Landslide is one of the most serious natural disasters that cause casualties and economic loss. In Taiwan, most of the landslides are triggered by heavy rainfall brought by typhoons and monsoons. Thus, the evaluation of landslide susceptibility and the establishment of early warning threshold for rainfall-induced landslide are important issues. This study uses a great quantity of SPOT images before and after 16 typhoon events from 1996 to 2011 for the interpretation of landslides induced by rainfall. We also extract geomorphological characteristics such as gradient, slope roughness, terrain curvatures from 5-meter resolution DEM, and geological characteristics such as rock strength, dip slope, fault density, fold density from 1:50,000 geological map. Besides, rainfall data of each typhoon event is also collected as triggering factor.

Logistic regression is adopted and predisposing factors mentioned above are used for landslide susceptibility analysis in this study. For the purpose of establishing early warning threshold for landslides, this study chooses 24-hour accumulated rainfall (R₂₄) and the 3-hour mean rainfall intensity (I₃) as long-term and short-term rainfall index respectively. The early warning signals are divided into 4 categories including red, orange, yellow and green according to the concept of hazard matrix which contains the magnitude of landslide (landslide susceptibility and landslide ratio of slope units) and the occurrence possibility of landslide (30%, 60% and 90% thresholds determined by historical disaster records). Validation with landslides caused by Typhoon Soudelor (6^th to 9^th, August, 2015) in northern Taiwan shows that these thresholds can issue warning signals 1 to 3 hours prior to the occurrence of landslides. This may provide crucial information for the evacuation and reduce damages. An early-warning system is also built in this study for regional land-use planning and disaster prevention.