Research On Landslide Hazard Assessment Based On Multi-Source Spatiotemporal Data

Landslides are the most frequent geological disasters in China,which hinder the development of national urbanization and regional economic construction and bring huge hidden dangers to the safety of people’s lives and properties.Landslide hazard assessment is an evaluation of the possibility of landslide occurrence within a certain location,time and intensity range,and is obtained through comprehensive calculation of spatial probability,time probability and intensity probability.Carrying out landslide hazard assessment is of great significance to the national disaster prevention and mitigation work.Although scholars have carried out some landslide hazard assessment work for different regions,there are still some problems in practical research due to the complex and extensive landslide hazard assessment mechanism.Spatial probability is the most complex part of landslide hazard assessment because it involves a large number of complex geographical environment factors.The calculation of landslide spatial probability is called landslide susceptibility assessment.The supervised machine learning model is currently the most commonly used method in landslide susceptibility assessment.However,the natural imbalance of the proportion of landslide and non-landslide samples in the training of the susceptibility assassment model has not been solved.There is no consensus on the best machine learning models for susceptibility assessment.In this paper,the landslide sample equalization research is carried out from the positive and negative perspectives through the landslide positive sample oversampling method based on Synthetic Minority Oversampling Technique and the sensitivity analysis-based landslide negative sample undersampling method.Both datasets are experimented with multiple machine learning models.Experiments show that the susceptibility evaluation based on multilayer perceptron neural network and undersampled dataset has the best comprehensive performance.Because the distribution of geographic information data involved in spatial probability varies greatly among different regions,the landslide susceptibility assessment model is often limited to a specific spatial range,which also restricts the scope of hazard assessment.In this paper,the domain adaptation mechanism in the field of computer vision is introduced into the landslide susceptibility assessment,and a multi-layer perceptron landslide susceptibility assessment transfer model based on domain adaptation is constructed.Model transfer experiments are carried out between the study area and the transfer area.The experimental results show that the proposed domain-adapted multilayer perceptron susceptibility assessment transfer model has better performance in both model accuracy and hierarchical graph credibility.In the calculation of spatial probability,time probability and intensity probability of landslide hazard assessment,a variety of mathematical models and a variety of multi-source heterogeneous spatio-temporal data are required.The processing efficiency of spatial and temporal data is relatively low,and the unified management and reuse capabilities of mathematical models are relatively weak.In this paper,knowledge graph technology is introduced into landslide hazard assessment,and a multi-source spatiotemporal knowledge graph is constructed for landslide hazard assessment.It can improve the analysis efficiency of multi-source heterogeneous spatiotemporal data in the hazard assessment,and enhance the unified management and reuse ability of mathematical models.In addition,this paper also conducts a specific case application study with dynamic precipitation as a trigger event based on the spatiotemporal knowledge graph for landslide hazard assessment,which verifies the effectiveness of the method.