University of Sistan and Baluchestan Journal of Natural Environmental Hazards 2676-4377 14 45 2025 09 23 Evaluation of Random Forest and Support Vector Machine Models in Landslide Risk Mapping (Case study: Tajan Basin, Mazandaran Province) Evaluation of Random Forest and Support Vector Machine Models in Landslide Risk Mapping (Case study: Tajan Basin, Mazandaran Province) 133 154 8783 10.22111/jneh.2025.50031.2071 FA Sayed Hamid Sadati PhD Student of Watershed Management, Department of Watershed Engineering, Faculty of Natural Resources, Sari Agricultural Sciences and Natural Resources University, Sari, Iran Seyed Ramazan Mousavi Assistant Professor of Watershed Management, Department of Watershed Engineering, Faculty of Natural Resources, Sari Agricultural Sciences and Natural Resources University, Sari, Iran Ghorban Vahabzadeh Kebria Associate Professor of Watershed Management, Department of Watershed Engineering, Faculty of Natural Resources, Sari Agricultural Sciences and Natural Resources University, Sari, Iran Sayed Hussein Roshun PhD Graduate of Watershed Management, Department of Watershed Engineering, Faculty of Natural Resources, Sari Agricultural Sciences and Natural Resources University, Sari, Iran Journal Article 2024 10 13 The development of landslide susceptibility maps using machine learning is an effective tool for managing land in vulnerable regions. This study generates a landslide susceptibility map for the Tajan watershed using machine learning techniques. Twenty-one factors influencing landslides were identified and categorized into geological, climatic, environmental, topographical, and hydrological factors. Raster data was prepared using ENVI 5.6, SAGA GIS, and ArcGIS software. Field surveys documented 155 landslide locations, converted to point layers in ArcGIS. This data, along with the training layer, was imported into R software in ASCII format. For model training, Support Vector Machine (SVM) and Random Forest (RF) algorithms were applied, using 70% of the data (109 samples) for training and the remaining 30% (46 samples) for testing. Evaluation of the RF model using the ROC curve showed high predictive accuracy, with scores of 0.972 for training and 0.949 for testing. Analysis of the RF model identified key factors influencing landslides, including aspect, distance from streams and roads, slope, and the Topographic Position Index. The SVM model results indicated a greater proportion of high-susceptibility areas in the watershed than the RF model. AUC values for the SVM model were slightly lower, at 0.906 for training and 0.831 for testing. The SVM model highlighted elevation classes, rainfall, aspect, and distance from streams and roads as significant factors but underperformed compared to the RF model in mapping landslide susceptibility. Risk classification with the RF model showed that 10.19% of the area is very high risk, 4.17% high risk, 10.76% moderate risk, 15.62% low risk, and 59.26% very low risk. Conversely, the SVM model predicted smaller very high-risk areas at 5.51%, high risk at 15.58%, moderate risk at 5.33%, low risk at 4.47%, and very low risk at 69.09%. The development of landslide susceptibility maps using machine learning is an effective tool for managing land in vulnerable regions. This study generates a landslide susceptibility map for the Tajan watershed using machine learning techniques. Twenty-one factors influencing landslides were identified and categorized into geological, climatic, environmental, topographical, and hydrological factors. Raster data was prepared using ENVI 5.6, SAGA GIS, and ArcGIS software. Field surveys documented 155 landslide locations, converted to point layers in ArcGIS. This data, along with the training layer, was imported into R software in ASCII format. For model training, Support Vector Machine (SVM) and Random Forest (RF) algorithms were applied, using 70% of the data (109 samples) for training and the remaining 30% (46 samples) for testing. Evaluation of the RF model using the ROC curve showed high predictive accuracy, with scores of 0.972 for training and 0.949 for testing. Analysis of the RF model identified key factors influencing landslides, including aspect, distance from streams and roads, slope, and the Topographic Position Index. The SVM model results indicated a greater proportion of high-susceptibility areas in the watershed than the RF model. AUC values for the SVM model were slightly lower, at 0.906 for training and 0.831 for testing. The SVM model highlighted elevation classes, rainfall, aspect, and distance from streams and roads as significant factors but underperformed compared to the RF model in mapping landslide susceptibility. Risk classification with the RF model showed that 10.19% of the area is very high risk, 4.17% high risk, 10.76% moderate risk, 15.62% low risk, and 59.26% very low risk. Conversely, the SVM model predicted smaller very high-risk areas at 5.51%, high risk at 15.58%, moderate risk at 5.33%, low risk at 4.47%, and very low risk at 69.09%. https://jneh.usb.ac.ir/article_8783_d006d23dfbdc681cc24a7790354b3152.pdf