Landslide susceptibility prediction method, device, equipment and medium

Abstract

The invention relates to a landslide susceptibility prediction method, apparatus and device, and a medium, belongs to the technical field of artificial intelligence data analysis and processing, and solves the problem of low prediction precision caused by difficult utilization of multiple scales and multiple factors of disaster-inducing factors in landslide susceptibility prediction in the prior art. According to the technical scheme, the method mainly comprises the following steps: acquiring influence factor data of an area, wherein the influence factor data is three-dimensional image data comprising spatial distribution data of a plurality of influence factor categories of the area on the area; the influence factor data are input into a landslide susceptibility prediction model, the landslide susceptibility prediction model carries out convolution on the influence factor data to obtain feature vectors, a landslide probability distribution diagram is output based on the feature vectors, and convolution comprises pyramid expansion three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model; and post-processing the landslide probability distribution map to visualize the landslide probability distribution map.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence data analysis and processing, and specifically relates to a landslide susceptibility prediction method, device, equipment and medium. Background Art

[0002] Landslide susceptibility refers to the possibility or probability of a landslide occurring in a certain area under specific conditions. Landslide is a common geological disaster, usually caused by the combined effects of natural factors (such as rainfall, earthquakes, geological structures) and human activities (such as land development and road construction). Assessing landslide susceptibility is of great significance for disaster prevention, risk management and land use planning.

[0003] Landslide hazard is an environmental geological problem with spatiotemporal scale attributes. As a natural phenomenon occurring in a specific time and space, landslide hazard itself must contain scale-related properties. At the same time, as a controlling factor of landslides under regional natural conditions, its characterization characteristics and evolution process are also affected by scale. Therefore, it is crucial to consider the scale issue in landslide susceptibility analysis (LSA). By selecting an appropriate scale, the geographic information of landslide susceptibility can be better captured and understood, thereby providing a more targeted basis for landslide prevention and control strategies.

[0004] In geographic information science, scale is a key indicator for measuring the informatics mechanism of geographic phenomena, processes and changes, corresponding to levels and levels. GIS discretizes continuous geographic phenomena, which is a resampled expression at a certain accuracy and level. From the perspective of cognitive science, scale reflects the depth and breadth of people's cognition of geographical objects and spatial phenomena. The definitions of scale in geography and ecology tend to describe the scope, coverage and time span of geographical and ecological phenomena. However, the concept of scale in geographic information science and cartography describes more the accuracy and depth of phenomenon expression, which can be understood as the granularity of spatial data expression, the level of abstraction of geographical laws, the level of user information perception, the depth of field of geographic scene expression, and the level of digital twin element derivation.

[0005] The scale effect of geospatial data of hazard factors limits the performance of statistical landslide susceptibility models. Given this limitation, single-scale models are therefore unable to accurately calculate results and measure the importance of landslide factors. One solution is to determine the impact of multiple scales on landslide susceptibility and construct different scales for landslide susceptibility. However, combining multi-scale landslide hazards is challenging: ① lack of data of different scales; ② limited understanding of the causes of scale effects; ③ inability to find a suitable fusion multi-scale model. Some solutions are to compare the accuracy of LSA models at different scales and select the optimal single scale as the LSA method. For example, Wei et al. used Depthwise Separable Convolution (DSC) to extract spatial features to prevent confusion of multi-factor features. There are also a large number of studies that show that considering spatial scale and hazard factor correlation characteristics in the process of model construction can improve the accuracy of landslide susceptibility models. In addition, multi-scale spatial classification methods are suitable for LSA because different scale regions contain complementary but interrelated classification information. Multi-scale spatial LSA methods can be divided into two categories: LSA that integrates multi-scale data and constructs multi-scale analysis LSA models.

