Study On The Technology On The Occurrence Probability Of Coseismic Landslides

China is the country with the most serious geological disasters and the most threatened population in the world,because of its vast territory,complex physical and geographical conditions,and strong neotectonic movement and easy to induce all kinds of geological disasters(including landslides,collapses,debris flows,ground collapses,ground fissures and other disasters).The situation of geological disaster risk prevention,emergency disposal and comprehensive disaster reduction is very serious.The evaluation of the earthquake geological disaster vulnerability can provide scientific decision basis for the earthquake geological disaster prevention,earthquake emergency preparation and post disaster emergency rescue command.The earthquake landslide susceptibility mapping is one of the important parts of earthquake landslide loss prediction.The earthquake landslide susceptibility mapping is to analyze the possibility of landslide in the study area according to the terrain,geology and environmental conditions of the study area.At present,there are many results of earthquake landslide susceptibility mapping,which are involved from small region to global scale,and the absolute probability of earthquake landslide is used.In recent years,the research mainly improves the modeling model used in seismic landslide susceptibility analysis,and the main evaluation models include physical model and machine learning model.In recent decades,data-driven models such as machine learning algorithms have received extensive attention.And it has been proved that it is superior to the deterministic method in large area evaluation.However,in the current risk assessment of earthquake landslide disaster based on machine learning,it is not appropriate to use absolute probability value to replace the probability of earthquake landslide.All the evaluation results are not guided by the idea of probability,only consider the relative risk of landslide,and do not link the risk with the actual landslide,resulting in regional restrictions or different classification standards.That is to say,the high landslide risk in different areas may actually vary greatly,which often leads to inaccurate estimation of the loss of landslide disaster.The application of earthquake landslide absolute probability can not meet the requirements of a series of earthquake geological comprehensive disaster reduction work such as earthquake landslide loss estimation and rapid post disaster loss assessment.It is necessary to study how to accurately give the earthquake landslide occurrence probability.Secondly,due to the lack of limited coseismic landslide database,there are fewresearches on real data-driven landslide evaluation and prediction based on real and complete landslide database.In the emergency response of earthquake geological landslide,it is impossible to give the results of the quick assessment of the vulnerability of the near real-time rapid landslide in a region at the first time.It is very urgent to study and apply the near real-time evaluation model of earthquake landslides.In this paper,based on the real and complete earthquake co seismic landslide database,through the introduction of naive Bayesian sampling method and the combination of landslide susceptibility evaluation based on machine learning,the calculation method of earthquake landslide occurrence probability is established.Combined with independent repeated experiments to eliminate the uncertainty and randomness of the model,the quantitative effects of different fault models,sampling methods and data resolution on the model results are discussed.At the same time,combining the model with different ground motion peak acceleration conditions,the prediction results of landslide probability under different ground motion conditions are directly given,which provides technical support for earthquake landslide risk assessment and loss prediction.The research content mainly involves the following four aspects(1)Exploring the effects of the sample size and proportion on the actual probability estimator of logistic regression-based landslideDue to the influence of sampling strategy,the resulting probability often deviate considerably from the actual areal percentage of the co-seismic landslide.In order to achieve a more reasonable prediction result,it is necessary to revise the forecasting results compulsively,but this method has obvious limitations.In this study,we selected the Lushan earthquake as the study area.Based on the Bayesian theory,we proposed a sampling method that selects the sliding samples and non-sliding samples based on the ratio of stable area to the landslide area.Using this sampling method,we explored 15kinds of sampling density(1,10,20,30,40,50,60,70,80,90,100,200,500,1000,2000 samples/km²)and 12 kinds of non-sliding/sliding sample proportion(1,5,10,25,50,75,100,125,150,175,200,228,respectively).Ten factors including elevation,slope gradient,aspect,topographic wetness index(TWI),peak ground acceleration(PGA),distances to epicenter,distances to rivers,distances to roads,lithology and annual precipitation are considered as the influencing factors.Based on Logistic regression(LR)model,we run our models based on the 15 kinds of sampling density and each performing 200 times.And we chose 12 kinds of sample proportion,150 times model training were performed respectively.Finally we constructed 4800 resulting models.The results show that the different sampling density has little effect on the total predicted landslide area,the higher density of the samples,the more stable the predicted results.Especially when the sampling density reach 1000 samples/km²,the total area of the predictive model is about 17.1 km²,which is close to the real area with 17.16km²,the difference is controlled within 2%.But the different proportion of non-sliding/landslide samples has a great influence on the occurrence probability of co-seismic landslides.When the ratio of non-sliding samples to sliding samples is 1:1,the predicted landslide area(A_p)is between 1265 and 1290 km²,with an average of 1280km²,which is 75 times to the actual landslide area.The functional relationship between the ratio of predicted area to real area(R_pa)and the ratio of non-sliding samples to non-sliding samples(R_ns)is finally obtained,which is=99.156×..This study indicates that the non-sliding/sliding sample proportion determined by the ratio of stable areas to the landslide area is the basis of constructing the resulting models that predict the landslide areal percentages.This resulting probability is very consistent with the actual areal percentage of the co-seismic landslides.