ارزیابی خطر وقوع ناپایداریهای دامنه‌ای گسل بناروان با استفاده از مدل شبکه‌های عصبی مصنوعی (MLP)

Introduction
Faults‌ are among the most common‌ geological conditions ‌in rock masses‌, which are considered as one of the most significant examples‌ of rock discontinuity. The earth’s instability is often due to the presence of faults in or near them. Bozqush Mountains are the most important‌ landscapes‌ in southern Azerbaijan, including the south-western‌ border of Tabriz fault’s extension, and south-eastern border‌ of Mianeh-Ardebil fault. Benaravan fault in the eastern part of this mountain range is part of the Mianeh-Ardabil fault‌. In the present study‌, a study was conducted titled as Zoning‌ the Risk of Hillside‌ Instability Using Artificial Neural Network with MLP Model. This was the purpose of selecting this model to investigate areas with the potential of hillside‌ instability occurrence.
Data and Methodology‌
In this study, aerial photos‌ and ETM satellite‌ images of the region, topographic maps‌ with a scale of 25000/1, and geological‌ map with a scale of 100000/1 were used. There were also‌ several field visits to the area.
Artificial Neural Networks‌ Model
Neural network‌ models are a kind of simplistic‌ modeling‌ of real neural systems‌ that are used extensively in solving various problems in science. The best way to solve complex problems is to break it down into a simpler ‌sub-question‌, and each of these sub-sections‌ can be easily‌ understood and described. In fact, the network is a collection‌ of these simple structures that together describe‌ the ultimate complex system. It must be pointed out that in this study, a variety of multilayer perceptron networks‌ were used.
Data Analysis
To investigate the relationship between the factors affecting the occurrence of hillside‌ instabilities in the studied area, after providing dispersion map of instability points, the dispersion of these points to nine factors affecting the occurrence of hillside‌ instabilities was investigated. Each information layer‌ was classified‌ into five classes, and based on the degree of sensitivity to hillside‌ instabilities, each of the rating classes was given a score between 1 to 5‌. 5 was given to a class‌ that had the greatest hillside ‌instability.
The Argument and the Findings
The maps‌ of the factors affecting the hillside‌ instability, which are independent variables in the instability occurrence, were entered into IDRISI software‌ and were processed. In constructing an artificial neural network, the first task is to determine the type of the network.  In this study, an artificial neural network with a multi-layer perceptron structure (MLP) with 1 input layer, including 9 neurons, and a hidden layer with 16 neurons, and an outlet layer were used. The algorithm used in this network was the same after error algorithm, and the sigmoid function was used as an activity function. In order to achieve better zoning‌ and proper output with high precision, a different structure of the neural network was tested by changing the number of neurons‌ and other parameters. This optimal number was achieved at 20000th recurrence with a training error of 440/0 and a test error of 0.0622. As stated, the best way for an appropriate network architecture is achieved‌ through trial and error. In this study, using the trial and error method, the best architecture was selected with 1 input layer including 9 neurons, a hidden layer including 16 neurons, and an output layer (9,16,1). The study used 3181 sliding pixels data for network training and testing. Of these pixels, 2544 pixels were used for training and 636 pixels for network testing. Multiple training rates and momentum factors were tested, and finally, the training rates between 0.014-0.01 with amounts and momentum factors of 0.5 were selected. After performing the above steps, using neural network method, the zoning‌ map of hillside‌ instabilities‌ occurrence risk was prepared,
Conclusion
When a hillside‌ is liable to instability, the occurrence‌ of instability at its surface will be inevitable‌. Theoretically, if steep hillsides‌ that are formed by loose and detached materials, get adequate moisture or water, they are prone to instability. Then, if one or more secondary factors interfere, the instability will occur. Heavy rainfall‌, movement of faults‌, earthquake‌, Earth’s roughness and so on are among such factors. Given these cases, the likelihood of hillside‌ instabilities ‌occurrence‌ in an area, in case of occurrence conditions, is expected. For zoning‌ hillside‌ instabilities occurrence risk using artificial neural network, first at the testing stage, in order to avoid error increase and over-training of the network, each of the parameters of the artificial neural network was determined by trial and error. In this study, a network‌ with an architecture of an input layer with 9 neurons including categories as altitude, gradient, gradient direction‌, lithology, distance from fault, distance from canal, distance from road, land use, vegetation‌, and an intermediate layer with 16 neurons and an output layer (9,16,1) that shows‌ areas with potential for the occurrence of hillside‌ instability, was used. The results showed‌ that 81.5%, 95.12%, 38.19%, 06.27%, and 78.34% of the area were located in very high, high, medium, low and very low risk groups‌, respectively. The zoning‌ map of hillside‌ instability‌ risk shows‌ that areas with high and very high risk, cover about 76.18% of the study area‌, which are more consistent with the high altitudes, fault zone, and high gradient of the area‌.