Mapping land enclosures and vegetation cover changes in the surroundings of Kenya's Dadaab refugee camps with very high resolution satellite imagery

The immediate surroundings of refugee camps in drylands are among the areas exposed to highest pressure on natural resources including vegetation and soil. Understanding the dynamics of land fencing in these areas is critical for sustainable camp management and can help to improve the knowledge about land management in drylands in general. Very high resolution satellite imagery provides a means to observe such areas over time and to document land cover and use changes. This study uses satellite images to map fenced areas, which can be divided into pastoral enclosures and the so called ‘green belts’ (areas fenced for afforestation) around the Hagadera Camp in Dadaab (Kenya). It then analyses change dynamics between 2006 and 2013, a period where the refugee camp has been subject to high oscillations in camp population, due to a combination of conflicts and droughts in Somalia. The applied methodology allows detailed fence mapping and shows a large increase in fenced area (56%) over the 7‐year period. Although new pastoral enclosures expanded into more densely vegetated surroundings, land cover density inside already fenced areas either decreased or remained stable. Green belt areas grew at a similar rate (58%) but did not show evidence of greening over time and their longer term success is strongly dependent on maintenance. The settlement area did also expand remarkably in the same time (65%), and human and animal movements in the surroundings intensified with a negative impact on vegetation density. The study could not fully investigate the socio‐economic drivers and impacts linked to the rapid increase of enclosures, which are inextricably linked to evolutions in local agro‐pastoral practices. However, by documenting spatial and temporal dynamics of fenced areas, it adds new evidence to their increasing relevance in rangeland management, and opens the way to a number of hypotheses, stimulating the debate about long‐term ecological and socio‐economic impact.


