Explainable deep learning powered building risk assessment model for proactive hurricane response

Climate change and rapid urban development have intensified the impact of hurricanes, especially on the Southeastern Coasts of the United States. Localized and timely risk assessments can facilitate coastal communities’ preparedness and response to imminent hurricanes. Existing assessment methods focused on hurricane risks at large spatial scales, which were not specific or could not provide actionable knowledge for residents or property owners. Fragility functions and other widely utilized assessment methods cannot model the complex relationships between building features and hurricane risk levels effectively. Therefore, we develop and test a building‐level hurricane risk assessment with deep feedforward neural network (DFNN) models. The input features of DFNN models cover the meta building characteristics, fine‐grained meteorological, and hydrological environmental parameters. The assessment outcomes, that is, risk levels, include the probability and intensity of building/property damages induced by wind and surge hazards. We interpret the DFNN models with local interpretable model‐agnostic explanations (LIME). We apply the DFNN models to a case building in Cameron County, Louisiana in response to a hypothetical imminent hurricane to illustrate how the building's risk levels can be timely assessed with the updating weather forecast. This research shows the potential of deep‐learning models in integrating multi‐sourced features and accurately predicting buildings’ risks of weather extremes for property owners and households. The AI‐powered risk assessment model can help coastal populations form appropriate and updating perceptions of imminent hurricanes and inform actionable knowledge for proactive risk mitigation and long‐term climate adaptation.


INTRODUCTION
Hurricane events have caused annual billion-dollar impacts in the United States over the recent years (NCEI, 2021;Office of Coastal Management, 2021). Specifically, IPCC (2021) claimed that it is likely that the frequency and intensity of tropical cyclones (especially hurricanes, i.e., tropical cyclones of Category 3 to 5 with maximum sustained surface winds of 74 mph or greater) increased over the last four decades globally. The expected destructive hazards of severe hurricanes increase the risks to coastal households Coastal populations can mitigate the adverse impacts of hurricanes by conducting appropriate proactive measures before imminent hurricanes, such as preparing the emergency supplies (e.g., food, water, and medical supplies) and evacuating to safe areas (Dash & Gladwin, 2007;Lindell & Hwang, 2008;Morss et al., 2016). The appropriate decision making for the prehurricane proactive measures is potentially guided by the accurate assessment of individual households' and property owners' risk levels (Gao et al., 2022;Lazo et al., 2010;National Research Council, 2013). However, hurricane risk assessment at the individual household level is currently not available, and the lack of specific risk information can lead to inappropriate risk perception, insufficient hurricane preparedness, failed responses, and even fatalities (Baker et al., 2012;Burnside et al., 2007;Cole & Fellows, 2008;Lindner et al., 2018;Meyer et al., 2014;Senkbeil et al., 2019;Senkbeil et al., 2020).
Existing studies and practices lacked effective methods for assessing hurricane risks of individual households and property owners, hindering the development of relevant and accurate risk information. Existing methods of hurricane risk assessment widely adopted fragility function-based models to estimate buildings' physical damage and households' economic loss (Khajwal & Noshadravan, 2020;Lin & Shullman, 2017). Fragility functions may not represent the performance of building structures resisting hurricane hazards because their form and parameters' value were based on experimental data (such as wind tunnel experiments) instead of real-world conditions (Vickery et al., 2006). To avoid the limitations of fragility functions, some studies have proposed to utilize statistical machine learning models to assess buildings' hurricane risks with real-world damage datasets and consider buildings' features, environmental features, and meteorology features (Spekkers et al., 2014;Szczyrba et al., 2021). However, these naïve models may not be able to learn the complex and interactive relationships between the high-dimension input features and hurricane risk levels sufficiently (Kim & Yoon, 2018;Kulkarni et al., 2018;Nateghi et al., 2014;Sejnowski, 2020). Deep-learning models, which consist of multilayer neural networks, have the potential to overcome this limitation and advance the performance of naïve machine learning models (Lin & Cha, 2020;Pilkington & Mahmoud, 2016). However, few studies have evaluated the performance of deep learning models in assessing individual buildings' hurricane risk levels, and the lack of transparency and explainability in neural networks may prevent some individuals from trusting the predicted risk outputs (Spiegelhalter, 2020). Therefore, we develop deep feedforward neural network (DFNN) models to assess building-level risks of imminent hurricanes. The risk levels are measured by the probability and intensity of potential physical damage caused by (i) high winds and (ii) storm surges during hurricanes. We focus on these two hazards of hurricanes due to their severe damage and economic losses to the hurricaneprone regions, especially in the Southeastern Coasts of the United States (Pei et al., 2014). The input features of our developed DFNN models include (i) buildings' structural characteristics, (ii) meteorological environmental data, and (iii) surge-related features surrounding individual buildings. To make the DFNN-based risk analysis models transparent and understandable, we employ the local interpretable model-agnostic explanations (LIME), an explainable AI (XAI) model, to identify the features with significant influence on predicting household buildings' risk levels (Ajayi et al., 2020;Ribeiro et al., 2016). We evaluate the DFNN models with three baseline statistical machine-learning models (i.e., decision tree, random forest, and support vector machine) and demonstrated the DFNN models' usefulness in a simulated prehurricane scenario for buildings' risks in Cameron County, Louisiana with the updating weather forecasts in response to Hurricane Delta. Our research can contribute to the knowledge body of hazard risk management by proposing deep-learning-based risk assessment models that provide relevant, timely, and actionable knowledge for coastal households and property owners for more proactive hazard response and climate adaptation.

