cells from the overall count (Almeida et al., 2008). Approaches considering neighbourhood contexts are useful in comparing maps that do not exactly match on a cell-by-cell basis, but still present similar spatial patterns within certain cell vicinity (Soares-Filho et al., 2002). The method retrieves a fuzzy similarity index dened inside a window that is gradually expanded, allowing the assessment of model performance at multiple resolutions. This fuzzy similarity index is based on the concept of fuzziness of location, in which a representation of a cell is inuenced by the cell itself and by the cells in its vicinity (Hagen, 2003). The comparison results in a map that species for each pixel the degree of similarity on a scale of 01, where zero represents total disagreement and one represents identical maps. 3.4. Simulation of future scenarios After calibration, the model is executed using a recent LULCM as an initial landscape, producing simulated landscape scenarios as output. In this study, the LULCM from the year 2003 was considered to be the initial landscape, and the model was applied to simulate two different scenarios up to 2030. One exploratory and one prescriptive scenario were simulated. A more detailed description of each scenario is presented below: Business as Usual (BAU): This scenario was simulated using an exploratory approach. An exploratory scenario is a sequence of emerging events (Alcamo, 2001). Here, the average agricultural expansion rates observed from 1987 to 2003 in the study area were used to build an exploratory scenario with stationary behaviour for the year 2030. Hence, in this case the agricultural expansion rates are considered a static input for the model and are not affected by policy decisions. Governance (GOV): In this case, a prescriptive scenario was simulated. Prescriptive scenarios are established a priori by the modeller in accordance with a targeted future (Alcamo, 2001). Thus, the agricultural expansion rates were associated with ctitious governance policies, in which the LUCCs were constrained according to the availability of Renewable Freshwater Resources (RWR). RWR is dened as the water that is continuously recharged in the hydrological cycle. Here, it was represented by the annual average rainfall volume. The assumption made was that annual Irrigation Water Requirements (IWR) could not exceed 70% of the total RWR, leaving the remaining 30% to be used for residential or commercial purposes. Consequently, the agricultural expansion rates are gradually decreased through simulated time as the IWR would approach the RWR limit. The 70% threshold was based on the global average distribution of water resources withdraws (FAO, 2005) and used as a virtual limit of water consumption. This approach, however, does not represent any real policy or water management strategies. The equation used to calculate the transition rates is written as follows:

evapotranspiration loss from a cropped eld (Allen et al., 1998), and is represented here by the Crop Evapotranspiration (ETc). Evapotranspiration (ET) is dened as the combination of two separate processes, in which water is lost on the one hand from the soil surface by evaporation and on the other hand from the crop by transpiration (Allen et al., 1998). In cases where all the water needed for optimal growth of the crop is provided by rainfall, irrigation is not required and the Irrigation Water Requirement (IWR) is equal to zero. In cases where all water has to be supplied by irrigation the IWR is equal to the CWR (ETc). However, when part of the CWR is supplied by rainfall and the remaining part by irrigation, the IWR is equal to the difference between the ETc and the Effective Precipitation (Peff). In such cases, the IWR was computed using the following equation (FAO, 1997):

IWRm ETcm 30 Peff m ;

where IWRm is the monthly average crop water requirement in month m, [mm]; ETcm is mean daily crop evapotranspiration in month m, [mm day1]; Peffm is the average effective precipitation in month m, [mm]. The Peff is dened as the fraction of rainfall retained in the root zone, which can be effectively used by plants: that is, the portion of precipitation that is not lost by runoff, evaporation or deep percolation (Brouwer and Heibloem, 1986). The ET was calculated using the Hargreaves model (Hargreaves and Samani, 1985). A detailed illustration of this method, as well as the procedures used to parameterize and calibrate the ET model, are beyond the scope of this paper, and can be found in Maeda et al. (2010b). 4. Results 4.1. Global transition rates The overall accuracy of the 2003 land cover map was 89%, with a Kappa index for agreement of 0.87. The class specic producers and users accuracy assessment results are shown in Table 1. The producers accuracy indicates the probability of a reference pixel

