CN116205522B

CN116205522B - Landslide vulnerability evaluation method and system based on multidimensional CNN coupling

Info

Publication number: CN116205522B
Application number: CN202310046502.7A
Authority: CN
Inventors: 路文娟; 毛曦; 程瑶; 马维军; 王继周; 高武俊; 赵占骜
Original assignee: Chinese Academy of Surveying and Mapping
Current assignee: Chinese Academy of Surveying and Mapping
Priority date: 2023-01-31
Filing date: 2023-01-31
Publication date: 2023-10-20
Anticipated expiration: 2043-01-31
Also published as: CN116205522A

Abstract

The invention discloses a landslide vulnerability assessment method and system for multi-dimensional CNN coupling, which relate to the technical field of landslide vulnerability assessment.

Description

Landslide vulnerability evaluation method and system based on multidimensional CNN coupling

Technical Field

The invention relates to the technical field of landslide susceptibility evaluation, in particular to a multi-dimensional CNN coupled landslide susceptibility evaluation method and system.

Background

Landslide is one of common natural disasters, and when landslide occurs, the landslide not only damages the natural environment, but also damages villages, urban houses, infrastructure and the like, and when serious, a large amount of casualties are caused. The landslide vulnerability evaluation can discover the complex relation between landslide and disaster-pregnant environment through the selection of the model and data, gives the occurrence probability of unknown landslide in similar environment through comprehensive judgment, is an effective visualization technology for landslide area positioning and sustainable management of land resources, and can provide powerful technical support for decision managers. The modeling process of quantitative landslide susceptibility evaluation involves several important problems such as acquisition of landslide and non-landslide records, extraction of landslide related environmental factors, and model selection, and from the aspect of methods, it is still a challenge to apply a reliable and effective model to judge the landslide susceptibility level.

Currently, landslide susceptibility evaluation methods based on statistical analysis have become mainstream, existing statistical methods can be classified into classical statistical methods and Machine Learning (ML) based methods, and in consideration of the complexity of landslide occurrence mechanisms, classical statistical methods such as a frequency ratio method, a comprehensive index method and the like summarize the relationship between landslide and its induction factor in a linear form, traditional ML models such as Logistic Regression (LR), support Vector Machines (SVM), random Forests (RF) and the like are often inaccurate, and have been widely applied to landslide susceptibility evaluation, and compared with classical statistical methods, higher precision results are revealed. With the development of technology, a Deep Learning (DL) model has an effect superior to that of a traditional ML model in aspects of image classification, voice recognition and the like, DL is used as a resumption development of a neural network in ML, and the structure of a multi-hidden layer of the DL enables the DL to fully mine advanced features of a data set and to add nonlinear representation fitting abstract features. As one of the most mature and popular DL frameworks, convolutional neural networks (convolutional neural network, CNN) have also been widely used in the field of geography, such as landslide detection using remote sensing images, seismic waveform features identification from seismograms, etc., while CNN is still in the development stage in terms of landslide vulnerability assessment. CNN is a kind of feedforward neural network with depth structure, and can be divided into one-dimensional convolutional neural network (one-dimensional convolutionalneural network, 1D-CNN), two-dimensional convolutional neural network (two-dimensional convolutional neural network, 2D-CNN) and three-dimensional convolutional neural network (three-dimensional convolutional neural network, 3D-CNN) according to different definition modes of the dimension or convolutional kernel of the received feature map. In the first landslide susceptibility evaluation based on CNN, wang et al try three architectures respectively, and the results show that the overall accuracy of the three architectures and the indexes such as the Markis correlation coefficient are obviously improved compared with the traditional ML method, and the applicability of CNN in landslide susceptibility evaluation is shown.

The improvement of the model precision is an important direction of CNN in the development of landslide susceptibility evaluation, the deepened network hierarchy is the operation of DL model improvement precision, more levels can be iteratively extracted from low-level features to higher-level representations, but for CNN, the deeper network is more prone to generate gradient disappearance or explosion, so that the model is difficult to converge, the number of layers is increased along with the increase of parameters, the model training difficulty or the result overfitting and other problems are caused, therefore, other models and CNN are combined to improve the model precision, the model combination mode can generally obtain more comprehensive feature learning, the model scale is increased, the model parameter quantity and the calculated quantity are greatly increased, the model training fitting difficulty is high, the prediction reasoning speed is low, and the overall efficiency is reduced in landslide emergency disaster treatment. In a comprehensive view, the CNN is still in an exploration stage in landslide susceptibility evaluation application at present, and the result presentation and the precision improvement still have development space. The 1D-CNN and the 2D-CNN are widely applied to landslide susceptibility evaluation due to strong feature extraction capability, however, in the current method for improving the precision, model parameters and calculated amount are inevitably increased by means of hierarchical deepening or model combination, so that model training is difficult and results are over-fitted, and model efficiency is low.

Therefore, how to guarantee model efficiency while improving accuracy in landslide susceptibility evaluation is a problem to be solved in practical application.

Disclosure of Invention

The invention aims to provide a landslide vulnerability evaluation method and system of multidimensional CNN coupling, which can improve prediction accuracy and ensure model efficiency.

In order to achieve the above object, the present invention provides the following solutions:

a landslide vulnerability assessment method of multidimensional CNN coupling, the landslide vulnerability assessment method comprising:

acquiring a factor distribution diagram of each landslide influence factor in a target area; the landslide impact factors comprise topography factors, geological condition factors and environmental condition factors;

using all the factor distribution graphs as input, and determining landslide probability of each position point of the target area by using a trained multidimensional CNN coupling model so as to evaluate landslide susceptibility of the target area; the trained multidimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence; the two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer.

A multi-dimensional CNN coupled landslide vulnerability assessment system, the landslide vulnerability assessment system comprising:

the data acquisition module is used for acquiring a factor distribution diagram of each landslide influence factor in a target area; the landslide impact factors comprise topography factors, geological condition factors and environmental condition factors;

the evaluation module is used for determining landslide probability of each position point of the target area by using the trained multidimensional CNN coupling model by taking all the factor distribution graphs as input so as to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence; the two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

the invention provides a landslide susceptibility evaluation method and system for multi-dimensional CNN coupling, wherein a trained multi-dimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are sequentially connected, so that the one-dimensional convolutional neural network and the two-dimensional convolutional neural network are coupled by utilizing asymmetric aggregation, the network depth is maintained to limit model parameters and reduce calculated amount, and then a factor distribution diagram of a plurality of landslide influence factors in a target area is taken as input, and the trained multi-dimensional CNN coupling model is utilized to evaluate the landslide susceptibility of the target area, thereby improving the prediction precision and guaranteeing the model efficiency.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a method for evaluating landslide susceptibility according to embodiment 1 of the present invention;

FIG. 2 is an overall flow chart of constructing a data set according to embodiment 1 of the present invention;

fig. 3 is a schematic diagram of a network structure of a multidimensional CNN coupling model according to embodiment 1 of the present invention;

FIG. 4 is a schematic view of the range of the Gordon trench domain provided in example 1 of the present invention;

FIG. 5 is a graph showing historical landslide distribution of the color Dongpu ditch provided in embodiment 1 of the present invention;

FIG. 6 is a factor distribution sample graph of the landslide impact factor of Isodon japonicus provided in example 1 of the present invention;

FIG. 7 is a graph of the susceptibility to landslide of the Dongpu ditch using different models provided in example 1 of the present invention;

FIG. 8 is a comparison of training results of different models provided in example 1 of the present invention;

FIG. 9 is a schematic diagram of the result of confusion matrix for different model test sets according to embodiment 1 of the present invention;

FIG. 10 is a schematic diagram of the result of the multiple co-linearity analysis provided in example 1 of the present invention;

FIG. 11 is a graph showing the frequency ratio of each landslide impact factor attribute interval according to embodiment 1 of the present invention;

fig. 12 is a system block diagram of a landslide vulnerability evaluation system provided in embodiment 2 of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.

