Evaluation of the efficiency of artificial intelligence and bivariate statistical models in determining landslide prone areas in West Azerbaijan

Landslide is one of the natural hazards that lead to a lot of human and financial losses. Researchers on the subject of landslide susceptibility are investigating the possibility of landslides with respect to topographic and geo-environmental conditions, and the obtained information is critical in landslide risk management Preparation of landslide sensitive points is an essential tool for assessing landslide risk and is very useful in better planning and management of these areas. In this research, models based on artificial intelligence and two statistical variables in determining landslide sensitive points in West Azerbaijan province have been studied and compared.

Methods based on artificial intelligence and two statistical variables were used to prepare landslide-sensitive points in the province of West Azerbaijan, which is located in northwestern Iran. This study was conducted in four stages. The first stage: the study of landslides in the studied region based on the database of the Forests, Rangelands and Watershed Organization of Iran (FRWO) and the identification of 110 landslides through field surveys, interpretation of aerial photographs and Google Earth satellite images, the second stage: data collecting and creating a spatial databases of effective factors, the third stage: applying the Frequency Ratio (FR), Shannon Entropy (SE), Bagging (BA), Random Forest (RF) and hybrid model (RF-BA) and stage four: methods validating using the system performance curve (ROC). Based on field surveys and similar studies, 12 factors affecting landslide occurrence including altitude, slope angle, slope direction, distance from fault, distance from river, distance from road, drainage density, road density, rainfall, soil, land use and lithology were identified. In the field survey, 110 landslides were identified in West Azerbaijan. 70 percent of the data were randomly selected and used for modeling and 30 percent of the data were used for validation.

In terms of geographical directions, the southern direction with a weight of 1.49 had the greatest impact on the occurrence of landslides in the province. The least weight was related to flat areas where no landslide occurred. The results of slope factor showed that the middle slopes had the greatest effect on the occurrence of landslides, so that in low slopes due to low gravity, less landslides occur and too much slopes were related to mountainous areas that were covered with rocks and there was very thin soil that is not suitable for landslide. The study of land use factor showed that 48 percent of landslides occured in agricultural areas. The results showed that most of the landslides occurred near rivers and faults. Also, in some areas, the closest distances to the road had the greatest risk of landslide

The results of this study showed that the artificial intelligence models (RF and the combined model RF-BA) had the higher efficiency than the statistical models (FR and SE). The accuracy of the combined models was higher than the single models. The ROC curve results showed the accuracy of 0.92, 0.91, 0.89 and 0.88 with RF-BA, RF, FR and SE models, respectively.