Evaluating the impact of climate change on hurricane wind risk: A machine learning approach

Abstract

In the residential sector, hurricane winds are a major contributor to storm‐related losses, with substantial annual costs to the US economy. With the potential increase in hurricane intensity in changing climate conditions, hurricane impacts are expected to worsen. Current hurricane risk management practices are based on the hurricane risk assessment without considering climate impact, which would result in a higher level of risk for the built environment than expected. It is crucial to investigate the impact of climate change on hurricane risk to develop effective hurricane risk management strategies. However, investigation of future hurricane risk can be very time‐consuming because of the high resolution of the models for climate‐dependent hazard simulation and regional loss assessment. This study aims to investigate the climate change impact on hurricane wind risk on residential buildings across the southeastern US coastal states. To address the challenge of computational inefficiency, we develop surrogate models using machine learning techniques for evaluating wind and rain‐ingress losses of simulated climate‐dependent hurricane scenarios. We collect historical hurricane data and use selected climate variables to predict changing hurricane attributes under climate change. We build the surrogate loss model using data generated by the existing fragility‐based loss model. The loss estimation of synthetic events using the surrogate model shows an accuracy with a 0.78 R‐squared value compared to Hazard U.S. – Multi Hazard (HAZUS‐MH) estimation. The results demonstrate the feasibility of utilizing surrogate models to predict risk changes and underline the increasing hurricane wind risk due to climate change.

1. INTRODUCTION

Hurricanes pose significant threats to coastal communities, with wind damage being a major contributor to overall losses. According to Congressional Budget Office (2019), hurricane winds account for over 40% of storm‐related losses in the residential sector, causing $14 billion in expected annual costs to the US economy. As climate change continues to influence global weather patterns, there is growing concern about its potential impact on hurricane intensity and frequency, which could lead to increased risks and damages. This study aims to investigate the climate change impact on hurricane wind risk on residential buildings across the southeastern US coastal states, including consequential wind‐induced rain‐ingress damage, using machine learning techniques to develop surrogate models for efficient hurricane simulation and loss estimation.

The global temperature is expected to continuously increase due to climate change (IPCC, 2018; Xu et al., 2018), and future hurricane profiles have been predicted to change accordingly (Balaguru et al., 2023; Hallegatte, 2007; Holland & Bruyère, 2014; IPCC, 2013; Snaiki et al., 2020; Vecchi et al., 2008). Global warming has an impact on hurricane activities such as intensity, frequency, rainfall, and translation speed. Studies project increased hurricane intensity and design wind speeds under various climate change scenarios (Elsner et al., 2008; Emanuel, 2008; Knutson et al., 2010; Oouchi et al., 2006). These findings have prompted recommendations for enhanced design wind speeds (Mudd et al., 2014; Rosowsky et al., 2016), with regional analyses predicting significant increases in design wind speeds and return period precipitation rates across coastal areas of the United States (Esmaeili & Barbato, 2021; Orcesi et al., 2022). Climate change is also expected to affect hurricane‐related precipitation, a key factor in wind‐induced rain‐ingress damage. The warmer air holds more moisture, which induces higher precipitation (Coumou & Rahmstorf, 2012; IPCC, 2013). Studies project significant increases in tropical cyclone‐related rainfall under climate change scenarios, with intensification particularly pronounced for hurricane‐strength storms and extreme events (Knutson et al., 2019; Mudd et al., 2017; Reed et al., 2022). This increased rainfall, coupled with wind damage to building envelopes, could lead to more severe combined losses.

Research shows varying findings regarding hurricane frequency and translation speed under climate change. While studies largely agree on an increasing trend in the proportion of high‐intensity storms (Bender et al., 2010; Bengtsson et al., 2007; IPCC, 2013; Knutson et al., 2020, 2015), with observations indicating a 25%–30% increase in category 4 and 5 hurricanes per 1°C warming (Holland & Bruyère, 2014), there remains uncertainty in hurricane translation speed trends, with conflicting findings of both decrease (Kossin, 2018) and increase (Kim et al., 2020) in global measurements. This study will examine the sensitivity of hurricane risk to translation speed while monitoring changes in storm intensity distribution under climate change.

While studies indicate the potential impact of climate change on hurricane characteristics and hazards, a full picture of the future hurricane risk, connecting changing hurricane attributes and the consequent hurricane risk to a broader region, is needed to assist in the development of effective regional risk mitigation strategies. Few studies focus on the potential change in the risk (e.g., expected hurricane losses). Since hurricane loss distribution is a more direct measure of risk compared to hurricane hazard characteristics, investigating the climate change impact on hurricane losses would provide more useful information for risk management. Studies demonstrate significant climate change impacts on hurricane‐induced losses, considering multiple hazard factors such as wind damage, storm surge, and sea level rise (Bjarnadottir et al., 2014; Frame et al., 2020; Wang & Rosowsky, 2018). Recent research has expanded to assess specific vulnerabilities and risks to coastal infrastructure and power systems (Park, 2021; Snaiki & Parida, 2023a). These findings highlight the need to consider climate change in structural design and risk management.

To meet this requirement, efficient tools for accurately estimating hurricane hazards and resulting losses are essential. While existing studies focus on smaller regions, the development of efficient assessment tools would enable large regional studies that can provide a more thorough understanding of regional climate change impacts. Current regional hurricane loss assessment models such as the HAZUS‐MH (Vickery et al., 2006) and the Florida Public Hurricane Loss Model (FPHLM) (Pinelli et al., 2008) conduct detailed load and resistance comparisons on structural components to evaluate exterior physical damage and the consequent wind penetration. The process requires iterations to assess the final damage state with balanced structure internal pressure. These high‐resolution models can be used to predict the consequence of a single event well. However, these models need a more strict designation of initial and boundary conditions and more computation time, which is attributed to model complexity, making them unsuitable for assessing hurricane risk, which requires a large number of simulations. Without an adequate future hurricane risk assessment, built‐infrastructure hurricane risk management practices continue based on the risk assessment with stationary hurricane occurrence assumption. The design wind load specified in ASCE 7–16 was derived based on long‐term expected hurricane statistics and did not consider possible future climate conditions. Such practice might underestimate the hurricane risk in the changing climate conditions and put the nation's built environment at risk. Therefore, to effectively inspect the hurricane risk in changing climate conditions, conducting climate‐dependent hurricane simulations and efficiently estimating the consequent losses for a vast region risk assessment is valuable.

To address these challenges and to effectively inspect the hurricane risk in changing climate conditions, this study employs machine learning methods to develop surrogate models for climate‐dependent hurricane simulation and efficient regional loss estimation. Specifically, artificial neural networks (ANNs) are applied to develop surrogate loss estimation models for efficiently assessing regional hurricane wind and rain‐ingress losses, including consequential rain‐ingress damage from simulated hurricane scenarios across the southeastern US coastal states.

2. CLIMATE‐DEPENDENT HURRICANE PARAMETER MODELING

2.1. Stationary and nonstationary hurricane parameters

To simulate hurricane scenarios for future hurricane risk assessment, this study adopts the approach proposed by Lin and Cha (2020) with a few modifications in nonstationary hurricane parameter modeling. In this hurricane simulation model, we simulate hurricanes from their landfall with parameters including annual occurrence rate, approach angle, translation speed, central pressure difference (CPD), radius to the maximum wind, and filling rate constant. Then, we use transition matrices of translation speed and approaching angle to simulate subsequent hurricane tracks, with the filling rate constant determining the decay of hurricane intensity.