Table 1 2003 Land use and land cover map classication accuracy assessment. Land cover class Cropland Shrubland Woodland Plantation forest Broadleaved forest Grassland Bare soil Built-up area Water Overall accuracy Overall kappa index of agreement # Reference test data sites 122 67 47 72 33 31 59 25 20 Producer accuracy (%) 95.9 64.2 91.5 97.2 97.0 71.0 84.7 96.0 100.0 User accuracy 81.8% 82.7% 91.5% 94.6% 100.0% 95.7% 90.9% 92.3% 100.0% 89.0% 0.87

Ryab

Riab IWRy 0; 7 RWR Riab ; IWRi 0; 7 RWR

where Ryab is the transition rate from a to b in year y; Riab is transition rates for the same land use types in the beginning of the simulation; IWRy is annual average irrigation water requirements during year y; IWRi is annual average irrigation water requirements in the beginning of the simulation; and RWR is the annual average renewable freshwater resources. Some important concepts are central to a better understanding of this approach. For instance, Crop Water Requirement (CWR) is dened as the amount of water required to compensate the

Table 2 Annual average agricultural expansion rates (baseline 19872003). Original vegetation Shrubland Woodland Plantation forest Broadleaved forest Grassland Annual conversion rate (%) (baseline 19872003) 1.305 2.013 1.161 0.289 0.310

614

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

for a particular category on the map being correctly classied, and is a measure of the omission error. The users accuracy indicates the probability that a pixel classied on the map actually represents that category on the ground, and is a measure of the commission error. It can be seen that, for the most part, accuracies were good with the croplands class, for example, having a producers accuracy of 96% and a users accuracy of 82%. The exception was

the lower producer accuracies of the shrubland and grassland classes, due to misclassication errors with certain areas of cropland where either shrub-like or grass-like crops gave very similar spectral and textural characteristics in the SPOT imagery. Because of a lack of timely ground reference test data or aerial photography, the accuracy of the 1987 classication could not be assessed directly. However, as the same classication methodology was applied to

Fig. 3. Graphics showing the W+ values attributed for each range of six landscape attributes most related to the shrublands to croplands transition.

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

615

both scenes, the 1987 map accuracy was assumed to be similar to that of the 2003 map. The annual average agricultural expansion rates observed from 1987 to 2003 are shown in Table 2. The highest conversion rates were observed in the transition from woodlands to agriculture. However, considering absolute numbers, shrubland areas are the most affected, given that currently it represents the predominant vegetation type in the region. The small regions covered with broadleaved forests were nearly untouched, presenting low conversion rates; the total area decreased from 7.7 to 6.9 km2 during the observed period. 4.2. Local transitional probabilities The most relevant W+ values obtained during the model calibration are shown in Figs. 35. This information represents the attraction between a determined landscape transition and a certain landscape attribute. Several attributes were important in the conversion from shrublands to croplands: distance to rivers, insolation, distance to croplands, DEM, distance to roads and distant to markets were particularly associated with this transition (Fig. 3). The distance to croplands is an important driving factor for all transitions (Figs. 35) indicating that the proximity to previously established croplands is a key factor for agricultural expansion in

this region. Areas with low insolation were indicated as having lower probability of being converted to croplands. Although areas close to rivers did not exhibit high positive W+ values, the importance of water bodies for croplands is clearly reected in distance to rivers, where high negative W+ values are observed. Hence, the results indicate that patches further than 1 km from water bodies have lower probability of being converted to cropland (Figs. 3 and 4). Distance to roads also demonstrated a clear pattern in inuencing the transition from shrublands to croplands (Fig. 3). That is to say, the probability of a shrubland area being converted to cropland linearly decreases with distance away from the roads. Nevertheless, DRo did not show very high W+ values, possibly due to the fact that the Taita Hills includes a relatively dense road network, diminishing the contrasts between areas near to and away from roads. The distance to markets, here represented by the Euclidean distance to the main villages, was the most representative driving force for shrublands to croplands transition (Fig. 3). Shrublands patches within up to 100 m from markets have particularly high probability of being converted to croplands, while patches farther than 10 km from markets are unlikely to be used as agricultural lands. Particular to the woodlands to croplands transition is the inuence of the slope parameter on the transition probability. On slopes

Fig. 4. Graphics showing the W+ values attributed for each range of four landscape attributes most related to the woodland to croplands transition.

