Co-seismic landslide prediction method and system fusing physical information and landslide scale constraints

Abstract

The application relates to the field of coseismic landslide prediction, and discloses a coseismic landslide prediction method fusing physical information and landslide scale constraints, which comprises the following steps: a. acquiring landscape data and seismic motion parameters of a research area, and taking a slope unit as a basic analysis unit; b. based on the landscape data and seismic motion parameters, calculating physical information parameters of each slope unit by adopting a physical model based on the law of conservation of energy, wherein the physical information parameters are used for representing the dynamic characteristics of materials in the slope; and c. adopting a staged neural network model for prediction. By adopting an improved Heim energy line model in the calculation of physical information, the PGA of seismic motion parameters is integrated into the energy conservation equation, so that the velocity field and other physical information reflecting the internal dynamic characteristics of each slope unit can be calculated without acquiring regional rock-soil mechanical parameters.

Description

Technical Field

[0001] This invention relates to the field of coseismic landslide prediction, specifically to a method and system for predicting coseismic landslides that integrates physical information with landslide scale constraints. Background Technology

[0002] Coseismic landslides are among the most destructive types of earthquake-induced secondary disasters, especially near the epicenter, steep mountainous areas, and rupture zones, severely impacting human safety and geomorphological stability. Therefore, reliable landslide prediction can provide crucial emergency decision-making solutions after an earthquake, assisting in the decision-making process for landslide risk reduction. Since the concept of landslide prediction was introduced, the literature has widely recognized the key role of data-driven landslide sensitivity in landslide prediction. Landslide sensitivity refers to the potential likelihood of a landslide occurring in a given area under the influence of natural and human factors. This can also be understood as the probability of a landslide occurring in a given unit. For a long time afterward, this research dominated, primarily because it could predict the location of future landslides, providing insights for landslide relief efforts.

[0003] Predicting regional coseismic landslide susceptibility using existing technologies typically faces numerous challenges. On one hand, while traditional statistical models and machine learning methods are widely used, their prediction processes are often considered "black boxes," lacking clear physical mechanisms and resulting in poor model interpretability. Furthermore, the reliability of predictions is difficult to guarantee when extrapolated to new regions. On the other hand, while physical mechanics-based analytical methods have clear physical meanings, they often require detailed, spatially heterogeneous geotechnical parameters. These parameters are difficult to obtain accurately at the regional scale using conventional methods, and the investigation costs are high and uncertain, significantly limiting the widespread applicability of physical models in practical work. Moreover, most existing prediction models treat the probability of landslide occurrence (susceptibility) and the potential scale of landslides (magnitude) as two independent problems, neglecting the inherent physical coupling between them. This separation leads to predictions that fail to reflect the impact of disaster scale, potentially misjudging areas with frequent small-scale landslides as extremely high-risk areas, thus causing biased risk perception. Meanwhile, existing technologies often employ simple feature overlay methods when integrating multiple influencing factors, making it difficult to effectively distinguish the contribution weights of different information (such as physical information and scale information) to the final susceptibility judgment, thus limiting further improvements in model performance. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a coseismic landslide prediction method and system that integrates physical information and landslide scale constraints. This solves the problems of insufficient interpretability of prediction models due to unclear physical mechanisms, limited applicability of physical analysis methods due to excessive reliance on hard-to-obtain soil and rock parameters, and biased risk perception caused by assessing landslide susceptibility and potential scale separately and ignoring their inherent coupling relationship.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The first aspect of this invention provides a coseismic landslide prediction method that integrates physical information and landslide scale constraints. This method improves the reliability and physical interpretability of the prediction results by constructing a joint modeling framework that couples landslide scale prediction with susceptibility prediction and introducing physical information that does not require detailed geotechnical parameters.

[0007] In one embodiment, the method includes the following steps:

[0008] a. Data Acquisition and Unit Division Steps: Acquire landscape data and seismic parameters of the study area. The landscape data includes topographic data, geological data, hydrological data, and human element data; the seismic parameter is peak ground acceleration (PGA). Simultaneously, the study area is divided into multiple slope units, which serve as the basic spatial units for all subsequent analyses and calculations.

[0009] b. Physical Information Calculation Steps: For each slope unit, a physical model based on the law of conservation of energy is used to calculate a set of physical information parameters. This physical model is an improved Heim energy line model, which, based on characterizing the conversion of gravitational potential energy and kinetic energy, further incorporates the influence of the peak ground acceleration (PGA) parameters obtained in step a to simulate material movement under co-seismic conditions. The calculated physical information parameters, such as travel speed, travel distance, and travel angle, are used to quantitatively characterize the dynamic properties of potential landslide material on the slope unit. This step does not require input of regional geotechnical parameters, such as the cohesion or internal friction angle of the geotechnical mass.

