Modeling pastoralist movement in response to environmental variables and conflict in Somaliland: Combining agent-based modeling and geospatial data

Pastoralism is widely practiced in arid and semi-arid lands and is the primary means of livelihood for approximately 268 million people across Africa. Critical environmental, interpersonal, and transactional variables such as vegetation and water availability, conflict and ethnic tensions, and private/public land delineation influence the movements of these populations across space and time. The challenges of climate change and conflict are widely observed by nomadic pastoralists in Somalia, particularly in the regions of Somaliland and Puntland, where resources are scarce, natural disasters are increasingly common, a protracted conflict has plagued communities for decades, and over 65% of the population rely on pastoralism as a primary livelihood. Bereft of necessary, real-time data, researchers and programmatic personnel often turn to post hoc analysis to create an understanding of the interaction between climate, conflict, and migration, and design programs to address the needs of nomadic pastoralists and those that drop out of pastoralism in search for alternate livelihoods. By designing an Agent-Based Model (ABM) that simulates the movement of nomadic pastoralists based on aggregated, typologically-diverse, historical data of environmental, interpersonal, and transactional variables in Somaliland and Puntland between 2008 and 2018, this study intends to identify how pastoralists respond to complex, changing environments over time. The subsequent application of spatial analysis, through Choropleth maps, Kernel Density Mapping and Standard Deviational Ellipses, characterizes the resultant pastoralist population densities in response to these spatio-temporal variables. Outcomes of these analyses demonstrate a large scale spatio-temporal trend of pastoralists migrating to the southeast of the study area with high density areas manifesting in the south of Nugaal, the northwest corner of Sool, and along the Ethiopian border. While minimal inter-seasonal variability is seen, multiple analyses support the consolidation of pastoralists to specifically favorable regions. While this ABM does produce compelling associations between pastoralist movements and terrestrial and conflict variables, it relies heavily on assumptions and incomplete data and is not necessarily representative of on-the-ground realities. Given the paucity of data regarding pastoralist decision-making and migration, validation remains challenging with current methods based on heuristics and descriptions in literature.