616

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

Fig. 5. Graphics showing the W+ values attributed for each range of two landscape attributes most related to the plantation forest to croplands transition.

higher than 30% woodland patches are unlikely to be converted to croplands. This can be explained by the fact that woodlands may act as a soil conservation component in steep slopes areas, and at the same time sloping terrains are less appropriate for agriculture. For the transition from plantation forests to croplands, the results show that areas with low precipitation (i.e. lower than 550 mm/year) have higher probability of being converted (Fig. 5). Some factors may explain this result. Firstly, the low availability of areas suitable for agricultures in the highlands, where the precipitation rates are higher, drives the agricultural expansion to the remaining areas with lower rainfall. Secondly, although croplands can, to some extent, migrate to lower precipitation areas, drier lands are not likely to be suitable for plantation forests. Moreover, most plantation forests in the Taita Hills are government owned, and for this reason cannot be converted to croplands. Once the driving forces of LUCC are dened, the model is iteratively executed and calibrated, using as an initial landscape the LULCM from the year 1987. The objective of this approach was to execute the model to simulate the landscape observed in the year 2003 and compare with the reference LULCM obtained from the satellite image classication. An important parameter to be calibrated during this process is the size and characteristics of new cropland patches. It was found that areas converted to agriculture have an average size of 3 ha, with a standard deviation of 3.2 ha. From the new cropland patches stochastically allocated in the simulations, 90% were handled by the patcher algorithm and 10% by the expander algorithm. In actual fact this means that, on average, 90% of new agricultural areas are created to be separate from existing crop elds, and the remaining 10% are a result of the expansion of current elds. The evaluation of the LUCC model performance is illustrated in Fig. 6, which shows the fuzzy similarity indices achieved using different window sizes. The maximum fuzzy similarity indices ranged from around 55% at a spatial resolution of 20 m to 90% at a spatial resolution of 380 m. 4.3. Scenarios simulation After the model is calibrated and the role of each landscape variable dened, transition probability maps are created for each simulated year. In Fig. 7, an example of a transition probability map for the year 2003 is illustrated. In the map, the light colors represent areas with higher probabilities of being converted to croplands.

Fig. 6. Fuzzy similarity indices based on multiple size windows obtained in the model tting evaluation.

Fig. 7. Map illustrating the probability and area being converted to cropland during the years 2003/2004.

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

617

Once spatial probabilities are dened, the new agricultural patches are stochastically allocated by the expander and patcher algorithms. In Fig. 8, the historical land use maps for 1987 and 2003 are displayed (upper left and upper right) together with the land use maps for 2030 resulting from the BAU and GOV scenarios simulations (lower left and lower right). It is observed that, in 1987, croplands were already clearly established in the highlands (central area in the maps). This is explained by the favorable climatic and edaphic conditions for agricultural activities (e.g. high precipitation rates), which resulted in the clearance of large areas of forest during the last century. The protected areas, situated in the northeast (grassland area) and southwest parts of the study area (Fig. 8) were effective in containing the agricultural expansion between 1987 and 2003. The inuence of this variable was reected in both simulated scenarios, where a clear boundary constraining the expansion of croplands is observed in the limits of the protected areas. Hence, although a high environmental pressure is present along the fringes of the conservation areas, such as Tsavo East National Park, the areas inside the park are likely to be preserved.

Between 1987 and 2003, croplands started to be implemented with higher intensity in the lowlands, given that suitable areas for agriculture activities in the highlands were already almost entirely taken. This trend is clearly reected in the LUCC simulation results. Although each simulated scenario was created independently, using different transitions rates, the spatial distribution of new cropland patches followed the same patterns in all simulations. As suitable agricultural areas in highland disappeared, the expansion of new patches was distributed in the foothills. Among the main driving forces of such distribution were the distance to markets (here represented by villages or towns), distance to roads and distance to rivers. Distance to markets and roads played an interesting role in croplands distribution, in the sense that the effects of these two variables in the landscape dynamic were closely related. Towns and villages acted as core points, which are interconnected by roads creating axes in which new cropland patches were settled. Such patterns were observed mainly in the southern and southwestern parts of the study area. It is also important to notice the enhanced importance attributed to rivers in the land use dynamic. Given that in this region

Fig. 8. Historical land use maps for 1987 and 2003 (upper left and upper right) and simulated scenarios for 2030 (lower left and lower right).

