(2018). Impact of earthquake source complexity and land elevation data resolution on tsunami hazard assessment and fatality estimation.

This study investigates the impact of model complexity in source characterization and digital elevation model (DEM) resolution on the accuracy of tsunami hazard assessment and fatality estimation through a case study in Padang, Indonesia. Two types of earthquake source models, i.e. complex and uniform slip models, are adopted by considering three resolutions of DEMs, i.e. 150 m, 50m, and 10m. For each of the three grid resolutions, 300 complex source models are generated using new statistical prediction models of earthquake source parameters developed from extensive ﬁ nite-fault models of past subduction earthquakes, whilst 100 uniform slip models are constructed with variable fault geometry without slip heterogeneity. The results highlight that signi ﬁ cant changes to tsunami hazard and fatality estimates are observed with regard to earthquake source complexity and grid resolution. Coarse resolution (i.e. 150 m) leads to inaccurate tsunami hazard prediction and fatality estimation, whilst 50-m and 10-m resolutions produce similar results. However, velocity and momentum ﬂ ux are sensitive to the grid resolution and hence, at least 10-m grid resolution needs to be implemented when considering ﬂ ow-based parameters for tsunami hazard and risk assessments. In addition, the results indicate that the tsunami hazard parameters and fatality number are more sensitive to the complexity of earthquake source characterization than the grid resolution. Thus, the uniform models are not recommended for probabilistic tsunami hazard and risk assessments. Finally, the ﬁ ndings con ﬁ rm that uncertainties of tsunami hazard level and fatality in terms of depth, velocity and momentum ﬂ ux can be captured and visualized through the complex source modeling approach. From tsunami risk management perspectives, this indeed creates big data, which are useful for making effective and robust decisions.


Introduction
Recent advances in geosciences have drastically enhanced the accuracy of numerical simulations of complex natural phenomena, such as earthquake rupture processes and consequent earthquake motions and tsunamis Ismail-Zadeh et al., 2017). In tsunami sciences, the improvements lead to sophistication in earthquake source modeling and availability of high-resolution bathymetry and digital elevation models (DEMs), which enhance the quality of tsunami simulations (Ridente et al., 2014;Ramírez-Juidías et al., 2017). Consequently, these progresses in turn have promoted more rigorous treatment and visualization of uncertainties associated with tsunami disaster risk management.
Among the three key elements of tsunami risk, tsunami hazard assessment can be carried out in three stages: tsunami source modeling, propagation, and inundation (Srisutam and Wagner, 2010;Sleeter et al., 2017). Typically, the tsunami hazard assessment results are significantly affected by the source and inundation modeling because those are subjected to large uncertainties (Mueller et al., 2015;Sch€ afer and Wenzel, 2017). On the other hand, a well-established methodology, e.g. linear/ non-linear shallow water equations, has been verified and widely used for tsunami propagation modeling (King, 2015;Park and Cox, 2016;Feng et al., 2017). Given accurate offshore bathymetry data, e.g. GEBCO and SRTM15-plus, propagation components involve relatively less uncertainties and can be regarded as solved problems (Feng et al., 2017).
Nowadays, so-called source inversion analyses have been conducted using various types of data from teleseismic, strong motion, GPS, and tide gauge stations. One example of such developments can be seen by examining the fast-growing size of earthquake source model databases, such as SRCMOD (Mai and Thingbaijam, 2014; http://equake-rc.info/ SRCMOD/). The SRCMOD database is an online database of finite-fault rupture models of past earthquakes around the world. As of June 2017, 334 finite-fault models from 169 earthquakes are available. About 109 out of 334 models are for tsunamigenic earthquakes generated from subduction zones and its number, especially for large tsunamigenic events, is increasing steadily (Fig. 1A). One notable trend of the SRCMOD database is the change of the resolution of source model descriptions over time (Fig. 1B); more models have been developed by capturing finer details of the rupture processes. Moreover, gathered finite-fault models can be used to characterize source parameters (geometry and slip features) for subduction earthquakes and therefore can be adopted to develop new statistical prediction models of the parameters for future  A. Muhammad, K. Goda Computers and Geosciences 112 (2018) 83-100 Table 1 Scaling relationships of the earthquake source parameters . The prediction error terms of the scaling relationships are represented by epsilons, which are the standard normal variables (i.e. zero mean and unit standard deviation).

