Dynamic Updating of Building Loss Predictions Using Regional Risk Models and Conventional Post-Earthquake Data Sources

Abstract

Earthquakes can cause widespread damage to the built environment, disrupt the function of many residential buildings to provide safe housing capacities and thus, potentially induce severe long-term societal consequences. Rapid recovery significantly improves the short-term resilience of communities after an earthquake. However, time pressure and scarce information on the severity and the spatial distribution of damage complicate the decision-making. Therefore, early damage estimates are produced using regional earthquake risk models with rapid earthquake intensity data and typological building vulnerability functions. While the precision of the former depends, amongst other issues, on the density of seismic network stations and the region-specific geological knowledge, the typological classification of buildings often involves attribution models correlating exposure data, such as building height and age, with typological seismic vulnerability classes. Typological attribution models are approximate and locally add to the uncertainties resulting from the average representation of buildings forming one building class. Employing probabilistic machine-learning tools, the continuous inspection data inflow is leveraged to dynamically update initial regional earthquake risk predictions by updating simultaneously the functions that govern typological attribution and building damage. Hence, while completing inspection of all affected buildings may take several weeks, the limited information becoming available in the first days following an earthquake helps constraining underlying uncertainties. This leads to more reliable rapid estimates of losses of building functions and their respective spatial distribution. The framework is demonstrated on a region in Switzerland subjected to a fictitious earthquake scenario.

Dynamic post-earthquake updating of regional damage estimates using Gaussian Processes

“…Experts may provide insights into the proportions of building types in the region, for instance, by specifying the evolution of the regional percentage of RC shear wall buildings. In Bodenmann et al [34], latent functions account for variability and correlations associated to the building attribute combinations to incorporate typological information gathered during post-earthquake inspections. By analogy to the multi-class GP classification scheme (Section 2.2), we employ 𝑐 𝑎 independent latent functions 𝐠 = [𝑔 (1) , … , 𝑔 (𝑐 𝑎 ) ] and impose a GP prior on them.…”

Fusing damage-sensitive features and domain adaptation towards robust damage classification in real buildings

“…Hence, supervised ML approaches that require large amount of labeled data cannot be directly applied. Recently, the combination of a limited number of buildings for damage assessment at larger scales has been proposed as an alternative to overcome the lack of historic labeled data [38][39][40].…”