[0010] c. Phased Joint Prediction Step: A phased neural network model is used to comprehensively predict landslide hazards. This neural network model can be a Physical Information Neural Network (PINN) based on a Residual Network (ResNet) architecture. This step specifically includes:

[0011] i. First Stage: Landslide Scale Prediction. Using the landscape data obtained in step a and the physical information parameters calculated in step b as model input, the landslide scale information for each slope unit is predicted and output through feature extraction and mapping via a neural network. In a specific implementation, this landslide scale information is a predicted landslide area value conforming to a log-Gaussian distribution.

[0012] ii. Second Stage: Landslide Susceptibility Prediction. The landslide scale information output from the first stage is used as a new physical constraint feature, which, along with the original landscape data and physical information parameters, serves as the input to the model in this stage. A neural network is used to process the data and predict and output the final landslide susceptibility for each slope unit. In a specific implementation, this landslide susceptibility is a probability value that follows a Bernoulli distribution.

[0013] In a more specific embodiment, in the second stage, a feature fusion attention mechanism can also be used to effectively integrate landslide scale information features with other input features. This mechanism uses a self-attention network to calculate and assign weights to landslide scale information features and other features respectively, adaptively adjusting the influence of each feature in susceptibility assessment, and finally generating a fused feature representation for subsequent susceptibility probability prediction.

[0014] A second aspect of the present invention provides a coseismic landslide prediction system that integrates physical information and landslide scale constraints, the system being configured to perform the aforementioned method. The system includes:

[0015] The data acquisition module is configured to acquire landscape data of the study area and ground motion parameters, represented by peak ground acceleration (PGA), from external data sources.

[0016] The physical information calculation module, which is connected to the data acquisition module, is configured to use the slope element as the basic unit and call an improved Heim energy line model. This model integrates seismic parameters to calculate physical information parameters characterizing the dynamic properties of slope materials, such as velocity and travel distance.

[0017] The data aggregation module, which is connected to the aforementioned module, is configured to uniformly summarize and format the acquired raw data and the calculated physical information parameters at the scale of the slope unit to generate an input dataset suitable for the neural network model.

[0018] The joint modeling and prediction module, the core computing unit of the system, integrates a phased physical information neural network. This module receives data processed by the data aggregation module and is configured to perform the following functions:

[0019] Run a scale prediction program, which corresponds to the first prediction stage in the method, to predict and generate landslide scale information for each slope unit based on the input dataset.

[0020] Run a susceptibility prediction program, corresponding to the second prediction stage of the method. This program receives the original input dataset and simultaneously receives landslide scale information generated by the scale prediction program. It uses this scale information as a constraint input, performs comprehensive analysis, and finally predicts and outputs the landslide susceptibility result for each slope unit.

[0021] This invention provides a method and system for predicting coseismic landslides by integrating physical information and landslide scale constraints. It has the following beneficial effects:

[0022] 1. This invention employs an improved Heim energy line model in physical information calculation, integrating the ground motion parameter PGA into the energy conservation equation. This enables the calculation of velocity field physical information reflecting the intrinsic dynamic characteristics of each slope element without requiring the acquisition of regional soil and rock mechanical parameters. This method reduces dependence on hard-to-obtain data and significantly enhances the model's universality and operability in different geological backgrounds.

[0023] 2. This invention constructs a phased neural network model, placing landslide scale prediction and landslide susceptibility prediction within a unified framework. The landslide scale information predicted in the first phase is used as a constraint term for the susceptibility prediction in the second phase, thus establishing an inherent physical correlation between the two. This design ensures that the final susceptibility assessment result is reasonably constrained by the potential disaster scale, effectively avoiding the problem of traditional models misclassifying small-scale landslide-prone areas as extremely high-risk areas, and improving the physical consistency and reliability of the prediction results.

[0024] 3. This invention uses slope units as objective analysis carriers and designs a feature fusion attention mechanism to dynamically fuse original features containing physical information with predicted scale features. This allows the model to adaptively learn the contribution weight of scale information to susceptibility judgment. Compared to simple feature stitching, this method can more intelligently handle the complex relationships between different features, achieving more refined feature interaction learning on analysis units that better fit the real terrain, thereby improving the model's learning efficiency and prediction accuracy.