Parameter
Prediction models W log 10 W ¼ À0:4877 þ 0:3125Mw þ 0:1464εW L log 10 L ¼ À1:5021 þ 0:4669Mw þ 0:1717εL D a log 10 Da ¼ À5:7933 þ 0:7420Mw þ 0:2502εDa D m log 10 Dm ¼ À4:5761 þ 0:6681Mw þ 0:2249εDm A z log 10 Az ¼ À1:0644 þ 0:3093Mw þ 0:1592εAz A x log 10 Ax ¼ À1:9844 þ 0:4520Mw þ 0:2204εAx Box-Cox power A normal variable with mean equal to 0.312 and standard deviation equal to 0.278 Hurst number A value of 0.99 with probability of 0.43 and a normal variable with mean equal to 0.714 and standard deviation equal to 0.172 with probability of 0.57 Table 2 Linear correlation coefficients of regression residuals of the scaling relationships for the six earthquake source parameters. Note that the Box-Cox parameter and the Hurst number are considered to be independent. tsunamigenic scenarios . These new statistical prediction models may be used to generate the heterogenous/complex slip models and hence, can be used to rigorously study the effects of earthquake source complexity on the tsunami hazard assessment and fatality estimation.
For inundation modeling, a major source of uncertainty can be attributed to land DEM data in coastal areas (Griffin et al., 2015;Farahani et al., 2017), which have been significantly improved in the last two decades. The spatial resolution of global elevation data has been improved from~900 m (GTOPO30) in 1990s to~30 m (SRTM1) in 2014 (see Fig. 2). Moreover, data with resolutions up to 10 m or less (e.g. 5 m in Padang areas) are available at the regional scale using the advanced satellite and aerial photography. Such improvements lead to accurate evaluations of tsunami hazard metrics, such as depth and flow velocity   A. Muhammad, K. Goda Computers and Geosciences 112 (2018) 83-100 (De Risi et al., 2017). Building upon the above findings, this study aims at evaluating the impact of source characterization complexity and DEM resolution on the accuracy of tsunami hazard assessment and fatality estimation. This study is focused upon Padang, West of Sumatra, Indonesia which is one of the most tsunami-prone areas around the world due to future tsunamigenic events from the Mentawai segment of the Sunda subduction zone (Diana et al., 2014;Alvina et al., 2015;Muhammad et al., 2016Muhammad et al., , 2017. Two future scenarios (M w 9.0) of earthquake sources in the Mentawai-Sunda zone are generated by considering (1) uniform-slip source models with variable fault geometry and mean slip and (2) complex source models with heterogeneous slip distribution. 100 uniform-slip source models are used, whilst 300 stochastic source models are generated using new statistical prediction models of source parameters that are based on finite-fault models of the past subduction earthquakes (Muhammad et al., 2016). Moreover, for inundation modeling, three elevation resolutions, i.e. 150 m, 50 m and 10 m, are considered and hence, a total of 1200 scenarios are developed in this study. It is A. Muhammad, K. Goda Computers and Geosciences 112 (2018) 83-100 noteworthy that stochastic source models are particularly useful for capturing the uncertainties associated with earthquake ruptures, and that use of stochastic sources in tsunami hazard and risk assessments allows a new way of summarizing the hazard and risk analysis results. In other words, new probabilistic approaches create big data for more robust and effective tsunami risk management. In this study, a tsunami hazard is the main focus without accounting for other earthquake impacts (e.g. shaking damage and ground failures) and hence, the fatality estimation also excludes the earthquake effects. Therefore, the results reported in this paper constitute the lower bound estimate of fatalities due to the mega-thrust subduction earthquakes. Potential earthquake impacts have been discussed in Muhammad et al. (2017). The scope of this study was limited partly due to unavailability of extensive building information data for Padang areas (Schlurmann et al., 2010;Muhari et al., 2010Muhari et al., , 2011, which is necessary for evaluating both earthquake and tsunami losses quantitatively.

