CN116205522B - A multi-dimensional CNN coupled landslide susceptibility evaluation method and system - Google Patents

Abstract

The invention discloses a multi-dimensional CNN coupled landslide susceptibility evaluation method and system, which relates to the technical field of landslide susceptibility evaluation. The trained multi-dimensional CNN coupling model includes sequentially connected two-dimensional convolutional neural networks, two-dimensional asymmetric aggregation module, one-dimensional convolutional neural network, one-dimensional asymmetric aggregation module and fully connected layer module to use asymmetric aggregation to couple the one-dimensional convolutional neural network and the two-dimensional convolutional neural network, maintain the network depth and limit the model parameters And reduce the amount of calculation. Subsequently, the factor distribution map of multiple landslide influencing factors in the target area is used as input, and the trained multi-dimensional CNN coupling model is used to evaluate the landslide susceptibility of the target area, which can improve the prediction accuracy while ensuring Model efficiency.

Description

A landslide susceptibility evaluation method and system based on multi-dimensional CNN coupling

Technical Field

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

Background Art

Landslide is one of the common natural disasters. When a landslide occurs, it not only destroys the natural environment, but also causes damage to villages, urban houses and infrastructure, and in severe cases, causes a large number of casualties. Landslide susceptibility assessment can explore the complex relationship between landslides and disaster-prone environments through the selection of models and data. It gives the probability of occurrence of unknown landslides in similar environments through comprehensive evaluation. It is an effective visualization technology for landslide regional positioning and sustainable management of land resources, and can provide strong technical support for decision-makers. The modeling process of quantitative landslide susceptibility assessment involves several important issues such as the acquisition of landslide and non-landslide records, the extraction of landslide-related environmental factors, and the selection of models. From a methodological perspective, it is still a challenge to use reliable and effective models to judge the level of landslide susceptibility.

At present, the landslide susceptibility assessment method based on statistical analysis has become the mainstream. The existing statistical methods can be divided into classical statistical methods and machine learning (ML) methods. Considering the complexity of the landslide mechanism, classical statistical methods such as frequency ratio method and comprehensive index method are often not accurate enough to summarize the relationship between landslide and its inducing factors in a linear form. Traditional ML models such as logistic regression (LR), support vector machine (SVM), random forest (RF) have been widely used in landslide susceptibility assessment, showing higher accuracy results than classical statistical methods. With the development of technology, deep learning (DL) models have achieved better results than traditional ML models in image classification and speech recognition. As a continuation of neural networks in ML, DL has a multi-hidden layer structure that allows it to fully mine the advanced features of the data set and add nonlinear representation to fit abstract features. As one of the most mature and popular DL frameworks, convolutional neural network (CNN) has also been widely used in the field of geoscience, such as landslide detection using remote sensing images and identification of seismic waveform features from seismic images. However, in terms of landslide susceptibility assessment, CNN is still in the development stage. CNN is a type of feedforward neural network with deep structure. According to the different dimensions of the accepted feature map or the definition of the convolution kernel, it can be divided into one-dimensional convolutional neural network (1D-CNN), two-dimensional convolutional neural network (2D-CNN) and three-dimensional convolutional neural network (3D-CNN). In the first landslide susceptibility assessment based on CNN, Wang et al. tried three architectures respectively. The results showed that the overall accuracy and Matthews correlation coefficient of the three architectures were significantly improved compared with the traditional ML method, showing the applicability of CNN in landslide susceptibility assessment.

Improving model accuracy is an important direction for CNN in the development of landslide susceptibility assessment. Deepening the network layer is the simplest operation to improve the accuracy of DL models. More layers can iteratively extract higher-level representations from low-level features. However, for CNN, deeper networks are more likely to produce gradient vanishing or explosion, making it difficult for the model to converge. The increase in the number of layers is also accompanied by an increase in the number of parameters, which leads to problems such as difficulty in model training or overfitting of results. Therefore, combining other models with CNN to improve model accuracy is considered more. This type of model combination method can usually obtain more comprehensive feature learning, but the increase in model size often leads to a significant increase in the number of model parameters and calculations, making model training and fitting difficult, and slow prediction and reasoning speed, resulting in a decrease in its overall efficiency in landslide emergency disaster management. Overall, CNN is still in the exploratory stage in the application of landslide susceptibility assessment, and there is still room for development in result presentation and accuracy improvement. 1D-CNN and 2D-CNN are widely used in landslide susceptibility assessment due to their powerful feature extraction capabilities. However, the current approach to improving their accuracy relies on layer deepening or model combination, which inevitably leads to a significant increase in the number of model parameters and the amount of calculation, resulting in difficulty in model training and overfitting of results, which in turn leads to low model efficiency.

Therefore, how to improve the accuracy of landslide susceptibility assessment while ensuring model efficiency is an urgent problem to be solved in practical applications.

Summary of the invention

The purpose of the present invention is to provide a multi-dimensional CNN coupled landslide susceptibility evaluation method and system, which can improve the prediction accuracy while ensuring the model efficiency.

To achieve the above object, the present invention provides the following solutions:

A multi-dimensional CNN coupled landslide susceptibility evaluation method, the landslide susceptibility evaluation method comprising:

Obtaining a factor distribution map of each of the multiple landslide influencing factors in the target area; the landslide influencing factors include topographic factors, geological conditions factors and environmental conditions factors;

Taking all the factor distribution maps as input, a trained multidimensional CNN coupling model is used to determine the landslide probability of each location point in the target area, so as to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model includes 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 fully connected layer module connected in sequence; the two-dimensional asymmetric aggregation module includes a two-dimensional maximum pooling layer and a two-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer respectively; the one-dimensional asymmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer respectively.

A multi-dimensional CNN coupled landslide susceptibility evaluation system, the landslide susceptibility evaluation system comprising:

A data acquisition module, used to obtain a factor distribution map of each of a plurality of landslide influencing factors in a target area; the landslide influencing factors include topographic factors, geological factors, and environmental factors;

An evaluation module is used to use all the factor distribution maps as inputs and use a trained multidimensional CNN coupling model to determine the landslide probability of each location point in the target area to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model includes 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 fully connected layer module connected in sequence; the two-dimensional asymmetric aggregation module includes a two-dimensional maximum pooling layer and a two-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer respectively; the one-dimensional asymmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer respectively.

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

The present invention is used to provide a multi-dimensional CNN coupled landslide susceptibility evaluation method and system. The 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 connected in sequence, so as to couple the one-dimensional convolutional neural network and the two-dimensional convolutional neural network by utilizing asymmetric aggregation, maintain the network depth while limiting the model parameters and reducing the amount of calculation, and subsequently, taking the factor distribution map of multiple landslide influencing factors in the target area as input, the trained multi-dimensional CNN coupling model is utilized to evaluate the landslide susceptibility of the target area, so as to improve the prediction accuracy while ensuring the model efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a method for evaluating landslide susceptibility provided by Example 1 of the present invention;

FIG2 is an overall flow chart of constructing a data set provided by Embodiment 1 of the present invention;

FIG3 is a schematic diagram of the network structure of a multi-dimensional CNN coupling model provided in Example 1 of the present invention;

FIG4 is a schematic diagram of the Sedongpugou watershed provided in Example 1 of the present invention;

FIG5 is a schematic diagram of the distribution of historical landslides in Sedongpu Valley provided in Example 1 of the present invention;

FIG6 is a sample diagram of factor distribution of the influencing factors of the Sedongpugou landslide provided in Example 1 of the present invention;

FIG7 is a landslide susceptibility map of Sedongpugou obtained by using different models provided in Example 1 of the present invention;

FIG8 is a schematic diagram showing a comparison of training results of different models provided in Example 1 of the present invention;

FIG9 is a schematic diagram of confusion matrix results of different model test sets provided in Example 1 of the present invention;

FIG10 is a schematic diagram of the multicollinearity analysis results provided in Example 1 of the present invention;

FIG11 is a schematic diagram of the frequency ratio of the attribute intervals of various landslide influencing factors provided in Example 1 of the present invention;

FIG. 12 is a system block diagram of a landslide susceptibility assessment system provided in Example 2 of the present invention.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

The purpose of the present invention is to provide a multi-dimensional CNN coupled landslide susceptibility evaluation method and system, which can improve the prediction accuracy while ensuring the model efficiency.

