Using Artificial Neural Networks to Assess Earthquake Vulnerability in Urban Blocks of Tehran

Abstract

The purpose of this study is to assess the vulnerability of urban blocks to earthquakes for Tehran as a city built on geological faults using an artificial neural network—multi-layer perceptron (ANN-MLP). Therefore, we first classified earthquake vulnerability evaluation criteria into three categories: exposure, sensitivity, and adaptability capacity attributed to a total of 16 spatial criteria, which were inputted into the neural network. To train the neural network and compute an earthquake vulnerability map, we used a combined Multi-Criteria Decision Analysis (MCDA) process with 167 vulnerable locations as training data, of which 70% (117 points) were used for training, and 30% (50 points) were used for testing and validation. The Mean Average Error (MAE) of the implemented neural network was 0.085, which proves the efficacy of the designed model. The results showed that 29% of Tehran’s total area is extremely vulnerable to earthquakes. Our factor importance analysis showed that factors such as proximity to fault lines, high population density, and environmental factors gained higher importance scores for earthquake vulnerability assessment of the given case study. This methodical approach and the choice of data and methods can provide insight into scaling up the study to other regions. In addition, the resultant outcomes can help decision makers and relevant stakeholders to mitigate risks through resilience building.