Optimization of the Landslide Susceptibility Model based on Deep Forest and Twice-Sample

In the analysis of regional landslide susceptibility, the geological data used usually have the characteristics of multiplicity and nonlinearity. At the same time, in the past researches, the selection of non-landslide units in the training set is usually random or subjective, and there is no guarantee sample accuracy. In this paper, the deep forest algorithm combined with the twice-sample method is used to evaluate the landslide susceptibility in the Zigui Badong area of the Three Gorges Reservoir area. First, 11 landslide impact factors were selected according to the geological data to initially construct a sample set, and the cleaned data set was obtained by twice-sample. After dividing the training set and the test set, the deep forest model was used to predict the landslide susceptibility and evaluate the accuracy. The results show that the AUC of the susceptibility prediction model built using the deep forest is 1%-10% higher than that of GBDT, MLP and other methods, while the Twice-Sample method improves the AUC by 2%.