Landslide hazard zoning using random forest and support vector machine models (Case study: Talar basin)

Introduction
Landslide is a geomorphological phenomenon with high potential for human and financial losses that occurs due to the sudden and rapid movement of soil, rock, and other materials on low to medium slopes down the slope. This phenomenon is considered one of the most destructive natural disasters in steep areas. Given the importance and sensitivity of the subject, this article is dedicated to landslide risk zoning in the Talar Basin. Studies conducted in countries with similar climatic conditions, global experiences, and perspectives on landslide risk zoning can best help in planning for management and damage reduction. Nearly 70 percent of Iran's land area is above 1000 meters in altitude (Zomordian, 2004) and 34 percent has a slope of more than 10 percent (Jedari Eyvazi, 1995). This indicates that Iran is a mountainous country and prone to slope movements. Active neotectonics, together with young Tertiary geological formations and climatic conditions, provide the basis for facilitating slope movements, especially landslides. Meanwhile, the Talar watershed is also prone to landslides due to its foothill and mountainous conditions in the central and southern parts of the basin. The high altitudes of the central Alborz and deep valleys are the main features of the ruggedness of this region. The vegetation of the region includes Hyrcanian forests, mountain pastures, and agricultural lands. The climate of the Talar basin is temperate and humid. Annual precipitation in this region varies on average between 200 - 1000 mm, and most precipitation occurs in the autumn and winter seasons. Due to its abundant rainfall and specific topography, this region is prone to floods and landslides. The Talar watershed has significant geological diversity. Limestone and schist are among the dominant rocks in this region. Due to the presence of active faults and complex geological structures, the potential of the region for landslides is high. Land use changes, deforestation, water pollution, and Mass movements are among the main problems of this region. Conservation and management measures are necessary to reduce the negative effects of these problems. Given the diversity and complexity of the natural, climatic, and human factors of this region, accurate knowledge of environmental characteristics can help in better planning and management of natural and environmental resources. The aim of the present study is to use past landslide data and investigate the factors affecting the occurrence of this geomorphological phenomenon, to prepare a landslide zoning map using random forest and support vector methods, and to further evaluate the accuracy and efficiency of the aforementioned models in predicting and identifying landslide-prone areas using the ROC curve. Unfortunately, due to the high sensitivity of the formations of the northern Alborz range and the tendency of non-native immigrants to settle and spend their leisure time in the mountainous and summer slopes overlooking the plain in the Talar basin, conditions have been created for drastic changes in land use, which in the event of a landslide will cause high and irreparable losses in terms of life and property.
Therefore, investigating and identifying landslide hazard zones can be effective in raising awareness and controlling landforms in the face of slope processes in the Talar basin. On the other hand, the aforementioned models have been used separately or in combination with other models in landslide hazard zoning, but a comparison of the efficiency of the two aforementioned models to identify the best model in an area with semi-humid environmental conditions in the north of the country has not been conducted.
 
Materials and Methods
The Talar watershed is one of the ten largest basins in the central part of Mazandaran Province, with an area of ​​3227.4 square kilometers, located in Mazandaran Province and south of the city of Qaemshahr.  The minimum and maximum elevation of this basin is between -26 and 4002 meters above sea level. The average elevation of the basin is 767 meters. The ruggedness units of the basin can be divided into three classes: plain (up to 200 meters), hills (200-500 meters), and mountains (more than 500). Since various factors are effective in landslide occurrence, in this study, natural factors and variables such as topographic factors including altitude, slope, slope direction, topographic moisture index, profile and plan curvature, along with distance from rivers, distance from faults, lithology, precipitation, normalized difference vegetation index (NDVI) and human factors such as distance from road and land use have been used. To produce the altitude, slope, slope direction, topographic moisture index, profile curvature, plan curvature and distance from rivers layers, the United States Geological Survey (USGS) digital elevation model with 30-meter pixel size was used. To produce the distance from faults and lithology layers, the 1/100,000 geological map of the Geological Survey of Iran was used. To produce the precipitation contour layer, data from rain gauge stations in the Talar basin were used. To prepare the vegetation layer (NDVI), the vegetation index was used, and to prepare the land use layer, the Landsat OLI satellite images were used with the supervised maximum likelihood classification method in Envi5.6 software. Using the Global Positioning System (GPS), the distribution layer of landslide occurrence points was prepared. To do this, first, by field surveying, the location of landslides that were physically accessible was recorded by GPS. Of course, due to the topographic conditions of the study area and due to their impassability and inaccessibility, Google Earth software was used to identify some landslide points, and finally, 61 landslide points were collected and recorded in the study area. 70% of the points (43 points) were randomly selected as training points and 30% of the points (18 points) were randomly selected for model validation. Training data was used for modeling and validation data was used for accuracy. Two methods, random forest and support vector machine, were used for landslide zoning and modeling.
 
