Physical information guided machine learning landslide susceptibility evaluation method for scientific consistency

Abstract

The invention discloses a scientific consistency-oriented physical information guided machine learning landslide susceptibility evaluation method, and relates to the technical field of landslide susceptibility evaluation, and the method comprises the steps: constructing a preliminary training data set; carrying out optimization processing on the preliminary training data set based on a semantic-space consistency joint rejection and topology-residual rejection method; calculating physical information parameters of the target landslide area, and merging the physical information parameters into the optimization training data set; constructing a loss function containing physical consistency and risk consistency constraints; training a probabilistic physical information machine learning model in combination with the loss function, and optimizing the probabilistic physical information machine learning model by using a spatial cross validation method; and performing landslide susceptibility evaluation on the target landslide area by using the optimized probability physical information machine learning model to obtain an evaluation result. According to the method, a more stable and reliable modeling path is provided for regional landslide risk assessment, and a basis is provided for decision-making of related departments of land resource management.

Description

Technical Field

[0001] This invention relates to the field of landslide susceptibility assessment technology, and more specifically, to a physical information-guided machine learning method for landslide susceptibility assessment oriented towards scientific consistency. Background Technology

[0002] Landslides are a common natural disaster worldwide and one of the most frequent geological hazards in my country. Extreme conditions such as heavy rainfall and earthquakes exacerbate landslide susceptibility. While most natural landslides are shallow, only a few meters deep, their highly concealed occurrence, proximity to areas of human engineering activity, and severe consequences make them a significant threat to human life and property. Therefore, landslide prediction before disasters occur is extremely important. Landslide susceptibility assessment (LSA) is a crucial tool for regional disaster risk management, identifying areas most prone to landslides.

[0003] Landslide susceptibility assessment quantifies the probability of landslides in a specific area by analyzing the relationship between key factors such as topography, geology, climate, and human activities and historical landslide events. It is mainly divided into two categories: methods based on physical mechanics models and data-driven methods based on statistical analysis or machine learning (ML) models. Physical mechanics modeling methods are based on landslide mechanics analysis and combine slope limit equilibrium theory with hydrogeological models to simulate the impact of triggering factors such as rainfall and earthquakes on slope stability. However, due to the complexity of the terrain and incomplete information, physical models often rely on existing parameters, neglecting the inherent uncertainties of the parameters. This often leads to an overestimation of landslide susceptibility, reducing the reliability of the model in regional landslide prediction.

[0004] In recent years, probabilistic methods have been gradually introduced into landslide susceptibility analysis to improve the robustness and scientific validity of the models. Although this method provides a clear mechanistic explanation based on physical models, its inherent simplification assumptions still limit its application at the regional scale and restrict its scope of application.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] In view of this, the present invention provides a physical information-guided machine learning method for assessing landslide susceptibility with scientific consistency, in order to solve the aforementioned problems.

[0007] To solve the above problems, the specific technical solution adopted by the present invention is as follows:

[0008] A physical information-guided machine learning method for landslide susceptibility assessment, oriented towards scientific consistency, includes the following steps:

[0009] S1. Obtain geographic data of the target landslide area and input the geographic data into open-source geographic information system software. Use buffer analysis to construct a preliminary training dataset.

[0010] S2. Optimize the initial training dataset to obtain the optimized training dataset;

[0011] S3. Obtain soil parameters and rainfall parameters of the target landslide area, and calculate the physical information parameters of the target landslide area in combination with the instantaneous infiltration module model. Then, merge the physical information parameters into the optimized training dataset to obtain the merged training dataset.

[0012] S4. Based on the simplified transient infiltration model and the first-order reliability algorithm framework, a loss function containing physical consistency and risk consistency constraints is constructed.

[0013] S5. Based on the merged training dataset, a probabilistic physical information machine learning model is trained using a loss function, and the probabilistic physical information machine learning model is optimized using the spatial cross-validation method.

[0014] S6. Use the optimized probabilistic physical information machine learning model to assess the landslide susceptibility of the target landslide area and obtain the assessment results.

[0015] Preferably, the step of acquiring geographic data of the target landslide area and inputting the geographic data into open-source geographic information system software, and constructing a training dataset using buffer analysis includes the following steps:

[0016] S11. Input the geographic data of the target landslide area into the open-source geographic information system software, and based on the spatial analysis function of the open-source geographic information system software, construct a buffer zone with a preset radius centered on the preset landslide point.

[0017] S12. Remove all grid cells within the buffer zone from the total area of ​​the target landslide area, and use a stratified random sampling method to stratify the remaining area after the buffer zone is removed according to slope grade and lithology.

[0018] S13. Within each layer, randomly select geographical units that do not overlap with the preset landslide samples as negative samples, and extract the corresponding factor values ​​from the preset environmental factor raster data based on the geographical coordinates of the negative samples; construct a training dataset based on the extracted factor values.

[0019] Preferably, optimizing the initial training dataset to obtain an optimized training dataset includes the following steps:

[0020] S21. Based on each sample point in the preliminary training dataset, select several geographic entities according to a preset range;

[0021] S22. Based on the selected geographic entities, calculate the semantic similarity between each sample point and the preset standard geographic entities;

[0022] S23. Based on the selected geographic entities, perform graph convolution aggregation on the samples to obtain sample node representations, and calculate the semantic consistency probability based on the sample node representations.

[0023] S24. The semantic similarity and semantic consistency probability are weighted and fused to obtain the weighted fusion result. The sample points are judged to be semantically abnormal by using the preset threshold, and the sample points with semantically abnormality are removed to obtain the secondary training dataset.

[0024] S25. Using the topological consistency test and combined with the M-estimation residual screening method, false neighboring points caused by topological barriers in the secondary training dataset are removed to obtain the optimized dataset.

[0025] Preferably, the step of calculating the semantic similarity between each sample point and a preset standard geographic entity based on the selected geographic entities includes the following steps:

[0026] S221. Encode geographic entities into vector representations and perform semantic embedding on the geographic entity encoding results to generate entity semantic embedding vectors.

[0027] S222. Based on the spatial distance between each sample point and the geographic entity, and combined with the entity semantic embedding vector, calculate the domain weight and normalize it to obtain the sample semantic vector.

[0028] S223. Perform cosine similarity calculation between the sample semantic vector and the preset standard prototype vector to obtain the semantic similarity between the sample point and the standard geographic entity.

[0029] Preferably, the step of performing graph convolution aggregation on the samples based on the selected geographic entities to obtain sample node representations, and calculating the semantic consistency probability based on the sample node representations, includes the following steps:

[0030] S231. Determine the spatial adjacency relationship based on the selected geographic entities, and construct a heterogeneous graph with the spatial adjacency relationship as the edge and the selected geographic entity as the vertex.

[0031] S232. Based on the constructed heterogeneous graph, the sample node representation is obtained by using a relational graph convolutional network for aggregation processing.

[0032] S233. Based on the sample node representation and combined with the preset nonlinear activation, calculate the semantic consistency probability.

[0033] Preferably, the steps of obtaining soil parameters and rainfall parameters of the target landslide area, calculating physical information parameters of the target landslide area in conjunction with the instantaneous infiltration module model, and merging the physical information parameters into the optimized training dataset to obtain the merged training dataset include the following steps:

[0034] S31. Obtain soil parameters for the target landslide area, analyze the ratio of soil shear strength to sliding shear stress and soil thickness using the instantaneous infiltration module model, and calculate the failure probability and soil reliability index by combining the ratio of soil shear strength to sliding shear stress.

