دانشگاه اصفهان جغرافیا و برنامه ریزی محیطی 2008-5362 37 1 2026 03 21 Received: 30/11/2025 Accepted: 26/04/2026 Landslide Assessment and Prediction Using Advanced Machine-Learning Algorithms (Case Study: Alvand Watershed) ارزیابی و پیش‌بینی حساسیت وقوع زمین‌لغزش با استفاده از الگوریتم‌های پیشرفته یادگیری ماشین (محدوده مورد ‌‌‍مطالعه: حوضه آبخیز الوند) 1 22 30370 10.22108/gep.2026.147111.1752 FA سجاد تبارکی دانش‌آموخته ‌کارشناسی ارشد‌ سنجش از دور وGIS، دانشکده‌ علوم جغرافیایی و برنامه‌ریزی، دانشگاه اصفهان، اصفهان، ایران رضا ذاکری نژاد استادیار گروه جغرافیای طبیعی، دانشکده علوم جغرافیایی و برنامه‌ریزی، دانشگاه اصفهان، اصفهان، ایران ایمان خسروی استادیار گروه مهندسی نقشه‌برداری، دانشکده مهندسی عمران و حمل و نقل، دانشگاه اصفهان، اصفهان، ایران Journal Article 2025 11 30 Abstract Natural hazards are among the most pressing environmental problems, causing annual fatalities, injuries, and homelessness for millions worldwide. Landslides represent one of the most destructive natural hazards, threatening human lives and inflicting substantial damage to natural resources. Unlike other natural events that occur probabilistically over time, landslides have the potential for instantaneous occurrence. In mountainous regions and steep slopes, landslides not only lead to loss of life, but also alter and degrade the environment each year. This study evaluated and predicted landslide hazard in the Alvand watershed of Kermanshah using data mining algorithms. 14 parameters were employed to model the probability of landslide hazard: elevation, slope, slope aspect, geology, land use, Topographic Wetness Index (TWI), Stream Power Index (SPI), distance from stream, drainage density, distance from fault, fault density, mean annual rainfall, distance from road, and Normalized Difference Vegetation Index (NDVI). 4 data-mining algorithms—Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Rotation Forest (RoF)—were applied for modeling. Based on the results obtained from Overall Accuracy (OA) and the Kappa coefficient, the models achieved the following performance: DT (OA = 82.91%, Kappa = 0.78), RF (OA = 96.47%, Kappa = 0.93), RoF (OA = 87.47%, Kappa = 0.82), and SVM (OA = 93.90%, Kappa = 0.90). These findings demonstrated the high accuracy of the RF and SVM methods in predicting landslide hazard in the Alvand watershed. Keywords: Landslide, Data-Mining, Alvand Watershed. Introduction Natural hazards are among the environmental problems that annually lead to the death and injury of many people and cause homelessness for millions around the world. Landslides, in particular, are one of the most destructive natural hazards. In addition to threatening human lives, they cause further damage to natural resources and exert pressure on the economic development of countries, especially developing nations. The financial losses resulting from natural hazards can lead to stagnation in economic growth and prosperity. Slope instability in steep mountainous lands is a major challenge for land managers worldwide. Furthermore, landslides result from complex and diverse processes and are considered among the most serious and destructive natural hazards in mountainous regions globally. A distinctive feature of landslides is their potential for instantaneous occurrence. The Alvand watershed in Kermanshah Province, given its topographic, climatic, and morphological conditions, is one of the areas where investigating landslide probability is essential. The aim of this research was to zone areas susceptible to landslide occurrence and compare the performance of Decision Tree (DT), Random Forest (RF), Rotation Forest (RoF), and Support Vector Machine (SVM) models for landslide prediction in the study area The Alvand River watershed is located in the southwest of Kermanshah Province. The Alvand River is a border river between Iran and Iraq and the region lies between Mesopotamia and the Iranian plateau. The study area covers 112,332.77 hectares and is situated between 34°34' to 34°15' N latitude and 45°35' to 46°09' E longitude. Elevation in the basin ranges from 329 to 2,672 meters above sea level with a mean elevation of 1,098 meters. Materials & Methods To model the probability of landslide hazard, 14 physiographic, hydrological, geological, and environmental parameters were selected based on a review of relevant literature and the specific characteristics of the study area. These parameters included elevation, aspect, slope, land use and land cover, Topographic Wetness Index (TWI), Stream Power Index (SPI), stream