[0006] In recent years, neural networks (CNNs) have achieved great success in image-related tasks. As a powerful deep learning technique, the deeper model structure in CNN has great potential to solve the above problems. However, only a few studies have attempted to apply CNN models to map landslide susceptibility, mainly due to the limitation of data representation. The input of CNN models is usually one or more two-dimensional images, however, the hazard factors associated with landslide occurrence in LSA are usually points or irregular polygons. To alleviate this problem, in the work of Wang et al., landslide hazard factors were converted into different dimensions (1D, 2D and 3D) to meet the needs of CNN as "valid" input. Similar studies related to deep learning also use one-dimensional data as input to deep models. However, one-dimensional data and converted two-dimensional data only utilize local information and lack spatial location information, such as environmental information around the landslide. In addition, due to the large number of zero values ​​involved, the converted two-dimensional data becomes sparse, which is not conducive to the training of CNN models.

[0007] In summary, the existing technology for landslide susceptibility prediction has the problem of low prediction accuracy due to the difficulty in utilizing the multi-scale and multi-factor characteristics of disaster-causing factors, which needs to be solved. Summary of the invention

[0008] In view of the above analysis, the embodiments of the present invention aim to provide a landslide susceptibility prediction method, device, equipment and medium to solve the problem of low prediction accuracy in the prior art for landslide susceptibility prediction due to the difficulty in utilizing the multi-scale and multi-factor of disaster-causing factors.

[0009] The embodiment of the present invention provides a method for predicting landslide susceptibility, comprising the following steps:

[0010] Acquire impact factor data of a region, wherein the impact factor data is three-dimensional image data including spatial distribution data of a plurality of impact factor categories of the region in the region;

[0011] The influencing factor data is input into a landslide susceptibility prediction model, the landslide susceptibility prediction model performs convolution on the influencing factor data to obtain a feature vector, and outputs a landslide probability distribution map based on the feature vector, wherein the convolution includes a pyramid dilation three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model;

[0012] The landslide probability distribution map is post-processed to visualize the landslide probability distribution map.

[0013] In some embodiments, the influencing factor categories include one or a combination of aspect, slope, lithology, elevation, distance to faults, distance to rivers, land use, distance to roads, and rainfall.

[0014] In some embodiments, the training method of the landslide susceptibility prediction model includes:

[0015] Constructing a training data set, wherein the samples in the training data set include regional influencing factor data and regional historical landslide point distribution labels, wherein the regional influencing factor data includes spatial distribution data of several influencing factor categories in the region;

[0016] The neural network model includes a convolution module and an output module, the influencing factor data is input into the convolution module to obtain a feature vector, and the convolution module includes a pyramid expansion three-dimensional convolution module;

[0017] The feature vector is input into an output module, the output module obtains a probability vector based on the feature vector through a fully connected layer and an activation function, and maps the elements of the probability vector to the spatial coordinates of the region to obtain the landslide probability distribution map;

[0018] The loss value is determined based on the difference between the landslide probability distribution map and the historical landslide point distribution labels, and the neural network is back-propagated and trained iteratively to obtain the landslide susceptibility prediction model.

[0019] In some embodiments, inputting the impact factor data into a convolution module to obtain a feature vector comprises:

[0020] Inputting the influencing factor data or the feature map into the pyramid dilated 3D convolution module, wherein the pyramid dilated 3D convolution module includes a basic 3D convolution kernel and a dilated 3D convolution kernel, wherein the dilated 3D convolution kernel is larger than the basic 3D convolution kernel in at least the length and width dimensions of the feature map;

[0021] The influencing factor data or the feature map is convolved using the basic three-dimensional convolution kernel and the expanded three-dimensional convolution kernel respectively.

[0022] In some embodiments, the dilated three-dimensional convolution kernel has the same amount of non-zero data as the basic three-dimensional convolution kernel, and the position where the dilated three-dimensional convolution kernel fills the non-zero data is uniformly expanded along the dilation dimension;

[0023] The expanded three-dimensional convolution kernel includes a first expanded three-dimensional convolution kernel and a second expanded three-dimensional convolution kernel.

