Research Of Dynamic Spatio-temporal Graph Neural Network And Its Application In Landslide Prediction

Landslides are one of the common geological hazards that cause serious damage to people’s life and property safety every year.Over the decades,researchers have made great efforts to predict landslides.With the development of the technological progress over time,the technical of landslide prediction has gone through a long progress from highly relying on experts and technicians for analysis,to using statistical models for prediction,and finally to the current prediction methods based on artificial intelligence and deep learning,which have achieved great success and gradually got rid of the high reliance on human labor towards automation and intelligence.Land deformation prediction is the main research target of landslide prediction,which is in the scope of spatio-temporal data prediction.The current research paradigm of spatiotemporal data prediction uses convolutional neural networks or graphical neural networks to aggregate spatial features,and uses temporal models to predict the temporal dimension and finally completes the modeling of spatio-temporal dependencies.In some fields such as traffic flow prediction,there are already many mature theoretical methods and excellent practical results,but in land deformation prediction,there are still some urgent problems and challenges:(1)unable to learn the latent node relationships and graph structures on unlabeled data,and the lack of metrics for the intrinsic relationships between nodes?(2)unable to learn the manifold structure of mountain surface and model complex spatial relationships?(3)unable to generate dynamic spatio-temporal embedding representations and model dynamic spatial features?(4)unable to combine spatial dependence and temporal dependence under a unified theoretical framework.To address these challenges,we propose two models successively.First,to address the challenges(1-2),we propose the SA-GNN model and the WLLE manifold learning method,which effectively combines geological features and manifold learning,which is one of metric learning.It constructs a better adjacency for aggregation of spatial features by mapping nodes to manifold space,where the distance between nodes can better measure the intrinsic spatial relationship between nodes.Based on SA-GNN,we further improve its shortcomings by proposing the Dy Land model,which attempts to give a solution to challenges(1-4).Dy Land first models the dynamic manifold embedding using continuous normalizing flow,and then models the coevolution of temporal and spatial features using the neural ordinary differential equations to efficiently learn the spatio-temporal dependencies.We also provide a high level abstraction and overview of the spatio-temporal modeling process and refine a unified theoretical framework for spatio-temporal data prediction,which is also informative for other researchers.To verify the effectiveness and reliability of the model,we conduct sufficient experiments on In SAR datasets from continuously monitored landslide-prone slopes.After comparisons and analyses,we demonstrate that SA-GNN and Dy Land outperform other comparative baseline methods in terms of prediction accuracy and model interpretation.