Abstract: Landslide susceptibility assessment is the basis of disaster monitoring and early warning. How to scientifically and reasonably screen feature indicators and optimize assessment samples is still a difficult and easily ignored problem. The authors took Longshan County of Hunan Province as an example, and screened high-quality indicators by principal component analysis (PCA), correlation analysis and collinearity diagnosis, based on 15 feature indicators such as elevation, slope and slope direction. An optimize negative samples (ONS) method was proposed to construct assessment samples, and then certainty factor-random forest (CF-RF) model was used to map landslide susceptibility. The accuracy of prediction results was tested according to receiver operating characteristic (ROC) curve and rationality analysis. The results show that ONS-CF-RF model can significantly enhance the accuracy of model assessment. The area under curve (AUC) of this model has increased by 10.64% compared to the AUC of CF-RF model. The high landslide prone area is concentrated in human gathering area in the northwest corner of the study area, while the low landslide prone area is distributed in the high-altitude mountainous area which is less affected by human activities. The research results could provide scientific guidance for the prevention and control of landslide disaster in Longshan County, and could also provide references for the landslide prone zoning in the same type of region.