Catastrophic models for natural disaster risk assessment have been developed in the public sector at a national level as the demand for quantification of the financial losses due to natural disasters has increased. Among the others, the new Multi-Risk and Multi-Risk Assessment Method (MATRIX) in Europe, the Central American Probability Risk Assessment (CAPRA) in South America, the Florida Public Hurricane Loss Model (FPHLM) in Florida, United States, The RiskScape in New Zealand and the US Federal Emergency Management Agency (FEMA) Hazus Multi-Hazard (MH) are demonstrative examples. Mentioned catastrophic models simulate, evaluate, and quantify direct and indirect losses caused by natural disasters such as earthquakes, floods, and windstorms that occur at the national or regional level. The models are seen as essential means to improve formulating responses and mitigation strategies. Thus, efforts are being made for constant development and advancement (Kim et al., 2016). However, the aforementioned models are ineffective when applied to other countries and regions since they are based on specific data such as the vulnerability of buildings, topography, and disaster patterns for the relevant country or region. In the private sector, vendors such as RMS (Risk Management Solution), Risk Frontiers, EQECAT, and AIR (Applied Insurance Research) develop and provide their own models for each natural disaster (Sanders et al., 2002; Kunreuther et al., 2004). The vendors’ models are widely used by insurers and reinsurers around the world as vital tools for cumulative risk management, insurance underwriting, as well as holding and establishing business strategies through the risk assessment of loss due to natural disasters. Nonetheless, the models’ usefulness is limited to the respective countries, and the evaluation for other regions is not possible.

Furthermore, the presence of financial constraints lessens the chances for small and medium-sized companies to introduce and use catastrophic models, as they must pay high annual fees. Additionally, since the black box module of the models determines the natural disaster loss evaluation, verifying whether the loss calculation result is appropriate appears problematic. Moreover, although the above models may be referential for the market, the plan of their comparison is required because each company’s portfolios, risk appetite, and capital are different. Therefore, an in-house model, which considers the vulnerability and exposure of the country or the region, other than the commercialized models, is required to evaluate the risk of natural disasters of a country or a company.

As mentioned above, there is a strong possibility that the amount of loss from natural disasters and the frequency and severity of natural disasters will increase. Naturally, a more sophisticated and refined in-house loss prediction model that reflects the complexity and uncertainty of natural disaster damage as well as the circumstances and characteristics of each user is required. Recently, the fourth revolution technology, consisting of Internet of Things, artificial intelligence, robots, unmanned transportation, and big data, has been introduced to a wide range of industries, and its effectiveness has been recognized (Gledson and Greenwood, 2017). Thus, a natural disaster loss assessment model using a deep learning algorithm, which can be widely used for analysis and prediction, appears to be a suitable solution (Kim et al., 2020a). The demand for deep learning algorithms burgeons progressively as they can adapt and respond quickly and precisely to changes in the environment owing to constructing and integrating big data easier (Kim et al., 2021a; Guha et al., 2022). In the past, various studies have been conducted aiming to develop and advance natural disaster risk assessment models using various deep learning algorithms (Kim et al., 2020b; Khosravi et al., 2020; Shane Crawford et al., 2020; Yi and Zhang, 2020; Kim et al., 2021a; Al Najar et al., 2021; Moishin et al., 2021; Berezina and Liu, 2022; Kaur et al., 2022; Klepac et al., 2022; Irwansyah et al., 2023). For example, deep learning technology was applied for the detection of damages after natural disasters such as tornadoes and landslides (Shane Crawford et al., 2020; Yi and Zhang, 2020; Al Najar et al., 2021; Haggag et al., 2021). Similar techniques were applied to analyze changes in coastal areas after natural disasters such as windstorms. In the described study, the depth of the sea was precisely estimated by simulating remote sensing images through deep learning (AI Najar et al., 2021). Yi et al. (2020) predicted the risk of landslides due to earthquakes by analyzing satellite images using the Convolutional Neural Network (CNN) technique (Yi and Zhang, 2020). Kim et al. (2020b) use a deep learning algorithm to assess the loss before and after an earthquake at a regional scale. Ren et al. (2022) used a deep-learning technique to predict typhoon tracks accurately. In addition, when tornados occur, the wind speed is estimated through the analysis of the satellite image data on trees using the CNN method, and damage to buildings is classified (Shane Crawford et al., 2020). Kaur et al. (2022) analyzed satellite images based on the convolutional neural network to detect the damages induced by severe hurricanes. Another previous study adopted deep learning technology to predict and manage the risk of flooding, one of the major natural disasters (Khosravi et al., 2020; Moishin et al., 2021). Moishin et al. (2021) projected flood occurrence utilizing past daily rainfall data based on a hybrid deep learning algorithm. The described algorithm was a combination of long short-term memory networks, CNN, and advanced flood risk management (Moishin et al., 2021). Moreover, a flood sensitivity map was created using CNN based on a vast amount of data concerning the past flooding area, such as elevation, geology, vegetation, distance from rivers and roads, land use, rainfall, slope, curvature and soil moisture (Khosravi et al., 2020). Kim et al. (2021a) proposed a model for expected maintenance and repair costs using a deep neural network (DNN) algorithm as a method for the economic management of facilities. The mentioned study used a DNN algorithm based on factors affecting loss from natural disasters and loss of data from natural disasters for learning, testing, and verifying the new model. As a result of comparison with other models, the developed model was considered to be effective in predicting maintenance and repair costs (Kim et al., 2021b). Furthermore, when loss of data on construction projects caused by natural disasters has been predicted through deep learning, the prediction error was lower than the statistical model generally used for prediction (Kim et al., 2021a).

