滑坡、泥石流等地质灾害给人民生产生活和社会经济发展带来了严重影响，因此急需对地质灾害进行易发性预测，以防患于未然，减少灾害带来的损失。目前的易发性预测方法主要采用基于统计或机器学习的算法。这些方法根据历史上的危险点位置和地理环境测绘数据，并依据统计规律来确定容易发生地质灾害的区域；对于监督学习类方法，需要在危险点信息的基础上标出安全点，但是安全点的选取质量难以控制，导致易发性预测的准确率较低。针对上述问题，本文依托“贵州省地质灾害防治指挥平台”项目，开展了基于深度学习的地质灾害易发性预测算法研究。论文的主要工作与成果如下： （1）对贵州省地理数据进行了数字化建模，并设计了安全点样本迭代方法，运用随机森林、支持向量机等机器学习方法对地质灾害易发性预测进行了分析。根据模型易发性预测结果来调整安全点分布，减少安全点出现在高危地区的频率，从而迭代产生安全点样本。采用实际地理数据进行了仿真实验，结果表明整体预测效果得到了提升。（2）将地理数据转化为三通道图片数据，并建立了基于卷积神经网络（CNN）的地灾易发性预测模型。将多维地理数据转化为“高度+坡度+第三特征”的三通道数据，保留了高度和坡度的主要地理特征，用“第三特征”整合了其他稀疏特征，从而精简了CNN模型结构。在实际地理数据上的仿真结果表明，基于三通道数据的CNN易发性预测模型相比随机森林等方法具有更高的准确率。（3）利用CNN提取的深度特征，建立了基于CNN与图神经网络（GNN）相结合的地灾易发性预测模型。在图结构建模中，把每个图片视为节点，将CNN所提取的深度特征作为节点的特征，通过连接相似节点来构建GNN模型。GNN通过和相连接的点进行信息融合，增强了模型对每个点所连接区域的整体感知能力。在实际地理数据上的仿真实验结果表明，该模型增强了预测危险区域的效果。本文对地质数据的样本处理、深度特征提取、图结构建模、图网络计算等方面进行了研究和探索，建立了“CNN+GNN”的地灾易发性预测模型，对贵州省地质灾害危险点的预测准确率较高，相比随机森林、SVM等方法取得了显著提升。

Landslides, mudslides and other geohazards have caused serious impact on people‘s production, life, and social economic and development. Therefore, it’s crucial to study the prediction of the geohazards suscepbility, which can prevent them earlier and reduce the losses caused by these disasters. In the present, the prediction methods of geohazards suscepbility are mainly based on statistics or machine learning. These methods predict geohazards suscepbility by statistical law according to the geographical location of historical dangerous points and environment surveying and mapping data. For supervised machine learning methods, geological safety points need to be marked and selected based on information of dangerous points, but it is always difficult to control the sample data quality of the safety points, which causes low prediction accuracy of geohazards suscepbility. In order to solve these problems, research efforts in this paper have focused on geohazards suscepbility prediction algorithm based on deep learning methods, which is relying on the project "Geohazards prevention and control platform of Guizhou Province". The main work and achievements of this paper are as follows: (1) Digital models of geographic data in Guizhou Province are built, an iterative method for sampling safety point is designed, and then prediction models of geohazards suscepbility are built by using random forest, support vector machine and other machine learning methods. According to the prediction results of original model, the distributions of safety points is adjusted to reduce the frequency of safety points near high-risk areas, thus iteratively generating safety point samples. Simulation results on the actual geographic data show that the overall prediction effects of the machine learning methods have been improved by new safety point samples.(2) By transforming geographic data into three-channel picture data, a new prediction model of geohazards suscepbility is proposed based on convolution neural network (CNN). The multi-dimensional geographic data is transformed into the three-channel data of "height, slope, third feature", in which the main geographical features of height and slope are retained and other sparse features are integrated as "third feature", simplifying the model structure. Simulation results on real geographic data show that the CNN prediction model based on the three-channel data significantly improves the prediction accuracy compared with methods such as random forests.(3) Based on the deep features extracted by CNN, a mixed prediction model combining CNN and graph neural network (GNN) is established. During the graph structure modeling, each picture is regarded as a node of GNN, the deep features of CNN are taken as the features of the nodes, and the GNN model is constructed by connecting those similar nodes. Through each node interacted with connected nodes, GNN enhances the overall perception ability of the prediction model. The simulation results on the actual geographic data show that the model enhances the effectiveness of predicting dangerous areas. This paper has studied and explored a series of methods of predicting geohazards suscepbility, such as the data sample processing, deep feature extraction, map structure, map network calculation, and so on. The prediction model of geohazards suscepbility based on "CNN+GNN" is established, the risk points of geological disasters in Guizhou Province is predicted quite well by the model, and the prediction accuracy of the model is higher than random forests and SVM.