618

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

Fig. 9. (a) Percentage of the main land use/land cover classes during the years 1987, 2003 and in both scenarios simulated for 2030; (b) LUCC simulation results from 2010 to 2030.

the foothills typically have higher average temperatures and lower precipitation volumes, the proximity to water bodies is essential to the establishment of agricultural activities. The numerical results of the simulations are presented in Fig. 9a and b, which display the percentage of the main land use/land cover classes during the years analyzed, and total cropland areas in the study site from 2010 to 2030, respectively. In the BAU scenario, the cropland areas expanded to around 515 km2 in 2030, corresponding to about 60% of the study area (Fig. 9a). This represents an increase of 40% in comparison to the year 2003, when croplands occupied around 365 km2. Although the effects of the governance scenario cannot be easily identied in the map showed in Fig. 8, the simulated land use policies resulted in a signicant reduction in agricultural expansion (Fig. 9b). The total area used for agriculture in 2030 for the GOV scenario was approximately 485 km2.

5. Discussion In contrast to studies aiming to assess the social and economical component of LUCC, this study focused on the inuence of physical landscape attributes and infrastructure components on the spatial distribution patterns of agricultural expansion. Although socioeconomic studies are helpful to better understand the uctuations and tendencies in land change rates (Rudel, 2009), they may be inadequate to numerically analyse the spatial probabilities of changes and the consequent development of plausible landscape scenarios. The results obtained in the LUCC driving forces analysis closely agree with previous studies carried out in other locations in Kenya. For instance, studying the Narok District in Kenya, Serneels and Lambin (2001) found that the expansion of small holder agriculture is mainly controlled by proximity to permanent water, land suitability and vicinity to villages. Moreover, the authors found differences in the factors driving mechanized agriculture, where conversions are mainly driven by distance to the markets and agroclimatic potential. In another similar work, Mertens and Lambin (2000) designed a spatially explicit model driven by a spectrum of socio-economic and infrastructure variables to simulate deforestation in Southern Cameroon. According to the authors, in this study case the results suggested that roads mostly increased the accessibility of the forest for migrants rather than providing incentives for the establishment of market oriented farming systems. Therefore, the comparison with previous studies highlights the fact that, although many similarities can be found at large scales,

some characteristics of the landscape dynamic are intrinsic, and can only be assessed locally. According to Lambin et al. (2003), land changes are in general caused by multiple interacting factors, which vary in time and space, according to specic human-environment conditions. In this context, similarities and disparities are also observed when comparing the driving forces of LUCC in the Taita Hills with agricultural expansion areas in other tropical forests. For instance, in the rain forests of South America, distance to roads and distance to markets arise as common factors affecting croplands expansion (Aguiar et al., 2007). In agreement with the present study, Alves (2002) indicates that new agricultural patches in the Brazilian Amazon spring up predominantly in the surroundings of pioneer areas. On the other hand, in contrast to the industrial agricultural activities in the Brazilian Amazon, new cropland patches in the Taita Hills are predominantly small, targeting subsistence or local production. New cropland patches in the Taita Hills were found to have an average size of 3 ha, with a variance of 10 ha, while in the agricultural expansion areas in So Feliz do Xingu, Brazilian Amazon, new agricultural patches have an average size of 300 ha, with a standard deviation of 22.3 ha (Ximenes et al., 2008). The performance achieved in the LUCC model calibration is considered satisfactory. In the multiple resolution tting procedure, the simulations achieved spatial ttings from 75%, at a spatial resolution of 100 m, up to 90% at a spatial resolution of 380 m. Such results are consistent with previous studies carried out with Dinamica-EGO. For instance, aiming to simulate the landscape dynamics in an Amazonian colonization frontier, Soares-Filho et al. (2002) calibrated the model for two study sites, reaching average spatial tting values between 63.3% and 82.4% at a spatial resolution of 100 m. In another similar study, the CLUE-S model was applied to simulate deforestation in the Klang-Langat watershed, Malaysia (Verburg et al., 2002). The study evaluated the deforestation process as a function of socio-economic and biophysical driving factors and also considered driving forces related to the urbanization process, such as the accessibility of the area by different road types. In this study case, the model t ranged from approximately 65% at a window size of 150 m, to 84% with a window of 10.5 km. Although the remaining biodiversity hotspot forests in the Taita Hills are likely to be preserved, the simulated scenarios clearly indicate a tendency towards the continuity of habitat fragmentation caused by agricultural expansion. It is known that habitat fragmentation can signicantly reduce species richness and change ecosystem functioning (Flynn et al., 2009). Likewise, the destruction of natural habitats may signicantly increase the dependence of species on protected areas (Jackson and Gaston, 2008). Studying