We classify the hurricane parameters as stationary or nonstationary and model differently depending on the stationarity classification. The stationary hurricane parameters—including annual occurrence rate, approach angle, translation speed, and filling rate constant—are assumed to follow the distribution derived from historical observations under current climate conditions. Recent studies broadly support these stationarity assumptions: Global tropical cyclone frequency has remained remarkably steady (Sobel et al., 2021), track changes show significant regional variations with low confidence in systematic shifts (Nakamura et al., 2017). Translation speed trends lack consensus (Gong et al., 2022), and while some changes in decay rates have been observed (Li & Chakraborty, 2020), regional analyses indicate the traditional constant filling rate model remains effective in many areas (Zhu et al., 2021). These parameters exhibit either minimal trends, high uncertainty in detected changes, or strong regional dependencies that make stationary assumptions appropriate for engineering purposes.

The nonstationary hurricane parameters are affected by climate change. The CPD is assumed to be nonstationary in changing climate conditions. Recent assessments and modeling studies strongly support this assumption. Satellite data analysis shows significant upward trends in very intense hurricanes globally (Kossin et al., 2020), and the (Intergovernmental Panel on Climate Change [IPCC], 2023) projects higher peak winds and lower central pressures under continued warming. The World Meteorological Organization's assessment also confirms likely increases in the intensity of the strongest storms by the late 21st century (Walsh et al., 2016). In Lin and Cha (2020), two hurricane parameters, namely, CPD and the ratio of high‐intensity hurricanes to all hurricanes (RHH), were considered nonstationary. In this study, to avoid double counting the climate change effect on hurricane intensity, the RHH is not modeled as a separate hurricane parameter, and only the CPD is considered nonstationary. Furthermore, the radius of the maximum wind speed is modeled as a function of CPD (Vickery & Wadhera, 2008) and thus is considered a nonstationary parameter. Recent research has strengthened this modeling choice, with new analyses demonstrating robust relationships between cyclone intensity and radial size (Sun et al., 2022) and studies showing that stronger storms tend to have different wind field structures (Chan & Chan, 2015). The site‐specific annual mean CPD (denoted as MCPD) of hurricanes making landfall at the coastline of a study region is used for climate‐dependent hurricane modeling. The increase in CPD projection will correspondingly lead to an increase in the frequency of high‐intensity hurricanes, as suggested by several studies (IPCC, 2023; Knutson et al., 2020; Kossin et al., 2020; Vecchi et al., 2021; Bhatia et al., 2019; Walsh et al., 2019). We predict the MCPD under climate change scenarios using ANNs with selected climate variables. We project the difference in the MCPD relative to the current climate conditions and use it to shift the mean of the CPD distributions. In each simulation, we sample the number of simulated hurricanes from the Poisson distribution with a stationary annual occurrence rate in each study region, and for each hurricane, we simulate its CPD at landfall from the shifted CPD distribution, which will then determine the intensity of the hurricane.

We use historical hurricane events from 1944 to 2016 to construct the probabilistic models of stationary and nonstationary hurricane parameters under current climate conditions. We obtain hurricane data from the International Best Track Archive for Climate Stewardship (IBTrACS v03r10) data set published by Knapp et al. (2010). The IBTrACS data set provides storm coordinates, central pressure, and wind speed every 6 h as well as at landfall, and we use the storm track information to derive the hurricane parameters for hurricane simulation. In this study, the southeastern US coastal region is divided into four study regions for deriving the site‐specific probabilistic models of stationary hurricane parameters and current conditions of nonstationary hurricane parameters to simulate hurricanes on specified coastlines. Basic information on study regions is provided in Table 1. Note that the total exposure comprises the exposure of wooden and concrete masonry single‐family residential buildings in the study regions extracted from the HAZUS‐MH based on the 2010 census data. The nonstationary hurricane parameters for future hurricane simulation are predicted and applied to all study regions due to the constraint of data availability. Training neural networks for time series projection for nonstationary parameters involves greater complexity and parameter adjustments compared to simple statistical modeling for stationary parameters. Therefore, the nonstationarity is modeled using all study regions' data and then aggregated with the region‐specific modeling parameters. The nonstationary parameter projections are validated in the following section by comparing the CPD projections with existing literature, confirming that the predictions are within a reasonable range.

TABLE 1.

Study region information.

Regions	Coastlines of hurricane simulation	Length of coastlines (km)	Area of regions (km2)	Total exposure (US $1B)
Region 1	Texas (TX) state	595	685,841	2,367
Region 2	Louisiana (LA), Mississippi (MS), and Alabama (AL) states	993	378,770	1,089
Region 3a	Florida state by the coast of the Gulf of Mexico (FL_G)	1,173	146,843	1,684
Region 3b	Florida state by the coast of the Atlantic Ocean (FL_A)	809
Region 4	Georgia (GA), South Carolina (SC), and North Carolina (NC) states	974	360,137	2,560

2.2. Climate‐dependent hurricane parameter prediction

We conduct climate‐dependent hurricane simulation using climate change scenarios proposed by the IPCC (2013) and IPCC (2014). The IPCC‐projected climate scenarios are characterized by four representative concentration pathway (RCP) scenarios, the projection of greenhouse gas concentration. Four pathways are proposed for climate‐related research: RCP2.6, RCP4.5, RCP6.0, and RCP8.5. The number associated with RCP represents the radiative forcing in a unit W/m², quantifying the difference between sunlight absorbed by the earth and the energy radiated back to space (Shindell, 2013). Higher radiative forcing results in higher sea surface temperature and might induce higher intensity hurricanes. Among the four projected pathways, the most extreme one, RCP8.5, which represents the greenhouse gas emissions continuously increasing throughout the 21st century, is considered for climate‐dependent hurricane parameter prediction in this study. This study primarily adopts the RCP 8.5 scenario because it provides a worst‐case, upper‐bound scenario for assessing hurricane wind risks under extreme climate conditions. While RCP 8.5 is considered less plausible regarding emission trajectories, it remains widely used for hazard and risk studies to understand potential high‐impact outcomes. This choice ensures comparability with prior studies and offers valuable insights for decision‐makers when planning mitigation strategies. However, results derived from this scenario should be interpreted as upper‐bound estimates rather than the most likely future outcomes, with limitations acknowledged in the context of its assumptions. The global SST increment projection based on concentration‐driven Coupled Model Intercomparison Project Phase 5 (CMIP5) simulation is used. For the RCP8.5 scenario, global mean SST is projected to increase by 3.1°C with the lower bound and upper bound being 2.1°C and 4.0°C, respectively, for the long‐term projection (2081‐2100) using 39 Atmosphere‐Ocean General Circulation Models (AOGCMs) from the CMIP5. For simplicity, the expected value, lower and upper bound of global SST, are assumed to linearly increase to its long‐term projection relative to the reference period, 1986 to 2005.