[0035] S32. Obtain rainfall parameters during rainfall in the target landslide area and calculate the final infiltration amount using the infiltration calculation formula;

[0036] S33. The ratio of soil shear strength to sliding shear stress, final infiltration, soil thickness, failure probability, and soil reliability index are used as physical information parameters of the target landslide area. The training dataset is optimized and merged to obtain the merged training dataset.

[0037] Preferably, before analyzing the ratio of soil shear strength to sliding shear stress and soil thickness using the instantaneous infiltration module model, the method includes: dynamically correcting the soil parameters of the target landslide area using a parameter adaptive analysis method;

[0038] The method of dynamically correcting soil parameters in the target landslide area using adaptive parameter analysis includes:

[0039] Based on the fitting residual of the measured rainfall infiltration rate and soil moisture content curve, the soil saturation unit weight parameter is adjusted in real time using the exponential decay function.

[0040] Based on the changes in water content at different depths from soil profile moisture sensors, the advance rate of the wetting front is calculated, and the depth parameters of the wetting front are updated.

[0041] By measuring the suction and moisture content data of the soil matrix, the key parameters of the soil water characteristic curve are updated using a real-time inversion method, and the relationship between pore water pressure and depth is corrected. Based on the correction results, the pore water pressure parameters are adjusted.

[0042] Preferably, the loss function that incorporates physical consistency and risk consistency constraints includes:

[0043] Constructing a scientific consistency loss based on physical consistency loss and probabilistic risk loss;

[0044] Construct a supervised loss to measure the difference between the prediction results of the probabilistic physical information machine learning model and the actual observations;

[0045] By combining scientific consistency loss with supervision loss, a loss function is constructed to measure the probabilistic physical information machine learning model.

[0046] Preferably, the step of training a probabilistic physical information machine learning model based on a merged training dataset and a loss function, and then optimizing the probabilistic physical information machine learning model using spatial cross-validation includes the following steps:

[0047] S51. Use the digital elevation model to derive various terrain factors and extract the geographical location features of landslide points to obtain multi-dimensional feature vectors. Add the multi-dimensional feature vectors to the merged training dataset to obtain the integrated training dataset.

[0048] S52. Input the integrated training dataset into the probabilistic physical information machine learning model, and train the probabilistic physical information machine learning model by combining the loss function.

[0049] S53. Optimize the trained probabilistic physical information machine learning model to obtain the optimized probabilistic physical information machine learning model.

[0050] Preferably, the step of optimizing the trained probabilistic physical information machine learning model to obtain the optimized probabilistic physical information machine learning model includes the following steps:

[0051] S531. Based on the spatial distribution characteristics of the target landslide area, the spatial region corresponding to the integrated training dataset is divided into several clusters;

[0052] S532. Divide one group of clusters into a validation set and the remaining clusters into a training set, and re-perform multiple rounds of cross-validation on the trained probabilistic physical information machine learning model.

[0053] S533. Calculate the index of the prediction results of multiple rounds of cross-validation, and optimize the parameters of the probabilistic physical information machine learning model based on the index calculation results to obtain the optimized probabilistic physical information machine learning model.

[0054] The beneficial effects of this invention are as follows:

[0055] 1. This invention effectively bridges the contradiction between data dependence and scientific consistency in traditional machine learning in geological disaster modeling by integrating a regularization mechanism based on physical stability and probabilistic risk information. It provides a more robust and reliable modeling path for regional landslide risk assessment and provides a basis for decision-making by relevant departments of land and resources management.

[0056] 2. This invention proposes a spatial distribution map of landslide susceptibility generated based on a probabilistic physical model, and uses a spatial cross-validation strategy to conduct model validation, further quantitatively evaluating the predictive performance of the PRL-STIM model. This solves the problem that traditional pure data-driven models have obvious instability and insufficient generalization ability when dealing with areas with strong spatial heterogeneity. In contrast, the model that introduces physical constraints shows stronger robustness in cross-regional migration.

[0057] 3. The PIML model used in this invention, as the regularization strength increases, enhances the guiding role of physical and probabilistic constraints, and the model output gradually tends to be more consistent with the prediction results of geotechnical engineering principles. Based on the advantages of integrating data-driven learning and probabilistic physical modeling, the PIML model takes into account both prediction performance and scientific consistency, and shows good generalization and interpretability. The PIML model has strong classification ability in various spatial clustering regions, which is superior to probabilistic physical models and machine learning models. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a diagram of a probabilistic physical information-based machine learning framework in a physical information-guided machine learning method for landslide susceptibility assessment based on scientific consistency, according to an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram of the PRL-STIM model in the physical information-guided machine learning landslide susceptibility assessment method based on an embodiment of the present invention.

[0061] Figure 3 This is the spatial distribution of landslide and non-landslide samples in five clusters in the physical information-guided machine learning landslide susceptibility assessment method based on an embodiment of the present invention.

[0062] Figure 4 This is a schematic diagram of high-resolution satellite images before and after extreme rainfall in the physical information-guided machine learning landslide susceptibility assessment method for scientific consistency according to an embodiment of the present invention.

[0063] Figure 5 This is a map of landslide influencing factors in the study area of ​​the landslide susceptibility assessment method guided by physical information for scientific consistency according to an embodiment of the present invention.

[0064] Figure 6This is a landslide susceptibility map of the study area based on PPM in the physical information-guided machine learning landslide susceptibility assessment method for scientific consistency according to an embodiment of the present invention.

[0065] Figure 7 This is a PRL-STIM performance graph evaluated on the entire dataset in the physical information-guided machine learning landslide susceptibility assessment method for scientific consistency according to an embodiment of the present invention.

[0066] Figure 8 This is a performance comparison chart of NN models with different clustering regions based on spatial cross-validation in the physical information-guided machine learning landslide susceptibility assessment method for scientific consistency according to an embodiment of the present invention.

[0067] Figure 9 This is a landslide susceptibility map of the study area drawn by a combined NN model based on a single spatial CV repetition in the landslide susceptibility assessment method for physical information guided by scientific consistency according to an embodiment of the present invention.

[0068] Figure 10 This invention relates to a physical information-guided machine learning method for landslide susceptibility assessment based on scientific consistency, as described in an embodiment of the invention. phy =λ risk L under change sc A schematic diagram illustrating the impact on PIML model performance;

[0069] Figure 11 This is a schematic diagram illustrating the impact of the scientific consistency-guided loss function on model prediction in the physical information-guided machine learning landslide susceptibility assessment method based on scientific consistency, according to an embodiment of the present invention.

[0070] Figure 12 This is a performance graph (λ) of the PIML model based on spatial cross-validation in the physical information-guided machine learning landslide susceptibility assessment method for scientific consistency according to an embodiment of the present invention. phy =λ risk =0.5). Detailed Implementation

[0071] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0072] According to embodiments of the present invention, a physical information-guided machine learning method for landslide susceptibility assessment oriented towards scientific consistency is provided.

[0073] Specifically, firstly, for complex environmental information inputting topographic factors, vegetation factors (NDVI), and soil factor geographic data, a landslide probability analysis framework (PRL-STIM) based on a simplified transient infiltration model and a first-order reliability algorithm is constructed, and a machine learning model loss function containing physical consistency and risk consistency constraints is built; a probabilistic physical information machine learning (PIML) model is proposed.

