不同机器学习模型和输入变量的崩岗易发性评价适宜性研究OA

[Objective]To investigate the suitability of different machine learning models and input variables for assessing collapsing gully susceptibility.[Methods]Taking Shicheng County,Jiangxi Province as the study area,an indicator system was constructed using geodetector(GD)for factor screening.The original values,frequency ratio(FR),and neighborhood frequency ratio(NFR)were used as input variables for the multilayer perceptron(MLP)and random forest(RF)models.The adaptability of these different models and input variables for Benggang susceptibility assessment was studied.[Results]1)The AUC values of susceptibility assessment results from MLP and RF models under the NFR input variables were 0.854 and 0.860,respectively.Both models demonstrated good assessment performance,indicating that NFR was a suitable input variable.2)The RF model generally outperformed the MLP model.Specifically,the Benggang densities in high susceptibility areas of original value-RF,NFR-RF,and FR-RF models were 3.93,3.83,and 3.69,respectively.The original value-RF model demonstrated the strongest capability in identifying extremely high and high susceptibility areas.3)The Benggang density was highest in the extremely high susceptibility area.Both high and extremely high susceptibility areas were concentrated in the northwest,closely matching the actual distribution pattern of Benggang.[Conclusion]NFR is a highly generalizable input variable.Compared with original values and FR,NFR exhibits the highest robustness in both MLP and RF models.The RF model is more suitable than the MLP model for assessing Benggang susceptibility.