The nonstationary hurricane parameter, MCPD, under climate change scenarios is modeled as a time series using the ANN, specifically, a nonlinear autoregressive network with exogenous inputs (NARX ANNs), with selected climate variables. Compared to a usual feedforward back‐propagation ANN, the NARX connects the output to the input layer to achieve recurrent dynamic modeling. The recurrent output is used in the time series regression, and the external input is used to achieve dynamic feedback of the network. The procedure of using the selected climate variables to predict nonstationary hurricane parameters is shown in Figure 1. Two categories of climate variables are considered, which are SST variables and relative humidity variables. Monthly values of both SST and relative humidity data are reported on a 2° × 2° grid for the period between 1944 and 2016, where SST variables are derived from the Extended Reconstructed Sea Surface Temperature (ERSST) Version 5 data set published by the National Centers for Environmental Information (Huang et al., 2017) and relative humidity variables are derived from International Comprehensive Ocean‐Atmosphere Data Set (ICOADS) Release 3.0 provided by Freeman et al. (2017). This study computes the annual average, maximum and minimum of the spatial mean, maximum and minimum monthly values for both SST and relative humidity for six basins, including the tropical Atlantic Ocean, the nontropical Atlantic Ocean, the tropical Indian Ocean, the nontropical Indian Ocean, the tropical Pacific Ocean, and the nontropical Pacific Ocean. Thus, we obtain 54 variables for each SST and relative humidity. Two other variables, annual mean SST differences between the tropical Atlantic Ocean and the tropical Indian Ocean and between the tropical Atlantic Ocean and the tropical Pacific Ocean, are also considered variable candidates because of their potential influence on vertical wind shear, which is directly related to the formation and intensity of the resulting hurricanes (Latif et al., 2007). We conduct variable selection to identify the most influential variables within those 110 climate variables for nonstationary hurricane parameters.

FIGURE 1.

Schematic view of variable selection and nonstationary hurricane parameter prediction.

To depict the trend of future hurricane parameters, essential climate variables that predict the nonstationary hurricane parameter, MCPD, should be identified. To ensure the identification of nonlinear relationships, this study utilizes feature importance in gradient boosting, as introduced in Hastie et al. (2009) to select the climate variables, which is different from Lin and Cha (2020). The details of the gradient boosting algorithm can be found in Friedman (2001, 2002) and Mason et al. (2000). This study first establishes a gradient boosting model for the prediction of MCPD from all climate variables. Fivefold cross‐validation mean square error (MSE) is used to determine the hyperparameters of the model. In fivefold cross‐validation, the training data are randomly separated into five portions. In the Kth iteration, the Kth fold of data is held as the test data, and the remaining data are used to train the model. The cross‐validation error, the average of the test error of each holdout test data, is then used to evaluate the model performance. The hyperparameters considered for tuning the ensemble model are the maximum number of splits, the learning rate, and the number of trees. The maximum number of trees decides the number of individual models stacked up in the final model. The maximum number of splits decides the depth of each tree, which specifies the level of variable interaction in a tree. The learning rate or shrinkage reduces the contribution of individual trees and hence slows down the learning process. Based on the minimum cross‐validation error for MCPD prediction, the selected maximum number of splits, learning rate, and the number of trees are 5, 0.0005, and 1500, respectively. After fitting the model, the variables are sorted in descending order based on the relative importance, and the number of selected climate variables are determined by using the elbow point on the relative importance plot. Twelve climate variables are selected for MCPD prediction, and the selected climate variables are shown in Table 2.

TABLE 2.

Selected climate variables for mean central pressure difference (MCPD) prediction.

Selected climate variables

Annual mean of SSTD between TA and TI Annual mean of NA regional mean SST Annual minimum of TI regional minimum SST Annual minimum of NI regional minimum RH Annual maximum of TP regional minimum SST Annual minimum of NA regional maximum SST Annual minimum of TA regional minimum RH Annual mean of NP regional maximum RH Annual maximum of NP regional minimum SST Annual maximum of NA regional mean RH Annual minimum of NP regional mean SST Annual mean of SSTD between TA TP

Note: The scope of the above basins is detailed in Lin and Cha (2020).

Abbreviations: SST = sea surface difference; SSTD = SST difference; RH = relative humidity; TA = tropical Atlantic; TI = tropical Indian Ocean; TP = tropical Pacific Ocean; NA = nontropical Atlantic; NI = nontropical Indian Ocean; NP = nontropical Pacific Ocean.

NARX network attributes, including training algorithm, activation function in the hidden layer, number of hidden layers, number of hidden nodes in each hidden layer, and number of steps of feedback delay, are also investigated to find the most suitable network configuration for each NARX model. There are 96 combinations of network setups considered. A network ensemble is adopted to improve the model performance, and nested cross‐validation is adopted to find the most suitable network configuration for the ensemble model for the time series prediction. As discussed in Varma and Simon (2006), nested cross‐validation can provide a closely unbiased estimate of the true error.

To implement this procedure, the observation time series of MCPD is divided into six sections to form five pairs of training and test sets. For the 73‐year‐long MCPD observations, the first five sections contain 12 observations, and the last one has 13 observations. In the first pair, the first section (years 1 to 12) is used to train the network, and the second section (years 13 to 24) is used to test the model. In the second pair, the first two sections (years 1–24) are used to train, and the third section (years 25–36) is used to train, and so on. This procedure is an expanding window approach, and a discussion of this approach can be found in Hyndman (2016) and Bergmeir et al. (2018). Then, the performance of the ensemble model with a certain network configuration is determined based on the average MSE of the five test sets.

The corresponding climate variables and hurricane parameters are predicted using the IPCC projected global SST. As shown in Figure 1, we first predict the selected climate variables using ensemble NARX I models with the IPCC projected global SST as the input variable. The predicted climate variables are then used as the input variable to predict the MCPD using the NARX II model. The 5‐year moving average of prediction of MCPD for mean RCP8.5 scenario (solid line), as well as the upper bound (dashed line) and lower bound (dash‐dotted line), are shown in Figure 2. The shaded area represents the prediction interval representing the parameter uncertainty estimated using neural network dropout as described in and Gal and Ghahramani (2016), Pearce et al. (2018), Zhu and Laptev (2017). Under the mean RCP8.5 scenario, compared to the current climate conditions (mean value between the years 1944 to 2016), MCPD increases 37% and 49% for the near‐term projection (2050–2060) and the long‐term projection (2090–2100), respectively. Knutson et al. (2015) suggested a 4.5% increase in hurricane maximum wind speed under RCP4.5 in the late 21st century in the Atlantic Ocean based on dynamic downscaling of the CMIP5 model, and Knutson et al. (2020) suggested a global 5% increase in maximum wind speed under RCP8.5 in the year 2055. Given that the wind speed can be approximately estimated by the square root of CPD (Holland, 2008), the CPD prediction in this study is reasonable compared to these previous findings.

FIGURE 2.

Prediction of mean central pressure difference (MCPD) under the mean, upper bound, and lower bound climate change scenario (RCP 8.5) projections.