Hurricane risk assessment at different spatial scales
Hurricane risk has been assessed at different spatial scales, including regions, local communities, and individual buildings (Davidson & Lambert, 2001;Legg et al., 2010). The region-level hurricane risk represents the potential hurricane damage to the whole area from the physical and socioeconomic perspective (Davidson & Lambert, 2001). Specifically, the physical damages of hurricanes can be measured based on (i) the overall characteristics of the local building inventory, that is, the number, type, and general status of buildings in the community (Jain et al., 2005), and (ii) the forecasted hurricane parameters, for example, wind speed and central pressure difference (Lin & Cha, 2020). The community-level hurricane risk assessment is similar to the region-level assessment, but the data on the local built environment and hurricane parameters are more fine-grained (Park et al., 2013). Specifically, the community building inventory features are more detailed, including the overall number of buildings with specific types of components, such as roof coverings and windows . The hurricane risk assessment for individual buildings is based on the buildings' detailed characteristics, such as buildings' structural features and characteristics of the buildings' surrounding environment (Dong & Li, 2017).
The assessment methods of regional and communitylevel hurricane risks generally estimate the potential damage caused by hurricane events by approximating the economic loss based on fragility functions and features of local building inventories (Guikema, 2020). With fragility functions, existing studies have estimated the community's hurricane risks based on the hazard intensity and the community's exposure to the hurricane hazards (Lin & Shullman, 2017;Nofal & van de Lindt, 2020). Similar to the risk assessment at larger spatial scales, existing studies have also assessed the potential hurricane damage to individual buildings by developing fragility functions for estimating the overall building damage or components' damage (Dong & Li, 2017). These studies utilized fragility functions to describe the relationship between the hazard features and the failure of building components or structures (Park et al., 2013). The parameters in the fragility functions were generally calculated based on the historical data of hurricane events, experiment outcomes, and datasets of building features (Masoomi et al., 2018).
Although several studies have analyzed the hurricane risks at different spatial scales and provided valuable insights into hurricane risk mitigation, their limitations were also evident. First, the hurricane risk levels of large spatial regions cannot fully represent the risk levels of households of different housing types (e.g., single-family houses, multifamily houses, or mobile houses), which were the basis of the risk mitigation behaviors of local households (Chatterjee & Mozumder, 2014). Second, because the detailed features of individual buildings and their surrounding environments were rarely considered (Gao & Wang, 2021), the hurricane risk assessment at large spatial scales may only describe the general hurricane impact instead of the damages to specific individual buildings. Additionally, the fragility functions mainly captured the building components' performance in ideal experiment scenarios, while the real-world building performance in hurricane events cannot be fully represented by these fragility functions.

