Study On The Geo-Hazards Susceptibility Assessment Based On A Novel Ensemble Learning Framework

Geo-hazards seriously threaten the safety of life and property,resulting in irretrievable damages and huge economic losses worldwide.With the development of human society and the increase of human engineering activities,geo-hazards have occurred frequently in recent years,and there is still an aggravating trend.Lushui County is a subordinate Autonomous County of Nujiang Lisu Autonomous Prefecture in the northwest of Yunnan Province.Due Its complex and changeable natural geography,meteorology,hydrology and geological environment,Lushui area is prone to various disasters.The susceptibility mapping of geo-hazards is considered to be an effective tool for predicting and preventing the occurrence of disasters,which is helpful for local governments to take timely and reliable early warning and decision-making.Therefore,the development of accurate and efficient susceptibility assessment model has become a very critical task.With the improvement of computer technology,machine learning has been widely used in geo-hazards susceptibility assessment.However,the performance of machine learning is often limited by its own algorithm property,so it is necessary to explore another evaluation model with high generalization capability.Ensemble learning is capable to reinforce the prediction ability of machine learning,which has been successfully applied in medicine,computer,remote sensing and other fields drawing support from its advantages.However,the use of ensemble learning method is really limited in the field of geo-hazards spatial prediction.This paper is devoted to applying the Stacking ensemble method for geo-hazards susceptibility assessment in Lushui County.For this,spatial database was constructed on the basis of detailed field survey at 1:50000 and various sources.The main achievements of this paper are as follows:(1)The physical geography,meteorology,hydrology and geological environment conditions in the study area were synthetically analyzed.Based on these natural conditions and field investigation,the characteristics,types,spatial and temporal distribution and mechanism of geo-hazards are analyzed,and the development law of geo-hazards is summarized,so that the relationship between geo-hazards and the conditions in Lushui County was clear to be understood.(2)Based on the full analysis of the temporal and spatial distribution of geo-hazards and the related conditions,sixteen factors including the elevation,slope angle,slope aspect,plane curvature,profile curvature,geomorphological type,land use,rock and soil type,distance to roads,road density,distance to rivers,river density,distance to faults,fault density,annual average rainfall and normalized difference vegetation index were extracted from field investigation,secondary data,digital website and remote sensing images and other sources,which were considered as the conditioning factors of geo-hazards in the study area.The temporal and spatial distribution characteristics of these factors and their relationship with the occurrence of geo-hazards are analyzed.(3)After preparation of conditioning factors,it is necessary to unify the data type.The continuous factors were discretized in a supervised way.The information value method was used to quantify the importance of each factor and the interval of each factor to geo-hazards,so as to single out the important factors and eliminate the relatively unimportant factors.According to the calculation results,out of the sixteen,nine most important factors were selected as the input data for the modeling.The IV vales of these nine factors from large to small is distance to roads,elevation,road density,land use,normalized difference vegetation index,geomorphological type,distance to faults,fault density,rock and soil type.(4)On the basis of analysis of the regularity of geo-hazards and the influence degree of each factor,the selected factors were input into different single machine learning models for training,and the performance of each single model was compared.In this paper,three classical machine learning models,Support Vector Machine(SVM),Artificial Neural Network(ANN)and Classification and Regression Tree(CART),were selected as the base-learners to construct the ensemble framework.The performance of the ensembles composed of these three base-learners was compared with that of each base-learner,and the AUC value is taken as the main criterion to measure the performance of models.The validation results show that the ensemble model(SVM-ANN-CART)outperformed than all individuals.Therefore,the susceptibility maps of geo-hazards in the study area was completed by using ensembles in GIS software,and was then divided into five levels: very low,low,moderate,high and very high.Ensemble learning can enhance the spatial prediction ability,and plays an important role in improving the accuracy of geo-hazard susceptibility assessment,especially in identifying high-risk areas.Therefore,the proposed ensemble model in this paper is recommended to be widely used in geo-hazard analysis in order to identify vulnerable areas more accurately to take timely measures to effectively reduce harmful impact.