[0024] In some embodiments, the convolution module further includes a downsampling module and a two-dimensional convolution module, and the downsampling module is used to downsample the feature map output by the pyramid expansion three-dimensional convolution module;

[0025] Extracting features of the influencing factors by combining a plurality of pyramid dilated three-dimensional convolution modules and the downsampling module to obtain a first feature map;

[0026] The first feature map is convolved by the two-dimensional convolution module to obtain the feature vector.

[0027] In some embodiments, the post-processing includes a natural breaks method.

[0028] The present invention also provides a landslide susceptibility prediction device, comprising:

[0029] An acquisition module, used for acquiring the impact factor data of a region, wherein the impact factor data is three-dimensional image data including spatial distribution data of a plurality of impact factor categories of the region in the region;

[0030] A prediction module, inputting the influencing factor data into a landslide susceptibility prediction model, the landslide susceptibility prediction model convolves the influencing factor data to obtain a feature vector, and outputs a landslide probability distribution map based on the feature vector, wherein the convolution includes a pyramid dilation three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model;

[0031] The post-processing module is used to post-process the landslide probability distribution map to make the landslide probability distribution map visual.

[0032] The present invention also provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the landslide susceptibility prediction method as described in any of the above embodiments is implemented.

[0033] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the landslide susceptibility prediction method as described in any of the above embodiments is implemented.

[0034] The above embodiments of the present invention have at least the following beneficial effects:

[0035] The present invention superimposes the regional distribution of each category in the influencing factor data into three-dimensional image data, takes the influencing factor data in the form of three-dimensional image data as model input, and extracts data features using a pre-trained neural network model using pyramid dilation three-dimensional convolution, so that the features can contain the associations between influencing factors of different categories, and can obtain the associations between influencing factors from different scales, thereby improving the accuracy of the model in predicting landslide susceptibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0037] Figure 1 A schematic diagram of a flow chart of a method for predicting landslide susceptibility provided by an embodiment of the present invention;

[0038] Figure 2 Schematic diagram of the neural network structure of the landslide susceptibility prediction model in an embodiment of the present invention;

[0039] Figure 3 Schematic diagram of the 3D convolution kernel used in the pyramid dilation 3D convolution module;

[0040] Figure 4 Schematic diagram of ROC curves of four models;

[0041] Figure 5 The statistical diagram of the proportion of landslide points in each prone area of ​​the four models;

[0042] Figure 6 A schematic diagram of the structure of the landslide susceptibility prediction device provided by the present invention;

[0043] Figure 7 This is a schematic diagram of the electronic device architecture provided by the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. It should be noted that, in the absence of conflict, the embodiments and features in the embodiments disclosed in this disclosure can be combined, separated, interchanged and / or rearranged with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0045] The terms used here are for the purpose of describing specific embodiments, and are not intended to be restrictive. As used here, unless the context clearly indicates otherwise, the singular forms "one (kind, person)" and "said (the)" are also intended to include plural forms. In addition, when the terms "comprise" and / or "include" and their variations are used in this specification, it is explained that there are stated features, integral bodies, steps, operations, parts, assemblies and / or their groups, but it is not excluded that there are or add one or more other features, integral bodies, steps, operations, parts, assemblies and / or their groups. It should also be noted that, as used here, the terms "substantially", "approximately" and other similar terms are used as approximate terms and not as degree terms, so that they are used to explain the inherent deviations of the measured values, calculated values ​​and / or the values ​​provided that will be recognized by those of ordinary skill in the art.

[0046] The present disclosure is described below through several specific embodiments. In order to keep the following description of the embodiments of the present disclosure clear and concise, the present disclosure omits detailed descriptions of known functions and known components. Figure 1 The embodiment of the present invention provides a method for predicting landslide susceptibility, comprising the following steps:

[0047] S1. Acquire influencing factor data of a region, wherein the influencing factor data is three-dimensional image data including spatial distribution data of several influencing factor categories of the region.

