Earthquake landslide susceptibility evaluation method and storage medium based on semi-supervised incremental strategy

Abstract

The present invention relates to the technical field of earthquake landslide susceptibility assessment, and specifically discloses an earthquake landslide susceptibility assessment method and storage medium based on a semi-supervised incremental strategy. The present invention cleverly combines the core advantages of data-driven and physical mechanics methods, and constructs a set of earthquake landslide susceptibility assessment methods based on semi-supervised incremental learning strategies. The present invention is particularly focused on solving the difficult problems of limited sample number and real-time data processing in earthquake environments. It can mine valuable information under limited data conditions, and continuously integrate new data to upgrade the model. The present invention shows obvious advantages in earthquake landslide susceptibility prediction. It is worth noting that even in the early stage when data is scarce, this method can still provide relatively accurate prediction results, provide reliable support for initial decision-making in post-disaster rescue, and achieve instant and high-precision earthquake landslide prediction.

Description

Technical Field

[0001] The invention relates to the technical field of earthquake landslide susceptibility evaluation, and in particular to an earthquake landslide susceptibility evaluation method and a storage medium based on a semi-supervised incremental strategy. Background Art

[0002] my country is located at the intersection of the Pacific Rim Seismic Belt and the Himalayan Seismic Belt. Earthquake activities not only directly cause huge casualties and property losses, but also frequently cause landslide disasters, which exacerbate this destructive impact. Landslide disasters caused by earthquakes seriously threaten people's lives and safety, infrastructure stability, and the effective implementation of post-disaster rescue work. Therefore, timely clarifying the stability of natural slopes under the action of earthquakes and potential landslide risk areas has a strategic significance that cannot be ignored for formulating scientific prevention and control measures and effectively reducing the impact of geological disasters.

[0003] Currently, there are two common methods for evaluating the susceptibility of earthquake landslides: data-driven methods and physical mechanics methods. The data-driven method, especially with the help of machine learning methods, has demonstrated broad practical value and remarkable results in large-scale regional landslide susceptibility assessments with its excellent pattern recognition capabilities and deep correlation mining capabilities. However, a major limitation of this method is its high dependence on landslide databases. The suddenness and unpredictability of earthquakes themselves lead to obstacles to instant analysis of this method when facing earthquake emergencies. In contrast, the physical mechanics method directly evaluates the stability of slopes under earthquakes based on the principles of geomechanics, without the support of historical landslide data, and is theoretically more suitable for rapid response needs after earthquakes. However, this method faces the problem of insufficient parameter accuracy when applied on a large scale, which affects the prediction accuracy and practicality.

[0004] Although there are many research methods for earthquake landslide susceptibility, there is still a lack of a comprehensive optimization method that can adapt quickly and predict with high accuracy to meet the specific requirements of immediate emergency response after earthquake disasters. Therefore, there is an urgent need for a new generation of earthquake landslide susceptibility evaluation method that can efficiently integrate the advantages of data-driven and physical mechanics, overcome their respective limitations, and optimize the design specifically for earthquake emergency response needs. Summary of the invention

[0005] To solve the problems existing in the prior art, the present invention provides an earthquake landslide susceptibility evaluation method and storage medium based on a semi-supervised incremental strategy, which can achieve instant and high-precision earthquake landslide prediction and solve the problems mentioned in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for evaluating earthquake landslide susceptibility based on a semi-supervised incremental strategy, comprising the following steps:

[0007] S1. Obtain data on the target earthquake-affected area, including basic geological data, hydrogeological data, seismic data, and data on human engineering activities;

[0008] S2. Select the appropriate Newmark permanent displacement calculation model according to the type of earthquake-causing fault;

[0009] S3. Calculate the slope safety factor Fs and critical acceleration a based on the data collected in step S1. c , and the critical acceleration a is obtained c Then the permanent displacement value Dn is calculated using the Newmark permanent displacement calculation model, and the initial landslide susceptibility distribution map is formed according to the Dn value.