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620

619

the effects of habitat fragmentation on seed dispersal of an interior tree in the Taita Hills, Lehouck et al. (2009) show that forest deterioration reduced avian visitation and seed removal rates, creating long-term effects on tree communities. In this context, Hall et al. (2009) state that in order to ensure the efcacy of conservation efforts in the EAM hotspots, one needs to consider the extent of habitat changes within and across the entire elevational ranges found in these ecosystems. The agricultural expansion simulated in the presented study will have direct impacts on soil and water conservation issues by the year 2030. The natural vegetation protects the soil against the impacts of rainfall and provides organic matter to the soil. These factors improve water inltration and recharging of groundwater reservoirs. When this vegetation cover is displaced the soil becomes vulnerable to compaction and the water that would otherwise have inltrated the soil is turned into surface runoff, which can potentially carry out sediments and nutrients to rivers (Maeda et al., 2009). Such changes in the landscape dynamic can potentially cause environmental problems such as erosion, silting of rivers, eutrophication, and water contamination, among others (Maeda et al., 2008, 2010a). The benets of increasing agricultural areas for the local economy need to be carefully analyzed. Although the local economy is mainly based on agricultural activities, the results of this study show that croplands are currently expanding throughout the foothills, that is to say, in the direction of areas with lower precipitation and higher temperatures. This dynamic will likely lead to a higher water resources demand for irrigation (Maeda, 2009), having the potential to generate water conicts. In addition, Thornton et al. (2010) warn that, due to continuing population increases, together with climate change, agricultural systems may have to undergo substantial intensication during the next decades. Therefore, alternative solutions, such as the implementation of drought resistant crops, should be explored in order to meet agricultural production requirements. Such approaches can increase crop yield and potentially spare land for wild nature (Ewers et al., 2009). Hence, agricultural expansion will likely increase land pressure, having direct and indirect consequences on the characteristics of farming systems. For instance, Fermont et al. (2008) show that, during the last few decades, increasing populations and lower availability of arable land contributed to signicant changes in the crops cultivated in parts of mid-altitude zones of East-Africa. Traditional farming systems based on millet, cotton, sugarcane and/or banana have evolved into continuously cultivated cassava or cassava/maize-based systems. The authors state that the declining soil fertility and food shortage is likely the primary trigger for this transformation, given that cassava can recycle nutrients and tolerate poor soils. In this sense, the scenarios simulated in this study can offer a valuable tool to guide informed policy decisions in the direction of the improvement of agricultural systems, land use allocation and environmental protection. An immediate and straightforward application of the results of this study will be in supporting local authorities and researchers in implementing new water quality control and stream ow measuring stations. Water quality measurement stations aiming to assess the impacts of agricultural expansion are expected to be implemented during the coming years. The results presented here can elucidate the priority regions where new croplands are likely to occur and, consequently, the approximate location where measurements need to be taken in order to achieve consistent results. The use of scenarios to improve agricultural systems strategies has been extensively reported in the literature. Such studies demonstrate the importance of this approach in simulating major directions of agricultural development, alternative strategies for

soil conservation, irrigation and land use allocation. The present study represents an important contribution in this context. It can be considered an additional step in the direction of lling an evident gap in the understanding of environmental changes at local scales in Eastern Africa.

