A Strategic Framework for Natural Disaster-Induced Cost Risk 1 Analysis and Mitigation: A Two-Stage Approach Using Deep 2 Learning and Cost-Benefit Analysis

. Due to gradual increases in the frequency and severity of natural disasters, risks to human life and property from 16 natural disasters are exploding. To reduce these risks, various risk mitigation activities have been widely conducted. Risk 17 mitigation activities are becoming more and more important for economic analysis of risk mitigation effects due to limited 18 public budget and the need for economic development. To respond to this urgent need, this study aims to develop a strategic 19 evaluation framework for natural disaster risk mitigation strategies. The proposed framework predicts natural disaster losses 20 using a deep learning algorithm (stage I) and introduces a new methodology that quantifies the effect of natural disaster 21 reduction projects adopting cost-benefit analysis (stage II). To achieve the main objectives of this study, data of insured loss 22 amounts due to natural disasters associated with the identified risk indicators were collected and trained to develop the deep 23 learning model. The robustness of the developed model was then scientifically validated. To demonstrate the proposed 24 quantification methodology, reservoir maintenance projects affected by floods in South Korea were adopted. The results and 25 main findings of this study can be used as valuable guidelines to establish natural disaster mitigation strategies. This study will help practitioners quantify the loss from natural disasters and thus evaluate the effectiveness of risk reduction projects. This study will also assist decision-makers to improve the effectiveness of risk mitigation activities.

have also increased. The Intergovernmental Panel on Climate Change (The Fifth Assessment Report, 2014) has already warned 48 of an increase in global average temperature, average sea level escalation, heating, and acidification. In many countries, severe 49 weather events such as typhoons and heavy rains and changing patterns of meteorological disasters have already increased the 50 loss of many lives and property. These damages are expected to accelerate in the future (Kim et al., 2020). 51 52 Therefore, lives and property worldwide are threatened by natural disasters. Such threats will increase. To reduce these threats, 53 numerous non-governmental organizations and countries are investing a lot of time, budget, and manpower to mitigate risks 54 from natural disasters. Mitigation of risks can reduce the loss by decreasing vulnerability or by decreasing the frequency and 55 severity of causal factors (Rose et al., 2007). For risk mitigation, the execution of financial resources should be carried out 56 quickly and extensively. In practice, the efficiency and amounts of financial resources should be considered due to limited 57 resources. Hence, it is important to grasp the amount of risk and the effect of risk reduction at the same time to achieve the 58 ultimate reduction and mitigation of risk through an efficient use of limited resources. In other words, it is essential for risk 59 mitigation against potential risks by predicting the exact amount of risk, which aims to make an active investment to reduce 60 the predicted risk, and to find out the economic effect of the risk reduction. Consequently, as part of a case study on risk 61 mitigation costs, this study developed a strategic framework by developing a natural disaster damage prediction model using