3. DEVELOPMENT OF ANN‐BASED REGIONAL WIND AND RAIN‐INGRESS LOSS ESTIMATION MODEL

To evaluate the hurricane wind damage of simulated hurricanes, it is necessary to develop a surrogate loss estimation model. In the previous study by Lin and Cha (2020), hurricane wind damage and economic loss were estimated by using HAZUS‐MH. However, due to computational complexity and inefficiency, wind loss estimation using HAZUS‐MH is limited. The computational complexity arises from manual study region selection and input of storm track information, which makes the automatic loss estimation of simulated hurricane scenarios unachievable. Moreover, wind loss estimation using HAZUS‐MH is also highly time‐consuming due to the high‐resolution loss estimation and the large scale of the hazard impact. Therefore, the number of hurricane simulations is limited, which results in a high variation in risk assessment. This study develops a computationally efficient surrogate model for hurricane wind and rain‐ingress loss estimation using ANN to assess hurricane wind risk. The loss assessment process using the proposed ANN‐based surrogate model is shown in Figure 3. The surrogate model takes the hurricane hazard and vulnerability of the built environment as inputs and loss ratios calculated by using regional aggregated loss as output. The loss ratio represents the proportion of losses relative to the total exposure at risk. We evaluate the loss of hurricane wind and rain‐ingress at the census tract level. The hurricane hazard variables include the maximum surface wind speed (3‐s wind gust) and maximum rainfall intensity predicted at the centroid of the census tract, and the vulnerability variables include building type composition, dollar exposure of each building type, and topographic information of the census tract. This study considers building and content losses for the loss assessment, where the building loss includes structure and interior losses.

FIGURE 3.

Hurricane wind and rain‐ingress loss assessment using the artificial neural network (ANN)‐based surrogate model.

We develop the ANN‐based surrogate model for hurricane wind and rain‐ingress loss assessment using the loss data generated by the hurricane wind and rain‐ingress loss assessment model proposed by Pant and Cha (2018, 2019). The use of generated data is due to the lack of real loss data. The model by Pant and Cha (2018, 2019) adopts existing methods (Cope, 2004; FEMA, 2013; Gurley et al., 2005; Vickery et al., 2006) to estimate hurricane wind and rain‐ingress losses of 32 prototypes of residential buildings and evaluates hurricane wind damage, wind‐borne debris damage, and subsequent rain‐ingress damage, by considering the interdependency between the component failure mechanisms. This model considers the dependency between the component failure mechanisms. For example, the roof cover fails by default if the sheathing has already failed. Also, the damage to building components would lead to recalculation of the balance between the building's internal and external pressures, resulting in further wind damage and subsequent higher rain‐ingress damage. Using this model for the affected census tracts, we evaluate mean losses of simulated hurricane events. The wind and rainfall intensities are recorded at the centroid of the affected census tracts, and the computed aggregated losses are used to calculate the regional building and content loss ratios. It is important to note that the model by Pant and Cha (2018, 2019) was validated using actual loss data from two historical hurricanes. The authors compared the predicted mean loss ratios with insurance claim data for these events and found the predicted loss ratios align well with the actual loss ratios from the data. However, they did not provide specific error metrics. Using this model as the foundation for our surrogate model may lead to the propagation of errors. Any inherent uncertainties or biases in Pant and Cha's model could be carried forward into our data set and subsequently into our trained ANN model. This could potentially lead to biased estimates and increased uncertainty in our climate change impact assessment. To address this limitation, future work should focus on conducting sensitivity analyses and implementing uncertainty quantification techniques to provide a more comprehensive understanding of how these potential errors might affect our climate change impact projections.

The synthetic data set has 35 input variables, including recorded 3‐s wind gust, maximum rain rate, surface roughness length, and 32 building type ratios, and has two output variables, including building and content loss ratios. The model development process is shown in Figure 4. To generate the data set for training the model, we generate 1000 synthetic landfalling hurricanes, and correspondingly, 57,460 rows of data are generated to construct the surrogate wind and rain‐ingress loss model. The generated data are split into train and test sets to train the surrogate model, and cross‐validation is conducted on the training data to find the most suitable network configuration. The split ratio of 70% to 30% is used for training to test data sets. Fivefold cross‐validation is conducted within the training data to evaluate the various network configurations and select the optimal configuration. The tested network configurations are shown in Table 3. Based on the cross‐validation performance, we adopt the Levenberg–Marquardt back‐propagation as a network training algorithm and adopt the hyperbolic tangent sigmoid function as an activation function in hidden neurons. The network structure with three hidden layers and five hidden neurons per layer is selected. We choose the square roots of the regional building and content loss ratio as the dependent variables for training the model. The candidate model with the selected configuration is then used for prediction to check if the performance on hold‐out data is acceptable, and the result of prediction versus target values is shown in Figure 5. The model performance on the test data set is comparable to that on the training data set. The R‐squared value and test root mean square error (RMSE) are 0.962 and 0.0420, respectively, for the training data set and 0.9613 and 0.0423, respectively, for the test data set.

FIGURE 4.

Flowchart of wind and rain‐ingress loss estimation artificial neural network (ANN) development.

TABLE 3.

Investigated network setups for the wind and rain‐ingress loss prediction artificial neural network (ANN).

Network attributes	Tested setups
Training algorithm	1) Levenberg–Marquardt back‐propagation 2) Resilient back‐propagation 3) Gradient descent with momentum and adaptive learning rate back‐propagation 4) Scaled conjugate gradient back‐propagation
Transfer function in hidden layer	1) Hyperbolic tangent sigmoid function 2) Log‐sigmoid transfer function 3) Normalized radial basis transfer function 4) Positive linear transfer function (ReLU) 5) Elliot symmetric sigmoid transfer function
Number of hidden layers	1 or 3 hidden layers
Number of hidden nodes	5, 10, or 15 hidden nodes
Output transformation	1) Response = loss ratio 2) Response = (loss ratio)0.5 3) Response = (loss ratio)0.25

FIGURE 5.

Test of the artificial neural network (ANN)‐based surrogate wind and rain‐ingress loss model prediction.

This study further verifies the proposed ANN‐based wind and rain‐ingress loss estimation model by using the loss data estimated by HAZUS‐MH. We use the proposed wind and rain‐ingress loss model to estimate 328 synthetic hurricane scenarios generated in a previous study (Lin & Cha, 2020). In the previous study, the synthetic hurricanes were simulated at landfall, and the associated losses were estimated using HAZUS‐MH. The property losses estimated by HAZUS‐MH, including building and content losses, serve as the target values for validation in our study. The comparison of the wind and rain‐ingress loss estimations by the proposed model and HAZUS‐MH is shown in Figure 6. While the surrogate model has a tendency to overestimate the loss for low‐wind‐loss cases, a significant correlation between HAZUS estimation and surrogate model prediction is observed in the comparison result. Therefore, we conclude that the proposed surrogate model is sufficient to predict the change in the wind risk under climate change. Note that the developed model only serves the purpose of large‐scale climate change impact risk assessment in this study but does not intend to replace the operational tools.

FIGURE 6.

Comparison of the loss estimations by the proposed artificial neural network (ANN)‐based model and HAZUS‐MH.