6. Conclusions The research presented here was able to identify the main landscape attributes driving the agricultural expansion in the Taita Hills. A connected relationship between villages and roads is evident in the denition of new cropland patches. Proximity to already established crop elds is also one of the key factors driving agricultural expansion. If current trends persist, it is expected that agricultural areas will occupy 60% of the study area by 2030. LUCC simulations indicate that agricultural expansion will likely take place predominantly in lowlands and foothills throughout the next 20 years. Such dynamics will increase the spatial dependence on distance to rivers and other water bodies due to the higher potential evapotranspiration in these areas. On the other hand, given that the most favorable areas for agriculture are already taken, agro-climatic factors (e.g. precipitation) will have a decreased inuence on the spatial distribution of croplands. Current trends indicate that the small residual areas of tropical cloud forest, home for a great part of the biodiversity in the Taita Hills, will likely remain intact throughout the coming years. Nevertheless, the impact of increasing habitat fragmentation on such biodiversity, caused by the agricultural expansion discussed in this paper, is a relevant issue that must be addressed in further research. The results described in this study have good potential to be used by policy makers in improving the identication of priority regions from the point of view of land use allocation and environmental risks. Moreover, the integrated modelling presented represents an important tool for researchers to understand the humanenvironment relations in this region. Acknowledgements Part of this research was carried out during the Young Scientists Summer Program (YSSP) at the International Institute for Applied System Analysis (IIASA), Austria. The authors kindly thank the IIASAs researchers Dr. Marek Makowski, Dr. David Wiberg and Dr. Tatiana Ermolieva for their comments. Thanks are also given to Dr. Cludia Maria de Almeida from the National Institute for Space Research (INPE), Brazil, for her contribution in improving part of the methodological descriptions presented in this paper. The research was funded by the Centre of International Mobility (CIMO), University of Helsinki and Academy of Finland for TAITATOO project. References
Aguiar, A.P.D., Cmara, G., Escada, M.I.S., 2007. Spatial statistical analysis of land-use determinants in the Brazilian Amazonia: exploring intra-regional heterogeneity. Ecological Modelling 209 (24), 169188. Alcamo, J., 2001. Scenarios as tools for international environmental assessments. In: Experts Corner Reports Prospects and Scenarios No 5. European Environment Agency, Copenhagen. Allen, R.G., Pereira, L.S., Raes, D., Smith, M., 1998. Crop Evapotranspiration. Guidelines for Computing Crop Water Requirements. FAO Irrigation and Drainage Paper 56. FAO, Rome. Almeida, C.M., Batty, M., Monteiro, A.M.V., Camara, G., Soares-Filho, B.S., Cerqueira, G.C., Pennachin, C.L., 2003. Stochastic cellular automata modeling of urban land use dynamics: empirical development and estimation. Computers, Environment and Urban Systems 27 (5), 481509.