Example 1:

the embodiment is used for providing a landslide vulnerability evaluation method coupled by multidimensional CNN, as shown in fig. 1, the landslide vulnerability evaluation method comprises the following steps:

s1: acquiring a factor distribution diagram of each landslide influence factor in a target area; the landslide impact factors comprise topography factors, geological condition factors and environmental condition factors;

In this embodiment, the topography factors may include elevation, gradient, slope, plane curvature, section curvature and topography humidity index, the geological condition factors may include lithology, distance from fault and distance from earthquake, the environmental condition factors may include normalized vegetation index and distance from channel, and in this embodiment, initial vector data of fault, earthquake and channel are used as input, and distance from fault, distance from earthquake and distance from channel can be obtained by using euclidean distance calculation method in ArcGIS software. And for each landslide influence factor, the value of the landslide influence factor at each position point of the target area is the factor distribution diagram of the landslide influence factor in the target area.

S2: using all the factor distribution graphs as input, and determining landslide probability of each position point of the target area by using a trained multidimensional CNN coupling model so as to evaluate landslide susceptibility of the target area; the trained multidimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence; the two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer.

According to the embodiment, all factor distribution graphs are divided by using a moving window method, a window of the corresponding position of each factor distribution graph is used as input, and the landslide probability of a position point located at the center of the window in a target area is calculated, so that the position point-by-position prediction is carried out on the target area, the landslide probability of each position point in the target area is finally obtained, and if the landslide probability is larger than a preset threshold value, the position point is determined to be a landslide point, otherwise, the position point is a non-landslide point.

Before S2, the landslide susceptibility evaluation method of the embodiment further includes a step of training the multidimensional CNN coupling model to obtain a trained multidimensional CNN coupling model, where the step may include:

(1) And acquiring a data set, and dividing the data set into a training set, a verification set and a test set according to a preset proportion. The data set comprises a plurality of groups of factor sample sets and labels corresponding to each group of factor sample sets, the factor sample sets comprise factor sample graphs of each landslide influence factor, and the labels are landslide or non-landslide.

Accurate sample data is a precondition of model learning, in order to meet the requirement of model training while ensuring reliable information, the embodiment adopts a window extraction mode to construct a data set, the whole flow is shown in fig. 2, and specifically, the data set acquisition can include:

1) And acquiring a factor distribution sample graph of each landslide influence factor in the target area and a landslide cataloging graph of the target area, wherein the landslide cataloging graph identifies the landslide position and the non-landslide position of the target area.

The method for acquiring the landslide cataloging diagram comprises the following steps: the method comprises the steps of respectively obtaining a digital elevation model before and after landslide disaster occurs in a target area by utilizing a satellite remote sensing technology or an unmanned aerial vehicle platform, observing the digital elevation model before landslide disaster occurs and the digital elevation model after landslide disaster occurs, taking an area with the front and rear elevation change exceeding 10m as a landslide area to determine the landslide position and the non-landslide position of the target area, and marking the landslide position and the non-landslide position on each position point of the target area to obtain a landslide cataloging diagram.

The method for obtaining the factor distribution sample graph of each landslide impact factor in the target area comprises the following steps: and for each landslide influence factor, acquiring the value of the landslide influence factor at each position point of the target area before landslide hazard occurs, and forming a factor distribution sample graph of the landslide influence factor in the target area.

Preferably, the method for evaluating landslide susceptibility according to the present embodiment further includes: in the preprocessing stage, histogram equalization is carried out on the factor distribution sample graph to obtain an equalized image, binarization processing is carried out on the landslide catalogue graph to obtain a binarized image, and the equalized image and the binarized image are used as new factor distribution sample graph and the landslide catalogue graph to execute the step 2).

The histogram equalization refers to removing larger or smaller sparse values in a factor distribution sample graph (which can be a grid graph), namely, observing overall distribution trend of data in the factor distribution sample graph, removing the sparse values with extremely small number under the condition of not affecting overall distribution rule so as to display variation condition of the values with large number, normalizing pixels of the factor distribution sample graph after removing the sparse values, and stretching the pixels of the factor distribution sample graph after normalization processing in a preset proportion so as to dynamically stretch pixel difference to 0-255. The preset proportion of different landslide impact factors is the same when stretching, namely stretching is carried out in the same proportion. And the factor distribution sample graph of each landslide influence factor is subjected to histogram equalization and then pixel differences are dynamically stretched to be between 0 and 255, so that the landslide influence factor has rich and obvious characteristics under a single-band bitmap, and the subsequent model reading and processing are convenient. The binarization processing means that the pixel value of the landslide position in the landslide cataloging diagram is set to 255, the pixel value of the non-landslide position is set to 0, so that the landslide cataloging diagram is processed into a black-white image for display, the landslide cataloging diagram is in two types of landslide and non-landslide after being subjected to binarization processing, so that data screening is carried out, positive and negative samples required by model training are screened, the positive samples are landslide samples, and the negative samples are non-landslide samples.

2) And extracting a plurality of windows from the landslide cataloging graph by using a moving window method. For each window, respectively extracting a region corresponding to the window position from each factor distribution sample graph to obtain a factor sample graph of each landslide influence factor, and forming a factor sample set corresponding to the window; taking the landslide condition of the central pixel of the window as a label corresponding to the window; and the factor sample sets and the labels corresponding to all the windows form a data set, and the number of the factor sample sets, of which the labels are landslide, in the data set is the same as the number of the factor sample sets, of which the labels are non-landslide.

The traditional method for extracting the factor attribute according to the landslide point position is seriously dependent on the accuracy of landslide positioning, the landslide belongs to a complex physical process, the occurrence of the landslide is often closely related to the surrounding environment, and therefore, the embodiment adopts a moving window to extract the landslide space neighborhood to construct a data set. The window side length is set, and the window can be moved by one pixel step length each time when the window moves in the landslide cataloging graph, so that a plurality of windows are determined. The embodiment can obtain 1058×963 landslide cataloging diagram of the target area under the resolution of 10m, and finally establishes 13 pixels as window side length through test, so that the result shows that the adopted image size and window size can contain enough characteristics and avoid data redundancy.

For each window, respectively extracting a region corresponding to the window position from each factor distribution sample graph, namely extracting a region which is completely corresponding to the window position from the factor distribution sample graph, so as to obtain a factor sample graph of each landslide influence factor, wherein the factor sample graph of all the landslide influence factors forms a factor sample set corresponding to the window, the landslide condition of a central pixel of the window is used as a label corresponding to the window, the label is landslide or non-landslide, the association of the surrounding environment and the central pixel is determined, one sample is formed by the factor sample set corresponding to one window and the label, the label is a positive sample, and the label is a non-landslide sample. And the factor sample sets and the labels corresponding to all the windows form a data set.

In consideration of the problems that small-area landslide is less and samples are insufficient, the method for evaluating landslide susceptibility according to the embodiment introduces data enhancement processing, and expands a factor sample set corresponding to a limited positive sample through the data enhancement processing, namely after the data set is obtained, the method for evaluating landslide susceptibility according to the embodiment further comprises the following steps: and carrying out data enhancement on each group of factor sample sets with landslide labels in the data set to obtain a plurality of groups of enhanced sample sets, wherein the data enhancement comprises horizontal overturning and vertical overturning so as to expand the data set to obtain an expanded data set, the expanded data set comprises a plurality of groups of enhanced sample sets besides the data in the original data set, the labels of each group of enhanced sample sets are landslide so as to expand positive samples in the data set, and the expanded data set is used as a new data set to execute a subsequent model training process.