Modeling of tsunami sources
Two critical factors to determine earthquake sources for the future tsunamigenic events are earthquake location and source parameters (e.g. fault geometry and slip distribution). In general, the locations of earthquake source scenarios are defined based on the fault rupture areas of the past seismic events in a region of interest (Martínez-Alvarez et al., 2015;Abril and Peri añez, 2017). On the other hand, the earthquake source parameters are determined using empirical scaling relationships that are typically generated based on historical events (e.g. Wells and Coppersmith, 1994;Somerville et al., 1999;Mai and Beroza, 2002;Strasser et al., 2010;Murotani et al., 2013;Thingbaijam et al., 2017). The empirical relationships between moment magnitude (M w ) and other earthquake source parameters, e.g. rupture length, rupture width, rupture area, and slip statistics, have been developed extensively since 1950s (e.g. Tocher, 1958;Iida, 1959). Among the existing global scaling laws, the scaling laws developed by Goda et al. (2016) may be considered as one of the most comprehensive and reliable scaling laws to determine earthquake source parameters of future tsunamigenic events from the Mentawai segment of the Sunda subduction areas because of the following reasons. First, it was developed using extensive finite-fault models of the past subduction earthquakes around the world (i.e. 226 models). Second, it incorporated the uncertainty and dependency of essential earthquake source parameters using multivariate probabilistic models, i.e. fault geometry parameters (the fault width, W, and fault length, L), slip statistics parameters (mean slip, D a , maximum slip, D m , and Box-Cox number, λ and spatial heterogeneity slip parameters (correlation lengths along strike, A x, and dip, A z, directions and Hurst number, H), and third, It was successfully modelled the future tsunamigenic earthquake source scenarios generated from the Mentawai-Sunda zone, Indonesia by considering the regional seismological characteristic of the Sunda subduction zone (see Muhammad et al., 2016 andMuhammad et al., 2017). Table 1 lists the scaling relationships for W, L, D a , D m , A z , and A x , respectively, by Goda et al. (2016). In the equations, epsilon terms represent the prediction errors of the equations and thus the relationships are probabilistic. Furthermore, their correlation coefficients are given in Table 2. Stochastic slip models can be simulated by taking into account all uncertainty and dependency of the earthquake source parameters presented in Table 2 Muhammad et al., 2016;Mori et al., 2017). A spectral synthesis method is adopted to model the stochastic earthquake slip by characterizing earthquake slip distribution with wavenumber spectra (see Mai and Beroza, 2002;Goda et al., 2014 for details). Subsequently, a total of 300 stochastic models and 100 uniform models of future tsunamigenic earthquakes in the Mentawai-Sunda zone are generated using the new scaling relationships. The generated stochastic source models have heterogeneous spatial slips over the fault plane simulated through stochastic earthquake slip synthesis (see the top panel of Fig. 3A). On the other hand, the uniform models have variations of fault geometry and mean slip with a constant slip value over the fault plane taken from the generated mean slip (see the bottom panel of Fig. 3A). The numbers of source models for complex and uniform cases, i.e. 300 and 100, respectively, were determined based on the convergence tests of the various tsunami outputs. The convergence tests were carried out by evaluating the stability of tsunami simulation results (i.e. maximum tsunami depth and height) against the number of simulations for each type of earthquake source complexity (see De Risi and Goda, 2016).

Elevation data for tsunami simulation
In this study, local data of 5 m-DEM (DEM5) and 3 m-bathymetry (Bathy3m) are used to produce elevation data for tsunami simulation. DEM5 was generated through the German Indonesian Tsunami Early Warning System for the Indian Ocean (GITEWS) project. The DEM5 was developed by combining extensive remote sensing data including highresolution satellite imagery from the IKONOS and GPS measurements.
On the other hand, to obtain the Bathy3m data, ship sounding surveys were carried out near shore Padang with depth ranges from 3 m to 50 m. An accuracy assessment of DEM5 and Bathy3m shows that the vertical errors of these elevation data are about 0.5 m. The vertical error for land elevation in DEM5 was calculated from the comparison with the observed data using GPS measurements. The error of ocean depth in the Bathy3m dataset was estimated based on the accuracy of echo sounder device used for ocean depth measurements, noting that the error of the sounding depth is only~0.2% (Taubenbock et al., 2009;Schlurmann et al., 2010). Therefore, it is considered as a reliable dataset for tsunami hazard and risk assessments. Moreover, missing land elevation and bathymetry data which are not covered in those datasets are supplemented from GDEM2 (https://asterweb.jpl.nasa.gov/gdem.asp) and GEBCO2014 (http://www.gebco.net/data_and_products/gridded_ bathymetry_data/), respectively.
