ABSTRACT
The effect of influencing factors on landslide occurrences was considered as a constant for the whole study area using statistical-based landslide susceptibility assessment methods. However, spatial variability in the effect may exist at different parts of the study area. The purpose of this study attempts to explore the feasibility of applying the Geographically Weighted Logistic Regression (GWLR) model for earthquake-induced landslide susceptibility mapping on different geographic scales. The Logistic Regression (LR) and GWLR models were applied in Wenchuan County and Caopo Town, which were severely affected by the Wenchuan earthquake on May 12th, 2008, for landslide susceptibility mapping. The landslide inventory was randomly divided into training and validation datasets. Nine landslide influencing factors were considered, including elevation, slope, relative relief, curvature, Newmark displacement values, Topographic Wetness Index (TWI), distance from drainages, distance from faults and distance from roads. Different statistics such as Akaike’s Information Criterion (AIC), Receiver Operation Characteristic (ROC) curve and confusion matrix were used to assess the modeling results. The results show that five explanatory variables, namely, elevation, relative relief, distance from drainages, distance from faults, and distance from roads, and three explanatory variables, namely, elevation, slope, and distance from faults, were finally selected based on the stepwise regression and the multi-collinearity detection for Wenchuan County and Caopo Town, respectively. These explanatory variables exhibit spatial variability in the influence on the occurrence of landslides. The comparison results of the GWLR and LR models on two geographic scales both demonstrated that the GWLR models increase the explanation capability of the landslide susceptibility assessment model and are suitable for landslide susceptibility mapping at various geographic scales. Furthermore, the evaluated statistics showed that the GWLR models yield a significant improvement compared to the LR models of Wenchuan County both in training and validation process but only show a minor improvement in Caopo Town. This suggests that the GWLR model performs more efficiently compared to the LR model at larger spatial scales.