[0048] It should be understood that the three-dimensional image data here refers to the three-dimensional image data formed by superimposing the two-dimensional spatial distribution data of each influencing factor category, which includes not only the two dimensions in the two-dimensional space, but also the dimension on the influencing factor category. The embodiment of the present invention expects to perform feature extraction and prediction on the influencing factor data as a whole to improve the accuracy of the prediction results. Based on this, it should be understood that the influencing factor data is pre-processed data, including but not limited to aligning the two-dimensional spatial dimensions of the data of each category.

[0049] S2. Inputting the influencing factor data into a landslide susceptibility prediction model, the landslide susceptibility prediction model convolves the influencing factor data to obtain a feature vector, and outputs a landslide probability distribution map based on the feature vector, wherein the convolution includes a pyramid dilation three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model.

[0050] S3. Post-processing the landslide probability distribution map to make the landslide probability distribution map visual.

[0051] Scale features are crucial for landslide susceptibility mapping because they can more accurately represent the terrain and its likelihood of landslides. Scale features can include the size of the mapped area, the resolution of the data used, and the type of data used. The present invention extracts features of influencing factors through pyramid dilation three-dimensional convolution. On the one hand, the pyramid structured convolution module can extract features of different scales of each influencing factor category. On the other hand, the three-dimensional convolution can be used to extract the hard-to-capture correlation information between each influencing factor category, thereby helping the model to better and more accurately predict the likelihood of landslides.

[0052] In some embodiments, the influencing factor categories include one or a combination of aspect, slope, lithology, elevation, distance to faults, distance to rivers, land use, distance to roads, and rainfall.

[0053] In one embodiment of the present invention, data of all nine influencing factors are used, so the depth of the formed three-dimensional image data is nine.

[0054] In some embodiments, the training method of the landslide susceptibility prediction model includes:

[0055] A training data set is constructed, wherein samples in the training data set include regional influencing factor data and regional historical landslide point distribution labels, wherein the regional influencing factor data include spatial distribution data of several influencing factor categories in the region.

[0056] It should be understood that the sample data format in the training data set should also include the above three-dimensional image data and also include the historical landslide point distribution labels that are spatially aligned with the three-dimensional image data.

[0057] The neural network model includes a convolution module and an output module. The influencing factor data is input into the convolution module to obtain a feature vector. The convolution module includes a pyramid expansion three-dimensional convolution module. After the influencing shadow data is input into the convolution module, feature extraction is performed through one or more layers of the pyramid expansion three-dimensional convolution module.

[0058] The feature vector is input into an output module, and the output module obtains a probability vector based on the feature vector through a fully connected layer and an activation function, and maps the elements of the probability vector to the spatial coordinates of the region to obtain the landslide probability distribution map; preferably, in some embodiments, the activation function adopts a hyperbolic tangent function (tanh), and it should be understood that other activation functions can be selected as needed, and the present invention does not make any special limitation on this.

[0059] The loss value is determined based on the difference between the landslide probability distribution map and the historical landslide point distribution labels, and the neural network is back-propagated and trained iteratively to obtain the landslide susceptibility prediction model.

[0060] Preferably, during training, mean square error can be used as the loss function, and stochastic gradient descent can be used as the optimizer to train the neural network, with the initial learning rate, momentum, and decay set to 0.0005, 0.9, and 0.000001, respectively. The batch size is set to 80. In one embodiment, the model converges after about 350 training rounds.

[0061] In some embodiments, inputting the impact factor data into a convolution module to obtain a feature vector comprises:

[0062] Inputting the influencing factor data or the feature map into the pyramid dilated 3D convolution module, wherein the pyramid dilated 3D convolution module includes a basic 3D convolution kernel and a dilated 3D convolution kernel, wherein the dilated 3D convolution kernel is larger than the basic 3D convolution kernel in at least the length and width dimensions of the feature map;

[0063] The influencing factor data or the feature map is convolved using the basic three-dimensional convolution kernel and the expanded three-dimensional convolution kernel respectively.

