Image-based preliminary emergency assessment of damaged buildings after earthquake: Taiwan case studies

Abstract

Structural collapses that occur during earthquake disasters and subsequent aftershock periods often result in high numbers of fatalities. Rapid preliminary building evaluations are necessary to determine whether buildings are safe for continued occupation. Currently, structural damage is evaluated manually in on-site, expert inspections that are time-intensive, uncertain, inconsistent, and heavily reliant on the experience of the expert team. Therefore, TL-EfficientNet, which combines EfficientNet and Transfer Learning, was used in this research to create an image recognition system for use in an Emergency Assessment Damaged Building Model (EADBM). The integration of hybrid models and building safety assessments presents a pioneering approach to replace manual damage assessments. This system classifies images of structural members into four damage levels based on Taiwan's emergency assessment code, then calculates a structural damage index (SDI) based on the assessed level of damage to columns and structural walls. The accuracy and overall F1 score obtained by the proposed model were both high, with the accuracy of validation results reaching 97% and 91% for columns and structural walls, respectively. Furthermore, the model was applied to two actual damaged building case studies, with the highest accuracy and overall F1 score for columns reaching 90% and 95%, respectively. The SDI value indicates that the first case study was correctly classified as dangerous, while the second case study was categorized as restricted. This model has been proven to overcome subjective and inconsistent results associated with manual assessment, achieving rapid and accurate preliminary assessment in post-earthquake building inspections and safety evaluations.