[0025] 4. This invention, through its phased prediction architecture, can simultaneously generate two interconnected core products: a landslide scale prediction map and a landslide susceptibility prediction map. This provides a more comprehensive information dimension for disaster risk assessment. Users can not only learn about the spatial probability (susceptibility) of landslides, but also understand their potential magnitude (scale). This dual information output greatly enhances the practical value of the prediction results, providing more targeted scientific decision support for disaster prevention and mitigation, engineering planning, and emergency management.

[0026] 5. This invention combines physical model calculations based on the law of conservation of energy with a data-driven model based on deep learning, and uses objective slope units for spatial division, to achieve a systematic and multi-faceted analysis of coseismic landslide problems. This method utilizes physical laws to enhance the interpretability and extrapolation ability of the model, while leveraging the powerful nonlinear fitting capabilities of deep learning to uncover complex patterns in the data. Simultaneously, it avoids subjective analysis unit division, making the entire prediction process more objective and rigorous, and the resulting outcomes more convincing. Attached Figure Description

[0027] Figure 1 This is a structural framework diagram of the PINN model in an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of the prediction method of the present invention;

[0029] Figure 3 This is a schematic diagram illustrating the principle of the improved Heim energy line model of the present invention;

[0030] Figure 4 This is a schematic diagram of the phased physical information neural network structure of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example:

[0033] Please see the appendix Figure 1 - Appendix Figure 4 This invention provides a method for predicting coseismic landslides by integrating physical information and landslide scale constraints, including:

[0034] S1. In this embodiment, step S1 is data preparation and slope unit division. The purpose of this step is to prepare a complete, spatially aligned, multi-source input dataset that covers topography, geological environment, earthquake triggering factors, and historical disaster records for subsequent physical information calculation and joint prediction model construction, and to establish a unified geospatial analysis unit that can objectively reflect the real terrain morphology.

[0035] First, multi-source heterogeneous data is acquired and preprocessed. This multi-source heterogeneous data is used to comprehensively characterize the disaster-prone environment and disaster-causing factors within the study area.

[0036] The data specifically includes landslide inventory data. In one implementation, this landslide inventory data can be obtained by collecting publicly available information such as disaster catalog reports and academic literature records related to earthquake events that have occurred in the study area. Preferably, it is combined with high-resolution remote sensing images acquired after the earthquake for secondary manual visual interpretation and verification. The remote sensing images can be historical images provided by WorldView series, GF series, or Google Earth. This process aims to create an accurate vectorized dataset containing the spatial location and geometric boundary range of each landslide body. This landslide inventory data is the fundamental basis for sample unit annotation (i.e., distinguishing stable units from unstable units) and extracting real landslide scale information in the subsequent model training stage of this invention, and its accuracy is directly related to the final performance of the model.

[0037] The data also includes a set of landscape data describing the background environment of the landslide. This set of landscape data covers several core aspects affecting slope stability:

[0038] Regarding topographic features, based on the Digital Elevation Model (DEM), a series of quantitative topographic parameters are derived using spatial analysis tools provided by professional Geographic Information System (GIS) platforms (GRASSGIS or ArcGIS). These parameters include, but are not limited to, slope, aspect, plane curvature, profile curvature, topographic relief, and topographic moisture index (TWI). These topographic features collectively characterize the geometry and hydrodynamic conditions of slope units from different dimensions, forming the basis for judging slope stability. Among them, the topographic moisture index (TWI), as a parameter characterizing the cumulative effect of topography on water flow, can be calculated based on the following formula:

[0039] ;

[0040] In the formula, The upstream catchment area is the length of a unit contour line. This refers to a local slope.

[0041] In terms of geological elements, the main components include engineering geological rock group maps and active fault distribution maps of the study area. Among them, lithological data reflects the material basis and inherent strength characteristics of the slope; while the distance to the faults characterizes the intensity of regional tectonic activity and the degree of rock mass fragmentation.

[0042] Other environmental factors may include hydrological factors, such as distance from river networks, and human activity factors, such as distance from road networks and land use / land cover type (LULC). These factors together constitute the external environmental conditions that influence slope stability.

[0043] The data further includes ground motion parameters. Preferably, peak ground acceleration (PGA) is used as the core physical indicator characterizing coseismic triggering intensity. Its acquisition method includes collecting strong ground motion records from seismic stations deployed within and around the study area, and performing necessary baseline correction and bandpass filtering preprocessing on the raw records. Subsequently, spatial interpolation methods such as Kriging interpolation or inverse distance weighted (IDW) interpolation are used to generate a continuous spatial distribution map of ground motion intensity covering the entire study area based on the PGA values ​​of discrete stations. The purpose of this step is to provide a quantitative input of the disaster-causing triggering intensity for each spatial location. Specifically, the PGA value extracted for each slope unit is denoted as follows: This parameter will serve as a key input to the improved Heim energy line model in subsequent step S2, directly used to calculate the work done by the seismic force. Its role in the energy balance equation is reflected as follows: This, in turn, affects the final velocity field calculation.