[0064] In some embodiments, the dilated three-dimensional convolution kernel has the same amount of non-zero data as the basic three-dimensional convolution kernel, and the position where the dilated three-dimensional convolution kernel fills the non-zero data is uniformly expanded along the dilation dimension;

[0065] The expanded three-dimensional convolution kernel includes a first expanded three-dimensional convolution kernel and a second expanded three-dimensional convolution kernel.

[0066] In some embodiments, the convolution module further includes a downsampling module and a two-dimensional convolution module, and the downsampling module is used to downsample the feature map output by the pyramid expansion three-dimensional convolution module;

[0067] Extracting features of the influencing factors by combining a plurality of pyramid dilated three-dimensional convolution modules and the downsampling module to obtain a first feature map;

[0068] The first feature map is convolved by the two-dimensional convolution module to obtain the feature vector.

[0069] Specifically, in some embodiments, Figure 2 , which is a schematic diagram of the neural network structure for training the landslide susceptibility prediction model. In the illustrated embodiment, for the convenience of illustration, the input is a three-dimensional image data with a size of 15*15 in length and width and a depth of 9, that is, the influencing factor data.

[0070] First, feature expansion is performed on the input to obtain five superimposed impact factor data, that is, 45@15*15 data, that is, the two-dimensional scale is 15*15 and the depth is 45.

[0071] Then, if Figure 2 As shown, a pyramid dilation 3D convolution is performed on the feature expanded data, and after passing through the first pyramid dilation 3D convolution module, a 35*3@12*12 feature map is obtained.

[0072] Three-dimensional convolution is performed respectively through three-dimensional convolution kernels of three scales, one of which has the smallest scale and is a three-dimensional convolution kernel with full data filling, namely the above-mentioned basic three-dimensional convolution kernel. The other two are the first expanded three-dimensional convolution kernel and the second expanded three-dimensional convolution kernel. It should be understood that the pyramid expanded three-dimensional convolution of the present invention does not limit the number of expanded three-dimensional convolution kernels. The scales and parameter distributions of the three three-dimensional convolution kernels in this embodiment are as follows: Figure 3 As shown. It is worth noting that the second and third expanded 3D convolution kernels are uniformly diffused as the convolution kernels are expanded in length and width dimensions of the image. Uniform diffusion here means that, for example, the parameters of the basic 3D convolution kernel are densely distributed in 3*3 in 2D space, while the first expanded 3D convolution kernel is expanded to 5*5 in 2D space, but the parameters are distributed in 9 points in the center and the edge. Similarly, the second expanded 3D convolution kernel is also similarly distributed.

[0073] It is worth noting that in the illustrated embodiment, since the depth of the input data is 9, the third dimension is not expanded when the convolution kernel in this embodiment is expanded. It should be understood that if there is a lot of data in the impact factor category dimension, the convolution kernel during the pyramid expansion 3D convolution can be expanded in all three dimensions to improve the perception field of the third dimension.

[0074] Furthermore, the feature map is downsampled to obtain a 35*3@6*6 feature map, which is then input into the second pyramid dilated 3D convolution module to obtain a 15*9@4*4 feature map.

[0075] Then down-sample to obtain a 15*9@2*2 feature map, and finally perform two-dimensional convolution through two two-dimensional convolution kernels to obtain a 270-dimensional feature vector.

[0076] The embodiment of the present invention avoids the expansion of convolution kernel parameters while expanding the convolution kernel during convolution, and extracts features at different scales through similar parameter distribution, thereby improving the richness of feature information while maintaining a low amount of calculation.

[0077] In some embodiments, the post-processing includes a natural break method. When the landslide susceptibility probability map is generated using a trained model, these probability values ​​are usually a continuous value (e.g., between 0 and 1), indicating the possibility of landslides in different regions. The natural break method can divide these continuous probability values ​​into several categories (such as low susceptibility, medium susceptibility, high susceptibility, etc.). Through the natural break method, the values ​​in the probability map will be divided into several intervals, so that the difference in values ​​within each interval is minimized, but the difference between intervals is maximized.