620

E.E. Maeda et al. / Agricultural Systems 103 (2010) 609620 Lambin, E.F., Geist, H.J., Lepers, E., 2003. Dynamics of land-use and land-cover change in tropical regions. Annual Review of Environment and Resources 28, 205241. Lehouck, V., Spanhove, T., Colson, L., Adringa-Davis, A., Cordeiro, N.J., Lens, L., 2009. Habitat disturbance reduces seed dispersal of a forest interior tree in a fragmented African cloud Forest. Oikos 118 (7), 10231034. Maeda, E.E., 2009. Impacts of agricultural expansion on irrigation water requirements in Taita hills, Kenya. Interim Report (IR-09-056) International Institute for Applied Systems Analysis, Laxenburg, Austria. Maeda, E.E., Formaggio, A.R., Shimabukuro, Y.E., 2008. Impacts of land use and land cover changes on the sediment yield in Brazilian Legal Amazon drainage basin. Geoscience & Remote Sensing 45, 443453. Maeda, E.E., Formaggio, A.R., Shimabukuro, Y.E., Kaleita, A.L., 2009. Impacts of agricultural expansion on surface runoff: a case study of a river basin in the Brazilian Legal Amazon. International Journal of Geoinformatics 5 (3), 3341. Maeda, E.E., Pellikka, P.K.E., Siljander, M., Clark, B.J.F., 2010a. Potential impacts of agricultural expansion and climate change on soil erosion in the Eastern Arc Mountains of Kenya. Geomorphology, doi:10.1016/j.geomorph.2010.07.019. Maeda, E.E., Wiberg, D.A., Pellikka, P.K.E., 2010b. Estimating reference evapotranspiration using remote sensing and empirical models in a region with limited ground data availability in Kenya. Applied Geography. doi: 10.1016/j.apgeog.2010.05.011. Mertens, B., Lambin, E.F., 2000. Land-cover-change trajectories in Southern Cameroon. Annals of the Association of American Geographers 90 (3), 467494. Myers, N., Mittermeier, R.A., Mittermeier, C.G., Fonseca, G.A.B., Kent, J., 2000. Biodiversity hotspots for conservation priorities. Nature 403 (6772), 853858. Pellikka, P.K.E., Ltjnen, M., Siljander, M., Lens, L., 2009. Airborne remote sensing of spatiotemporal change (19552004) in indigenous and exotic forest cover in the Taita Hills, Kenya. International Journal of Applied Earth Observations and Geoinformation 11 (4), 221232. Republic of Kenya, 2001. The 1999 Population & Housing Census. Central Bureau of Statistics. Ministry of Planning and National Development, Kenya. Rudel, T.K., 2009. How do people transform landscapes? A sociological perspective on suburban sprawl and tropical deforestation. American Journal of Sociology 115 (1), 129154. Serneels, S., Lambin, E.F., 2001. Proximate causes of land-use change in Narok district, Kenya: a spatial statistical model. Agriculture, Ecosystems and Environment 85 (13), 6581. Soares-Filho, B.S., Pennachin, C.L., Cerqueira, G., 2002. DINAMICA a stochastic cellular automata model designed to simulate the landscape dynamics in an Amazonian colonization frontier. Ecological Modelling 154 (3), 217235. Soares-Filho, B.S., Garcia, R.A., Rodrigues, H.O., et al., 2007. Coupling socioeconomic and demographic dimensions to a spatial simulation model of deforestation for the Brazilian Amazon. Paper presented at the LBA-ECO 11th Science Team Meeting. Soares-Filho, B.S., Rodrigues, H.,Costa, W., 2009. Modeling Environmental Dynamics with Dinamica EGO. ISBN: 978-85-910119-0-2 <www.csr.ufmg.br/dinamica>. Soini, E., 2005. Livelihood capital, strategies and outcomes in the Taita Hills of Kenya. In: ICRAF Working Paper No. 8. World Agroforestry Centre, Nairobi, Kenya, p. 48. Thornton, P.K., Jones, P.G., Alagarswamy, G., Andresen, J., Herrero, M., 2010. Adapting to climate change: agricultural system and household impacts in East Africa. Agricultural Systems 103 (2), 7382. Turner, B.L., Lambin, E.F., Reenberg, A., 2007. The emergence of land change science for global environmental change and sustainability. Proceedings of the National Academy of Sciences 104, 2066620671. Veldkamp, T., Lambin, E.F., 2001. Predicting land-use change. Agriculture, Ecosystems & Environment 85 (13), 16 (Editorial, Special issue). Verburg, P.H., Soepboer, W., Limpiada, R., Espaldon, M.V.O., Sharifa, M.A., Veldkamp, A., 2002. Modelling the spatial dynamics of regional land use: the CLUE-S model. Environmental Management 30, 391405. Washington, C., Pijanowski, B., Campbell, D., Olson, J., Maitima, J., Musili, A., Kibaki, T., Kaburu, H., Hayombe, P., Owango, E., Irigia, B., Gichere, S., Mwangi, A., 2010. Using a role-playing game to inform the development of land-use models for the study of a complex socio-ecological system. Agricultural Systems 103 (3), 117126. Ximenes, A.C., Almeida, C.M., Amaral, S., et al., 2008. Dynamic deforestation modeling in the Amazon. Boletim de Cincias Geodsicas 14, 370391 (in Portuguese).