2.2
Building-scale hurricane risk assessment with machine learning models To specify the hurricane risks of individual buildings and overcome the limitations of fragility functions, some research has utilized machine learning models to model the relationships among building features and hurricane risk levels (Spekkers et al., 2014;Szczyrba et al., 2021). Compared to the methodology of fragility functions, the studies with machine-learning models have some advantages. First, the mathematical relationships between the relevant features and potential damages were determined during the model training with the postevent damage datasets. In this way, the machine-learning model can describe the relationships in a way that is much closer to the real-world situation (Pilkington & Mahmoud, 2016). Second, the machinelearning models can include large numbers of variables and then automatically discard those that lack prediction power (Aziz & Dowling, 2019). This advantage of machinelearning models makes it possible for researchers to utilize the collected datasets fully and reduce the extent of theorizing needed to determine suitable independent variables (Aziz & Dowling, 2019). Specifically, DFNN-based models have frequently been developed to predict the potential loss of buildings based on multiple categories of features, such as the meteorological features and buildings' physical characteristics (Pilkington & Mahmoud, 2016;Lin & Cha, 2020).
However, existing studies with machine learning models were still insufficient for the building-level hurricane risk assessment, partly due to the lack of training data and outcome validation measures (Guikema, 2020). Specifically, it was challenging to make the decisionmakers of hurricane risk mitigation understand and use the outcome of machinelearning risk assessments, especially the model accuracy and the delivered uncertainty (Guikema, 2020). Additionally, although model interpretation was critical for evaluating and validating deep learning models, related methods have rarely been applied to hazard-related studies, which may hinder individuals' trust in the prediction outcomes of deep learning models of hazard risk analysis.

DEVELOPING DEEP RISK ASSESSMENT MODELS
This study develops DFNN models of hurricane risk assessment for individual buildings and properties in the Southeastern Coasts of the United States. DFNN models are suitable for the risk assessment because they can learn the nonlinear and complex relationships between static input features (i.e., characteristics of buildings and their surrounding environments) and overall risk levels. The DFNN models can predict the given buildings' hurricane risk levels under wind and surge hazards. To make the prediction process transparent and explainable to the public, we interpret the DFNN models by LIME (Ribeiro et al., 2016).

