Abstract: Machine learning has been widely used in the evaluation of landslide susceptibility,and has achieved good performance. However,there are still many problems in the evaluation of large areas. The problems include a large number of database samples and high computing power. When the impact factors are classified,their correlation with the landslide mechanism is not considered. In order to reduce the demand for database samples,this paper proposes to construct a slope state that includes three states:a slope that has already experienced instability,a slope that is in an unstable state,and a slope with a low probability of instability,and a database of landslides. The critical value is divided numerically to highlight the landslide,which is convenient for the model to identify the landslide more accurately and greatly reduces the amount of data. Aiming at the problem of grading mechanization of influencing factors,a mathematical statistical method based on frequency distribution diagram,cumulative curve and its derivative diagram is proposed to describe the relationship between factors and landslide susceptibility in a more precise manner. Taking the landslide disaster in Xinjiang as an example,the applicability of the "database containing three slope states" and the "description method based on mathematical statistics" is verified. A susceptibility zoning map of Xinjiang is obtained. Under the premise of remaining the accuracy,the sample size is reduced by 90%. The description method based on mathematical statistics can draw a more detailed landslide risk zoning map. Landslide susceptibility in Xinjiang is mainly controlled by active faults and land surface fluctuations.