基于信息量优化策略和CNN-LSTM模型的地质灾害易发性制图：以西盟县为例

Mapping Geological Hazard Susceptibility Using an Information Content Optimization Strategy and a CNN-LSTM Model: A Case Study of Ximeng County

摘要: 云南省西盟佤族自治县位于“中印缅孟经济走廊”战略枢纽，开展精细化地质灾害易发性评价研究，对于保障该区域重大基础设施建设安全、优化国土空间开发格局具有重要意义。针对一般评价流程中负样本选取策略对评价模型精度的影响以及传统深度学习方法中的数据依赖，分类结果的不可解释性等问题，利用信息量法(Information Value, IV)结合卷积神经网络(CNN)进行静态空间特征深度提取以及长短期记忆网络(LSTM)捕捉降雨时序动态，创新性地提出了一种IV-CNN-LSTM耦合模型架构。通过与CNN-LSTM模型、RF模型和XGBoost模型的预测结果进行对比评估表明：信息量模型的应用有效提高了CNN-LSTM模型的预测效果，IV-CNN-LSTM模型的ROC曲线下面积(area under curve, AUC)值为0.9845，具有更好的预测效果，同时其他四项指标也均优于单一模型。灾害点的94.51%落入中易发等级以上区域，71.95%落入高易发等级以上区域，易发等级为高以上的区域主要沿公路网分布，结合SHAP分析方法，揭示了人类工程活动是该区域地质灾害的主要诱因。本文研究结果可为滇西南边境地区灾害链式防控提供了科学决策依据和有效参考。

Abstract: Ximeng Wa Autonomous County in Yunnan Province is in the “China-India-Myanmar-Bangladesh Economic Corridor” strategic hub. It is of great importance to conduct a geological hazard vulnerability assessment study to ensure the safety of the major infrastructure construction in the region and to optimise the spatial development pattern of the national territory. Considering the influence of the negative sampling strategy on the accuracy of the evaluation model in the general evaluation process, as well as the data dependency and non-interpretability of the classification results in the traditional deep learning method, an innovative approach is proposed by using the Information Value (IV) method in combination with the Convolutional Neural Network (CNN) for the deep extraction of the static spatial features and the Long Short-Term Memory (LSTM) for the acquisition of the rainfall temporal sequences. Dynamics, an innovative coupled IV-CNN-LSTM model architecture, is proposed. The evaluation results are compared and evaluated with the prediction results of the CNN-LSTM model, RF model and XGBoost model. The research results show that applying the informative model effectively improves the prediction effect of the CNN-LSTM model. The IV-CNN-LSTM model, with a value of 0.9845 for the area under the curve (AUC) of the ROC curve, has a better prediction effect, while the other four indicators are also better than the single model. 94.51% of the disaster sites fall into the area above the medium susceptibility level, 71.95% fall into the area above the high susceptibility level, and the areas with high susceptibility levels are mainly distributed along the highway network, which combined with the SHAP analysis method proves that the main causative factor of geohazards in this region is human engineering activities. The results of this paper can provide a scientific basis for decision-making and a practical reference for preventing and controlling the disaster cascade in the border area of southwest Yunnan.