Landslide Susceptibility Mapping using Bayesian Model: A Case Study of some Regions of Talar Watershed, Mazandaran Province

Mass movements are a type of morphodynamical phenomena that are usu ally related to various factors and occur on slopes in mountainous areas. Every year, damages caused by landslides lead to financial loss and death over the world (Fathi et al., 2018). According to previous studies, landslides cause 17% of the world's natural disasters. Mortality rates from 1903 to 2004 for different continents, including Asia, United States, Europe, Africa, and Australia were 29%, 39%, 30%, 1%, and 1%, respectively (Kohorest et al., 2005). Over the last three decades, several research studies have been conducted on landslide susceptibility mapping using different methods for developing their classification. All mapping methods are classified into five different groups including landslide distribution analysis, qualitative, statistical, deterministic, and frequency analyses (Vanvestern, 2003). Iran is prone to landslide phenomena due to natural conditions such as mountain topography, high tectonic and seismicity activity, geological and climatic diversity. Environmental factors affecting the occurrence of landslides are slope degree, aspect, plan curvature, elevation, land use, lithology, distance from the road, river and fault, topographic moisture index, slope length index or sediment transport and vegetation cover (Silakhori et al., 1400).

Firstly, 134 points were identified using the Iranian landslide database and field survey in order to prepare a landslide susceptibility map using Bayesian theory. In the present study, 12 factors were used including elevation, slope and aspect, plan curvature, distance from the fault, road, and river, land use, geology, sediment transport index (STI), topographic wetness index (TWI), and vegetation cover. Then, topographic (1:50000), geology (1:100000), soil maps, and satellite imagery (Indian Remote Sensing) for 2012, were prepared and classified in ArcMap and ENVI environments. In order to evaluate Bayesian theory in landslide risk analysis, the relative performance curve of the relative efficiency of variables (ROC) was applied. This index is used to determine the accuracy and efficiency of the model (Egan, 1975; Williams et al., 1999). The area under the ROC curve represents the predicted value by describing its ability which accurately estimates the occurred events (landslide occurrence) and its non-occurred events (non-landslide occurrence). Therefore, the area under the curve is used as the model accuracy assessment. In the present study, 134 points of the landslide’s phenomena were used for modeling (70%) and accuracy assessment (30%) (Pourghasemi et al., 2013).

The results of the factors affecting the occurrence of landslides using Bayesian theory in the study area showed that most of the landslides occurred in the class of 15-30 degrees with a weight of 1.28, which is also reported by Eracanoglu and Gokceolu (2004). The reason is that human intervention on these slopes causes more susceptibility (Yalcin et al., 2011). The study of aspect shows that most of the landslides occurred in the west and south directions with a weight of 2.21 and 2.41, respectively, which is confirmed by the results of Shams and Alizadeh (2019). The results of plan curvature showed that most of the landslides happened with a weight of 1.1 in convex slopes, which is close to the results of Pourghasemi (2013). Convex slopes usually have the highest landslides which is also reported in previous studies (Vanvewsten et al., 2003; Jaaferi et al., 2014). The altitude of 500-1000m covered by sandstones, chile and siltstones showed a significant relationship with a high number of landslides (Ayalew & Yamagishi, 2005). Among different land uses, rangeland (weight of 3.48) indicated the most significant relationship in landslide occurrence, which is reported by Shams and Alizadeh (2019).  
The results of distance from roads, rivers, and faults showed that most landslides occurred at distances of more than 0-100 meters which confirms the results of previous studies (Pourghasemi & Mohammadi 2016). Also, the relationship between vegetation cover and landslides showed that the highest percentage occurred in class 0.5-0.3 with a weight of 3.56. In addition, the highest slippage for topographic wetness index and sediment transport index (river capacity) occurring were related to 6.39- 11.29 and 11.84-19.6 classes (weights of 1.009 and 1.17), respectively. The area under the curve was calculated at 85.56% in the model validation for the Bayesian model, which was classified as a very good performance. Therefore, the results of our study can play an important role in the management and planning of the Talar watershed.