The increase of the number ratio of positive and negative samples will reduce the accuracy of positive sample prediction, which is contrary to the landslide early warning goal, so that the embodiment adopts a random equalization sampling strategy to set the number of non-landslide samples (namely negative samples), namely, a plurality of negative samples are cut at random positions in a landslide inventory map in the same number as the positive samples, so that the number of positive samples is the same as the number of negative samples, and the number of factor sample sets with labels of landslide in the data set is the same as the number of factor sample sets with labels of non-landslide. When there is data enhancement processing, in order to achieve this, the present embodiment may first determine a plurality of positive samples in the landslide catalog, then perform data enhancement processing on the positive samples to expand the positive samples, determine the number of positive samples after expansion, and then determine negative samples in the landslide catalog, which are the same as the number of positive samples after expansion, thereby constructing and obtaining a data set.

After acquiring the data sets, the data sets are divided into training sets, verification sets and test sets according to preset proportion after being disordered, wherein the preset proportion can be 6:2:2 or 8:1:1. the training set is used for model learning landslide features and training internal parameters of the model; the verification set is used for checking the training state of the model and is used as a reference for optimizing the super parameters of the model; the test set is used for checking the final generalization error of the model and evaluating the robustness of the model.

(2) And training the multidimensional CNN coupling model by taking the training set as input to obtain a trained model.

According to the embodiment, a new landslide susceptibility evaluation model (namely a multidimensional CNN coupling model) is constructed by utilizing a 2D-CNN and 1D-CNN coupling mode, the advantages of two convolution neural networks are combined to realize automatic learning of the spatial neighborhood of landslide factors and related characteristics among different factors, and under a coupling structure, the model maintains depth while the parameter quantity and the calculated quantity are prevented from being greatly increased, so that the model efficiency can be ensured, and the model prediction precision can be improved. The multidimensional CNN coupling model of the embodiment is mainly divided into five parts, and comprises a two-dimensional convolutional neural network (2D-CNN), a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network (1D-CNN), a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence, wherein the network structure is shown in figure 3, and in figure 3, C is the number of characteristic channels.

The two-dimensional convolution neural network comprises a plurality of two-dimensional convolution pooling blocks and a two-dimensional convolution layer which are sequentially connected, wherein the two-dimensional convolution pooling blocks comprise a two-dimensional convolution layer and a two-dimensional maximum pooling layer which are sequentially connected. In fig. 3, the number of two-dimensional convolution pooling blocks is 2, the two-dimensional convolution layer of the first two-dimensional convolution pooling block adopts a convolution kernel of 2×2, and the pooling window of the two-dimensional maximum pooling layer is 1×2; the two-dimensional convolution layer of the second two-dimensional convolution pooling block adopts a convolution kernel of 3×3, the pooling window of the two-dimensional maximum pooling layer is 1×2, and the two-dimensional convolution layer connected behind the two-dimensional convolution pooling blocks adopts a convolution kernel of 3×3. The model receives image input of a data set, different landslide influence factors are overlapped in the channel dimension to form a landslide factor multispectral image with various characteristics, the image comprises spatial neighborhood characteristics of landslide or non-landslide occurrence, and shallow characteristic learning is carried out through a two-dimensional convolution layer and a two-dimensional maximum pooling layer. By adopting the 1 multiplied by 2 asymmetric pooling window, features are aggregated in one dimension, and only a representative one of two pixels is aggregated at a time due to the fact that the pooling window is smaller, so that the defect of data loss can be reduced. As the feature map is reduced, the model calculation amount is reduced by times. With the deepening of the network hierarchy, deep features in the feature map are extracted by using more convolution kernels, and spatial characteristics formed by different dimensions are more fully learned.

The two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer. In fig. 3, the pooling window of the two-dimensional maximum pooling layer is 1×3, and the pooling window of the two-dimensional average pooling layer is 1×3. The two-dimensional maximum pooling layer and the two-dimensional average pooling layer are used for respectively acquiring representative information of aggregation dimension, landslide influence factor graph characteristics and interrelation characteristics among factors are acquired in the aggregation dimension through convolution kernel operation, the two pooling modes are respectively used for acquiring the representative information in different modes, for example, maximum pooling acquires maximum pixel values in a pooling window, average pooling acquires average pixel values in the pooling window, the pooling layer is used for reducing data redundancy, the purpose of acquiring the representative information in two modes is to weaken the influence of data loss, two pooling results are integrated through a concatate layer to reduce the loss of spatial semantic information caused by pooling operation, and meanwhile, the capacity of a model for containing abnormal information values is improved, so that model overfitting is avoided.

The one-dimensional convolutional neural network comprises a Reshape layer, a plurality of one-dimensional convolutional pooling blocks and a one-dimensional convolutional layer which are sequentially connected, wherein the one-dimensional convolutional pooling blocks comprise a one-dimensional convolutional layer and a one-dimensional maximum pooling layer which are sequentially connected. In fig. 3, the number of one-dimensional convolution pooling blocks is 2, the convolution kernel size adopted by the one-dimensional convolution layer of the first one-dimensional convolution pooling block is 3, and the pooling window of the one-dimensional maximum pooling layer is 3; the one-dimensional convolution layer of the second one-dimensional convolution pooling block adopts a convolution kernel size of 3, the pooling window of the one-dimensional maximum pooling layer is 3, and the one-dimensional convolution layer connected behind the one-dimensional convolution pooling blocks adopts a convolution kernel size of 3. The 1D-CNN receives two-dimensional input composed of characteristic length and characteristic channel number in a data form, the Reshape layer compresses the dimension with the length of 1 in the two-dimensional asymmetric aggregation module, the dimension is reduced on the premise of keeping the characteristic quantity unchanged, so that the dimension is reduced, then the residual dimension characteristics are learned through a one-dimensional convolution kernel with the window size of 3, the model not only learns the local characteristics of a single landslide influence factor, but also couples the characteristic association among different factors due to the characteristic sharing characteristic of the convolution kernel, and the whole network has a deeper structure along with deepening of the 1D-CNN layer, so that more abstract landslide and non-landslide characteristics can be extracted. The characteristic of the parameter sharing of the convolution kernels is the existing characteristic of CNN, namely the weight of each convolution kernel is fixed in the process of one iteration, the weight in the convolution kernel cannot be changed due to different positions in the image, each convolution kernel slides in the image to learn the independent characteristic, the model parameters are reduced due to the weight sharing, and the training difficulty of a network is greatly reduced.

The one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer. In fig. 3, the pooling window of the one-dimensional maximum pooling layer is 3, and the pooling window of the one-dimensional average pooling layer is 3. The characteristics are aggregated through average pooling and maximum pooling, and the integral characteristics (including abstract characteristic aggregation information represented by numbers) after being connected through a concatate layer are transmitted into a full-connection layer module.

The full-connection layer module comprises two full-connection layers which are sequentially connected. The fully connected layer maps the received feature vectors after flattening, so that the neural nodes are related to each other and have global features. The model nonlinear expression capacity can be improved by deepening the full-connection layer, but the model training cost is increased and the phenomenon of overfitting is caused, so that the model is only provided with two full-connection layers, the last full-connection layer outputs by a single neuron, and the output value is between 0 and 1, namely the probability of occurrence of landslide in the point positions is represented.

