Spatial modeling of landslides in Khorasan Razavi Province using machine learning methods

Humanity is facing many environmental challenges. Natural disasters are among these problems, causing the death and injury of hundreds of thousands of people and rendering millions homeless worldwide every year. Geomorphological hazards, particularly mass movements and landslides, are considered some of the most potentially harmful phenomena. Mass movement refers to the outward or downward movement of domain-forming materials under the influence of gravity. These movements are primarily triggered by gravity, natural factors such as heavy rainfall, earthquakes, and soil saturation with water, as well as human activities like deforestation and improper engineering operations. According to the 2012 World Natural Hazards Report, landslides were ranked among the seven most dangerous natural disasters globally. The occurrence of natural hazards, including landslides, exerts considerable pressure on the economic development of countries, especially in developing regions, with financial damages hindering economic growth and prosperity. Iran, with its mountainous terrain, high tectonic and seismic activity, and diverse geological and climatic conditions, has created natural conditions conducive to a wide range of landslides.

The aim of this study is to spatially model landslide susceptibility using machine learning techniques, including random forest, support vector machine, and enhanced regression tree, in Razavi Khorasan province. Initially, the distribution map of landslides in the region was prepared through field visits and data from the national landslides database. In the next step, 70% of the identified landslides were used for model development, while 30% were reserved for model evaluation. The information layers for altitude, slope, slope direction, distance from waterway, waterway density, distance from roads, distance from faults, land use, vegetation index, surface curvature, profile curvature, precipitation, and selected lithological units were prepared and mapped using ArcGIS.

Prioritization of the factors affecting landslide occurrence using the random forest model showed that precipitation and altitude had the greatest impact on landslides in the study area. Additionally, the evaluation of the machine learning models using the relative operating characteristic (ROC) curve indicated that the landslide potential map generated by the random forest method had the highest accuracy (0.97). Based on this map, more than 25% of the area was classified into high and very high-risk zones.

This map can assist environmental planners in construction projects and help prevent land-use changes and construction in high-risk areas. Additionally, public awareness campaigns can reduce harmful human activities in these zones. While controlling landslides may not always be feasible or is often very expensive, proper management can help mitigate or reduce risks. By identifying the key factors in mass movement occurrences and zoning the areas accordingly, it is hoped that this research will contribute to the development of effective risk management plans and reduce the damage caused by landslides.