Global scale strong earthquake induced landslide space distribution probability intelligent prediction method

Abstract

The invention discloses a global scale strong earthquake induced landslide space distribution probability intelligent prediction method. The prediction method comprises the following steps: S1, selecting x times of earthquake induced landslide events under a global scale as a main target, and artificially constructing an accurate landslide label; s2, selecting four types of feature factors including a seismic oscillation factor, a geological factor, a topographic factor and an environmental factor, enabling the four types of feature factors to be in one-to-one correspondence with landslide labels in geographic spatial positions, and converting all the feature factors into a TIFF format by using a GDAL open source library; according to the method, topographic factors, geological factors, environmental factors and seismic oscillation factors are comprehensively considered, and a novel full-convolution regression neural network model is adopted, so that the spatial distribution probability of landslide induced by each region in the global under the condition of an earthquake can be efficiently and accurately predicted; and all correlation factors can be automatically processed quickly when an earthquake occurs, so that the model achieves nearly real-time prediction performance.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster risk prediction, and in particular to a global-scale strong earthquake-induced landslide spatial distribution probability intelligent prediction method. Background Art

[0002] Earthquakes are one of the most persistent and serious natural disasters, claiming tens of thousands of lives every year and causing devastating impacts on the human living environment. In particular, landslides caused by earthquakes are one of the important factors leading to disastrous consequences. However, in addition to directly causing casualties, landslides induced by strong earthquakes often hinder rescue operations, making it difficult for aid workers and supplies to reach the disaster area in time. Faced with this urgent challenge, it is crucial to grasp the distribution of landslides in mountainous areas after earthquakes in real time to reduce casualties and economic losses.

[0003] With the advancement of earth observation technology, remote sensing surveys have become an effective means of identifying the spatial distribution of landslides, but the weather conditions or operating environment in earthquake-affected areas often limit the possibility of obtaining real-time images. To address this problem, researchers have been exploring various prediction methods for the spatial distribution probability of landslides, which are mainly divided into two categories: physics-based models and data-driven models. Physics-based models analyze slope deformation and rupture under the action of earthquake forces through dynamic mechanisms to infer the distribution of landslides. However, these models require a deep understanding of the landslide causal mechanism and are less efficient in large-scale predictions. In contrast, data-driven models reveal the distribution pattern of landslides through statistical analysis of historical events, reducing the difficulty of data collection and processing. Although sensitivity methods have been adopted, accurately locating and estimating the spatial distribution of landslides remains a challenge due to the uncertainty of model construction and event variability. One of the main obstacles is the lack of a universal prediction method for the spatial distribution of strong earthquake-induced landslides. Current models rely on a list of landslides from strong earthquake-induced landslide events that have already occurred, and on the other hand, they rely on traditional machine learning algorithms for modeling, making it difficult to achieve near-real-time predictions for global or macro-regions. Summary of the invention

[0004] The purpose of the present invention is to provide a global-scale strong earthquake-induced landslide spatial distribution probability intelligent prediction method to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a global scale strong earthquake induced landslide spatial distribution probability intelligent prediction method, the prediction comprises the following steps: S1, select x earthquake-induced landslide events on a global scale as the main target to manually construct accurate landslide labels; S2, select four types of characteristic factors, namely, seismic factors, geological factors, terrain factors and environmental factors, correspond one by one to landslide labels in geographic space, and convert all characteristic factors into TIFF format using GDAL open source library; S3, combining the small categories of seismic factors, geological factors, terrain factors and environmental factors into a TIFF file, cropping the TIFF file data and landslide labels using a sliding window cropping method, and flipping, symmetric and rotating the cropped TIFF file data and landslide labels based on computer vision technology to achieve data enhancement operations; S4. Use Python language and tensorflow open source architecture to design a fully convolutional regression neural network algorithm, independently design loss functions, and train artificial intelligence models; S5. Use GDAL to design a large-area geographic spatial TIFF large image deep learning ignoring edge prediction method, and use an adaptive threshold division method to perform threshold division on the spatial distribution probability prediction results of strong earthquake-induced landslides.

