Rock fall Susceptibility Mapping Using Artificial Neural Network, Frequency Ratio, and Logistic Regression: A Case Study in Central Iran, Taft County

Abstract

This study aims at evaluating and comparing the ability of Artificial Neural Network (ANN), Logistic Regression (LR), and Frequency Ratio (FR) methods to generate a rock fall susceptibility map for predicting the probability of rock fall occurrences in Taft County, a central region of Iran. The maps were prepared by assuming an association between rock fall susceptibility and nine factors including slope angle, slope aspect, elevation, land use, lithology, precipitation, distance from faults, distance from roads, and distance from streams. The performance of the methods was evaluated using the area under the operating characteristic curve (ROC), the precision of the predicted results (P), seed cell area index (SCAI) and statistical measures, including sensitivity, specificity, and accuracy. The area under the ROC was calculated for the frequency ratio, the artificial neural network, and logistic regression methods, and it was found to be 0.899, 0.893, and 0.883, respectively. An assessment of the parameter P also showed the high precision of all the three methods (particularly the frequency ratio method) for identifying high-susceptibility areas. It was also found that high-susceptibility classes had low SCAI values in all the methods, while low-susceptibility classes had higher SCAI values indicating acceptable performance of models. Overall, the results showed that the model developed by the FR method has better prediction accuracy than the ANN and LR methods. Decision makers can effectively use the findings of the present study to mitigate the financial and human cost resulting from the rockfalls.