Introduction

Landslides are among the most destructive geomorphological hazards, causing extensive damage to human lives, infrastructure, and natural environments, particularly in mountainous regions. Triggered by factors such as intense rainfall, seismic activity, geological conditions, and human interventions, landslides have become increasingly frequent due to urbanization, deforestation, and climate change. In Iran, where mountainous terrains and tectonic activity dominate, landslides pose a persistent challenge, inflicting significant economic and social losses annually. Mapping landslide susceptibility is therefore essential for risk reduction, sustainable land-use planning, and hazard management. Recently, machine learning algorithms such as Random Forest (RF) and Support Vector Machine (SVM) have been widely adopted for landslide susceptibility mapping (LSM), owing to their ability to capture nonlinear and complex relationships among conditioning factors. Additionally, the use of Google Earth Engine (GEE) provides an efficient cloud-based platform for processing multi-source geospatial data, enhancing model scalability and accuracy. The Firouzabad watershed is located in the south of Ardabil province, covering an area of 1599 square kilometers of the province. This basin is located at a geographical location of 48 degrees 14 minutes to 48 degrees 32 minutes east longitude and 37 degrees 25 minutes to 37 degrees 54 minutes north latitude (Figure 1). The Firouzabad watershed is bounded by the Gharesu watershed from the north, the Qarnqu River watershed from the west, the Talesh Mountains from the east, and the foothills of the Qezel Ozan River from the south (Abadi et al., 2013). The Firuzabad watershed in Khalkhal County, as one of the landslide-prone areas in Ardabil Province, has been facing an increase in the occurrence of this phenomenon in recent years, and the identification and zoning of high-risk areas in this area is of great importance. Given the complexity of the factors affecting landslides and the nonlinear interactions between them, the use of classical methods alone in identifying high-risk areas cannot meet today's needs. In this regard, machine learning algorithms such as random forest (RF) and support vector machine (SVM) have attracted the attention of researchers as new and efficient tools in landslide hazard modeling due to their high accuracy and ability to analyze large and diverse data. Also, the use of Google Earth Engine (GEE) as a cloud computing platform for analyzing remote sensing data provides the basis for fast, accurate, and spatially based analyses and enables better modeling at different scales.

Methodology

This study was conducted in the Firoozabad watershed, Ardabil Province, northwestern Iran, which is highly prone to landslides due to its rugged topography, lithological diversity, heavy precipitation, and human-induced land-use changes. A landslide inventory containing 300 confirmed events was compiled from field surveys, Google Earth image interpretation, and official reports. Twelve conditioning factors were selected: slope, aspect, elevation, lithology, land use/land cover, NDVI, topographic wetness index (TWI), curvature, distance to roads, distance to rivers, distance to faults, and precipitation. These were derived from DEMs, remote sensing products, and climatic datasets. The dataset was divided into training (70%) and validation (30%) subsets. The RF and SVM algorithms were implemented using GEE, supplemented with statistical analyses in R. Both models generated susceptibility maps, which were classified into five categories (very low, low, moderate, high, and very high susceptibility). Model performance was assessed using the area under the ROC curve (AUC) and classification accuracy.

Results and Discussion

Both RF and SVM successfully delineated landslide-prone zones, though their predictive accuracies differed. The RF model classified approximately 1059.72 km² (66.31%) as very high risk and 132.32 km² (8.28%) as high risk. The SVM model similarly identified 1059.89 km² (66.32%) as very high and 133.53 km² (8.35%) as high risk. While spatial patterns were comparable, the models differed in factor importance. RF identified slope (14%) and elevation (11%) as the most influential variables, whereas SVM emphasized slope (34%) and precipitation (10%). Validation metrics confirmed the superiority of RF. The AUC value for RF was 0.895, compared to 0.798 for SVM, indicating higher predictive reliability. This aligns with prior research suggesting RF’s robustness in handling nonlinear relationships, multicollinearity, and overfitting, making it more effective than SVM in landslide hazard modeling.

Conclusion

The study highlights the effectiveness of machine learning algorithms in landslide susceptibility mapping, with RF outperforming SVM in predictive accuracy for the Firoozabad watershed. Scientifically, these findings confirm that advanced algorithms can capture the complex interactions among environmental factors influencing landslide occurrence. Practically, the susceptibility maps produced can guide policymakers and land managers in prioritizing preventive measures, including restricting construction in high-risk zones, slope stabilization, improved drainage systems, and vegetation reinforcement. The study also acknowledges limitations, including reliance on static datasets for precipitation and land use, as well as uncertainties introduced by IDW interpolation of rainfall. Future studies should integrate high-resolution, time-series datasets and explore ensemble or deep learning methods to enhance dynamic susceptibility modeling. Overall, this research demonstrates that combining multi-source geospatial data, machine learning, and cloud-based platforms provides a powerful framework for landslide risk assessment, supporting sustainable watershed management and disaster risk reduction in Iran and similar regions worldwide.