戴勇, 孟庆凯, 陈世泷, 等. 2024. 基于可解释神经网络的中巴公路沿线区域工程扰动滑坡危险性评价[J]. 工程地质学报, 32(3): 935-946. doi: 10.13544/j.cnki.jeg.2024-0097.
    引用本文: 戴勇, 孟庆凯, 陈世泷, 等. 2024. 基于可解释神经网络的中巴公路沿线区域工程扰动滑坡危险性评价[J]. 工程地质学报, 32(3): 935-946. doi: 10.13544/j.cnki.jeg.2024-0097.
    Dai Yong, Meng Qingkai, Chen Shilong, et al. 2024. Regional engineering disturbance landslide hazard evaluation along China-Pakistan highway based on interpretable neural network[J]. Journal of Engineering Geology, 32(3): 935-946. doi: 10.13544/j.cnki.jeg.2024-0097.
    Citation: Dai Yong, Meng Qingkai, Chen Shilong, et al. 2024. Regional engineering disturbance landslide hazard evaluation along China-Pakistan highway based on interpretable neural network[J]. Journal of Engineering Geology, 32(3): 935-946. doi: 10.13544/j.cnki.jeg.2024-0097.

    基于可解释神经网络的中巴公路沿线区域工程扰动滑坡危险性评价

    REGIONAL ENGINEERING DISTURBANCE LANDSLIDE HAZARD EVALUATION ALONG CHINA-PAKISTAN HIGHWAY BASED ON INTERPRETABLE NEURAL NETWORK

    • 摘要: 为提高滑坡危险性评价精度、解释工程扰动滑坡风险评估过程,本文以中巴公路沿线区域为例,提出了一种DNN-SHAP可解释神经网络模型。首先选取了距道路距离、坡度等12个危险性评估因子,计算因子间的皮尔逊相关系数,剔除强相关因子,其次构建DNN模型进行滑坡预测,并综合对比随机森林(RF)、支持向量机(SVM)和逻辑回归(LR)模型,最后利用SHAP模型获取DNN预测过程中各因子的影响贡献,完成工程扰动滑坡危险性评价,并解释影响因子间的依赖耦合关系。研究结果表明:本文提出的DNN-SHAP模型预测精度上相比其他3种模型除精准度(Precision)略低于SVM模型以外,其余评价指标均为最高,且该方法可定量揭示道路-岩性、道路-坡度、道路-坡度-地形起伏度等共同作用是该区域工程扰动滑坡灾害的主控因素,为完善滑坡危险性评价方法提供了新的研究思路和技术参考。

       

      Abstract: To enhance the precision of landslide hazard evaluations and elucidate the assessment process of landslide risks influenced by engineering activities, this study introduces a DNN-SHAP interpretable neural network model using the area along the China-Pakistan Highway as a case study. Initially, 12 risk assessment factors such as the distance to the road and slope gradient were selected. The Pearson correlation coefficients among these factors were calculated to exclude highly correlated ones. Subsequently, a DNN(Deep Neural Network) model was developed for landslide prediction and comprehensively compared with Random Forest(RF), Support Vector Machine(SVM), and Logistic Regression(LR)models. Finally, the SHAP(SHapley Additive exPlanations) model was utilized to determine the contributory influence of each factor in the DNN prediction process. This completes the evaluation of landslide hazards due to engineering disturbances and explains the dependent coupling relationships between influencing factors. The research outcomes demonstrate that, except for the Precision metric where it is slightly outperformed by the SVM model, the proposed DNN-SHAP model surpasses the other three models in prediction accuracy. Moreover, the method quantitatively reveals that the synergistic effects of the distance to the road, lithology, and topographic relief are the dominant factors in controlling the engineering-disturbed landslide hazards in this region. This provides new method and technical references for the evaluation of landslide hazard risks.

       

    /

    返回文章
    返回