Background
Pastoralism is widely practiced in arid and semi-arid lands (ASALs) and is the primary means of livelihood for approximately 268 million people across Africa (FAO, 2018). Pastoral mobility is largely driven by the availability and quality of fodder and water to maintain livestock herds, influencing the spatial and temporal variability of pastoral migration patterns across landscapes (Pas, 2018;Sakamoto, 2016). While the non-sedentary lifestyle of pastoralism can be considered an effective adaptation technique to environmental changes, dependence on natural resources contributes to the risk-averse nature of pastoralism (FAO, 2018;Pas, 2016). Unpredictable climatic changes contribute to the increase in the severity and frequency of natural hazards, irregular rainfall patterns, extreme temperatures, and changing land cover, affecting the availability of natural resources required to support livestock herds (Avis & Herbert, 2016;Onyango, 2016;Sakamoto, 2016). These environmental effects are compounded by ongoing land degradation, land privatization, conflict, and numerous other factors, weakening pastoral systems (FAO, 2018;Onyango, 2016).
The challenges experienced by pastoralist communities are unevenly felt across the continent but are particularly pertinent in Somalia, where approximately 65% of the population relies on pastoralism as a primary source of livelihood (Carr-Hill, R. A., & Ondijo, D. 2012). In the last decade alone, Somalia has experienced numerous devastating environmental shocks over a short period of time. Between 2010 and 2012, a catastrophic drought with subsequent food insecurity and famine that affected a large part of the Horn of Africa affected 13 million people, many of whom were pastoralists (Slim, 2012). Of all the countries affected by the drought, Somalia was arguably the hardest hit in the region, which was aggravated by the inability to provide timely assistance due to instability, conflict, and lack of humanitarian coordination (ibid). Shortly after this period, Somalia experienced another drought between 2016 and 2017 that was followed by lower-than-normal rainfall in the following years, resulting in growing numbers of internally displaced populations (IDPs). The Food and Agriculture Organization (FAO) (2018) noted that it takes approximately five years for a livestock-dependent household to fully recover from the effects of severe drought. The severe droughts in Somalia in 2010 -2012 and 2016 -2017 then theoretically did not not give pastoral households time to fully recover. A variety of adaptation techniques are used to protect livestock-related livelihoods in times of drought or conflict, including forming agreements and alliances with members of the community, sharing of necessary resources, and diversification of livelihoods (Shaughnessy, 2018). At times, however, the expansion of private land and environmental degradation results in resource conflicts and occasional livestock death (Onyango, 2016).
The impact of environmental hazards is compounded by ongoing conflict and political instability that has troubled the country for decades. Violence has manifest itself in Somalia in numerous ways. Ongoing armed conflict between the Somali militia, the African Union Mission in Somalia, and non-state actors has led to widespread insecurity and displacement, particularly in Southern Somalia (IDMC, 2018). In the northern part of the country, communal violence prompted by clan/ethnic differences and the quest for political autonomy has been an ongoing problem in the contested regions of Puntland and Somaliland. In this region, Somaliland is working to gain international recognition for being an independent country while quarreling with neighboring Puntland over disputed land (Avis and Herbert, 2016). In 2018 alone, an estimated 578,000 people were displaced due to conflict, while an additional 548,000 people were displaced by regional disasters in Somalia (IDMC, 2018). To date, the total number of IDPs in Somalia is thought to be more than two million people, and this does not account for the many refugees who have fled the country altogether (ibid). Numerous organizations have documented conflict and disaster as the primary drivers of displacement for Somalis, including pastoralists (IDMC, 2018;UNHCR PRMN, 2019).
Pastoral migration, whether voluntary or forced, is a complex phenomenon that scientists continue to grapple with. While the drivers of migrations are broadly understood, less is known about how environmental and conflict variables affect patterns of movement across space and time in the past and how they may continue to evolve in the future. To the knowledge of the authors, there is very limited public information available about pastoral populations in Somalia or the self-declared autonomous region of Somaliland. Without information that captures the dynamic nature of pastoral populations, it is extremely challenging to understand the evolving patterns of movement in response to environmental changes and ongoing conflict on a regional scale. To date, several studies use computer models as a means to explore pastoral migration through virtual environments, which will be referred to in the following paragraph. Computer models enable the simulation of scenarios in which the relationships of variables are tested and analyzed over time and/or space providing insight into the observed phenomenon.
Agent-based models (ABM) are a subset of computer modeling techniques that are increasingly being utilized to examine the stochastic realities of human migration in response to environmental and conflict variables . ABMs can simulate the actions of agents, in this case pastoralists, based on the interactions among agents and their surrounding environment. The bottom-up nature of ABMs enables the capture of granular patterns of movement that can be summarized to a systems-wide level. While some models aim to gain a holistic understanding of the adaptive behaviors of pastoralists in response to their environment (Sakamoto, 2016), others consider the possible effects of climate and/or conflict on pastoral movement (Ginetti et al., 2015, Smith, Kniveton, Wood, & Black, 2011, as well as the interplay between pastoral movement, land privatization , Lesorogol & Boone, 2016, and disease transmission (Xiao, Cai, Moritz, Garabed, & Pomeroy, 2015). Computer models rely on a combination of data sources, both qualitative and quantitative evidence, predictive datasets, and/or informed assumptions to generate simulation scenarios. The extensive spatial and temporal variability of pastoral movement and the factors that trigger it warrants the implementation of a simulation model that captures its nuances and the wide-ranging variability.
The integration of geospatial data and agent-based modeling to study pastoral migration is less explored in literature. The existing models and findings, however, have contributed substantially to an ever-growing body of literature on the subject. Sakamoto (2016) integrates low-resolution multi-temporal satellite imagery analysis and agent-based modeling to study pastoral access to resources in dryland vegetation in northeastern Nigeria while the Center for Social Complexity and Department of Computational Social Science at George Mason University has examined the intersection of GIS and agent-based modeling through the development of a range of models on the ABM platform MASON, including the HerderLand, AfriLand, RiftLand, and RebeLand models. The models consider a range of scenarios in Eastern Africa that examine resource contention, the effect of environmental changes, availability of watering holes, and the effect of private land on pastoral movement on multiple scales . To the knowledge of the authors however, no ABM has been designed for or applied to Somaliland while southern Somalia is well studied.
The purpose of this study is to examine the relationship between environmental change, conflict, and pastoral movement in Somaliland between 2008 to 2018 through agent-based modeling. The research generates synthetic movement patterns for nomadic pastoralists in the region, which are influenced by a series of environmental, interpersonal, and transactional variables. Evidently, computer modeling presents high levels of uncertainty but also has immense potential to improve humanitarian preparedness and response. To improve modeling capacities, it is necessary to continue developing the input data and methodology to identify the successes and limitations of the model. This paper adds to a growing body of literature that considers the use of predictive modeling and geospatial analysis to address issues in the humanitarian sector where climate and conflict variables have significant impact on population movement.

