Multi-Variable Models Enhance Tsunami Preparedness

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jun 13 2024

In a paper published in the journal Communications Earth & Environment, researchers highlighted the necessity of comprehensive multi-variable models for effective disaster preparedness and management in coastal communities at risk from tsunami inundation. They critiqued the limitations of traditional univariate fragility functions.

The team leveraged extensive ex-post damage surveys from the 2011 Great East Japan tsunami alongside hydrodynamic modeling and advanced machine learning (ML) techniques. They explored the intricate factors affecting building vulnerability to tsunamis, focusing on hydrodynamic effects during tsunami propagation on land.

Novel synthetic variables representing shielding and debris impact mechanisms were introduced as practical proxies for water velocity, offering a solution for rapid damage assessments in post-event scenarios or large-scale analyses. ML was showcased as a promising tool for addressing vulnerability assessment complexities, providing valuable and interpretable insights.

Related Work

Past research has emphasized the need for multi-variable models to enhance disaster preparedness in tsunami-prone coastal communities, critiquing traditional univariate fragility functions. By analyzing extensive post-2011 Great East Japan tsunami data with hydrodynamic modeling and ML, researchers highlighted that building vulnerability to tsunamis is influenced by factors like shape, construction material, and environmental characteristics.

Relying solely on water depth fails to capture complex damage mechanisms, advocating for additional measures such as flow velocity and momentum flux. Many researchers demonstrated that ML could effectively model these interactions, incorporating novel variables like shielding and debris impacts to improve damage assessments.

Dataset Expansion Analysis

The original Ministry of Land, Infrastructure, Transport and Tourism (MLIT) of Japan dataset contained polygon shapefiles detailing building-scale information on observed damage, categorized into seven classes ranging from no damage to washed away. It included various explicative factors for damage, such as inundation depth, structural type, number of floors, and intended building use.

In this study, tsunami flow velocity (v_c) was indirectly estimated using Bernoulli's theorem, expressed as a function of inundation depth. The dataset was extended by introducing additional variables, including building geometry and coast-related parameters like distance to the coastline, building orientation, tsunami wave direction, and coastal type. The team added synthetic variables to account for hydrodynamic effects, such as shielding and debris impact mechanisms, calculated using a buffer geometry around each building to represent potential barriers and sources of debris.

Hydrodynamic data for the 2011 tsunami were derived from a two-dimensional nonlinear shallow water model coupled with a tsunami-sediment transport model (TUNAMI-STM). The model used a nesting grid system with varying spatial resolutions to optimize computational efficiency and resolution, focusing on Region R6 for detailed analysis. The simulations covered tsunami propagation and inundation for three hours post-earthquake, capturing maximum values for inundation depth and velocity at building locations. The tsunami source model was developed through tsunami waveform inversion, and STM was modeled using a transport formula with uniform grain size.

Multi-variable models were trained using the scikit-learn library to predict damage levels based on input features from the extended dataset. The models' accuracy was evaluated using the relative hit rate and normalized confusion matrices. The analysts conducted a random search to fine-tune hyperparameters and assessed feature importance by considering the relative mean decrease in accuracy.

Tsunami fragility functions with confidence intervals were generated by sampling data points from the extended MLIT dataset and constructing structured datasets for plotting. The study used inundation depth as the primary intensity measure and calculated the probabilities of reaching each damage state to create fragility functions. The researchers iterated the process multiple times to account for predictive uncertainty.

Tsunami Damage Analysis

The study employed an extensive analysis to understand the impact of various factors on predicting damage levels in tsunami-affected areas. Initially, the focus was on evaluating the influence of water velocity-related variables, such as the hydrodynamic model of the event (v_sim) and v_c, and synthetic variables, like shielding and debris impact proxies derived from the extended MLIT dataset.

Through iterative model training and feature combination exploration, the research aimed to discern the incremental improvements in predictive accuracy attributed to each component under consideration. The results highlighted the importance of incorporating direct and proxy velocity information, along with other explanatory variables, to enhance the predictive capability of the models.

The study moved beyond traditional fragility functions, using ML to offer explicit insights into tsunami vulnerability. Fragility functions, with confidence intervals, depicted the probability of damage states based on inundation depth and varied factors like velocity, coastal morphology, and building characteristics.

Hydrodynamic effects and coastal morphology significantly influenced damage predictions, alongside structural attributes and distance from the coast. It highlighted the importance of multi-variable models in accurately assessing tsunami risk, revealing nuanced relationships among factors. While certain trends emerged, like reinforced concrete buildings being less vulnerable, the study emphasized the need for comprehensive modeling to capture the complexities of tsunami damage assessment.

Conclusion

To summarize, the study highlighted the importance of considering various factors reflecting tsunamis' hydrodynamic impact on buildings, from direct velocity measurements to synthetic indicators like shielding and debris impact proxies. These synthetic variables proved effective for rapid damage assessment, especially in post-event scenarios where acquiring extensive velocity data was challenging.

The authors also underscored the limitations of traditional univariate fragility functions in capturing the complexities of tsunami inundations, advocating for ML approaches for more comprehensive vulnerability assessment. However, the lack of comprehensive damage datasets hindered the widespread application of such models, emphasizing the need for standardized data collection and sharing procedures within the research community.