Determination of effective predisposing factors using random forest-based gini index in landslide susceptibility mapping

Abstract

Landslide susceptibility mapping is a multi-phase procedure that includes several key steps, one of which is the correct determination of landslide predisposing factors. In the current literature, however, there is no global consensus or framework about the selection of these factors. In this study, the effectiveness of predisposing factors was investigated using the Random Forest-based Gini index to generate landslide susceptibility models. For this purpose, 16 predisposing factors, representing the morphological, lithological, and environmental characteristics of the study area, were initially utilized and measured their importance scores calculated by utilizing the Gini index. Then, three models (RF-1, RF-2, and RF-3) including 50%, 75%, and the whole of the factors, were produced based on the importance scores. To select the optimum one among these models, their performances were assessed employing two accuracy assessment metrics, namely overall accuracy (OA) and area under curve (AUC). The validation results revealed that AUC obtained using RF-1, RF-2, and RF-3 models were calculated as 85.85%, 96.70%, and 90.66% respectively. Also, the statistical significances of the models were evaluated using McNemar’s test, which revealed that all models were statistically different from each other.