[0006] As a further improvement of the present invention, in S3: Sub-factors of the ground motion factor include peak ground motion acceleration; Sub-categories of geological factors include distance from faults and lithology; Subcategories of topographic factors include slope, surface roughness, elevation, curvature, and topographic position index; Subcategories of environmental factors include land cover, soil type, and surface hydrological index.

[0007] As a further improvement of the present invention, the S1 comprises the following steps: S101. Use the Create Features tool in ArcGIS to create a vector file, manually draw the landslide boundary, save it as a Shapefile, create an attribute table in the Shapefile, set the header to value, and set the value of all the annotated landslide vector files to 1; S102. Use the Erase tool in ArcGIS to cut out the drawn landslide vector file within the seismic image to obtain the space without landslide, and assign the attribute table value to 0.

[0008] As a further improvement of the present invention, S2 includes the following steps: S201. Comprehensively consider the topographic characteristics, geological characteristics, environmental characteristics and seismic characteristics of the study area within the spatial range of the impact of n earthquakes, select eleven small factors including seismic peak acceleration, distance from fault, lithology, slope, surface roughness, elevation, curvature, terrain position index, surface cover, soil type and surface hydrological index, and use the GDAL open source library to unify the spatial resolution and geographic coordinates, with a spatial resolution of 30m and a geographic coordinate system of WGS1984; S202, based on step S201, a group of small class factors are normalized by using the GDAL library and the Numpy library, and all data types are unified into Float32 bits, the data range is 0-255, and the data format is TIFF; S203, convert the vector file created in step S101 into a TIFF binary tag using the ogr.Getlayer() function and the GDAL.GetDiverByName() function in the GDAL library and the ogr library, and the TIFF type data is consistent with the spatial resolution and geographic coordinates of the selected feature factors.

[0009] As a further improvement of the present invention, S3 comprises the following steps: S301, using the GDAL library, synthesize the eleven small category factors from top to bottom in the order of peak seismic acceleration, fault distance, lithology, slope, surface roughness, elevation, curvature, terrain position index, surface cover, soil type and surface hydrological index into an eleven-channel TIFF format file, and store it as a training database; S302, using the GDAL library to store the landslide labels as a training label library, and corresponding them one by one with the training data names; S303, use GDAL and Numpy libraries to perform sliding window cropping on the data and labels, set the Overlap parameter to 0.1, crop the data to 448×488×11 TIFF and save the labels to 448×488×1 TIFF, and the data type is Float32; S304: Based on computer vision technology, the cropped data and labels are flipped, symmetrical and rotated to achieve data enhancement.

[0010] As a further improvement of the present invention, the S4 comprises the following steps: S401. Design a fully convolutional regression neural network algorithm based on Python language and Tensorflow open source architecture. The algorithm mainly consists of three parts: encoder, channel attention and spatial attention, and decoder; S402, based on the mean square error, a loss function MSE_loss is designed for the image regression neural network algorithm, and its basic formula is as follows: ; in, is the number of samples for pixel value, is subordinate to The true label of the sample, 0 or 1, is subordinate to The predicted label value of the sample; S403, using Adam as the optimizer, R 2 The cosine annealing scheme is used to optimize the learning rate. In the initial stage of training, the learning rate is set to a larger value so that it can quickly converge to the local minimum. As the training steps proceed, the learning rate gradually decreases according to the curve of the cosine function, so that the model can adjust the parameters more finely when it is close to the optimal solution, avoiding falling into the local minimum too early and being unable to jump out. The loss function does not change in every 10 epochs, and the learning rate is cosine transformed.

[0011] As a further improvement of the present invention, the S5 comprises the following steps: S501, according to the ignoring edge prediction method, overlappingly cropping TIFF and adopting the ignoring edge strategy when stitching, directly inputting a large TIFF file of an independent geographic space into the model for result prediction, and the ignoring edge prediction method is written by the GDAL library; S502. Based on computer vision technology and statistical learning strategies, adaptive threshold division is performed on the predicted results of the spatial distribution probability of strong earthquake-induced landslides to divide them into extremely high probability areas, high probability areas, medium probability areas, low probability areas and extremely low probability areas of landslides.