According to the embodiment, through designing the structure of the multidimensional CNN coupling model, the parameter quantity can be halved through asymmetric aggregation of the feature map of the 2D-CNN, and the parameter quantity and the calculated quantity can be effectively reduced through 1D-CNN learning of residual dimension data, so that the great increase of the parameter quantity and the calculated quantity can be restrained while the depth of the model is ensured, the problems of difficult model training and overfitting of results are avoided, and the model efficiency is ensured while the prediction precision is improved.

Model training mainly refers to the process of fully fitting features by iteratively updating weights between model layers, wherein the weights are regarded as parameters to be trained in a model, the model parameters determine the model scale, and the calculation amount between the parameters influences the model reasoning speed, and the model reasoning speed and the model parameters are related to structural design. The model parameters of the CNN are mainly concentrated in a convolution layer and a full connection layer, the convolution layer is taken as a main part, the function of the convolution layer is realized by means of weighted aggregation of convolution kernels on a feature map, the parameter quantity is related to the size of the convolution kernels and the number of channels of the input/output feature map, and a calculation formula is shown in the following formula (1). The 1D-CNN adopted in the model design only contains a single-dimensional convolution kernel, so that the parameter number can be reduced to a certain extent, and a deeper network becomes practical. The main calculation mode of the convolution layer is the sum of the cumulations and the operation among the matrixes, the calculated amount can be measured by floating point operands (floating pointoperations, FLPs), the FLPs are related to the feature map size, the number and the convolution kernel size, the calculation formula is shown in the following formula (2), and the feature map size is reduced in multiple in the 2D-CNN part of the model due to the asymmetric pooling effect, so that the calculated amount is reduced. Therefore, the multidimensional CNN coupling structure of the embodiment reduces the calculated amount while limiting the parameter quantity, effectively avoids the problem of training difficulty caused by deepening of the layer number of the neural network, and ensures the model efficiency.

Param＝C _out ×(k ² ×C _in +1)； (1)

In the formula (1), param represents the parameter quantity required by a convolution layer, k is the convolution kernel size, C _in And C _out Representing the number of channels of the input and output feature maps, respectively.

FLOPs＝H _out ×W _out ×(k ² ×C _in )×C _out ； (2)

In the formula (2), H _out And W is _out Representing the length and width of the output profile, respectively.

The embodiment can also add Dropout in the structure of the multidimensional CNN coupling model.

Part of element arrangement of the model needs to be considered besides the structural design of the model. For example, the linear function mapping between layers in the conventional neural network is difficult to express the real relationship between complex variables such as landslide, so the embodiment uses the modified linear unit (rectified linearunit, reLU) function in each part of convolution layer and full connection layer to activate the interlayer mapping, and effectively fits the output result and real data by using the nonlinear relationship. Meanwhile, the ReLU function can help to overcome gradient disappearance, is more efficient in training, further shortens the network learning period and reduces the calculated amount. In the final output of the model, single-value output of a sigmoid function is adopted, and the sigmoid function can map the output result in a distribution interval of 0 to 1 to represent landslide occurrence probability. That is, in this embodiment, the two-dimensional convolution layer, the one-dimensional convolution layer, and the full-connection layer all use ReLU functions.

The loss function is used for comparing the difference between the model predicted result and the real result, and further balancing the optimization direction of the model quality. In this embodiment, since the output result is in the form of landslide probability, a mean square error (mean squared error, MSE) is selected as the model loss function. The MSE may simply measure the "average error" of the predicted value and the true value, and its calculation formula is shown in the following formula (3). In the training process, when the model obtains a result close to fitting through a sigmoid function, the curvature of the MSE function is stable, and the approach of the model to a classification result is avoided, so that the overfitting of the model is avoided. Overall, the degree of variation of the MSE reflects the quality of the model, and the smaller the value of the MSE is, the better accuracy of the prediction model description experimental data is shown. In terms of network optimization, the present embodiment uses Adam Optimizer instead of the traditional random gradient descent method to update network weights more effectively. In this embodiment, when the multidimensional CNN coupling model is trained, the mean square error is used as a loss function, and the Adam optimizer is used to update the network parameters of the multidimensional CNN coupling model.

In the formula (3), m is the number of samples; y is _i A predicted value for the i-th sample; Is the label (i.e., actual value) of the i-th sample.

According to the embodiment, dropout is added in the model structure, the maximum pooling and average pooling splicing structure is utilized to avoid information loss to the greatest extent, meanwhile, the model structure has good containing capacity for abnormal information values, and the MSE and the output layer sigmoid function are jointly applied, so that the model has a more obvious gradient inhibition effect under the condition that the model is close to fitting, differentiation of results is avoided, and research is facilitated.

(3) Taking the verification set as input, and calculating the verification error by using the trained model.

In this embodiment, the MSE loss function is also selected when calculating the verification error.

(4) Judging whether an iteration termination condition is reached or not to obtain a judgment result; if the judgment result is yes, ending iteration, and taking the trained model as a trained multidimensional CNN coupling model; if the judgment result is negative, continuing iteration, and judging whether the verification error of the current iteration is the same as the verification error of the previous N iterations; if yes, the learning rate in the training process in the next iteration is adjusted, the trained model is used as a multidimensional CNN coupling model in the next iteration, and the step of training the multidimensional CNN coupling model by taking a training set as input to obtain a trained model is returned; if not, the step of taking the trained model as a multidimensional CNN coupling model of the next iteration, returning to the step of training the multidimensional CNN coupling model by taking the training set as input to obtain a trained model.

The iteration termination condition may be that a maximum number of iterations is reached. In this embodiment, the learning rate in the training process is dynamically adjusted by the callback function, the initial learning rate is set to be 0.001, and when the iteration result with the same performance level as the three performance levels appears along with the observation of the verification error in the iteration process, that is, when the verification error of the previous iteration is the same as the verification error of the previous 2 iterations, the learning rate is attenuated by 0.5 multiplying power.

Convolutional neural networks are widely used for landslide susceptibility evaluation due to their strong feature extraction capabilities. However, with the needs of scene diversification and high precision, the CNN algorithm is continuously improved, and the method of improving the precision by deepening the network hierarchy or combining other models tends to greatly increase the model parameters and the calculated amount, so that the model training is difficult or the result is over-fitted, further, the model efficiency cannot be ensured, and the practical application of the model is limited. The embodiment proposes to construct a multidimensional CNN coupling model to solve the problems, and connect a one-dimensional convolutional neural network and a two-dimensional convolutional neural network through asymmetric aggregation of feature graphs, so that the network depth is maintained to limit model parameters and reduce the calculated amount, and the model efficiency can be ensured while the accuracy is improved. The deep coupling characteristics among different dimensions and factors of each landslide influence factor can be further captured by utilizing multidimensional convolution kernel parameter sharing, so that the characteristic is fully utilized, and the overfitting is avoided.

Preferably, before acquiring the factor distribution map of each landslide impact factor in the target area, the landslide susceptibility evaluation method of the embodiment further includes: and calculating the variance expansion factor of each initial factor, and selecting the initial factor with the variance expansion factor smaller than a preset threshold as a landslide influence factor, so that the independence of the landslide influence factors is ensured.

In the following, the embodiment takes the Tibetan crow as an experimental area, selects 11 landslide influence factors to analyze the landslide vulnerability, comprehensively considers the model robustness, the reasoning speed and the result generalization capability, makes new attempts for CNN in landslide vulnerability evaluation application, and has important significance for regional engineering planning and disaster prevention and reduction.

