پهنه‌بندی حساسیت وقوع زمین‌لغزش با استفاده از الگوریتم‌های یادگیری ماشین (منطقه‌ مورد مطالعه: بخشی از حوزه‌ آبخیز هراز)

زمین‌لغزش یکی از انواع پدیده‌های زمین‌شناسی در سراسر جهان است که هر ساله تلفات جانی و خسارات اقتصادی زیادی را به همراه دارد. بنابراین، این پژوهش به‌منظور ارزیابی پهنه‌بندی حساسیت وقوع زمین‌لغزش با استفاده از الگوریتم‌های مختلف یادگیری ماشین از نوع ماشین پشتیبان بردار (svm) و رگرسیون فرآیند گاوسی (svm) با دو کرنل (puk و rbf) و جنگل تصادفی (rf) در بخشی از حوزه‌ آبخیز هراز، ایران انجام شده است. در پژوهش حاضر از نُه عامل شیب، جهت، ارتفاع، زمین‌شناسی، کاربری اراضی، فاصله از گسل، فاصله از جاده، فاصله از رودخانه و بارش به‌عنوان پارامترهای ورودی و نقاط لغزشی و غیرلغزشی به‌عنوان پارامتر خروجی برای مدل‌سازی و پهنه‌بندی حساسیت وقوع زمین‌لغزش استفاده شد. از مجموع 148 نقاط لغزشی و غیرلغزشی، 70 درصد برای مرحله‌ آموزش و 30 درصد برای مرحله‌ آزمایش مدل‌سازی استفاده شد. برای ارزیابی کارایی مدل‌ها و انتخاب مدل بهینه از معیارهای سنجش خطای مدل accuracy، f1-score و auc و برای تحلیل حساسیت از روش حذفی استفاده شد. نتایج به‌دست ‌آمده نشان داد که مدل rf (با 9/0accuracy =، 957/0f1-score= و 999/0auc=) در بخش آزمایش در مقایسه با دیگر مدل‌ها به‌عنوان بهترین مدل برای پهنه‌بندی حساسیت وقوع زمین‌لغزش انتخاب شد. بر اساس نتایج نقشه‌ پهنه‌بندی مشخص شد که به‌ترتیب 86/31، 16/32، 38/13، 73/9 و 84/12 درصد در طبقات با حساسیت خیلی کم، کم، متوسط، زیاد و خلیی زیاد قرار دارد. علاوه‌براین نتایج تحلیل حساسیت مدل نشان داد که جهت شیب، حساس‌ترین پارامتر در پهنه‌بندی خطر وقوع زمین لغزش است. مقایسه‌ نتایج مدل‌ها نشان داد که ارتباط معناداری بین مقادیر پیش‌بینی ‌شده و مقادیر مشاهداتی با استفاده از مدل‌های استفاده شده وجود ندارد. بر اساس نتایج به‌دست آمده از نقشه‌ پهنه‌بندی حساسیت وقوع زمین‌لغزش می‌توان به اولویت‌بندی و مدیریت مناطق پایدار و با حساسیت کم به وقوع حرکت‌های توده‌ای برای اجرای عملیات عمرانی پرداخت.

landslide susceptibility mapping using various soft computing techniques (case study: a part of haraz watershed)

introductiona landslide is one of the mass movements on the top surface of the earth. landslides have resulted in notable injury and damage to human life and destroyed infrastructure and property. landslides represented approximately nine percent of the natural disasters worldwide during the 1990s. according to studies, this trend is expected to continue due to increased human development. many studies have been done to determine the factors affecting mass movement. in large part of iran including the mountain areas, tectonic activity and seismic high with diverse geological and weather conditions led to many countries prone to landslide. landslides cause wide damage to natural resources, human settlements, infrastructure, mud floods, and filling reservoirs. landslides cause extensive property damage and occasionally result in loss of life. besides, should not be ignored the social and environmental impacts resulting from the occurrence of this phenomenon, such as immigration and unemployment. one of the strategies for reducing losses due to a range of movements is the identification and management of unstable slope areas. to identify unstable regions pay to landslide hazard mapping. the main purpose of this research is to assess the effective parameter on landslide occurrence and to compare different machine learning models including svm, gp regression, and rf for landslide susceptibility zoning. materials and methods the study area is a part of the haraz watershed, mazandaran province, iran, occurrence many landslides are damaged after each heavy rain. so, it was selected as a suitable watershed to evaluate the landslide susceptibility mapping (lsm). the vegetation covers and land mainly consists of rangeland. the geology of the study area consists mainly of quaternary and shemshak formations. the first step for the assessment of landslide susceptibility is gathering the necessary data and preparing information. these data were determined based on several factors. considering the literature review, the local conditions, and previous studies. in this study, nine parameters such as slope angle, slope aspect, elevation, geology, land use, the distance of fault, the distance of the road, the distance of the river, and precipitation were identified as key factors for the prediction of landslide susceptibility. to assess the effectiveness of gp-puk, gp-rbf, svm-puk, svp-rbf, and rf to estimate the landslide susceptibility map (lsm), data used in the present study were taken from field data. in this study, the dataset contains 148 observations of landslide occurrence and landslide non-occurrence points. the landslide data have been randomly separated into training (70% of landslides; 103) and testing (30% of the landslides; 45). to judge the performance of the soft computing techniques, statistical evaluation parameters were used. in this research, three statistical evaluation parameters were used. these parameters are the correlation coefficient (c.c.), root mean square error (rmse), and nash -sutcliffe model efficiency (nse). results and discussionaccording to the results of the comparison of methods, rf was the best model and the accuracy of the rf model was more suitable for the estimation of the landslide occurrence. so, in this study, rf was used for the landslide susceptibility map. single-factor anova test suggests that there is an insignificant difference between observed and predicted values of landslide occurrence and landslide non-occurrence using gp_puk, gp_rbf, svm_puk, svm_rbf and random forest approaches. according to the results of the comparison of methods, rf was the best model and the accuracy of the rf model was more suitable for the estimation of the landslide occurrence. the map of landslide susceptibility map was divided into five classes from none susceptible to very high susceptibility. according to the final landslide susceptibility map, the area belonging to the ldquo;non-susceptible rdquo; class covers 35.86 km2, ldquo;low susceptibility rdquo; class 36.19 km2, ldquo;moderate susceptibility rdquo; class 15.06 km2, ldquo;high susceptibility rdquo; class 10.95 km2 and ldquo;very high susceptibility rdquo; class 14.46 km2 of haraz watershed. sensitivity analysis was performed to find the most significant input parameter in the prediction of landslide occurrence and landslide non-occurrence. the result shows that aspect has a major role in predicting landslide occurrence and landslide non-occurrence in comparison to other input parameters, respectively. conclusion due to all results, some zones are potentially dangerous for any future habitation and development. thus, there is an immediate need to implement mitigation measures in the very high-hazard and high-hazard zones, or such zones need to be avoided for habitation or any future developmental activities. the results of this research can be used by the local authority to manage properly, and systematically and plan development within their areas.