Setting
This agent-based model simulates the movement of nomadic pastoralists in response to conflict and environmental variability in northern Somalia, specifically in the regions of Somaliland and Puntland, between

Data collection
Data utilized for this study (Table 1) were obtained from diverse sources, are of disparate typologies, and were quantified, normalized, and utilized in various ways to generate agents, assign attributes to those agents, and develop an environmental model. The following is a description of those data sources, manipulations, and the mechanism for normalization.

Manipulation and Normalization
The model environment is comprised of terrestrial variables including slope, surface water, points of artificial water sources, and vegetation rigor; interpersonal components including ethnic boundaries and locations of conflict; and a transactional variable that delineates private and public land. All data values were rasterized and aggregated to a 1 km² grid. Variables were then normalized to range values between 0-1, utilizing the equation, below.
Wherein V is the variable in consideration.
Slope was calculated from remotely-sensed elevation data obtained from DIVA GIS using the Slope Spatial Analyst in ArcMap 10.6.1. The presence or absence of surface water was ascribed a binary score of 1/0 (presence=1) based upon data downloaded through Google Earth Engine. Artificial water sources, such as wells and boreholes, were gleaned from a survey-based, geocoded dataset provided by SWALIM (W. Stephen, personal communication, April 18 2019), with cells containing more than 1 water source having higher values than pixels that contained only one or none.
The vegetation score was calculated on a raster grid containing seasonal median pixel values. To calculate the median pixel values, all available MODIS satellite data was seasonally aggregated, and the median pixel value was calculated. Following this calculation, the Soil-Adjusted Vegetation Index (SAVI), equation below, was applied to those median pixel values for each season in the study period to create a map of proxied vegetation availability.
This index has proven more appropriate for the presence of vegetation in arid regions (Vani and Mandla, 2017).
In this equation, NIR stands for near-infrared values and L equals the canopy background adjustment factor, which was set at 0.5 to minimize soil brightness variations and eliminate the need for further calibrations around soil-type (cite).
Ethnic boundaries obtained from a static map provided by the Kenya Somalia Consortium were digitized, and rasterized, with buffers applied to the borders at 10 and 20 meters, with descending weighted values of 0.5 and 0.25, respectively. Conflict point data was gleaned from the ACLED database and rasterized. Grid cells that had higher incidence of conflict were attributed higher values, and a temporal lag was assigned within the model, with conflict occurring the previous season having half the effect on environmental favorability the following season.
In the absence of reliable data, the delineation of private and public lands was determined through expert opinion (HHI-UNICEF workshop, Nairobi, Kenya, June 3-4, 2019). Eighty percent of land in Somaliland was estimated to be publicly-owned, and the distribution of private land was assumed to be land which extends approximately 15 kilometers from the centroid of large metropolitan centers and five kilometers from smaller settlement centers.
To adhere to the 80% threshold reported, the private land boundaries around cities and towns were modified to 14 and 4 kilometers, respectively.