Developing deep feedforward neural network models
We define the hurricane risk of individual buildings as the intensity levels and the corresponding probability for specific levels of structural damage based on the definition in Hernández et al. (2018). We focus on individual buildings of multiple types, such as single-family houses, multi-family houses, and mobile houses. We develop paralleled DFNN risk assessment models separately for the two major hurricane hazards: wind and storm surge hazards. The input features of the two models have been specified based on existing studies about building damages caused by hurricane hazards. The structures of the two models are illustrated in Figure 1.
The input features are listed in Table 1. Each building is recorded with one numerical/categorical value for each building-related feature. The values of large-area features, that is, the meteorological and surge features, are assigned to individual buildings within their scales. Notably, the meteorological and hydrological features update every 6 hours. The updated value can be input into the risk assessment DFNN models to update the risk outcomes.
• 0.6 mile × 0.6 mile • Updated every six hours Surge-related Features • Inundation, the elevation of water/wave, and the max speed of the storm (NOAA, 2021).
• One record for an individual building. • Static.
The DFNN-based models for both wind risk and surge risk assessment consist of an input layer, three hidden layers, and an output layer. Each hidden layer includes multiple neurons that receive the values of neurons from the prior layer, process the input with activation functions, and transmit the outcomes to the neurons in the next layer with specific weights. We set the number of hidden layers and neurons for each layer with a trial-and-error process used in similar DFNN-based research (Jiang et al., 2013;Leung et al., 2003;Shi et al., 2021). The process includes (i) arbitrarily determining the initial setting of neural network structure (i.e., number of layers and number of neurons in each layer) and (ii) adjusting the setting to improve the performance of the neural network. To specify the number of hidden layers, we initially set the number of neurons in each hidden layer as 60 and gradually add the number of hidden layers (from "no hidden layer" to "multiple hidden layers") and compare the accuracy, precision, and recall of the neural network with the certain number of hidden layers. We finally decided to have three hidden layers for the DFNN models because the model performance does not improve heavily with more hidden layers. The accuracy of models with a certain number of hidden layers is shown in Figure 2.
After selecting the number of layers, we decide the numbers of neurons in the three hidden layers for both wind and surge models, which are set as 100, 50, and 30 after a trial-and-error process. We select the activation function for neurons in the DFNN models by summing the inputs to neurons in the prior layer with weights (Agatonovic- Kustrin & Beresford, 2000). The activation functions in the hidden layer neurons in our DFNN models are set as hyperbolic tangent (i.e., tanh) for the first hidden layer and as rectified linear units (i.e., ReLU) for the other hidden layers and output layer. These activation functions are mathematically convenient and allow the DFNN models to learn the relationships between input features and risk levels more efficiently (Harirchian et al., 2020;Samikwa et al., 2020).
To train the DFNN models, we utilize the Adam optimizer to optimize the weight setting for the connection between neurons in the network layers (Kingma & Ba, 2014). Adam is an algorithm for the first-order, gradient-based optimization of stochastic objective functions based on adaptive estimates of lower-order moments. We select the Adam optimizer because, compared to other optimizers (e.g., stochastic gradient descent), it is computationally efficient, requires little memory, and is suitable for a large volume of training data and model parameters (Kingma & Ba, 2014). Adam stores both the exponentially decaying average of past squared gradients and the exponentially decaying average of past gradients (see Eq. 1).
In Equation 1, t and t+1 are the weight values for a connection between two neurons in t and t+1 round. m t is an exponentially decaying average of past gradients in t round. v t is an exponentially decaying average of past squared gradients in t round. 1 and 2 are the decay rates. is the learning rate, that is, how much to change the model in response to the estimated error each time the model weights update. We can adjust the decay rates and learning rate for the Adam optimizer. In this study, we set their values as default: equals 0.001, 1 equals 0.9, and 2 equals 0.999. In each round of training, the weight matrices in the DFNN model are updated based on Equation 1.
The DFNN models output the potential damage levels of individual buildings under the wind and surge hazards as well as their probability, ranging from Level 0-5 (wind damage) and Level 0-6 (surge damage) respectively based on the existing classification of hurricane hazards' damage levels (Kennedy et al., 2020). Specifically, Level 0 represents no physical damage being done to the building's components, and the higher levels represent more severe physical damage being done to the building's components. The raw outcomes of the DFNN models are values of neurons in the output layer with a softmax activation function. The softmax activation function is suitable for the output layer of neural networks for multi-class classification due to its advantage in generating the probability of the assessment target belonging to each class (Bishop, 1995). The softmax activation function is calculated based on Eq. 2.
In Equation 2, z i and z j represents the raw value of the ith and jth neuron in the output layer, while n represents the number of damage levels (six levels for wind risk assessment and seven for surge risk assessment). The raw value is calculated as the weighted sum of the neuron values in the prior hidden layer. exp () represents the exponential function. With the softmax function, the final output of each neuron in the output layer represents the probability that an individual building's wind or surge risk belongs to specific levels. The number of neurons in the output layer depends on the number of categories of hazard risk levels. Each neuron in the output layer delivers the predicted probability that an individual building is under a certain risk level.
Based on the assessment outcomes, the risk level with the highest probability will be regarded as the potential risk level, which can inform proactive actions at the household level in response to the most-possible damages. For example, the wind hazard model predicts the probabilities of an individual building across risk levels are {0%, 10%, 10%, 40%, 30%, 10%}, among which, Level 3 has the highest value (40%), so the wind risk level of this individual building is Level 3 with a probability of 40%. The prediction of individual buildings' surge risk levels and related probabilities is under the same process.

Model interpretation
To make our DFNN models transparent and explainable to the potential users, for example, coastal households and property owners, we interpret our DFNN models via LIME (Ribeiro et al., 2016), a widely used post hoc XAI method. LIME is essentially a local surrogated model that approximates the predictions of deep learning models. The core of the LIME model can be described in Equation 3: In Equation 3, g is the explanation model for instance x, aiming to minimize the loss L, which measures the distance between the explanation and the prediction of the original model f . x describes the locality (e.g., distance function) concerning given x. Ω(g) describes the model complexity, such as the maximum number of features that the original model (DFNN in this study) may utilize. To utilize LIME for model interpretation, we first select several instances (i.e., the case building) randomly. For each instance building, we generate a new dataset as the building's "neighborhood," in which each "neighbor's" feature values are slightly different from those of the instance building. Then, we utilize DFNN models to predict the risk levels for the instance building's "neighbors." Utilizing the prediction outcomes of the instances and their "neighbors," we train a multivariant model and utilize the parameters of each input feature in the multivariant functions as their importance for manually explaining the DFNN models. With the interpretable model, we can compare the influence of adjusting the value of each input feature on the DFNN models' prediction; then, the critical features can be identified.