The nested grid systems having five levels (connected with a grid-size ratio of 3) are further developed: 1350 m, 450 m, 150 m, 50 m, and 10 m (see Fig. 3B). Subsequently, to study the effects of elevation data resolution on tsunami hazard assessment and fatality estimation, three finest grid levels of land elevation data, i.e. 150 m, 50 m, and 10 m, are considered. The 150-m grid is used to represent the coarse resolution, whilst 30-m to 50-m is a typical finer resolution for tsunami simulation in a region near Padang (Yue et al., 2014;Griffin et al., 2015). On the other hand, the 10-m grid is taken into account since the other hazard parameters, such as flow velocity and momentum flux, are found to be sensitive to the tsunami simulation resolution even though the computational cost is expensive (De Risi and . Fig. 4 presents the DEM profiles of three grid resolutions for tsunami simulation. Visually, the DEM appearances for the 10-m and 50-m grid resolutions are similar, whilst it is relatively blur for the 150-m grid. The initial deformation of sea bed is calculated based on the generated source models by considering both horizontal and vertical displacements of the seafloor using Okada (1985) and Tanioka and Satake (1996) formulae (see Fig. 3C). Tsunami simulation is then carried out by solving non-linear shallow water equation with run up (Goto et al., 1997). In this study, the effect of roughness coefficient on tsunami hazard assessment is not investigated due to lack of detailed information for the coefficient in Padang. The roughness coefficient of 0.025 m -1/3 s is used for ocean because the simulated offshore tsunami wave height using this value can be matched well with the observation data from the past events, such as the 2004 Indian Ocean tsunami (Imamura, 2009;Sugawara and Goto, 2012). Moreover, the roughness coefficient for land has relatively small impact when the value is changed from 0.025 m -1/3 s to 0.08 m -1/3 s (Griffin et al., 2015) and hence, a value of 0.06 m -1/3 s is adopted for land because the inundation inland produces from this value agreed well with the observed tsunami inundation depth from the 2004 and the 2011 Tohoku events (Kaiser et al., 2011;Griffin et al., 2015). The fault rupture is assumed to occur instantaneously. Moreover, the duration of the tsunami simulation is set to 2 h and the time step is varied from 0.1 s to 1 s depending on the grid resolution. This set up satisfies the Courant-Friedrichs-Lewys (C.F.L.) criterion for the bathymetry and elevation data of the Mentawai-Sunda region. The Monte Carlo tsunami simulation is finally performed using different initial deformation profiles calculated from the generated uniform and complex source models.

Tsunami hazard assessment
Results from the Monte Carlo tsunami simulation based on the generated earthquake source models can be used to produce probabilistic tsunami hazard results in a region of interest and further used for risk assessment (see Fig. 3D and E). In this study, three tsunami hazard parameters, i.e. maximum tsunami depth (H max Þ, maximum flow velocity (V max ), and momentum flux ðM FÀmax Þ; are considered. Tsunami depth/ height and velocity are popular parameters to represent the tsunami hazard level in the affected areas and have been extensively used to assess the tsunami hazard (Charvet et al., 2015;Park and Cox, 2016). On the other hand, momentum flux, calculated as the squared velocity multiplied by inundation depth (see Equation (3)), has been found to represent the tsunami hydrodynamic force on structures and are related to the building damage probability (FEMA, 2012;Park and Cox, 2016;Park et al., 2017). The spatial data of such tsunami force and impact parameters generated from the future tsunamigenic events will enable coastal planners to identify high risk regions and locate new development in safer areas.
The maximum tsunami inundation depth (H max ) at each location (x and y) is calculated by taking the difference between the maximum tsunami height (η) and the elevation (H land ) at a point of interest (Equation (1)).