Overview of the Computational Model
This agent-based model is developed in RePast (Recursive Porous Agent Simulation Toolkit) Simphony 2.6 using a Java-based simulation environment. Repast is a leading, open-source ABM development toolkit specifically designed for social science applications and has been well regarded in comparison to other ABM platforms (Railsback 2006). The Repast development framework provides all the basic functionality required to support the execution of an ABM, including scheduling mechanisms and diverse modeling functions. The established framework enables researchers to add components to customize the model to fit their needs, allowing the environment to be modified and incorporate a range of dynamic variables.
This agent-based model includes two entities: 1) agents, each representing a single nomadic pastoralist household unit, and 2) the physical environment, which is a geospatial landscape composed of both dynamic and static attributes. At the start of the simulation each generated agent is assigned attributes, including their geographic position at the start of the simulation, the name of the administrative unit they fall within, their ethnicity, and clan association. The number of agents generated per administrative unit was informed by a Population Estimation Survey conducted in 2014 (UNFPA, 2014). Within any given administrative unit, the agent start position was randomly generated with two constraints: 1) the agent must not be in unsuitable landscapes including water bodies and areas of bare soil (i.e. sand), and 2) the agent must be located 14 kilometers or more outside a major city and four kilometers or more outside of smaller settlements, i.e. on 'public land'. For every simulation run, the agent start position remains identical.
The gridded physical environment is composed of 1 km² grid cells and has a spatial extent of approximately 490,000 km², which covers the administrative regions in Somaliland and Puntland. The environment was designed to include eight variables, grouped into three thematic components: terrestrial variables, interpersonal variables, and transactional variables ( Table 2). The environment variables are categorized to be either pull (attractors) or push (detractors) factors for nomadic pastoralists whose patterns of movement are influenced by the availability of water and suitable grazing land to support their herd. The factors were identified and included based on literature, workshops, and/or discussions with regional experts. Intuitively, the presence of vegetation (as proxied by SAVI) and availability of water are considered to be attractors, while steeper terrestrial slope, proximity to conflict or potential ethnic tension (as proxied by proximity to ethnic borders) are designated detractors.
Variables, such as surface water availability, conflicts, and vegetation cover, are subject to seasonal changes. In this model, four seasons exist per annum. The dry season from December to March is locally referred to as Jilaal, which is followed by the long rainy season, Gu, from April to June. The dry season that follows, Hagaa, spans from July to September while the short rainy season that takes place between October and November is known as Deyr. Regarding water availability, during the dry seasons, the surface water layer is disabled, as nomadic pastoralists tend to rely on man-made water sources such as wells and boreholes during these times and when surface water is extremely sparse (HHI-UNICEF workshop, Nairobi, Kenya, June 3-4, 2019). The conflict environmental surface changes seasonally during the model's timeline, with conflict point data being aggregated to each season and changing at the end of the ascribed three-month period. Vegetation availability scores also change seasonally, based upon the the medial pixel value of the imagery available for each defined season, as described above, with applied SAVI scores. The inclusion of these real, longitudinal data creates a reflection of not only the seasonal changes associated with wet and dry seasons, but also the changing environment due to climate and human variables over the course of the simulation.
The favorability score for each land parcel is artificially created through the additive equation:

Simulating Movement
Agents move throughout the simulated geographic environment based on the fundamental premise that nomadic pastoralists rely on livestock to support their livelihoods and are therefore heavily dependent upon access to water and vegetation (Figure 2). At every time tick, i.e. a month, the agent searches for a cell in the environmental grid with the highest favorability score within a search radius (referred to as scouting range) of its surrounding environment. This scouting range was a random distance generated between a 15-and 30-kilometer radius, based on expert consensus regarding the monthly mobility capacity of nomadic pastoralists. Once the most favorable cell has been identified, the agent moves to this location and determines whether it is located on private or public land. If the cell is public land, they are free to move to that location without any further delay and have the option to stay there for the duration of the season (between one and three months), provided this location continues to have the best favorability score. Once the season changes, the agent is required to seek out further land. This latter constraint is an effort to model resource depletion of that grid given the grazing requirements of a pastoralist herd.
However, if the cell happens to be on private land, the agent must negotiate a land-sharing deal with the local landowners. If the agent successfully makes a deal with the landowner, they are able to remain at that location for the duration of the season. However, if the agent is unable to obtain access, they are required to move to another grid cell within the same time tick, consider whether the new grid location is private or public land, and repeat the process described above. This secondary selection of a cell is determined by the next best score in the gridded environment.
If the agent is unable to make a deal on three separate occasions in any given season, the agent state switches from pastoralist to IDP, at which point the agent exits the simulation or 'drops out'. This is done with the assumption that the lack of access to grazing land leads to the death of livestock, and the pastoralist agent is required to seek alternative livelihoods and will potentially drop out of a purely pastoralist lifestyle. protocol, which has been amended to include details regarding human decision making (ODD+D) (Müller et al, 2013), can be found in Appendix 1. Given the constraints of computational power, completing the simulation for all 225,767 pastoralist agents was not possible. Thus, a simple random sampling of 10% of the population in each administrative region was conducted, and the agent-based model was executed with this subset of the generated agents (Table 3). as artificial water sources, vegetation cover and conflict allow for associations regarding the interplay between changes in resource availability, conflict dynamics and pastoralist movements.
All spatial data was projected in WGS 1894 UTM Zone 38N and all analysis was performed utilizing the ArcGIS 10.7 platform by ESRI.