As stated above, various deep learning technologies were applied in order to develop and advance natural disaster risk assessment models due to higher prediction accuracy compared to the traditional approaches (Kim et al., 2020b; Berezina and Liu, 2022; Kaur et al., 2022; Klepac et al., 2022). Described studies exposed the shortcomings of existing models or showed advances in innovative technologies. Additionally, models designed via deep learning technologies are far more sophisticated and reliable than the existing ones and can serve as a cornerstone of better natural disaster risk mitigation strategies. Therefore, this study suggests a framework for developing a model for building loss due to typhoons through a deep learning algorithm.

Internet of Things, artificial intelligence, robots, unmanned transportation, big data, nanotechnology, etc., are applied to various industrial fields to prove their value and lead innovation in related fields (Cabello et al., 2021; Kim et al., 2022b). The amount of damage is expected to increase due to changes in the frequency and severity of natural disasters. Furthermore, to embrace the complexity and uncertainty of natural disaster damage as well as the various situations and characteristics of each user, a more sophisticated and efficient prediction model is essential. Moreover, an increasing usage of Internet of Things, Information and Communications and sensor technology for the management of buildings and facilities is expected to enrich complex and diverse big data. Also, for such an extensive analysis, the introduction of deep learning technology is rudimental. Deep learning algorithm is a part of machine learning and is applied to various industries and research, such as type classification of input data, regression, recognition, and prediction fields. A deep learning algorithm has components such as output layers, input layers, hidden layers, an activation function, neurons, and weights. Countless combinations of these components are possible, enabling applying to numerous data depending on the combination (Allison et al., 2019; Cabello et al., 2021). Deep learning algorithms are divided according to their arrangement and treating methods, such as Deep Neural Network (DNN), Convolutional Neural Network (CNN), Generative Adversarial Network (GAN), Recurrent Neural Network (RNN), and Auto Encoder (AE). For example, DNN allows construction of artificial neural networks with innumerable structures according to the number of hidden layers and combinations of components, enabling the modeling of complex nonlinear interactions. Due to these characteristics, DNN performs excellently in analyzing and predicting nonlinear data with high uncertainty and complexity like natural disasters (Kim et al., 2022a; 2021b).

Therefore, this study proposes a framework for developing a predictive model based on a DNN algorithm for analyzing big data generated by typhoons (Figure 1). Besides, this study aims to improve future natural disaster prediction through the prediction model based on the DNN algorithm, to reduce the monetary loss from natural disasters by more advanced loss prediction, and ultimately to contribute to the reduction and prevention of natural disaster risk.

This study proposed a framework for building a predictive model using the DNN algorithm, one of the deep learning algorithms, to predict the loss of buildings caused by typhoons. For data collection, data on loss of buildings that occurred by typhoon Maemi was collected from the insurance company, and variables were collected. For the development of an optimal DNN algorithm model, different network scenarios and hyperparameters learned the input and output variables, and a combination with a minimum cost function was identified. Table 3 shows the learning results according to the network structure scenario and dropout. Among the learning results, the scenario in which the MAE and RMSE values ​are the minima was decided as the final one. It was found that, in general, a smaller loss function was exhibited when the dropout was 0 than when the dropout was 0.2. Furthermore, as the number of hidden layer nodes increased, MAE and RMSE became smaller, and both MAE and RMSE had the smallest values in the scenario where the number of hidden layer nodes equaled 600-600-600. Therefore, the dropout was 0, the network structure scenario and the number of hidden layer nodes were determined to be 600-600-600, as depicted in Figure 6. The final network structure and hyper-parameters are summarized in Table 4.

