基于证据权和卡方自动交互检测决策树的滑坡易发性预测

Landslide susceptibility prediction modeling based on weight of evidence and chi-square automatic interactive detection decision tree

滑坡与其环境因子间的非线性关联计算影响滑坡易发性预测建模的不确定性。为研究不确定性因素下易发性建模规律,以中国延长县为例,获取82处滑坡和14种环境因子,通过频率比(Frequency Ratio,FR)和证据权(Weight of Evidence,WOE)等关联法与卡方自动交互检测(Chi-squared Automatic Interaction Detector,CHAID)决策树相耦合进行建模,并用原始环境因子(称为“原始因子数据”)作为输入变量的单独CHAID决策树进行对比。使用精度、易发性指数均值、标准差和平均秩等评价易发性建模的不确定性。结果表明:WOE-CHAID模型预测的滑坡易发性不确定性低于FR-CHAID模型,可见WOE具有较优秀的非线性关联性能;单独CHAID决策树预测的易发性精度整体略低于WOE-CHAID和FR-CHAID模型,但其建模效率较高;在体现滑坡与其环境因子空间关联性方面,考虑FR和WOE关联法的CHAID决策树模型优势显著。WOE是更优秀的关联分析法,CHAID决策树预测性能好且预测效率高,WOE-CHAID决策树模型的易发性预测不确定性较低且更符合实际滑坡概率分布特征。

The calculation of the non-linear correlation between the landslide inventories and their environmental factors is an important factor that affects the uncertainty of the landslide susceptibility prediction(LSP)modeling.In order to study the changing patterns of LSP under the influence of the uncertain factors,taking Yanchang County of China as example,82 landslides and 14 environmental factors are obtained,and the frequency ratio(FR)and weight of evidence(WOE)connection methods are coupled with the chi-squared automatic interaction detector(CHAID)decision tree model to carry out LSP.Then the original environmental factors data(hereinafter referred to as"original data")is used as the input variable to compare the individual CHAID decision tree model to realize the analysis of LSP modeling pattern.ROC accuracy,mean,standard deviation,and average rank are adopted to analyze the uncertainty characteristics in the LSP modeling process.Results show that:1)LSP uncertainty of the WOE-CHAID model is lower than that of the FR-CHAID model,and WOE has relatively excellent nonlinear correlation performance.2)The prediction accuracy of individual CHAID decision tree model is slightly lower than that of the WOE-CHAID and FR-CHAID models,but it has higher modeling efficiency.3)In terms of reflecting the spatial correlation between landslides and its environmental factors,the CHAID decision tree model coupled with FR and WOE connection methods have significant advantages.Generally,WOE is a better connection method and CHAID decision tree model has good prediction performance and high prediction efficiency.Susceptibility prediction by the WOE-CHAID decision tree model is less uncertain and more in line with the actual landslide probability distribution characteristics.