[0010] S4. Identify and mark the extremely high and extremely low landslide susceptibility areas from the initial landslide susceptibility distribution map, randomly extract a small number of representative potential disaster points in the extremely high susceptibility area; extract a large number of representative potential non-disaster points in the extremely low susceptibility area to form a basic disaster sample set;

[0011] S5. Based on the deep integration of on-site feedback, remote sensing interpretation and drone aerial photography data, the landslide disaster points are accurately identified and integrated with the basic disaster sample set to form a new disaster sample set;

[0012] S6. Use the Pearson correlation coefficient PCC to quantitatively evaluate and screen the landslide influencing factors. After eliminating the influencing factors with high correlation, extract the influencing factor values ​​of the disaster sample points to form a new landslide training set.

[0013] S7, select the incremental algorithm model, and use the Bayesian optimization algorithm to select the best hyperparameters including learning rate, batch size, and number of layers, and import the new disaster sample set into the model for training and learning;

[0014] S8. Use the trained model to predict the entire target earthquake-stricken area and obtain the landslide susceptibility probability value. Use the natural breakpoint method to divide the landslide susceptibility probability value into five susceptibility levels, and finally obtain the overall spatial distribution of earthquake landslide hazards in the disaster area.

[0015] Preferably, in step S1, the basic geological data and hydrogeological data include: topography, slope, effective cohesion of rock and soil, effective friction angle, bulk density, thickness of potential landslide body and proportion of saturated part in landslide body; the earthquake data include: magnitude, epicenter location, depth, peak acceleration of seismic motion, peak velocity of seismic motion and Arias intensity data; the human engineering activity data include: surface of human activities and road distribution.

[0016] Preferably, in step S2, selecting a suitable Newmark permanent displacement calculation model according to the type of seismogenic fault specifically includes the following:

[0017] S21. Strike-slip faults are mainly horizontal dislocations. They are identified by observing linear geological structures and lateral displacements during earthquakes. Strike-slip faults often show significant horizontal stress release characteristics. When the seismogenic fault type is a strike-slip fault, select the following model:

[0018]

[0019] S22. Non-strike-slip faults include reverse faults and normal faults. Their main feature is that there is an obvious vertical displacement of rock blocks on both sides of the fault. In reverse faults, the hanging plate moves upward relative to the footplate, while in normal faults, the footplate moves upward relative to the hanging plate. The key to identifying non-strike-slip faults is to identify the fault zone on the surface, the abnormal terrain undulation, and the vertical displacement information reflected in the seismic waveform. When the seismogenic fault type is a non-strike-slip fault, select the following model:

[0020]

[0021] S23. The characteristic of pulse type near fault is the instantaneous release of energy during the earthquake, which is manifested as strong pulse fluctuations in the seismic waveform. The basis for identification depends on seismic waveform analysis, especially looking for waveforms with obvious transient characteristics in seismic records and stress release patterns reflected in the focal mechanism solution. When the seismogenic fault type is pulse type near fault, the following model is selected:

[0022]

[0023] Where PGA is the peak acceleration of the earthquake, PGV is the peak velocity of the earthquake, Ia is the Arias intensity, and a c is the critical acceleration.

[0024] Preferably, in step S3, the critical acceleration a c The calculation formula of slope safety factor Fs is as follows:

[0025] a c =(Fs-1)gsinα;

[0026]

[0027] In the formula, a c is the critical acceleration; g is the acceleration caused by the earth's gravity; α is the angle at which the center of mass of the potential landslide block starts to move from the horizontal direction, that is, the slope; Fs is the slope safety factor; c is the effective cohesion of the rock and soil (kPa); γ is the weight of the rock and soil (kN / m 3);t is the potential sliding body thickness (m); is the effective internal friction angle of the rock and soil mass (°); m is the ratio of the saturated part of the potential sliding mass to the total sliding mass thickness; γ w It is the groundwater concentration.