[0074] Specifically, the research area of ​​this invention is located in a small basin in the northeastern part of a region in my country in July 2013. Due to the complex topography and diverse ecological and geomorphological types of this region, it is highly representative. Data from 1981 to 2010 shows an average annual precipitation of approximately 500.7 mm and an average relative humidity of 66%. 85% of the precipitation is concentrated between April and October, especially between July and September, accounting for approximately 68% of the total annual precipitation. Heavy rainfall is the main triggering factor for landslides.

[0075] The landslide was triggered by continuous heavy rainfall. Considering both the duration and amount of rainfall, the concentrated rainfall period before the landslide was selected as the study period for physical modeling. The landslide influencing factors in the study area were analyzed, with the main influencing factors in the main case study being... Figure 5 As shown, a physical information-guided machine learning method for landslide susceptibility assessment, oriented towards scientific consistency, is employed to prevent the recurrence of disasters. The specific process is as follows:

[0076] The files required for the topographic spatial data of this region are simplified. This example mainly includes the following parts:

[0077] By comparing high-resolution satellite images before and after extreme rainfall, such as Figure 4 As shown, the study area is mainly distributed in regions with an altitude of 1506–2047 m and a slope of 22°–33°. The landslide bodies are usually composed of weathered rock, mudstone layers, or gravelly soil with a thickness of less than 3 m, exhibiting obvious spatial aggregation. The heaviest rainfall occurred from noon on July 21 to the morning of July 22, which was the direct trigger for the landslide.

[0078] To accurately characterize the impact of topography, hydrology, and soil properties on landslide occurrence, it is necessary to rationally determine environmental factors such as... Figure 5 Twelve representative factors were selected, along with three main control indicators: topographic factors (elevation, slope, aspect, topographic humidity index TWI, plan curvature and profile curvature), vegetation factors (NDVI), and soil factors (soil particle composition, bulk density, saturated hydraulic conductivity Ks, and field capacity).

[0079] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the landslide susceptibility assessment method guided by physical information based on scientific consistency according to an embodiment of the present invention includes the following steps:

[0080] S1. Obtain geographic data of the target landslide area and input the geographic data into open-source geographic information system software. Use buffer analysis to construct a preliminary training dataset.

[0081] in, Figure 1 The fully connected layers shown are used to construct deep neural network structures in probabilistic physical information machine learning (PIML) models. They linearly transform input features (such as environmental factors like elevation, slope, and NDVI) through weight matrices and bias parameters. The fully connected structures of different layers can abstract and extract deep features of input factors layer by layer, which can be used to better learn the relationship between landslide occurrence and environmental factors.

[0082] As a preferred embodiment, the steps of acquiring geographic data of the target landslide area and inputting the geographic data into open-source geographic information system software, and constructing a training dataset using buffer analysis include the following:

[0083] S11. Input the geographic data of the target landslide area into the open-source geographic information system software, and based on the spatial analysis function of the open-source geographic information system software, construct a buffer zone with a preset radius centered on the preset landslide point.

[0084] S12. Remove all grid cells within the buffer zone from the total area of ​​the target landslide area, and use a stratified random sampling method to stratify the remaining area after the buffer zone is removed according to slope grade and lithology.

[0085] S13. Within each layer, randomly select geographical units that do not overlap with the preset landslide samples as negative samples, and extract the corresponding factor values ​​from the preset environmental factor raster data based on the geographical coordinates of the negative samples; construct a training dataset based on the extracted factor values.

[0086] Specifically, by directly inputting text files containing environmental information such as topographic factors, vegetation factors (NDVI), and soil factors into the open-source Geographic Information System (QGIS) software, and within areas far from known landslides, using the spatial analysis functions of QGIS or ArcGIS, a buffer analysis method is employed. A buffer zone with a radius of R is constructed centered on the known landslide point (R is determined based on the DEM resolution of the study area and the landslide impact range, generally 500–1000m). All raster cells within the buffer zone are then removed from the total study area. Within the remaining area, a stratified random sampling method is used, stratifying the data according to slope grade and lithology. Within each layer, geographical units that do not overlap with known landslide samples are randomly selected as negative samples. The corresponding factor values ​​are then extracted from the environmental factor raster data (such as slope, curvature, NDVI, and soil saturated hydraulic conductivity) based on the sample geographical coordinates to construct a complete training dataset. This method combines spatial buffers and stratified random sampling, which ensures the spatial independence of negative samples and landslide areas, and provides balanced coverage in terms of terrain and geological distribution, thereby improving the model's ability to generalize to different environmental conditions.

[0087] S2. Optimize the initial training dataset to obtain the optimized training dataset;

[0088] Data is collected and processed by inputting terrain factors, and then spatial cross-validation is performed, such as... Figure 3 As shown, the k-means clustering algorithm is used to divide the samples into five spatial clusters based on their geographical coordinates. When dividing the spatial clusters, in addition to removing spatial outliers by density clustering (DBSCAN), a better approach is to use a two-stage method that combines semantic-spatial consistency removal and topology-residual removal.

[0089] In a preferred embodiment, optimizing the initial training dataset to obtain an optimized training dataset includes the following steps:

[0090] S21. Based on each sample point in the preliminary training dataset, select several geographic entities according to a preset range;

[0091] S22. Based on the selected geographic entities, calculate the semantic similarity between each sample point and the preset standard geographic entities;

[0092] In a preferred embodiment, calculating the semantic similarity between each sample point and a preset standard geographic entity based on the selected geographic entities includes the following steps:

[0093] S221. Encode geographic entities into vector representations and perform semantic embedding on the geographic entity encoding results to generate entity semantic embedding vectors.

[0094] S222. Based on the spatial distance between each sample point and the geographic entity, and combined with the entity semantic embedding vector, calculate the domain weight and normalize it to obtain the sample semantic vector.

[0095] S223. Perform cosine similarity calculation between the sample semantic vector and the preset standard prototype vector to obtain the semantic similarity between the sample point and the standard geographic entity.

[0096] S23. Based on the selected geographic entities, perform graph convolution aggregation on the samples to obtain sample node representations, and calculate the semantic consistency probability based on the sample node representations.

[0097] In a preferred embodiment, the step of performing graph convolution aggregation on the samples based on the selected geographic entities to obtain sample node representations, and calculating the semantic consistency probability based on the sample node representations, includes the following steps:

[0098] S231. Determine the spatial adjacency relationship based on the selected geographic entities, and construct a heterogeneous graph with the spatial adjacency relationship as the edge and the selected geographic entity as the vertex.

[0099] S232. Based on the constructed heterogeneous graph, the sample node representation is obtained by using a relational graph convolutional network for aggregation processing.

[0100] S233. Based on the sample node representation and combined with the preset nonlinear activation, calculate the semantic consistency probability.

[0101] S24. The semantic similarity and semantic consistency probability are weighted and fused to obtain the weighted fusion result. The sample points are judged to be semantically abnormal by using the preset threshold, and the sample points with semantically abnormality are removed to obtain the secondary training dataset.

[0102] S25. Using the topological consistency test and combined with the M-estimation residual screening method, false neighboring points caused by topological barriers in the secondary training dataset are removed to obtain the optimized dataset.

