| The southern part of Sichuan is located in the Himalayan seismic belt,with the strong geological movements,heavy rainfall,complex geological characteristics and complex climatic characteristics,and there are thus frequent landslides.Based on the geological characteristics,hydrological characteristics,meteorological characteristics,surface characteristics and engineering data,the geographical distribution of landslide influencing factors are analyzed.The landslide influencing factors are selected by Chi-square test,out-of-bag(OOB)error and multicollinearity test.Landslide spatial susceptibility(LSS)is evaluated using the deep belief network(DBN),simulated annealing algorithm(SAA)-DBN,particle swarm optimization(PSO)-DBN,sparrow search algorithm(SSA)-DBN and logistic regression(LR)model,respectively.Landslide temporal susceptibility(LTS)is evaluated by analyzing the induced effects of rainfall and earthquakes on landslides using the effective rainfall model(ERM)and peak ground acceleration(PGA).The landslide spatiotemporal susceptibility(LSTS)zoning map is obtained by comprehensively considering the LSS and LTS.The influence of factors on landslide susceptibility is analyzed by multi-temporal factor analysis,secondary-factor interval division analysis and secondary-factor combination analysis.Taking the southern Sichuan as an example,the research contents are as follows:(1)The landslide influencing factors are analyzed and selected,and the index system of LSS assessment are thus established.Based on the relevant literature and field study,10 factors,namely the lithology,elevation,slope,profile curvature,relief degree of the land surface,topography,topographic wetness index,normalized difference vegetation index,average annual rainfall,distance to the railway and distance to a structure line,are selected using chi-square test,OOB error and multicollinearity test from 15 factors.(2)The optimal hyperparameters of DBN model are searched using SAA,PSO and SSA.The zoning maps of LSS are obtained using the SAADBN,PSO-DBN and SSA-DBN models and compared with that obtained using the DBN and commonly used method,namely the LR model.It indicates that the spatial distribution of various susceptibility levels in the four maps is basically the same.The assessment results obtained using various models are compared using receiver operating characteristic(ROC)curve and seed cell area index(SCAI).It illustrates that assessment result obtained using the intelligent optimization model is more accurate,and the SSA-DBN is the optimal model.(3)The ERM and PGA are respectively introduced to indicate the impacts of rainfall and earthquake on landslides,and the LTS is thus evaluated.The LSS and LTS are analyzed comprehensively to obtain the LSTS zoning map.(4)Combining the DEM in 1992,2000 and 2009 with the actual landslide from 2000 to 2009 and from 2010 to 2016,the landslide susceptibility is analyzed in detail using the multi-temporal factor,namely the elevation,slope,profile curvature and relief degree of the land surface.It is verified that the change of DEM has a significant impact on LSS.Landslide susceptibility zoning obtained using the recent DEM is more accurate.The landslide susceptibility level of the area after the landslides will decrease,and the geo-hazards of the area will be mitigated.(5)The impact of the secondary-factor interval division on the landslide susceptibility assessment is analyzed finely,and the scientific methods of secondary-factor interval division are studied.The optimal interval division is studied using the genetic algorithm based on the entropy theory which takes the historical landslide density in the various secondary-factor intervals as a parameter.The method can significantly improve the accuracy of assessment compared with the natural breakpoint method and equidistant statistical method.(6)An efficient method for mining the most unfavorable secondaryfactor combination is studied based on intelligent big data mining technology.The Apriori algorithm,frequent pattern tree algorithm and Eclat algorithm are optimized using the historical landslide area ratio as a characteristic parameter.And the most unfavorable secondary-factor combinations prone to landslides are thus mined efficiently using the optimization algorithm.The correlations between the combinations and landslides are verified using the chi-square value and the frequency ratio which are calculated using the historical landslide area of various combinations. |