As shown in FIG. 4, the region is the region of the Goodyear ditch, the Goodyear ditch is located in Milin county of Linzhi, an autonomous region of Tibet, in China, the Yalu canyon core region is located in the place, the ditch opening is connected with Yalu canbujiang, and the area of the ditch region is about 67.15km ² The main trench is 7.4km long. The northern upstream region of the color Dongpu ditch has wide topography, steep topography, glacier development, comprises the highest point in the domain, namely Gaola Bai Leifeng (7294 m), the downstream channel is narrowed, the gradient is slowed down, the lowest point, namely the ditch mouth (2746 m), and the whole height difference 4548m, and belongs to a typical mountain gorge valley region. The rock body in the region is seriously crushed, the shear strength is low, and complex topography is combined with strong movement under the influence of the structural stress of the Qinghai-Tibet plateau and the Himalayan mountain, so that a large number of landslide disasters are inoculated in the region. Meanwhile, due to the funnel effect of the area, the interaction of the collapse body, a large amount of glacier melted water and precipitation is easy to become a chip flow, the chip flow rushes out along a ditch and blocks the elegance Tibetan river to form a dam with a dam body, and high-level mountain floods are finally initiated when the dam body breaks. Currently, the east of color Thousands of people are affected by common ditch disaster chains, tens of villages and towns are destroyed, and local traffic logistics, engineering construction and the like are seriously affected. With the evolution of global climate change and geological structures, the risk of extremely large geological disaster chains caused by high-level collapse landslide still exists in the color Dongpu ditch. Existing studies analyze and explore the chromatopril from the aspects of formation conditions, occurrence processes, failure mechanisms and the like, but lack specific indications for the starting point of future disaster chains, i.e., the future landslide occurrence positions. Therefore, it is necessary to evaluate the susceptibility to landslide in this area, and it is recommended to perform long-term monitoring and early protection in an extremely highly susceptible area, and to fundamentally suppress occurrence of the crow-east jew disaster chain.

Because of the dangerous terrain of the color east Pugou, the traditional geological survey means are difficult to implement in the area, and the specific position, scale and volume of landslide cannot be determined. Therefore, the embodiment completes the screening of the historical landslide range by means of satellite remote sensing technology, unmanned plane platform, DEM differential and visual interpretation means. In 10 months in 2018, two events of brucella Jiang Dujiang caused by debris flow caused by landslide collapse occur in the ground, so that the focusing research period of the embodiment is between 2017 and 2020 before and after disaster occurrence, a future landslide occurrence risk area is explored from landslide development characteristics in the period, and the data used are satellite images of 2017, 12 months and resource No. three, satellite images of 2019, 8 months and ZC-3 unmanned aerial vehicles and satellite images of 2020, 12 months and resource No. three, provided by the national satellite remote sensing application center of the natural resource department and the China mapping science institute. The satellite of the third resource is the first civil high-resolution stereo mapping satellite of China, a digital elevation model (digital elevation model, DEM) with the spatial resolution of 10m and Gao Chengzhong error of 5m and including a research area can be constructed through stereo image pairs, and the elevation precision meets the requirements of state 1:1 ten thousand mountainous regions and high mountain regions DEM precision requirements, a research area DEM is extracted from the requirements, the unmanned aerial vehicle aerial image resolution is designed to be 0.1 meter, and the acquisition area is about 12km ² The mapping result satisfies state 1:2000 topography precision requirements, the unmanned aerial vehicle acquisition and construction DEM is resampled to the same spatial resolution and geographic registration is carried out through a feature point matching algorithm by taking the resource number three data as a reference, so as toFor subsequent calculation. The landslide of the color east Prussian channel is mainly represented by shear-slip deformation generated by loose mass of the trailing edge channel. Due to the large catchment area and poor rock integrity, landslide is overhead disintegrated into a detritus stream at the "start" instant. In order to comprehensively identify the source region of the chip flow of the color east Pu ditch, referring to the generalized landslide definition, the phenomenon that substances forming a slope move downwards and outwards is regarded as landslide, and the method is particularly applied to DEM changes in different time periods. However, due to the influence of shooting angles and geometric distortion, errors exist between the construction of the DEM and the true value, so that according to data processing results and the geomorphic characteristics of a research area, the error limit in the elevation of the mountain area is used as a dividing basis, namely, the area with the front-back change of the elevation exceeding 10m is used as a landslide range. The difference of the DEM constructed by the two-stage satellite images is obtained to obtain most landslide, and an unmanned aerial vehicle DEM difference result which covers the main ditch domain and has higher precision is added to supplement the result to obtain more accurate results near the main ditch. The deformation ranges under three time periods of three years are aggregated, then the smooth landslide edges are visually interpreted by combining images to accord with actual development characteristics, finally, the historical landslide distribution drawing of the Dongpu ditch is completed, and as a result, as shown in fig. 5, fig. 5 (a) is a landslide range determined based on 2017 and 2019 data, fig. 5 (b) is a landslide range determined based on 2019 and 2020 data, fig. 5 (c) is a landslide range determined based on 2017 and 2020 data, and fig. 5 (d) is a landslide range comprehensively determined based on 2017, 2019 and 2020 data. It is counted that 70 landslide points are totally identified in the research area, and the maximum complete landslide area is about 1.78km ² The minimum area is 3971m ² Most landslide is distributed near the middle channel and shows aggregation. Based on this, a landslide catalog of the color east Pu ditch can be constructed.

The landslide susceptibility evaluation is established based on the assumption that the future landslide is in the same environment as the previous landslide, so that the selection of an appropriate landslide influence factor is also an important link for successful landslide susceptibility modeling. Because the range of the research area is smaller and the information content is less, the embodiment reflects the landslide characteristics from the angles as much as possible, and respectively selects the landform class (elevation, gradient, slope direction, plane curvature, section curvature and terrain humidity index), the geological condition class (lithology, distance from fault and distance from earthquake), and the environmental condition class (normalized vegetation index and distance from channel) to reflect the landslide characteristics by 11 factors. To guarantee detailed information over the area, all landslide factors are at a resolution of 10m by 10m, and the results are shown in fig. 6. The data sources are respectively as follows: the elevation is generated by a 2017 resource third remote sensing image stereopair, and other topography related factors are derived by DEM; lithology and fault data nationwide 1: intercepting a 25 ten thousand geological map; the channel draws vectors according to the optical remote sensing images and the historical research, and the related distance is calculated according to Euclidean distance; the seismic point data is provided by the Chinese seismic platform network disclosure; NDVI is publicly provided by the national ecological data center resource sharing service platform. Based on this, a factor distribution sample graph of each landslide impact factor as shown in fig. 6 can be obtained.

To verify model accuracy and running effect, the present example was quantitatively evaluated by a comparative experiment. In this embodiment, four independent experiments are performed, which are respectively a proposed multidimensional CNN coupling model, and 1D-CNN, shallow 2D-CNN, and deep 2D-CNN, which are in the same order of magnitude as the parameters of the multidimensional CNN coupling model, are used to compare the difference between the existing method and the coupling method proposed in this embodiment in feature learning, and the deep 2D-CNN is used to verify the efficiency of the network proposed in this embodiment under the condition that the model scale is approximately the same. The specific parameters of the four models are shown in table 1.

TABLE 1

The four model structures are set according to the characteristics of the model structures and simultaneously keep variable control, and the four model structures adopt the same activation function, loss function, the same optimizer, learning rate and the like. In the network training process, in order to reduce the over-fitting phenomenon, a Dropout layer is also introduced into each network, and the Dropout operation plays an important role in improving the prediction performance, and the neural network units can be temporarily discarded according to a specific probability in the training process, so that the cooperative adaptation among hidden units is reduced, and the generalization capability of a prediction method is enhanced.

