A Comparison Among ANFIS, MLP, and RBF Models for Hazard Analysis of Rockfalls Triggered by the 2004 Firooz Abad-Kojour, Iran, Earthquake

Abstract

Rockfall hazard is a very common phenomenon mainly occurring in mountainous slopes, coastal cliffs, volcanoes, riverside, and trenches. In the present research, a model is developed for hazard analysis of seismic rockfalls on a regional scale. For this purpose, three models including Adaptive Neuro-Fuzzy Inference System (ANFIS), multilayer perceptron artificial neural network (MLP), and radial basis function artificial neural network (RBF) were utilized. Firooz Abad-Kojour earthquake of 2004 was used as the benchmark and the model base. The rockfall-susceptible zones predicted by ANFIS, MLP, and RBF methods were compared with the database (distribution map) of seismic rockfalls. The results showed a good overlap between MLP-predicted rockfall hazard zones and database (distribution map) of seismic rockfalls. To evaluate the statistical results of ANFIS, MLP, and RBF models, the verification parameters with high accuracy such as density ratio, quality sum (Qs), and Receiver Operating Characteristic Curve were employed. By analyzing the hazard maps and considering the Qs index obtained by ANFIS (26.76) and MLP (49.19), and RBF (13.84), it could be observed that the calculated Qs of MLP were higher than those of ANFIS and RBF. Moreover, based on the obtained value of the area under the curve from ANFIS (0.984), MLP (0.986), and RBF methods (0.884), it is seen that the MLP network, compared to ANFIS and RBF models, provided a higher accuracy in hazard analysis of rockfalls caused by the earthquake of Firooz Abad-Kojour of 2004.