ارزیابی عملکرد الگوریتم‌های آماری لجستیکی و ناپارامتریکی به‌منظور مدیریت مناطق حساس به حرکات توده‌ای در حوضه آبریز گویجه بل

شناسایی مناطق مستعد حرکات توده‌ای ازجمله زمین‌لغزش از طریق مدل‌سازی خطر با مدل‌های مناسب و کارا، یکی از اقدامات اساسی در کاهش خسارت احتمالی و مدیریت خطر است. زمین‌لغزش به‌عنوان یکی از انواع حرکات توده‌ای، فرایند پیچیده‌ای است که تحت تاثیر پارامترهای داخلی و خارجی روی می‌دهد که شناخت این پارامترها و میزان تاثیرشان در وقوع مخاطرات و استفاده از ابزاری مناسب برای کمی سازی، برنامه‌ریزان و مدیران را در برنامه‌ریزی‌های توسعه و مدیریت بهینه منطقه به‌ویژه مناطق کوهستانی در مقیاس‌های منطقه‌ای و محلی یاری می‌کند. هدف از مطالعه حاضر بررسی وقوع بالقوه زمین‌لغزش در حوضه گویجه‌بل با استفاده از مدل‌های رگرسیون لجستیک و شبکه عصبی مصنوعی با الگوریتم پرسپترون چند لایه به‌منظور شناخت مناطق حساس به وقوع پدیده مذکور می‌باشد. برای مدلسازی از 9 پارامتر مستقل اعم از لایه بارش، لیتولوژی، پوشش و کاربری اراضی، ارتفاع، شیب، جهت شیب، فاصله از شبکه زهکشی، فاصله از گسل و فاصله از جاده استفاده گردید. بعد از استانداردسازی فازی هر یک از پارامترها، نه فاکتور به‌عنوان متغیر مستقل و زمین‌لغزش‌های رخ داده نیز به‌صورت یک لایه باینری و به‌عنوان متغیر وابسته برای مدل رگرسیون لجستیک؛ همچنین فاکتورهای استاندارد شده به‌عنوان نرون های ورودی و زمین‌لغزش‌های رخ داده به‌عنوان آموزش دهنده مدل شبکه عصبی مصنوعی با الگوریتم پرسپترون چند لایه معرفی گردید. نتیجه اعتبارسنجی roc نشان می‌دهد مساحت زیر منحنی در مدل شبکه عصبی مصنوعی بیشتر از مدل رگرسیون لجستیک بوده است و دقت برابر با 91/0 را نسبت به رگرسیون لجستیک با دقت 89/0 نشان می‌دهد. همچنین 9 درصد از مساحت منطقه مورد مطالعه در پهنه‌های خطر زیاد و بسیار زیاد و 5/9 درصد جزو پهنه‌های خطر متوسط می‌باشد. پهنه‌های خطر متوسط می‌توانند با سومدیریت و ساخت و سازهای عوارض انسانی ازجمله جاده تحت تاثیر قرار گرفته و به پهنه‌های خطر زیاد و بسیار زیاد تبدیل گردند.

Evaluating the Performance of Logistics and Nonparametric Statistical Algorithms in order to Manage Areas Vulnerable to Mass Movements in Goijabel Catchment

. IntroductionMass movements are among the most destructive natural disasters in mountainous areas which cause great damages of over millions of dollars to human beings ' life and infrastructures across the world. Landslide, as a type of mass movement, is a complex process that occurs under the influence of internal and external parameters. The main effective parameters in hillside instability are earthquakes, precipitation and human activities. Landslide risk assessments estimate the probability of their occurrence in a place with a return period. Most of the methods based on educational samples and assignment of a weight to each of the parameters and subparameters establish the relationship between effective factors in landslides and spatial analyses. Among these methods are logistic regression, multivariate statistical analysis, artificial neural network, fuzzy model, and neural fuzzy model. In this study, for optimal management in Goijabel basin, we used the logistic regression method, as a statistical method which creates a mutual relationship between the dependent parameter and independent parameters, as well as artificial neural networks with perceptron multilayer algorithm, which allocates a weight to each of the factors in landslide using educationaltesting and training data based on nonlinear functions. 2. Study areaThe study area is located 10 km from the south west of Ahar, in East Azerbaijan Province. The application site lies between the latitudes 38° 21′ 42″ N to 38° 27′ 39″ N, and the longitudes 46° 47′ 21″ E to 46° 56′ 53″ E, and covers an area of 74.62km2.3. Material and MethodsTo implement the methods used in this study, firstly, the maps of effective factors in landslide occurrence were provided and extracted, including lithology static factors, distance from fault, elevation, slope and aspect and dynamic factors of distance from road, distance from river, land use and land cover.Fuzzification of effective factors, logistic regression, parameters such as Chi Square, Pseudo R Squared and ROC are used for validation of the logistic regression model. The other method used in this research is perceptron neural network. Accordingly, effective factors in landslide occurrence in the area under study were prepared as the models ' input. In the next step, each extracted layer was sectioned with the landslides layers, and based on the histogram and the area of landslides in each of the layers, a reclassification was done in all of the layers.4. Results and Discussion Validation of the models used was by ROC method. The results of this model are based on statistical classification analyses. This method indicates higher area under the curve in the neural network model. In other words, the neural network model with the numerical average of 0.91% is introduced as a more efficient model compared to the logistic regression model with the numerical average of 0.89%. Using educational data and hidden nodes and other indexes such as coefficient of acceleration, this model can provide better modeling of the areas sensitive to landslides occurrence.5. Conclusion In this study, for an optimized management in the region and avoidance of potential damages caused by mass movements, especially landslides, the zoning of the areas sensitive to these movements was performed using logistic regression models and artificial neural network with perceptron multilayer algorithm. For the modeling, 9 independent parameters were used including precipitation, Lithology, land use and land cover, elevation, slope, aspect, distance from drainage network, distance from fault and distance from road. After standardization of each of the parameters, 9 factors were introduced as the independent variable, the landslides as binary layer for logic model, and standardized factors as input neurons and landslide layer as training of artificial neural network model with perceptron multilayer algorithm. The result of ROC validation shows that the area under the curve in the artificial neural network model has been higher than that of the logistic regression model, indicating a precision of 0.91% compared to logistic regression with a precision of 0.89%. Therefore, an area of over 9% is in high and very high risk zones and 9/5% of the area is in mediumrisk zones. Mediumrisk zones can be affected by mismanagement and human constructions such as roads and be changed into high and very high risk zones, considering the fact that in this modeling, the areas close to manmade effects in both models are zoned as areas sensitive to landslide risk.