Population Counts and Differences
Population counts were created by identifying the location of each pastoralist agent at the aforementioned time periods and summed to generate aggregate population counts across the study region. These data were then spatially joined with administrative districts within Somaliland and Puntland. Choropleth maps delineating population counts per administrative district were created excluding those agents that lay outside of boundary lines. Classification was done using Jenks Natural Breaks without justification between the two time periods, as justification would obscure the significant difference in population counts in each district. Differences between the January 2008 population counts and January 2018 and October 2018 counts were calculated to create a change layer. These calculations were undertaken to explore the change in population counts between 2008 and 2018, both taking into account season (January 2008 verses January 2018) and the beginning and end of the simulation. Choropleth maps were subsequently created, and justified Jenks Natural Breaks were applied to create population change classifications.

Kernel Density Maps
Only a 10% random sampling of each of the regional populations were utilized in the ABM simulation, therefore there are a great number of agents (i.e. pastoralists) that go unrepresented when population distribution is mapped. Through the production of a kernel density surface, this underrepresentation is addressed by creating a predictive distribution surface map of pastoralist populations given known spatial inputs. Population density is calculated using a quadratic formula, seen below, with the highest weight ascribed to the known point location and tapering to zero at the edges of the search radius with the predicted population at any given cell in the output raster map being an accumulation of the values for each of the calculated surfaces.
Wherein: i=1,...,n are the input points; popi is the population field value of point I, and disti is the distance between point i and the (x,y) location (ESRI, n.d.). The cell size (x,y) is 0.0089831528, 0.0089831528 decimal degrees, which is equivalent to a 1 km². Given the UTM projection of the environmental basemap, the analysis was parameterized to a planar method of measurement. By creating these kernel density maps of the pastoralist population over the four seasons of 2008 and 2018, a spatio-visual interpretation of how pastoralist population has evolved over the course of ten years and its seasonal permutations may be developed.

Standard deviational ellipse
An unweighted, standard deviational ellipse analysis was undertaken to characterize the spatial distribution - While SDEs have been utilized in the past to explore the relationships between environment and criminal activity (Kent andLeitner, 2007 andChainey et al., 2008), to characterize racial segregation (Wong, 1998), and to assist in outbreak surveillance (Eryando et al., 2012), the application of SDEs to ABM outputs, specifically in the context of pastoralist migration, is novel.

