Landslide susceptibility assessment using a novel hybrid model of statistical bivariate methods (FR and WOE) and adaptive neuro-fuzzy inference system (ANFIS) at southern Zagros Mountains in Iran

Abstract

The aim of landslide susceptibility mapping (LSM) is to produce the most important and basic information required for overall landslide disaster planning and mitigation. Different statistical bivariate methods such as frequency ratio (FR) and weights-of-evidence (WOE) have been widely used for LSM. Although results of these aforementioned statistical methods are generally acceptable, however, they can be improved further by fine tuning the conditioning factor’s classes. The purpose of this paper is to overcome some drawbacks of the bivariate models by developing a novel hybrid method using adaptive neuro-fuzzy inference system (ANFIS) and statistical bivariate methods (FR and WOE) in geographical information system. The provinces of southern Zagros Mountains (Iran) are chosen as a case study to implement the proposed method. First, landslide inventory map was produced using various data source such as historical landslides locations, remote sensing images and land surveying techniques. Second, the inventory data were divided into a ratio 70:30 for training and testing the models. Third, twelve landslide conditioning and triggering factors (such as altitude, slope, aspect, plan and profile curvatures, distance to roads, distance to streams, distance to faults, rainfall, seismicity, land use and lithology) were selected and categorized in two groups consisting of numerical and nominal values. Then, each conditioning factor was classified and the weight of each class was determined by using FR and WOE models. The outputs of individual statistical and hybrid methods were applied to determine nominal and continuous numerical data, respectively. In the hybrid approach, the calculated weights of each class were allocated to the center of each class, and the rest of the weights were determined by ANFIS. Landslide locations which were not used in training the models were used for validation. The produced susceptibility maps were validated and compared using area under the curve (AUC). Results indicated that predictive performances of hybrid models (FR-ANFIS and WOE-ANFIS) have better performance than the statistical models (FR and WOE). The AUCs of success rates are 85, 86, 87 and 87% and the AUCs of predicted rates are 82, 82, 85 and 84% for FR, WOE, FR-ANFIS, and WOE-ANFIS, respectively. Additionally, frequency ratio plots (FRP) and seed cells area index SCAI methods were also employed for a second round of validation. The FRP and SCAI graphs illustrated higher performance for the hybrid LSM models. The proposed approach can be applied in other types of natural hazard modeling such as flood, land subsidence, and gully erosion.