4. FUTURE HURRICANE RISK ASSESSMENT CONSIDERING CLIMATE CHANGE

4.1. Overview of hurricane simulation

We simulate synthetic hurricanes for climate scenarios listed in Table 4 to investigate the impact of climate change on hurricane wind risks. Under a stratified sampling framework, hurricanes are simulated at each of five sections of coastline as defined in Table 1. Correspondingly, there are five sets of region‐specific hurricane parameters. We modify the hurricane parameters from (Lin & Cha, 2020). In Lin and Cha (2020), hurricanes are simulated using CPD distributions of high‐ and low‐intensity hurricanes separately with the use of a hurricane parameter defining the ratio of high‐intensity hurricanes. In this study, the CPD distributions are derived using hurricanes of all categories. All region‐specific hurricane parameters listed in Table 5 passed the Kolmogorov–Smirnov goodness‐of‐fit test (K‐S test) with a significance level of 0.05, demonstrating that the probability models used in this study can be employed to generate representative samples of hurricanes. The loss of each region under each scenario is estimated by aggregating losses resulting from hurricanes simulated across all regions. For hurricane simulation, we adopt Latin hypercube sampling to sample hurricane parameters to reduce the required number of simulations, and hurricanes are simulated for a 10‐year span. If a 1‐year time span is used for sampling, a higher sampling number is required to avoid biased results due to the nonoccurrence of hurricanes. Using a 10‐year time span for sampling reduces this issue and allows for a lower sample size. A total of 1000 10‐year span simulations are executed, resulting in 10,000 years of hurricane events simulated for each scenario listed in Table 4, and the required number of simulations is verified in the convergence analysis in a preliminary study. In the 10‐year period of hurricane simulation, the number of landfalling hurricanes follows the Poisson distribution, and the location of the hurricanes is uniformly distributed along the coastline of each study region. The initial state of a simulated hurricane is described by location, translation wind speed, approach angle, filling rate constant, the radius of maximum wind speed, and CPD. We simulate hurricane parameters using statistical models specified in Table 5. The predicted MCPD is used to calculate the MCPD change relative to current climate conditions (i.e., ΔMCPD) and then to update the MCPD distributions of each study region under the changing climate conditions, as shown in Equation (1).

$$MCP{D}_{F,i}=MCP{D}_{C,i}+\mathrm{\Delta }MCPD,$$ (1)

where the $MCP{D}_{F,i}$ denotes projected MCPD for study region $i$ and $MCP{D}_{C,i}$ represents the MCPD for study region $i$ under current climate conditions. To simulate a storm track before and after landfall for the subsequent loss estimation, we model the translation wind speed and the approach angle of the storm with the discrete Markov chain. Past hurricane data are used to derive the transition probabilities. Note that simulating hurricanes at landfall using landfalling hurricane statistics might underestimate hurricane intensity since hurricanes usually have higher intensity before landfall and govern the simulated wind speed. This systematic underestimation is acceptable for evaluating changes in hurricane risk. To predict the accumulated rainfall, hurricanes are simulated starting from 24 h before their landfall. The before‐landfall transition matrices are developed based on historical records, and hurricane intensity is assumed to be constant before making landfall. We model the gradient wind field by the function provided by Vickery et al. (2000):

$$\begin{array}{ccc}\hfill {\mathrm{V}}_g& =& \frac{1}{2}\left(c·\sin \left(\alpha \right)-fr\right)\hfill \\ & & +\sqrt{\frac{1}{4}{\left(c·\sin \left(\alpha \right)-fr\right)}^2+\frac{B\mathrm{\Delta }p}{\rho }{\left(\frac{R_{max}}{r}\right)}^{B}exp\left[-{\left(\frac{R_{max}}{r}\right)}^B\right]},\hfill \end{array}$$ (2)

where $\alpha$ is the heading angle, $f$ is the Coriolis parameter, $r$ is the distance from the storm center to the location of interest, $\rho$ is the air density, $\mathrm{\Delta }p$ is the CPD, $B$ is the pressure profile parameter, $R_{max}$ is the radius of the maximum wind speed adopted from Vickery and Wadhera (2008):

• (i) landfalling hurricanes in the Gulf of Mexico:

  $$\ln \left(R_{max}\right)=3.558;{\sigma }_{lnRMW}=0.457,$$ (3)
• (ii) landfalling hurricanes along the Atlantic coast:

  $$\begin{array}{ccc}\hfill \ln \left(R_{max}\right)& =& 2.556-5.963\times {10}^{-5}\mathrm{\Delta }{p}^2\hfill \\ & & +0.0458\psi ;{\sigma }_{lnRMW}=0.456,\hfill \end{array}$$ (4)

where $\psi$ is the latitude of the landfalling hurricane and ${\sigma }_{lnRMW}$ is the model error.

TABLE 4.

Climate scenarios for hurricane wind risk investigation.

Scenario	Description
1	Current climate condition (1944–2016)
2	2050–2060 under RCP8.5 lower‐bound
3	2090–2100 under RCP8.5 lower‐bound
4	2050–2060 under RCP8.5 mean
5	2090–2100 under RCP8.5 mean
6	2050–2060 under RCP8.5 upper‐bound
7	2090–2100 under RCP8.5 upper‐bound

TABLE 5.

Probability distributions for stationary hurricane parameters.

Hurricane parameter	Distribution.	Distribution Parameter	Region 1 (TX)	Region 2 (LA+MS+AL)	Region 3a (FL_G)	Region 3b (FL_A)	Region 4 (GA+SC+NC)
AOR	Poisson	λ	0.236	0.403	0.333	0.264	0.375
AAG (degree)	Normal	μ	−30.134	−5.372	36.465	−57.512	−2.511
		Σ	35.616	33.301	37.342	41.665	33.577
CPD (mb)	Weibull	U	48.889	55.165	53.804	58.148	51.040
		K	2.732	2.418	3.055	3.212	2.679
TWS (m/s)	Lognormal	Λ	1.352	1.708	1.930	1.700	1.917
		Ζ	0.391	0.547	0.466	0.333	0.501
FRC	Normal	μ	−2.868	−3.280	−3.340	−3.355	−3.120
		σ	0.490	0.508	0.649	0.393	0.732

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; NC = North Carolina.

Note that the simplification of using a symmetric wind field is to reduce the computational cost. To improve the precision of wind speed estimation, more advanced models considering asymmetric hurricane structure (Guo & van de Lindt, 2019; Mattocks & Forbes, 2008; Olfateh et al., 2017) should be utilized.

This study adopts the modified rainfall prediction model from Lin and Cha (2021) to assess the hurricane‐induced rain‐ingress damage due to the hurricane wind damage. This modification is pivotal for addressing the original R‐CLIPER (Marks Jr & DeMaria, 2003) model's limitations in capturing extreme rainfall intensities, employing historical hurricane data and gradient boosting regression. This methodology improves the spatial accuracy of predicted rainfall, enhancing the model's reliability for simulating hurricane‐induced rainfall patterns. With the R‐CLIPER estimated rainfall, this model takes the coordinates and elevation of the location of interest and distance to the coast as additional inputs to predict the rainfall at the centroid of affected census tracts for simulated hurricane events.