(2)Effects of seismogenic fault on the predictive mapping of probability to earthquake-triggered landslidesThe seismogenic fault is crucial for the spatial prediction of co-seismic landslides.On the one hand,earthquake-induced landslides are densely distributed along the seismogenic fault;On the other hand,different sections of the seismogenic fault often have distinct landslide triggering capabilities due to their different characteristics mechanical property.While how the features of a fault influence landslide occurrence probability mapping remain unclear.Relying on the landslide data of the 2013 Lushan,China Mw 6.6 earthquake,this study attempts to further address this issue.This study quantifies the seismogenic fault effects of landslides into three modes:the distance effect,the different part effects,and the combined effects of the two.Four possible cases are taken into consideration:with zoning the study area vertical to fault(case2),with zoning the study area parallel to the fault(case3),with the combined both(case1),and without such study-area zonations(case4),combined with LR model,yielding predictive landslide probability maps.In addition,the model also fully considers other influencing factors of earthquake landslides,such as elevation,slope,aspect,TWI,PGA,lithology,rainfall,distance from the epicenter,distance from the road,and distance from the river.Then cross comparisons and validations are conducted to these maps.Results show that the success rates of earthquake-triggered landslides for former three scenarios are 85.1%,84.2%,and 84.7%,respectively,while that of the model without considering the effect of relative positions to the seismogenic fault effect only84%.And the prediction rates of the four LR are 84.45%,83.46%,84.22%,and 83.61%,respectively,shown by comparing the test dataset and the landslide probability map.It means that the effects of the seismogenic fault,that are represented by study-area zonations vertical and parallel to the fault proper,are significant to the predictive mapping of earthquake-induced landslides.(3)Effects of raster resolution on the occurrence probability predictionlity of co-seismic landslidesRaster resolution has been playing a significant role in the prediction of landslideoccurrence probability,so choosing the appropriate resolution is a still worth considering issue.This study used the landslides data of the 2013 Minxian earthquake to further address this issue.Based on the Bayesian theory,according to a sampling method that selects the sliding samples and non-sliding samples based on the ratio of stable area to the landslide area,we studied the real probability of co-seismic landslides at different resolution datasets(2.5m,5m,10m,20m,40m,80m)for the analysis of the logistic regression(LR)model.Eight factors were considered in this analysis,including elevation,slope angle,profile curvature,topographic wetness index(TWI),distance to fault,distance to epicenter,distance to roads,and lithology.In terms of 6 kinds of resolution datasets(2.5m,5m,10m,20m,40m,80m),the samples was trained 20 times,respectively using the LR model,yielding 120 predicted results.The landslide predicted area(A_p),maximum predicted probability(P_max)and AUC/AIC of the different resolutions were compared.The results show that the total predicted landslide area at the different resolutions is roughly same,and the best resolution is medium resolution(10m,20m).It means that when the grid size is closest to the average area of the landslide,the LR model has the best prediction accuracy.Meanwhile,we used cross-scaling technique to discuss the applicability of above six resulting models on different resolution datasets.Finally,36 probability results were obtained.The results show that the satisfactory result could be achieved with high resolution DEMs while the accuracy of landslide prediction becomes increasingly worse with coarser resolutions.(4)Calculation of landslide occurrence probability in Taiwan area under different ground motion conditionsIn this study,Bayesian probability method and machine learning model are used tostudy the real occurrence probability of earthquake induced landslide risk in Taiwan region.The analyses were based on the 1999 Taiwan Chi-Chi Earthquake,the largest earthquake in the history of Taiwan in a hundred years,which provides better control on the prediction accuracy of the model.This seismic event has detailed and complete seismic landslide inventories identified by polygons,including 9272 seismic landslide records.Taking into account the real earthquake landslide occurrence area,the difference in landslide area and the non-sliding/sliding sample ratios and other factors,a total of 13,656,000 model training samples were selected.We also considered other seismic landslide influencing factors,including elevation,slope,aspect,TWI,lithology,distance to fault,PGA,and rainfall.Bayesian probability method and machine learning model were combined to establish the multi-factor influence of earthquake landslide occurrence model.The model is then applied to the whole study area of Taiwan using different ground motion peak accelerations(from 0.1 g to 1.0 g with 0.1 g intervals)as a triggering factor to complete the real probability of earthquake landslide map in Taiwan under different peak ground accelerations,and the functional relationship between different PGA and their predicted area is obtained.