In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

Embodiment 1:

This embodiment is used to provide a multi-dimensional CNN coupled landslide susceptibility assessment method, as shown in FIG1 , the landslide susceptibility assessment method includes:

S1: Obtain a factor distribution map of each of a plurality of landslide influencing factors in a target area; the landslide influencing factors include topographic factors, geological factors, and environmental factors;

In this embodiment, topographic factors may include elevation, slope, aspect, plane curvature, profile curvature and topographic moisture index, geological conditions factors may include lithology, distance from fault and distance from epicenter, environmental conditions factors may include normalized vegetation index and distance from channel, and this embodiment uses the initial vector data of fault, epicenter and channel as input, and adopts the Euclidean distance calculation method in ArcGIS software to obtain the distance from fault, distance from epicenter and distance from channel. For each landslide impact factor, the value of the landslide impact factor at each position point in the target area constitutes the factor distribution map of the landslide impact factor in the target area.

S2: Taking all the factor distribution maps as input, using the trained multidimensional CNN coupling model to determine the landslide probability of each location point in the target area, so as to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model includes 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 fully connected layer module connected in sequence; the two-dimensional asymmetric aggregation module includes a two-dimensional maximum pooling layer and a two-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer respectively; the one-dimensional asymmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer respectively.

This embodiment uses a moving window method to divide all factor distribution maps separately, takes the window of the corresponding position of each factor distribution map as input, calculates the landslide probability of the position point located at the center of the window in the target area, and thus predicts the target area point by point, and finally obtains the landslide probability of each position point in the target area. If the landslide probability is greater than a preset threshold, the position point is determined to be a landslide point, otherwise, the position point is a non-landslide point.

Before S2, the landslide susceptibility assessment method of this embodiment further includes the step of training the multi-dimensional CNN coupling model to obtain a trained multi-dimensional CNN coupling model, which may include:

(1) Obtain a data set and divide the data set into a training set, a validation set, and a test set according to a preset ratio. The data set includes multiple groups of factor sample sets and labels corresponding to each group of factor sample sets. The factor sample set includes a factor sample map of each landslide influencing factor, and the label is landslide or non-landslide.

Accurate sample data is a prerequisite for model learning. In order to ensure information reliability and meet the needs of model training, this embodiment uses a window extraction method to construct a data set. The overall process is shown in Figure 2. Specifically, obtaining a data set may include:

1) Obtain a factor distribution sample map of each landslide influencing factor in the target area and a landslide inventory map of the target area. The landslide inventory map identifies the landslide location and non-landslide location of the target area.

The methods for obtaining a landslide catalog map include: using satellite remote sensing technology or an unmanned aerial vehicle platform to obtain digital elevation models of the target area before and after the landslide disaster, observing the digital elevation model before and after the landslide disaster, and taking the area with an elevation change of more than 10m before and after as the landslide area to determine the landslide location and non-landslide location in the target area, marking the landslide location and non-landslide location at various locations in the target area, and thus obtaining a landslide catalog map.

The method of obtaining a factor distribution sample map of each landslide impact factor in the target area includes: for each landslide impact factor, obtaining the value of the landslide impact factor at each location point in the target area before the landslide disaster occurs, and forming a factor distribution sample map of the landslide impact factor in the target area.

Preferably, the present embodiment may also preprocess the acquired factor distribution sample map and landslide catalog map, that is, the landslide susceptibility assessment method of the present embodiment also includes: in the preprocessing stage, performing histogram equalization on the factor distribution sample map to obtain a balanced image, performing binarization on the landslide catalog map to obtain a binarized image, and using the balanced image and the binarized image as new factor distribution sample map and landslide catalog map, executing 2).

Among them, histogram equalization refers to the removal of larger or smaller sparse values in the factor distribution sample map (which can be a raster map), that is, observing the overall distribution trend of the data in the factor distribution sample map, and removing the sparse values with a very small number without affecting the overall distribution law, so as to show the changes in the values with a large number, and normalizing the pixels of the factor distribution sample map after removing the sparse values, and then stretching the pixels of the normalized factor distribution sample map at a preset ratio to dynamically stretch the pixel difference to between 0 and 255. Different landslide influencing factors have the same preset ratio when stretching, that is, they are stretched at the same ratio. After histogram equalization, the pixel difference of the factor distribution sample map of each landslide influencing factor is dynamically stretched to between 0 and 255, so that it still has rich and obvious characteristics under the single-band bitmap, which is convenient for subsequent model reading and processing. Binarization processing refers to setting the pixel value of the landslide position in the landslide catalog map to 255 and the pixel value of the non-landslide position to 0, so as to process the landslide catalog map into a black and white image for display. After binarization processing, the landslide catalog map is presented as two types, landslide and non-landslide, for data screening, and the positive and negative samples required for model training are screened out. The positive samples are landslide samples, and the negative samples are non-landslide samples.

2) Multiple windows are extracted from the landslide catalog map using the moving window method. For each window, the area corresponding to the window position is extracted from each factor distribution sample map to obtain the factor sample map of each landslide influencing factor, which constitutes the factor sample set corresponding to the window; the landslide situation of the central pixel of the window is used as the label corresponding to the window; the factor sample sets and labels corresponding to all windows constitute a data set, and the number of factor sample sets with the label of landslide in the data set is the same as the number of factor sample sets with the label of non-landslide.

The traditional method of extracting factor attributes based on landslide points is heavily dependent on the accuracy of landslide positioning. Landslides are complex physical processes, and their occurrence is often closely related to the surrounding environment. Based on this, this embodiment uses a moving window to extract the landslide spatial neighborhood and construct a data set. The window side length is set, and the window can move one pixel step each time it moves in the landslide catalog map to determine multiple windows. This embodiment can obtain a 1058×963 landslide catalog map of the target area at a resolution of 10m. After testing, 13 pixels are finally set as the window side length. The results show that the image size and window size used can contain enough features and avoid data redundancy.

For each window, the area corresponding to the window position is extracted from each factor distribution sample map, that is, the area completely corresponding to the window position is extracted from the factor distribution sample map to obtain the factor sample map of each landslide influencing factor. The factor sample maps of all landslide influencing factors constitute the factor sample set corresponding to the window, and the landslide situation of the central pixel of the window is used as the label corresponding to the window, the label is landslide or non-landslide, and the association between the surrounding environment and the central pixel is determined. The factor sample set and label corresponding to a window constitute a sample. The sample with the label of landslide is a positive sample, and the sample with the label of non-landslide is a negative sample. The factor sample set and label corresponding to all windows constitute a data set.

Taking into account the problem that landslides occur less frequently in small areas and samples are insufficient, this embodiment introduces data enhancement processing, and expands the factor sample sets corresponding to the limited positive samples through data enhancement processing. That is, after obtaining the data set, the landslide susceptibility evaluation method of this embodiment also includes: performing data enhancement on each group of factor sample sets labeled as landslide in the data set to obtain multiple groups of enhanced sample sets, and the data enhancement includes horizontal flipping and vertical flipping to expand the data set to obtain an expanded data set. In addition to the data in the original data set, the expanded data set also includes multiple groups of enhanced sample sets, and the label of each group of enhanced sample sets is landslide, so as to expand the positive samples in the data set, and use the expanded data set as a new data set to perform subsequent model training process.

