Application Of Kernel Learning Methods On Regional Landslide Susceptibility Assessment

Landslide disasters not only seriously destroy the natural environment,but also threat the safety of people's life and property.The hilly area of Sichuan province is one of the most precipitation-induced landslide dense areas in China.This area has dense population and developed economy.Therefore,it is of great research value and practical significance to study the regional landslide susceptibility assessment method of this area and provide scientific basis for local disaster prevention and mitigation work.With the rapid development of computer,GIS and artificial intelligence technologies,the kernel function method has gain rapid development.Particularly,the support vector machine(SVM)is widely used in the field of landslide susceptibility assessment,and its superior performance has been recognized by the majority of researchers.Relevance vector machine(RVM)is a new Bayesian probability model based on SVM.Both the classification results and the probability distribution can be obtained by RVM,which are quit suitable for landslide susceptibility assessment.Therefore,based on the theory of multi-kernel function method,a new evaluation system of regional landslide susceptibility based on RVM is proposed.In order to improve the performance of RVM,two improved algorithms are provided based on the multi-kernel learning theory,including the multi-kernel relevance vector machine(MKRVM)model based on cloud particle swarm optimization(CPSO-MKRVM)and the MKRVM model based on second-order cone programming(SOCP)algorithm(SOCP-MKRVM).Finally,two new evaluation systems of regional landslide susceptibility can be given based on these two models.The main work and innovations are as follows:(1)Study on the landslide data acquisition and feature selection.According to the special geological environment of the hilly area of Sichuan province,the occurrence mechanism of landslide in the area is analyzed.And then,eight landslide-predisposing factors(elevation,slope,lithology,relief amplitude,and et al.)are adopted in terms of the landslide characteristics in this area.And they are graded and quantified by different steps.Based on this,combined with ArcGIS technology,the training set and the regional landslide set are built to obtain the complete data set of landslide susceptibility assessment.(2)Study on the regional landslide susceptibility evaluation system based on common kernel function method.SVM has been widely used in the regional landslide susceptibility assessment,while RVM has been seldomly used in this field.Based on the basic theory,principle and performance optimization of the kernel function method,combined with the technique of SVM in landslide susceptibility assessment,a new regional landslide susceptibility evaluation system based on RVM is proposed.After that,the evaluation system is used to map landslide susceptibility of Sichuan hilly area.And the rationality and superiority of this new evaluation system are verified by using receiver operating characteristic(ROC)curve and landslide dot density.(3)Study on the MKRVM based on CPSO algorithm and its application.In order to solve the selection problem of kernel and its parameters,a linear adaptive MKRVM method is proposed based on the multi-kernel learning theory.By introducing the kernel weight coefficients,the kernel selection problem can be transformed into the multi-kernel parameters optimization problem.In this way,based on the cloud model and PSO algorithm,the cloud PSO(CPSO)optimization algorithm is proposed to optimize the kernel parameters and weight coefficients of MKRVM.Based on the CPSO-MKRVM model,a new regional landslide susceptibility evaluation system is built.The prediction results on UCI data sets show that the proposed method can both accelerate the convergence speed and improve the prediction accuracy.The rationality and credibility of the evaluation system are verified by the practical application results which meet the actual landslide distribution law and have high reference value.(4)Study on the MKRVM method based on SOCP and its application.With the increase of particles size or population size in CPSO algorithm,the executive efficiency is reduced greatly.Therefore,a new MKRVM method based on SOCP is proposed.First of all,for a specific sample space,CPSO-RVM is used to select the optimal basic kernel functions of multi-kernel architecture.Then the optimum setting problem of kernel weight coefficients is transformed into SOCP form for calculating.And the calculated results are brought into the multi-kernel structure to build the MKRVM model(SOCP-MKRVM).The performance of SOCP-MKRVM is evaluated by UCI data sets and the practical application.For one thing,the results of UCI show that the prediction accuracy can be improved and lower time and fewer resources are requested with the SOCP-MKRVM model,compared with the CPSO-MKRVM model.For another thing,the SOCP-MKRVM method is adopted to improve the regional landslide susceptibility evaluation system.The practical application results show that this evaluation system has higher execution efficiency than CPSO-MKRVM method,which is more suitable for real-time landslide susceptibility assessment.