Using the hourly precipitation data from automatic meteorological stations and meteorological observation stations corresponding to 61 landslides from 2010 to 2014 in Tongren, the eastern Guizhou, the types of rainfall before and after landslide are analyzed by using statistical analysis. The prediction method of landslides is discussed by using different combinations between cumulative rainfall and triggering rainfall. It is found that the forecast accuracy of the model is the best, and the false alarm rate is the smallest when using cumulative rainfall from two days before the day landslide occurred and the triggering maximum rainfall in 3 hours of the day landslide occurred. The model using the cumulative rainfall of the day landslide occurred and the triggering maximum rainfall in 3 hours of the day landslide occurred takes the second place. The discriminate curve whether the landslide occurred can be obtained in both cases. A landslide can be predicted according to the discriminate curve and precipitation forecasts for the next 24 hours or 3 hours.