[0028] Preferably, in step S3, the specific criteria for forming the initial landslide susceptibility distribution map according to the Dn value division are as follows:

[0029] ① When the Dn value is <1 cm, it indicates an extremely low-risk area;

[0030] ②1≤Dn value <5 cm, indicating a low-risk area;

[0031] ③5≤Dn value <10 cm indicates a moderately susceptible area;

[0032] ④When 10≤Dn value <15 cm, it indicates a high-risk area;

[0033] ⑤When the Dn value is ≥15 cm, it indicates an extremely high incidence area.

[0034] Preferably, in step S6, the landslide influencing factors are nine influencing factors, specifically including: altitude, slope aspect, terrain undulation, land use classification, normalized difference vegetation index NDVI, rainfall, distance from river, distance from road and distance from fault; the quantitative evaluation and screening specifically eliminates the influencing factors with PCC value ≥ 0.8, that is, highly correlated factors.

[0035] Preferably, in step S7, the incremental algorithm model is an incremental support vector machine model ILSVM or an incremental convolutional neural network model ILCNN.

[0036] Preferably, in step S8, the natural breakpoint method is used to divide the landslide susceptibility probability value into five susceptibility levels, specifically referring to: by calculating the variance minimization objective function of the data distribution, the optimal breakpoint position is determined, and then the landslide susceptibility probability value is divided into five susceptibility levels: extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

[0037] On the other hand, to achieve the above-mentioned purpose, the present invention also provides the following technical solution: a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the earthquake landslide susceptibility assessment method based on a semi-supervised incremental strategy.

[0038] The beneficial effects of the present invention are as follows: the present invention skillfully combines the core advantages of the two major methods of data-driven and physical mechanics, and constructs a set of earthquake landslide susceptibility evaluation methods based on a semi-supervised incremental learning strategy. The present invention particularly focuses on solving the difficult problems of limited sample number and real-time data processing in earthquake environments. It can mine valuable information under limited data conditions, and continuously integrate new data to upgrade the model. The present invention shows obvious advantages in earthquake landslide susceptibility prediction. It is worth noting that even in the early stages when data is scarce, this method can still provide relatively accurate prediction results, providing reliable support for initial decision-making in post-disaster rescue. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The figure is a schematic diagram of the earthquake landslide susceptibility assessment process based on the semi-supervised incremental strategy in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] 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.

[0041] See also Figure 1 The present invention provides a technical solution: a method for evaluating the susceptibility of earthquake landslides based on a semi-supervised incremental strategy, comprising the following steps:

[0042] S1. Obtain data on the target earthquake-stricken area, including basic geological data, hydrogeological data, seismic data, and data on human engineering activities.

[0043] The basic geological data and hydrogeological data include: topography, slope, effective cohesion of rock and soil, effective friction angle, bulk density, thickness of potential landslide body and the proportion of saturated part in the landslide body; the earthquake data include: magnitude, epicenter location, depth, peak acceleration of seismic motion, peak velocity of seismic motion and Arias intensity data; the human engineering activity data include: distribution of human activity surface and roads.

[0044] S2. Select the appropriate Newmark permanent displacement calculation model based on the type of earthquake fault. Specifically, it includes the following:

[0045] S21. Strike-slip faults are mainly horizontal displacements and can be identified by observing linear geological structures and lateral displacements during earthquakes. In addition, in seismic wave records, strike-slip faults often show significant horizontal stress release characteristics. When the seismogenic fault type is a strike-slip fault, select the following model:

[0046]

[0047] S22. Non-strike-slip faults include reverse faults and normal faults. Their main feature is that there is a significant vertical displacement of rock blocks on both sides of the fault. In a reverse fault, the upper plate moves upward relative to the lower plate, which is common in plate compression areas; while in a normal fault, the lower plate moves upward relative to the upper plate, which often occurs in a crustal extension environment. The key to identifying non-strike-slip faults is to identify the fault zones on the surface, the anomalies of the terrain, and the vertical displacement information reflected in the seismic waveform; when the type of earthquake-causing fault is a non-strike-slip fault, select the following model:

[0048]

[0049] S23. The characteristic of pulse-type near-fault is the instantaneous release of energy during an earthquake, which is manifested as a strong pulse fluctuation in the seismic waveform. The basis for judgment mainly relies on seismic waveform analysis, especially looking for waveforms with obvious transient characteristics in seismic records, and stress release patterns reflected in the focal mechanism solution. When landslide events are mainly concentrated in areas within 20 kilometers of the fault, this phenomenon becomes a typical sign of pulse-type near-fault, suggesting the direct impact of fault activity on the surrounding environment and the high incidence of geological disasters caused by it; when the earthquake-causing fault type is a pulse-type near-fault, select the following model:

[0050]

[0051] Where PGA is the peak acceleration of the earthquake, PGV is the peak velocity of the earthquake, Ia is the Arias intensity, and a c is the critical acceleration.

[0052] S3. Calculate the slope safety factor Fs and critical acceleration a based on the data collected in step S1. c , and the critical acceleration a is obtained c The Newmark permanent displacement calculation model is then used to quickly analyze the impact of earthquakes on slope stability, calculate the permanent displacement value Dn, and form an initial landslide susceptibility distribution map based on the Dn value.

[0053] Furthermore, the critical acceleration a c The calculation formula of slope safety factor Fs is as follows:

[0054] a c =(Fs-1)gsinα;

[0055]

[0056] In the formula, a cis the critical acceleration; g is the acceleration caused by the earth's gravity; α is the angle at which the center of mass of the potential landslide block starts to move from the horizontal direction, that is, the slope; Fs is the slope safety factor; c is the effective cohesion of the rock and soil (kPa); γ is the weight of the rock and soil (kN / m 3 ) ; t is the potential sliding body thickness (m); is the effective internal friction angle of the rock and soil mass (°); m is the ratio of the saturated part of the potential sliding mass to the total sliding mass thickness; γ w It is the groundwater concentration.

[0057] The above formula can also be used to obtain a preliminary data set of slope safety factors. Check the area where Fs < 1 and replace the Fs in this area with the area where Fs > 1 with the area where the safety factor is 1.01. Then merge the area where Fs > 1 with the area where the safety factor is 1.01 to obtain the revised slope safety factor distribution map. Finally, based on the revised Fs value, the critical acceleration a is calculated. c Distribution situation.

[0058] Furthermore, the specific criteria for forming the initial landslide susceptibility distribution map based on the Dn value are as follows:

[0059] ① When the Dn value is <1 cm, it indicates an extremely low-risk area;

[0060] ②1≤Dn value <5 cm, indicating a low-risk area;

[0061] ③5≤Dn value <10 cm indicates a moderately susceptible area;

[0062] ④When 10≤Dn value <15 cm, it indicates a high-risk area;

[0063] ⑤When the Dn value is ≥15 cm, it indicates an extremely high incidence area.

[0064] S4. Identify and mark the areas with extremely high and extremely low landslide susceptibility from the initial landslide susceptibility distribution map, randomly extract a small number of representative potential disaster points in the extremely high susceptibility areas; and extract a large number of representative potential non-disaster points in the extremely low susceptibility areas to form a basic disaster sample set.

[0065] From the initial landslide susceptibility distribution map, the extremely high landslide susceptibility area, i.e., the area with Dn ≥ 15 cm calculated in step S3, is identified and marked, and N high-confidence landslide hazard points are extracted from it; from the extremely low susceptibility area, i.e., the area with Dn < 1 cm calculated in step S3, 5N high-confidence non-landslide hazard points are extracted from it to form a basic disaster sample set.

[0066] S5. Based on the deep integration of on-site feedback, remote sensing interpretation and drone aerial photography data, the landslide disaster points are accurately identified and integrated with the basic disaster sample set to form a new disaster sample set.