[0012] As a further improvement of the present invention, the encoder comprises: Convolutional layer, size 3×3, pooling layer, size 3×3 and residual layer, size 1×1, 3 convolution operations and 2 pooling operations; CBAM consists of two parts: channel attention layer and spatial attention layer. Channel attention includes: Global average pooling with size 1×1, fully connected layers and normalization layers; The spatial attention consists of a convolutional layer with size 3×3, a global average pooling layer with size 1×1, a fully connected layer, and a normalization layer.

[0013] As a further improvement of the present invention, the decoder mainly comprises: The deconvolution layer is of size 3×3, the upsampling layer is of size 3×3 and the residual layer is of size 1×1, which performs 3 deconvolution operations and 2 upsampling operations.

[0014] As a further improvement of the present invention, in the process of training the model in S403, the batch_size parameter is set to 24, the epochs parameter is set to 400, the initial learning rate parameter is set to 0.00001, and the pre-training model is not used.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention constructs an automated solution for characteristic factor preprocessing and vector label preprocessing using GDAL. Through the solution, all relevant factors can be quickly and automatically processed when an earthquake occurs, so that the model can achieve near real-time prediction performance. The fully convolutional regression neural network algorithm is used to fully consider the spatial correlation between earthquake-induced landslides and their inducing factors, and can also mine the nonlinear relationship between disasters and characteristic factors from a global perspective. The edge-ignoring prediction solution and the sliding window clipping method for large TIFF data are used, and the adaptive threshold division method is used to effectively reduce the noise of the prediction results and accurately obtain the threshold range of the prediction results, ultimately achieving the purpose of constructing a global-scale strong earthquake-induced landslide spatial distribution probability prediction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the present invention; Figure 2 This is a schematic diagram of GDAL synthesizing multi-channel TIFF data in the present invention; Figure 3 It is a binary schematic diagram of the vector label of the present invention; Figure 4 A schematic diagram of the present invention ignoring the edge prediction scheme and the sliding window cropping of large TIFF data; Figure 5 This is a schematic diagram of ignoring edge prediction in the present invention; Figure 6 This is a schematic diagram of cutting the sliding window of the present invention; Figure 7 The present invention calculates the spatial distribution probability prediction result diagram of strong earthquake-induced landslides. DETAILED DESCRIPTION

[0017] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] like Figure 1The figure is a flow chart of the overall method of the present invention. The global scale strong earthquake-induced landslide spatial distribution probability intelligent prediction method of this embodiment includes the following steps (): S1. Select x (x=38) earthquake-induced landslide events that occurred on a global scale from 1976 to the present as the main target to manually construct accurate landslide labels. The specific steps are as follows: S101. Use the Create Features tool in ArcGIS to create a vector file, manually draw the landslide boundary, save it as a Shapefile, create an attribute table in the Shapefile, set the table header to value, and set the value of all the annotated landslide vector files to 1.

[0019] S102. Use the Erase tool in ArcGIS to cut out the drawn landslide vector file within the seismic image to obtain the space without landslide, and assign the attribute table value to 0.

[0020] S2. Select four types of characteristic factors, namely, seismic factors, geological factors, terrain factors and environmental factors. These four types of characteristic factors are obtained through the following channels: ① Seismic factors: China Earthquake Administration (https: / / www.cea.gov.cn / ); ② Geological factors, terrain factors and environmental factors: Geospatial Data Cloud (https: / / www.gscloud.cn / #page1 / 4). The sub-factors of seismic factors include peak seismic acceleration, the sub-factors of geological factors include distance from faults and lithology, the sub-factors of terrain factors include slope, surface roughness, elevation, curvature and terrain position index, and the sub-factors of environmental factors include surface cover, soil type and surface hydrological index. They correspond to landslide labels one by one in geographic spatial position. Use the GDAL open source library to convert all characteristic factors into TIFF format. The specific steps are as follows: S201. Within the spatial range of the x-th earthquake, the topographic characteristics, geological characteristics, environmental characteristics and seismic characteristics of the study area were comprehensively considered. Eleven factors, including peak seismic acceleration, distance from the fault, lithology, slope, surface roughness, elevation, curvature, terrain position index, surface cover, soil type and surface hydrological index, were selected. The GDAL open source library was used to unify the spatial resolution and geographic coordinates (spatial resolution 30m, geographic coordinate system WGS1984).