[0044] After completing the above data preparation, this step also requires dividing the study area into slope units. The slope unit (SU) is established as the unique and unified basic spatial unit for expressing all subsequent analysis, calculation, and prediction results in this invention. As an objective division method based on topography, the slope unit's boundary is composed of both catchment and watershed lines. Compared to regular grid units, it can more reasonably express the macroscopic geomorphological characteristics of a single slope, providing a more realistic analytical framework for subsequent physical information calculations and susceptibility assessments.

[0045] In one specific implementation, the division of slope units can be achieved by utilizing a module embedded in the open-source geographic information system software GRASSGIS, which automatically generates units based on preprocessed digital elevation model (DEM) data. This module uses a simulated surface water flow path algorithm to identify ridgelines and valley lines, thereby defining relatively independent slope units in the terrain. After the division is completed, preferably, a visual inspection and correction can be performed to remove units that are too small or too fragmented in shape, ensuring the rationality and effectiveness of the unit division.

[0046] Ultimately, step S1 outputs a complete set of preprocessed data layers with uniform spatial resolution, and a slope unit vector layer defining all basic analysis units within the study area. All data layers (including the aforementioned landscape data and seismic parameters) will be aggregated into each slope unit in subsequent step S3, forming the input feature vector required by the neural network model in subsequent step S4. Its composition can be represented as This data includes various landscape features and physical information characteristics for subsequent calculations. All data layers are strictly aligned spatially with the slope unit layer, laying a solid data foundation for subsequent steps.

[0047] S2. In this embodiment, step S2 involves calculating physical information parameters. The core technical task of this step is to calculate and generate a set of parameters that can quantitatively characterize the dynamic properties of potential material movement under co-seismic conditions for each slope unit divided in step S1, using a model based on physical laws. A significant feature of this step is that it eliminates the need to input geotechnical mechanical parameters such as cohesion and internal friction angle, which are difficult to obtain accurately at the regional scale, thereby effectively improving the feasibility and universality of this invention in different regions.

[0048] To achieve the above objectives, this embodiment preferably employs an improved Heim energy line model. This model is chosen as the basis for physical information calculation because it is based on the universal law of conservation of energy, providing a clear physical framework for describing the motion of landslides. The original Heim model simplifies the motion of a landslide as a rigid block sliding down a slope, with its energy conversion process primarily following the conversion of gravitational potential energy into kinetic energy.

[0049] In the original Heim energy line model (Kinetic load) In the original Heim energy line model, It is a kinetic load, defined as:

[0050] ;

[0051] Physical meaning:

[0052] Indicates the landslide mass at a certain point The ratio of kinetic energy to gravitational acceleration, expressed in meters, can be considered as the "equivalent height" of the landslide's kinetic energy.

[0053] It stems from the principle of energy conservation: when a landslide slides down from the top (highest point) of the slope, the potential energy lost (compared to the height difference) is... (Related) energy is converted into kinetic energy, and some of the energy is dissipated through friction.

[0054] It is a point The vertical distance relative to the energy line represents the remaining potential energy height.

[0055] Velocity formula:

[0056] According to the definition of kinetic energy Derivation of speed:

[0057] ;

[0058] here, The speed of the landslide at a certain point is reflected by calculating the difference between the position of the landslide body on the slope and the height of the energy line.

[0059] The function of energy lines:

[0060] The energy line is a straight line from the top of the landslide to the furthest point of deposition (tip), at an angle of... .

[0061] At a certain point The kinetic energy of a landslide is related to its vertical distance from the energy line. Proportional Usually by And frictional dissipation estimation. Summary: It is an intermediate parameter in the Heim model used to simplify kinetic energy calculation. It is based on the conversion of potential energy to kinetic energy and ignores the influence of other external forces (such as seismic forces).

[0062] 2. When discussing coseismic landslides, the original Helm model mentioned above neglected the issue of seismic forces (such as peak ground acceleration, PGA). Therefore, an improved formula was established, incorporating a seismic force term. The derivation logic of this formula is as follows:

[0063] The original Helm model's velocity formula assumes that the landslide's kinetic energy comes entirely from the conversion of gravitational potential energy. The velocity formula is:

[0064] ;

[0065] :point The vertical distance relative to the energy line represents the height of potential energy lost by the landslide body at that point. Potential energy per unit mass is converted into kinetic energy.