To verify the problem of overfitting the final DNN algorithm model, this study simulated the validation and test data and associated the predicted results. For these purposes, the regression analysis (R ²) results of the optimized DNN structure for the training, validation, test and total datasets are plotted in Figure 7. The regression results indicate that the optimized DNN model can accurately capture the relationship between the independent and dependent variables. The consistent correlation coefficients (R ² > 0.8) across the different datasets (training, validation, testing, and combined) suggest the model’s generalizability and reliability in predicting the dependent variable. Moreover, comparative verification with the MRA model was performed for additional verification. As shown in Table 5, MAE, RMSE and R ² values ​of the validation and test data were calculated and compared to verify the final DNN model. As a result of the simulation of validation data and test data, the MAE, RMSE, and R ² values of the validation data were 0.133, 0.222, and 0.825, and the test data values were 0.131, 0.215, and 0.832, respectively. Thus, the overfitting of the DNN model was insignificant since the alteration between the result values ​of the verification data and the test data was unimportant. Table 5 shows the comparison of the results of the DNN model and the MRA model. As a result of comparing the two models (i.e., DNN/MRA (%) for test data sets), the DNN model showed 82.5% lower MAE and 64.8% lower RMSE than the MRA model, which leads to the conclusion that the DNN model displays a lower prediction error rate than the MRA model. Furthermore, it was observed that the R ² values for validation and test cases were notably higher in the DNN model compared to the MRA model. Thus, the higher R ² values in the DNN model suggest that it captured the underlying patterns and relationships in the data more accurately than the MRA model. Wei and Yang (2021) also noticed the better performance of the backpropagation algorithm compared to the MRA model. Therefore, it was observed that the DNN algorithm model has higher reliability and robustness in predicting the loss of buildings due to typhoons. To sum up, the non-parametric DNN algorithm is more suitable than the parametric MRA model for non-linearity and data with high complexity and uncertainty, such as natural disaster losses (Schmidhuber, 2015; Nguyen et al., 2019).

Risk managers and facility managers in public or private sector will be capable of predicting losses due to typhoons or developing an in-house model through the DNN model proposed in this study. As well, since the DNN algorithm model exhibits a lower prediction error rate than the existing model, the loss prediction will be more accurate and reliable. Therefore, risk managers or facility managers will be able to reduce expected losses and establish investment strategies to prevent damage. Into the bargain, the model can be useful as a guideline for preparing an emergency reserve in reflection of the expected typhoon loss for sound and continuous financial planning. Furthermore, establishing a new strategy for transferring the risk of loss by using this model is possible. For example, the prediction done by the DNN model can be applied as a criterion for judging the appropriate level of insurance premiums currently insured according to the estimated risk of loss and can be utilized to enhance the adjusting of insurance coverage. Besides, insurance companies can use it to manage the accumulated risk of typhoons. Hence, it is conceivable to predict the economic loss due to typhoons through the DNN model and framework and eventually reduce the loss from typhoons by upgrading the damage prevention and risk transfer strategy.

Additionally, the developed models and framework can be adopted as reference materials for loss analysis and forecasting in public and private sectors in other industries and academia, as well as for the prediction of other natural disasters. Abnormal climate and natural disasters caused by global warming, such as windstorms, tornados, heavy snow, heavy rain, hail and droughts, are occurring all over the world and have adverse effects on many fields, so many studies are conducted to overcome their repercussions (Carreira-Perpinan and Hinton, 2005; Kwon, 2019). The proposed model was built on numerous introduced indicators such as the distance from the coastline, maximum wind speed, number of basement floors, floors, occupancy, construction type, and total property value. Initial result that the risk indicators are closely related to the loss from natural disasters was re-examined. Furthermore, expected rise of the frequency and severity of natural disasters such as typhoons due to the global warming is anticipated to have a significant antagonistic effect on the management of buildings and facilities in the future (Kim et al., 2020a). Besides, Information and Communications (ICT), Internet of Things (IoT), and sensor technology have recently become widely used for building and facility management. Thus, similar research can be conducted to analyze and utilize big data accumulated in these devices and technologies. Therefore, as the need for advanced prediction of loss owing to typhoons will grow greater and sources of information become gradually more accessible, the results and framework of this study can be recognized as rudiments of meeting these needs.

Nevertheless, in this study, only data on loss of buildings caused by typhoon Maemi of a major insurance company in South Korea were accepted. As a result, additional research is desired to secure supplementary data on loss due to other typhoons from other insurance companies and other countries and comparative verification to upgrade the proposed model. Furthermore, discovering additional variables such as topographical indicators according to the characteristics of natural disaster losses is essential not only for improving the prediction error of the model but also in order to expand the model’s capacity of predicting human casualties apart from damage to buildings due to typhoons. Although the DNN algorithm created in this study presents reduced prediction error compared to the existing model, it has the disadvantage of the weights and correlations between nodes being black boxes. The problem stays unresolved as the deep learning algorithms unilaterally provide through a black box without screening the basis and process of deriving prediction results, which may raise questions about the objectivity and dependability of the results and models. Therefore, further research is required by introducing XAI (eXplainable AI). XAI is also an artificial intelligence technology that explains AI algorithm classification, prediction, causality, and decision-making to users (Chollet, 2017). Consequently, in the future study of deep learning algorithm models, it is crucial to gain users’ trust and improve understanding of models and data through a glass box rather than a black box by using XAI.