And reassigning each pixel point of the research area by using the trained model to prepare a landslide susceptibility map of the whole research area. Inputting landslide influence factors of a research area into each model for prediction, and dividing the obtained landslide probability into five landslide susceptibility grades by using a natural breakpoint method by using an ArcGIS after prediction is completed: very low, medium, high, very high. FIG. 7 is a graph of the susceptibility of the Isodon japonicus to landslide obtained with different models. There are differences among the four model results, but there is still some commonality. Firstly, the predicted positions of the four models for the high-sensitivity area are approximately the same, most of the historical landslide is contained in the extremely high-probability area predicted by the four models, the reliability of the predicted result is shown, and the landslide range of the Dongpu ditch is still mostly distributed near the main ditch in a comprehensive view, so that a large amount of material sources still can be fused into the ditch along with glacier water melting and precipitation, the risk of blocking the Yaruu Tibetan river is high, and importance should be drawn. Among the four different results, 1D-CNN has more abundant texture features, but has very high susceptibility to region distribution, and cannot highlight typical problems, and the other three models are obvious in distinguishing very low and very high susceptibility, which is related to the classification performance of CNN. In contrast, the network proposed in this embodiment is more specific to the characterization of landslide locations. The method of the embodiment shows that the area of the extremely high landslide easy area of the color Dongpu ditch is about 5.75km ² Accounting for 7.1 percent of the total area of the area.

In order to quantify the statistical result of landslide vulnerability evaluation, the embodiment compares the historical landslide distribution with the overall evaluation result, and objectively evaluates the result by the ratio of the area of the historical landslide to the area of each vulnerability class partition distributed in each vulnerability class partition. The results are shown in Table 2, wherein models a-d correspond to the results of (a) - (d) in FIG. 7, respectively, the very high incidence areas of the four models all comprise more than 90% of historical landslide, and the frequency ratio of each model is in an ascending trend along with the improvement of the incidence level, which indicates that the four models can effectively evaluate the incidence of landslide of the Gongdong ditch. The 1D-CNN result is divergent, and the number of the history landslide is large in low, medium and high susceptibility partitions; the shallow layer 2D-CNN misjudges a part of historical landslide as non-landslide due to the model differentiation characteristic; in contrast, the method provided by the embodiment has more reliable results, has the maximum frequency ratio in the extremely high-frequency area, accurately covers the 98.460% historical landslide range only in 7.513% of the area of the research area, and shows stronger characteristic learning capability.

TABLE 2

To further explore the reasons for the differences in model results, this embodiment makes further analysis of model mechanisms from the model training process. The related neural network experiments of the embodiment are all realized based on Python under a Tensorflow framework, and the experimental results are obtained on a host computer equipped with an Intel (R) Xeno (R) Silver 4214 processor and an NVIDIA Quadro P2200 display card. The experiment utilizes a moving window to extract 30720 landslide space neighborhood data from the landslide factor grid data of the color Dongpu ditch, and through data enhancement processing, the whole data set is constructed by randomly selecting equal quantity of non-landslide data, and the total quantity of the final data is 184320. To take into account the influence of the diversity of different data volumes and practical application situations, the whole data set is adopted with 6:2:2 and 8:1:1 two proportional forms divide training set, validation set and test set.

FIG. 8 shows the learning curves of each model verification set in the training process, 1D-CNN iterates 120 times, the rest of models iterate 64 times starting with 2D-CNN, the batch_size is uniformly sized to 128, the rising curve represents the training process accuracy ACC, the falling curve represents the training process loss function MSE, and the model learning capacity can be primarily evaluated through curve difference comparison. The verification set curves of the four models in the first proportion form are shown in (a) to (D) of fig. 8, gradual changes of the two curves of ACC and MSE reflect gradual improvement of learning conditions of the models, the four models are reversely propagated through errors of a gradient descent algorithm, weights among neurons are adjusted through errors between actual values and predicted values of training samples, therefore minimum values of target loss functions and optimal weights of the models are calculated gradually, the final steady curves express that the models form an orderly, stable and decision-making structure, in comparison of the four models, the learning ability of 1D-CNN is found to be worse, the highest accuracy rate is only about 0.932, the rest three models are higher than 0.980, the learning ability of shallow 2D-CNN is limited under limited data, the final ACC stays at 0.9825, MSE is 0.0152, compared with the deep 2D-CNN with the model parameter amount, the deep 2D-CNN is more advanced than the deep 2D-0120, the overall error of the ACC is found to be more than the MSE 2.0120, and the deep 2D-CNN is more advanced, and the error of the deep 2D-CNN is more advanced than the ACC is more slowly trained at the angle of the MSE 2.0120.0120. Notably, the network provided by the embodiment obtains better learning results under the condition that the reference quantity is less than that of deep 2D-CNN, and the landslide characteristics are proved to be more fully utilized under the multi-dimensional coupling structure.

In the second scale (fig. 8 (e) - (h)), the training set data size is increased, the model obtains relatively more learning references, and the overall performance of the four models is improved. 1D-CNN (ACC: 0.9544, MSE: 0.0370), shallow 2D-CNN (ACC: 0.9842, MSE: 0.0135), deep 2D-CNN (ACC: 0.9862, MSE: 0.0119), the proposed network of this embodiment (ACC: 0.9870, MSE: 0.0111), it can be seen that the proposed network of this embodiment still maintains the optimal level. The more parameter amount always accompanies the increase of training cost, and the time consumption of deep 2D-CNN is the greatest when processing more data, but the network proposed by this embodiment has a little different time from shallow 2D-CNN, and the result advantage is obvious. This proves that the overall calculation amount of the network model in the embodiment does not bring higher requirements to model training, and the model deepening is more feasible.

The test set data may be used to evaluate the ultimate generalization ability of the model. And respectively transmitting the test set data divided by the two different proportions to a trained model, and displaying the comparison condition of the predicted result and the true value through the confusion matrix. The confusion matrix is a common index for summarizing model classification prediction results in machine learning, and the model prediction results can be compared with historical real landslide by using the confusion matrix, so that evaluation of model learning results is formed. The confusion matrix form is shown in table 3.

TABLE 3 Table 3

Wherein TP, FN, FP, TN represents the number of model predictors divided into four cases by threshold, respectively. From these values, the following index can be calculated:

the overall Accuracy (Accuracy) represents the correct overall prediction ratio in the test set, and the calculation formula is as follows:precision, which represents the ratio of the actual positive sample predictions in the test set, is calculated as:Recall (Recall) represents the actual positive ratio in samples predicted to be positive, calculated as: The F1 score (F1-score) represents the harmonic mean of the precision and recall,the calculation formula is as follows:The Kappa coefficient represents the consistency of the predicted result and the whole actual value, and the calculation formula is as follows:Where n is the sum of the number of columns of the confusion matrix (total number of categories); x is X _ii For the ith row and the ith column of the confusion matrix, the number of samples is correctly classified; x is X _i+ 、X _+i The total number of samples in the ith row and the ith column are respectively; n is the total number of samples used for accuracy evaluation.

The working characteristics (receiver operating characteristic, ROC) curves of the subjects are the synthesis of confusion matrixes under different thresholds, are widely applied to landslide susceptibility evaluation result evaluation, the ROC takes the false positive rate as an abscissa, the true positive rate as an ordinate, continuous changes of data specificity and sensitivity are reflected, the area under the curve (area underthe curve, AUC) of the ROC can intuitively reflect the result, and the closer the AUC value is to 1, the better the model effect is indicated.