An increase in the ratio of the number of positive and negative samples will reduce the prediction accuracy of positive samples, which is contrary to the landslide warning goal. Therefore, this embodiment adopts a random balanced sampling strategy to set the number of non-landslide samples (i.e., negative samples), that is, multiple negative samples are cut out at random positions in the landslide catalog map with the same number as the positive samples, so that the number of positive samples and negative samples is the same, ensuring that the number of factor sample sets labeled as landslides in the data set is the same as the number of factor sample sets labeled as non-landslides. When there is data enhancement processing, in order to achieve this purpose, this embodiment can first determine multiple positive samples in the landslide catalog map, and 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 the same number of negative samples as the positive samples after expansion in the landslide catalog map, so as to construct a data set.

After obtaining the data set, the data set is shuffled and divided into training set, validation set and test set according to the preset ratio, which can be 6:2:2 or 8:1:1. The training set is used for the model to learn the landslide characteristics and train the internal parameters of the model; the validation set is used to test the model training status as a reference for optimizing the model hyperparameters; the test set is used to test the final generalization error of the model and evaluate the robustness of the model.

(2) Using the training set as input, the multi-dimensional CNN coupling model is trained to obtain the trained model.

This embodiment proposes to use 2D-CNN and 1D-CNN to couple to build a new landslide susceptibility assessment model (i.e., a multi-dimensional CNN coupling model), combining the advantages of the two convolutional neural networks to achieve automatic learning of the spatial neighborhood of landslide factors and related features between different factors. Under the coupling structure, the model maintains depth while avoiding a substantial increase in the number of parameters and the amount of calculation, thereby ensuring the model efficiency and improving the model prediction accuracy. The multi-dimensional CNN coupling model of this embodiment is mainly divided into five parts, including a two-dimensional convolutional neural network (2D-CNN) connected in sequence, a two-dimensional asymmetric aggregation module, a one-dimensional convolutional neural network (1D-CNN), a one-dimensional asymmetric aggregation module, and a fully connected layer module. The network structure is shown in Figure 3, where C is the number of feature channels.

The two-dimensional convolutional neural network includes a plurality of two-dimensional convolutional pooling blocks connected in sequence and a two-dimensional convolutional layer. The two-dimensional convolutional pooling block includes a two-dimensional convolutional layer and a two-dimensional maximum pooling layer connected in sequence. In Figure 3, there are two two-dimensional convolutional pooling blocks. The two-dimensional convolutional layer of the first two-dimensional convolutional pooling block uses a 2×2 convolution kernel, and the pooling window of the two-dimensional maximum pooling layer is 1×2; the two-dimensional convolutional layer of the second two-dimensional convolutional pooling block uses a 3×3 convolution kernel, and the pooling window of the two-dimensional maximum pooling layer is 1×2. The two-dimensional convolutional layer connected after multiple two-dimensional convolutional pooling blocks uses a 3×3 convolution kernel. The model accepts image inputs of the data set, and different types of landslide influencing factors are superimposed in the channel dimension to form a multi-spectral image of landslide factors with multiple features. This image contains the spatial neighborhood features of landslides or non-landslides, and shallow feature learning is performed through the two-dimensional convolutional layer and the two-dimensional maximum pooling layer. Using a 1×2 asymmetric pooling window, features are first aggregated in one dimension. Since the pooling window is defined to be small, only one of the two pixels that are representative is aggregated each time, thus reducing the disadvantage of data loss. As the feature map is reduced, the amount of model calculation will be reduced exponentially. As the network level deepens, more convolution kernels are used to extract deep features in the feature map, and the spatial characteristics composed of different dimensions are more fully learned.

The two-dimensional asymmetric aggregation module includes a two-dimensional maximum pooling layer and a two-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer respectively. In Figure 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 representative information of the aggregation dimension is obtained through the two-dimensional maximum pooling layer and the two-dimensional average pooling layer respectively. The landslide impact factor map features and the relationship features between factors are obtained in the aggregation dimension through convolution kernel operation. The two pooling methods obtain representative information in different ways. For example, the maximum pooling obtains the maximum pixel value in the pooling window, while the average pooling obtains the average pixel value in the pooling window. The role of the pooling layer is to reduce data redundancy, and the purpose of obtaining representative information in two ways is to weaken the impact of data loss. The two pooling results are integrated through the concatenate layer to reduce the loss of spatial semantic information caused by the pooling operation, while improving the model's tolerance for abnormal information values and avoiding model overfitting.

The one-dimensional convolutional neural network includes a Reshape layer, multiple one-dimensional convolutional pooling blocks and a one-dimensional convolutional layer connected in sequence. The one-dimensional convolutional pooling block includes a one-dimensional convolutional layer and a one-dimensional maximum pooling layer connected in sequence. In Figure 3, there are 2 one-dimensional convolutional pooling blocks. The convolution kernel size used in the one-dimensional convolutional layer of the first one-dimensional convolutional pooling block is 3, and the pooling window of the one-dimensional maximum pooling layer is 3; the convolution kernel size used in the one-dimensional convolutional layer of the second one-dimensional convolutional pooling block is 3, and the pooling window of the one-dimensional maximum pooling layer is 3. The convolution kernel size used in the one-dimensional convolutional layer connected after multiple one-dimensional convolutional pooling blocks is 3. 1D-CNN accepts two-dimensional input data in the form of feature length and feature channel number. The Reshape layer compresses the dimension of length 1 in the two-dimensional asymmetric aggregation module, reduces the dimension while keeping the number of features unchanged, so as to perform downscaling connection, and then learns the remaining dimensional features through the one-dimensional convolutional kernel with a window size of 3. Due to the parameter sharing characteristics of the convolutional kernel, the model not only learns the local features of a single landslide influencing factor, but also couples the feature associations between different factors. As the 1D-CNN layer deepens, the overall network has a deeper structure and can extract more abstract landslide and non-landslide features. The convolution kernel parameter sharing feature is an existing feature of CNN, which means that the weight of each convolution kernel is fixed during one iteration, and the weight within the convolution kernel will not change due to different positions in the image. Each convolution kernel slides in the image to learn individual features. Weight sharing reduces model parameters and greatly reduces the difficulty of network training.

The one-dimensional asymmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer, respectively. In Figure 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. After the average pooling and maximum pooling are aggregated, the overall features (including the abstract feature aggregation information represented by numbers) after being connected by the concatenate layer are passed to the fully connected layer module.

The fully connected layer module includes two fully connected layers connected in sequence. The fully connected layer maps the received feature vector after flattening, so that the neural nodes are interconnected and have global characteristics. Deepening the fully connected layer can improve the nonlinear expression ability of the model, but it will increase the model training cost and overfitting. Therefore, this model only sets two fully connected layers, and the last fully connected layer outputs a single neuron with an output value between 0 and 1, which indicates the probability of a medium landslide occurring at the point.

In this embodiment, by designing the structure of the above-mentioned multi-dimensional CNN coupling model, the asymmetric aggregation of the feature maps of 2D-CNN can halve the number of parameters, and by using 1D-CNN to learn the remaining dimensional data, the single-dimensional convolution kernel can also effectively reduce the number of parameters and the amount of calculation, thereby suppressing a significant increase in the number of parameters and the amount of calculation while ensuring the depth of the model, avoiding the problems of model training difficulties and overfitting of results, and ensuring the model efficiency while improving the prediction accuracy.

