Research And Application Of Landslide Susceptibility Prediction Based On Long Short-term Memory Deep Neural Network

Landslide prediction and prevention is a worldwide problem.In the past decade,with the rapid rise of artificial intelligence,machine learning has been widely used in landslide susceptibility prediction.However,the landslide susceptibility prediction model based on machine learning has certain limitations because the characteristics of landslides are usually irrelevant or nonlinearly related,as well as the weak generalization ability and insufficient feature extraction of traditional machine learning models.Deep learning has been widely used in many fields such as smart healthcare,smart city,smart transportation,and network security due to the strong learning ability,wide coverage,and good adaptability.In this paper,two landslide susceptibility prediction models based on the Long Short-Term Memory(LSTM)network has been proposed and applied to the actual collected landslide spatial dataset.The interpretability analysis of landslide environmental factors and prediction model is also carried out.The main research contents are as follows:(1)A landslide susceptibility prediction model cp LSTM-CRF(cascade-parallel Long Short-Term Memory and Conditional Random Fields,cp LSTM-CRF)based on cascade-parallel Long Short-Term Memory and Conditional Random Fields is proposed.Firstly,a cascaded parallel LSTM feature extraction network is constructed by using a cascade structure,so that the network can fully extract the nonlinear features among the landslide environmental factors.Secondly,constructed a CRF to optimize the extracted features and smooth the mutation prediction results to optimize the prediction results.Finally,the cp LSTM-CRF is applied to the landslide susceptibility evaluation in Shicheng County,Ganzhou City,Jiangxi Province,and compared the results with Logistic Regression(LR),Decision Tree(DT)and Multi-Layer Perceptron(MLP).The experimental results show that the Negative Predictive Rate(NPR),landslide prediction accuracy and Area Under Curve(AUC)of cp LSTM-CRF are 80%,75.67% and 0.868,respectively,which are better than C5.0DT(69.73%,75.67%,0.838),LR(70.83%,70.94%,0.833)and MLP(71.64%,71.61%,0.826).Therefore,the cp LSTM-CRF prediction model proposed in this paper effectively overcomes the limitations of traditional machine learning,extracts nonlinear features more fully,and improves the prediction performance of landslide susceptibility.(2)A landslide susceptibility prediction model SBi LSTM-CRF based on selfscreening bidirectional long-short-term memory and conditional random fields is proposed.Firstly,use Bi-LSTM and full connection to build a pre-classification selfscreening network to screen the data with wrong labeling results due to uncertain factors.At the same time,pre-evaluation of the whole study area is carried out to reduce the uncertainty in the process of susceptibility evaluation.Secondly,reliable data were selected from the pre-assessment results of the whole study area for equal supplementation.Next,a cascaded Bi-LSTM feature extraction network is constructed to extract spatial information more fully in all directions.Then,use CRF to obtain smoothed landslide probability output between grids.Finally,SBi LSTM-CRF is applied to landslide susceptibility evaluation in Yanchang County,Yan’an City,Shanxi Province,and is combined with cp LSTM-CRF and Random Forest(RF),Support Vector Machine(SVM),LR,Stochastic Gradient Descent(Stochastic Gradient Descent,SGD)and other machine learning models for comparison.The experimental results show that the PPR,NPR,landslide prediction accuracy and AUC of SBi LSTM-CRF are 83.07%,82.56%,82.82% and 0.9202,respectively,which are better than cp LSTMCRF(71.09%,76.03%,73.30%,0.7773),RF(75.84%,78.04%,76.89%,0.8478),SVM(71.82%,73.49%,72.62%,0.7800),LR(70.43%,70.63%,70.53%,0.7682),SGD(66.26%,73.95%,69.37%,0.7646).Therefore,the SBi LSTM-CRF proposed in this paper effectively reduces the uncertainty in the process of landslide susceptibility modeling,effectively extracts features from the three directions of forward grid correlation,reverse grid correlation and adjacent grid energy,significantly improves the prediction performance and has better landslide susceptibility evaluation performance.(3)Research on the interpretability of landslide susceptibility environmental factors.In view of the disadvantage of weak interpretability of machine learning algorithm,taking the study of Yanchang County as an example,the prediction distribution map and the integral gradient deeply explain the natural principles of landslides in the study area and the contribution of environmental factors to the neural network.Firstly,the relationship between the predicted results of SBi LSTM-CRF algorithm and the coupling of environmental one-factor and two-factor is studied,and the possible natural principle of landslide is expounded.Then,the sum of the contribution of each factor to the decision-making of SBi LSTM-CRF is calculated by using the integral gradient.The experimental results show that the landslide factors such as slope,elevation,lithology,surface relief and slope aspect control the development of accumulation layer landslides in Yanchang County.When the altitude is 866.6 m ～ 979.4 m,the slope is 21.9° ～ 41°,the terrain relief is 33.1 m ～ 84.6 m,the lithology is t3 y,and the terrain humidity is 0 ～ 0.123,it is easy to cause landslide development.The study provides a basis for further elaborating the causes of landslide.In summary,based on the deep neural network of long and short-term memory,two landslide susceptibility prediction models cp LSTM-CRF and SBi LSTM-CRF are proposed,which are applied to the geological data of Shicheng County,Ganzhou City,Jiangxi Province,and Yanchang County,Yan’an City,Shaanxi Province,respectively,and obtain high-performance landslide susceptibility prediction results.The two models effectively solve the problems of highly nonlinear landslide environmental factors,limitations of traditional machine learning,and uncertainty of landslide susceptibility modeling process.Taking Yanchang County as an example,the relationship between environmental factors and the natural causes of landslides is analyzed by an interpretable analysis.In short,the two prediction models proposed in this paper and the interpretability analysis method have certain innovation and versatility for landslide susceptibility prediction.