H max ðx; yÞ ¼ ½ηðx; y; tÞ max À H land ðx; yÞ The tsunami velocity ðV max Þ at each point (x and y) used in this work is the resultant velocity of the maximum velocity in X and Y directions (u and v, respectively) from each simulation time step, t, calculated using Equation (2).
In addition, the maximum momentum flux ðM FÀmax Þ is calculated by taking the maximum value of multiplication between the tsunami depth and the squared velocity at each point as presented in Equation (3  On the other hand, the percentage of elderly is relatively small (4%) and that for infants and children is about 17% (see Fig. 5A). Subsequently, the mobility during the evacuation is not critical. In this study, the population data obtained from the WorldPop database are adopted (http://www.worldpop.org. uk/). The WorldPop database includes high-resolution population maps (approximately 100 m by 100 m resolution) of Padang in year 2010, 2015, and 2020. By incorporating various geographical and demographic data, the WorldPop database has been constructed. The population numbers of WorldPop data (2015) in Padang (i.e. South, North, West, and East of Padang) are slightly smaller than the BPS data (2013; i.e. 500, 000 people for the WorldPop, whilst it is 510,000 people for the BPS). In this study, the population data in 2015 are used to estimate the fatality due to the future tsunamigenic event. Moreover, building structures in Padang are mainly categorized as unreinforced masonry, confined masonry, and reinforced concrete (RC) frames with masonry fills (see Fig. 5B for details; Wilkinson et al., 2012;Mulyani, 2013).

Fatality estimation
Several models in predicting fatality have been developed using observed tsunami hazard parameters (i.e. tsunami height, depth and building damage) from the past tsunami events, e.g. the 2004 Aceh-Andaman tsunamigenic earthquake and the 2011 Tohoku earthquake (Kawata, 1997;Marchand et al., 2009;Yeh, 2010;Suppasri et al., 2012Suppasri et al., , 2016. Typically, the death rate (DR) is used to estimate fatality in a region of interest. In order to determine appropriate fatality models for calculating fatality in Padang, three criteria are further adopted. First, the model is developed through extensive tsunami survey data from the historical tsunami events and/or is the most updated fatality model in current literature. Second, the model is developed by assuming only one parameter (i.e. tsunami depth or height) used to calculate the fatality without considering other parameters (e.g. people evacuation and building damage probability). Third, the model has been evaluated or successfully calibrated for the fatality number observed in Banda Aceh, Sumatra due to the 2004 Aceh-Andaman tsunami. The last criterion is because current situations of tsunami mitigation measures in Padang resemble those in Banda Aceh. Therefore, in this study, four fatality models (see Fig. 6) are taken into account: two variations of the models by Kawata (1997), a model by Suppasri et al. (2012), and a model by Suppasri et al. (2016). Model 1 and model 2 were developed by Kawata (1997) with and without considering the existing evacuation plan. Model 3 (Suppasri et al., 2012) revised model 1 to match with observational fatality data in Banda Aceh. The first three models use tsunami height to calculate death rate. In addition, model 4 was generated by Suppasri et al. (2016) considering tsunami height to calculate the death rate.
Total number of fatality (TC) is finally calculated by summing all fatalities occurred in each inundation point (x and y). Assuming tsunami hazard parameter (TP) at x and y represents maximum inundation height, η max , or maximum inundation depth, H max , and death rate, DR, occurred at the expected inundation point, the total fatality is further calculated using Equation (4).

Effect of DEM resolution on tsunami hazard assessment
Tsunami depth, velocity, and momentum flux maps of Padang areas for the three DEM resolutions using stochastic models are shown in Figs. 7, 9 and 11, respectively. For each tsunami hazard parameter map, three probability ranks, i.e. the 10th percentile rank, the 50 th percentile rank, and the 90 th percentile rank out of 300 source models, are included as summary statistics. In addition, total inundation areas and total inundation areas above 1 m depth are also presented in the maps. The 1m depth is adopted because at this water level, evacuation may be difficult. Visually, the tsunami depth maps produced from the 10-m and 50-m grids are similar, whilst for the 150-m grid, the depth maps for all three percentile ranks are slightly different compared to the other two specifically near the ocean and rivers (see Fig. 7). Moreover, the inundation areas indicated in Fig. 7 quantify the tsunami depth differences due to the use of three different DEM resolutions. In general, the higher elevation resolution results in larger inundation areas. However, the differences of inundation areas from the three grids are considered being small, i.e. the percentage differences of the inundation from the 50-m grid and the 150-m grid against the 10-m grid are only~5%.