[0103] Specifically, when employing a two-stage method that incorporates semantic-spatial consistency joint elimination and topology-residual elimination: semantic consistency scores can be constructed based on geographic entities such as road networks, administrative divisions, POI density, and river networks. Firstly, the similarity between sample points and standard geographic entities can be calculated using a semantic embedding model, including:

[0104] First, encode the geographic entities (roads, administrative regions, river networks, POIs, etc.) near the sample point into vector representations, and denote the sample point as x. i The set of entities within its neighborhood is (x i Given entity type t ∈ {road, district, river, POI}, the entity semantic embedding vector v is obtained through Node2Vec / Word2Vec / category embedding. e ∈R d Neighborhood weight The calculation formula is:

[0105] ;

[0106] In the formula, For sample x i Spatial distance to geographic entity e For kernel function, Weights for geographic entity types, The scale or bandwidth parameter of the kernel function is used to determine the sample semantic vector. Next, the neighborhood weights are calculated and normalized. Then, cosine similarity is calculated between the vector and the standard prototype vector r. The calculation formula is as follows:

[0107] ;

[0108] In the formula, Indicates sample x i The cosine similarity calculation results of the standard prototype vector r. This represents the normalized similarity score, which linearly maps the original similarity (value range [-1,1]) to the [0,1] interval, and is used to represent the "semantic consistency probability". Represents sample point x i The neighborhood weights for a geographic entity e are obtained by comprehensively considering spatial distance and entity type weights.

[0109] Secondly, roads, administrative regions, and river networks are further modeled as a heterogeneous graph G=(V,E), where vertices represent geographic entities and edges represent spatial adjacency relationships. A graph neural network is used to aggregate the samples through graph convolution, connecting edges according to relationships such as "nearest neighbor / containment / topographic adjacency," and then using relational graph convolution (R-GCN / GAT) for aggregation to obtain the sample node representation z. i The aggregation formula is:

[0110] ;

[0111] In the formula, This represents the nodes that are adjacent to v according to the relation r. Represents the normalization constant. Indicates learnable weights, Indicates nonlinear activation. Indicates the first Layer hidden representation, Indicates the first Hidden representation; output semantic consistency probability The two are combined in a weighted fusion form as follows: v e Indicates entity semantic embedding, Represents the neighborhood weight, and represents the sample point x. i The overall weight between the geographic entity e and the relation R represents the set of relations, including various spatial relations such as "nearest neighbor", "containment", and "terrain adjacency", r represents a single relation type, which belongs to a specific relation in the set R, and u represents a neighbor node that has a relation r with node v.

[0112] After determining semantic similarity and graph neural network consistency and removing semantic outliers, a superior approach is to further utilize the topological consistency test based on the ratio of the shortest path in the road / river network to the Euclidean distance, combined with M-estimation residual screening, to remove spurious neighbor points caused by topological barriers. Subsequently, to ensure consistency between spatial proximity and environmental similarity, a multi-index fusion clustering distance is defined:

[0113] ;

[0114] In the formula, D represents the cluster distance. This represents geographic spatial distance, typically Euclidean distance or spatial distance constrained by road networks. The distance between environmental factors is represented by Mahalanobis distance, which measures the differences between different environmental indicators (slope, lithology, land use, etc.). sem This indicates the introduction of semantic differences between roads / administration / land use and the shortest path term in the graph, d. dyn This represents the DTW distance using NDVI and soil moisture time series. The distances of hydrological factors are represented; after normalization of each sub-distance, the weights {α,β,γ,δ,ζ} are adaptively determined using the information entropy method or attention mechanism. Based on this, weighted k-means clustering is performed using the metric matrix 𝑀, and spatial topological regularization is added.

[0115] ;

[0116] In the formula, A represents the DEM adjacency matrix. First, environmental, hydrological, and dynamic process indicators are fused and embedded using CCA / autoencoder, and then the above weighted clustering is performed to improve cross-regional generalization stability and physical interpretability. The feature vector representing the sample point, Indicates sample The center of the cluster to which it belongs, The coefficient represents the topology regularization factor, which controls the importance of spatial topology constraints. Indicates and The center of an adjacent cluster.

[0117] Further, a superior multidimensional weighted k-means clustering method integrating topographic and geological features is employed. Environmental factors such as slope, lithology, and rainfall intensity are combined with geographic coordinates to calculate the clustering distance. Outlier removal includes: constructing semantic consistency scores based on road networks, administrative divisions, POI density, and river networks, and performing cascaded judgments using the outlier scores output by a classifier. For example, within the normalized score range [0,1], when the comprehensive outlier score is below 0.3, it is judged as a semantic-spatial inconsistent sample and removed. Furthermore, a topological consistency test using the ratio of the shortest path to the Euclidean distance on the road / river network map is used in conjunction with M-estimation residual screening; if the residual exceeds 3σ, it is removed as a pseudo-neighbor. The clustering distance D simultaneously considers geographic, environmental, hydrological, semantic, and dynamic process indicators, calculated according to the clustering distance formula. The calculations are performed, sub-distances are normalized, and weights are determined using either information entropy or attention mechanisms. Clustering is then performed using a weighted k-means algorithm based on the metric matrix M, and a topological regularization term based on the adjacency matrix 𝐴 of the DEM is introduced. This method ensures high consistency in spatial proximity and environmental similarity among spatial clusters. It effectively reduces the interference of outliers on cluster centers and improves the generalization stability and physical interpretability of clustering results in cross-regional landslide susceptibility prediction.

[0118] S3. Obtain soil parameters and rainfall parameters of the target landslide area, and calculate the physical information parameters of the target landslide area in combination with the instantaneous infiltration module model. Then, merge the physical information parameters into the optimized training dataset to obtain the merged training dataset.

[0119] In a preferred embodiment, the steps of obtaining soil parameters and rainfall parameters of the target landslide area, calculating physical information parameters of the target landslide area in conjunction with the instantaneous infiltration module model, and merging the physical information parameters into the optimized training dataset to obtain the merged training dataset include the following steps:

[0120] S31. Obtain soil parameters for the target landslide area, analyze the ratio of soil shear strength to sliding shear stress and soil thickness using the instantaneous infiltration module model, and calculate the failure probability and soil reliability index by combining the ratio of soil shear strength to sliding shear stress.

[0121] It should be noted that the soil parameters in the target landslide area include soil saturation unit weight parameters, wetting front depth parameters, pore water pressure parameters, soil thickness, slope angle parameters, soil cohesion parameters, effective friction angle parameters, and effective stress parameters. Specifically, the parameter adaptive analysis method is used to dynamically correct the soil saturation unit weight parameters, wetting front depth parameters, and pore water pressure parameters. Based on the corrected soil saturation unit weight parameters, wetting front depth parameters, and pore water pressure parameters, combined with slope angle parameters, soil cohesion parameters, and effective friction angle parameters, the ratio of soil shear strength to sliding shear stress is analyzed using an instantaneous infiltration module model.

[0122] Specifically, the formula for calculating the ratio of soil shear strength to sliding shear stress is as follows:

[0123] ;

[0124] In the formula, γ represents the ratio of soil shear strength to sliding shear stress. sat The value represents the saturated weight of soil, and θ represents the slope angle. Indicates the cohesion of the soil. Indicates the effective friction angle. The value represents the depth of the wetting front, and μ represents the pore water pressure parameter.

