Effects of variable selection on the landslide susceptibility assessment using machine learning techniques

Abstract

This study aims to produce landslide susceptibility map (LSM) using landslide conditioning attributes selected by different feature selection methods and compare predictive capability. Among the total 140 landslide locations, 98 locations (70%) were selected randomly for model training and remaining 42 locations (30%) were used to validate. Fourteen landslide conditioning attributes related to topography, hydrology, and forestry factors were considered. These factors were analyzed importance using four feature selection methods, such as information gain, gain ratio, Chi-squared, and filtered subset evaluator. From the results, the top seven attributes were selected and the LSMs were produced by random forest model. The results showed that the all LSMs had a prediction rate of more than 0.80 that yielded higher accuracy than the LSMs produced using all attributes. In addition, the LSM produced using attributes selected by gain ratio performed slightly better than another LSMs. These results indicate that the produced LSMs had good performance for predicting the spatial landslide distribution in the study area. In addition, selection of input attributes using feature selection methods was contributed to improve model performance. The produced LSMs could be helpful for establishing mitigation strategies and for land use planning in the study area.