[0066] In coseismic landslides, the introduction of seismic force (PGA) provides an additional driving force as an external force, influencing the movement of the landslide mass. PGA can be decomposed into:

[0067] Components along the slope: ,in It is the angle between the slope and the PGA direction (usually assumed that the PGA is mainly horizontal acceleration, and the slope angle is...). ).

[0068] Effect: The work done by the seismic force along the slope increases the kinetic energy of the landslide. The work done by the seismic force can be expressed as:

[0069] ;

[0070] Landslide mass. The landslide mass from the starting point to the point The sliding distance along the slope. The contribution of seismic acceleration along the slope direction is equivalent to an additional "acceleration". "Distance" item.

[0071] 3. Improved Energy Conservation: Considering seismic forces, the total kinetic energy of the landslide comes not only from gravitational potential energy but also from the work done by the seismic forces. Therefore, the total energy contribution is:

[0072] ;

[0073] First item : The contribution of gravitational potential energy (original Heim model).

[0074] Second item The work done by the seismic force along the slope. The kinetic energy formula is:

[0075] ;

[0076] Divide by mass Thus, the kinetic energy per unit mass is obtained:

[0077] ;

[0078] Multiply both sides by 2 and take the square root to get the velocity:

[0079] .

[0080] Introduction :

[0081] Physical basis: Seismic force (PGA), as an external force, exerts additional acceleration along the slope direction during the movement of the landslide mass. Its contribution is similar to gravitational acceleration. However, the direction and size depend on the slope angle. and sliding distance .

[0082] Simplifying assumptions:

[0083] Assuming PGA is primarily horizontal acceleration, with the slope component as follows: .

[0084] Assume that the PGA is constant during the sliding process (in reality, the PGA is time-varying and requires further complex processing).

[0085] This represents the cumulative distance the landslide body slides along the slope, and the work done by the seismic force is proportional to this distance.

[0086] This formula constructs the geometry of the slope element (through...). and (Implication), sliding path (through) (reflected) and external earthquake triggering intensity (through) This reflects the direct physical connection between the internal material movement speed and the physical relationship between the material and the internal material movement speed.

[0087] In the specific calculation process, for each slope element, the highest and lowest points within the element are first determined based on its DEM data to define the energy line. Then, parameters characterizing the macroscopic geometric features of the element, such as the travel angle, horizontal travel distance, and vertical travel height, are calculated. Next, using the improved velocity formula described above, the potential velocity values ​​at different locations within the element are calculated.

[0088] Finally, the output of step S2 is to assign a set of physical information parameters to each slope element. These parameters include not only the aforementioned travel angle, travel distance, and travel height, but more importantly, statistics on all calculated velocity values ​​within the element. Preferably, the average velocity, velocity standard deviation, and maximum velocity can be used as a comprehensive representation of the overall dynamic characteristics of the element. These calculated physical information parameters will be aggregated in subsequent step S3 and used as physical constraint features, inputting them into the staged neural network model to enhance the model's understanding of the landslide physical process, thereby improving the reliability and interpretability of the final prediction results.

[0089] S3. In this embodiment, step S3 is data aggregation at the slope unit scale. This step is a key link connecting data preparation and model building. Its core purpose is to spatially unify and attribute-merge the landscape and seismic data from various sources and formats acquired and preprocessed in step S1, as well as the physical information parameters calculated in step S2. Ultimately, on the slope unit established in step S1 as the basic analysis carrier, a structured feature dataset is formed that can be directly read and processed by the subsequent neural network model.

[0090] Specifically, this step uses spatial analysis to calculate and assign a complete feature vector to each individual slope unit. This process employs targeted aggregation strategies based on the type of data variables to maximize the preservation of the effectiveness of the original information.

[0091] For continuous variables, such as slope, curvature, and topographic relief in topographic features, distance from rivers / roads in hydrological and human factors, and peak ground acceleration (PGA) in seismic parameters, this embodiment preferably uses statistical generalization for aggregation. Specifically, for a specific slope unit... If it contains Each raster cell corresponds to a set of raster values ​​for a given variable. Then, the calculation method for its average value can be expressed as:

[0092] ;

[0093] Meanwhile, to quantify the spatial variability of this variable within a cell, its standard deviation can be calculated as follows:

[0094] ;

[0095] Using both the mean and standard deviation to characterize the data allows for a more comprehensive depiction of the distribution characteristics of continuous variables across each slope unit.

