Comparison of Landslide Susceptibility Evaluation by Deep Random Forest and Random Forest Model: A Case Study of Lueyang County, Hanzhong City

To address the problem of low prediction accuracy of landslide susceptibility evaluation model due to the difficulty of knowledge reuse and generalization of shallow machine learning model, this paper takes Lueyang County, Hanzhong City, Shaanxi Province as the study area and uses deep random forest to build a regional geological disaster susceptibility evaluation model to improve the prediction accuracy. Firstly, based on the research results of landslide mechanism in Lueyang County, nine factors such as slope, relative height difference, slope direction, slope type, engineering geological rock group, fault distance, river system distance, road and railroad distance, and vegetation cover are selected as susceptibility evaluation indexes; secondly, the study area is divided into 5 m × 5 m raster cells and the values of evaluation factors are extracted and input into the depth random forest evaluation model; finally, the susceptibility evaluation map of the study area is obtained. Based on the evaluation results, geological hazards in Lueyang County can be classified into four levels: very high susceptibility, high susceptibility, medium susceptibility, and low susceptibility, with the proportion of area being 5.31%, 22.97%, 42.11%, and 29.61%. The classification results are consistent with the actual development of geological hazards and reasonably reflect the overall characteristics of geological hazard distribution in the study area. In addition, the area under the ROC curve of the geological hazard susceptibility prediction model of deep random forest is 91.2%, which is higher than 86.3% of the random forest prediction model, indicating that the model is reasonable and feasible, and can provide new ideas for the evaluation of regional landslide susceptibility.