[0021] S202, based on step S101, the data of the small class factors of the same class are normalized by using the GDAL library and the Numpy library, and all data types are unified into Float32 bits, the data range is 0-255, and the data format is TIFF. See the feature factor data preprocessing Figure 2 .

[0022] S203, convert the vector file created in step S101 into a TIFF binary tag using the ogr.Getlayer() function and the GDAL.GetDiverByName() function in the GDAL library and the ogr library, and the TIFF type data is consistent with the spatial resolution and geographic coordinates of the selected feature factors. Figure 3 Schematic diagram of binary label conversion.

[0023] S3. The sub-category factors of seismic factors, geological factors, terrain factors and environmental factors are combined into a TIFF format file. The TIFF format file data and landslide labels are cropped using a sliding window cropping method. The cropped TIFF format file and landslide labels are flipped, symmetrical and rotated based on computer vision technology to achieve data enhancement. The specific steps are as follows: S301. Use the GDAL library to synthesize the eleven sub-category factors into an eleven-channel TIFF format file from top to bottom in the order of peak seismic acceleration, fault distance, lithology, slope, surface roughness, elevation, curvature, terrain position index, surface cover, soil type and surface hydrological index, and store it as a training database.

[0024] S302, using the GDAL library to store the landslide labels as a training label library, and corresponding them to the training data names one by one.

[0025] 303. Use GDAL and Numpy libraries to perform sliding window cropping on data and labels. The Overlap parameter is set to 0.1. The data is cropped and saved as TIFF of (448×488×11) size, and the label is cropped and saved as TIFF of (448×488×1) size. The data type is Float32 bit.

[0026] S304: Based on computer vision technology, the cropped data and labels are flipped, symmetrical and rotated to achieve data enhancement.

[0027] S4. Use Python language and tensorflow open source architecture to design a fully convolutional regression neural network algorithm, independently design the loss function, and train the artificial intelligence model (see the model structure diagram for details). Figure 4 ), the specific steps are: S401, based on Python language and Tensorflow open source architecture, designed a fully convolutional regression neural network algorithm. The algorithm mainly consists of three parts: encoder, channel attention and spatial attention (CBAM), and decoder. The encoder includes: Convolutional layer (size 3×3), pooling layer (size 3×3) and residual layer (size 1×1), performing 3 convolution operations and 2 pooling operations; CBAM consists of two parts: channel attention layer and spatial attention layer. Channel attention includes: Global average pooling (size 1×1), fully connected layers, and normalization layers; The spatial attention consists of a convolutional layer (of size 3×3), a global average pooling layer (of size 1×1), a fully connected layer, and a normalization layer; The decoder mainly includes: It consists of a deconvolution layer (size 3×3), an upsampling layer (size 3×3) and a residual layer (size 1×1), performing 3 deconvolution operations and 2 upsampling operations.

[0028] S402, based on the mean square error, a loss function MSE_loss is designed for the image regression neural network algorithm, and its basic formula is as follows:

[0029] in, is the number of samples (pixel values), is subordinate to The true label of the sample (0 or 1), is subordinate to The predicted label value of the sample.

[0030] S403, using Adam (adaptive moment estimation) as the optimizer, R 2 (R-squared score) and MAE (mean absolute error) are used as accuracy indicators, and the cosine annealing scheme is used to optimize the learning rate. The cosine annealing scheme means that in the initial stage of training, the learning rate is set to a larger value so that it can quickly converge to the vicinity of the local minimum. As the training steps proceed, the learning rate gradually decreases according to the curve of the cosine function, so that the model can adjust the parameters more finely when it is close to the optimal solution, avoiding falling into the local minimum too early and being unable to jump out. The loss function does not change in every 10 epochs, the learning rate is cosine transformed, the batch_size parameter is set to 24, the epochs parameter is set to 400, the initial learning rate parameter is set to 0.00001, and the pre-trained model is not used.

