Comparative study on landslide susceptibility mapping based on different ratios of training samples and testing samples by using RF and FR-RF models

• DOI: 10.1016/j.nhres.2023.07.004
• Publication Date: 2023-07-21T02:34:32Z
• Authors: Ke Xu; Zhou Zhao; Wei Chen; Jianquan Ma; Fei Liu; Yihao Zhang; Zijun Ren
• Keywords: Lithology

Evaluation of landslide susceptibility is essential to planning land and space utilization. For this purpose, the paper presents a case study from Fugu County, Shaanxi Province, China. Firstly, geological environment current state landslides in County were investigated. Then, slope, aspect, terrain relief, curvature, lithology, type, normalized difference vegetation index (NDVI) considered as condition factors, correlation between these carried out by using Multicollinearity Analysis method. Next, non-landslide samples divided into training testing according sample ratios 8/2, 7/3, 6/4, 5/5, respectively. The mapping was Random Forest (RF) model Frequency Ratio coupled with (FR-RF) model, Lastly, density (LD), frequency ratio (LFR), area under curve (AUC) receiver operator, other indicators used validate rationality, accuracy, performance maps produced different models ratios. results indicated that all are reasonable, except map when 5/5. each map, regardless ratios, LD LFR greatest zones classed having very high susceptibility, followed those high, moderate, low, low classes. In test set not at same time its extremely sensitivity size FR value respectively 7/3 (201.026) > 8/2 (154.440) 6/4 (93.696) >5/5 (136.364) (4.806) (3.692) (3.260) 5/5 (2.240); Inall sets proportion (145.693) (127.151) (122.857) (113.263) (3.334) (3.073) (2.811) (2.592). What else, comparison ROC curves, accuracy two higher than Similarly, ensemble (A combination learning abilities.) more reasonable single which reflects weaker learner (Frequency here) stronger (Random can diminish model.