Selecting and weighting spatial predictors for empirical modeling of landslide susceptibility in the Darjeeling Himalayas (India)

In this paper, we created predictive models for assessing the susceptibility to shallow translational rocksliding and debris sliding in the Darjeeling Himalayas (India) by empirically selecting and weighting spatial predictors of landslides. We demonstrate a two-stage methodology: (1) quantifying associations of individual spatial factors with landslides of different types using bivariate analysis to select predictors; and (2) pairwise comparisons of the quantified associations using an analytical hierarchy process to assign predictor weights. We integrate the weighted spatial predictors through multi-class index overlay to derive predictive models of landslide susceptibility. The resultant model for shallow translational landsliding based on selected and weighted predictors outperforms those based on all weighted predictors or selected and unweighted predictors. Therefore, spatial factors with negative associations with landslides and unweighted predictors are ineffective in predictive modeling of landslide susceptibility. We also applied logistic regression to model landslide susceptibility, but some of the selected predictors are less realistic than those from our methodology, and our methodology gives better prediction rates. Although previous predictive models of landslide susceptibility indicate that multivariate analyses are superior to bivariate analyses, we demonstrate the benefit of the proposed methodology including bivariate analyses.

Research Highlights
► Spatial association methods can select suitable predictors of landslide susceptibility.
► Inter-predictor weights are determined using indices of spatial association and AHP.
► Predictors are integrated by weighted multiclass index overlay for susceptibility.
► Results are compared and evaluated with multivariate predictive mapping method.