Model training mainly refers to the process of fully fitting features by iteratively updating the weights between model layers. These weights are regarded as parameters to be trained in the model. The number of model parameters determines the model scale, while the amount of calculation between parameters affects the model reasoning speed. Both are related to the structural design. The model parameters of CNN are mainly concentrated in the convolution layer and the fully connected layer, with the convolution layer as the primary part. The function of the convolution layer is realized by weighted aggregation of the convolution kernel on the feature map. The number of parameters is related to the size of the convolution kernel and the number of input and output feature map channels. The calculation formula is shown in the following formula (1). The 1D-CNN used in this model design only contains a single-dimensional convolution kernel, which can reduce the number of parameters to a certain extent, making deeper networks practical. The main calculation method of the convolution layer is the cumulative multiplication and addition operation between matrices. The amount of calculation can be measured by the number of floating point operations (FLOPs). FLOPs is related to the size and number of feature maps and the size of the convolution kernel. The calculation formula is shown in the following formula (2). Due to the asymmetric pooling effect, the size of the feature map decreases exponentially in the 2D-CNN part of this model, and the amount of calculation is also reduced accordingly. Therefore, the multi-dimensional CNN coupling structure of this embodiment reduces the amount of calculation while limiting the number of parameters, effectively avoiding the problem of training difficulties caused by the deepening of the number of neural network layers, and ensuring the efficiency of the model.

Param＝C _out ×(k ² ×C _in +1); (1)

In formula (1), Param represents the number of parameters required for a convolutional layer, k is the convolution kernel size, _Cin and _Cout represent the number of input and output feature map channels respectively.

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

In formula (2), H _out and W _out represent the length and width of the output feature map respectively.

This embodiment can also add Dropout in the structure of the multi-dimensional CNN coupling model.

In addition to the structural design of the model, some elements of the model need to be set. For example, the linear function mapping between layers in traditional neural networks is difficult to express the real relationship between complex variables such as landslides. Therefore, in the present embodiment, the rectified linear unit (rectified linear unit, ReLU) function is used in each part of the convolution layer and the fully connected layer to activate the inter-layer mapping, and the nonlinear relationship is used to effectively fit the output results and real data. At the same time, the ReLU function can also help overcome the gradient disappearance, which is more efficient in training, further shortens the network learning cycle, and reduces the amount of calculation. In the final output of the model, the sigmoid function single-value output is adopted, and the sigmoid function can map the output result to a distribution interval of 0 to 1 to represent the probability of landslide occurrence. That is, in the present embodiment, the two-dimensional convolution layer, the one-dimensional convolution layer and the fully connected layer all use the ReLU function.

The loss function is used to compare the difference between the model prediction result and the actual result, and then weigh the optimization direction of the model quality. In this embodiment, since the output result is in the form of landslide probability, the mean square error (MSE) is selected as the model loss function. MSE can simply measure the "average error" between the predicted value and the true value, and its calculation formula is shown in the following formula (3). During the training process, when the model obtains a result close to the fit through the sigmoid function, the curvature of the MSE function is relatively stable, which avoids the model from approaching the classification result, thereby avoiding overfitting of the model. Overall, the degree of change of MSE reflects the quality of the model. The smaller the value of MSE, the better the accuracy of the prediction model in describing the experimental data. From the perspective of network optimization, this embodiment uses Adam Optimizer instead of the traditional stochastic gradient descent method to more effectively update the network weights. That is, in this embodiment, when training the multidimensional CNN coupling model, the mean square error is used as the loss function, and the network parameters of the multidimensional CNN coupling model are updated using Adam optimizer.

In formula (3), m is the number of samples; _yi is the predicted value of the i-th sample; is the label of the i-th sample (i.e., the actual value).

This embodiment adds Dropout to the model structure, and utilizes the concatenated structure of maximum pooling and average pooling to avoid information loss to the greatest extent while also having good tolerance for abnormal information values. The joint application of MSE and the output layer sigmoid function enables a more obvious gradient suppression effect when the model is close to fitting, thus avoiding result differentiation and facilitating research.

(3) Take the validation set as input and use the trained model to calculate the validation error.

This embodiment also uses the MSE loss function when calculating the verification error.

(4) Determine whether the iteration termination condition is met and obtain the determination result; if the determination result is yes, terminate the iteration and use the trained model as the trained multi-dimensional CNN coupling model; if the determination result is no, continue the iteration and determine whether the verification error of the current iteration is the same as the verification error of the previous N iterations; if so, adjust the learning rate in the training process of the next iteration, use the trained model as the multi-dimensional CNN coupling model of the next iteration, and return to the step of "using the training set as input to train the multi-dimensional CNN coupling model to obtain the trained model"; if not, use the trained model as the multi-dimensional CNN coupling model of the next iteration and return to the step of "using the training set as input to train the multi-dimensional CNN coupling model to obtain the trained model".

The iteration termination condition may be reaching the maximum number of iterations. In this embodiment, the learning rate in the training process is dynamically adjusted through a callback function, and the initial learning rate is set to 0.001. During the iteration process, as the verification error is observed, when every three iterations with equal performance appear, that is, when the verification error of the current iteration is the same as the verification error of the previous two iterations, the learning rate decays at a rate of 0.5.

Convolutional neural networks are widely used in landslide susceptibility assessment due to their powerful feature extraction capabilities. However, with the diversification of scenarios and the demand for high precision, the CNN algorithm is constantly improving. The practice of improving accuracy by deepening the network layer or combining other models often greatly increases the number of model parameters and the amount of calculation, resulting in difficulty in model training or overfitting of the results, and thus the model efficiency cannot be guaranteed, limiting its practical application. This embodiment proposes to construct a multidimensional CNN coupling model to solve the above problems. The one-dimensional convolutional neural network and the two-dimensional convolutional neural network are connected by asymmetric aggregation of feature graphs, maintaining the network depth while limiting the model parameters and reducing the amount of calculation, thereby improving the accuracy while ensuring the model efficiency. The multidimensional convolution kernel parameter sharing can also be further used to capture the deep coupling characteristics between the different dimensions of each landslide influencing factor and the factors, so as to fully utilize the features and avoid overfitting.

Preferably, before obtaining the factor distribution map of each landslide influencing factor in the target area among multiple landslide influencing factors, the landslide susceptibility evaluation method of this embodiment also includes: calculating the variance inflation factor of each initial factor, and selecting the initial factor with a variance inflation factor less than a preset threshold as the landslide influencing factor, thereby ensuring the independence of the landslide influencing factor.

Below, this embodiment takes Sedongpugou in Tibet as the experimental area, selects 11 landslide influencing factors to analyze the landslide susceptibility of the area, comprehensively considers the model robustness, reasoning speed and result generalization ability, and makes a new attempt for the application of CNN in landslide susceptibility evaluation, which is of great significance to regional engineering planning and disaster prevention and mitigation.

As shown in Figure 4, it is the drainage area of Sedongpugou. Sedongpugou is located in Milin County, Nyingchi City, Tibet Autonomous Region, China. It is located in the core area of Yarlung Zangbo Grand Canyon. The mouth of the gully connects to the Yarlung Zangbo River. The gully area is about ^67.15km2 and the main gully is 7.4km long. The upper reaches of the northern part of Sedongpugou have wide terrain, steep terrain, and developed glaciers, including the highest point in the area, Jiala Bailei Peak (7294m). The downstream channel is narrower and the slope is reduced, including the lowest point, the gully mouth (2746m). The overall height difference is 4548m, which belongs to a typical high mountain canyon area. Affected by the tectonic stress of the Qinghai-Tibet Plateau and the Himalayas, the rock mass in the area is severely broken and has low shear strength. The complex terrain and strong movement have caused a large number of landslide disasters in the area. At the same time, due to the "funnel effect" in the region, the landslide body interacts with a large amount of glacial meltwater and precipitation to easily become a debris flow, rushing out along the ditch and blocking the Yarlung Zangbo River, forming a barrier dam. When the dam body collapses, it will eventually trigger a high-level mountain torrent. At present, the disaster chain in Sedongpugou has caused thousands of people to be affected, dozens of villages and towns have been destroyed, and local transportation and logistics, engineering construction, etc. have been seriously affected. With global climate change and the evolution of geological structures, Sedongpugou still has the risk of a high-level collapse and landslide causing a large-scale geological disaster chain. Existing studies have analyzed and explored Sedongpugou from the perspectives of formation conditions, occurrence process and destruction mechanism, but there is a lack of specific indications for the starting point of the future disaster chain, that is, the location of future landslides. Therefore, it is necessary to carry out a landslide susceptibility assessment in the area, and it is recommended to carry out long-term monitoring and early protection in extremely high-susceptibility areas to curb the occurrence of the Sedongpugou disaster chain from the root.