To investigate the differences of tsunami depth, more local areas of Padang, i.e. Teluk Bayur (see the red rectangle in the left panel of Fig. 7A), are focused upon. Fig. 8 shows that the depth maps for the 10-m and 50-m grids are similar. Nevertheless, the 50-m resolution may not capture detailed spatial features of the tsunami hazard along the coastal line (see white rectangle areas in Fig. 8A and B). In addition, the 150-m grid results are not accurate along the coast. Several significant tsunami depth areas found in the 10-m and the 50-m grids disappear in the 150-m grid cases.
On the other hand, the maximum tsunami velocity in Padang is about 7 m/s which is mainly found in the areas near the coast and rivers (see Figs. 9 and 10). In general, the tsunami wave velocity is higher at finer elevation resolution (i.e. 10 m). Moreover, the 150-m grid may not capture accurate tsunami velocity profiles specifically near the coast.  A. Muhammad, K. Goda Computers and Geosciences 112 (2018) 83-100 grids is almost covered by red color (~250 m 3 /s 2 ). In contrast, the 50-m and the 150-m grids visualize relatively smaller momentum flux, e.g. several parts are in yellow color. In addition, Fig. 13 and Fig. 14 show the comparisons of the three tsunami hazard parameters of two grids (coarser grid versus finer grid) and their residual plots against the distance from the coastal line, respectively. The residual is calculated by taking the difference between the larger grid and the smaller grid. A diagonal dashed-black line in Fig. 13 corresponds to the perfect positive correlation coefficient, r, equal to 1, whilst the blue line in Fig. 14 represents a zero-difference base line. In general, the figures show that most of the depth data for the 50-m and 10-m grids are almost identical, whilst the 150-m resolution results are typically overestimated in comparison to the base case (i.e. the 10-m grid). However, the differences are mostly minor, i.e. less than 0.5 m. On the other hand, the velocity comparison confirms that a larger grid (i.e. the 50-m and the 150-m grids) underestimates flow velocity in comparison to a finer grid (i.e. 10-m grid). In addition, the variability of velocity differences between the 10-m and the other grids is larger than the tsunami depth parameter. The results also exhibit that momentum flux is the most sensitive to the DEM resolutions among the three hazard parameters. The main reason is because the momentum flux calculation incorporates both depth and velocity parameters (see Equation (3)) that can significantly increase the variability of momentum flux differences. Another important finding is that the three tsunami hazard parameters are sensitive to the DEM resolution when it is located closer to coastal line (<1 km). To examine the effect of earthquake source model complexity on tsunami hazard assessment, tsunami hazard map profiles in terms of depth, flow velocity, and momentum flux produced from the 10-m DEM using uniform models are presented in Fig. 15 and Fig. 16 for Padang and Teluk Bayur, respectively. The counterparts of the stochastic source models are shown in Figs. 7A, 9A and 11A for Padang and Figs. 8A, 10A and 12A for Teluk Bayur. The map comparisons in Fig. 7 versus Fig. 15A clearly indicate that the inundation areas predicted in the uniform models is two to three times smaller than the inundation areas calculated in the stochastic models. The residuals calculated by taking the difference between the uniform models over the stochastic models for all percentiles shown in Fig. 17A also confirm that the uniform model majorly underestimates the depth in comparison to the stochastic models. For flow velocity and momentum flux, the uniform models also substantially underestimate the hazard parameters. As presented in the zoom red rectangle areas in Teluk Bayur for both parameters (see Fig. 16B and C), the results from the uniform models predict much smaller velocity and momentum flux in comparison to the stochastic models (Figs. 10A and 12A). Flow velocities at all probability ranks in Fig. 17 are majorly overestimated specifically for larger velocity values. The differences in terms of residual for velocity also corroborate this trend. The differences reach up to 10 m/s. For momentum flux, the differences even become greater (up to 4 times differences) since it is calculated by considering both tsunami depth and velocity parameters.