[0125] It should be noted that, according to the GIS-FORM model, based on the ratio of soil shear strength to sliding shear stress ( This study calculates the impact of uncertainties in soil cohesion, internal friction angle, slope, and rainfall intensity parameters on slope stability. Further calculations of failure probability and geotechnical reliability indices are performed. A real-time parameter adaptive module is introduced to dynamically correct the input parameters of the instantaneous infiltration module (STIM), including soil saturated hydraulic conductivity, wetting front depth, and pore water pressure curve parameters, to improve slope stability. With P f The accuracy and timeliness of the (failure probability index) calculation are important. The correction of soil saturated hydraulic conductivity can be based on the fitting residual of the measured rainfall infiltration rate and soil moisture content curve, using an exponential decay function to adjust the soil saturated hydraulic conductivity value in real time. Combined with the moisture content changes at different depths of the soil profile moisture sensor, the wetting front advance rate is calculated and the wetting front depth value is updated to reflect the relationship between the current infiltration depth and time. The pore water pressure curve parameter correction can be achieved by using measured soil matrix suction (tensiometer data) and moisture content data, using a real-time inversion method to update the key parameters (such as α, λ) of the soil water characteristic curve (SWCC) and correct the depth relationship.

[0126] In addition, the Z-model was used to generalize the soil thickness as a function of elevation Z, and combined with digital elevation model data, the soil thickness at each topographic point in the target landslide area was calculated.

[0127] It should be noted that the Z-model for the thickness of the sliding soil mass (i.e., the upper boundary of the potential sliding surface) generalizes the soil thickness as a function of elevation Z as follows:

[0128] ;

[0129] In the formula, h max and h min These represent the maximum and minimum soil thicknesses, respectively; h i Z represents the soil depth of a given grid cell. i Z represents the altitude of each unit; max and Z min These represent the maximum and minimum altitudes, respectively.

[0130] In addition, based on multi-starting point parallel search and particle swarm optimization, the GIS-FORM model is optimized. Combining the ratio of soil shear strength to sliding shear stress, the optimized GIS-FORM model is used to analyze the impact of uncertainty parameters on slope stability and calculate the failure probability and geotechnical reliability index.

[0131] As a preferred embodiment, before analyzing the ratio of soil shear strength to sliding shear stress and soil thickness using the instantaneous infiltration module model, the method includes: dynamically correcting the soil parameters of the target landslide area using a parameter adaptive analysis method.

[0132] The method of dynamically correcting soil parameters in the target landslide area using adaptive parameter analysis includes:

[0133] Based on the fitting residual of the measured rainfall infiltration rate and soil moisture content curve, the soil saturation unit weight parameter is adjusted in real time using the exponential decay function.

[0134] The correction of the soil saturated unit weight parameter can be achieved by using an exponential decay function to adjust the soil saturated unit weight parameter value in real time based on the fitting residual of the measured rainfall infiltration rate and soil moisture content curves. During the calculation of effective infiltration and wetting front advancement, the hydraulic conductivity parameter is indirectly reflected by influencing the wetting front depth and pore water pressure curves. Therefore, the dynamic correction of soil saturated hydraulic conductivity in this invention serves as an external calibration basis for the STIM model input parameters (wetting front depth and pore water pressure parameters).

[0135] Based on the changes in water content at different depths from soil profile moisture sensors, the advance rate of the wetting front is calculated, and the depth parameters of the wetting front are updated.

[0136] By measuring the suction and moisture content data of the soil matrix, the key parameters of the soil water characteristic curve are updated using a real-time inversion method, and the relationship between pore water pressure and depth is corrected. Based on the correction results, the pore water pressure parameters are adjusted.

[0137] S32. Obtain rainfall parameters during rainfall in the target landslide area and calculate the final infiltration amount using the infiltration calculation formula;

[0138] It should be noted that rainfall parameters during rainfall in the target landslide area include effective infiltration and rainfall intensity. Since the evolution path of pore water pressure during instantaneous rainfall infiltration is closely related to the changes in slope shear strength and soil water characteristic curve, the effective infiltration I during rainfall is not equivalent to the rainfall intensity I. R The final infiltration amount I is:

[0139] ;

[0140] In the formula, the pore water pressure μ(z) is closely related to the soil water characteristic curve, and its parameters include soil porosity n, initial and final saturation S0 and S2. f Rainfall Intensity I R , and empirical constants related to SWCC (such as α, λ, etc.), where α represents the scale parameter of soil matrix suction and λ represents the porosity distribution index.

[0141] S33. The ratio of soil shear strength to sliding shear stress, final infiltration, soil thickness, failure probability, and soil reliability index are used as physical information parameters of the target landslide area. The training dataset is optimized and merged to obtain the merged training dataset.

[0142] Specifically, based on GIS-FORM, the variable vector X is updated iteratively. k Transformation matrix T k Transformation matrix: A random variable is transformed into its equivalent mean variable MV, which is normally distributed. The correlation matrix R of the random variables can be used to approximate the most likely failure point using a multi-starting-point parallel search strategy. In the parameter space, based on the range of the variable's mean and standard deviation, Latin Hypercube Sampling (LHS) or uniform grid partitioning is used to generate N initial points (N≥5), ensuring that the initial solution is uniformly distributed in the parameter space. These N initial points are then used as the starting points for the GIS-FORM iteration, running simultaneously in a multi-threaded or distributed computing environment, recording the iteration path of each starting point and the current optimal F. S P fThe GIS-FORM algorithm combines with Particle Swarm Optimization (PSO). It selects the K best solutions in the current iteration as the initial particle swarm positions for PSO and uses the limit state function value of GIS-FORM as the fitness function of PSO. PSO searches for new candidate solutions in the parameter space through a global information sharing mechanism, prioritizes exploring high-risk areas that have not yet been covered, and automatically switches to particle swarm optimization when convergence is slow, thereby improving the search efficiency for the global optimum.

[0143] In addition to the Instantaneous Infiltration Module (STIM), the GIS-FORM used in this model is capable of analyzing F S Based on this, the impact of uncertainties in parameters such as soil cohesion, internal friction angle, slope, and rainfall intensity on slope stability is further evaluated. This is calculated using the following formula:

[0144] ;

[0145] ;

[0146] ;

[0147] ;

[0148] In the formula, vector x = (c s ,tanφ,β,n,γ sat Let f(x) represent the set of random variables that form the input parameters, and let f(x) represent the joint probability distribution function of the random variable x. s (x) represents the safety factor. Represents the standard normal cumulative distribution function. This represents the MPP value of the i-th random variable; and R represents the equivalent normal mean and standard deviation of the variable, R represents the correlation matrix, and T represents the correlation matrix. k Represents the transformation matrix. The equivalent normal standard deviation of the i-th random variable. Indicator of failure probability The limit state function is used to distinguish between the safety and failure of a structure / slope. This represents the joint probability density function of the input random variable set 𝑥. This represents the reliability index, and n represents the direction vector (normalized gradient vector). This represents the design point at the (k+1)th iteration. The limit state function g(x) represents the state at the k-th iteration point X. k The gradient vector at that point.

[0149] Furthermore, based on the inclusion of physical laws from the geotechnical engineering field in the model training process, and considering the deterministic and risk-based uncertainties in slope stability and failure probability assessment, F... S and The monotonic relationship between them can be expressed as:

[0150] ;

[0151] In the formula, P f Indicates the probability of failure. and Representing the predicted landslide susceptibility values ​​for two samples i and j respectively, the performance evaluation image of PRL-STIM over the entire physical set is shown below. Figure 7 As shown, This represents the safety coefficient of the i-th sample. Indicates the first Safety coefficient of each sample Let represent the failure probability of the i-th sample. Indicates the first The failure probability of each sample is calculated by ranking the characteristics of the training samples from high to low, based on the physical laws governing slope stability learned by the model. The formula is as follows:

[0152] ;

[0153] ;

[0154] ;

[0155] In the formula, This represents the difference between the model predictions of two adjacent samples. This indicates that negative values ​​violate the expected monotonic trend and are penalized. This represents the risk consistency loss, where n represents the number of samples.

