A. H. Pasha; A. Sorbi; S. Behzadi
Abstract
Mass movements, especially landslides, are one of the natural hazards that to a large extent occur, are controlled, or are prevented by human. It is obvious that human interferences in nature regardless of stability conditions and its natural balance leads to physical reactions from the environment to ...
Read More
Mass movements, especially landslides, are one of the natural hazards that to a large extent occur, are controlled, or are prevented by human. It is obvious that human interferences in nature regardless of stability conditions and its natural balance leads to physical reactions from the environment to return the sustainability and balance. Damages caused by the landslides, which have been growing in recent decades, have made humans to find appropriate solutions to reduce and control this phenomenon. Zonation of areas susceptible to landslide is one of the most widely used methods to avoid hazardous areas or applying controlling methods in hazardous areas. This research uses artificial neural network for zonation of landslide susceptibility in the Qazvin-Rasht quadrangle area. The studied area is one of the most susceptible areas for landslide event in terms of topography, climate, and geology, as the history of the area shows 338 recorded landslides. Fifteen variables studied in other researches as effective variables in occurrence of landslides were selected to investigate this area. By combining these variables and the map of existing landslides, value of each of the 15 variables was extracted for sliding points. In the next stage, a number of points (1000 points) were randomly selected from the area and values of these variables were extracted for them. Each of the two data sets was divided into two training (70%) and test (30%) categories. We combined each of the two training and test categories, and used their output for training and testing the network. The number of internal layers of the neural network was determined to be 9 layers based on trial and error method and calculation of the root mean square error value (RMSE = 0.4041). The constructed neural network is of feedforward networks type with back-propagation algorithm and its training algorithm is of Levenberg-Marquardt back-propagation training algorithm type. After training and testing the network and conducting necessary corrections on it, the constructed neural network was used to predict the sensitivity of landslides in studied area. We placed results of this prediction in a range from 0 to 1 and obtain the best zonation map of the landslide susceptibility by choosing a threshold. Final evaluation of the zonation map of landslide susceptibility in the Qazvin-Rasht quadrangle shows an error of approximately RMSE = 0.4164 and the constructed neural network identifies 298 out of 338 occurred landslides in the high-risk zone, indicating the accuracy of 88.1%.