Simulated hurricane attributes are summarized in Tables 6, 7, 8. The total number of hurricanes simulated for each study region is shown in Table 6, where NMH denotes non‐major hurricanes (Saffir–Simpson hurricane wind scale category 1 and 2) and MH denotes major hurricanes (Saffir–Simpson hurricane wind scale category 3 and above). Table 7 shows the ratio of the number of major hurricanes to the number of all simulated hurricanes for each scenario. When the total number of hurricanes remains approximately constant, the increased intensity of hurricanes caused by climate change will significantly raise the proportion of major hurricanes. In the most extreme case (scenario 7), the upper bound projection of RCP8.5 scenario in the year 2090 to 2100, it is predicted that over 40% of the landfalling hurricanes along the Atlantic coast of Florida will be major hurricanes. In scenario 4 (RCP8.5 2050–2060 projection), the overall ratio of major hurricanes increases from 7% to 18%, which is a 157% increase in the occurrence probability of major hurricanes. Kossin et al. (2020) found that the probability of a hurricane reaching a category 3 or higher has increased by 49% per decade in the North Atlantic basin from 1979 to 2017. The predictions of this study generally align well with this observed trend, indicating that the projections fall within a reasonable range. Table 8 shows the changes in the mean values of maximum CPD, maximum sustained wind speed, maximum accumulated rainfall, and maximum rain rate under climate scenarios 4 and 5 relative to current climate conditions. Since the predicted change in MCPD is applied to all study regions, the change in hurricane attributes related to the hurricane intensity is relatively uniform across study regions. The maximum accumulated rainfall not only depends on hurricane intensity but also depends on local topography, hurricane tracks, and so forth. Therefore, the change in simulated accumulated rainfall exhibits different trends across study regions.

TABLE 6.

Total number of major and non‐major hurricanes simulated for hurricane wind and rain‐ingress loss estimation.

Region	Scenario 1		Scenario 2		Scenario 3		Scenario 4
	NMH	MH	NMH	MH	NMH	MH	NMH	MH
Region1: TX	2301	58	2256	105	2227	134	2159	202
Region2: A+MS+AL	3658	370	3452	577	3378	650	3219	808
Region3a: FL_G	3127	204	2967	365	2932	399	2779	555
Region3b: FL_A	2385	252	2223	415	2183	454	1975	663
Region4: GA+SC+NC	3503	247	3307	439	3247	503	3060	688
Overall	14974	1131	14205	1901	13967	2140	13192	2916

	Scenario 5		Scenario 6		Scenario 7
	NMH	MH	NMH	MH	NMH	MH
Region1: TX	2063	297	2027	336	1931	430
Region2: A+MS+AL	3044	982	2937	1090	2764	1265
Region3a: FL_G	2609	725	2531	803	2316	1019
Region3b: FL_A	1784	857	1742	895	1526	1112
Region4: GA+SC+NC	2859	887	2820	930	2547	1204
Overall	12359	3748	12057	4054	11084	5030

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; NC = North Carolina.

TABLE 7.

Simulated ratio of the number of MH to the number of all hurricanes for each study region under climate scenarios.

Region	Scenario
	1	2	3	4	5	6	7
Region1: TX	2.5%	4.4%	5.7%	8.6%	12.6%	14.2%	18.2%
Region2: LA+MS+AL	9.2%	14.3%	16.1%	20.1%	24.4%	27.1%	31.4%
Region3a: FL_G	6.1%	11.0%	12.0%	16.6%	21.7%	24.1%	30.6%
Region3b: FL_A	9.6%	15.7%	17.2%	25.1%	32.4%	33.9%	42.2%
Region4: GA+SC+NC	6.6%	11.7%	13.4%	18.4%	23.7%	24.8%	32.1%
Overall	7.0%	11.8%	13.3%	18.1%	23.3%	25.2%	31.2%

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; NC = North Carolina.

TABLE 8.

Percentage increase in mean values of maximum central pressure difference (CPD), maximum sustained wind speed, maximum accumulated rainfall, and maximum rainfall rate distributions of simulated hurricanes.

Region	Maximum CPD		Maximum sustained wind speed		Maximum accumulated rainfall		Maximum rain rate
	Sce. 4	Sce. 5	Sce. 4	Sce. 5	Sce. 4	Sce. 5	Sce. 4	Sce. 5
Region1: TX	28.1%	36.8%	14.2%	18.1%	5.3%	6.7%	15.1%	18.5%
Region2: LA+MS+AL	24.3%	31.9%	12.6%	16.2%	7.2%	7.3%	12.3%	17.0%
Region3a: FL_G	18.7%	27.1%	9.3%	13.3%	3.0%	4.1%	8.8%	12.1%
Region3b: FL_A	23.4%	30.9%	11.7%	15.2%	2.4%	2.3%	13.2%	15.8%
Region4: GA+SC+NC	25.7%	34.4%	13.1%	17.1%	2.7%	3.1%	15.5%	19.7%
Overall	23.8%	31.9%	12.1%	15.9%	4.6%	5.2%	12.7%	16.4%

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; NC = North Carolina.

Freimarck (2022) utilized the same hurricane hazard model used to derive the US design wind speeds and combined it with environmental variables under the RCP8.5 scenario for climate change impact investigation. The prediction indicated that between 2045 and 2055, the region with the largest increase in the 700‐year hurricane wind speed would be Florida, with a projected increment of 5%–7% along the Atlantic coast. In our study, although the region with the greatest increase in maximum wind speeds under the RCP8.5 2050–2060 scenario is Texas, the predicted increase in maximum wind speed along the Florida Atlantic coast is 11.7%. In comparison, the trends observed in the two studies are in agreement. Our research's model predictions can present a credible range of projected increases in wind speeds under similar climatic conditions.

4.2. Validation of wind speed simulation and the change in wind hazards

We verify the hurricane simulation model by comparing the simulated wind speed with the HAZUS predicted wind speed (FEMA, 2013) under current climate conditions. For each county, the simulated 10‐year maximum wind speeds are fitted to an extreme value distribution, and the annual maximum wind speed distribution is then derived using the inverse transformation method. Finally, wind speeds corresponding to different AEPs are calculated. Figure 7 presents the scatter plot of HAZUS predicted annual maximum 3‐s gust wind speeds (Vickery et al., 2006) versus the corresponding wind speeds simulated by this study for annual exceedance probabilities (AEPs) of 1/300, 1/1700, and 1/3000. Note that the AEPs 1/300, 1/1700, and 1/3000 correspond to the mean recurrence intervals (MRIs) of 300, 1700, and 3000 years, respectively, in a stationary climate condition, which is used for design wind speed definitions in minimum building design load standards (ASCE 7). Table 9 summarizes the difference between the two wind speeds, where MPD is the mean percentage difference, SDPD is the standard deviation of percentage difference, and RMSD is the root mean square deviation. N_C represents the number of the counties within the study region of 839 counties that have (1) more than 300 decadal maximum wind speed samples from the 1000 10‐year simulations and (2) the wind speed at a given AEP reaching hurricane levels (i.e., ≥ 74 mph). For the 839 counties in our study region, as the AEP decreases, the wind speeds for each county increase, resulting in a higher N_C. As AEP decreases, MPD and RMSD also increase, indicating a larger prediction difference for rarer events, likely due to the difficulty in accurately predicting extreme events. On the other hand, SDPD measures the dispersion of difference, with a lower SDPD indicating more consistent predictions. Although MPD increases with lower AEP, SDPD slightly decreases, suggesting that while prediction differences grow, their variability decreases, leading to more consistent prediction differences for rarer events. Moreover, the high R‐squared and low p‐value indicate that this hurricane simulation model can robustly predict changes in wind hazard under climate change.

FIGURE 7.

HAZUS predicted wind speed versus simulated wind speed at the county level for annual exceedance probabilities (AEPs) of 1/300, 1/1700, and 1/3000 under the current climate conditions.

TABLE 9.

Summary of errors between simulated wind speed and HAZUS predicted wind speed (units: Mph).