[0156] S4. Based on the simplified transient infiltration model and the first-order reliability algorithm framework, a loss function containing physical consistency and risk consistency constraints is constructed.

[0157] Specifically, a landslide probability analysis framework (PRL-STIM) based on a simplified transient infiltration model and a first-order reliability algorithm is constructed, and a machine learning model loss function containing physical consistency and risk consistency constraints is built. The model is constructed from two aspects: the simplified transient infiltration model (STIM) and FORM based on the uncertainty and spatial variability of geotechnical parameters to quantify the probability of landslide failure.

[0158] As a preferred embodiment, the construction of the loss function that includes physical consistency and risk consistency constraints includes:

[0159] Constructing a scientific consistency loss based on physical consistency loss and probabilistic risk loss;

[0160] Construct a supervised loss to measure the difference between the prediction results of the probabilistic physical information machine learning model and the actual observations;

[0161] By combining scientific consistency loss with supervision loss, a loss function is constructed to measure the probabilistic physical information machine learning model.

[0162] It should be noted that, by incorporating supervision loss, physical consistency loss, and probabilistic risk loss, a balance between accuracy and scientific consistency can be achieved in landslide susceptibility prediction. The supervision loss function is used to optimize the model prediction results based on labeled landslide samples, and the formula for calculating the loss function is as follows:

[0163] ;

[0164] ;

[0165] ;

[0166] In the formula, This indicates the actual observed landslide occurrence (label). This represents the landslide probability predicted by the model. This represents the scientific consistency loss, which combines physics-based loss and risk-based consistency loss, with a regularization coefficient λ. phy and λ ris k controls the degree of influence of physical constraints and risk constraints on model training, respectively, and λ phy Strengthen the protection of F S Learning monotonicity relations (L) phy This ensures that the model adheres to basic slope stability principles. λ risk Strengthen the protection of P f The responsiveness of the model enables its output to provide a reliable probabilistic risk interpretation. The loss function represents the probabilistic physical information machine learning model. Let represent the supervised loss, n represent the total number of samples, and i represent the index of the i-th sample. For example... Figure 10 As shown, when the value of λ increases (the value of λ here is the regularization coefficient λ in the formula), phy With λ risk , respectively corresponding to physical consistency loss L phy Consistency loss with risk L risk (weighting), the SI index decreased significantly, such as Figure 11The impact of the loss function guided by scientific consistency on model predictions shows that the model predictions more closely follow the parameters set by F. S and P f The defined monotonicity relation (i.e., F) S The higher or P f The lower the value, the lower the landslide susceptibility. Simultaneously, classification performance also improves, reaching its optimum around λ1=λ2≈0.5, where λ1 represents the regularization coefficient λ for physical consistency loss. phy, λ² represents the regularization coefficient λ of the risk consistency loss. risk The optimal point is reached in the vicinity.

[0167] S5. Based on the merged training dataset, a probabilistic physical information machine learning model is trained using a loss function, and the probabilistic physical information machine learning model is optimized using the spatial cross-validation method.

[0168] Specifically, various topographic factors derived from the Digital Elevation Model (DEM) (such as slope, curvature, and topographic moisture index) are combined with the geographical location features of landslide points as multidimensional feature vectors to construct a corresponding model. Five-fold spatial cross-validation (SCCV) is used, where the model is trained on four clusters and validated on the remaining cluster to test its predictive ability for unseen areas. Spatial CV is repeated five times under different random seeds, and the prediction results from the five rounds of CV are fused using an average ensemble strategy. To improve the model's generalization stability under different spatial partitions, an average ensemble strategy is introduced. Let p be the prediction result of the i-th round of spatial CV across all grid cells. i (x), the mean of the predicted values ​​from multiple rounds of cross-validation can reduce the variance caused by differences in data partitioning, thus obtaining more robust prediction results.

[0169] After obtaining five rounds of predictions, the prediction standard deviation for each grid cell is calculated. A landslide susceptibility distribution map is generated, and the prediction standard deviation is calculated to quantify model uncertainty. The standard deviation, as a quantitative indicator of model uncertainty, reflects the stability of prediction results under different training / validation partitions. Samples in high-uncertainty regions are resampled and weighted in the next round of training, or the physical constraint coefficients are adjusted to enhance the model's constraint strength in that region, thereby reducing uncertainty. The PPM model outputs data; the entire process is PIML model development, using a PIML model trained based on spatial CV, multiple averaging experiments, and generating a landslide susceptibility model map.

[0170] Among them, such as Figure 6 This is a spatial distribution map of landslide susceptibility generated based on the Probabilistic Physical Model (PRL-STIM). The physical model (PRL-STIM) is as follows: Figure 2 Spatial distribution results show that F SThe low-value area highly coincides with the locations of landslides that have occurred in the landslide inventory, P f The high-value areas also exhibit clustering characteristics consistent with the actual landslide distribution.

[0171] in, Figure 2 In the diagram, β represents the slope angle; H represents the initial groundwater depth; Z w Indicates the depth of the groundwater level after infiltration; L represents the length of the infiltration influence; W represents the distance from the wetting front to the groundwater surface; T represents the thickness of the wetting zone; u w Indicates pore water pressure; γ w u represents the unit weight of water. w / γ w This indicates the pore water pressure head. w / γ w =-H·cos 2 β represents the pore water pressure distribution before infiltration; u w / γ w =z w cos 2 β represents the pore water pressure distribution after infiltration; -h c This indicates matrix suction.

[0172] As a preferred embodiment, the step of training a probabilistic physical information machine learning model based on a merged training dataset and a loss function, and then optimizing the probabilistic physical information machine learning model using spatial cross-validation includes the following steps:

[0173] S51. Use the digital elevation model to derive various terrain factors and extract the geographical location features of landslide points to obtain multi-dimensional feature vectors. Add the multi-dimensional feature vectors to the merged training dataset to obtain the integrated training dataset.

[0174] S52. Input the integrated training dataset into the probabilistic physical information machine learning model, and train the probabilistic physical information machine learning model by combining the loss function.

[0175] S53. Optimize the trained probabilistic physical information machine learning model to obtain the optimized probabilistic physical information machine learning model.

[0176] In a preferred embodiment, optimizing the trained probabilistic physical information machine learning model to obtain an optimized probabilistic physical information machine learning model includes the following steps:

[0177] S531. Based on the spatial distribution characteristics of the target landslide area, the spatial region corresponding to the integrated training dataset is divided into several clusters;

[0178] S532. Divide one group of clusters into a validation set and the remaining clusters into a training set, and re-perform multiple rounds of cross-validation on the trained probabilistic physical information machine learning model.

[0179] S533. Calculate the index of the prediction results of multiple rounds of cross-validation, and optimize the parameters of the probabilistic physical information machine learning model based on the index calculation results to obtain the optimized probabilistic physical information machine learning model.

[0180] It should be noted that spatial cross-validation is used to evaluate the generalization ability of machine learning models (NN models) in different clustering regions, such as... Figure 8 The model's predictive performance across various spatial clusters is demonstrated, compared to the physically constrained PRL-STIM model (see [link to documentation]). Figure 7 The overall performance of the NN model is slightly lower. After selecting one round of spatial cross-validation, the process is repeated. Based on the five sub-models in this round, a final model is constructed through average ensemble, denoted as the NN ensemble model. The NN ensemble model is adaptively weighted based on the performance (e.g., AUC, SI metrics) of each sub-model on the spatial validation set. Let the performance of the k-th sub-model be π. k The weight is , This represents the performance of the j-th sub-model.

