Mapping Landslide Hazard Risk Using Machine Learning Algorithm In Guixi,Jiangxi,China

The frequently occurred landslide hazards in Guixi,Jiangxi have seriously threatened the safety of lives and property of local residents.Research on landslide hazards is hence critical for decision-makers to implement disaster prevention and mitigation measures.Taking Guixi as a case study,this thesis analyzed the characteristics of landslide hazards and their relationship with various hazard-causative factors based on field investigation.According to the conditions of landslide occurrence and our field knowledge,we selected21 geo-environmental factors including lithology,faults,rainfall,slope,roads and land cover types,etc.,for achieving our research.After preprocessing with GIS platform including digitization and weight assignment of the categorical factors,Random Forests(RFs)and Support Vector Machines(SVMs)algorithms were utilized to conduct landslide risk modeling and prediction in the study area,and the following procedures were followed and the main results were obtained:1)The development characteristics and distribution of landslides in the study area and their relationship with different geo-environmental factors were analyzed.Most of the landslides area are small in scale,which are mainly distributed in the southeast part and along the roads of the study area.Our analysis revealed that concentrated rainfall,slope and human activities are the main factors causing landslides.2)Digital preprocessing of landslide survey data includes vectorization of the 273 field landslides points and stochastic division of these points into two groups,70% as training set and 30% as validation set.And 380 points were randomly sampled in the non-risk stable area(slope < 3°)and integrated into these two sets in the same ratio as stable non-risk samples.These datasets were converted into raster of 30 m in resolution.Non-digital categorical factors such as lithology and land cover were assigned a weight according to their resistance of rock types or land cover type to landslide;and linear features,e.g.,faults and roads were buffered and each buffer zone was assigned a weight in terms of the scale of the linear feature and its proximity to this feature.This weighting provides a basis for subsequent quantitative analysis and modeling.3)Within the En Map-Box package developed using IDL language,the RF and SVM algorithms were applied for parameterization,or rather,landslide risk modeling.The obtained models were calibrated against the validation set,and the Kappa Coefficient(KC)and Overall accuracy(OA)revealed that both RF and SVM models are of high accuracy and reliability,and especially,RF model performed better.These models were applicable for landslide risk prediction and mapping,at the same time,demonstrating the contribution of each factor to the landslide events.4)The probability of occurrence of landslides in the study area was classified into five levels: No-risk,Low-risk,Medium-risk,High-risk,and Extremely high-risk areas.The landslide risk map is consistent with the field survey,and thus,the results may serve as reference or technical support for the local governments while taking action on disaster prevention and early warning.