AEP	NC	MPD	SDPD	R‐squared	RMSD	p‐value
1/300	630	3.404	8.972	0.785	10.490	9.32E‐212
1/1700	726	5.234	8.695	0.800	12.693	2.18E‐255
1/3000	728	6.231	8.502	0.801	13.826	1.69E‐256

To compare the impact of climate change on wind hazard, the percentage of changes in wind speeds under the IPCC RCP8.5 scenario relative to current climate conditions is calculated, as shown in Figure 8. The projected changes in the wind speeds under the IPCC RCP8.5 scenario reveal both spatial and temporal variations. For the near‐term period (2050–2060), AEP 1/300 events show moderate increases (5%–10%) in most regions, with higher increases (10%–15%) along the coast. In the long‐term period (2090–2100), these trends intensify, with widespread increases across all AEPs, particularly in coastal regions where increases often exceed 20%. The increases for AEP 1/1700 and 1/3000 events are more extensive and significant in both near‐term and long‐term projections, especially in coastal and some inland areas. The results underscore the necessity for mitigation strategies to address future wind risks, especially for the buildings designed with lower AEPs in coastal areas where the impacts will be most severe.

FIGURE 8.

Percentage changes in annual maximum 3‐s wind gusts corresponding to annual exceedance probabilities (AEPs) of 1/300, 1/1700, and 1/3000 under the Intergovernmental Panel on Climate Change (IPCC) RCP8.5 scenario for near‐term (2050–2060) and long‐term (2090–2100) projections. NED stands for not enough data, which is noted for the counties with the number of data less than three hundred.

4.3. Future wind and rain‐ingress loss prediction

We estimate the hurricane wind and rain‐ingress loss of simulated hurricane events under climate change scenarios using the surrogate hurricane wind and rain‐ingress loss model. The wind and rain‐ingress loss is predicted at the census tract level and aggregated for the county‐ and region‐level estimated losses, as shown in Equation (5).

$$aggregatedloss=\sum _{i=1}^{N_t}{f}_{WL}\left(V_{s,i},R{R}_i,S{R}_i,RB{T}_{j,i}\right)D_i,$$ (5)

where $f_{WL}$ is the ANN‐based regional wind and rain‐ingress loss estimation function; $V_{s,i},R{R}_i,S{R}_i$, $D_i$ are wind speed, rainfall rate, surface roughness, and total dollar exposure for the ith affected census tract; $RB{T}_{j,i}$ is the ratio of the jth building type in the ith census tract and j = 1–32; and $N_t$ is the total number of census tracts affected by the simulated event. The changes in hurricane hazard extreme values and expected wind and rain‐ingress loss under climate scenario 5, the long‐term projection of the IPCC RCP8.5 scenario, are demonstrated at the county level in Figure 9. Table 10 summarizes the percentage change shown in Figure 9 from 10 selected counties across study regions. During hurricane simulation, the decadal maximum values of a 3‐s wind gust, wind and rain‐ingress loss, rainfall, and rain rate are recorded at the county level. The maximum values are then fitted with extreme value distribution, and the corresponding annual maximum values are derived. Figure 10 shows the number of hurricane damage samples used for fitting the extreme value distributions for each county, which indicates the reliability of the extreme value distribution fitting. Counties with more samples have more reliable wind risk estimates. Corresponding to Figure 9, counties with fewer than 300 hurricane damage samples are marked as not enough data (NED) in the figure. There are more wind damage samples closer to the coastline, as expected. The number of samples can be used to indicate confidence in prediction when comparing the change in hurricane risk at the county level.

FIGURE 9.

Percentage change in (A) annual maximum rain rate (annual exceedance probability [AEP] = 0.01), (B) annual maximum accumulated rainfall (AEP = 0.01), (C) annual maximum 3‐s wind gust (AEP = 1/700 or 0.00143), and (D) expected 10‐year accumulated wind and rain‐ingress loss at county‐level under IPCC RCP8.5 long‐term projection.

TABLE 10.

Percentage change in (a) annual maximum rain rate (AEP = 0.01), (b) annual maximum accumulated rainfall (AEP = 0.01), (c) annual maximum 3‐s wind gust (AEP = 1/700 or 0.00143), (d) expected 10‐year accumulated wind and rain‐ingress loss, and (e) maximum wind and rain‐ingress loss at selected counties under IPCC RCP8.5 long‐term projection.

	County	State	(a)	(b)	(c)	(d)	(e)
Region 1	Harris	TX	19.5%	20.2%	15.7%	66.0%	70.0%
	Dallas	TX	20.8%	15.6%	13.9%	94.3%	49.7%
Region 2	Orleans	LA	14.9%	19.6%	16.5%	92.4%	116.8%
	Hinds	MS	14.6%	16.6%	11.8%	88.5%	53.6%
	Jefferson	AL	16.4%	21.2%	14.0%	74.0%	55.8%
Region 3	Orange	FL	17.2%	18.4%	16.0%	54.8%	81.1%
	Miami‐Dade	FL	16.9%	14.7%	15.0%	56.3%	80.7%
Region 4	Fulton	GA	18.1%	18.6%	15.9%	75.9%	61.8%
	Charleston	SC	18.1%	23.8%	17.6%	106.9%	152.9%
	Mecklenburg	NC	15.4%	24.5%	15.6%	91.7%	59.0%

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; NC = North Carolina.

FIGURE 10.

Spatial distribution of the number of hurricanes for each county from a 10,000‐year hurricane simulation under the Intergovernmental Panel on Climate Change (IPCC) RCP8.5 long‐term projection.

Wind gust and rain rate show similar change patterns, as shown in Figure 9A and C, respectively, which is due to the direct relationship between the two variables. The accumulated rainfall depends on topography, storm movement, and so forth. Thus, the changing pattern in accumulated rainfall (Figure 9B) is observed to be different from those in wind gust and rain rate. According to Figure 9D, climate change has a higher impact on the coastal area than on the inland area for the wind risk. The percentage change in hurricane wind and rain‐ingress loss increases significantly along the coastal region under climate change scenario 5. The difference in the individual locations might be the result of differences in the vulnerability of residential buildings, terrain conditions, and increment of rainfall amount. In Table 10, the significant rise in projected hurricane wind and rain‐ingress losses (e.g., Charleston, SC) is likely due to the dependence of percentage change on the baseline value. Charleston shows relatively low risk at present, resulting in a high percentage change. Also, increased wind speeds and the high vulnerability of local infrastructure can contribute to this.

Table 11 shows the expected 10‐year accumulated hurricane wind and rain‐ingress loss and corresponding standard error (SE) for four study regions, and Table 12 shows the percentage change relative to current climate conditions. Furthermore, by comparing Table 10 column (d) and Table 11 (Scenario 5), differences in risk variations across different observation scales can be observed. For the same study region, the risk changes at the county level do not necessarily reflect the risk changes at the state level, and vice versa. The average loss percentage increment for the three selected counties (i.e., Fulton, GA; Charleston, SC; and Mecklenburg, NC) in region 4 is higher than the average increment for the two selected counties (i.e., Harris and Dallas, TX) in region 1 in Table 10, while the overall loss percentage increments for region 1 is greater than for region 4 in Table 12. These disparities highlight that understanding the risk changes at different levels will assist decision‐makers at various levels in risk management and resource allocation.

TABLE 11.

Predicted expected values and standard errors for the 10‐year accumulated losses from hurricane wind damages (US $1 M).