[0181] The final prediction result is: , The predicted outputs of each sub-model are represented, and a dynamic weighted ensemble is constructed. For further optimization, a meta-model ensemble can be selected to construct a secondary training set. The predicted outputs of each sub-model [P1(x),…,P5(x)] are used as inputs, and the real landslide labels are used as outputs to train the meta-model and learn the optimal combination. The training process can be designed in stages. Stage 1: Single-modal feature pre-training is performed for spatial features (such as geographical location and slope), semantic features (such as landform classification and land use type), and dynamic features (such as rainfall intensity time series and NDVI time series) to ensure that the spatial representation of each modality feature is sufficient.

[0182] Phase Two: A multimodal fusion mechanism is introduced during meta-model training. The sub-model predictions and multimodal features are jointly input into the meta-model. Through attention or gating mechanisms, the contribution of different modalities is adaptively learned, resulting in an ensemble model with stronger spatial awareness. Specifically, the sub-model prediction output vector [P1(x),…,P5(x)] is combined with the spatial feature embedding f... sp Semantic feature embedding f sem Dynamic feature embedding f dyn All modal embeddings are input into the meta-model. Each modality embedding is first mapped to a unified dimension, and then weighted and fused using an attention or gating mechanism to obtain the fused vector f. fusion , will ffusion The prediction results from the sub-models are then integrated into the meta-model.

[0183] Alternatively, asynchronous updates or alternating optimization methods can be adopted. This involves fixing the sub-model prediction weights in the early stages of training and optimizing only the semantic-spatial feature fusion parameters. Subsequently, the sub-model weights and fusion parameters are gradually unfrozen for joint training, improving model stability and avoiding overfitting. During training, a phased unfreezing strategy is preferred: Phase one freezes the sub-model weights and trains only the fusion module to ensure sufficient representation of different modalities; Phase two gradually unfreezes the sub-models using a learning rate scaling factor λ. t The update rate is controlled to smoothly transition from 0 to a normal learning rate; the three-stage training method involves joint training of the sub-models and the fusion module. This method overcomes the limitation of traditional average ensemble methods that only linearly combine the outputs of sub-models, achieving dynamic adjustment of cross-modal information coupling and spatial awareness, significantly improving the generalization ability and physical interpretability of cross-regional landslide susceptibility prediction. This meta-model training method breaks through the limitation of traditional ensemble methods that only perform simple linear combination of sub-model outputs, enabling dynamic adjustment of cross-modal information coupling and spatial awareness, significantly improving the generalization performance and physical interpretability of cross-regional landslide susceptibility prediction. Furthermore, additional features such as spatial cluster location and terrain factors can be combined to obtain an ensemble model with spatial awareness capabilities.

[0184] Through the above improvements, the ensemble prediction results can better reflect the differences in performance of sub-models across different spatial clusters, improve cross-regional generalization performance, and avoid the performance degradation problem caused by average ensemble. Figure 9 The landslide susceptibility map of the study area was generated by the NN ensemble model based on a single spatial CV repetition. The average AUC of the NN ensemble model in the independent test area was only 0.68, which was significantly lower than the fitting performance on the training set. It had limited generalization ability to geographical distribution changes and was difficult to effectively transfer to unseen areas.

[0185] Based on dataset preparation and spatial cross-validation, random samples are combined with extracted coordinate environmental factors. Spatial cross-validation (spatial CV) is further used to visualize the feature space using t-SNE nonlinear dimensionality reduction technology. Based on the confusion matrix, four types of prediction results are calculated (true positive TP, true negative TN, false positive FP, and false negative FN), and five key performance indicators are derived.

[0186] Specifically, based on dataset preparation and spatial cross-validation, random samples are combined with extracted coordinate environment factors. Spatial cross-validation (spatial CV) is then used, followed by t-SNE nonlinear dimensionality reduction for visual analysis of the feature space. Four types of prediction results are calculated using the following formula:

[0187] ;

[0188] ;

[0189] The five key classification indicators are accuracy, precision, recall, F1 score, and scientific inconsistency (SI). TPR is the true positive rate, and FPR is the false positive rate. True positive (TP) refers to the number of landslide samples correctly classified as landslides, and true negative (TN) refers to the number of non-landslide samples correctly classified as non-landslides. False positive (FP) represents the number of non-landslide samples misclassified as landslides, and false negative (FN) represents the number of landslide samples misclassified as non-landslides. Indicators of scientific inconsistency. For example... Figure 12 As shown in the ROC results of spatial cross-validation, the PIML model exhibits strong classification ability in all spatial clustering regions, with an average AUC of 0.75, outperforming the probabilistic physics model and the machine learning model. The three models are further illustrated in Table 1. In practical applications, these metrics are not only used for model validation but also as criteria for rapid evaluation: for example, when Accuracy and AUC reach preset thresholds, a regional landslide susceptibility distribution map can be directly generated; when SI is high, the rapid correction modules TPR and FPR are triggered, which can be used to dynamically set the risk zoning threshold, thus forming a rapid evaluation mechanism that balances predictive performance and physical consistency.

[0190] Table 1 Performance of the three models

[0191] Specifically, the PIML model introduces a physical consistency constraint (λ). phy >0) or risk consistency constraint (λ) risk After λ > 0, the prediction results show that the model is not only clearer and more ordered in spatial distribution, but also that the prediction uncertainty is significantly reduced, and the generalization performance is significantly improved. phy =λ risk With a joint constraint setting of 0.5, the PIML ensemble model exhibits optimal performance.

[0192] S6. Use the optimized probabilistic physical information machine learning model to assess the landslide susceptibility of the target landslide area and obtain the assessment results.

[0193] In summary, by leveraging the above-mentioned technical solutions of this invention, the present invention effectively bridges the contradiction between data dependence and scientific consistency in traditional machine learning modeling of geological disasters by integrating a regularization mechanism based on physical stability and probabilistic risk information. This provides a more robust and reliable modeling path for regional landslide risk assessment and offers a basis for decision-making by relevant departments of land and resources management. This invention proposes a spatial distribution map of landslide susceptibility generated based on a probabilistic physical model, employs a spatial cross-validation strategy to validate the model, and further quantifies and evaluates the predictive performance of the PRL-STIM model. This addresses the significant instability and insufficient generalization ability of traditional pure data-driven models when dealing with highly spatially heterogeneous regions. In contrast, the model incorporating physical constraints exhibits stronger robustness in cross-regional migration. The PIML model used in this invention, as the regularization strength increases, enhances the guiding role of physical and probabilistic constraints, and the model output gradually tends to be more consistent with the prediction results of geotechnical engineering principles. Based on the advantages of integrating data-driven learning and probabilistic physical modeling, the PIML model takes into account both prediction performance and scientific consistency, and shows good generalization and interpretability. The PIML model has strong classification ability in various spatial clustering regions, which is superior to probabilistic physical models and machine learning models.