Data description
The training datasets for the DFNN models are listed in Table 2, including the building-level damage assessment data, meteorological data, and the datasets of surge-related indicators from several hurricane events. The damage assessment data was from NSF structural extreme events reconnaissance (StEER, 2021). This study utilized the Field Assessment Structural Teams' (FASTs') damage assessment data collected for 2017 Hurricane Harvey, 2018 Hurricane Florence, 2018 Hurricane Michael, 2020 Hurricane Laura, and 2020 Hurricane Sally (Figure 3), including the assessment outcomes for 3736 buildings under wind damage and 959 buildings under the surge damage. The basic information, affected areas, and the number of buildings with damage data for each hurricane event are listed in Table 3. The maximum wind speed reached Category 1-4, and the maximum storm surge reached Category 1-3 during hurricanes (Navy, 2021).

Model training outcomes
We trained the risk assessment model of wind hazard and surging hazard separately. We split the overall datasets into training datasets (80% of overall data) and testing datasets (20% of all the data) randomly. We regarded every thirtyseven records (around 1%) in the training data as one batch for the wind hazard model and ran the training process for 200 rounds. The batch size and training times were determined during the model tuning process, during which we identified that this set of batch size and training times can maintain the model accuracy/loss to a stable and high level. For the model of surge hazard, we regarded eight records (around 1%) in the training data as one batch and ran the training process for 100 rounds. The training outcomes of wind and surge DFNN models are shown in Figure 4. The prediction accuracy of the wind hazard model reached 83.24% after training, and the accuracy value of the surge hazard model was 92.18% after training. The training outcomes indicated that these two models could accurately predict the risk of wind and surge hazards for an individual building based on features of its physical components and environment (i.e., local atmosphere and hydrology environment).

Comparing DFNN models with baseline models
To evaluate whether the DFNN models can more accurately predict the overall risk levels of individual buildings than naïve statistical machine learning models, we compared the prediction performance of DFNN with typical machine learning models, including the decision tree, random forest, and support vector machine (SVM) models. We trained these typical machine-learning models for wind and surge risk assessment with the same training datasets. The parameters of these models were set as shown in the Supporting Information. Specifically, for the SVM model, we set the kernel function as a radial basis function (i.e., RBF). The precision, recall, and overall accuracy of these wind and surge risk assessment models were listed in Wind field data ARA (2021) High-resolution wind speed and gust data were generated based on the NOAA stations' observations.
Storm surge data CERA (2021) High-resolution datasets about the seawater elevation, wave elevation, inundation, and wind speed. The datasets were generated based on NOAA stations' observation of wind field and estimation from the Advanced Circulation Model (ADCIRC). of performance indicators are highlighted with bolder fonts). Because the prediction was multiclass, we calculated the precision and recall for each class (i.e., risk level) separately. The comparison outcomes indicated that the DFNN models outperformed other machine learning models in terms of the precision and recall of the predictions for wind and surge risk levels, and the accuracy levels were closed to the decision tree and random forest. We can also identify the different performances of the DFNN models regarding different risk levels.

TA B L E 3 Hurricane case description and size of building datasets
For the model of wind risk assessment, DFNN's precision tended to be high for Risk Level 1 and Level 5. For the model of surge risk assessment, DFNN's precision was generally at high levels and closed to 1, while the recall levels tended to be low for Risk Level 0 and 3, but the recall levels were high for Risk Level 4 and 6.

DEMONSTRATING THE UPDATING RISK ASSESSMENT WITH HURRICANE FORECASTS
To compute individual buildings' risks with updating weather forecasts, our DFNN models can update the risk assessment outcomes over the hurricane warning stage and intake temporally update data of meteorological and hydrological data. Hurricane forecast information is updated in the National Digital Forecast Database from National Weather Service every six hours (NHC, 2017). Specifically, the updated time on each day of hurricane periods includes 3:00 a.m. UTC, 9:00 a.m. UTC, 3:00 p.m. UTC, and 9:00 p.m. UTC (NHC, 2017).
To demonstrate the process of updating risk assessment outcomes, we applied our models to predict dynamic risks of one case building in Cameron County, Louisiana. We generated the virtual scenario based on the NHC forecasted wind field and CERA surge data of Hurricane Delta. Hurricane Delta made landfall on the two counties between October 9 and October 10, 2020, at Category 2. The input data were generated by NHC and CERA at 3:00 p.m. October 8, 2020, including the forecasted data of meteorological and surgerelated features surrounding the case building at (i) 3:00 p.m. October 9, 2020, (ii) 9:00 p.m. October 9, 2020, and (iii) 3:00 a.m. October 10, 2020. During these periods, the case building was forecasted to be influenced by the hypothetical imminent hurricane. We utilized DFNN models to predict the hurricane risk level of case buildings at each forecast time.

