Spatiotemporal assessment of landslide susceptibility in Southern Sichuan, China using SA-DBN, PSO-DBN and SSA-DBN models compared with DBN model

Abstract

Landslide is one of the most serious geo-hazards, and the landslide susceptibility assessment (LSA) is an existing effective method to efficiently mitigate the loss caused by landslides. This study develops the improved deep belief network (DBN) models by selecting the optimal model hyperparameter to improve the accuracy of LSA. Evaluation factors of the LSA are selected from fifteen influencing factors by chi-square test, multicollinearity test and out-of-bag error. 30% data of the study area are selected randomly as the training data to assess the landslide susceptibility of each grid in the study area. The spatial LSA is then obtained by integrating the DBN models with three optimization algorithms, namely the simulated annealing (SA), particle swarm optimization (PSO) and sparrow search algorithm (SSA). The assessment results obtained using DBN and improved DBN models are thus compared and verified using the receiver operating characteristic (ROC) curve and seed cell area index. It shows that the three improved DBN models outperform the DBN model, which demonstrates the ability of optimization algorithms to improve model performance, and the SSA-DBN model achieves the highest assessment accuracy, followed by the PSO-DBN and SA-DBN models. Meanwhile, the effective rainfall model and peak ground acceleration are respectively employed to evaluate the impact of two inducing factors, namely the rainfall and earthquake, and the temporal LSA is thus obtained. The spatiotemporal LSA map is then generated by coupling the optimal spatial LSA map and temporal LSA map. Therefore, the present study further explores the proposed improved methods and offers instructions for spatiotemporal LSA.