Stochastic procedure to extract and to integrate landslide susceptibility maps: an example of mountainous watershed in Taiwan

Abstract

The Generalized Likelihood Uncertainty Estimation (GLUE) is here incorporated into a deterministic landslide model (SHALSTAB) to generate 4000 landslide susceptibility maps which enclose various combinations of full range parameters. Furthermore, an improved index is adopted into GLUE as a criterion to measure model performance, and through that, 200 maps holding top 5% performance are retrieved. Proper ranges for parameters are obtained through GLUE yet they only perform well if combined appropriately. The 200 better maps are overlapped to construct an integrated landslide susceptibility map. Instead of giving a single parameter set or a single susceptibility map, the merit of extracting and integrating procedure is to envelope uncertainties inherited in model structure and input parameters. Bias due to subjective parameter input is potentially reduced. The entire procedure is applied to the Chi-Jia-Wan, a mountainous watershed in Taiwan. The integrated map shows high-risk area (>50% predicted landslide probability) only occupies 16.4% of the entire watershed while able to correctly identify 60% of the actual landslides. For areas above 2100 m height the map is even more successful (projects 77 of the 98 actual landslides). Interactions among parameters are discussed to highlight the unsolvable equifinality problem and improperness of presenting a single model result.