F I G U R E 5 Hurricane risk assessment outcomes for the case building during Hurricane Delta
The prediction outcomes of the updating building's wind and surge risk levels were shown in Figure 5.
For the wind-related risk, the DFNN model regarded Level 1 as the most possible risk level for the case building based on the forecast at 3:00 p.m. October 9, while the risk level changed to Level 2 with high-level probability with the updated hurricane forecast at 9:00 p.m. October 9 and 3:00 a.m. October 10. Comparatively, for the surge-related risk, the DFNN model did not update the building's risk level and kept it at Level 2, and the predicted probability for the surge risk level was generally high and closed to 100%. We utilized LIME to identify the top ten critical features that contributed to DFNN's prediction at each prediction time for wind and surge risk. The lists of critical features at each prediction time are shown in Table 5.
Based on the output of LIME (Table 5), we found that the update of the wind feature forecast caused the change in wind risk levels of the case building. At 3:00 p.m. October 9, the forecasted wind gust was 75 mph, while the value at 9:00 p.m. October 9 increased heavily and became 85 mph. Comparatively, the building's surge-related risk level was relatively stable at all the prediction times, which may be because that DFNN regarded the inundation level as the most important feature that determined the building's surge risks, and inundation did not exist surrounding the case building. The interpretation outcomes can facilitate the household in the case building to conduct response measures. Specifically, the wind-related risk assessment indicated that roof-related features contributed most to the building's wind-related risk level, such as the roof slope, roof cover, and roof shape. The household can prepare for the hurricane wind by strengthening the structure of their building roof and utilizing targeted tools and materials to mitigate the risk caused by weak roof cover. Additionally, to prepare for the potential surge, the household of the case building did not need to concern about the inundation above the ground, but they may need to strengthen their wall cladding to mitigate the potential damage of surges to the building's external walls.