In general, the above-mentioned results indicate that the tsunami hazard parameters, i.e. depth, velocity, and momentum flux, are more sensitive to the earthquake slip models than the grid resolution. Note that the differences between the complex and uniform models are attributed to earthquake slip heterogeneity only (other earthquake source parameters, e.g. geometry and mean slip, are identical). Substantial underestimation of tsunami depth/height may produce inaccurate tsunami hazard maps that are currently in use for developing tsunami evacuation plans. Subsequently, the consequences due the use of uniform models to develop tsunami evacuation plans may be significant. On the other hand, flow velocity and momentum flux are also essential to determine the building damage state and hence, by using the uniform models, tsunami risk assessments may be inaccurate. Accordingly, mitigation policies producing from such analyses may also be unreliable and may not be implemented effectively in coastal community.

Effect of DEM resolution and earthquake source model complexity on fatality estimation
Using tsunami height and depth parameters calculated from all cases, the fatality is further estimated (see the Fatality estimation section). To calculate the fatality, all tsunami height/depths produced from different DEM resolutions are linearly interpolated at the grid points of WorldPop. Using Equation (4), the total fatality is then calculated by considering the death rate based on the four fatality models (Fig. 6).
Generally, the fatality estimation clearly shows the importance of reducing the fatality in coastal areas because the estimated fatality numbers are generally high. Existing of tsunami evacuation plan in coastal areas is also essential to reduce the fatality number. As shown in Table 3, for the worst scenario (the 90 th probability rank), the fatality could be reduced significantly (up to~97%) when the people are effectively evacuated. Although the fatality models used in this study may involve large uncertainties in calculating the fatality (e.g. only considering one tsunami hazard parameter), it clearly highlights that tsunami evacuation drills are essential to be carried out in coastal community. By increasing the awareness of tsunami evacuation plan in a tsunami-prone region, the fatality could be reduced substantially. Moreover, Table 3 also underlines huge human loss due to the future megathrust earthquake from the Mentawai segment in Padang areas. The fatality could be as much as 200,000 people as estimated using model 1 which is almost 75% of the total population of Padang in coastal areas. Emergency managers in Padang should be aware of potentially devastating tsunami losses and strive to reduce the human losses in the near future due to the tsunamigenic events.
Regarding the DEM resolution effects on fatality estimation, Table 4 presents the percentage difference between the finer grid and the coarser grid (i.e. the 10-m grid vs the 50-m grid and the 10-m grid vs the 150-m grid). Negative percentage in the table means that the finer grid underestimates the fatality in comparison to the coarser grid. In general, all comparisons confirm only~5% differences in terms of fatality specifically for the 50 th and 90 th percentile ranks. However, the coarser grid seems to underestimate the fatality for the 10th percentile rank. It is mainly due to significant values of depths above 1 m that are underestimated for the coarser grid in the 10th percentile rank compared to the finer grid (see Fig. 13). The fatality differences tend to be substantial A. Muhammad, K. Goda Computers and Geosciences 112 (2018) 83-100 when the comparison is carried out between the complex and the uniform models, i.e. it ranges from 78% to 100% (see Table 4). The results on fatality estimation confirm that the complexity of earthquake source characterization significantly influences the fatality number estimation. Subsequently, realistic tsunamigenic earthquake source models (i.e. stochastic models) generated by taking into account all uncertainty of earthquake source parameters, are essential to produce accurate and reliable tsunami risk estimation. Lack of earthquake source modeling, in particular slip heterogeneity modeling, may lead to inaccurate prediction of human losses. Such inaccurate prediction may further influence on the development of tsunami mitigation policies in coastal community. For instance, tsunami evacuation shelters may not be built in high risk human losses areas due to underestimating the fatality number in that region. Thus, major undesirable consequences might be caused from inaccurate earthquake source modeling.

