基于BPNN-SHAP模型的滑坡危险性评价:以伊犁河流域为例

Landslide hazard evaluation based on BPNN-SHAP model:A case study of the Yili River Basin,Xinjiang Province

为进一步提高滑坡危险性预测模型精度、增强模型可解释性,本文以新疆伊犁河流域为研究区,选取8个影响滑坡发生的危险性因子,在反向传播神经网络(BPNN)基础上,借鉴博弈论思想,构建一种可解释BP神经网络模型(BPNNSHAP),解决神经网络滑坡危险性评价的“黑箱”问题。将数据集分为70%训练集和30%测试集,采用5折交叉验证提高模型稳定性,对比深度神经网络(DNN)、随机森林(RF)和逻辑回归(LR)3个模型的评价精度,并探讨BPNNSHAP预测结果的可解释性,完成区域滑坡危险性评价。研究结果表明:相较于其他模型,BPNN-SHAP模型的5个精度评价指标均为最高,分别是:准确率(A)=0.904、精准度(P)=0.911、召回率(R)=0.919、F1分数(F1_(Score))=0.915、曲线下面积(SAUC)=0.901;研究区滑坡极高、高危险区分别占比11.96%、15.53%,其中新源县和巩留县极高、高危险区占比最高,分别为51.1%、45.6%;滑坡主控因子为高程、坡度、降雨量和峰值地面加速度(PGA),定量揭示高程在1500~2000 m、坡度大于14°、年降雨量在260~310 mm、PGA大于0.23 g的区域对滑坡发生起促进作用,表明该区域滑坡可能为高程和坡度主控的降雨型、地震型滑坡。本研究方法可为滑坡危险性评价提供新的技术参考,为伊犁河流域防灾减灾韧性建设提供理论支撑。

To further improve the accuracy of landslide hazard prediction models and enhance their interpretability,this study selected 8 influencing factors of landslide occurrence,taking the Yili River Basin,Xinjiang province as an example.An interpretable BPNN-SHAP model,based on the back propagation neural network(BPNN)model and the game theory with the aim of addressing the'black box'issue,was constructed.Firstly,the dataset was divided into 70%training set and 30%test set,and 5-fold crossvalidation was used to enhance the robustness of the BPNN-SHAP model.Then,the evaluation accuracy of this model was compared with three other models:Deep Neural Network(DNN),Random Forest(RF),and Logistic Regression(LR).Finally,regional landslide hazard assessment was completed,and the interpretability of BPNN-SHAP was also discussed.The results showed that the BPNN-SHAP model achieved the highest statistical values in the following metrics:Accuracy(A)=0.904,Precision(P)=0.911,Recall(R)=0.919,F1_(Score)=0.915,and SAUC=0.905.The very high and high danger areas for landslides in the study region accounted for 11.96%and 15.53%,respectively.Among these regions,Xinyuan and Nileke County occupy the highest proportions,at approximately 51.1%and 45.6%,respectively.The primary controlling factors for landslides were elevation,slope,rainfall,and peak ground acceleration(PGA).Specifically,areas with an elevation of 1500 m to 2000 m,slopes greater than 14°,annual rainfall between 260 mm and 310 mm,and PGA greater than 0.23 g are prone to landslides,indicating that the predominant types of landslides are rainfallinduced and earthquake-induced.Our research method is expected to provide a new technical reference for landslide hazard assessment and theoretical support for disaster prevention,mitigation,and resilience construction in the Yili River Basin.