| As a sudden regional geological disaster,landslide has the characteristics of wide distribution,high frequency and strong damage.China is one of the countries most seriously affected by landslide disasters in the world.Landslide disasters bring huge losses to national construction,people’s lives and property every year.With the rapid development of economy and society,human interference to the natural environment has intensified,leading to a continuous increase in the frequency and intensity of landslide disasters.The landslide susceptibility analysis method can evaluate the spatial distribution and occurrence probability of landslides,and its prediction results have a key scientific support role in the planning of landslide disaster risk management,urban planning and landslide monitoring.There are several important limitations in the existing quantitative analysis methods of landslide susceptibility: the selection of non-landslide samples is fuzzy and uncertain,and the construction of non-landslide negative samples is non-uniform due to general conditions;The construction of landslide susceptibility function leads to oversimplification of the model,ignoring the role of landslide spatial relationship features and shape features,and lacking the fusion of spatial scale features.In response to the above problems,this paper uses deep learning technology as the main method,combined with the theory of landslide disaster risk analysis,to complete the study of landslide susceptibility analysis methods that take into account sample quality and spatial characteristics.On the basis of research and analysis of technical achievements in fusion of geographic information spatial features,optimize the negative sample construction strategy to eliminate the interference of sample selection and trigger factor correlation,and study the spatial feature expression,spatial feature fusion and spatial feature analysis technology of landslide susceptibility data,and constructed a susceptibility evaluation system taking into account the spatial characteristics,the main achievements and conclusions are as follows:(1)By analyzing the spatial distribution of landslide historical disaster points and related trigger factors information,we have summarized the current selection method of negative samples of non-landslide points,and a landslide sample construction strategy based on fuzzy C-means classification method is constructed.Fuzzy C-means clustering can deal with ambiguity and uncertainty through the concept of membership,and construct training sample data with the constraint of maximizing the variance between positive and negative samples,so as to obtain a more robust negative sample set.The susceptibility model constructed by support vector machine is used to compare and analyze three sample optimization strategies.The analysis results prove that the sample distribution based on fuzzy C-means is more reasonable,and the susceptibility model has better accuracy.(2)A three-dimensional convolutional neural network landslide susceptibility analysis method is proposed that takes into account the landslide trigger factor-spatial characteristics.This method reconstructs the trigger factors and the spatial attributes of neighboring pixels into threedimensional layer data,using The three-dimensional convolution kernel extracts the trigger factorspatial features,and at the same time fuses the spatial neighborhood features of the landslide unit and the correlation features of the trigger factors.Compared with support vector machines and convolutional neural networks,the proposed method shows higher accuracy and better prediction results.(3)Construct a pyramid dilated convolutional neural network landslide susceptibility analysis model.By combining the advantages of pyramidal convolution multi-scale feature extraction and dilated convolution with large receptive field and low parameters,a pyramid dilated convolution coupling module is constructed to increase multi-scale spatial feature information.At the same time,the module supports embedding a 3D convolutional neural network model.Compared with single-scale convolutional neural network and 3D convolutional neural network models,the convolutional neural network and 3D convolutional neural network that integrate multi-scale spatial feature pyramid expansion convolution The susceptibility model significantly improves prediction accuracy.(4)Combining landslide edge and spatial shape features into the convolutional neural network landslide susceptibility analysis model,the influence of different expressions of landslide boundary and spatial shape on the landslide susceptibility model is studied.Through the data representation of four different landslide boundaries and spatial shapes of point,circle,square and polygon,respectively combined with the convolutional neural network model,the degree of uncertainty of the four types of landslide susceptibility is evaluated.Experimental results prove that the convolutional neural network model based on polygonal shape features has the best performance.The paper has 64 pictures,19 tables,and 284 references. |
