A Comparison of Fuzzy Analytic Hierarchy Process, Artificial Neural Network and Area Density in Quantitative Evaluation and Landslide Susceptibility Mapping within GIS Framework (Case Study: Simereh Homiyan Watershed(

Landslide is one of the major geomorphologic processes affecting the landscape's evolution in mountainous areas, which can lead to the catastrophic events (Hattanji & Moriwaki, 2009). Landslide Hazard Zonation (LHZ) is defined as dividing land into several regions and the classification of these areas, based on the actual or probable degree of susceptibility to landslides or other displacements of slopes (Varnes, 1984). For this purpose, the use of new technologies such as Geographic Information System and computational intelligence algorithms can be useful in preparing accurate maps of landslide zonation. In this study, three methods including Fuzzy Analytic Hierarchy Process (FAHP), Artificial Neural Network (ANN) and Area Density were employed as the representatives of two groups of decision making frameworks, i.e. non-deterministic computational and statistical methods in quantitative assessment. The efficiency of these approaches were examined, as well. This study was conducted in a part of Simereh Homiyan watershed with a large number of registered landslides to present an appropriate model for determining a geologic based efficient approach in landslide zonation. . 2.1. Study Area The study area is located in the west of Lorestan province with a geographical longitude of 47˚ 22' to 47˚ 52' and a geographical latitude of 33˚ 34' to 34˚ 9'. The study region with an area of ​​128,000 hectares has an average altitude of 1620 meters above sea level (Baharand and Surri, 2015). 2.1.1. Information Layers In order to determine the effective criteria for landslide susceptibility mapping, we should use the factors that are able to solve the problem, and take into account the local and general situation of the region and existing constraints. The information layers used in this study, based on a consultation with watershed managers and geology experts were topographic layers, vegetation, relative humidity, depth of soil fracture, proximity to river, fault and road, and geological sensitivity layer. 2.1.2. Layer standardization Measurement of criteria in the form of information layers takes place with a wide range of scales. Therefore, the values ​​in the various layers must be converted into the units that are comparable and proportionate. In order to evaluate the accuracy of the results, the data of the landslides recorded in the study area were used based on simple random sampling in the areas with mass movements. Generally, 50% of the landslide records were used to train the employed techniques and the remaining were utilized in validation analysis. Due to the number of layers provided for this study, the corresponding calculations were performed for each method. In sum, according to the type of training and experimental data used, the accuracy of the three methods used can be considered convincing; however, at the same time, the validity of the ANN and area density is significantly better than that of the FAHP. It could be due to the poor performance of the FAHP in pair comparisons. In other words, the opinions of the experts that should be involved in, impose some uncertainties in FAHP process. The highest accuracy for the ANN can be resulted from the proper functioning of this approach, which has been able to produce appropriate outputs by using appropriate training data and finding the internal relations of the target values ​​and inputs. Based on the results, the northern regions of the study area are prone to occurrence of landslide. On the other hand, the values of the verification parameters confirm the higher accuracy of the results of both neural network and surface area densities, respectively, with the overall accuracy of 0.73 and 0.71. Based on the comparison of the results, the FAHP method with a total accuracy of 0.58 shows better performance. In general, the neural network method based on the validation statistics used include general accuracy, user authenticity and manufacturer's accuracy which are respectively with values of 0.73, 0.8, and 0.59 and have the highest accuracy.