[0067] Through on-site surveys, rescue operations, and reporting of disaster feedback, we obtain first-hand disaster information, including the location and scope of landslide disasters, laying the foundation for sample updates. Subsequently, we call on high-resolution remote sensing image interpretation technology to conduct a detailed scan of the disaster area and accurately identify landslide disaster points. In addition, with the help of drone aerial photography technology, we take close-up, high-definition photos of suspected landslide areas to further verify and refine the remote sensing interpretation results and enhance the reliability and integrity of the samples. Finally, we deeply integrate on-site feedback, remote sensing interpretation, and drone aerial photography data, and systematically integrate them with the basic disaster sample set to generate a more detailed and accurate new disaster sample set.

[0068] S6. Pearson correlation coefficient PCC is used to quantitatively evaluate and screen landslide influencing factors. After eliminating the influencing factors with high correlation, the influencing factor values ​​of the disaster sample points are extracted to form a new landslide training set.

[0069] In step S6, the semi-supervised framework constructed by physical methods in the previous steps S2 to S4 obtained the permanent displacement value Dn, which is obtained by the slope gradient, stratum lithology and seismic parameters. Therefore, among the conditions considered in step S6, the influencing factors related to Dn are first eliminated, and then the nine influencing factors currently generally recognized are selected for comprehensive evaluation. The nine influencing factors specifically include: altitude, slope aspect, terrain relief, land use classification, normalized difference vegetation index NDVI, rainfall, distance from rivers, distance from roads and distance from faults. The Pearson correlation coefficient (PCC) is used to quantitatively evaluate and screen the factors. PCC is the ratio of the covariance of the pixel values ​​between two layers to the product of the standard deviation. Its value <0.3 indicates that the factors are not related, 0.3-0.5 indicates low correlation, 0.5-0.8 indicates moderate correlation, and ≥0.8 indicates high correlation. In order to prevent the model from overfitting or superposition of redundant information, strict statistical principles were followed, and the influencing factors with high correlation (PCC value ≥ 0.8) were eliminated to ensure the refinement and efficiency of the model. The influencing factor values ​​of the disaster sample points were extracted to form a new landslide training set.

[0070] S7. Select the incremental algorithm model and use the Bayesian optimization algorithm to select the optimal hyperparameters (such as learning rate, batch size, and number of layers), and import the new disaster sample set into the model for training and learning.

[0071] Select an efficient incremental algorithm model to achieve continuous optimization and updating of the landslide susceptibility model. The incremental algorithm model can update the existing model with new data points without retraining the entire model, which significantly improves learning efficiency and adaptability. Especially in the field of landslide susceptibility assessment, it can integrate new observation data in real time, such as the latest landslide location, scale and environmental factors. There is no need to train the model from scratch, but to gradually fine-tune it. The model continues to evolve over time to maintain or even improve the prediction performance. The incremental algorithm model can choose the incremental support vector machine model (Incremental Learning Support Vector Machine, ILSVM) and the incremental convolutional neural network model (Incremental Learning Convolutional Neural Network, ILCNN), among which ILSVM can be the first choice.

[0072] Furthermore, the decision function of ILSVM is as follows:

[0073] First, assume that a set of training data D = {(x1, y1), (x2, y2), ..., (x n ,y n )},y i ∈(-1,1). Using the linear equation w T X+b=0 defines its decision boundary. According to the distance from the point to the line, the classification interval distance is 2 / ||w||. In order to maximize the interval and consider the boundary constraints, the basic formula of SVM is obtained:

[0074]

[0075] Where: w is the normal vector; b is a scalar; m is the total number of training samples.

[0076] Introducing the insensitive loss function and Lagrangian function, we get the final linear fitting function:

[0077]

[0078] Where: a i , is the Lagrange multiplier; K(x i ,x) is the kernel function.