Due to the steep terrain and sparse population in Sedongpugou, traditional geological survey methods are difficult to implement in the area, and the specific location, scale and volume of the landslide cannot be determined. Therefore, this embodiment uses satellite remote sensing technology, drone platform, DEM difference and visual interpretation methods to complete the screening of historical landslide ranges. In October 2018, there were two incidents of debris flows caused by landslide collapse that blocked the Yarlung Zangbo River. Therefore, this embodiment focuses on the research period from 2017 to 2020 before and after the disaster. The development characteristics of the landslide in this period are used to explore the risk areas of landslides in the future. The data used are the December 2017 Resource-3 satellite images provided by the Land Satellite Remote Sensing Application Center of the Ministry of Natural Resources and the Chinese Academy of Surveying and Mapping, the ZC-3 drone collection results in August 2019, and the Resource-3 satellite images in December 2020. Ziyuan-3 is China's first civilian high-resolution stereo mapping satellite. Through stereo image pairs, a digital elevation model (DEM) with a spatial resolution of 10m and an elevation error of 5m can be constructed, including the study area. The elevation accuracy meets the national 1:10,000 mountain and high mountain DEM accuracy requirements. The DEM of the study area was extracted from it. The resolution of the drone aerial photography image was designed to be 0.1m, and the collection area was the main gully area with an area of about ^12km2 . The mapping results met the national 1:2000 topographic map accuracy requirements. Based on the Ziyuan-3 data, the DEM constructed by the drone collection was resampled to the same spatial resolution and georeferenced through the feature point matching algorithm for subsequent calculations. The Sedongpugou landslide is mainly manifested as shear-slip deformation caused by the loose body of the trailing edge channel. Due to the large catchment area and poor rock mass integrity, the landslide was overhead collapsed into debris flow at the moment of "starting". In order to fully identify the debris flow source area of Sedongpugou, referring to the general definition of landslide, this embodiment regards the downward and outward movement of the materials constituting the slope as landslides, which is specifically reflected in the changes in DEM in different time periods. However, due to the influence of shooting angle and geometric distortion, there is an error between the constructed DEM and the true value. Therefore, according to the data processing results and the geomorphological characteristics of the study area, the error limit in the high mountain elevation is used as the basis for division, that is, the area with an elevation change of more than 10m before and after is regarded as the landslide range. The difference of the DEM constructed from two phases of satellite images has obtained most of the landslides. In order to obtain more accurate results near the main ditch, the difference results of the drone DEM covering the main ditch area and with higher accuracy are added as a supplement. The deformation ranges in three time periods over three years were aggregated, and then the landslide edges were smoothed in combination with visual interpretation of images to conform to the actual development characteristics. Finally, the distribution of historical landslides in Sedongpugou was mapped. The results are shown in Figure 5. Figure 5(a) shows the landslide range determined based on the data of 2017 and 2019, Figure 5(b) shows the landslide range determined based on the data of 2019 and 2020, Figure 5(c) shows the landslide range determined based on the data of 2017 and 2020, and Figure 5(d) shows the landslide range determined based on the data of 2017, 2019 and 2020. According to statistics, a total of 70 landslides were identified in the study area, with the largest complete landslide area of about ^1.78km2 and the smallest area of ^3971m2 . Most of the landslides are distributed near the middle channel, showing aggregation. Based on this, a landslide catalog map of Sedongpugou can be constructed.

Landslide susceptibility assessment is based on the assumption that future landslides are in the same environment as previous landslides. Therefore, the selection of appropriate landslide influencing factors is also an important part of the successful landslide susceptibility modeling. Since the study area is small and the amount of information is small, this embodiment reflects the landslide characteristics from as many angles as possible, and selects 11 factors including topography (elevation, slope, slope aspect, plane curvature, profile curvature, terrain moisture index), geological conditions (lithology, distance from fault, distance from epicenter), and environmental conditions (normalized vegetation index, distance from channel) to reflect landslide characteristics. In order to ensure detailed information within the regional scope, all landslide factors use a resolution of 10m×10m, and the display results are shown in Figure 6. The data sources are as follows: elevation is generated by the stereo image pair of Resource-3 remote sensing images in 2017, and other terrain-related factors are derived from DEM; lithology and faults are intercepted from the national 1:250,000 geological map; channel vectors are drawn based on optical remote sensing images and historical research, and related distances are calculated using Euclidean distances; earthquake point data are publicly provided by the China Earthquake Networks Center; and NDVI is publicly provided by the National Ecological Data Center Resource Sharing Service Platform. Based on this, the factor distribution sample map of each landslide influencing factor can be obtained as shown in Figure 6.

In order to verify the accuracy and operation effect of the model, this embodiment is quantitatively evaluated through comparative experiments. In this embodiment, a total of four independent experiments were conducted, including the proposed multi-dimensional CNN coupling model, 1D-CNN, shallow 2D-CNN and deep 2D-CNN with the same order of magnitude of parameters as the multi-dimensional CNN coupling model. 1D-CNN and shallow 2D-CNN are used to compare the differences in feature learning between the existing methods and the coupling method proposed in this embodiment, and deep 2D-CNN is used to verify the efficiency of the network proposed in this embodiment when the model scale is roughly the same. The specific parameters of the four models are shown in Table 1.

Table 1

The four model structures are set according to their own characteristics while maintaining variable control. The same activation function, loss function, optimizer and learning rate are used in the four structures. In the process of network training, in order to reduce overfitting, the Dropout layer is introduced in each network. The Dropout operation plays an important role in improving the prediction performance. It can temporarily discard the neural network units according to a specific probability during the training process to reduce the co-adaptation between hidden units and enhance the generalization ability of the prediction method.

The landslide susceptibility map of the entire study area can be made by reassigning values to each pixel in the study area using the trained model. The landslide influencing factors in the study area are input into each model for prediction. After the prediction is completed, ArcGIS is used to divide the obtained landslide probability into five types of landslide susceptibility levels using the natural breakpoint method: extremely low, low, medium, high, and extremely high. Figure 7 shows the landslide susceptibility map of Sedongpugou obtained using different models. There are differences between the results of the four models, but there are still certain commonalities. First, the four models predict roughly the same location for highly sensitive areas, and most of the historical landslides are included in the extremely high susceptibility areas predicted by the four models, indicating the reliability of the prediction results. Overall, the landslide range of Sedongpugou is still mostly distributed near the main ditch, which indicates that with the action of glacial meltwater and precipitation, a large amount of material will still be integrated into the ditch, and the risk of blocking the Yarlung Zangbo River is relatively high, which should be taken seriously. Among the four different results, 1D-CNN has richer texture features, but the extremely high-prone area is widely distributed and cannot highlight typical problems. The other three models are obviously differentiated into extremely low and extremely high-prone areas, which is related to the classification performance of CNN. In contrast, the network proposed in this embodiment is more specific in the description of landslide locations. After calculation, the method of this embodiment shows that the extremely high landslide prone area in Sedongpugou is about ^5.75km2 , accounting for 7.1% of the total area of the region.