Cost-benefit analysis of natural disaster risk mitigation 95
Mitigating the risk with efficient investment and operation of resources is a challenging task because resources are finite while 96 risk reduction should be done quickly and extensively. To address these issues, cost-benefit analysis has been widely adopted 97 (FEMA, 2005;Rose et al., 2007). For instance, efficient use of public resources is indicated when total estimated profits of a 98 risk mitigation activity surpass the entire cost or are parallel to earnings on investment of both private and public. 99 100 Disaster risk mitigation represents mitigating social, environmental, and economic damage caused by natural disasters. Since 101 economic losses due to natural disasters are hard to minimize or avoid separately, there is an increasing public demand for risk 102 reduction investments to reduce these economic losses (Bouwer et al., 2007;Shreve and Kelman, 2014). Since resources for 103 risk mitigation investment are restricted, it is critical to estimate economic costs and benefits in order to determine the 104 effectiveness and appropriateness of the investment. For instance, the Federal Emergency Management Agency of the United 105 States has reported that the average benefit cost ratio is 4 for risk mitigation investment (e.g., structural defense measures 106 against floods and typhoons, building renovations in preparation for earthquakes, etc.) after reviewing 4,000 natural disaster 107 risk reduction programs in the United States ( Kunreuther et al., 2012;Rose et al., 2007). In addition, studies in developing 108 countries have shown a high benefit cost ratio in a study of 21 investment activities such as re-establishment of schools and 109 forestry in preparation for tsunami (Bouwer et al., 2014). 110 111 Despite these high potential benefits, investment in risk reduction for residents living in areas at risk of natural disasters is 112 restricted (Bouwer et al., 2014). According to Hochrainer-Stigler et al. (2010), since natural disaster risk reduction measures 113 are focused on short-term outcomes, only about 10% of residents in areas vulnerable to natural disasters receive natural disaster 114 risk reduction measures in the United States. In the case of a natural disaster risk reduction project, a large initial investment 115 is required, which reduces the expected profit if performance indicators need to be met in a short period of time. As a result, 116 policy makers and politicians are reluctant to make bold investments in natural disaster risk reduction. They prefer to provide 117 economic support after disasters (Cavallo et al., 2013). This phenomenon is also reflected in the budget distribution of disaster 118 management funds of donations and development agencies. Most (98%) of the budget is allocated to reconstruction or relief. 119 Only the remaining budget (2%) is allocated to risk reduction (Mechler, 2005). As such, while the need for pre-disaster risk 120 reduction through proactive disaster investment is widely recognized, the economic impact of natural disaster risk reduction 121 is often not fully considered in decision-making. Moreover, although cost-benefit analysis (CBA) is the main decision-making 122 tool commonly used in public sector investment and financial evaluation, natural disaster risk is not sufficiently applied in 123 CBA (Hochrainer-Stigler et al., 2010). Natural disasters in public sector investment projects are often overlooked or not 124 evaluated in CBA assessments (Kreimer et al., 2003). Hence, in this study, the effectiveness of a natural disaster reduction 125 project was determined through a case study of cost-benefit analysis conducted by the Korean government is considered and 2 Research objectives and methods 128 To reduce economic losses caused by natural disasters, it is necessary to quantify losses caused by natural disasters and make 129 active investments to reduce risks. Therefore, for economic analysis of losses from natural disasters, this study attempted to 130 examine the investment effects, predict losses caused by natural disasters. The main objectives of this study are to develop a 131 strategic framework that predicts natural disaster losses using a deep learning algorithm and introduces a methodology to 132 quantify the effect of natural disaster reduction projects using cost-benefit analysis. To achieve the main objective of this study, 133 a two-stage approach was adopted. 134

135
In Stage I, this study collected reliable storm and flood damage insurance data and natural disaster risk indicators, created a 136 predictive model based on a deep learning algorithm, and verified. This study proposed a deep learning modeling framework 137 that could accurately learn and predict multiple natural disaster indicators known to affect losses caused by natural disasters. 138 The first research objective was achieved through the following steps: 139 1) To collect data on loss caused by natural disasters, this study collected data on claim payout for storm and flood In Stage II, data on natural disaster risk reduction projects conducted by national institutions were collected and cost-benefit 150 analysis was performed for cost of natural disaster risk reduction. This study intended to propose a framework for quantifying 151 the economic cost of natural disaster risk reduction. To realize the goal of this study, the following steps were used. In addition, 152 this study intended to propose a framework for quantifying the economic cost of natural disaster risk reduction. The second 153 objective of this study was achieved through the following steps: 154 1) Among natural disaster risk reduction projects carried out by the South Korean government, information on disaster 155 risk reservoir maintenance projects completed in 2009-2019 was collected. 156 2) The loss rate of storm and flood insurance in the region where the flood damage occurred after the completion of the 157 maintenance project was investigated through the Korea Insurance Development Institute (KIDI). 158 3) The amount of precipitation before and after the disaster risk reservoir maintenance project was investigated. insurance professional service organization that develops insurance products, calculates insurance rates, and protects the rights 166 of policyholders. It also collects and manages various statistical data such as insurance information and losses of each insurance 167 company (Choi and Han, 2015). Storm and flood damage insurance, which reflects the loss amount, is an insurance that 168 compensates for property damage caused by natural disasters (e.g., typhoons, floods, heavy rains, tsunamis, strong winds, 169 storms, heavy snow, earthquakes, and so on). It has been implemented since 2006 under the initiative of state and local 170 governments (Kwon and Oh, 2018). The insurance payout amount is determined by objective analysis of certified loss 171 assessment service according to standardized procedures for each insurance company. Its reliability is high (Kim et al., 2020). 172 The prediction model was trained, tested, and validated using losses and natural disaster risk indicators. 173

174
The cost of loss due to natural disasters was divided by the total net premiums to calculate the ratio and then log-transformed. 175 In addition, natural disaster risk indicators affecting insurance loss due to natural disasters were collected. For natural disaster 176 risk indicators, building type, wind speed, total rainfall, and peak ground acceleration were selected as variables through past  Table  178 1. Building types were set as dummy variables that consist of residential buildings and greenhouses. Wind speed and the 179 maximum value of rainfalls were collected from the Korea Meteorological Administration (KMA). Peak ground accelerations 180 were collected from the National Oceanic and Atmospheric Administration (NOAA). Descriptive statistics of variables are 181 displayed in Table 2