| INTRODUCTION
Drylands cover around 40% of the global land area, host nearly one third of the human population and 50% of the livestock. To a large extent, they were traditionally used and managed by pastoralists through communal or common property rights-based land tenure systems (McDermott, Staal, Freeman, Herrero, & Van de Steeg, 2010;United Nations Environment Management Group, 2011). In sub-Saharan Africa (SSA), 40% of the total available land is utilised by 25 million pastoral and 240 million agro-pastoral farmers for livestock keeping, which is their primary source of livelihood (Neely, Bunning, & Wilkes, 2009). Many dryland areas have a history of overgrazing and degraded lands with low productivity, recurrent famines, resources conflicts, and economic and political marginalisation of pastoralist communities (Nori, Taylor, & Sensi, 2008;Opiyo, Mureithi, & Ngugi, 2011). Dryland areas also face the challenges posed by the combination of climate variability and change and rising demand for livestock products due to human population growth, urbanisation, and changing dietary preferences. This increased pressure has exposed drylands to greater levels of degradation in SSA, leading to the displacement of large numbers of people, their livestock, and a consequent impact on land management practices. In pastoral areas, and in particular, in the proximity of areas experiencing a reduction of nomadic livelihoods, the establishment of enclosures or fencing of land (Behnke, 1985;Beyene, 2009;Nyberg et al., 2015;Mureithi et al., 2015) is an increasingly common practice to protect and manage livestock, to increase and protect fodder production, to demarcate and claim land tenure as a follow up of sedentarisation, and for land rehabilitation among other reasons. Woodhouse (2003) even claims that the establishment of enclosures is a 'default' development when population pressure increases, which might link to the broader property rights theory (Ostrom, 1990). However, private enclosure systems in drylands are often disputed and can increase the risk of conflict for land. For example, in many of the pastoral livelihoods, which according to Catley 1999, can be associated to one large 'Somali pastoral ecosystem' (Figure 1), it is claimed that enclosures fragment the land, hinder FIGURE 1 Extent of the Somali pastoral ecosystem according to Catley, 1999 (red contour line) and grazing zones (light green: low density grassland; dark green: high density grassland derived from existing land cover/land use datasets) according to Vancutsem Conflicts, political instability, or natural disasters are additional drivers for displacing people and reducing the mobility of nomadic pastoralists by pushing them towards mixed agro-pastoral livelihoods and increasing sedentarisation. Most African countries host refugees from neighbouring areas or internally displaced populations in refugee camps. These are generally large temporary settlements of elementary infrastructure causing pressure on the host community and on natural resources (UNHCR, n.d.) and it can be expected that pastoral intensification strategies such as fencing occur at an accelerated rate in such environments as compared with areas with no refugee influx. The environmental impact of refugee camps in semi-arid regions is at the centre of research and impact monitoring (Lodhi, Echavarria, & C, 1998;Berry, 2008;Kariuki, Machua, Luvanda, & Kigomo, 2008, Braun & Hochschild 2015, Braun, Lang, & Hochschild, 2016, and a variety of mitigation strategies and interventions have been proposed in particular by international agencies and non-governmental organisations (NGOs) (Hoerz, Chege, Jacobsen, Kimani, & Nyandiga, 1999;Lahn & Grafham, 2015). The largest refugee complex in the world is located in a semi-arid area of north eastern Kenya called Dadaab (with over 350,000 inhabitants in 2014). The Dadaab Camps in Kenya are of particular interest in understanding environmental impact of mass displacements, due to their long history (since the early nineties), their size, and the rich bibliography on sustainability questions (Beaudou, Cambrézy, & Souris, 1999;Bizzarri, 2010;Enghoff et al., 2010;Lindley, 2011). The originary pastoral system of Dadaab's host population has adapted to close interconnections with the Somali refugees with a large portion of the host population converting from nomadic pastoralists to livestock producers supplying meat and milk products to the camps. (Kamau & Fox, 2013). Pastoralists from the local community also tend to aggregate around refugee camps such as Dadaab due to the better access to social services and to drinking water (De Montclos & Kagwanja, 2000). Such changes in local pastoral resource management carries relevant consequences on the overall land cover and resource access, and grazing dynamics around the camps are complex and often conflictual. Finally, since the opening of the camp, the dependency of the refugees on wood for cooking and building has exacerbated the conflicts for natural resources in an already fragile ecosystem.
For decades, satellite-based earth observation has provided means to monitor land cover dynamics and status of natural resources over relatively large areas. Previous remote sensing studies on refugee camps in the area have focused primarily on the estimation of camp population through the identification of dwellings and used different data sources in terms of resolution and sensors (Füreder, Hölbling, Tiede, Zeil, & Lang, 2012, Gorsevski, Kasischke, Dempewolf, Loboda, & Grossmann, 2012, Baker, Card, & Raymond, 2013. Despite the high interest on evidence and trends in environmental impact, analyses in this sense carried out by several teams and mainly based on Landsat data, have not been very conclusive due to limited image availability and insufficient spatial resolution for identifying vegetation changes linked to firewood collection and grazing (Enghoff et al., 2010). To overcome these limitations, other studies have used very high resolution imagery (Hagenlocher, Lang, & Tiede, 2012;Johannessen et al., 2001), whereas another approach proposed by Braun et al. in    Soils are classified as red sand, reddish sandy soils, sandy, and loamy-clayey soils ranging from deep soils in seasonally flooded areas to very shallow and superficial in more arid zones Famine Early Warning Systems Network, 2011).  (Enghoff et al., 2010). As opposed to traditional pastoralism in the region, the herd mobility is strongly reduced with few herds grazing further away than daily movements and the loss of mobility is partially compensated by private fencing of grazing land (Enghoff et al., 2010). Part of the herd is kept around the settlements and fed with relief fodder, thorn-fenced enclosures are used for grazing of familyowned livestock and fodder production especially during the dry season ( Figure 2). Other fenced plots common in the immediate camp proximity and generally with more regular shape than pastoral enclosures are afforestation areas called 'green belts' , whose construction was supported since the 90's mainly by NGOs, for increasing soil cover and firewood availability (Hoerz et al., 1999;UNHCR, 2015).
Both pastoral enclosures and green belt fences are usually made of thorny branches of Acacia and Commiphora spp. as well as invasive tree species such as Prosopis juliflora (Bizzarri, 2010;Bradford 2016 personal communication). For green belts, live hedges are also common ).

| Remote sensing datasets
The selection of images relies on previous activations of the Emer- Radiance values were obtained first, using absolute radiometric calibration coefficients. Subsequently, top of atmosphere (TOA) spectral reflectance was obtained normalising the TOA radiance with the solar incoming irradiance, and a solar angle correction (European Union, 2015). Pan-sharpening aims at obtaining high resolution multispectral imagery integrating low-resolution multispectral information with the high resolution panchromatic band. The QuickBird scene was acquired already pan-sharpened where the UNB (University of New Brunswick) fusion algorithm (Zhang, 2004) was applied.