density, fault density, distance from faults, average annual rainfall, distance from roads, vegetation density index (NDVI), and geology. All parameter layers were converted into raster format with a consistent spatial resolution and were resampled to ensure uniform geographic extent and cell size for integration into the modeling process. In this research, 4 efficient data-mining algorithms were employed for modeling: Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Rotation Forest (RoF). These algorithms were selected due to their proven performance in landslide susceptibility studies and their ability to handle high-dimensional, non-linear environmental data. The landslide inventory map was prepared using field surveys, historical records, and interpretation of satellite imagery. A total of landslide locations were identified and randomly divided into two independent datasets: training samples (70% of the data) and test samples (30% of the data). The training samples were used to calibrate the models by learning the relationships between the landslide occurrence and the fourteen conditioning factors. The test samples, which were not involved in the model building process, were used to evaluate the predictive accuracy and generalizability of the developed models. Model performance was assessed using the Overall Accuracy (OA) and the Kappa coefficient, both of which were derived from the confusion matrix comparing predicted and observed landslide occurrences in the test dataset. Research Findings Based on the results obtained from the Overall Accuracy (OA) and Kappa coefficient, the performances of the four models were as follows: DT Model: OA = 82.91%, Kappa = 0.78 RF Model: OA = 96.47%, Kappa = 0.93 RoF Model: OA = 87.47%, Kappa = 0.82 SVM Model: OA = 93.90%, Kappa = 0.90 These findings indicated the high accuracy of the RF and SVM methods in predicting landslide hazard in the Alvand watershed. Landslide susceptibility maps of the study area were generated by using all four methods. Discussion of Results & Conclusion The performances of the DT and RoF methods were comparatively weaker than those of the other two algorithms. In contrast, the best results were achieved by using the RF and SVM methods. The high accuracy exceeding 90% (97% for RF and 94% for SVM), along with their Kappa coefficients above 0.9, indicated excellent modeling performance for landslide susceptibility zoning in the study area. The superior performance of these two ensemble-based and kernel-based methods could be attributed to their robustness in handling complex, non-linear relationships among the fourteen conditioning factors, as well as their abilities to reduce overfitting compared to single decision tree models, such as DT. In conclusion, this study demonstrated the high predictive accuracy of the RF and SVM models for landslide hazard assessment in the Alvand watershed. The outstanding accuracy—above 90% for both methods—coupled with their high Kappa values (>0.9), confirmed the suitability of these two approaches for landslide hazard zoning in the region. Future research could explore the integration of deep learning models or the application of these methods to other watersheds with similar topographic and climatic conditions to further validate their generalizability. زمین‌لغزش یکی از مخاطرات طبیعی تخریب‌کننده است که علاوه بر تهدید زندگی انسان‌ها، آسیب جدید به منابع طبیعی وارد می‌کند. پدیده زمین‌لغزش دارای پتانسیل وقوع لحظه‌ای است؛ درحالی‌که سایر حوادث و مخاطرات طبیعی هرچند وقت یک‌بار احتمال وقوع آن وجود دارد. در این پژوهش با استفاده الگوریتم‌های داده‌کاوی به ارزیابی و پیش‌بینی خطر زمین‌لغزش در حوضه آبخیز الوند در استان کرمانشاه پرداخته شد. برای مدلسازی احتمال خطر زمین‌لغزش از 14 پارامتر استفاده شده است. این پارامترها عبارتند از ارتفاع، شیب، جهت شیب، زمین‌شناسی، کاربری اراضی، شاخص رطوبت توپوگرافی (TWI)، شاخص قدرت آبراهه (SPI)، فاصله از آبراهه، تراکم آبراهه، فاصله از گسل، تراکم گسل، میانگین بارش سالانه، فاصله از جاده و شاخص تراکم پوشش گیاهی (NDVI) استفاده شده است. در ادامه این تحقیق، چهار الگوریتم کارای داده‌کاوی شامل روش ماشین بردار پشتیبان (SVM)، درخت تصمیم (DT)، جنگل تصادفی (RF) و جنگل دورانی (RoF) برای مدلسازی به کار گرفته شد. با توجه به نتایج به‌دست‌آمده از شاخص‌های آماری دقت کلی و ضریب کاپا مدل‌های DT، RF، RoF و SVM دارای دقت کلی به‌ترتیب 91/82، 47/96، 47/87 و 90/93 درصد و ضریب کاپا 78/0، 93/0، 82/0، 90/0 و 90/0 هستند که این نتایج نشان از دقت زیاد روش RF و SVM در پیش‌بینی خطر زمین‌لغزش در حوضه آبخیز الوند است.

https://gep.ui.ac.ir/article_30370_817d85d0f28e7a01fa4778c93d45016f.pdf