Investigating the Efficiency of Neuro-Fuzzy Inference System Models in Landslide Susceptibility Mapping (Case Study: Sardarabad Watershed, Lorestan Province)

Landslide is introduced as one of the most important natural dangers in mountainous regions and induces heavy life and financial damages in life of mankind (Bui et al, 2012). Landslides in Iran are taken into account as one of the most common natural dangers, too and annually, induce many life and financial damages to the country and induces high cost to rebuild damaged regions on budget of the country (Zare et al, 2011). In view of undesired effect that this phenomenon has on social, economical and natural systems, knowing talented regions of slide is very necessary in level of the country. In view of dominant territory on the region, geology, being mountainous of Sardarabad watershed, non- principle building of roads and existence of 15 villages having inhabitants, Sardarabad watershed is such as talented watersheds of landslide and in view of landslides than other natural disasters like flood, volcano, and earthquake have capability of more management, so preparing map of zoning of landslide of the region is taken into account as one of important factors and tools in management and control of landslide danger. Using results of this research can prevent to further occur landslide and damage due to it in the region. Goal of the present research is to zone sensitivity of landslide by using adaptive neuro-fuzzy inference system with structure of fuzzy clustering inference system in Sardarabad watershed of Lorestan province.

The study area is the Sardarbad watershed in Khoram-Abad city, Lorestan Province. In order to carry out this research, first maps and basic images were obtained from the Natural Resources Office of Lorestan province. Spacing spots were recorded using aerial photographs (scale of 1: 20000) and field observations (using GPS) and entered the ArcGIS 10.2 software to provide land slide mapping in the area. Information layers for slope, altitude, gradient, distance from the road, distance from fault, distance from the waterway, rainfall, lithology and land use (as the affecting factors for landslide occurrence) were provided based on the basic maps (topography, geology and satellite imagery). The land use information layer was prepared from the Office of Natural Resources and Watershed of Lorestan Province. The geologic map and the distance from the fault are also provided by the geological map of the region (scale of 1: 100,000). The waterway layer was extracted from the topographic map of the study area. The gradient map was provided by topography map (1: 25000) and its digitizing was done in ArcGIS10.2 environment in three classes (concave, flat and convex).
In order to determine the amount of rainfall in the basin, all stations have been used with 30 years data. Due to the importance of rainfall in the occurrence of landslide, the rainfall map was prepared using rainfall data from existing stations from 1981 to 2011. In the Sardarabad watershed, the land use map was extracted from the ETM satellite data and completed by field operations (according to the report of the Natural Resources Department of Lorestan Province). In total, 109 sliding positions were identified using the database of the Office of Natural Resources and Watershed Management of the province, Google Earth images and field observations in the study area. After determining the effective factors, the mentioned maps and providing the data bank, landslide susceptibility map was obtained using the Anfis Model in ArcGIS (SAGA-GIS).

Based on the frequency ratio equation, the obtained weights were applied in 10 effective factors in landslide occurrence. Finally, using this equation, the level of correlation between the landslides and the effective factors (with a cell size of 30 meters) were obtained and was considered as the input weight of the Anfis model. After providing the weighted maps based on the model's equations, the landslide sensitivity map was obtained.
Finally, the landslide sensitivity map was obtained from the ANFIS method. Based on the natural fractures, landslide sensitivity map was classified to four classes (low, moderate, high sensitivity, Too much). A rock curve was used to obtain a suitable regional model for Sardarabad watershed. As the points are closer to the top and left (closer to 1), the model result is more appropriate and the predictions of model is closer to reality (Yesilnacar, 2005).
Overall, the division of 0.9-0.1 is excellent, 0.9-0.8, very good; 0.8-0.7, good; 0.6-0.7, moderate and 0.6-0.5; weak; (Zhu et al., 2009; Yesilnacar,2005). For this purpose, among the 109 identified landslides, 32 cases have been prepared for the model evaluation. According to Rock's results, the surface area under the curve of the study area using the adaptive neuro-fuzzy inference system is 89.11%, which indicates that the model lies in a very good class.

In this study, a nero-fuzzy model was used for landslide hazard zonation. The results of the Roc-Curve assessment in gaussian membership function showed that the surface under the curve is equal to 0.891. This means that the model sensitivity (89.1%) is correct. Also, due to the fact that the surface area under the curve ranged from 0.8 to 0.9, the accuracy of the model was in very good class. Overall, the study area has a high potential for landslide occurrence and without preventing and managing, these landslides annually cause damage to the road, residential areas and other resources in the region.
Geological factors, geomorphologic characteristics and the network of the canals are constant, and the only way to prevent their damage is bypassing these areas and not provoking these areas. Road factors and land use are one of the most effective factors in the occurrence of landslides in the area. They have more management ability. By constructing a road based on the conditions of the ecosystem and preventing the construction of unconventional roads and proper use in these areas, it is possible to prevent from their movement in the study area. The obtained map showed that the adaptive neuro-fuzzy inference system model has a high performance to provide the landslide sensitivity map.