[0078] In some embodiments, by comparing the accuracy evaluation analysis, ROC curves, and proportion of LSM landslide points of four models, namely, a single-scale two-dimensional convolutional neural network model (CNN), a pyramid dilated convolutional neural network model (PDCNN), a three-dimensional convolutional neural network model (3DCNN), and a pyramid dilated three-dimensional convolutional neural network model (PD3DCNN) adopted in the present invention, the performance of the PD3DCNN model provided by the present invention can be intuitively seen.

[0079] As shown in Table 1, the PDCNN model and PD3DCNN outperform the CNN model and 3DCNN model in terms of OA, Precision, Recall, and F1-score, showing that their classification performance is better than the latter two models. Specifically, the accuracy (OA) of the PDCNN and PD3DCNN models reached 84.6% and 87.4%, which is 3.7% and 3.5% higher than the accuracy of CNN and 3DCNN models of 80.8% and 84.0%, respectively. This shows that the model with the added pyramid dilated convolution module can classify samples more accurately and more accurately judge the landslide susceptibility level of the study area. The precision (Precision) of the PDCNN and PD3DCNN models is 83.7% and 86.3%, which is 3.1% and 4.9% higher than that of CNN and 3DCNN, respectively. This shows that the PDCNN and PD3DCNN models have higher accuracy and lower proportion of misclassification when judging high and higher landslide susceptibility levels. The recall rates of the PDCNN and PD3DCNN models reached 85.8% and 89.1%, which are 3.7% and 4.0% higher than those of the CNN and 3DCNN models, respectively. This shows that the PDCNN and PD3DCNN models have a wider range of recognition for high and relatively high landslide susceptibility levels, and can correctly identify more landslide areas. The F1-score of the PDCNN and PD3DCNN models reached 84.8% and 87.6%, which are 3.7% and 3.5% higher than those of the CNN and 3DCNN models, respectively. This comprehensively reflects that the fusion of multi-scale windows adds more feature information in terms of accuracy and recall, and the classification effect is better. In summary, the PDCNN and PD3DCNN models with the pyramid dilated convolution module outperform the CNN model and the 3DCNN model in terms of accuracy, precision, recall rate and F1-score. At the same time, the evaluation accuracy, precision, recall rate and F1-score of PD3DCNN are also improved by 2.9%, 2.5%, 3.2% and 2.9% respectively compared with PDCNN. This indicates that the dilated convolution module has a certain degree of independence, which can improve the performance of the model while ensuring the extraction ability of the spatial relationship features and disaster factor relationship features of 3DCNN.

[0080] Table 1 Model accuracy evaluation

[0081]

[0082]

[0083] In addition, the performance of the four models was compared by ROC curve and AUC value evaluation. The ROC curve reflects the model convergence speed, and the AUC value reflects the model accuracy. Figure 4As shown in the figure, the AUC values ​​of CNN, PDCNN, 3DCNN and PD3CNN models are 0.816, 0.844, 0.835 and 0.870 respectively, and the AUC values ​​of the four models are all greater than 0.5, proving that the four models are feasible. At the same time, the PD3DCNN and PDCNN with the pyramid dilated convolution module increased the AUC values ​​of the area under the curve by 2.8% and 3.5% respectively compared with the 3DCNN and CNN models, proving that adding the dilated convolution module, that is, increasing the receptive field size and scale range size, improves the model training accuracy.

[0084] Table 2 AUC curve values ​​of the model

[0085] Model Name Area under the curve (AUC value) CNN 0.816 PDCNN 0.844 3DCNN 0.835 PD3DCNN 0.870

[0086] Figure 4 The quantitative evaluation results of the models in the test dataset are shown. The prediction accuracy of the models with added pyramid dilated convolution is improved, especially PD3DCNN, which achieves the best landslide prediction ability with the highest AUC (0.870). Compared with the baseline model, PDCNN (AUC 0.870) is better than CNN, PD3DCNN (AUC 0.870) is better than 3DCNN (AUC 0.835), and the model with added pyramid dilated convolution module achieves satisfactory accuracy results, showing the superiority of adding scale features. Although the prediction performance of CNN and 3DCNN in their control group is poor, CNN (AUC 0.816) is worse than 3DCNN (AUC0.835), showing the advantage of 3D convolution in fitting the complex correlation between landslides and hazard factors.