Since the model output results are probability distributions of 0 to 1, the prediction results are divided into landslide and non-landslide by using 0.5 as a threshold, the confusion matrix division results obtained by the four models are shown in fig. 9, and the index results are shown in table 4. In the confusion matrix, each model has higher true positive rate and true negative rate under two data sets, indicating reliable classification when four models face strange data. The model proposed in this embodiment has the best classification result under both scale test sets, further demonstrating model learning ability. Particularly, under a 10% test set, sufficient training data enables each model to fully learn landslide and non-landslide characteristics, and for the landslide characteristics subjected to data enhancement, two types of 2D-CNNs are the same as network classification results proposed by the embodiment, and show differences when facing complex and changeable non-landslide samples, and actually show that the method proposed by the embodiment keeps the best classification results, and proves the poor resistance of the method. In terms of evaluation indexes, all evaluation indexes calculated by the confusion matrix listed in table 4 are displayed according to each value in the table, and the method proposed in the embodiment is optimal. In the second scale situation, the deep 2D-CNN with larger parameter amounts is lower than the shallow 2D-CNN in each index, and it is inferred that the over-fitting problem caused by the increase of the model parameters occurs. The model excessively learns the training set data and lacks generalization capability for the data of the test set which is not contacted, so that the practical application value of the model is low. From the comprehensive view of model training and result evaluation, the multidimensional CNN coupling structure has good application value, and the problems of model efficiency reduction and model overfitting caused by parameter quantity increase are avoided under the condition of improving model precision.

TABLE 4 Table 4

The result shows that the calculated amount is reduced, the multidimensional CNN coupling structure provided by the embodiment from the angles of landslide feature depth excavation and feature full learning is equivalent to shallow layer 2D-CNN with fewer parameters, the length is greatly reduced when the deep layer 2D-CNN is trained compared with the deep layer 2D-CNN with approximate parameter amount, and the model training cost is reduced. In addition, compared with independent 1D-CNN and 2D-CNN feature learning capability, the coupling model has the advantages that model accuracy is improved, and the coupling model has higher scores under each confusion matrix index of test set data, so that landslide susceptibility evaluation results with higher reliability are obtained. Therefore, the multidimensional CNN coupling model provided by the embodiment is a reliable method suitable for landslide susceptibility evaluation, and provides new theoretical guidance and technical support for further landslide disaster monitoring and prevention.

In order to explore a reliable landslide susceptibility evaluation method, the embodiment couples a 2D-CNN structure and a 1D-CNN structure, utilizes asymmetric pooling and convolution kernel parameters to share and learn deep features among different dimensions of landslide and landslide influence factors, and aims to improve the precision of landslide prediction results and avoid the problem of difficult model training caused by deepening network layers. The experiment takes Tibet color Dongpu ditch as a research object, obtains the local landslide deformation range through multi-source data, and considers 11 landslide influence factors from various environmental factors. In order to verify the model effect, a classical 1D-CNN model, a shallow 2D-CNN model and a deep 2D-CNN model with equivalent parameter amounts are established for comparison, and the following conclusion is obtained through model training and evaluation indexes such as confusion matrix:

(1) In the two data proportion forms established in the embodiment, the proposed coupling model has the highest model precision and the lowest loss function result in the verification set data. When the proportion of the training set is increased, the accuracy of the verification set is increased, and meanwhile, the model efficiency is reduced, and the model provided by the embodiment is equivalent to the shallow structure efficiency while extracting deep features, so that the network provided by the embodiment has advantages in feature extraction and model efficiency.

(2) Under the data of the test set, the model provided by the embodiment has optimal performance under each index in the confusion matrix, and the coupling structure adopted is proved to have stronger generalization capability, so that the problem of model overfitting is avoided. In the landslide susceptibility map generated by each model after training, the network provided by the embodiment is most accurate in identifying the historical landslide, and the obtained landslide prediction result has the highest reliability. Therefore, under comprehensive evaluation, the coupling structure is considered as a feasible landslide susceptibility evaluation method, and a new attempt is made for deep learning application on landslide susceptibility evaluation.

The embodiment can also analyze landslide impact factors, and specifically comprises the following steps:

(1) Multiplex collinearity analysis

When exploring the regression relationship between landslide and landslide influence factors, the independence among the factors is ensured first. When there is a strong linear relationship between factors, it will be difficult to find the true relationship between the individual factors and the occurrence of landslide, resulting in a large deviation or even complete opposite of the model prediction result. The multiple collinearity is typically evaluated by using a variance expansion factor (variance inflation factor, VIF) and a tolerance (T), the VIF formula is as follows:

in the formula (4), A ² Representing complex correlation coefficient of regression analysis of the independent variable on other independent variables, when the correlation degree of the independent variable and the other independent variables is higher, A ² The value is closer to 1, the VIF value is larger, and when VIF is>10, its reciprocal tolerance (T)<0.1, it indicates that there is a multiple collinearity problem.

(2) Frequency ratio analysis

The Frequency Ratio (FR) can accurately process the nonlinear response relationship between the landslide and the influence factors thereof, and represents the relative influence degree of each attribute interval of the landslide factors on the occurrence of the landslide. Multiple mechanisms of local landslide generation can be quantitatively judged through frequency ratio analysis, so that the method is convenient for local conditions. The frequency ratio calculation is shown in formula (5).

In the formula (5), A is the landslide grid number of each type of environmental factors in the interval, A 'is the total number of landslide grids in the interval, B is the grid number of the environmental factors in the interval, B' is the total number of grids in the research area, and the higher the FR value is, the greater the influence of the attribute factors on landslide is indicated.

The factor analysis method is applied to the color east spectral ditch, and the analysis result is as follows: in the SPSS software, the landslide factor feature analysis result selected in the embodiment is obtained. As shown in FIG. 10, the multiple collinearity analysis results show that the distance from the channel has the maximum VIF value (4.482) and the minimum tolerance value (0.223), all factors are within an acceptable threshold range, and the 11 factors have good independence and no collinearity problem, and can be used for deep learning model learning. Fig. 11 shows the frequency of occurrence of landslide for each factor class, with the different classes exhibiting variability.

(1) Topography and topography. The results show that the color Dongpu ditch has the highest FR value when the altitude is 3500-4000m and the gradient is less than 20 degrees, which indicates that landslide mostly occurs at the downstream in the ditch area and the slope gradient is lower; the influence degree of precipitation and solar radiation on the slope-related areas shows correlation in most directions; curvature is defined as the three-dimensional characteristic of a two-dimensional surface, plane curvature and section curvature can effectively reflect terrain complexity, and the color Dongpu has highest FR values at 0-1 and-1-0 respectively; the topography humidity index reflects the condition of runoff, and has the highest FR value in the maximum value interval, so that the condition that the landslide of the color Dongpu ditch is relatively related to precipitation and water melting is indirectly indicated.

(2) Geological conditions. The lithology rock type has obvious influence on the slope soil type, the slope structure and the soil shear strength, and as the area is smaller, the area is only displayed as glaciers and custard marble, and FR values indicate that landslide develops more at glaciers; different fault distances influence the mechanical structure of the landslide body, geological faults exist in the northeast, southeast, southwest and northwest directions near the color Dongpu ditch, and the distances 7500-8500m have a larger influence on the landslide under the comprehensive action of a plurality of faults.