[0031] S5. Use GDAL to design a large-area geographic spatial TIFF large image deep learning ignoring edge (windowing) prediction method, and use the adaptive threshold division method to perform threshold division on the spatial distribution probability prediction results of strong earthquake-induced landslides. The specific steps are as follows: S501, according to the ignoring edge prediction method, overlappingly cropping TIFF and adopting the ignoring edge strategy when stitching, directly inputting a large TIFF file of an independent geographic space into the model for result prediction, the ignoring edge prediction method is written by the GDAL library, Figure 5 Schematic diagram of the sliding window cropping scheme.

[0032] S502. Based on computer vision technology and statistical learning strategies, adaptive threshold division is performed on the predicted spatial distribution probability prediction results of strong earthquake-induced landslides to divide them into extremely high probability areas, high probability areas, medium probability areas, low probability areas and extremely low probability areas. Figure 6 As shown, this is the result of adaptive threshold division. Figure 7 An example of a global region test dataset.

[0033] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0034] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A global scale strong earthquake-induced landslide spatial distribution probability intelligent prediction method, characterized by: This forecast includes the following steps: S1, select x earthquake-induced landslide events on a global scale as the main target to manually construct accurate landslide labels; S2, select four types of characteristic factors, namely, seismic factors, geological factors, terrain factors and environmental factors, correspond one by one to landslide labels in geographic space, and convert all characteristic factors into TIFF format using GDAL open source library; S3, combining the small categories of seismic factors, geological factors, terrain factors and environmental factors into a TIFF file, cropping the TIFF file data and landslide labels using a sliding window cropping method, and flipping, symmetric and rotating the cropped TIFF file data and landslide labels based on computer vision technology to achieve data enhancement operations; S4. Use Python language and tensorflow open source architecture to design a fully convolutional regression neural network algorithm, independently design loss functions, and train artificial intelligence models; S5. Use GDAL to design a large-area geographic spatial TIFF large image deep learning ignoring edge prediction method, and use an adaptive threshold division method to perform threshold division on the spatial distribution probability prediction results of strong earthquake-induced landslides.

2. The global scale strong earthquake induced landslide spatial distribution probability intelligent prediction method according to claim 1 is characterized by: In the S3: Sub-factors of the ground motion factor include peak ground motion acceleration; Sub-categories of geological factors include distance from faults and lithology; Subcategories of topographic factors include slope, surface roughness, elevation, curvature, and topographic position index; Subcategories of environmental factors include land cover, soil type, and surface hydrological index.

3. The global scale strong earthquake induced landslide spatial distribution probability intelligent prediction method according to claim 1 is characterized by: The S1 comprises the following steps: S101. Use the Create Features tool in ArcGIS to create a vector file, manually draw the landslide boundary, save it as a Shapefile, create an attribute table in the Shapefile, set the header to value, and set the value of all the annotated landslide vector files to 1; S102. Use the Erase tool in ArcGIS to cut out the drawn landslide vector file within the seismic image to obtain the space without landslide, and assign the attribute table value to 0.

4. The method for intelligent prediction of spatial distribution probability of landslides induced by global-scale strong earthquakes according to claim 3 is characterized by: The S2 includes the following steps: S201. Comprehensively consider the topographic characteristics, geological characteristics, environmental characteristics and seismic characteristics of the study area within the spatial range of the impact of n earthquakes, select eleven small factors including seismic peak acceleration, distance from fault, lithology, slope, surface roughness, elevation, curvature, terrain position index, surface cover, soil type and surface hydrological index, and use the GDAL open source library to unify the spatial resolution and geographic coordinates, with a spatial resolution of 30m and a geographic coordinate system of WGS1984; S202, based on step S201, a group of small class factors are normalized by using the GDAL library and the Numpy library, and all data types are unified into Float32 bits, the data range is 0-255, and the data format is TIFF; S203, convert the vector file created in step S101 into a TIFF binary tag using the ogr.Getlayer() function and the GDAL.GetDiverByName() function in the GDAL library and the ogr library, and the TIFF type data is consistent with the spatial resolution and geographic coordinates of the selected feature factors.

