Comparing shallow landslide susceptibility maps in Northeastern Türkiye (Beşikdüzü, Trabzon): a multivariate statistical, machine learning, and physical data-based analysis

Abstract

This study endeavors to assess and compare the efficacy of various modeling approaches, including statistical, machine learning, and physical-based models, in the creation of shallow landslide susceptibility maps within the Besikduzu district of Trabzon province, situated in the Black Sea Region of Türkiye. The landslide inventory data, spanning from 2000 to 2018, was acquired through meticulous field surveys and analysis of Google Earth satellite imagery. Key topographic and geologic input parameters, such as slope, aspect, topographic wetness index, stream power index, plan and profile curvature, and geologic units, were extracted from a high-resolution 10 m spatial DEM (Digital Elevation Model) and a 1:25,000 scaled digital geology map, respectively. Additionally, soil unit weight and shear strength parameters, critical for the physical-based model, were determined through field samples. To evaluate landslide susceptibility, logistic regression, random forest, and Shalstab were employed as the chosen methods. The accuracy of susceptibility maps generated by each method was assessed using the area under the curve method, yielding impressive values of 0.99 for the random forest model, 0.97 for the logistic regression model, and 0.93 for the Shalstab model. These results underscore the robust performance of all three methods, suggesting their applicability for generating shallow landslide susceptibility maps not only in the Black Sea Region but also in analogous areas with similar geological characteristics.