A Comparison of Alternative GIS Data Model Methods for Landslide Susceptibility Mapping with XGBoost and SHAP

Abstract

Geographic Information Systems and machine learning algorithms suggest good alternatives for producing landslide susceptibility maps. In the process of producing these maps with machine learning, alternative data model options exist. Success rate of analyses may change according to the preferred data method. In this study, 6 different machine learning models were created by passing different data models with the XGBoost algorithm. Study area is located in the cities of Ordu and Giresun, Turkiye. 14 different factors and related geographic data layers were used. As a result of the study, the most successful model performance was achieved by taking the average values of all pixels of the combined landslide record polygons (Accuracy=0,88, Precision=0,86, F1 score=0,87). SHAP method was applied for better interpretation of machine learning results The susceptibility map produced with the ideal model, overlapped with 57.556 buildings in the region. The buildings were classified in 4 groups (low, moderate, high, and very high) and mapped, indicating their risk level.