205
The collected loss data were pre-processed using a z-score normalization method to adjust the unit and quantity of the data. 206 The pre-processed completed input data were divided into a training set, a verification set, and a test set of data. The training 207 set of data were used for learning of the DNN algorithm. The verification set of data were used to judge whether training was 208 optimal and the test set of data were used to verify whether the developed model was finally trained for the purpose. In this 209 study, considering the amount of data, 70% of the total data were set as training set of data and 30% of them were used as test 210 set of data. Then 30% of training data were utilized as verification data. 211 212 The DNN model selected the optimal combination through a trial-and-error method since the DNN model could update the 213 weights of neural network nodes with a backpropagation algorithm. Since various combinations were possible depending on 214 the input variable and the output variable, it was necessary to find the optimal combination through the trial-and-error method. 215 For such an optimal combination, it is necessary to define the network structure scenario for setting the number of layers and 216 nodes and defining hyper parameters such as optimizers, activation functions, and dropouts (Cavallo et al., 2013). This study 217 adopted a network structure scenario with three hidden layers considering data characteristics. Dropout is a regularization 218 penalty to avoid overfitting. It was set to reduce prediction errors caused by overfitting. In this study, making an allowance for 219 the amount of training data, dropout was set to 0 and 0.2 and simulated. The ReLu (Rectified Linear Unit) function was utilized 220 as the activation function, a method of adjusting the weight of each node for optimal learning. The ReLu function allows the 221 input value to change when the input value is greater than 0 or less than 0. It was established to resolve the problem of gradient  values of these two models were compared. The MRA method is widely adopted as an essential method for numerical 243 prediction models (Kim et al., 2021). Table 5    projects were initiated. Among them, a total of 12 areas were flooded before and after the completion of the disaster risk 269 reservoir maintenance project. Table 6 shows the loss rate and maximum precipitation at the time of flooding before and after 270 completion of the maintenance projects in these 12 areas. Data about the loss amounts from storm and flood insurance were 271 obtained from KIDI. Precipitation data were collected from KMA and the maximum daily precipitation at the time of the 272 flooding was used. Insured loss was expressed as a rate of the incurred loss divided by the accrued premium. The loss rate 273 before the maintenance project was 34.32% on average, while that after the maintenance project was completed was 5.9% on 274 average, showing a sharp decrease of 82.8% on average. However, when data of precipitation as the main cause of flooding accidents during flood damage were compared, the average precipitation was 331 mm/day before the maintenance project and 276 215 mm/day after the maintenance project. It could be seen that the amount of precipitation was decreased by 35% when flood 277 damage occurred after the maintenance project. The sharp decrease in the loss rate after the maintenance project could be due 278 to the effect of the maintenance project. It could also be attributed to a relatively small amount of precipitation compared to 279 that before the maintenance project. Therefore, it is difficult to conclude that the decreased loss rate is due to the effect of 280 reducing storm and flood damage caused by the maintenance project. 281 Therefore, cost-benefit analysis was conducted to analyze the economic effect. Equal-payment-series present-worth factor was 283 used for cost-benefit analysis. Equal-payment-series present-worth factor, assuming an annual loss rate i, is a coefficient used 284 to find the present value corresponding to annual equivalent loss A for the next n years. Eq. (1) presents a widely used concept The initial cost of each maintenance project was collected through The Public Data Portal and the average cost of the 295 maintenance project was calculated. For the loss rate, the average loss rate of the loss area was used. For the annual loss amount, 296 the average annual loss for the study period (2009-2019) was used as seen in table 7. However, it was assumed that no 297 additional costs incurred due to the maintenance project. Figure 1 shows calculation results before and after the maintenance 298 project. As can be seen from Figure 1, the loss amount becomes smaller after 8 years due to investment through the maintenance 299 project. 300

