Beyond tsunami fragility functions: experimental assessment for building damage estimation.

Area of Science:

• Disaster science
• Computational statistics
• Machine learning

Background:

• Tsunami fragility functions (TFF) are statistical models relating tsunami intensity to building damage probability.
• Research is shifting from statistical estimators to deep learning in disaster science.
• TFFs are valuable for disaster signatures but rarely used for direct damage estimation.

Purpose of the Study:

• Investigate TFF applicability for building damage estimation.
• Compare TFF application methodologies with a novel machine learning approach.
• Evaluate model generalization on out-of-domain datasets.

Main Methods:

• Selected three TFFs and two application methodologies for baseline damage estimation.
• Developed a machine learning method trained on physical parameters beyond TFF intensity measures.
• Tested methods on the 2011 Ishinomaki dataset (Great East Japan Earthquake and Tsunami) in binary and multi-class scenarios.

Main Results:

• Both TFF methods and the proposed machine learning model achieved good binary damage estimation results.
• TFF methods exhibited limitations in multi-class scenarios and out-of-domain generalization.
• The novel machine learning approach demonstrated superior generalization capabilities.

Conclusions:

• TFFs offer a baseline for binary tsunami damage assessment.
• A simple machine learning model trained on expanded physical parameters shows improved generalization for tsunami-induced building damage.
• Further research should focus on enhancing the generalization of damage estimation models for diverse tsunami events.