In order to quantify the statistical results of landslide susceptibility evaluation, this embodiment compares the distribution of historical landslides with the overall evaluation results, and objectively evaluates the results by the ratio of the historical landslide area distributed in each susceptibility level partition to the area of each susceptibility level partition. The results are shown in Table 2, where model a-model d correspond to the results of (a)-(d) in Figure 7, respectively. The extremely high susceptibility areas of the four models all contain more than 90% of the historical landslides, and the frequency ratio of each model shows an upward trend with the increase of the susceptibility level, indicating that the four models can effectively evaluate the susceptibility of Sedongpugou landslide. The results of 1D-CNN are relatively divergent, and a large number of historical landslides are distributed in low, medium and high susceptibility zones. Due to the differentiation characteristics of the model, the shallow 2D-CNN misclassifies some historical landslides as non-landslides. In comparison, the results of the method proposed in this embodiment are more reliable, with the largest frequency ratio in extremely high susceptibility zones, accurately covering 98.460% of the historical landslide range in only 7.513% of the study area, showing a strong feature learning ability.

Table 2

In order to further explore the reasons for the differences in model results, this embodiment further analyzes the model mechanism from the model training process. The relevant neural network experiments in this embodiment are all implemented based on Python under the Tensorflow framework, and the experimental results are obtained on a host equipped with an Intel (R) Xeno (R) Silver 4214 processor and an NVIDIA Quadro P2200 graphics card. The experiment uses a moving window to extract 30,720 landslide spatial neighborhood data from the Sedongpugou landslide factor raster data, and through data enhancement processing, randomly selects an equal amount of non-landslide data to construct an overall data set, and the final total data volume is 184,320. In order to consider the impact of different data volumes and the diversity of actual application situations, the overall data set is divided into training set, validation set and test set in two ratios of 6:2:2 and 8:1:1.

Figure 8 shows the validation set learning curves of each model during the training process. 1D-CNN iterates 120 times, and the other models starting with 2D-CNN iterate 64 times. The batch_size is uniformly 128. The rising curve represents the accuracy ACC of the training process, and the falling curve represents the loss function MSE of the training process. The learning ability of the model can be preliminarily evaluated by comparing the curve differences. The validation set curves of the four models in the first proportional form are shown in (a) to (d) of Figure 8. The gradual changes in the ACC and MSE curves reflect the gradual improvement of the model's learning. The four models use the error back propagation of the gradient descent algorithm to adjust the weights between neurons based on the error between the actual value and the predicted value of the training sample, thereby gradually calculating the minimum value of the target loss function and the optimal weight of the model. The final stable curve expresses that the model has formed an orderly, stable, and decision-making structure. In the comparison of the four models, it is found that 1D-CNN has poor learning ability, and the highest accuracy can only stay at around 0.932, while the other three models are all higher than 0.98 0, the shallow 2D-CNN has limited learning ability under limited data due to the small number of trainable parameters, so the final ACC stays at 0.9825 and the MSE is 0.0152. In comparison, the deep 2D-CNN with a large number of model parameters and the network proposed in this embodiment have stronger performance. The deep 2D-CNN (ACC: 0.9850, MSE: 0.0129) is slightly worse than the network proposed in this embodiment (ACC: 0.9857, MSE: 0.0127). From the perspective of model efficiency, the 1D-CNN with a small number of parameters is the fastest to train, followed by the shallow 2D-CNN, while the deep 2D-CNN with a large number of parameters and the network proposed in this embodiment are overall slow. It is worth noting that the network proposed in this embodiment obtains better learning results when the number of parameters is less than that of the deep 2D-CNN, proving that the landslide feature is more fully utilized in the multi-dimensional coupling structure.

In the second proportion form (Figure 8 (e) ~ (h)), the amount of training set data increases, 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 network proposed in this embodiment (ACC: 0.9870, MSE: 0.0111), it can be seen that the network proposed in this embodiment still maintains the optimal level. A large number of parameters is always accompanied by an increase in training costs. When processing more data, the deep 2D-CNN takes the most time, while the network proposed in this embodiment is similar to the shallow 2D-CNN in time, and the result is significantly superior. This proves that the overall computational amount of the network model proposed in this embodiment does not bring higher requirements to model training, making it more feasible to deepen the model.

The test set data can be used to evaluate the final generalization ability of the model. The test set data divided into two different proportions are passed to the trained model respectively, and the comparison between the predicted results and the true values is displayed through the confusion matrix. The confusion matrix is a common indicator for summarizing the classification prediction results of the model in machine learning. The confusion matrix can be used to compare the model prediction results with the historical real landslides, thereby forming an evaluation of the model learning results. The confusion matrix is shown in Table 3.

Table 3

Among them, TP, FN, FP, and TN represent the number of model prediction results in four situations according to the threshold. These values can be used to calculate the following indicators:

The overall accuracy (Accuracy) indicates the correct ratio of the overall prediction in the test set, and its calculation formula is: Precision refers to the ratio of samples in the test set that are actually positive to samples that are predicted to be positive. The calculation formula is: Recall represents the ratio of samples predicted to be positive to samples that are actually positive. The calculation formula is: The F1 score represents the harmonic mean of precision and recall, and its calculation formula is: The Kappa coefficient indicates the overall consistency between the predicted results and the actual values, and its calculation formula is: Among them, n is the sum of the number of columns in the confusion matrix (total number of categories); _Xii is the number of samples in the i-th row and i-th column in the confusion matrix, that is, the number of correctly classified samples; Xi ₊ and X _+i are the total number of samples in the i-th row and i-th column respectively; N is the total number of samples used for accuracy evaluation.

The receiver operating characteristic (ROC) curve is a synthesis of the confusion matrix under different thresholds and is widely used in the evaluation of landslide susceptibility assessment results. ROC uses the false positive rate as the horizontal axis and the true positive rate as the vertical axis to reflect the continuous changes in data specificity and sensitivity. The area under the ROC curve (AUC) can intuitively reflect the results. The closer the AUC value is to 1, the better the model effect.

Since the model output result is a probability distribution of 0 to 1, the prediction result is divided into landslide and non-landslide with a threshold of 0.5. The confusion matrix division results of the four models are shown in Figure 9, and the results of each index are shown in Table 4. In the confusion matrix, each model has a high true positive rate and true negative rate under the two data sets, indicating that the four models can reliably classify unfamiliar data. The model proposed in this embodiment has the best classification results under the two proportion test sets, which further proves the model learning ability. In particular, under the 10% test set, sufficient training data has allowed each model to fully learn the landslide and non-landslide features. For the landslide features after data enhancement, the two types of 2D-CNN and the network classification results proposed in this embodiment are the same, but when facing complex and changeable non-landslide samples, they show differences. It actually shows that the method proposed in this embodiment maintains the best division results and proves the anti-difference of the method. From the perspective of evaluation indicators, all evaluation indicators calculated by the confusion matrix listed in Table 4 are shown in the table. According to the values in the table, the method proposed in this embodiment is the best. This embodiment notes that in the second ratio situation, the deep 2D-CNN with more parameters is lower than the shallow 2D-CNN in all indicators. It is inferred that the overfitting problem should be caused by the increase in model parameters. The model has over-learned the training set data and lacks the ability to generalize to the test set data that has never been exposed. The practical application value of such models is low. From the comprehensive perspective of model training and result evaluation, the multi-dimensional CNN coupling structure avoids the problems of reduced model efficiency and model overfitting due to the increase in parameters while improving the model accuracy, and has good application value.

Table 4

The results show that due to the reduced amount of calculation, the multi-dimensional CNN coupling structure proposed in this embodiment from the perspective of deep mining of landslide features and full feature learning is as efficient as the shallow 2D-CNN with fewer parameters, and the training time is greatly reduced compared to the deep 2D-CNN with similar parameters, and the model training cost is reduced. In addition, compared with independent 1D-CNN and 2D-CNN, the coupled model has enhanced feature learning capabilities and improved model accuracy, and has higher scores under various confusion matrix indicators of the test set data, thereby obtaining a landslide susceptibility evaluation result with higher credibility. Therefore, the multi-dimensional CNN coupling model proposed in this embodiment is a reliable method for landslide susceptibility evaluation, which provides new theoretical guidance and technical support for further landslide disaster monitoring and prevention.

