پهنه‌بندی خطر زمین‌لغزش با استفاده از مدل شبکه عصبی مصنوعی در پایین دست سد سنندج

بررسی موقعیت و ارزش محیط‌های انسانی آسیب‌پذیر در برابر مخاطرات ژئومورفیک، از وظایف مهم دانش ژئومورفولوژی است. هدف تحقیق حاضر، امکان سنجی و ارزیابی خطر وقوع زمین‌لغزش با تولید نقشه‌های پهنه‌بندی خطر و درجه‌بندی حساسیت این پهنه‌ها به عنوان یکی از مخاطرات ژئومورفیک در منطقه ی پایین دست سد سنندج است. در این تحقیق با استفاده از نرم‌افزار arcgis و زبان برنامه‌نویسی پایتون، از مدل شبکه عصبی پرسپترون برای شناسایی پهنه‌های خطر زمین‌لغزش در منطقه استفاده شد؛ بدین منظور، 9 لایه ی ورودی درجه ی شیب، جهت شیب، لیتولوژی، کاربری اراضی، بارش، ارتفاع، فاصله از عوامل آبراهه، جاده و گسل در پهنه‌بندی خطر زمین‌لغزش بررسی شد. نقاط لغزشی و غیر لغزشی منطقه نیز با استفاده از تصاویر ماهواره ای، بازدیدهای میدانی و ... مشخص شد. در مدل شبکه عصبی، از وزن‌یابی درونی در تعیین وزن لایه ها استفاده شد. داده ها با استفاده از شبکه ی پرسپترون چندلایه با الگوریتم یادگیری آدام آموزش دیدند. از الگوریتم جستجوی شبکه ای نیز به منظور بهینه سازی و تنظیم فراپارامترهای شبکه عصبی استفاده شد. ساختار نهایی شبکه دارای 9 نرون در لایه ورودی، 30 نرون در لایه میانی و 1 نرون در لایه خروجی است. از 5 روش محاسبه ی میزان خطای مدل ها (امتیاز f1، دقت کلی، خطای تولیدکننده، خطای کاربر و ماتریس خطا) نیز برای صحت سنجی مدل استفاده شد. در نهایت، نقشه ی پهنه‌بندی خطر زمین‌لغزش در 5 کلاس خطر تهیه شد. بر اساس این پهنه‌بندی11.5 درصد از مساحت منطقه در کلاس خطر خیلی‌کم، 19.7 درصد در کلاس خطر کم، 29.6 درصد در طبقه متوسط، 31 درصد در طبقه زیاد و 8.1 درصد در طبقه خیلی زیاد قرار می‌گیرد. با توجه به نتایج صحت سنجی مدل نیز مشخص شد مدل شبکه عصبی پرسپترون با دقت 91.49 درصد، در پهنه‌بندی خطر زمین‌لغزش انطباق مناسبی با جغرافیای منطقه دارد.

Landslide hazard zonation using artificial neural network model downstream of Sanandaj Dam

Extended abstract1 Introduction Landslide as a morphological risk factor is the result of the operation of natural and environmental factors that, like other morphological irregularities, causes damage to human facilities, infrastructure, and property. One of the most significant works of geomorphology is the identification of stable landforms and places with minimal potential for catastrophic or lethal phenomena caused by environmental processes that have a detrimental effect on humans and their affinities (Rosenfeld, 2004, p. 423). Mass movements in mountainous areas are one of the natural processes and one of the most important causes of landscapes (Van Westen, 1993, p. 2) Landslide sensitivity zoning is one of the essential tools in risk management and decision making (Dahal, 2008, p 496). The purpose of this study is to identify landslide hazard zones in the downstream basins of Sanandaj Dam with an area of 970 km2. Annually, mass movements in the area cause damage to roads, power lines, rangelands and natural resources, farms and residential areas and increase the amount of soil erosion in the area. This basin is located on SanandajSirjan structural zone based on common structural divisions of Iran. Kurdistan mainly with mountainous topography, high tectonic activity, diverse geological and climatic conditions, has the major natural conditions for creating a wide range of mass movements. According to available statistics and also the research of MIRSANIE et al (2006), Kurdistan province is the third most landslide province after Mazandaran and Golestan. If it is a criterion for the ranking of the provinces, it will be ranked higher. NAIRI and KARAMI, 2018).2 Methodology In this study, in order to map the landslide hazard according to geological, geomorphological, hydrological, climatic and human and environmental factors of the area, 9 effective factors include slope, slope direction, distance from fault, distance from road, distance from Waterways, lithology, land use and precipitation were identified and evaluated. The required information layers were then prepared in Arc GIS 10.6 software environment. The first step in zoning is to identify the landslides that have occurred in the area. Using the 2016 Landsat 8 ETM + satellite imagery with a spatial resolution of 30 meters, Google Earth software and field studies identified 237 landslides. In this study, 89 nonslip points were prepared for using in training and testing stages of perceptron neural network in slopes of less than 5 degrees. Artificial neural networks have been created from a number of limitations of advanced content management services to the top walkway called Neuron, which works for problem solvers and can select information through synapses. Neural networks begin to learn using the data pattern entered into them. The learning of models, which in fact determine their internal parameters, is based on the error correction law, which is the generalization of the wellknown least squares method. In fact, in this method, by regularly correcting the error, the best weights that provide the most accurate output for the network are identified. The neurons fall into the input layer, the output layer, and the hidden or middle layer.3 Results By default, the software was selected to prevent interference and capture of the hidden layer. 70% of landslides in the study area were used to train the neural network and the remaining 30% were used as reference land data for testing and calibrating the model. Data were trained using a multilayer perceptron network with Adam's learning algorithm. In this study, a network search algorithm was used to optimize and adjust neural network metaparameters. Because of nonlinear relationships in landslide phenomenon, relu transfer functions were used. The coefficient of learning that controls the amount of weight change is 0.01. The final network structure has 9 neurons in the input layer, 30 neurons in the hidden layer and 1 neuron in the output layer. After preparing the neural network, the study area was analyzed with 970 square kilometers with 9 input variables that were converted to raster data in 30 x 30 pixels. The results of the analysis were plotted with five categories of landslide hazard. It uses 5 methods for model error detection4 Discussion ConclusionsThe downstream area of ​​Sanandaj Dam is one of the most active areas of Kurdistan province in terms of human activities. In the risk zoning maps, the optimal areas for human activities (very low and low risk areas) as well as the unfavorable areas (high and very highrisk areas) are identified. According to the neural network model, about 31% are in the range of desirable areas for human activity. Also, about 39% are in the area of ​​undesirable and very undesirable areas. The natural features of the region affect the occurrence of landslides in the region. The results of the neural network model are in line with the realities of the region's widerisk hazards, and highrisk areas are often located in the west and southwest of the basin. These areas correspond to the mountain unit and the steep slope. The western outskirts of Sanandaj have been affected by landslides and have been classified as highrisk and highrisk. The central and southern parts of the basin are arranged along the Qashlaq River to the exit of the basin in the areas with very low and low risk. The eastern part of the basin is affected by mass movements, especially precipitation and inflows, and is divided into areas with medium to high risk. The results of risk zoning validation show that in this region, the neural network model with 91.49% accuracy has very good accuracy. It is suggested that the resulting map be considered as a base map in order to carry out any executive actions in the area. The vastness of unfavorable areas in the basin indicates that the study area in general has a high potential for landslides.