5. The global scale strong earthquake induced landslide spatial distribution probability intelligent prediction method according to claim 1 is characterized by: The S3 comprises the following steps: S301, using the GDAL library, synthesize the eleven small category factors from top to bottom in the order of peak seismic acceleration, fault distance, lithology, slope, surface roughness, elevation, curvature, terrain position index, surface cover, soil type and surface hydrological index into an eleven-channel TIFF format file, and store it as a training database; S302, using the GDAL library to store the landslide labels as a training label library, and corresponding them one by one with the training data names; S303, use GDAL and Numpy libraries to perform sliding window cropping on the data and labels, set the Overlap parameter to 0.1, crop the data to 448×488×11 TIFF and save the labels to 448×488×1 TIFF, and the data type is Float32; S304: Based on computer vision technology, the cropped data and labels are flipped, symmetrical and rotated to achieve data enhancement.

6. The global scale strong earthquake induced landslide spatial distribution probability intelligent prediction method according to claim 1 is characterized by: The S4 comprises the following steps: S401. Design a fully convolutional regression neural network algorithm based on Python language and Tensorflow open source architecture. The algorithm mainly consists of three parts: encoder, channel attention and spatial attention, and decoder; S402, based on the mean square error, a loss function MSE_loss is designed for the image regression neural network algorithm, and its basic formula is as follows: ； in, is the number of samples for pixel value, is subordinate to The true label of the sample, 0 or 1, is subordinate to The predicted label value of the sample; S403, using Adam as the optimizer, R 2 The cosine annealing scheme is used to optimize the learning rate. In the initial stage of training, the learning rate is set to a larger value so that it can quickly converge to the local minimum. As the training steps proceed, the learning rate gradually decreases according to the curve of the cosine function, so that the model can adjust the parameters more finely when it is close to the optimal solution, avoiding falling into the local minimum too early and being unable to jump out. The loss function does not change in every 10 epochs, and the learning rate is cosine transformed.

7. The global scale strong earthquake induced landslide spatial distribution probability intelligent prediction method according to claim 1 is characterized by: The S5 comprises the following steps: S501, according to the ignoring edge prediction method, overlappingly cropping TIFF and adopting the ignoring edge strategy when stitching, directly inputting a large TIFF file of an independent geographic space into the model for result prediction, and the ignoring edge prediction method is written by the GDAL library; S502. Based on computer vision technology and statistical learning strategies, adaptive threshold division is performed on the predicted results of the spatial distribution probability of strong earthquake-induced landslides to divide them into extremely high probability areas, high probability areas, medium probability areas, low probability areas and extremely low probability areas of landslides.

8. The method for intelligent prediction of spatial distribution probability of global-scale strong earthquake-induced landslides according to claim 6 is characterized by: The encoder comprises: Convolutional layer, size 3×3, pooling layer, size 3×3 and residual layer, size 1×1, 3 convolution operations and 2 pooling operations; CBAM consists of two parts: channel attention layer and spatial attention layer. Channel attention includes: Global average pooling with size 1×1, fully connected layers and normalization layers; The spatial attention consists of a convolutional layer with size 3×3, a global average pooling layer with size 1×1, a fully connected layer, and a normalization layer.

9. The method for intelligent prediction of spatial distribution probability of landslides induced by global-scale strong earthquakes according to claim 6 is characterized by: The decoder mainly includes: The deconvolution layer is of size 3×3, the upsampling layer is of size 3×3 and the residual layer is of size 1×1, which performs 3 deconvolution operations and 2 upsampling operations.

10. The method for intelligent prediction of spatial distribution probability of global-scale strong earthquake-induced landslides according to claim 6, characterized in that: In the model training process in S403, the batch_size parameter is set to 24, the epochs parameter is set to 400, the initial learning rate parameter is set to 0.00001, and the pre-trained model is not used.

Images