In order to explore a reliable method for evaluating landslide susceptibility, this embodiment couples the 2D-CNN with the 1D-CNN structure, and uses asymmetric pooling and convolution kernel parameter sharing to learn the deep features of different dimensions of landslides and landslide influencing factors, aiming to improve the accuracy of landslide prediction results while avoiding the problem of model training difficulties caused by deepening the network layer. The experiment took Sedongpugou in Tibet as the research object, obtained the local landslide deformation range through multi-source data, and considered 11 landslide influencing factors from a variety of environmental factors. In order to verify the effect of the model, the classic 1D-CNN, shallow 2D-CNN, and deep 2D-CNN models with equivalent parameters were set up for comparison. The following conclusions were drawn through model training and evaluation indicators such as confusion matrix:

(1) Under the two data ratios established in this embodiment, the proposed coupling model has the highest model accuracy and the lowest loss function results in the validation set data. When the training set ratio increases, the validation set accuracy increases while the model efficiency decreases. However, the model proposed in this embodiment extracts deep features while still being as efficient as the shallow structure, proving that the proposed network has advantages in feature extraction and model efficiency.

(2) Under the test set data, the model proposed in this embodiment has the best performance under each indicator in the confusion matrix, which proves that the coupling structure adopted has strong generalization ability and avoids the problem of model overfitting. In the landslide susceptibility map generated by each model after training, the network proposed in this embodiment is the most accurate in identifying historical landslides, and the obtained landslide prediction results have the highest reliability. Therefore, under comprehensive evaluation, it is believed that the coupling structure is a feasible landslide susceptibility evaluation method, which makes a new attempt for the application of deep learning in landslide susceptibility evaluation.

This embodiment can also analyze the landslide influencing factors, as follows:

(1) Multicollinearity analysis

When exploring the regression relationship between landslides and landslide influencing factors, the independence of factors must be ensured first. When there is a strong linear relationship between factors, it will be difficult to find the true relationship between a single factor and the occurrence of landslides, resulting in a large deviation in the model prediction results or even a complete opposite. This phenomenon is called multicollinearity between factors in this embodiment. Multicollinearity is usually judged by variance inflation factor (VIF) and tolerance (T). The VIF formula is as follows:

In formula (4), ^A2 represents the multiple correlation coefficient of the independent variable for the other independent variables in the regression analysis. When the correlation between the independent variable and the other variables is higher, the ^A2 value is closer to 1 and the VIF value is larger. When VIF>10 and its reciprocal tolerance (T)<0.1, it indicates the existence of multicollinearity problem.

(2) Frequency ratio analysis

The frequency ratio (FR) can accurately handle the nonlinear response relationship between landslides and their influencing factors, and characterize the relative influence of each attribute interval of landslide factors on the occurrence of landslides. Through frequency ratio analysis, the multiple mechanisms of local landslide occurrence can be quantitatively determined, which is convenient for taking appropriate measures. The frequency ratio calculation is shown in formula (5).

In formula (5), A is the number of landslide grids that appear in each type of environmental factor in the interval, A' is the total number of landslide grids in the area, B is the number of grids of environmental factors in the interval, and B' is the total number of grids in the study area. The higher the FR value, the greater the impact of this type of attribute factor on the landslide.

The above factor analysis method was applied to Sedongpugou, and the analysis results are as follows: In SPSS software, the characteristic analysis results of the landslide factors selected in this embodiment were obtained. The multicollinearity analysis results are shown in Figure 10. The distance from the ditch has the largest VIF value (4.482) and the smallest tolerance value (0.223). All factors are within the acceptable threshold range, indicating that the 11 factors have good independence and no collinearity problems, which can be used for deep learning model learning. Figure 11 shows the frequency ratio of landslide occurrence in each factor category, and different categories show differences.

(1) Topography. The results show that Sedongpugou has the highest FR value at an altitude of 3500-4000m and a slope of less than 20°, indicating that landslides mostly occur in the middle and lower reaches of the gully area and at the lower slopes. The slope direction is related to the degree of influence of precipitation and solar radiation on the region, and it shows correlation in most directions. Curvature is defined as the three-dimensional feature of a two-dimensional surface. Plane curvature and profile curvature can effectively reflect the complexity of the terrain. Sedongpugou has the highest FR value at 0-1 and -1-0, respectively. The terrain humidity index reflects the runoff condition and has the highest FR value in the maximum value range, which indirectly indicates that the Sedongpugou landslide is more related to precipitation and meltwater.

(2) Geological conditions. Lithology and rock type have a significant impact on the slope soil type, slope structure and soil shear strength. Due to the small size of this area, only two types of rocks are shown: glacier and Nanga Parbat Group marble. The FR value shows that landslides are more common in glaciers. The distance from the fault affects the mechanical structure of the landslide. There are geological faults in the northeast, southeast, southwest and northwest directions near Sedongpugou. Under the combined effect of multiple faults, the distance of 7500-8500m has a greater impact on the landslide.

(3) Environmental conditions. As this area is affected by glacial meltwater and atmospheric precipitation all year round, collapse landslides often develop into debris flow disasters. In the channels formed by debris flows, the impact and scraping effect can further induce surrounding landslides. Therefore, according to relevant literature and image interpretation, the channel location of Sedongpugou was mapped as a regional characteristic factor. The FR value also shows a strong correlation between the channel and the landslide, with the maximum value at the closest distance of less than 300m. The correlation with earthquakes is reflected in the distance from the epicenter. The study selected the epicenter location of Milin County and its surrounding areas. Under the comprehensive effect, landslides are more frequent at a distance of 9000-10000m. NDVI reflects that the growth and coverage of vegetation indirectly affect the stability of the slope. The 0.2-0.4 interval with a low vegetation coverage has the highest FR value.

This embodiment proposes to construct a 1D-CNN and 2D-CNN coupling model, comprehensively utilizing the advantages of the two networks to learn the different dimensions of landslide factors and the deep correlation features between different factors, and at the same time reduce the model calculation amount through asymmetric aggregation of feature maps, thereby reducing the model training cost and ensuring model efficiency.

Embodiment 2:

This embodiment is used to provide a multi-dimensional CNN coupled landslide susceptibility evaluation system, as shown in FIG12 , the landslide susceptibility evaluation system includes:

The data acquisition module M1 is used to obtain a factor distribution diagram of each of the multiple landslide influencing factors in the target area; the landslide influencing factors include topographic factors, geological conditions factors and environmental conditions factors;

The evaluation module M2 is used to use all the factor distribution maps as inputs and use the trained multidimensional CNN coupling model to determine the landslide probability of each location point in the target area to evaluate the landslide susceptibility of the target area; the trained multidimensional CNN coupling model includes 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 fully connected layer module connected in sequence; the two-dimensional asymmetric aggregation module includes a two-dimensional maximum pooling layer and a two-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer respectively; the one-dimensional asymmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel, and a concatenate layer connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer respectively.

