The last Turkish census was completed in 2011 before the onset of Syrian mass migration. Therefore, most population models built from census-based sources do not account for the presences of Syrian refugees. This is not an intentional error (Gridded Population of the World version 4 dataset (GPWv4; ) explicitly states this particular shortcoming) but rather a systematic problem associated with infrequent data collection. Any forward-modeled population dataset for Turkey based on pre-2011 data will mischaracterize true populations unless refugees are explicitly included. Population models that incorporate migration at some level do exist, most notably Oak Ridge National Laboratory's LandScan^™ database , but they remain proprietary products.

As a framework for modifying regional census data for inter-period migration events, a geographic information systems (GIS) workflow was utilized to construct a regional refugee-inclusive gridded population model using freely available data from Turkey's Address Based Population Registration System (ABPRS) and the Turkish DGMM. The DGMM, part of the Turkish Department of the Interior, is responsible for regularly disseminating registered refugee population statistics. These statistics are widely used in refugee-related reporting by the European Commission, United Nations High Commissioner for Refugees, and the US Humanitarian Information Unit, among others.

The framework for this study employs a minimally modeled areal distribution process that disaggregates administrative population counts into cells of equal population. Turkish district level boundaries from the GADM database of Global Administrative Areas , clipped to the area of interest, were first converted into 3 km grid cells and equally distributed 2015 ABPRS populations according to the proportional number of cells in each district. Refugee migration data are monitored at the province level, one administrative boundary larger than the ABPRS estimates. As mentioned above, the exact position of non-camped Syrian refugees within their respective provinces is unknown. Accordingly, the existing district level population distribution was used as a proxy for refugee settlement patterns. The non-camped refugee population was distributed according to the relative percentages between district and province level populations. Camped refugee populations were assigned to the district corresponding to the camp location and removed from the populations otherwise distributed. The model was finalized by repeating the process used above for distributing ABPRS populations to allocate refugees into equally populated grid cells. The resulting gridded population model (Fig. 4) is spatially consistent and has discrete values for base population and registered refugee population.

Earthquake scenarios are an important tool for emergency management planning. Tools like the USGS' Prompt Assessment of Global Earthquake Risk (PAGER) system and the Federal Emergency Management Agency's (FEMA) HAZUS software have been used in the US for emergency planning at both the national and state level . As part of the earthquake fatality estimation process, synthetic ground motion fields were produced for a series of earthquake ruptures spanning five faults across southeastern Turkey. For each fault, moment magnitude 5.8, 6.4, and 7.0 earthquakes were simulated. This spread of earthquake magnitudes reflects moderate to major earthquakes within the magnitude range of historical earthquakes in the area as seen in earthquake catalogs covering Turkey . Five earthquake epicentral locations were selected along fault traces provided in the fault-source background model in the Seismic Hazard Harmonization of Europe (SHARE) project. It should be noted that the choice of exact epicentral location is somewhat arbitrary but can have an impact on fatality levels. Epicenters for this study were selected to represent geographically distributed fault segments and were chosen independently of refugee migration patterns.

The Global Earthquake Model's OpenQuake software platform was utilized to produce ground motion fields for each earthquake scenario. OpenQuake's scenario-based hazard assessment implements ground motion prediction equations to estimate the geographic distribution of shaking intensity for a user-specified fault rupture . An overview of rupture input parameters for each fault segment used in this study is available in Table 1. For each scenario, OpenQuake generates the rupture area internally from magnitude and rake using the area–magnitude scaling relationship defined in . The rupture area is allowed to float along its corresponding fault trace (Fig. 5). All of the scenarios in this study utilized the ground motion prediction equation detailed in , relevant to earthquakes in Europe and the Middle East. Site amplification was accounted for by using Vs30 estimates from the USGS Global Vs30 Map server, which estimates Vs30 from topographic slope . OpenQuake implements site parameters by assigning each observation grid cell the site parameters of the nearest measurement in the Vs30 grid . For each earthquake scenario, 10 ground motion fields were produced – each resampling the aleatory uncertainty in the ground motion prediction equation.

