Applying Feature-Selection Algorithm to Predict Landslide in the Southwest of Iran

Extended abstract 1- INTRODUCTION Nowadays people have an increased sensitivity towards landslides especially in mountainous areas using change in the land use and the expansion of communication networks (Gvrsysky et al., 2006). In the twentieth century, Asia has allocated the highest incident of landslides (220 landslides). Latin America has had the highest number of casualties (more than 2,500) and Europe has experienced the highest loss (Karami, 2012). Landslide is one of the significant phenomena in the environment, watershed management, and natural resources. The importance of landslides can be discussed and analyzed from various perspectives. The most important reason refers to human and financial loss (Rajab Zadeh, 2013). Research on the dynamic relationships between factors in landslides has a high role in the investigation of the respective risk. In fact, much research has been conducted in the realm of determining the relationship between environmental factors and the occurrence of landslide (Anbalagan, 1992, Liu Min, 2001, Ayvahashy et al., 2003, Yalv and Yamagyshy, 2005). Some of such research is consistent with the relationship between the distribution of geological and geomorphological factors and landslides observed. However, to analyze the results and predict the likelihood of landslides, there are common tools that are used in statistical calculations. 2- THEORETICAL FRAMEWORK A landslide, also known as a landslip, is a form of mass wasting that includes a wide range of ground movements, such as rock falls, deep failure of slopes, and shallow debris flows. Slope, fine sediments and ground moisture have important roles in the occurrence of landslides; given that many parameters affect the landslide, a more effective choice to reduce time and costs is important. The subject of feature selection is the one of the issues identified in the machine learning and statistics. The problem in many applications (such as classification) is very important. Because in these applications, there are a large number of features that many of them are unused. In fact, if they are not removed, these features will not create problems, but save a lot of useless and useful information together. 3- METHODOLOGY This study was carried out in the southwest of Iran (a part of Khuzestan, Khorramabad, Ilam, Kermanshah and Hamedan). It includes an area of about 154272.48 km2 and is located at the longitude of N 29° 56΄to 35° 46΄and the latitude of E 45° 24΄ to 52° 1΄. The altitude of the study area ranges from the lowest 30 m to the highest 4,415 m. There are different methods that try to find better subsets among the 2T subsets. In all of these methods, the selection of the subset is based on the type of application and type of definition that can optimize the value of an evaluation function. In fact, each way tries to make the best attributes of choices, but according to the extent of answers and increasing the answer with T, finding the optimal solution and T medium is costly. Feature selection process has four steps: Generation function: this function find sub-candidates for the procedure. Elevation function: it is based on data subset to be evaluated and a number as the method returns. Different methods try to find a subset that optimizes the amount. Stopping criterion: it is used to decide when to stop the algorithm. 4- RESULTS The results show that the case study is located in 6 classes increasing sensitivity as the number of class increases. So that the areas located in the South, East, and parts of the West regions are most sensitive to landslide. Motevali et al., (2008) show that using a new method such as LMT can prepare landslide map with low data. So, in the research of geomorphology and geology, feature selection can be used. Rasaie et al (2015) used regression in GIS software to prepare landslide map.  The results showed that using effective parameters of landslide can find landslide map easily and quirkily.   5- CONCLUSIONS AND SUGGESTIONS The results of feature selection method show that the Ranker method with Gain-Ratio-Attribute-Eval, with low error, with highly significant correlation (87.5), and with LMT classification is the best method for the selection of the most effective data to determine landslide. Also, the results indicated that some parts of North and South-East of the study area are located at greater risk of landslide. Also, principal components showed that curvature, profile, plan and SPI were the most important data for determining landslide. In the study, it was attempted to use low data selected by feature selection, and save time and money via determining important data for landslide. Using the data, landslide map was prepared spatially.