[0087] An embodiment of the present invention also statistically validates the generated visualization image of landslide susceptibility distribution through test data and compares four models. This embodiment further divides the landslide susceptibility predicted by the four models into five risk areas: Very Low, Low, Moderate, High and Very High through the natural breakpoint method. According to statistics, high-susceptibility areas are widely distributed in the southeast and northwest of the study area, accounting for more than 20% of the total area, while low-susceptibility areas account for about 10% of the total area. In addition, the proportion of landslide points is obtained by analyzing the proportion of historical landslides in different susceptibility levels. Figure 5 As shown, the higher the susceptibility level, the more landslide points there are. According to the statistical results, most landslide points are located in high susceptibility areas, while fewer landslide points appear in low susceptibility areas, which means that the method based on the present invention accurately predicts that most of the observed landslides fall into high susceptibility areas.

[0088] The LSM predicted by the CNN model shows that about 60.9% of the landslide points are distributed in extremely high and high susceptibility areas, and 5.72% of the landslide points are distributed in extremely low susceptibility areas. In the PDCNN model combined with pyramid dilated convolution, about 70.9% of the landslide points are distributed in extremely high and high susceptibility areas, and 3.73% of the landslide points are distributed in extremely low susceptibility areas. Based on the 3DCNN model, about 65.4% of the landslide points are distributed in extremely high and high susceptibility areas, and 3.98% of the landslide points are distributed in extremely low susceptibility areas. In addition, based on the LSM predicted by the PD3DCNN model, about 74.0% of the landslide points are distributed in extremely high and high susceptibility grade areas, and 1.74% of the landslide points are distributed in extremely low susceptibility grade areas. The above comparison shows that the model provided by the present invention can effectively improve the classification accuracy of landslide points.

[0089] In addition, the distribution of non-landslide points in each prone area of ​​PD3DCNN and PDCNN models is more reasonable than that of 3DCNN and CNN. The highest proportion of non-landslide samples in extremely low and low prone areas in PD3DCNN is 77.9%, which is 7.3% higher than 3DCNN 70.6%; the proportion of non-landslide samples in high prone areas and extremely high prone areas is as low as 10.4%, which is 4.0% lower than 3DCNN 14.4%. The highest proportion of non-landslide samples in extremely low and low prone areas in PDCNN is 70.4%, which is 6.2% higher than CNN 64.2%; the proportion of non-landslide samples in high prone areas and extremely high prone areas is 13.4%, which is 1.5% lower than 3DCNN 14.9%. It is proved that the positive and negative sample classification of the landslide susceptibility prediction model provided by the present invention is better, and the model performance is better.

[0090] The present invention also provides a landslide susceptibility prediction device, such as Figure 6 As shown, including:

[0091] An acquisition module, used for acquiring the impact factor data of a region, wherein the impact factor data is three-dimensional image data including spatial distribution data of a plurality of impact factor categories of the region in the region;

[0092] A prediction module, inputting the influencing factor data into a landslide susceptibility prediction model, the landslide susceptibility prediction model convolves the influencing factor data to obtain a feature vector, and outputs a landslide probability distribution map based on the feature vector, wherein the convolution includes a pyramid dilation three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model;

[0093] The post-processing module is used to post-process the landslide probability distribution map to make the landslide probability distribution map visual.

[0094] The present invention also provides an electronic device, such as Figure 7As shown, it includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the biomolecule pair matching analysis method as described in any of the above embodiments is implemented.

[0095] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the biomolecule pair matching analysis method as described in any of the above embodiments is implemented.