[0079] Furthermore, the decision function of ILCNN is as follows:

[0080] CNN is mainly composed of three parts: ① data input layer; ② a combination of n convolutional layers and pooling layers; ③ a fully connected multi-layer perceptron classifier (fully connected layer) and an output layer. The landslide influencing factor is used as input data, and the output layer is the corresponding landslide state. The convolution kernel of the convolution layer can be used to extract the primary features of the data. The output of the convolution operation is defined as follows:

[0081]

[0082] Where: C j is the output of the jth convolution kernel; f is the nonlinear activation function; v i is the input data of the convolution layer, * is the convolution operator; i is the spatial position of the convolution; k is the number of convolution kernels; w j and b j are weight and bias respectively.

[0083] The factor features are continuously extracted through the convolution-pooling operation, and the features are mapped to a low-dimensional space through the fully connected layer, so that the parameters between layers will be reduced accordingly. Ultimately, the purpose of reducing model complexity and improving computational efficiency is achieved. The widely used maximum pooling operation is used in the entire operation process, and the output of this operation is calculated as follows:

[0084]

[0085] Where: u(n,1) is the window function of the previous layer patch, a j is the maximum value in the patch. Next, these local representations extracted by convolution and pooling operations are reorganized through a fully connected layer. Finally, the fully connected layer is connected to the output layer, which consists of two neurons representing landslides and non-landslides. All parameters of the CNN layer are optimized using the back-propagation algorithm. The purpose of optimizing the parameters is to minimize the loss value calculated by the loss function, which is defined as follows:

[0086]

[0087] Where: m is the number of input landslide data, y i and The two variables represent the true label and predicted label of the i-th input sample respectively. The parameters are updated iteratively until the loss value converges.

[0088] Furthermore, in step S7, the Bayesian optimization algorithm is used to focus on optimizing the two core parameters of the kernel function of ILSVM - C value and gamma value. These two parameters directly affect the generalization ability and sensitivity to noise of the model, thereby ensuring that the ILSVM model maintains high classification accuracy while improving its adaptability to landslide data and computational efficiency; for the ILCNN model, the Bayesian optimization algorithm focuses on adjusting three key hyperparameters: MiniBatchSize (mini-batch size), InitialLearnRate (initial learning rate) and L2Regularization (L2 regularization coefficient). By optimizing MiniBatchSize, the best balance between computational efficiency and model convergence speed is sought; fine adjustment of InitialLearnRate ensures the stability and efficiency of the learning process; and optimization of L2Regularization helps to reduce model overfitting and improve generalization ability. The decision function of the Bayesian optimization algorithm is as follows:

[0089]

[0090] Where: x * is the best parameter combination for optimization; f(x) is the target optimization function; χ is the search space; φ is the probability function; μ(x) is the expectation; σ(x) is the variance; f(x + ) represents the optimal value of the objective function; ξ is a set of hyperparameters.

[0091] S8. Use the trained model to predict the entire target earthquake-stricken area and obtain the landslide susceptibility probability value. Use the natural breakpoint method to divide the landslide susceptibility probability value into five susceptibility levels, and finally obtain the overall spatial distribution of earthquake landslide hazards in the disaster area.

[0092] Furthermore, the natural breakpoint method is used to divide the landslide susceptibility probability values ​​into five susceptibility levels, specifically referring to: by calculating the variance minimization objective function of the data distribution, the optimal breakpoint position is determined, and then the landslide susceptibility probability values ​​are divided into five susceptibility levels: extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

[0093] In step S8, the landslide susceptibility probability value is calculated through model training. The susceptibility probability value essentially reflects the relative possibility of a landslide event in a certain area under given conditions. The larger the value, the higher the potential risk of landslide in the area. In order to express these probability values ​​more intuitively and scientifically, the natural breakpoint method is used for classification here. This method can identify the group boundaries that naturally exist in the data, thereby maximizing the differences between different levels while maintaining the internal consistency of each level. In specific operations, this method determines the optimal breakpoint position by calculating the variance minimization objective function of the data distribution, and then divides the landslide susceptibility probability value into five levels: extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area, each level represents a different level of landslide risk.