| Detection of pastoral enclosures and vegetation cover (object oriented image analysis)
In this study, eCognition software (Trimble Navigation Ltd.) was used to perform the object-based image analysis (OBIA; Baatz & SCHÄPE, 2000). In particular, the multiresolution segmentation algorithm was applied to the three acquired images. Over the last twenty years, object-based image analysis has gained rapid reputation in geospatial applications over per-pixel and subpixel analyses, with increased spatial resolution of images being one of the preconditions for detecting objects from groups of pixels significantly smaller than the object size. The group of pixels, also called segments, are regions defined by one or more criteria of homogeneity in one or more dimensions (Blaschke, 2010). Particularly, in the exploitation of multitemporal high resolution imageries, an OBIA approach is generally preferable to pixel-based methods for the detection of objects that can be classified based on texture, context (using different thematic layers), and geometry. This has been applied in literature for the detection of linear features such as fences, mapping land cover, and in particular land restoration interventions (Chepkochei, 2011;Spiekermann & Brandt, 2015;Fava, Pulighe, & Monteiro, 2016).
The approach was used for (a) a first automatic detection of pastoral enclosures and green belt fences and (b) for the classification of vegetation cover. The segmentation settings and the following rule parameters depended on the data type. As the optical properties of the images change due to sensor characteristics and time of acquisition, the reflectance values of the three images were not directly comparable and common thresholds valid for all images could not be used. This segmentation algorithm is based on the scale parameter, used to control heterogeneity and size of objects, and shape and compactness parameters that control relationships between spectral and spatial homogeneity. The scale, shape, and compactness parameters of the multiresolution segmentation were selected on empirical basis until the segments were well delineating visuallyobserved boundaries of enclosures.
The analysis started with WorldView-2 data, which was prioritised among the set of images for its better spectral and spatial characteristics (Table 1)   NDVI formula: All polygons not classified as fences or vegetation cover were assigned to the non-vegetated class.

| Validation of detection of pastoral enclosures and vegetation cover
In the absence of ground based measurements, the OBIA results for fences and vegetation cover were verified through comparison with visual photointerpretation on the same images conducted by an independent operator. Manual digitisation of features (enclosure fences and vegetation cover) was executed within 50 randomly selected square plots (30 × 30 m) distributed over the study area, accounting for roughly 10% of the area analysed, and used as reference data.
The size of the squares was selected by taking into account that most fenced areas are clearly smaller than 1 ha. Table 2

| Fenced area mapping
The first result of the image analysis is the semi-automatic detection of fences for the surveyed area and for the three inquired years.
An example of delineation of fenced areas is offered in Figure 4,   Figure 4b shows an example of classification of vegetation cover present both inside and outside the fenced areas.

| Main spatial and temporal dynamics
The monitoring of fenced areas over time shows a clear increase in terms of total fenced area during the observation period (Table 3).
The trends suggest a progressive expansion of land fencing through    (Table 4) The different clearing practices in the western and eastern parts of the study area could correspond to different use patterns by different population groups, where in the second case, there is a more conservative use of woody vegetation inside the enclosures. In any case, areas in between enclosures and corridors are exposed to higher livestock density and transit and show a strong decrease of vegetation density (Figure 5b). For green belts, a negative mean trend in vegetation density can be noticed by looking at all green belts for each date.
Although considering only the areas with green belt use for all three dates, cover density remains stable (Table 4). Both observations suggest that long-term success of green belts in terms of increasing biomass on degraded areas is challenging and requires intensive maintenance. The increasing pressure on resources and the public nature of green belts (as opposed to the generally private management of pastoral enclosures) might also be responsible for the observed absence of biomass increase in green belts, which is also confirmed by the find-    indicates the difficulties in the recovery of degraded areas and could possibly also be due to dwindling external support to afforestation around the camps.
The study also confirms the increase of settlement area over time as documented by other sources but more interestingly adds information on woody vegetation density inside the settlement area. The results seem to indicate that although vegetation density slightly decreases over the initial settlement area, it rapidly increases for more recently built areas, possibly indicating an increased use of fast growing species and suggesting a stronger attention towards tree planting inside the settlement area.
In general, all results reflect the increasing demand for fodder