Improving Geospatial Agreement by Hybrid Optimization in Logistic Regression-Based Landslide Susceptibility Modelling

Abstract

This study aims to develop a logistic regression model of landslide susceptibility based on GeoDetector for dominant-factor screening and 10-fold cross validation for training sample optimization. First, Fengjie county, a typical mountainous area, was selected as the study area since it experienced 1,522 landslides from 2001 to 2016. Second, 22 factors were selected as the initial conditioning factors, and a geospatial database was established with a grid of 30 m precision. Factor detection of the geographic detector and the stepwise regression method included in logistic regression were used to screen out the dominant factors from the database. Then, based on the sample dataset with a 1:10 ratio of landslides and nonlandslides, 10-fold cross validation was used to select the optimized sample to train the logistic regression model of landslide susceptibility in the study area. Finally, the accuracy and efficiency of the two models before and after screening out the dominant factors were evaluated and compared. The results showed that the total accuracy of the two models was both more than 0.9, and the area under the curve value of the receiver operating characteristic curve was more than 0.8, indicating that the models before and after screening factor both had high reliability and good prediction ability. Besides, the screened factors had an active leading role in the geospatial distribution of the historical landslide, indicating that the screened dominant factors have individual rationality. Improving the geospatial agreement between landslide susceptibility and actual landslide-prone by the screening of dominant factors and the optimization of the training samples, a simple, efficient, and reliable logistic-regression–based landslide susceptibility model can be constructed.