Recommendations for tsunami disaster risk management
Extensive tsunami hazard assessments by considering different DEM resolution and source model complexity may suggest several essential recommendations for tsunami disaster risk management purposes. Firstly, accurate earthquake slip modeling is essential to produce reliable tsunami hazard and risk assessments. In this study, the earthquake slip is found to be a major source of uncertainty affecting the tsunami simulation results significantly and hence, the implementation of stochastic earthquake source modeling is deemed essential to develop rigorous and robust tsunami hazard assessment and fatality estimation in a region of interest. Secondly, at least a 10-m or 50-m grid is recommended for tsunami hazard assessment purposes using depth parameter. However, detailed tsunami depth assessment along coastal lines should be carried out using the 10-m or finer resolution. Thirdly, when the flow velocity and momentum flux are taken into account, finer grids (<10-m) are needed specifically within the range of 1 km from the coastal line. Finally, the extensive tsunami hazard and risk analyses carried out in this study, demonstrate the possibility of creating new big data information for tsunami disaster risk management purposes. The tsunami hazard potential may be presented rigorously for different probability ranks (represented by a set of tsunami hazard maps, in sharp contrast with conventional single worst-case hazard map) and further explored to create big data information of tsunami hazard level in a region of interest. Such information is valuable for the emergency managers to develop effective tsunami mitigation policies in coastal community. Rigorous information of tsunami hazard parameters from wave height and depth to flow velocity and momentum flux in a region of interest may also be generated through extensive tsunami simulations specifically by incorporating stochastic earthquake source scenarios. Those hazard parameters are important to calculate the tsunami risk in terms of economic and human losses, as demonstrated in the case study of this work. This will allow the related managers to give strong recommendations for the local authorities to implement risk management strategies in coastal community.

Conclusions
The main purpose of this study was to examine the effects of DEM resolutions and complexity of earthquake source characterization on tsunami hazard assessment and fatality estimation. To investigate such purpose, three DEM resolutions including the 10-m, the 50-m, and the 150m grids were adopted using two different earthquake source models, i.e. the stochastic and the uniform source models. Both the stochastic and uniform models had identical geometry but varying in slip value. The stochastic earthquake source models incorporated heterogeneity of spatial earthquake slip parameter, whilst the uniform models only used mean slip as the only to represent earthquake slip over the fault-plane. Subsequently, the tsunami hazard assessments in terms of depth, velocity and momentum fluxes in Padang areas were assessed by considering both different DEM grid resolutions and earthquake source model complexity. The tsunami hazard maps for the three parameters were developed for three probability ranks, i.e. the 10th, the 50 th and the 90 th percentile ranks, and were used to evaluate the differences in terms of tsunami hazard potential. Finally, the fatality was estimated to study the effect of DEM resolutions and earthquake source model complexity on fatality estimation.
The results show that the tsunami depth is less sensitive to the DEM resolutions in comparison to the flow velocity and momentum flux. However, all hazard parameters are more sensitive to the earthquake slip models than the grid resolutions. The tsunami hazard parameters are significantly underestimated in the uniform models compared to the stochastic models. The fatality estimation is also found to be more sensitive to the earthquake source models, in particular, earthquake slip, than the grid resolution. This highlights that accurate and reliable tsunamigenic earthquake source models (e.g. stochastic models) generated by taking into account all uncertainty and dependency of earthquake source parameters are essential to produce tsunami risk estimation. Moreover, the results also show that the 10-m grid developed from extensive geosciences measurements is extremely useful to improve the quality assessment of tsunami hazard. The availability of such datasets improves the tsunami preparedness by providing rigorous and extensive  percentile  3007  0  1234  405  2878  0  1177  385  2759  0  1140  272  50 th percentile  71093  180  31821  5090  71547  177  32002  5007  69971  179  31303  4542  90 th percentile  201384  2702  103471  19999  196465  2639  99779  19252  189500  2678  96307  18481 Uniform model  information of tsunami hazard maps using various parameters from depth, velocity to momentum flux and hence, is essentially needed for developing reliable mitigation strategies. Lastly, although the tsunami hazard assessment and fatality estimation in Padang were carried out using the results of rigorous stochastic tsunami simulations, there are limitations that need to be addressed in future studies. These include: (1) taking into account the building information data for tsunami inundation modeling to capture the tsunami hazard inland more realistically, (2) incorporating earthquake impacts on the fatality estimation, and (3) other impact metrics, such as building damage and loss, should be taken into account when assessing the earthquake risk in Padang.