This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims (8)

1. A multi-dimensional CNN coupled landslide susceptibility evaluation method, characterized in that the landslide susceptibility evaluation method includes: Obtain the factor distribution map of each of the multiple landslide influencing factors in the target area; the landslide influencing factors include topographic factors, geological condition factors and environmental condition factors; Taking all the factor distribution maps as input, the trained multi-dimensional CNN coupling model is used to determine the landslide probability of each location point in the target area to evaluate the landslide susceptibility of the target area; the trained multi-dimensional The CNN coupling model includes 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 fully connected layer module connected in sequence; the two-dimensional asymmetric aggregation module includes a parallel The connected two-dimensional maximum pooling layer and the two-dimensional average pooling layer and the concatenate layer respectively connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer; the one-dimensional non- The symmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel and a concatenate layer connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer respectively; Before using all the factor distribution maps as input and using the trained multi-dimensional CNN coupling model to determine the landslide probability of each location point in the target area, the landslide susceptibility evaluation method also includes training the multi-dimensional CNN coupling model. , get the trained multi-dimensional CNN coupling model, including: Obtain a data set, and divide the data set into a training set, a verification set and a test set according to a preset ratio; the data set includes multiple groups of factor sample sets and labels corresponding to each group of the factor sample sets; The factor sample set includes a factor sample map of each landslide influencing factor; the label is landslide or non-landslide; Taking the training set as input, the multi-dimensional CNN coupling model is trained to obtain the trained model; Taking the verification set as input, using the trained model to calculate the verification error; Determine whether the iteration termination condition is reached and obtain the judgment result; if the judgment result is yes, then end the iteration and use the trained model as the trained multi-dimensional CNN coupling model; if the judgment result is no, continue the iteration, Determine whether the verification error of the current iteration is the same as the verification error of the previous N iterations; if so, adjust the learning rate during the training process in the next iteration, use the trained model as the multi-dimensional CNN coupling model of the next iteration, and return The step of "taking the training set as input, training the multi-dimensional CNN coupling model to obtain the trained model"; if not, using the trained model as the multi-dimensional CNN coupling model for the next iteration, and returning "using the training set as the multi-dimensional CNN coupling model" The training set is used as input, and the multi-dimensional CNN coupling model is trained to obtain the trained model" steps; The obtained data set specifically includes: Obtain the factor distribution sample map of each landslide impact factor in the target area and the landslide catalog map of the target area; the landslide catalog map identifies landslide locations and non-landslide locations in the target area; Using a moving window method to extract multiple windows in the landslide catalog map; For each window, extract the area corresponding to the window position in each factor distribution sample map, and obtain a factor sample map of each landslide impact factor to form a factor sample set corresponding to the window. ; Use the landslide situation of the central pixel of the window as the label corresponding to the window; all the factor sample sets and labels corresponding to the window form a data set, and the labels in the data set are the number and label of the landslide factor sample set The number of factor sample sets for non-landslides is the same. 2. The landslide susceptibility evaluation method according to claim 1, characterized in that the topographic and geomorphological factors include elevation, slope, slope aspect, plane curvature, section curvature and terrain humidity index; the geological condition factors Including lithology, distance from the fault and distance from the epicenter; the environmental condition factors include normalized vegetation index and distance from the channel. 3. The landslide susceptibility evaluation method according to claim 1, characterized in that, before obtaining the factor distribution map of each landslide influence factor in the target area among the plurality of landslide influence factors, the landslide susceptibility The evaluation method also includes: calculating the variance expansion factor of each initial factor, and selecting the initial factor whose variance expansion factor is less than a preset threshold as the landslide impact factor. 4. The landslide susceptibility evaluation method according to claim 1, characterized in that the two-dimensional convolutional neural network includes a plurality of two-dimensional convolution pooling blocks and one two-dimensional convolution layer connected in sequence, so The two-dimensional convolution and pooling block includes a two-dimensional convolution layer and a two-dimensional maximum pooling layer connected in sequence; The one-dimensional convolutional neural network includes a sequentially connected Reshape layer, a plurality of one-dimensional convolution pooling blocks and a one-dimensional convolution layer. The one-dimensional convolution pooling block includes a sequentially connected one-dimensional convolution layer. and one-dimensional max pooling layer; The fully connected layer module includes two fully connected layers connected in sequence; Wherein, the two-dimensional convolution layer, the one-dimensional convolution layer and the fully connected layer all use the ReLU function. 5. The landslide susceptibility evaluation method according to claim 1, characterized in that, before using a moving window method to extract multiple windows in the landslide catalog map, the landslide susceptibility evaluation method further includes: Perform histogram equalization on the factor distribution sample map to obtain an equalized image; perform binarization processing on the landslide catalog map to obtain a binary image; use the equalized image and the binarized image as New factor distribution sample plots and landslide catalog plots. 6. The landslide susceptibility evaluation method according to claim 1, characterized in that, after obtaining the data set, the landslide susceptibility evaluation method further includes: evaluating each group of factors labeled as landslide in the data set. The sample sets are all subjected to data enhancement to obtain multiple groups of enhanced sample sets to expand the data set to obtain an expanded data set, and use the expanded data set as a new data set; the data enhancement includes horizontal Flip and flip vertically. 7. The landslide susceptibility evaluation method according to claim 1, characterized in that when training the multi-dimensional CNN coupling model, the mean square error is used as the loss function, and the Adam optimizer is used to update the network of the multi-dimensional CNN coupling model. parameter. 8. A multi-dimensional CNN coupled landslide susceptibility evaluation system, characterized in that the landslide susceptibility evaluation system includes: The data acquisition module is used to obtain the factor distribution map of each of the multiple landslide impact factors in the target area; the landslide impact factors include topographic factors, geological condition factors and environmental condition factors; An evaluation module is used to take all the factor distribution maps as input and use the trained multi-dimensional CNN coupling model to determine the landslide probability of each location point in the target area, so as to evaluate the landslide susceptibility of the target area; The trained multi-dimensional CNN coupling model includes 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 fully connected layer module connected in sequence; the two-dimensional asymmetric aggregation module The symmetric aggregation module includes a two-dimensional maximum pooling layer and a two-dimensional average pooling layer connected in parallel and a concatenate layer connected to the output of the two-dimensional maximum pooling layer and the output of the two-dimensional average pooling layer respectively; The one-dimensional asymmetric aggregation module includes a one-dimensional maximum pooling layer and a one-dimensional average pooling layer connected in parallel and is respectively connected to the output of the one-dimensional maximum pooling layer and the output of the one-dimensional average pooling layer. Connected concatenate layer; Before using all the factor distribution maps as input and using the trained multi-dimensional CNN coupling model to determine the landslide probability of each location point in the target area, the landslide susceptibility evaluation system also includes training the multi-dimensional CNN coupling model. , get the trained multi-dimensional CNN coupling model, including: Obtain a data set, and divide the data set into a training set, a verification set and a test set according to a preset ratio; the data set includes multiple groups of factor sample sets and labels corresponding to each group of the factor sample sets; The factor sample set includes a factor sample map of each landslide influencing factor; the label is landslide or non-landslide; Taking the training set as input, the multi-dimensional CNN coupling model is trained to obtain the trained model; Taking the verification set as input, using the trained model to calculate the verification error; Determine whether the iteration termination condition is reached and obtain the judgment result; if the judgment result is yes, then end the iteration and use the trained model as the trained multi-dimensional CNN coupling model; if the judgment result is no, continue the iteration, Determine whether the verification error of the current iteration is the same as the verification error of the previous N iterations; if so, adjust the learning rate during the training process in the next iteration, use the trained model as the multi-dimensional CNN coupling model of the next iteration, and return The step of "taking the training set as input, training the multi-dimensional CNN coupling model to obtain the trained model"; if not, using the trained model as the multi-dimensional CNN coupling model for the next iteration, and returning "using the training set as the multi-dimensional CNN coupling model" The training set is used as input, and the multi-dimensional CNN coupling model is trained to obtain the trained model" steps; The obtained data set specifically includes: Obtain the factor distribution sample map of each landslide impact factor in the target area and the landslide catalog map of the target area; the landslide catalog map identifies landslide locations and non-landslide locations in the target area; Using a moving window method to extract multiple windows in the landslide catalog map; For each window, extract the area corresponding to the window position in each factor distribution sample map, and obtain a factor sample map of each landslide impact factor to form a factor sample set corresponding to the window. ; Use the landslide situation of the central pixel of the window as the label corresponding to the window; all the factor sample sets and labels corresponding to the window form a data set, and the labels in the data set are the number and label of the landslide factor sample set The number of factor sample sets for non-landslides is the same.