Spatial Non-Stationarity-Based Landslide Susceptibility Assessment Using PCAMGWR Model

Yange Li,Weidong Wang,Jiaying Li,Shuangfei Huang,Jianling Huang

doi:10.3390/w14060881

Abstract

Landslide Susceptibility Assessment (LSA) is a fundamental component of landslide risk management and a substantial area of geospatial research. Previous researchers have considered the spatial non-stationarity relationship between landslide occurrences and Landslide Conditioning Factors (LCFs) as fixed effects. The fixed effects consider the spatial non-stationarity scale between different LCFs as an average value, which is represented by a single bandwidth in the Geographically Weighted Regression (GWR) model. The present study analyzes the non-stationarity scale effect of the spatial relationship between LCFs and landslides and explains the influence of factor correlation on the LSA. A Principal-Component-Analysis-based Multiscale GWR (PCAMGWR) model is proposed for landslide susceptibility mapping, in which hexagonal neighborhoods express spatial proximity and extract LCFs as the model input. The area under the receiver operating characteristic curve and other statistical indicators are used to compare the PCAMGWR model with other GWR-based models and global regression models, and the PCAMGWR model has the best prediction effect. Different spatial non-stationarity scales are obtained and improve the prediction accuracy of landslide susceptibility compared to a single spatial non-stationarity scale.

Highlights

Landslide susceptibility refers to the occurrence possibility of a landslide under the combined effect of Landslide Conditioning Factors (LCFs) and predicts the location and probability of a landslide in a specific area [4]
Multivariate statistical analysis is mainly expressed in regression models, which consist of global regression models, such as logistic regression [20], and local regression models, such as Geographically Weighted Regression (GWR) [21]
Geospatial data may lead to the spatial non-stationarity process, and the scale at which each independent variable affects a dependent variable may vary according to the independent variables

Summary

Introduction

Landslides are one of the most destructive and catastrophic geohazards worldwide and threaten the safety of humans and property in mountainous areas [1,2]. Multivariate statistical analysis is mainly expressed in regression models, which consist of global regression models, such as logistic regression [20], and local regression models, such as Geographically Weighted Regression (GWR) [21]. The local regression models consider spatial variation in LCFs and landslides. The problem of spatial autocorrelation and spatial non-stationarity is well-known in LSA research. Since LSA is based on geospatial data analysis, the relationship between LCFs and landslides may be spatial non-stationarity [24]. The presence of spatial non-stationarity demonstrates that the traditional models are only applicable to the case of stationary spatial relations in the study area and cannot accurately fit the local relations [25]. The models which are widely used in LSA cannot express spatial non-stationarity

Methods

Results

Discussion

Conclusion