[0096] Computer readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0097] The professionals should further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0098] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0099] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting landslide susceptibility, characterized in that: The following steps are involved: Acquire impact factor data of a region, wherein the impact factor data is three-dimensional image data including spatial distribution data of a plurality of impact factor categories of the region in the region; The influencing factor data is input into a landslide susceptibility prediction model, the landslide susceptibility prediction model performs convolution on the influencing factor data to obtain a feature vector, and outputs a landslide probability distribution map based on the feature vector, wherein the convolution includes a pyramid dilation three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model; The landslide probability distribution map is post-processed to visualize the landslide probability distribution map.

2. The method for predicting landslide susceptibility according to claim 1, characterized in that: The influencing factor categories include one or a combination of slope aspect, slope, lithology, elevation, distance to faults, distance to rivers, land use, distance to roads, and rainfall.

3. The method for predicting landslide susceptibility according to claim 1, characterized in that: The training method of the landslide susceptibility prediction model includes: Constructing a training data set, wherein the samples in the training data set include regional influencing factor data and regional historical landslide point distribution labels, wherein the regional influencing factor data includes spatial distribution data of several influencing factor categories in the region; The neural network model includes a convolution module and an output module, the influencing factor data is input into the convolution module to obtain a feature vector, and the convolution module includes a pyramid expansion three-dimensional convolution module; The feature vector is input into an output module, the output module obtains a probability vector based on the feature vector through a fully connected layer and an activation function, and maps the elements of the probability vector to the spatial coordinates of the region to obtain the landslide probability distribution map; The loss value is determined based on the difference between the landslide probability distribution map and the historical landslide point distribution labels, and the neural network is back-propagated and trained iteratively to obtain the landslide susceptibility prediction model.

4. The method for predicting landslide susceptibility according to claim 3, characterized in that: The step of inputting the impact factor data into a convolution module to obtain a feature vector comprises: Inputting the influencing factor data or the feature map into the pyramid dilated 3D convolution module, wherein the pyramid dilated 3D convolution module includes a basic 3D convolution kernel and a dilated 3D convolution kernel, wherein the dilated 3D convolution kernel is larger than the basic 3D convolution kernel in at least the length and width dimensions of the feature map; The influencing factor data or the feature map is convolved using the basic three-dimensional convolution kernel and the expanded three-dimensional convolution kernel respectively.

5. The method for predicting landslide susceptibility according to claim 4, characterized in that: The dilated three-dimensional convolution kernel has the same amount of non-zero data as the basic three-dimensional convolution kernel, and the position of the dilated three-dimensional convolution kernel filling non-zero data is uniformly expanded along with the dilation dimension; The expanded three-dimensional convolution kernel includes a first expanded three-dimensional convolution kernel and a second expanded three-dimensional convolution kernel.

6. The method for predicting landslide susceptibility according to claim 3, characterized in that: The convolution module also includes a downsampling module and a two-dimensional convolution module, and the downsampling module is used to downsample the feature map output by the pyramid expansion three-dimensional convolution module; Extracting features of the influencing factors by combining a plurality of pyramid dilated three-dimensional convolution modules and the downsampling module to obtain a first feature map; The first feature map is convolved by the two-dimensional convolution module to obtain the feature vector.

7. The method for predicting landslide susceptibility according to claim 1, characterized in that: The post-processing includes the natural breaks method.

8. A landslide susceptibility prediction device, characterized in that: include: An acquisition module, used for acquiring the impact factor data of a region, wherein the impact factor data is three-dimensional image data including spatial distribution data of a plurality of impact factor categories of the region in the region; A prediction module, inputting the influencing factor data into a landslide susceptibility prediction model, the landslide susceptibility prediction model convolves the influencing factor data to obtain a feature vector, and outputs a landslide probability distribution map based on the feature vector, wherein the convolution includes a pyramid dilation three-dimensional convolution; the landslide susceptibility prediction model is a pre-trained neural network model; The post-processing module is used to post-process the landslide probability distribution map to make the landslide probability distribution map visual.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the method for predicting landslide susceptibility according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the landslide susceptibility prediction method according to any one of claims 1 to 7 is implemented.

Images