Landslide hazard potential zoning using the neural network method (Case study: Alamut watershed in Qazvin province)

One of the types of wide-ranging processes that cause many casualties and financial losses in many parts of the world and Iran every year is landslides. Landslide is a mass movement of soil or rock due to gravity on slopes, which is one of the important geological hazards. Landslides are morphodynamic processes that have a complex structure and various factors and variables play a role in its creation, so it is very difficult to assess the risks of this phenomenon. This natural phenomenon causes the destruction or damage to residential areas, all kinds of vital structures and arteries such as roads, power lines, water, gas, pastures, forests, and agricultural lands and will have many destructive environmental and social effects. Landslide mechanisms and their main mechanisms, in addition to internal and external factors (climate), are also affected by human (entropic) activities. In today's world, many methods have been introduced for zoning the risk of landslides, which are generally divided into three categories: statistical methods (two-dimensional, multivariate, logistic regression, information value), experimental (Anbalagan, Mora-Warson, Stevenson) and a combination. (Artificial neural network and fuzzy logic) are divided. Most of these methods are experimental and are presented regionally for specific conditions. In Iran, there are many landslides every year that cause a lot of damage (Afjeh Nasrabadi et al., 2008). Alamut River catchment area with an area of 74.82 square kilometers, with coordinates of 15 23 23 ° 50 to 54 52 52 ° 50 eastern longitude and 12 17 ° 36 to 17 33 ° 36 north latitude in northeastern Qazvin province in one The mountainous region is located. Due to the mountainous nature of the area and the great difference in altitude, various geological formations and a wide variety of lithology, have a great talent for creating wide-ranging movements, especially landslides. In this study, the potential for landslide risk in the Alamut River catchment area has been investigated using the perceptron artificial neural network method. Neural networks are computer algorithms that can extract our important relationships between a large number of linear and nonlinear parameters from a given bank. Pradhan Vali, 2010). Artificial neural network is one of the effective models in zoning the landslide. In this model, complex statistical analyzes have been avoided and based on nonlinear functions, each of the effective factors in landslides has been assigned weight. This model is based on the training of effective factors in landslides and by dividing the data in educational and experimental classes and using sigmoid functions, it proceeds to zoning the susceptible areas. Perceptron networks consist of an input layer, a number of hidden layers, and an output layer. In multilayer perceptron networks, the number of hidden layers can be any number, although in most applications a hidden layer is sufficient. To conduct this research, slides were first identified through the interpretation of aerial photographs and satellite images and field visits and a map of their distribution was prepared. In the next step, according to the location of landslides, seven effective factors of slope, slope direction, height, precipitation, land use, lithology and distance from fault were investigated and information layers were prepared by GIS and in MATLAB environment suitable structure for Landslide zoning was written using the artificial neural network method with a multilayer perceptron structure. In order to use the neural network method in MATLAB software, the following steps have been performed in order. 1. Provide data. 2. Normalize the data (0 to 1). 3. Squaring the area to 100 * 100 sides. 4. Teaching the neural network (using the previous slips and that the areas with a slope below 5 degrees and the areas inside the waterway do not have slips). 5. Convert data to Excel. 6. Enter the data into the network and get the output. 860 pixels of data were used to train and test the network throughout the region. Of these, 674 pixels were used for training and 184 pixels for network testing.The results of the artificial neural network outcomes in the experimental phase show that the network created was able to report 36 out of the 38 sliding pixels correctly, indicating of 91% sensitivity. Also, out of 150 non-slip pixels, the network was able to detect 143 experimental samples, which again achieved an accuracy of about 91%. Therefore, the total accuracy was calculated to be 91%. The number of repetitions was changed from 1000 to 15000, which was calculated with the number of repetitions of 10,000 as the minimum error value. According to the results of the neural network method, only 25.76% of the basin area is in the middle and upper class and 72.17% of the basin is in the low and very low class. Out of a total area of 435 square kilometers, landslides are 121 square kilometers, about 21.81 percent are in very low, low and medium classes and 314 kilometers, about 72.81 percent are in high and very high classes. Areas with high and very high risk classes are often located in the eastern and northern parts of the basin, and areas with very low and low risk classes are mostly located in the western and central areas.