Results
The results of the ABM yielded the location of each pastoralist agent at every time step, along with its scouting range, and the favorability score of the grid in which it inhabited. The following table (Table 4)   All districts portray a decline in pastoralist population counts, resultant of pastoralist agents 'dropping out' of purely pastoralist cycles (Figure 3). However, certain districts have significantly greater declines in population than others, with Nugaal demonstrating only a decrease of 125 pastoralists, and Woqooyi Galbeed having a decline of almost 3,500. Also notable is the difference in population counts between administrative districts over time. In January 2008, the difference in population count between the most (Sanaag) and least (Bari) populated districts was 2,196, as opposed to October 2018 in which the difference between the most (Nugaal) and least (Woqooyi Galbeed) populated districts was 2,429. In general, the greatest decline in population per district was noted in the north-east of the study region with very little change when accounting for seasonal variability between January and October of 2018.  Table 5. In January 2008, the density across the study area ranged from 0 to 0.55 agent per square-kilometer with a standard deviation of 0.71. The map in Figure 4A shows that the agent density is fairly dispersed, but the highest densities are found in Awdal, the westernmost administrative region, while, Bari, the easternmost administrative region has the lowest estimate density. It is important to consider that January 2008 is the first month of the simulation, directly before which the agents were randomly generated in each administrative region, so the spatial dispersion is likely a remnant of the agent generation. Four simulated months later, in the following May, the maximum population density increased to 0.96 agents per square kilometer, without a large change in standard deviation. Certain high-density areas noted in January 2008, specifically in the south of Nugaal and the northwest corner of Sool, persist as relatively dense areas throughout the remainder of the simulation ( Figure 4A). By August 2008, maximum density continues to increase, and additional clusters begin to form in Togdheer along the Ethiopian border as well as near the coast in Sanaag.
Towards the end of 2008, high-and low-density locations appear to have reached homeostasis, with both large scale spatial patterns and standard deviations of population density remaining similar.
By January 2018, the mean density of pastoralist agents has approximately halved (0.024 from 0.043) resultant of pastoralist 'drop out'. Areas of highest population density are now centralized predominantly within four districts: Sanaag, Togdheer, Sool and Nugaal ( Figure 4B). Certain early areas of pastoralist accumulation, such as that in the south of Sool and along the Ethiopian border persist as the most densely populated. In general, as the simulation progresses, pastoralists appear to tend towards clustering, with maximum densities at 1.5 and 1.4 pastoralists per 1km 2 in the final two seasons. These areas of high density seem to shift eastwards over the eleven-year period, or inversely, the density of pastoralists in the western districts of Awdal and Woqooyi Galbeed declined significantly.  In 2018, the area of the SDE ranged from 501,606 km² to 505,096 km² (Table 7)   Existing studies that use ABM to simulate pastoral mobility have typically used the availability of water and pasture as drivers of migration. Water sources can include surface water, artificial sources, or a range of other water sources. Some studies have even incorporated complex hydrological modeling and have incorporated information on rainfall. The studies that consider vegetation and pasture availability most commonly incorporate the Normalized Difference Vegetation Index (NDVI), which captures vegetation vigor in any given area. While NDVI is a widely used vegetation index in academic studies because it is highly correlated with leaf-area index and biomass, which are attributes of measuring greenness, it is acknowledged that soil reflectance interferes with NDVI calculations, not accurately capturing the vegetative ground conditions. Due to the known limitations of NDVI, this research instead uses SAVI to more accurately capture the vegetative cover of Somaliland and Puntland, which is sparsely vegetated. Beyond water and vegetation, some have incorporated historical conflict while others actually identify areas of possible contention instead.
Typically, studies have considered the incorporation of variables one by one, where the agent makes a decision based on a decision tree. This study however, normalized all the variables to have values between 0 and 1, after which they were aggregated to a single favorability score. The favorability score is considered in relation to the score of the surrounding grid cells prior to the agents physically moving to another cell. While there is one favorability score, the data layers can be weighted in a variety of combinations to understand the effects of individual variables rather than the collective whole. Following the generation of ABM results, geospatial analysis techniques were used to analyze the data. The integration of geostatistics and ABM is for this particular application is not commonly documented, so the use of these methods and the generation of the results contribute valuable information to the academic sphere and unifies the multiple drivers of migration through a spatial and geographic lens.
In building the conceptual model, we invoked a number of disciplines for our assumptions. Social science, climatology, conflict studies and spatial epidemiology all contributed to its development. Such a multi-disciplinary approach is required to understand real-world complex ecosystems and for us, to ideally produce data that would allow us to create inferences between change in terrestrial, interpersonal and transactional variables and population movement and density.

Limitations and challenges
The model in theory includes a large number of agents making monthly decisions based on the influence of eight variables over a 10-years period resulting in hundreds of thousands of potential data points, requiring a vast amount of computing power. For us, this is limiting for two reasons: first, it required us to use a sample of agents, introducing some degree of statistical error that we could not determine after running the model; second, we could only look at beginning and end point outcomes (2008 and 2018) and not iterative trends (points in time) over the study time period that could be linked to specific climate or conflict events. Arguably then, to get the benefit of a full simulation, ABMs must have the computing power available and this fact could significantly limit its field applicability.
The model also makes a number of assumptions about pastoralist behavior that while duly gleaned from informed local non-governmental service providers, have not been scientifically validated or have an evidencebased understanding of real-time migration beyond anecdotal observations. As a result, the variables are weighted based on these assumptions. As such migration patterns cannot be fully validated. Without clear behavioral and decision-making benchmarks, this makes the model output challenging to validate; rather, validation is done visually using heuristics from the social science domain literature and expert opinion. That said, the benefit of the agent-based model presented here provides some insight into pastoralist behavior and further invites theory and generates hypotheses for deeper study.