[0096] For categorical variables, such as engineering geological rock groups or land use types (LULC) in geological elements, the aggregation method is to determine the category that occupies the largest area within each slope unit, i.e., to calculate its mode. This approach aims to extract the categorical features that play a dominant role in the overall attributes of the slope unit and use them as the representative categorical attributes of the unit.

[0097] The physical information parameters calculated in step S2 are aggregated based on their scale attributes. For macroscopic geometric parameters such as travel angle, horizontal travel distance, and vertical travel height, which are calculated for the entire slope unit, their unique values ​​are directly assigned to the corresponding slope unit. For velocity field information that is spatially distributed within the unit, the same statistical generalization method is used. Preferably, the average velocity, velocity standard deviation, and maximum velocity within the unit are extracted as composite physical features that jointly characterize the dynamic properties of the unit.

[0098] Crucially, this step also requires generating target labels for supervised learning of the neural network model. This process is accomplished by spatially overlaying the landslide inventory vector data obtained in step S1 with the slope unit layer, and generating two different types of target labels for the subsequent phased prediction task:

[0099] First, there are susceptibility labels used for the final susceptibility prediction. Each slope cell is assigned a binary objective variable by determining whether it spatially intersects with any landslide surface in the landslide inventory. If spatial intersections exist, the slope element is considered an unstable element, and let... Conversely, if there are no spatial intersections, it is considered a stable unit, and let... The probability distribution of the target variable follows a Bernoulli distribution, and its probability mass function (PMF) is:

[0100] ;

[0101] in, This represents the true probability of a landslide occurring in that unit, and is the target that the subsequent susceptibility prediction model will learn.

[0102] Secondly, there are scale labels used for the first-stage scale prediction. These labels apply only to all slope units marked as unstable (i.e.,...). The unstable slope unit is generated by calculating the total area of ​​one or more historical landslide surfaces contained within each unstable slope unit. To make the data distribution closer to the normal distribution that is easily handled by neural network models, preferably, the calculated total area value... Perform a logarithmic transformation to obtain the scale label. :

[0103] ;

[0104] In terms of physical cognition and statistical laws, the area of ​​a landslide It usually conforms to a log-Gaussian distribution, that is, its logarithmic value... It follows a Gaussian distribution (normal distribution). The probability density function (PDF) of this distribution is:

[0105] ;

[0106] in, and These are the mean and standard deviation of the logarithmic landslide area, respectively, which are the two core parameters that the subsequent scale prediction model needs to predict.

[0107] Finally, the output of step S3 is a regularized feature matrix. In this matrix, each row represents an independent slope unit, and each column represents a specific feature dimension or target label. All aggregated landscape features, together with the physical information features calculated in step S2, constitute the model's input feature vector. Its composition can be formally represented as The feature matrix, together with the generated target labels, forms a complete training and validation dataset, providing direct and effective data support for the construction, training, and validation of the phased neural network in step S4.

[0108] S4. In this embodiment, step S4 involves the construction and training of a phased neural network model. This step is the core computational step for achieving co-seismic landslide joint prediction in this invention. Its fundamental purpose is to construct and train a unified, internally phased deep learning model. This model can not only learn complex nonlinear relationships from data, but its structural design also incorporates constraints from physical information and establishes an intrinsic coupling relationship between landslide scale and landslide susceptibility, thereby completing the end-to-end mapping from input features to the final prediction result.

[0109] Therefore, this embodiment preferably constructs a Physical Information Neural Network (PINN) and uses a ResNet as the backbone of its network structure. The ResNet architecture is adopted because its unique Residual Blocks can effectively alleviate the gradient vanishing problem that may occur during deep network training, allowing the model to be built deeper, thereby capturing and expressing deeper and more complex dependencies between input features.

[0110] The key innovation of this invention lies in the fact that the neural network model is logically divided into two closely connected and interactive stages: the first stage of landslide scale prediction and the second stage of landslide susceptibility prediction.

[0111] Phase 1: Landslide Scale Prediction

[0112] This phase aims to quantitatively predict the potential landslide size for each slope unit based on the input landscape and physical information characteristics.

[0113] In a specific implementation, the input to this stage is the feature vector generated in step S3 for each slope unit. The data first passes through an input linear transformation layer, transforming the original feature vector... The transformation process from the input space to the hidden feature space inside the network can be represented as:

[0114] ;

[0115] In the formula, and These are the weight matrix and bias vector of the linear transformation layer, respectively, and are the parameters that the network needs to learn.

[0116] To improve the stability and convergence speed of model training, a batch normalization layer is preferably set before the data flows through the deep network. This layer normalizes the mini-batch data, and its calculation can be expressed as:

[0117] ;

[0118] In the formula, and The mean and variance of the current small batch of data. It is a very small constant to ensure numerical stability, while and These are learnable scaling and translation parameters. The data can then be passed through a dropout layer to enhance the model's generalization ability and prevent overfitting.