Discussion 305
In Stage I, this study developed a model for predicting economic losses due to natural disasters using the DNN algorithm 306 among deep learning algorithms. For model development, insurance company's storm and flood damage insurance loss records 307 were used to collect economic losses caused by actual natural disasters. After developing a DNN algorithm model and training 308 it with collected data, the model was validated by comparing different models. In addition, network scenarios and hyper-309 parameters were found using the trial-and-error method to derive the optimal model. The DNN model was 15.2% less in the 310 MAE and 10.12% less in the RMSE than the MRA model. As shown in prediction results, the non-parametric model DNN 311 was more proper than the parametric model MRA model for the economic loss analysis of natural disasters with non-linear 312 characteristics. These results also indicate that the DNN model has higher reliability than other models in identifying financial 313 losses due to natural disasters. Due to the nature of natural disasters, the loss is very diverse. Thus, the prediction error value 314 can be very large. It can be seen that the DNN model reflects this diversity of natural disaster losses well. By using the 315 development model and the methodology described in this study, natural disaster risk managers will be able to predict the 316 financial loss cost of natural disasters or develop an optimal deep learning prediction model according to user conditions. It 317 can also be used as a reference when developing systems or models for predicting natural disaster losses in a public or private 318 sector. Based on this sophisticated economic loss prediction, it will be possible to make decisions for active risk reduction 319 investment. Such investment can strengthen natural disaster risk management and reduce the amount of risk, ultimately 320 reducing the economic loss caused by natural disasters. For example, it will be possible to calculate the amount of economic 321 loss in an area expected to be flooded in advance and establish a preventive strategy for loss measures and appropriate facility 322 investment according to the expected loss amount. Moreover, such loss forecasting can help prepare financial guidelines such 323 as emergency reserves and budgeting. It can also be used to prepare budget guidelines according to the calculated expected 324 loss and manage business continuity. In addition, according to established financial guidelines, it will be helpful for strategies 325 to avoid and transfer financial losses through insurance coverage or special purchases suitable for expected losses. These 326 activities can ultimately reduce the risk of financial loss due to natural disasters. Nevertheless, this study has some limitations. 327 First, owing to the limited data set, it was problematic to accumulate different data sets. Additional research in the future is 328 needed to parallel and prove loss records in other countries or regions. In addition, further research is required to increase the 329 amount of available data and upgrade the model through the introduction of additional variables to more precisely predict 330 losses from natural disasters using deep learning algorithms. 331 332 In Stage II, a methodology was proposed to quantify the effectiveness of natural disaster risk reduction projects using cost-333 benefit analysis. Among natural disaster risk reduction projects were implemented in South Korea, information was collected 334 and analyzed for the disaster risk reservoir maintenance project where flood damage occurred before and after completion. To 335 analyze benefits and costs, this study collected and analyzed the loss rate and precipitation from wind and flood damage before 336 and after the maintenance project in the target area and judged the efficiency of the maintenance project. As a result of CBA analysis, in the short term, the loss after the maintenance project was greater than that before the maintenance project. However, 338 this was reversed from 8 years after the maintenance project and the loss amount before the maintenance project was larger 339 than that after the maintenance project. Although it is difficult to expect profits from the maintenance project in the short term, 340 it can be seen that the maintenance project is economically beneficial in the long term (8 years or more). Results and 341 methodology of this study will be helpful for decision making of natural disaster management policy and natural disaster risk 342 reduction project investment. Evaluating the effectiveness of risk reduction through this analysis will lead to drastic investment, 343 which will ultimately reduce the amount of natural disaster risk. However, the study period was relatively short and cases that 344 could be analyzed were limited because all study subjects were from South Korea. In addition, it was assumed that the inflation 345 rate is identical during the study period. Therefore, it is necessary to conduct additional analyses considering various locations 346 venerable to natural disasters in other countries and more realistic financial loss values using a net present value concept. 347

Conclusion 348
Due to increasing threats to life and property from natural disasters, a variety of risk mitigation activities are being carried out 349 extensively to reduce these threats. Economic analysis of natural disaster risk mitigation effects is becoming increasingly 350 important due to limited public budget and economic feasibility. Therefore, in this study, a framework for developing a natural 351 disaster loss prediction model based on a deep learning algorithm for predicting natural disaster losses was presented and a 352 methodology for quantifying the effect of natural disaster reduction through cost-benefit analysis was presented as a case study. The cost-benefit analysis was conducted on the disaster risk reservoir maintenance project that occurred before and after the 364 completion of the flood damage. As the result, it was difficult to expect profits from the maintenance business in the short 365 term. However, in the long term (more than 8 years), it was found that the maintenance business was economically profitable. 366 Results and methodology of this study could be used as a guideline for decision-making of natural disaster management 367 policies and investment in natural disaster risk reduction projects. This study an also be used as a reference for application to 368 other types of loss. The suggested methodology can also be used to support the current knowledge framework.
Code and data availability. 371 The data presented in this research are available from the corresponding author by reasonable request.