AUTHOR=Wang Siying , Lin Xiaokun , Qi Xing , Li Hongde , Yang Jingjing 

TITLE=Landslide susceptibility analysis based on a PSO-DBN prediction model in an earthquake-stricken area

JOURNAL=Frontiers in Environmental Science

VOLUME=Volume 10 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2022.912523

DOI=10.3389/fenvs.2022.912523

ISSN=2296-665X

ABSTRACT=Landslide hazards, one of the main geological hazards, have had a great impact on the normal life of human beings. Deep belief networks (DBN) hyperparameter determination problem is the key to its application. To improve the accuracy of regional landslide susceptibility prediction, this paper introduces the particle swarm algorithm (PSO) to determine the hyperparameters of the DBN, and applies it to regional landslide susceptibility prediction. Firstly, PSO is used to optimize the hyperparameters of the DBN, and obtain a set of hyperparameters with the optimal fitness function. Secondly, a landslide susceptibility prediction model based on PSO-DBN is constructed, and the K-fold cross-validation method is used to determine the accuracy of the model. Then, the model is applied to landslide susceptibility prediction in the most impacted area of the Wenchuan earthquake to analyze the model’s accuracy. Finally, a model susceptibility analysis is performed. The research results show that the final optimal model accuracy of the PSO-DBN model is 95.52%, which is approximately 28.31% and 15.35% higher than that of the logistic regression (LR) model and the common DBN model, respectively. The Kappa coefficient is 0.883, which is higher than that of the LR model. Compared with the LR model and the common DBN model, it is improved by approximately 0.542 and 0.269 respectively, and the area under the curve (AUC) is 0.951, which is improved by approximately 0.201 and 0.080 compared to the LR model and the common DBN model. The susceptibility of the model to the inertia factor is low, and the average change in model accuracy (when the inertia factor changes by 0.1) is approximately 0.1%, and the overall stability of the model is high. The landslide susceptibility level is very high. The area includes 219 landslide points, which account for 39.2% of the total landslide points. In the area with the high landslide susceptibility level, there are 191 landslide points, accounting for 34.2% of the total landslide points. Together, the two contain approximately 73.4% of the landslide points. This indicates that the model prediction results are in good agreement with the spatial distribution characteristics of the landslide.