[A-HW21] Hydrological modelling to support water resources management and engineering designs

[AHW21-02] Application of Bayesian Approach on Prediction of Rainfall-Induced Landslide Events

★Invited Papers

*Chih Mei Lu¹、Li Chi Chiang¹ (1.Department of Bioenvironmental Systems Engineering, National Taiwan University)

キーワード：Bayesian spatial statistics analysis, landslide, SWAT, landslide factors

Slope failures are influenced by various factors such as geology, vegetation cover, rainfall amount and intensity, soil infiltration capacity, and human development. Specifically, rainfall-triggered landslides are predominantly linked to geological conditions, soil properties, and land use. Most researches have been focusing on analyzing landslide susceptibility in specific regions, while comparatively less attention has been devoted to evaluating the reoccurrence of landslides at the same location. This study aims to construct a hydrological model, Soil and Water Assessment Tool (SWAT), with an integration of landslide information, and assess the influence of various geophysical and hydrological factors on the reoccurrence of rainfall-induced landslide events by the Bayesian spatial statistics analysis packages (brms and tidybayes) in the R language. The Xiuguluan River basin in eastern Taiwan is selected as the study area. The observed landslide data covers the period of 2004 - 2022, including occurrence timing, location, and extent of each event. The geophysical and hydrological factors include land use/land cover (LULC), susceptibility of soil to landslides (SOIL_LS), area of the simulation unit (AREA), average slope (AVG_SLOPE), subbasin elevation (SubElev), daily rainfall (PCP), two-day cumulative rainfall (PCP2D), soil moisture ratio (SMR), soil moisture change (SMC), and percolation (PERC). In this study, non-landslide and landslide areas were assigned with 0 and 1 in Bayesian analysis, respectively. Normalized factors were used to identify the individual impact, and further estimate both group-level effects and fixed effects. The group-level effect is determined by calculating the effectiveness of a factor to identify discrete groups in landslide prediction. A higher value indicates a more pronounced clustering effect. On the other hand, the fixed effect reflects the impact of one factor among clusters on the likelihood of landslide occurrence. A fixed effect value greater than zero suggests the likelihood of landslide occurrence increases with a higher factor value, while a fixed effect value less than zero indicates the likelihood of landslide occurrence increases with a lower factor value. The preliminary results of spatial analysis on factors and landslide events showed that LULC, SOIL_LS, and AREA exhibited significant clustering effects (group-level effect = 0.13 – 0.36) on landslide types classification with reducing inter-group differences. However, factors such as SubElev, PCP, and PCP2D (group-level effect = 0.00) could not effectively classify the rainfall-induced landslide events. Moreover, factors typically considered directly related to landslide events, such as PCP, PCP2D, PERC, SMR, and SMC, were negatively correlated or unrelated to the occurrence of landslides (fixed effect = -0.636 – 0.002). It is suggested that factors directly related to rainfall (e.g., PERC, SMR, and SMC) are influenced by the input rainfall data, highlighting the importance of data quality from rainfall stations. Due to lack of weather stations in mountainous regions in the Xiuguluan River basin, rainfall distribution within the basin was further estimated by Thiessen polygon method with observed rainfall data of available 9 weather stations to ensure simulated factors aligning with real conditions and improve landslide prediction.