There are a variety of methods for estimating earthquake fatalities. specifies three primary categories: empirical, analytical, and hybrid approaches. The three categories differ in their input data. Empirical methods derive fatality rates from historical records, analytical methods use detailed structural engineering and building occupancy information, and hybrid (semi-empirical) approaches use empirical estimates of building collapse rates and occupancy. The choice of methodology is usually dictated by data availability and the scale of analysis .

In this study, a semi-empirical methodology was used to estimate fatalities in earthquake scenarios. Empirical approaches were deemed poorly suited to this particular problem because fatality rates are derived from numerous historical earthquake events. This study is based on the concept that earthquakes in the short term will have higher fatalities due to contextual population increases. Analytical approaches, while the most robust of the three methods, also have the highest data requirements. The structural performance and building occupancy data necessary to support an analytic approach are not available for Turkey, even before considering the challenges of including refugee populations. Accordingly, this study employs the semi-empirical approach detailed in , given by Eq. (1).

E[L]≈∑i=1n∑j=1mPi×fij×CRj(Si)×FRj This approach estimates fatalities given a series of n grid cells and m structural types. Each grid cell's population Pi is first broken out into a fractional percentage for a given structural type fij. Fatalities are then calculated based on the collapse rate of structural type j (CRj(Si)) at macroseismic intensity (Si) and the fatality rate FRj of structure type j under collapse .

Empirical data from the World Housing Encyclopedia (WHE)-PAGER project, phase I, were used to constrain collapse rates. provides estimates of the building stock distribution under the PAGER taxonomy along with estimated collapse percentages. It is noted that several of the collapse probabilities in are higher than estimates that have been generalized across the entire WHE-PAGER phase I dataset . Accordingly, when available, collapse rates were calculated using generalized fragility coefficients (listed in Appendix A) using Eq. (). For building types without published coefficients, values were estimated using the methodology in , minimizing the residual error of the power function in Eq. () fit to a single set of collapse rates at given intensities. Fatality rates were drawn from for building types with HAZUS-MH fatality rates and generalized Turkish values from in their absence. CRj(S)=Aj×10BjS-Cj

The results presented in Tables 4 and 5 were transferred directly from fatality calculations without applying any rounding. Non-rounded values were included to allow for closer comparisons to be drawn between individual scenarios. However, this choice may inadvertently suggest that the values presented are very precise – this is not the case. Every attempt has been made to utilize the best data available, but semi-empirical fatality estimations remain a fundamentally uncertain process and will not be perfectly accurate. Yet there is ample evidence to suggest that fatality estimations remain a useful procedure for determining disaster scale and response capacity needs . The US Geological Survey provides the following estimates for response levels at varying earthquake fatality thresholds:

1–100 fatalities: regional response required;

It is also stressed that the values shown in Tables 4 and 5 do not represent fatality predictions for future earthquakes. Rather, the fatality estimates are better interpreted as order-of-magnitude estimates for hypothetical earthquakes of varying size and location. Therefore, the conclusions drawn henceforth are scenario specific – they should only be generalized to other scenarios with appropriate caution.

Fatality estimations were first produced for non-refugee populations to provide baseline values. These baseline values, shown in Table 4, help determine how earthquake fatalities in southern Turkey vary with earthquake magnitude, location, and time of day. At all five fault locations, increasing earthquake magnitudes from 5.8 to 6.4 corresponded with larger fatality increases (241 % on average) than subsequent increases when magnitudes were changed from 6.4 to 7.0 (175 % on average). These results are expected given the logarithmic relationship between magnitude and intensity.