[0119] The processed data is fed into a deep feature extraction network composed of multiple stacked residual modules. After sufficient nonlinear transformation and feature learning by this network, high-dimensional abstract features capable of characterizing the scale of the landslide are obtained, denoted here as... .

[0120] The output layer in this stage is designed to contain two neurons and employ a linear activation function. Its purpose is to directly predict the two core parameters of the landslide area, which conforms to a log-Gaussian distribution and was established in step S3: namely, the mean of the logarithmic area. with standard deviation To achieve this goal, the training process in this phase uses the Huber loss function as its objective function. The Huber loss function was chosen because it combines the advantages of mean squared error (MSE) and mean absolute error (MAE). It exhibits MSE when the difference between the predicted and true values ​​is small, and MAE when the difference is large, thus providing stronger robustness to outliers in the data.

[0121] Phase Two: Landslide Susceptibility Prediction

[0122] The core of this stage is to use the landslide scale information generated in the first stage as a key physical constraint, and integrate it with the original input features to jointly predict the final landslide susceptibility probability.

[0123] The input in this stage consists of two parts: one is the original input feature vector. (After preliminary linear transformation) Secondly, the high-dimensional features extracted in the first stage that contain scale information. .Will As input, this constitutes the specific technical implementation of "using landslide scale information as a constraint".

[0124] To achieve efficient fusion of the two sets of information, this embodiment preferably employs a Feature FusionAttention mechanism. This mechanism is not a simple feature concatenation, but rather uses a self-attention network to allow the model to adaptively learn the relative importance of the two types of features for the final prediction task. The implementation process is as follows: first, the two sets of features... and By splicing, combined features are obtained. Then, The data is fed into a small multilayer perceptron (MLP) and processed by the Softmax function to generate two sets of data corresponding to... and attention weights and Ultimately, the fused features We obtain the following by weighted summation of the original features:

[0125] ;

[0126] Through this mechanism, the model is given the ability to dynamically adjust the impact of scale information.

[0127] Features after fusion Similarly, after batch normalization, random deactivation, and a series of residual modules, deep feature learning is performed for susceptibility prediction.

[0128] The final output layer in this stage is designed as a single-neuron unit, employing the sigmoid activation function. This activation function compresses any real-number input into the interval (0,1), and its output value... It can naturally be interpreted as the probability of a landslide occurring in that slope unit.

[0129] To train this phase, the binary cross-entropy loss function is used as its objective function. This loss function is a standard metric for evaluating the difference between the predicted probabilities of the binary classification model and the true labels (0 or 1).

[0130] In terms of the overall training strategy, the entire model can be jointly trained end-to-end. Its total loss function can be set as a weighted sum of the first-stage Huber loss and the second-stage binary cross-entropy loss. The Adam optimization algorithm is used to update all learnable parameters in the network using gradient descent based on the total loss. Before training, the total dataset generated in step S3 needs to be divided into a training set and a validation set in a certain ratio (7:3) to monitor model performance during training and adjust the strategy as needed to prevent overfitting.

[0131] Specifically, after the model training is completed, all learnable parameters within the model (such as the weights of each linear transformation layer) are... and bias and batch normalization layer and All features have been identified and solidified. This step switches the model to inference mode and provides the complete feature dataset prepared in step S3 for all slope units in the study area as input to the trained model.

[0132] For each slope element, its eigenvector After being input into the model, the data will propagate forward within the model according to the data flow path described in step S4, thereby generating two types of core prediction results. These two types of results correspond to the two stages of the staged prediction model, and together constitute the final output of this invention.

[0133] Firstly, there is the generation and output of landslide scale prediction results. This result is generated by the first prediction stage of the model. For each slope unit, the output layer of this stage directly predicts two key parameters of the log-Gaussian distribution that its potential landslide area follows: namely, the mean of the log-area. and standard deviation To generate intuitive size predictions, preferably, the predicted logarithmic mean can be used. This represents the most likely logarithmic size of the unit. By applying an exponential transformation, the predicted landslide area of ​​this slope unit can be obtained. :

[0134] ;

[0135] This predicted area value for all slope units within the study area. By associating these landslides with their corresponding spatial geometric locations, a continuous landslide scale prediction map covering the entire study area can be generated. This map visually demonstrates the spatial distribution of landslides of varying scales that may occur at different locations within the region under the triggering of earthquakes of similar intensity.

