The Southern Cross Island Highway was suffered from the landslide disaster in recent years. In order to reduce the reconstruction cost of broken road, the landslide susceptibility models were established in this study. The mileage ranging from 50 to 110 kilometers and buffer range of 1,400 meters of the Southern Cross Island Highway was selected as the study areas.
The inventories of landslide during typhoon Mindulle, Morakot, and 0719 rainfall event by Central Geological Survey were selected as the landslide data. The elevation, slope, slope aspect, slope high, lithology, terrain roughness, slope roughness, plan curvature, profile curvature, total curvature, distance of road, and distance of river were first chosen as the landslide causative factors, according to the previous studies. Secondly, the calibration and selection procedure were performed to select the factors efficiently. Each factor value was normalized from 0 to 1 for reducing the difference between the range and unit of various factors. Logistic regression method was used for establishing the landslide susceptibility model. Furthermore, the rainfall intensities of different rainfall duration were used as a landslide triggering factor in different rainfall events. The maps of potential landslide were delineated to discuss the influence of rainfall on the landslide susceptibility analysis. The landslide susceptibilities were separated into four levels, including high, medium, low, and steady. The classification error matrix and ROC-curve were used to evaluate the accuracy of landslide predicted by the present model.
As a result, the present model predicted landslide effectively. The overall accuracy of three rainfall events was higher than 70%. However, the results of Mindulle and 0719 rainfall event did not improve obviously by using landslide triggering factors. The overall accuracy of Morakot event increased 66.82% to 71.34% after considering the landslide triggering factor. Namely, the model with the rainfall factor increases the landslide predictive capability, once the rainfall event belonged to the type of long-duration and high-intensity, such as Morakot event. The other two events remain similar landslide predictive capability with and without the landslide causative factor in the model.