Predicting the probability of rockfalls occurrence caused by the earthquake of Changureh-Avaj in 2002 using LR, MLP, and RBF methods

Rockfalls pose serious threats to human beings, structures, and lifelines. In the present research, a model was developed to predict the possibility of seismic rockfalls on a regional scale. For this purpose, three models including logistic regression (LR), a multilayer perceptron artificial neural network (MLP), and a radial basis function artificial neural network (RBF) were employed. Bivariate logistic regression is a multivariable statistical method that provides a mathematical model using independent variables to predict the occurrence probability of a given phenomenon at a certain location. Although artificial networks of RBF and MLP are pretty similar, there are some structural differences in the components between these two neural networks. The earthquake of Changureh-Avaj in 2002 was used as a basis and benchmark for the model presented in this work. The sustainable zones predicted by LR, MLP, and RBF methods were compared with a database (distribution map) of seismic rockfalls. The results showed good overlap between RBF-predicted rockfall susceptibility zones and database (distribution map) of seismic rockfalls. Besides, in order to evaluate the statistical results of LR, MLP, and RBF models, the verification parameters with high accuracy such as density ratio ( Dr ), quality sum ( Qs ), and Receiver Operating Characteristic Curve ( ROC ) were used. By analyzing the susceptibility maps and considering the Qs index obtained by LR (2.94) and MLP (3.482), and RBF (4.344), it could be observed that the Qs of RBF were higher than that of LR and MLP. Moreover, based on the obtained value of AUC from LR (0.859), MLP (0.910), and RBF methods (0.956), it is seen that the RBF method provided a higher accuracy in predicting the probability of rockfalls occurrence caused by the earthquake of Changureh-Avaj in 2002 compared to LR and MLP methods.