[0136] Secondly, the generation and output of landslide susceptibility prediction results. This result is generated in the second prediction stage of the model and is the core output of this invention. In this stage, the final output layer of the model uses the Sigmoid activation function to calculate a value between 0 and 1 for each slope unit, denoted as . This value, in a physical sense, represents the probability that the slope unit will become unstable (i.e., a landslide) under given conditions; it is the landslide susceptibility index.

[0137] Predicted probability values ​​for all slope elements By assigning corresponding spatial units, a landslide susceptibility zoning map can be generated. It is important to emphasize that the generation process of this susceptibility result involves inputs containing high-dimensional features predicted and passed down from the first stage. Furthermore, it effectively fused features through a feature fusion attention mechanism, resulting in a higher final susceptibility probability. The calculation is constrained by information on the potential landslide scale. This design ensures the inherent physical consistency of the prediction results. That is, the instability probability of a slope unit predicted to have a very small potential landslide scale will also be reasonably modulated accordingly, thus avoiding the problem that may occur in traditional independent models, which may mistakenly identify small-scale landslide-prone areas as extremely high-risk areas.

[0138] Ultimately, the output of step S5 is a set of digital thematic map products. This product is preferably stored and delivered in a Geographic Information System (GIS) compatible format (raster file or vector polygon file), and mainly includes landslide scale prediction maps and landslide susceptibility maps. These map products provide a more reliable scientific decision-making basis with dual constraints of physical and scale information for subsequent land spatial planning, disaster risk assessment, early warning system construction, and emergency resource deployment.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A coseismic landslide prediction method integrating physical information and landslide scale constraints, characterized in that, Includes the following steps: a. Obtain landscape data and seismic parameters of the study area, and use the slope element as the basic analysis unit; b. Based on the landscape data and seismic parameters, a physical model based on the law of conservation of energy is used to calculate the physical information parameters of each slope unit. The physical information parameters are used to characterize the dynamic properties of the material within the slope. c. A staged neural network model is used for prediction, the process of which includes: Phase 1: Using the landscape data and physical information parameters as input, predict the landslide scale information for each slope unit; The second stage involves using the landslide scale information predicted in the first stage as a constraint, along with the landscape data and physical information parameters, as input to predict the final landslide susceptibility of each slope unit. The physical model based on the law of conservation of energy in step b is an improved Heim energy line model, which incorporates the influence of seismic motion parameters in the calculation. The phased neural network model is a physical information neural network, and its network architecture is based on a residual network.

2. The coseismic landslide prediction method integrating physical information and landslide scale constraints according to claim 1, characterized in that, The physical information parameters include at least velocity information that characterizes the speed of material movement.

3. The coseismic landslide prediction method integrating physical information and landslide scale constraints according to claim 2, characterized in that, In the second stage, the specific way to use the landslide scale information as a constraint is as follows: the landslide scale information is used as a new feature and input into the neural network model for landslide susceptibility prediction.

4. The coseismic landslide prediction method integrating physical information and landslide scale constraints according to claim 3, characterized in that, The method also employs a feature fusion attention mechanism to adaptively weight landslide scale information features and other input features to generate fused features for predicting landslide susceptibility.

5. The coseismic landslide prediction method integrating physical information and landslide scale constraints according to claim 1, characterized in that, The ground motion parameter is the peak ground acceleration.

6. The coseismic landslide prediction method integrating physical information and landslide scale constraints according to claim 1, characterized in that, The method does not require input of regional geotechnical parameters, including cohesion and internal friction angle, during implementation.

7. The coseismic landslide prediction method integrating physical information and landslide scale constraints according to claim 1, characterized in that, The landslide scale information predicted in the first stage is the landslide area that conforms to a log-Gaussian distribution; the final landslide susceptibility predicted in the second stage is the probability value that conforms to a Bernoulli distribution.

8. A system for predicting coseismic landslides based on the method for fusing physical information and landslide scale constraints as described in claim 1, characterized in that, include: The data acquisition module is used to acquire landscape data and seismic parameters of the study area; The physical information calculation module is used to calculate physical information parameters characterizing the dynamic properties of matter by calling the improved Heim energy line model, with the slope element as the unit. The data aggregation module is used to uniformly summarize and process all acquired and calculated data at the slope unit scale; The joint modeling and prediction module integrates a phased physical information neural network. This module receives the aggregated data and performs the following operations: Run a scale prediction program to predict and output landslide scale information for each slope unit based on input features; Run a susceptibility prediction program, take the scale information output by the aforementioned scale prediction program as the input of the constraint term, and predict and output the final landslide susceptibility map.

Images