[0094] According to the spatial distribution of earthquake landslide disasters in the disaster-stricken areas, reference suggestions and data support are provided for disaster rescue, and the landslide sample set is updated according to the on-site disaster feedback.

[0095] The spatial distribution characteristics of earthquake landslide disasters in the disaster-stricken area obtained in step S8 are used to further explore the concentrated areas and impact range of disasters, providing accurate reference suggestions and solid data support for front-line rescue and relief work. At the same time, combined with the real-time feedback of the disaster situation on the scene, the landslide sample set is updated in time to ensure the timeliness and accuracy of the database. The disaster feedback information collected in this step is used as the key input for optimizing and iterating step S5. Through the latest on-site feedback, the landslide sample set can be updated and expanded in time. This updated sample set will trigger the re-execution of the entire partial process, including extracting disaster sample point information to form a new landslide training set (step S6), Bayesian optimization of hyperparameters (step S7), deepening training of the susceptibility model, calculation and output of results (step S8), and providing accurate guidance and disaster feedback for on-site rescue operations again. This closed-loop mechanism ensures that the analysis model can continuously adapt to changing environmental conditions and continuously improve prediction accuracy and practicality.

[0096] Based on the same inventive concept as the above method embodiment, the present application embodiment further provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement

[0097] The earthquake landslide susceptibility evaluation method based on a semi-supervised incremental strategy is described.

[0098] The present invention is particularly focused on solving the problem of limited sample quantity and real-time data processing in earthquake environments. It can mine valuable information under limited data conditions and continuously integrate new data to upgrade the model. The present invention shows obvious advantages in earthquake landslide susceptibility prediction.

[0099] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for evaluating earthquake landslide susceptibility based on a semi-supervised incremental strategy, characterized in that: The steps include: S1. Obtain data on the target earthquake-affected area, including basic geological data, hydrogeological data, seismic data, and data on human engineering activities; S2. Select an appropriate Newmark permanent displacement calculation model based on the type of earthquake fault; specifically, the following: S21. Strike-slip faults are mainly horizontal dislocations. They are identified by observing linear geological structures and lateral displacements during earthquakes. Strike-slip faults often show significant horizontal stress release characteristics. When the seismogenic fault type is a strike-slip fault, select the following model: S22. Non-strike-slip faults include reverse faults and normal faults. Their main feature is that there is an obvious vertical displacement of rock blocks on both sides of the fault. In reverse faults, the hanging plate moves upward relative to the footplate, while in normal faults, the footplate moves upward relative to the hanging plate. The key to identifying non-strike-slip faults is to identify the fault zone on the surface, the abnormal terrain undulation, and the vertical displacement information reflected in the seismic waveform. When the seismogenic fault type is a non-strike-slip fault, select the following model: S23. The characteristic of pulse type near fault is the instantaneous release of energy during the earthquake, which is manifested as strong pulse fluctuations in the seismic waveform. The basis for identification depends on seismic waveform analysis, especially looking for waveforms with obvious transient characteristics in seismic records and stress release patterns reflected in the focal mechanism solution. When the seismogenic fault type is pulse type near fault, the following model is selected: Where PGA is the peak acceleration of the earthquake, PGV is the peak velocity of the earthquake, Ia is the Arias intensity, and a c is the critical acceleration; S3. Calculate the slope safety factor Fs and critical acceleration a based on the data collected in step S1. c , and the critical acceleration a is obtained c The permanent displacement value Dn is then calculated using the Newmark permanent displacement calculation model, and the initial landslide susceptibility distribution map is formed according to the Dn value. The specific criteria for forming the initial landslide susceptibility distribution map according to the Dn value are as follows: ① When the Dn value is <1 cm, it indicates an extremely low-risk area; ②1≤Dn value <5 cm, indicating a low-risk area; ③5≤Dn value <10 cm indicates a moderately susceptible area; ④When 10≤Dn value <15 cm, it indicates a high-risk area; ⑤When the Dn value is ≥15 cm, it indicates an extremely high-risk area; S4. Identify and mark the extremely high and extremely low landslide susceptibility areas from the initial landslide susceptibility distribution map, randomly extract a small number of representative potential disaster points in the extremely high susceptibility area; extract a large number of representative potential non-disaster points in the extremely low susceptibility area to form a basic disaster sample set; S5. Based on the deep integration of on-site feedback, remote sensing interpretation and drone aerial photography data, the landslide disaster points are accurately identified and integrated with the basic disaster sample set to form a new disaster sample set; S6. Use the Pearson correlation coefficient PCC to quantitatively evaluate and screen the landslide influencing factors. After eliminating the influencing factors with high correlation, extract the influencing factor values ​​of the disaster sample points to form a new landslide training set. S7, select the incremental algorithm model, and use the Bayesian optimization algorithm to select the best hyperparameters including learning rate, batch size, and number of layers, and then import the new disaster sample set into the model for training and learning; S8. Use the trained model to predict the entire target earthquake-stricken area and obtain the landslide susceptibility probability value. Use the natural breakpoint method to divide the landslide susceptibility probability value into five susceptibility levels, and finally obtain the overall spatial distribution of earthquake landslide hazards in the disaster area.