(3) Environmental conditions. Because the land is influenced by glacier water and atmospheric precipitation all the year round, the collapsed landslide often develops into a debris flow disaster, and peripheral landslide can be further induced by the impact scraping effect in a channel formed by the debris flow, the channel position of the color east Prain is drawn as a regional characteristic factor according to related literature and image interpretation, the FR value also shows strong correlation between the channel and the landslide, and the FR value has a maximum value at a distance of less than 300m from the nearest; the correlation with the earthquake is reflected in the distance from the earthquake center, the Milin county and the peripheral earthquake center positions are researched and selected, and landslide frequently occurs at the distance 9000-10000m under the comprehensive action; NDVI reflects that vegetation growth and coverage indirectly affect slope stability, and the lower vegetation coverage has the highest FR value in the interval of 0.2-0.4.

According to the embodiment, a 1D-CNN and 2D-CNN coupling model is constructed, the two network advantages are comprehensively utilized to learn different dimensionalities of landslide factors and deep association characteristics among different factors, meanwhile, the calculation amount of the model is reduced through asymmetric aggregation of feature graphs, further, the training cost of the model is reduced, and the model efficiency is guaranteed.

Example 2:

the embodiment is used for providing a multi-dimensional CNN coupled landslide susceptibility evaluation system, as shown in fig. 12, including:

the data acquisition module M1 is used for acquiring a factor distribution diagram of each landslide influence factor in a target area; the landslide impact factors comprise topography factors, geological condition factors and environmental condition factors;

the evaluation module M2 is used for determining landslide probability of each position point of the target area by using the trained multidimensional CNN coupling model by taking all the factor distribution graphs as input so as to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence; the two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer.

The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to assist in understanding the methods of the present invention and the core ideas thereof; also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In view of the foregoing, this description should not be construed as limiting the invention.

Claims

1. The multi-dimensional CNN coupled landslide vulnerability evaluation method is characterized by comprising the following steps of:

using all the factor distribution graphs as input, and determining landslide probability of each position point of the target area by using a trained multidimensional CNN coupling model so as to evaluate landslide susceptibility of the target area; the trained multidimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence; the two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer;

Before the landslide probability of each position point of the target area is determined by using the trained multidimensional CNN coupling model by taking all the factor distribution graphs as input, the landslide vulnerability evaluation method further comprises the steps of training the multidimensional CNN coupling model to obtain a trained multidimensional CNN coupling model, and specifically comprises the following steps:

acquiring a data set, and dividing the data set into a training set, a verification set and a test set according to a preset proportion; the data set comprises a plurality of groups of factor sample sets and labels corresponding to each group of factor sample sets; the factor sample set comprises a factor sample graph of each landslide impact factor; the label is landslide or non-landslide;

taking the training set as input, training the multidimensional CNN coupling model to obtain a trained model;

taking the verification set as input, and calculating a verification error by using the trained model;

judging whether an iteration termination condition is reached or not to obtain a judgment result; if the judgment result is yes, ending iteration, and taking the trained model as a trained multidimensional CNN coupling model; if the judgment result is negative, continuing iteration, and judging whether the verification error of the current iteration is the same as the verification error of the previous N iterations; if yes, the learning rate in the training process in the next iteration is adjusted, the trained model is used as a multidimensional CNN coupling model in the next iteration, and the step of training the multidimensional CNN coupling model by taking the training set as input is returned to obtain a trained model; if not, the step of taking the trained model as a multidimensional CNN coupling model of the next iteration, returning to the step of training the multidimensional CNN coupling model by taking the training set as input to obtain a trained model;

The acquiring the data set specifically comprises:

acquiring a factor distribution sample graph of each landslide influence factor in a target area and a landslide cataloging graph of the target area; the landslide cataloging graph marks the landslide position and the non-landslide position of the target area;

extracting a plurality of windows from the landslide cataloging graph by using a moving window method;

for each window, respectively extracting a region corresponding to the window position from each factor distribution sample graph to obtain a factor sample graph of each landslide impact factor, and forming a factor sample set corresponding to the window; taking the landslide condition of the central pixel of the window as a label corresponding to the window; and forming a data set by all the factor sample sets and the labels corresponding to the windows, wherein the number of the factor sample sets, of which the labels are landslide, in the data set is the same as the number of the factor sample sets, of which the labels are non-landslide.

2. The landslide vulnerability assessment method of claim 1, wherein the topography class factors comprise elevation, slope direction, plane curvature, section curvature and topography humidity index; the geological condition factors comprise lithology, fault distance and distance from the center of the earthquake; the environmental condition class factors include normalized vegetation index and distance from the channel.

3. The landslide susceptibility evaluation method of claim 1, further comprising, prior to acquiring the factor profile of each of the landslide impact factors in the target area: and calculating a variance expansion factor of each initial factor, and selecting the initial factor with the variance expansion factor smaller than a preset threshold as a landslide impact factor.

4. The landslide vulnerability evaluation method of claim 1, wherein the two-dimensional convolutional neural network comprises a plurality of two-dimensional convolutional pooling blocks and a two-dimensional convolutional layer which are sequentially connected, and the two-dimensional convolutional pooling blocks comprise a two-dimensional convolutional layer and a two-dimensional maximum pooling layer which are sequentially connected;

the one-dimensional convolutional neural network comprises a Reshape layer, a plurality of one-dimensional convolutional pooling blocks and a one-dimensional convolutional layer which are sequentially connected, wherein the one-dimensional convolutional pooling blocks comprise a one-dimensional convolutional layer and a one-dimensional maximum pooling layer which are sequentially connected;

the full-connection layer module comprises two full-connection layers which are sequentially connected;

wherein the two-dimensional convolution layer, the one-dimensional convolution layer and the full connection layer all adopt a ReLU function.

5. The landslide vulnerability assessment method of claim 1, further comprising, prior to extracting a plurality of windows in the landslide inventory map using a moving window method:

Performing histogram equalization on the factor distribution sample graph to obtain an equalized image; performing binarization processing on the landslide cataloging graph to obtain a binarized image; taking the equalization image and the binarization image as a new factor distribution sample graph and a landslide cataloging graph.

6. The landslide susceptibility evaluation method of claim 1, further comprising, after obtaining the data set: carrying out data enhancement on each group of factor sample sets, labeled as landslide, in the data set to obtain a plurality of groups of enhanced sample sets, expanding the data set to obtain an expanded data set, and taking the expanded data set as a new data set; the data enhancement includes a horizontal flip and a vertical flip.

7. The landslide vulnerability assessment method of claim 1, wherein when training a multidimensional CNN coupling model, a mean square error is used as a loss function, and an Adam optimizer is used to update network parameters of the multidimensional CNN coupling model.

8. A multi-dimensional CNN-coupled landslide vulnerability assessment system, comprising:

the evaluation module is used for determining landslide probability of each position point of the target area by using the trained multidimensional CNN coupling model by taking all the factor distribution graphs as input so as to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model comprises a two-dimensional convolutional neural network, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network, a one-dimensional asymmetric aggregation module and a full-connection layer module which are connected in sequence; the two-dimensional asymmetric aggregation module comprises a two-dimensional maximum pooling layer and a two-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional asymmetric aggregation module comprises a one-dimensional maximum pooling layer and a one-dimensional average pooling layer which are connected in parallel, and a concatate layer which is respectively connected with the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer;

Before the landslide probability of each position point of the target area is determined by using the trained multidimensional CNN coupling model by taking all the factor distribution graphs as input, the landslide susceptibility evaluation system further comprises the step of training the multidimensional CNN coupling model to obtain a trained multidimensional CNN coupling model, and the method specifically comprises the following steps of:

The acquiring the data set specifically comprises: