Landslide Susceptibility Prediction Based On High Confidence Data Selection

Landslide is the most common geological disaster in China,which causes serious casualties and economic losses every year.Therefore,how to effectively carry out landslide susceptibility prediction has become the focus of landslide direction research at this stage.Landslide susceptibility prediction has the disadvantages of difficulty in expanding landslide samples and low accuracy of subjective random selection of non-landslide samples.In this paper,we take Fuan City,Fujian Province as the study area and construct a model based on Semi-Supervised Learning(SSL)framework to select high-trust data points.The following are the main research results:(1)In landslide susceptibility prediction,if the environmental factors that are not assigned with weights for evaluation are used directly will cause unfavorable results on landslide susceptibility prediction and produce large errors.In this paper,we use the Max-Relevance and Min-Re-dundancy(MRMR)algorithm to assign weights to the landslide evaluation factors to improve the prediction accuracy of landslide susceptibility.(2)Landslide susceptibility prediction was performed using Extreme Learning Machine(ELM)and particle swarm optimization Extreme learning machine(PSO-ELM)models.The area under curve(AUC)of the ELM and PSO-ELM models were 0.710 and 0.788,respectively.It is shown that the particle swarm optimization algorithm improves the accuracy of the ELM model for landslide susceptibility prediction.(3)A Semi supervised particle swarm optimization Extreme learning machine(SS-PSO-ELM)model is proposed,which uses density peaks clustering by fast search and find of density peaks(DPC),Frequency Ratio(FR)and Random Forest(RF)model to select high-trust data to solve the problems of insufficient samples of landslide points and high cost of manual category labeling of non-landslide points.The high confidence data are used as the input of PSO-ELM model for landslide susceptibility prediction.The results show that the area under the curve of the SS-PSO-ELM model is 0.893 and the root-mean-square error(RMSE)is 0.370,which are better than the ELM and PSO-ELM models without semi-supervised framework.It indicates that the SS-PSO-ELM model can better evaluate landslide susceptibility.