Conclusion
The agent-based model, introduced here, is a useful tool to understand the behavior of individuals in a spatial and temporal context. Short of tracking individuals prospectively and manually querying their decision-making, the ABM, coupled with an ethnographic understanding of the factors and drivers of livelihood decisions, can provide a dynamic view of population ecosystems. In the case of nomadic pastoralists in the Horn of Africa where the history of seasonal migration lends itself to a simulated model, the ABM affords the ability to study the complexities of individual attributes in an aggregated and collective fashion with the added benefit of exploring layers of geographic and sociological variables.
In the humanitarian sphere, two critical independent and interdependent variables have the potential to trigger migration: climate-related environmental conditions, of which there are several, and conflict. Somaliland, with its long history of ethnic conflict over land use, its exposure to drought and water scarcity, and a basal level of livelihood migration amongst its nomadic pastoralists, offers a crucible in which to explore how these variables interplay in a model, assuming non-linearity and the need for non-parametric approaches to spatial analysis.
The use of ABM to understand pastoral migration in Somaliland and the ways it which it changes in response to environmental variability and conflict highlighted several important findings. This model was developed with limited data sources and capacity for large scale analysis, and still demonstrated the potential value added of the methodology in understanding pastoralist migration. In an attempt to evolve this model into one that is better representative of reality, the researchers intend to further refine the model, validating behavioural assumptions and including information and decisional pathways regarding economic markets, the effects of livestock disease, pastoral adaptation techniques, and resource depletion to develop a more holistic view of the pastoralist system in Somaliland and Puntland.

1.1.1.
What is the purpose of this study?
The purpose of this study is to understand the impact of seasonal, environmental changes and conflict on the movement of nomadic pastoralists in Somaliland between 2008 and 2018.

1.1.2.
For whom is the model designed?
This model is designed for humanitarian researchers studying the interaction between conflict, environment, and migration and those studying human migration more generally. This case study also adds to a growing body of literature addressing the value of computer models in the humanitarian sector.

Entities, state variables, and scales
1.2.1. What kind of entities are in the model?
Each modeled agent in the simulation represents a nomadic pastoralist household. The agent remains in a constant state unless the impact of exogenous variables is so severe that the agent state changes from nomadic pastoralist to internally displaced person (IDP).
The second entity in the model is the spatial environment. The environment is not only a modeled entity but also drives the behavior and movement of the agents.

1.2.2.
By what attributes (i.e. state variables and parameters) are these entities characterized?
The nomadic pastoralists' clan and ethnicity are identified according to the region in which the agent is generated. The spatial coordinates of the agent's position are recorded at each timestep in addition to the distance traveled to that location. Lastly, the grid cell value corresponding to the agent's position is also recorded as an attribute.
If the agent state changes from nomadic pastoralist to IDP, the agent effectively exits the simulation and no additional attributes are collected.
The entity representing the spatial environment is composed of aggregated variables in grid cells measuring 1 km². Each grid cell is characterized by a normalized value between 0 and 1. The pixel value reflects the state of the environment as characterized by the following variables: vegetation cover, water availability, terrain gradient, conflict events, and the presence of ethnic boundaries.

1.2.3.
What are the exogenous factors/drivers of the model?
The drivers of the model are largely captured in the spatial environment as described above. The presence of water and vegetation are considered pull factors whereas conflict, proximity to ethnic boundaries, and steep gradient are incorporated as push factors in the model. Once the agent has identified the most favorable grid cell, it must determine whether they are positions on private or public land. If it is public, they can continue their actions, however if the land is private, the agent must establish a deal with the landowner. If the deal is made the agent can remain on that parcel for the duration of a season, if they cannot make a deal then they must continue to search for alternate land. If they are unable to do so within a given season, the agent state changes from pastoralist to IDP and they exit the simulation.

