Study on landslide susceptibility mapping with different factor screening methods and random forest models

Abstract

The number of input factors affects the prediction accuracy of a model. Factor screening plays an important role as the starting point for data input. The aim of this study is to explore the influence of different factor screening methods on the prediction results. Taking the 2014 landslide inventory of Jingdong County as an example, a landslide database was constructed based on 136 landslide events and 11 selected factors, which were randomly divided into a training dataset and a test dataset according to a ratio of 7:3. Four factor screening methods, namely, the information gain ratio (IGR), GeoDetector, Pearson correlation coefficient and multicollinearity test (MT), were selected to screen the factors. A random forest (RF) model was then used in combination with each factor set for landslide susceptibility mapping (LSM). Finally, accuracy validation was performed using confusion matrices and ROC curves. The results show that factor screening is beneficial in improving the accuracy of the resulting model compared to the original model. Second, the IGR_RF model had the highest AUC value (0.9334), which was higher than that of the MT_RF model without factor screening (0.9194), and the IGR_RF model predicted the most landslides in the very high susceptibility zone (51.22%), indicating the good prediction performance of the IGR_RF model. Finally, the factor weighting analysis revealed that NDVI, elevation and aspect had the greatest influence on landslides in Jingdong County and that curvature had the least influence on landslides. This study can provide a reference for factor screening in LSM.