Reliable Landslide Susceptibility Evaluation Method Based On The Graph Neural Network With Constraints Of Environmental Similarity

Landslide disasters are the most frequent geological disasters that cause the greatest losses in China,especially in the mountainous areas of western China where the terrain conditions are extremely complex.Landslide disasters are widespread,seriously threatening the safety of human life and property and the safety of major infrastructure projects in these areas.Affected by special geomorphic conditions,landslide disasters are highly concealed,sudden,and uncertain,making it difficult to effectively detect disaster precursor information.At present,there is an urgent need for effective regional landslide susceptibility evaluation to provide reliable high-risk areas for key investigation and prevention of landslide hazards.The landslide formative conditions are unclear,the disaster triggers are random,disaster laws are unknown.The accurate identification and quantitative analysis of landslide-prone areas have become hot spots and difficulties in the field of international landslide research.With its powerful nonlinear relationship extraction capability,machine learning can automatically find the key features of landslide from many landslide factor data and landslide samples,and quickly predict the characteristics of regional landslide susceptibility without manual intervention,which has become the mainstream method for landslide susceptibility evaluation.However,existing machine learning methods,which consider only the first law of geography,still face the following difficulties in complex environments:(1)the contradiction between the complex landslide mechanism and the rare sample.Scarce landslide samples are difficult to support complex landslide predictions,and additional knowledge constraints need to be introduced;(2)the contradiction between the complex formative environment and the limited feature extraction ability,the existing methods do not fully consider the relationship between landslide and its formative environment,the potential correlation between landslides and environment factors cannot be deeply excavated,and it is difficult to apply it in a wide range of environmental heterogeneous areas;(3)the contradiction between the collected samples and the heterogeneous environment.The heterogeneous environment makes the landslide characteristics of different regions different,and it is difficult to accurately transfer the knowledge of landslide discrimination across regions.To solve the above problems,this paper takes the background knowledge that the more similar the environment is,the more similar the landslide characteristics as the starting point,and innovatively proposes a graph neural network with constraints of environment consistency for reliable landslide susceptibility evaluation.The specific research contents of this paper are as follows:(1)The landslide susceptibility evaluation knowledge graph.Firstly,the landslide ontology model is formed by analyzing the characteristics and interrelationships of related elements of the landslide.Then,focus on the characteristics and needs of landslide susceptibility evaluation tasks,a landslide susceptibility evaluation ontology model containing landslide elements,observation data,and landslide characteristics is extracted and formed to construct the schema layer of the knowledge graph;according to the ontology factors,Data and features,corresponding entity elements and attributes are extracted from multi-source and multi-modal landslide data,and the relationship between data entities is supplemented by semantic association measurement.So that the data layer of the knowledge graph is formed.Finally,the knowledge graph is stored in the graph database to provide an interface for machine learning to understand the prior knowledge of landslides.(2)Graph Neural Networks with Constraints of Environmental Similarity is constructed for Landslide Susceptibility Evaluation.Based on the knowledge graph,firstly,the geographic polygons are generated by terrain polygon approximation and defined as the evaluation unit entities in the knowledge graph.Then,based on the entity environment characteristics and unit features,the entities are associated to form a regional graph construction.Subsequently,a sampling strategy is designed to label non-landslide samples and prepare a balanced training set.A graph neural network constrained with environmental similarityis established to realize the aggregation of landslide feature information in similar geographical environments,and better extract landslide representations in different environment.The final node representations are used to determine the landslide susceptibility.(3)Environment-aware unsupervised transfer learning method is designed for landslide susceptibility evaluation without samples.Firstly,an environment-aware module is established based on the environmental classification and similarity of geographic unit,so that the sub-environments in different regions can be adaptively associated.Then a transfer learning network with multiple domain adversarial modules is established,so that the model knowledge trained in other regions with sufficient samples can be accurately applied to target regions without samples.Finally,in the area associated with the source domain and the target domain,typical feature units are selected as pseudo-labeled samples for training the domain classifiers,and the proposed transfer network is trained together with the landslide label samples in the source domain.The landslide susceptibility evaluation is generated in the region without landslide samples by the fitted model.Finally,this study takes the typical areas of Fengjie County,Fuling District and Qijiang District of Chongqing City as examples for experimentation.The results verify the effectiveness of the proposed method.