Nighttime fatalities were estimated higher than daytime fatalities at all fault locations (an average of 160 %). This indicates that the building stock distribution occupied during working hours is less susceptible to collapse than the building stock distribution occupied during nighttime hours. These results are supported by the occupancy patterns seen in Table 3, which shows that populations generally transition from vulnerable masonry buildings at night to concrete structures during working hours. Additionally, it is probable that the percentage of populations located outdoors is higher during working hours than during nighttime hours, especially in rural environments. These findings add to the growing volume of research stressing the importance of including temporal elements into natural-hazard studies .

Every earthquake scenario included in this study produces casualties that would require at minimum regional response. Many of the scenarios, especially at earthquake magnitudes 6.4 and higher, would likely require national or international response. These results indicate consistently high levels of seismic risk across most of southeast Turkey – a region with a deep history of deadly earthquake activity . The differences in fatality estimates between fault locations register the relative proximity of each fault segment to areas with high populations. The two fault segments with the highest fatality estimates, Kırıkhan and Göksun, are both located within some of the highest-population districts across southeast Turkey.

Median fatality estimates and median absolute deviations based on 500 building occupancy iterations are shown in Table 5. The median absolute deviations of refugee fatality estimations, based on adjustments in building occupancy percentages, range from 25 to 55 % of median estimates. These variations may have implications for the severity of a particular earthquake event, but in general they do not dramatically change the estimated impact levels due to refugee populations. At four of five fault locations, median fatality estimates reach over 100 fatalities for earthquakes above magnitude 6.4. Accordingly, refugee populations are sufficiently large to produce fatality estimates that would require local or regional response. On the Kırıkhan fault, refugee population fatalities are high enough to merit international response. Thus, it is clear that refugee populations in southeastern Turkey should be included in the fatality estimation process.

However, in comparison to baseline fatality estimates, refugee populations constitute relatively small portions of overall fatalities. The relative contributions of refugee and non-refugee populations for each scenario are compared in Fig. 6. The Kırıkhan fault scenarios have the highest refugee contributions, with 25–27 % of total scenario fatalities coming from refugee populations. The Göksun, Bozova, and Türkoğlu scenarios all have 7–9 % refugee fatalities, and the Pütürge scenario has only 1–2 % refugee fatalities. These differences reflect the distribution of refugees throughout the study region, which is similar but not identical to existing population distributions. As a result, refugee contributions to total fatality estimates are not tied to baseline fatalities. The relationship is fairly close for the scenarios in this study, but as a general rule, baseline fatality estimates should not be assumed to be good predictors of refugee fatality estimates.

Ground motion, population estimates, collapse rates, fatality rates, and occupancy patterns are all subject to varying levels of uncertainty in the semi-empirical model. In the context of this study, two particular sources of uncertainty are worth highlighting.

When compared across all countries, WHE collapse functions have shown tendencies towards overestimating fatalities, with more significant effects in smaller earthquakes .

A general shortcoming of fatality estimation processes is the difficulty in separating out individual uncertainty terms. As a result, uncertainties are often wrapped together into a total-model-uncertainty term . The USGS PAGER implementation of total model uncertainty specifies the probability P of estimated losses e falling between two thresholds a and b as Eq. ().

P(a<e<=b)=Φlog⁡(b)-log⁡(e)ζ-Φlog⁡(a)-log⁡(e)ζ

This implementation relies on a hindcasted country-specific residual error term, ζ, defined as the normalized standard deviation of the logarithmic ratio of expected to recorded losses .

Because fatality estimations are generally considered to be order-of-magnitude estimates, a and b are commonly set to 1 order of magnitude above and below median estimated fatalities . Using the ζ value for Turkey (1.52), the probability P of actual fatalities in a given scenario falling within 1 order of magnitude above and below median estimated fatalities is 49 %. Thus, there is a 25.5 % chance that actual fatalities are greater than 1 order of magnitude above median estimated values and a 25.5 % chance that actual fatalities are less than 1 order of magnitude below median estimate values. These relationships apply to every scenario in this study.