Approach To Landslide Susceptibility Prediction Based On Graph Convolutional And Long Short-Term Memory Neural Network

Landslide is a common geological hazard that causes a significant amount of infrastructure damage and casualties every year.Efficient and accurate landslide prediction is a key focus of researchers.Landslide susceptibility maps can identify potential unstable landslide areas and are one of the main visualization tools for the location and severity of landslides.Currently,landslide susceptibility prediction mainly relies on traditional machine learning methods,which make it difficult to extract spatial relationships and feature similarities between landslide factors,resulting in poor robustness and predictive performance of the model.Deep neural networks have powerful feature extraction capabilities and have been widely applied to data prediction tasks.In this paper,three prediction models based on graph convolutional neural networks and long short-term memory networks are proposed for the assessment of landslide susceptibility in actual regions.At the same time,the causes of landslide development are analyzed by interpretable method.The main research contents of this paper are as follows:(1)A landslide susceptibility prediction model SGCN-LSTM(Self-screening Graph Convolutional Network and Long Short-Term Memory)is proposed.Firstly,a pre-training network learning training set containing only one graph convolution and one fully connected layer is constructed,and the error samples caused by manual labeling or acquisition errors are eliminated by setting the threshold.The pre-training model is used to predict and evaluate the landslide susceptibility of the data in the whole study area,and the reliable data is selected to supplement the data,which improves the quality of the data.Then,the spatial information between the landslide grid points is extracted by the neighbor nodes of the cascaded GCN network,and the feature information of the long sequence samples is extracted by the cascaded LSTM network.The two types of information are fused to obtain landslide fusion features.Finally,the fusion feature is used as the input fully connected layer,and the binary classification prediction of sample point landslide-non-landslide is realized by Softmax function.The SGCN-LSTM landslide susceptibility prediction model was applied to the data of Anyuan County,Ganzhou City,Jiangxi Province,and compared with machine learning models such as CPLSTM-CRF depth model,RF(Random Forest),LR(Logistic Regression),SVM(Support Vector Machine),SGD(Stochastic Gradient Descent).The experimental results show that the prediction accuracy and AUC value of the SGCNLSTM model are the highest in the six models.Among them,the prediction accuracy is 92.38%,which is 5.88%,12.44%,19.65%,19.92% and 20.34% higher than the five comparison algorithm models respectively.The AUC value reached 0.9782,which was0.0305,0.0532,0.1875,0.1909 and 0.1829 higher than the other five models.It can be seen that the SGCN-LSTM prediction model proposed in this paper overcomes the problems existing in the machine learning model to a certain extent,can effectively weaken the interference of error samples,and has stronger feature extraction ability.(2)A landslide susceptibility prediction model SGCN-Transformer(Selfscreening Graph Convolutional Network and Transformer)based on graph convolution and Transformer deep neural network is proposed.Firstly,the environmental factors are classified according to continuous and discrete factors,and a parallel GCNTransformer feature extraction network is constructed.The continuous factor is used as the input of the GCN network,and the discrete classification feature factor is encoded by Transformer.The vectorization of the coding result is spliced with the output of the GCN as the input of the multi-layer perceptron.The SGCN-Transformer susceptibility prediction model was applied to the data of Poyang Lake research area in Jiangxi Province,and compared with SGCN-LSTM and RF,SVM,LR,SGD machine learning models.The experimental results show that the SGCN-Transformer model has the best PPR,NPR,total accuracy and AUC,which are 83.19 %,93.88 %,87.88 and 0.9263,respectively.It was higher than SGCN-LSTM,RF,SVM,LR and SGD models(0.25%,1.39%,0.76% and 0.0206),(7.13%,6.11%,6.96% and 0.0564),(11.33%,14.08%,12.6% and 0.1008),(10.93%,17.02%,13.49% and 0.1261)and(13.23%,19.82%,16.03% and 0.1726),respectively.It can be seen that the SGCNTransformer model proposed in this paper can process different types of data in parallel according to the characteristics of continuous and discrete data of landslide factors,so as to improve the prediction performance of landslide susceptibility.(3)A landslide susceptibility prediction model SGAT-LSTM(Self-screening Graph Attention and Long Short-Term Memory)based on graph attention and long short-term memory network is proposed,and the causes of landslide development are explained.Firstly,a parallel network of GAT and LSTM is constructed to obtain the spatial feature information of the factor.Then the SGAT-LSTM model is applied to the landslide susceptibility prediction in Anyuan County,Jiangxi Province.The prediction distribution map and integral gradient are used to explain the causes of landslide development from two aspects of local interpretability and global interpretability.The experimental results show that the PPR,NPR,total accuracy and AUC of SGAT-LSTM model are 92.61%,91.23%,91.90% and 0.9743,respectively.Only the PPR index is lower than RF,and other indicators are significantly higher than RF,SVM and LR models.In particular,the PPR index is improved on SGCN-LSTM and SGCNTransformer models.In terms of interpretability,landslide influencing factors such as elevation,plane curvature,distance from river and fault density play a major role in model decision-making.When the elevation frequency ratio is 1.1～2.06,the distance frequency ratio from the river is 1.56～2.3,the NDVI frequency ratio is 1.04～1.21,and the plane curvature frequency ratio is 1.33～1.53,it is conducive to landslide development.In summary,three landslide susceptibility prediction models based on graph convolution,long short-term memory network and attention mechanism are proposed in this paper,and applied to the landslide data of Anyuan County and Poyang Lake area,and the local interpretability and global interpretability analysis are carried out.The experimental results show that the performance of three landslide susceptibility prediction models based on deep learning network proposed in this paper is significantly improved compared with many machine learning models,and has great application value in the field of landslide prediction.