Scenario	Region 1: TX		Region 2: LA+MS+AL		Region 3: FL		Region 4: GA+SC+NC
	Wind and rain‐ingress total loss
	Mean	SE	Mean	SE	Mean	SE	Mean	SE
1	5,882	252	8,861	313	27,117	559	11,413	273
2	8,248	338	11,522	339	33,627	648	15,072	339
3	8,497	392	12,556	380	36,072	722	16,186	373
4	10,321	411	14,589	410	40,355	788	18,046	443
5	11,889	454	17,099	525	46,393	924	21,078	523
6	12,979	527	18,355	515	48,724	983	21,414	494
7	14,288	527	20,583	587	53,058	1075	24,557	567

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; NC = North Carolina.

TABLE 12.

Percentage increase in the expected 10‐year accumulated wind and rain‐ingress losses.

Scenario	Region 1:	Region 2:	Region 3:	Region 4:
	TX	LA+MS+AL	FL	GA+SC+NC
2	40.2%	30.0%	24.0%	32.1%
3	44.5%	41.7%	33.0%	41.8%
4	75.5%	64.6%	48.8%	58.1%
5	102.1%	93.0%	71.1%	84.7%
6	120.7%	107.2%	79.7%	87.6%
7	142.9%	132.3%	95.7%	115.2%

Abbreviations: TX = Texas; LA = Louisiana; MS = Mississippi; AL = Alabama; FL_G = Florida Gulf Coast; FL_A = Florida Atlantic Coast; GA = Georgia; SC = South Carolina; and NC = North Carolina.

The estimated losses are found to increase in all climate change scenarios, even under the lower bound projection of the IPCC8.5 scenario. Compared to current climate conditions, the expected wind and rain‐ingress loss increase by 24.0%–142.9% across the study region in six climate change scenarios. The amount of the change in risk varies by region: Region 1, Texas, has the highest percentage increase in expected wind and rain‐ingress loss for all climate change scenarios. Compared with other regions, region 3, Florida experiences relatively smaller increases in wind and rain‐ingress loss under climate change. This could be attributed to the significant losses that have already occurred due to the current climate conditions. Regions with historically high hurricane risks show relatively moderate percentage increases in risk under climate change scenarios compared to areas with previously lower risks. This trend is due to the already high baseline risk and the existing mitigation measures in these high‐risk areas. Conversely, areas with traditionally lower hurricane risks are projected to experience more significant percentage increases in risk. In most cases, the percentage increase in wind and rain‐ingress loss followed the same trend across study regions for near‐term and long‐term projections, with the order of severity being Region 1 > Region 2 > Region 4 > Region 3 for most of the climate change scenarios. It is worth noting that changes in hurricane wind and rain‐ingress losses are not directly correlated with changes in hurricane hazard intensity. It also depends on the extent of property exposure and building characteristics in different regions and the level of expected losses in current climate conditions.

We have investigated the potential impact of the change in hurricane translation speed on the hurricane risk, and no significant relationship between the translation speed and wind and rain‐ingress losses has been found. For the investigation, hurricane translation speed was increased by ± 10% and ± 30% under climate change scenario 5, the long‐term RCP8.5 projection, which was considered to be an additional climate change scenario in addition to the scenarios listed in Table 4. A decrease in hurricane translation speed could result in a decrease in peak wind gust and rain rate, which results in a reduction of wind and rain‐ingress losses. However, the decline in translation speed could lead to an increase in rainfall total and rain‐ingress total in coastal regions, which results in higher wind and rain‐ingress loss. Therefore, based on the result, no monotonical trend was observed between hurricane translation speed and wind and rain‐ingress losses.

5. CONCLUSIONS

Climate change may worsen the hurricane risk due to the increasing hurricane intensity because of the rising global temperature. Climate‐dependent hurricane risk assessment is essential for developing and evaluating climate change impact mitigation strategies. This study investigates the hurricane wind risk to residential buildings in the eight southeastern US coastal states, specifically wooden and concrete masonry single‐family residential buildings, under the IPCC projected RCP8.5‐related scenarios by conducting climate‐dependent hurricane impact simulation. To address the computation inefficiency of existing risk assessment tools, surrogate models are developed using machine learning methods for hurricane simulation and hurricane impact estimation.

To ensure the model validity, we developed the hurricane simulation and loss estimation models through rigorous validation processes. We trained and validated these machine learning models to ensure robust performance on unseen data sets. For wind speed and hurricane wind and rain‐ingress loss estimation, the models were additionally validated against predictions by a catastrophe model, HAZUS‐MH. These validations confirm the reliability of the models in assessing changes in hurricane risk under climate change scenarios.

The investigation revealed that climate change's impact on hurricane wind risk varies temporally and spatially. Under IPCC RCP8.5 mean projection, the near‐term (2050–2060) and long‐term (2090–2100) expected hurricane wind and rain‐ingress loss relative to current climate conditions increase by 49% to 76% and 71% to 102%, respectively. The highest increase in hurricane wind risk occurs in region 1 (TX), and the least in region 3 (FL). Considering the millions‐to‐billion dollar loss caused by a hurricane event, these risk increases are not negligible. The change in mean CPD is predicted to induce more high‐wind speed hurricanes. We found no monotonic relationship between hurricane translation speed and regional hurricane wind and rain‐ingress losses. Further high‐resolution analysis is needed to evaluate the impact of change in hurricane translation speed on hurricane risk of the individual locations of interest. Although inconclusive about translation speed, poleward migration of cyclones' lifetime‐maximum intensity (LMI) has been observed (Daloz & Camargo, 2018; Kossin et al., 2014) as well as the poleward migration of cyclones genesis location (Shan & Yu, 2020; Studholme et al., 2022).

County‐level hurricane risk analysis indicated highly nonuniform climate change impact within study regions and different regional prioritizations for climate change impact mitigation toward hurricane wind hazards. Within a state, climate change impact varies from coastal to inland counties. Integrating with the conclusions of Lin and Cha (2021), climate change generally has a higher impact on hurricane wind risk to the coastal counties. The discrepancy results from hurricane track attributes, region topography, building type composition, and so forth. It emphasizes the necessity of vast region risk assessment for federal‐ and state‐level resource allocation and risk mitigation planning considering the impact of climate change.

These findings contribute significantly to our understanding of climate change impacts on hurricane risks, providing valuable insights for policymakers, urban planners, and the insurance industry. The spatial and temporal variability in risk changes underscores the need for location‐specific approaches to hurricane risk management. Our methodological approach advances the field by enabling efficient large‐scale risk assessments, which are crucial for developing effective adaptation and mitigation strategies. To simplify and focus on climate change's impact on hurricane risks, we use current inventory and exposure in loss estimation, and changes in building stocks, population, inflation, and so forth, are not considered in this study. It is worth mentioning that the pledges to limit global warming to 1.5°C were made in the 2023 United Nations Climate Change Conference (COP28). Therefore, the hurricane risk analysis for the less severe climate change scenarios should be conducted in the future.

Lin, C.‐Y. , & Cha, E. J. (2025). Evaluating the impact of climate change on hurricane wind risk: A machine learning approach. Risk Analysis, 45, 4378–4396. 10.1111/risa.70042

DATA AVAILABILITY STATEMENT

Public data sources are provided in the article. Data, models, or code generated or used during the study are available from the corresponding author upon reasonable request.