Results and Discussion
In the study of natural and human variables, it was determined that 56% of landslides occurred at altitudes between 436 and 1126 meters. More than 78% of landslides are observed on slopes above 25 degrees. 49% of the landslides occurred in the east, northeast and 38% in the south and southwest directions. 75% of landslides occurred on shale, sandstone, conglomerate, and marl rocks in TRJs and Mmsl formations. Slopes with low to medium curvature with a frequency of 50% show the highest overlap with landslides. In terms of soil topographic moisture index, 80% of landslides are in values ​​above 1.6 to 8 indices. About 64% of landslides are located within 0 to 1.3 km of rivers, and 51% of landslides in the study area occur within 0 to 600 m of roads. About 78% of landslides are located within 2.5 km of faults. Approximately 50 and 45% of recorded landslides occur in pasture and forest areas in the study basin, respectively. Due to the high rainfall in the entire basin, landslide hotspots are visible, but in terms of landslide frequency, 30% of landslides occur in areas with rainfall greater than 581 mm. According to the zoning results and the information received, in the RF model, the highest risk class belongs to very low risk with a frequency of 54%, and the high risk class in this model is 18.4%. However, in the SVM model, the very low class shows the highest spread with a frequency of 40.6%, and the very high risk class accounts for 11% of the entire basin. 70% of the 61 landslide occurrence points were randomly selected for model training and 30% for model validation to evaluate the accuracy of the model with data that were not used in the training process. The evaluation results in the study area showed that the area under the curve in the SVM and RF models is 12.87 and 27.85, respectively. According to the classification provided for the area under the curve (excellent 0.9-1, very good 0.8-0.9, good 0.7-0.8, moderate 0.6-0.7, and poor 0.5-0.6), both models have very good accuracy, but in comparison, it can be said that the landslide zoning obtained from the support vector machine model in the study area has a higher level of accuracy. It seems that the slight difference in the validity of the models under study is related to the difference in the number and weighting of the criteria, as well as the difference in the climatic conditions of the study basins.
Conclusion
The studied area with 61 landslide zones is a very sensitive area to landslides, which has increased significantly due to increasing human activities. To evaluate landslide-prone areas, 12 factors affecting landslide occurrence (elevation, slope, slope direction, precipitation, distance from the river, distance from the fault, distance from the road, vegetation index, soil topographic moisture, curvature index, lithology and land use) were used. The overlap of landslide points and layers of effective factors showed that the lower parts of the basin due to the gentle slope and low altitude, the type of constituent rocks, are less sensitive to landslide occurrence, and in contrast, the middle and upstream parts of the Talar basin due to foothill and mountainous conditions, with high altitude and slope and sensitive formations, indicate much more favorable conditions for landslide occurrence. The results of the random forest model indicate very high and high risk areas as 18.4 and 14.9 percent, respectively. While the results of the support vector machine model indicate Listed hazard classes areas as 11 and 16.7 percent of the area, respectively. In the Talar basin, the risk classes mainly coincide in the southern and central highlands with steep slopes and on the edges of rivers and roads. Comparison of the aforementioned models showed that both models have high efficiency in landslide occurrence zoning, but the results of ROC curve evaluation showed that the area under the curve obtained for the support vector machine and random forest models are 87.1 and 85.3, respectively. According to the classification provided for the area under the curve, the support vector machine model has a higher accuracy in landslide susceptibility zoning in the study area. According to the results obtained from the two models, on average more than 40 percent of the basin is at medium to very high risk of landslide occurrence, and the need for planning for basin management and paying more attention to this phenomenon seems essential.