[0194] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0195] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A physical information-guided machine learning landslide susceptibility assessment method oriented to scientific consistency, characterized in that, The method includes the following steps: S1. Obtain geographic data of the target landslide area and input the geographic data into open-source geographic information system software. Use buffer analysis to construct a preliminary training dataset. S2. Optimize the initial training dataset to obtain the optimized training dataset; S3. Obtain soil parameters and rainfall parameters of the target landslide area, and calculate the physical information parameters of the target landslide area in combination with the instantaneous infiltration module model. Then, merge the physical information parameters into the optimized training dataset to obtain the merged training dataset. S4. Based on the simplified transient infiltration model and the first-order reliability algorithm framework, a loss function containing physical consistency and risk consistency constraints is constructed. S5. Based on the merged training dataset, a probabilistic physical information machine learning model is trained using a loss function, and the probabilistic physical information machine learning model is optimized using the spatial cross-validation method. S6. Use the optimized probabilistic physical information machine learning model to assess the landslide susceptibility of the target landslide area and obtain the assessment results.

2. The physically-informed machine learning landslide susceptibility assessment method oriented to scientific consistency according to claim 1, characterized in that, The process of acquiring geographic data of the target landslide area, inputting the geographic data into open-source geographic information system software, and constructing a training dataset using buffer analysis includes the following steps: S11. Input the geographic data of the target landslide area into the open-source geographic information system software, and based on the spatial analysis function of the open-source geographic information system software, construct a buffer zone with a preset radius centered on the preset landslide point. S12. Remove all grid cells within the buffer zone from the total area of ​​the target landslide area, and use a stratified random sampling method to stratify the remaining area after the buffer zone is removed according to slope grade and lithology. S13. Within each layer, randomly select geographical units that do not overlap with the preset landslide samples as negative samples, and combine the geographical coordinates of the negative samples to construct a training dataset.

3. The science-consistency-oriented physical information-guided machine learning landslide susceptibility assessment method according to claim 1, characterized in that, The optimization process for the initial training dataset to obtain the optimized training dataset includes the following steps: S21. Based on each sample point in the preliminary training dataset, select several geographic entities according to a preset range; S22. Based on the selected geographic entities, calculate the semantic similarity between each sample point and the preset standard geographic entities; S23. Based on the selected geographic entities, perform graph convolution aggregation on the samples to obtain sample node representations, and calculate the semantic consistency probability based on the sample node representations. S24. The semantic similarity and semantic consistency probability are weighted and fused to obtain the weighted fusion result. The sample points are judged to be semantically abnormal by using the preset threshold, and the sample points with semantically abnormality are removed to obtain the secondary training dataset. S25. Using the topological consistency test and combined with the M-estimation residual screening method, false neighboring points caused by topological barriers in the secondary training dataset are removed to obtain the optimized dataset.

4. The science-consistency-oriented physical information-guided machine learning landslide susceptibility assessment method according to claim 3, characterized in that, The step of calculating the semantic similarity between each sample point and a preset standard geographic entity based on the selected geographic entities includes the following steps: S221. Encode geographic entities into vector representations and perform semantic embedding on the geographic entity encoding results to generate entity semantic embedding vectors. S222. Based on the spatial distance between each sample point and the geographic entity, and combined with the entity semantic embedding vector, calculate the domain weight and normalize it to obtain the sample semantic vector. S223. Perform cosine similarity calculation between the sample semantic vector and the preset standard prototype vector to obtain the semantic similarity between the sample point and the standard geographic entity.

5. The landslide susceptibility assessment method guided by physical information based on scientific consistency as described in claim 3, characterized in that, The process of performing graph convolution aggregation on the samples based on the selected geographic entities to obtain sample node representations, and calculating the semantic consistency probability based on the sample node representations, includes the following steps: S231. Determine the spatial adjacency relationship based on the selected geographic entities, and construct a heterogeneous graph with the spatial adjacency relationship as the edge and the selected geographic entity as the vertex. S232. Based on the constructed heterogeneous graph, the sample node representation is obtained by using a relational graph convolutional network for aggregation processing. S233. Based on the sample node representation and combined with the preset nonlinear activation, calculate the semantic consistency probability.

6. The landslide susceptibility assessment method guided by physical information based on scientific consistency as described in claim 1, characterized in that, The process of obtaining soil and rainfall parameters for the target landslide area, calculating physical information parameters of the target landslide area using the instantaneous infiltration module model, and merging these physical information parameters into the optimized training dataset to obtain the merged training dataset includes the following steps: S31. Obtain soil parameters for the target landslide area, analyze the ratio of soil shear strength to sliding shear stress and soil thickness using the instantaneous infiltration module model, and calculate the failure probability and soil reliability index by combining the ratio of soil shear strength to sliding shear stress. S32. Obtain rainfall parameters during rainfall in the target landslide area and calculate the final infiltration amount using the infiltration calculation formula; S33. The ratio of soil shear strength to sliding shear stress, final infiltration, soil thickness, failure probability, and soil reliability index are used as physical information parameters of the target landslide area. The training dataset is optimized and merged to obtain the merged training dataset.

7. The landslide susceptibility assessment method guided by physical information for scientific consistency as described in claim 6, characterized in that, Before analyzing the ratio of soil shear strength to sliding shear stress and soil thickness using the instantaneous infiltration module model, the following steps are taken: dynamically correcting the soil parameters of the target landslide area using the parameter adaptive analysis method. The method of dynamically correcting soil parameters in the target landslide area using adaptive parameter analysis includes: Based on the fitting residual of the measured rainfall infiltration rate and soil moisture content curve, the soil saturation unit weight parameter is adjusted in real time using the exponential decay function. Based on the changes in water content at different depths from soil profile moisture sensors, the advance rate of the wetting front is calculated, and the depth parameters of the wetting front are updated. By measuring the suction and moisture content data of the soil matrix, the key parameters of the soil water characteristic curve are updated using a real-time inversion method, and the relationship between pore water pressure and depth is corrected. Based on the correction results, the pore water pressure parameters are adjusted.

8. The landslide susceptibility assessment method guided by physical information based on scientific consistency as described in claim 1, characterized in that, The loss function constructed, which includes constraints on physical consistency and risk consistency, includes: Constructing a scientific consistency loss based on physical consistency loss and probabilistic risk loss; Construct a supervised loss to measure the difference between the prediction results of the probabilistic physical information machine learning model and the actual observations; By combining scientific consistency loss with supervision loss, a loss function is constructed to measure the probabilistic physical information machine learning model.

9. The landslide susceptibility assessment method guided by physical information based on scientific consistency as described in claim 1, characterized in that, The process of training a probabilistic physical information machine learning model based on a merged training dataset and a loss function, and then optimizing the model using spatial cross-validation, includes the following steps: S51. Use the digital elevation model to derive various terrain factors and extract the geographical location features of landslide points to obtain multi-dimensional feature vectors. Add the multi-dimensional feature vectors to the merged training dataset to obtain the integrated training dataset. S52. Input the integrated training dataset into the probabilistic physical information machine learning model, and train the probabilistic physical information machine learning model by combining the loss function. S53. Optimize the trained probabilistic physical information machine learning model to obtain the optimized probabilistic physical information machine learning model.

10. The landslide susceptibility assessment method guided by physical information based on scientific consistency as described in claim 9, characterized in that, The process of optimizing the trained probabilistic physical information machine learning model to obtain the optimized probabilistic physical information machine learning model includes the following steps: S531. Based on the spatial distribution characteristics of the target landslide area, the spatial region corresponding to the integrated training dataset is divided into several clusters; S532. Divide one group of clusters into a validation set and the remaining clusters into a training set, and re-perform multiple rounds of cross-validation on the trained probabilistic physical information machine learning model. S533. Calculate the index of the prediction results of multiple rounds of cross-validation, and optimize the parameters of the probabilistic physical information machine learning model based on the index calculation results to obtain the optimized probabilistic physical information machine learning model.

Images