DISCUSSION
This study develops and trains DFNN models to predict building-level hurricane risks and focuses on the probability and severity of potential damages caused by hurricane winds and surges. The research shows that risk TA B L E 5 LIME-identified critical features for risk assessment comes 3:00 p.m., October 9, 2020 9:00 p.m., October 9, 2020 3:00 a.m., October 10, 2020 assessment can be more relevant to the confronted hurricane risks of individual buildings by considering detailed features of building components and the surrounding environment. The DFNN models showed acceptable performance in fitting the unstructured field-observation data and predicting individual buildings' hurricane risk levels. The XAI interpretation for the risk assessment outcomes can potentially facilitate coastal households to conduct targeted risk mitigation measures. Combined with the frequently updated forecasts of meteorological and hydrological features surrounding individual buildings, our models can effectively capture the vulnerable parts in the future hurricane period and update the buildings' hurricane risk levels in real-time. This capacity can help coastal households adjust their risk mitigation measures to fit the confronted risk levels timely. We evaluate the performance of DFNN models by comparing them with ordinary machine-learning models. The comparison results show that our DFNN models have overall higher levels of model performance. This study advances the existing knowledge body of hazard risk assessment in the following ways. Previous studies generally assessed hurricane risk levels at large spatial scales, such as regions, counties, and communities (Davidson & Lambert, 2001;Legg et al., 2010). These assessments of hurricane risks have little capacity on representing the confronted risk of local households or guiding their risk responses. Accordingly, this study assesses hurricane risks for individual buildings based on the detailed features of buildings, local meteorology parameters, and the status of the surrounding hydrology environments. With these detailed features, our risk assessment is more relevant to the confronted risks of individual households and property owners and more useful for their risk mitigation. Previous risk assessment methods calculated the risk levels of buildings based on fragility functions and lacked calibration with scalable real-world datasets (Khajwal & Noshadravan, 2020;Lin & Shullman, 2017). The development of fragility functions made their outcomes potentially different from the real-world situation of buildings' resistance to hurricane hazards. Comparatively, our DFNN models utilize the training data collected from the post-event field observation datasets and can generate predictions that are more closed to realworld conditions than the existing risk assessment methods. Additionally, similar to the human brain, the DFNN-based models are suitable for processing training data with multiple features. Compared to naïve statistical machine learning methods with simple structures, the DFNN models can better mine the potential patterns of the training datasets and fit with unstructured input data (i.e., features related to building status, meteorology, and hydrology). The model comparison outcomes indicate the considerable advancement of DFNN and other deep-learning models in assessing hazard risk with large-volume and detailed prediction features. To improve the model transparency and help the public understand the prediction outcomes, we also innovatively conduct a model interpretation using LIME to explain how the different input features contribute to the prediction outcomes of buildings' hurricane risk levels. Our study proves that it is practical to evaluate the predictions' reasonability based on the interpretation of data processing of deep-learning models. Making the risk assessment more explainable and transparent can potentially improve the public's understanding of the risk assessment process and outcomes and help them form a more appropriate perception of the hurricane risks. Their risk mitigation measures would also be more efficient as a result.
Limitations still exist in the DFNN-based risk assessment models. First, uncertainty still exists in existing weather forecast data, which may affect the accuracy of risk assessment outcomes, especially at the early stages of hurricanes. With the development of forecasting technology and AI-enabled forecast models, the input data for the wind and surge-related features are also expected to be more accurate. Our developed DFNN models have the agility to intake outputs from future forecast models as well; our future work will keep tuning the model with outputs from new weather forecast models and other real-time crowdsourcing data sources (Yao & Wang, 2020;Hao & Wang, 2021). Second, lacking data on the damage conditions of specific building components, our DFNN models only assess the overall level of each building's potential damages caused by hurricane winds and surges. Future studies can improve our DFNN models by calibrating the model to predict specific structural damages, proportions, and functional and economic loss of individual buildings with more nuanced output features and training datasets. Third, existing available training datasets do not include Category 5 hurricane cases due to the lack of reconnaissance filed investigation following the recent Category 5 hurricane in the Southeastern Coasts of the United States, which may affect the models' risk assessment accuracy for more extreme situations. Future studies can keep tuning the model by including building damage data collected from more destructive hurricanes.

CONCLUSION
Climate change has intensified extreme climate and weather events which makes it urgent to provide the relevant and timely risk assessments for households and property owners for motivating proactive risk mitigation and climate adaptation. Our research proposes DFNN models for predicting building-scale hurricane risk levels, showing the potential of deep learning in modeling the complex relationship between hurricane risks and features related to individual buildings and their surrounding built environment. Trained with realworld reconnaissance data, our DFNN-based risk assessment can accurately evaluate individual buildings' hurricane risk levels. We also utilized LIME to make the deep learningbased risk assessment explainable, which transformed the assessment outcomes into actionable knowledge, facilitating the proactive response of individual households and property owners. The research leads to an innovative risk assessment in response to destructive hurricane hazards (i.e., high wind and storm surge) for more equitable risk information and proactive climate extremes response. It tests the usefulness of household-level risk assessment in the face of SLR and severe storm surges, but the methodology is scalable to other types of extreme climate and weather events, such as wildfires, severe winter storms, and tornadoes. In addition to the model's usefulness for households and property owners, local emergency responders and voluntary community response teams also benefit from fine-scale risk assessment for immi-nent climate extremes, so they can provide point-to-point support adequately, especially for evacuating vulnerable populations with special needs. Specifically, our DFNN-based risk assessment and the LIME-based model interpretation can help identify the building characteristics that contribute to different risk levels of buildings when confronting an imminent weather extreme. The fine-granularity outcomes of our risk assessment models can assist in more agile risk mitigation actions and even long-term adaptation in response to climate and weather extremes. It can prioritize the limited retrofitting resources to households at the highest risk, inform the updating of building codes, and provide evidence-based planning tools for relocating people and businesses to safer grounds or investing in social capital to make hazard-prone communities more resilient.

A C K N O W L E D G M E N T S
This material is based upon work supported by the University of Florida faculty start-up funds. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the University of Florida.