A novel sampling approach for prediction of post-earthquake damage to buildings

Abstract

Predicting damage to buildings using field-based samples is important for crisis management after an earthquake. Because building damage in seismic regions is spatially correlated, neighboring samples will contain similar information. However, if they are treated as independent observations, the total sample set will not provide sufficient information. To prevent this, sample locations must be dispersed throughout the region and the spatial correlation of observations must be taken into account. The current study proposes a two-stage sampling approach to select well-dispersed samples of small size containing substantial information. The sample locations have been chosen using the Halton iterative partitioning method and individual buildings were randomly selected for each location. The selected sample buildings were employed in a kriging regression model to predict the damage ratio. Multiple factors were used in the model as predictor variables to increase the prediction accuracy. The proposed sampling approach was compared to other sampling methods in terms of spatial balance and prediction error using both simulated and real datasets from the 2017 Sarpol-e Zahab earthquake in Iran. The proposed approach provided better results than the other approaches. Although some factors affected the building damage in the actual data, a combination of fault strike angles and seismic intensity measures provided more accurate predictions than other combinations.