2. The earthquake landslide susceptibility assessment method based on a semi-supervised incremental strategy according to claim 1 is characterized in that: In step S1, the basic geological data and hydrogeological data include: topography, slope, effective cohesion of rock and soil, effective friction angle, bulk density, thickness of potential landslide body and the proportion of saturated part in the landslide body; the earthquake data include: magnitude, epicenter location, depth, peak acceleration of seismic motion, peak velocity of seismic motion and Arias intensity data; the human engineering activity data include: distribution of human activity surface and roads.

3. The earthquake landslide susceptibility assessment method based on a semi-supervised incremental strategy according to claim 1 is characterized in that: In step S3, the critical acceleration a c The calculation formula of slope safety factor Fs is as follows: a c =(Fs-1)gsinα; In the formula, a c is the critical acceleration; g is the acceleration caused by the earth's gravity; α is the angle at which the center of mass of the potential landslide block starts to move from the horizontal direction, that is, the slope; Fs is the slope safety factor; c is the effective cohesion of the rock and soil; γ is the weight of the rock and soil; t is the thickness of the potential sliding body; is the effective internal friction angle of the rock and soil; m is the ratio of the saturated part of the potential sliding mass to the total sliding mass thickness; γ w It is the groundwater concentration.

4. The earthquake landslide susceptibility assessment method based on a semi-supervised incremental strategy according to claim 1 is characterized in that: In step S6, the landslide influencing factors are nine influencing factors, including: altitude, slope aspect, terrain undulation, land use classification, normalized difference vegetation index NDVI, rainfall, distance from river, distance from road and distance from fault; the quantitative evaluation and screening specifically eliminates the influencing factors with PCC value ≥ 0.8, that is, highly correlated factors.

5. The earthquake landslide susceptibility assessment method based on a semi-supervised incremental strategy according to claim 1 is characterized in that: In step S7, the incremental algorithm model is an incremental support vector machine model ILSVM or an incremental convolutional neural network model ILCNN.

6. The earthquake landslide susceptibility assessment method based on a semi-supervised incremental strategy according to claim 1 is characterized in that: In step S8, the natural breakpoint method is used to divide the landslide susceptibility probability value into five susceptibility levels, specifically referring to: by calculating the variance minimization objective function of the data distribution, the optimal breakpoint position is determined, and then the landslide susceptibility probability value is divided into five susceptibility levels: extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

7. 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 method for evaluating earthquake landslide susceptibility based on a semi-supervised incremental strategy as described in any one of claims 1 to 6 is implemented.

Images