If applicable, how is space included in the model?
The modeled environment is a spatial grid, where each grid cell measures 1 km², collectively covering the full extent of Somaliland. The spatial environment is artificially created through the combination of GIS datasets and information derived from satellite imagery. The individual data layers are incorporated into the spatial environment through the following equation: The additive pixel value is therefore representative of the condition of the artificial environment, which is considered the primary driver of pastoral migration. Once an agent moves to a new grid cell, the agent determines whether they are located on private or public land. If the agent is on public land, they are free to remain at that location without restraints for the duration of a season. If the agent happens to be on private land, the agent must negotiate a land-sharing deal with the local landowner. If the agent successfully strikes a deal with the landowner, they are able to remain at that location for the duration of the season. However, if the agent is unable to obtain access, they are required to move to another grid cell within the same time tick and again consider whether the new grid location is private or public land and repeat the process described above. If the agent moves to grid cells in private land and is also unable to make a deal on three separate occasions in any given season, the agent state switches from pastoralist to IDP, at which point the agent exists the simulation. The data used to inform agent generation was disaggregated by administrative regions.
More granular data at smaller administrative units was largely unavailable.
The variables included in the spatial environment were available at several different spatial resolutions. For instance, the global surface water index was applied to data from Sentinel 2, which captures data at 10 m resolution. The vegetation index was calculated using data from MODIS Terra at 250 m resolution. However, to create one continuous spatial layer, the data layers were aggregated to 1 km resolution. The lower resolution was thought to adequately capture movement on a regional scale and also minimized the time and effort required by the machines running the model. The basic rationale behind the agent decision-making process is that the agent must identify a grid cell that is suitable for grazing, given a combination of factors. This is done by seeking out the grid cell with the highest additive score in a random proximity to the agent's location. Once the grid cell has been successfully identified, the agent must confirm that they are able to graze here, since the land may be privately owned.
If the agent is consistently able to occupy a grid cell without any problems, they are successful in their mission. While the agent is able to move at every time tick, the agent is also able to remain at any given grid cell for the duration of a full season, after which they will be required to move. This encourages the agent to follow typical seasonal movement patterns. None.

2.3.3.
What is the spatial scale of sensing?
The spatial extent of the sensing is determined through a randomized distance between 15 -30 km.

2.3.4.
Are the mechanisms by which agents obtain information modelled explicitly, or are individuals simply assumed to know these variables?
The individuals are assumed to know the variables they are sensing. Agents are homogeneous in their decision-making.

Stochasticity
2.8.1. What process (including initialization) are modelled by assuming they are random or partly random?
The agent start locations are randomly generated with two constraining rules as described in section 2.9.1

Observation
2.9.1. What data are collected from the ABM for testing, understanding and analyzing it, and how and when are they collected?
At every time tick, the following data are collected: 1) geographic position of the agent, 2) the score associated with the pixel within which the agent is situated, and 3) the clan and ethnic affiliation of the agent. The spatial analysis of these attributes will provide insight about where and how far agents move and what the relationship is between agent movement and individual variables. 2.9.2.
What key results, outputs or characteristics of the model are emerging from the individuals? (Emergence) We anticipate that varying spatial and temporal patterns of migrations will emerge from the model in response to changing environmental and conflict factors.

Implementation details
3.1.1. How has the model been implemented?
The model was implemented using Java in Repast.

3.1.2.
Is the model accessible, and if so where?
The source code is stored on GitHub.

3.2.1.
What is the initial state of the model world, i.e. at time t=0 of a simulation run?
The state of the model at t = 0 mimics the environment as it was in January 2008. The agents generated at t = 0 are generated randomly throughout Somaliland with two conditions: 1) Agents cannot be in areas that are labeled as water bodies or sand, and 2) agents cannot be within 14 km and 4 km radius from a major city or settlement, respectively.

3.2.2.
Is the initialization always the same or is it allowed to vary among simulations?
The simulation initialization is always the same.

3.2.3.
Are the initial values chosen arbitrarily or based on data?
The initial values of the simulation environment are based on aggregated GIS and remotely sensed data that were captured at the time of the simulation date.
The number of agents generated per administrative units are informed by data captured by UNPF in 2013.

Input data
3.3.1. Does the model use input from external sources such as data files or other models to represent processes that change over time?
The spatial environment changes on a seasonal basis, these files were preprocessed by analysts. The input variables and their sources are listed below: • Agent characteristics | Population distribution is derived from a UNFPA