Reza Talaei; Samad Shadfar
Abstract
Introduction
Landslides are one of the natural hazards in mountainous areas that threaten the safety of residents and the environment. In the past few decades, landslides have caused damage to natural and human resources in the Saqezchi Basin in the south of Ardabil Province. Landslides have occurred ...
Read More
Introduction
Landslides are one of the natural hazards in mountainous areas that threaten the safety of residents and the environment. In the past few decades, landslides have caused damage to natural and human resources in the Saqezchi Basin in the south of Ardabil Province. Landslides have occurred in more than 9.2% (2600 ha) of the area. In this basin, like other landslide areas, for land-use planning and management, it is necessary to analyze the whole area in order to estimate the probability of landslides occurrence in the future. It is possible to solve this problem by analyzing the geomorphology, topography, geology, land use, hydrology and climate factors of the basin in the form of information layers in the geographic information systems on a regional scale. Landslide susceptibility assessment has not been done with modern methods and with high accuracy in the Saqzachai Basin until now. The results of this research can be used in predicting the possible occurrence of landslides and reducing damage in the Saqzachai Basin.
Materials and methods
The research basin with an area of 27,918 ha is located in the south of Ardabil Province and in the southwest of Khalkhal City. In this basin, the inventory map was generated based on 113 landslides, the training dataset and validation dataset were, respectively, prepared using 70% landslides and the remaining 30% landslides. Ten landslide causative factors based on slope angle, slope aspect, distance to faults, distance to stream network, distance to the roads, distance to settlement area, lithology, land-use, peak ground acceleration (PGA) and average annual precipitation were applied for the models analysis. Two nonlinear methods of neural network called multi-layer perceptron with feed forward structure and logistic regression were used to predicting the susceptibility of landslide occurrence. The probability of landslide occurrence in each pixel was calculated based on both models. The prediction accuracy of the two models were evaluated using the Receiver Operating Characteristic (ROC) curve.
Results and discussion
In the neural network model, landslides triggering factors, including the average annual precipitation (0.136) and the peak ground acceleration (0.134), have been the greatest effect in predicting the probability of landslides. The factors of slope angle (0.067), slope aspect (0.069), distance to faults (0.110), distance to stream network (0.101), distance to the roads (0.109), distance to settlement area (0.096), lithology (0.109) and land-use (0.068) are respectively important in landslides susceptibility modeling to using artificial neural networks. Therefore, all ten factors were used in modeling by artificial neural networks. The results indicated that the probability of landslide occurrence varies from 0.00 to 0.961. In the classification of the watershed according to the degree of landslide susceptibility by the natural breaks method based on the estimated probability by the neural network method, 85.7% of the area is placed in the zones with low and very low susceptibility. In 6.6% of the area, there is a probability of moderate susceptibility, and in 7.7%, there is a high and very high landslide susceptibility. Landslide susceptibility analysis is started without independent variable and ended by adding variables in the tenth step using logistic regression method. The results show that only three levels of the factor of slope aspect are ineffective in the logistic regression model. Probability values were calculated between 0 and 1 for all pixels in the area based on the values of independent variables by estimating constant and coefficients related to logistic regression model. The landslide-prone areas of low and very low susceptibility, medium susceptibility and high to extremely high-susceptibility grades are 79.9%, 10.1%, and 10%, of area, respectively, by the natural breaks method in the logistic regression model. The accuracy and validity of the logistic regression and artificial neural network models based on the ROC curve and the area under it (AUC) are equal to 0.848 and 0.929, respectively. The findings of the models show good results with the accuracy of two models being higher than 84%. The results obtained from two methods in most studies in the world and in Iran indicate their ability to accurately estimate susceptibility to landslides occurrence, but the artificial neural network method is more accurate despite its specific complexities.
Conclusion
Landslides are an important limitation for development in the landslide areas in the south of Ardabil Province. The environmental conditions in the Saqzachi Basin are susceptible to the occurrence of new landslides or the reactivation of old landslides. The probability of landslides occurrence was simulated using effective factors and using logistic regression and artificial neural network models in the region. The results obtained from the artificial neural network model are the most accurate and better than the logistic regression model. The landslides triggering factors, including the average annual precipitation and the peak ground acceleration have the greatest impact to predicting the probability of landslide occurrence using the artificial neural network model. The findings of the models show good results with the accuracy of two models being higher than 84%. The artificial neural network method is superior in explaining the relationship between landslide occurrence and influencing factors. The landslide susceptibility map was prepared using this method by dividing into five class, namely: very low (71.4%), low (14.3%), moderate (6.6%), high (4.3%) and very high (3.4%) susceptibility zones. Therefore, it is recommended to use the artificial neural network models in landslide susceptibility assessment in the basin and similar regions to help decision makers, planners, land use managers and government agencies in hazard and damage reduction.