Almeida, C.M., Gleriani, J.M., Castejon, E.F., Soares-Filho, B.S., 2008. Using neural networks and cellular automata for modeling intra-urban land use dynamics. International Journal of Geographical Information Science 22 (9), 943963. Alves, D.S., 2002. Spacetime dynamics of deforestation in Brazilian Amazonia. International Journal of Remote Sensing 23 (14), 29032908. Batjes, N.H., Gicheru, P., 2004. Soil data derived from SOTER for studies of carbon stocks and change in Kenya (GEF-SOC Project; Version 1.0). Technical Report 2004/01. ISRIC World Soil Information, Wagningen. Bonham-Carter, G., 1994. Geographic Information Systems for Geoscientists: Modelling with GIS. Elsevier Science, Pergamon, Kidlington. Brink, A.B., Eva, H.D., 2009. Monitoring 25 years of land cover change dynamics in Africa: a sample based remote sensing approach. Applied Geography 29 (4), 501512. Brooks, T., Lens, L., De Meyer, M., Waiyaki, E., Wilder, C., 1998. Avian biogeography of the Taita Hills, Kenya. Journal of East African Natural History 87, 189194. Brouwer, C., Heibloem, M., 1986. Irrigation Water Management. Training Manual (FAO), vol. 3. Prov. ed. FAO, Rome, Italy (60p). Burgess, N.D., Butynski, T.M., Cordeiro, N.J., et al., 2007. The biological importance of the Eastern Arc Mountains of Tanzania and Kenya. Biological Conservation 134 (2), 209231. Burnett, C., Blaschke, T., 2003. A multi-scale segmentation/object relationship modelling methodology for landscape analysis. Ecological Modelling 168 (3), 233249. Clark, B.J.F., Pellikka, P.K.E., 2009. Landscape analysis using multiscale segmentation and object orientated classication. In: Rder, A., Hill, J. (Eds.), Recent Advances in Remote Sensing and Geoinformation Processing for Land Degradation Assessment. ISPRS Book Series in Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. 8. Taylor & Francis Group, pp. 323342. Di Gregorio, A., 2005. Land Cover Classication System (LCCS), Version 2: Classication Concepts and User Manual. FAO Environmental and Natural Resources Series 8, FAO, Rome, 208pp. Eva, H.D., Brink, A., Simonetti, D., 2006. Monitoring land cover dynamics in subSaharan Africa. In: A Pilot Study Using Earth Observing Satellite Data from 1975 and 2000. Institute for Environment and Sustainability, Joint Research Centre of the European Commission, p. 44. Ewers, R.M., Scharlemann, J.P.W., Balmford, A., Green, R.E., 2009. Do increases in agricultural yield spare land for nature? Global Change Biology 15 (7), 1716 1726. FAO Food and Agriculture Organization of the United Nations, 1997. Land and Water Development Division. Irrigation potential in Africa: A basin approach, FAO, 177p. FAO Food and Agriculture Organization of the United Nations, 2005. Land and Water Development Division. AQUASTAT Information System on Water and Agriculture: Online database. FAO, Rome. Fermont, A.M., van Asten, P.J.A., Giller, K.E., 2008. Increasing land pressure in East Africa: the changing role of cassava and consequences for sustainability of farming systems. Agriculture, Ecosystems and Environment 128 (4), 239250. Flynn, D.F., Gogol-Prokurat, M., Nogeire, T., Molinari, N., Richers, B.T., Lin, B.B., Simpson, N., Mayeld, M.M., DeClerck, F., 2009. Loss of functional diversity under land use intensication across multiple taxa. Ecology Letters 12 (1), 22 33. Hagen, A., 2003. Fuzzy set approach to assessing similarity of categorical maps. International Journal of Geographical Information Science 17 (3), 235249. Hall, J., Burgess, N., Lovett, J., Mbilinyi, B., Gereau, R.E., 2009. Conservation implications of deforestation across an elevational gradient in the Eastern Arc Mountains, Tanzania. Biological Conservation 142, 25102521. Hargreaves, G.H., Samani, Z.A., 1985. Reference crop evapotranspiration from temperature. Applied Engineering in Agriculture 1 (2), 9699. Himberg, N.L.M.A.Omoro., Pellikka, P., Luukkanen, O., 2009. The Benets and Constraints of Participation in Forest Management, vol. 187. The case of Taita Hills, Kenya. Fennia, pp. 6176. Hutchinson, M.F., 1995. Interpolating mean rainfall using thin plate smoothing splines. International Journal of Geographical Information Science 9, 305 403. Jackson, S.F., Gaston, K.J., 2008. Land use change and the dependence of national priority species on protected areas. Global Change Biology 14 (9), 21322138. Koomen, E., Rietveld, P., Nijs, T.de., 2008. Modelling land-use change for spatial planning support. The Annals of Regional Science 42 (1), 110.