Landslide disaster estimation method and device combining machine learning and physical mechanism

Abstract

The invention provides a machine learning and physical mechanism combined landslide disaster estimation method and device, and relates to the technical field of landslide disaster estimation, and the method comprises the steps: obtaining a landslide disaster basic data set of a target region; inputting the landslide disaster basic data set into a pre-constructed joint model; identifying a landslide identification result in the target area according to the landslide disaster basic data set through a landslide identification model; simulating a landslide disaster dynamic evolution process in the target area according to the landslide disaster basic data set and the landslide recognition result through a landslide fine analysis model to obtain a landslide simulation result; and fusing the landslide identification result and the landslide simulation result to obtain a landslide disaster estimation result corresponding to the target area. According to the invention, the future landslide disaster condition can be predicted more accurately and efficiently.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide disaster prediction, and in particular to a landslide disaster prediction method and device combining machine learning with physical mechanisms. Background Art

[0002] Landslides are common geological disasters worldwide, with complex mechanisms and diverse influencing factors. Existing landslide prediction methods are mostly based on single machine learning or physical models. However, single machine learning models and single physical models have their own limitations: although the former is good at processing big data and rapid prediction, it lacks the ability to explain the mechanism of landslide disasters; the latter can accurately simulate physical processes, but is limited by data acquisition, computing efficiency and dynamic simulation capabilities. The defects of both indicate that they cannot meet the comprehensive requirements of accuracy, efficiency and interpretability for landslide disaster warning. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a landslide disaster prediction method and device that combines machine learning and physical mechanisms, which can more accurately and efficiently predict future landslide disasters.

[0004] In a first aspect, the present invention provides a landslide hazard prediction method combining machine learning and physical mechanism, comprising:

[0005] Obtain basic data sets of landslide hazards in the target area;

[0006] Input the basic data set of landslide hazards into the pre-built joint model, which includes a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanisms;

[0007] Through the landslide identification model, the landslide identification results in the target area are identified according to the basic data set of landslide disasters. The landslide identification results are used to describe the probability of landslide disasters occurring in the sub-areas of the target area.

[0008] Through the landslide fine analysis model, the dynamic evolution process of landslide disasters in the target area is simulated according to the basic data set of landslide disasters and the landslide identification results, and the landslide simulation results are obtained. The landslide simulation results are used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-areas in the target area;

[0009] The landslide identification results and landslide simulation results are integrated to obtain the landslide hazard prediction results corresponding to the target area; among them, the landslide hazard prediction results are used to describe the landslide hazard levels corresponding to the sub-areas within the target area.

[0010] In one embodiment, before identifying the landslide identification result in the target area according to the landslide hazard basic data set through the landslide identification model, the method further includes:

[0011] Obtain historical landslide disaster basic data sets and historical landslide disaster event data;

[0012] Preprocess the historical landslide disaster dataset and historical landslide disaster event data;

[0013] Based on the preprocessed historical landslide disaster data set, data features that can express the triggering mechanism of landslide events are constructed, and the data features are used as model inputs, and the preprocessed historical landslide disaster event data is used as training labels to construct the initial training data set;

[0014] The landslide samples in the initial training data set are expanded to balance the ratio between landslide samples and non-landslide samples to obtain the target training data set;

[0015] A plurality of pre-built decision trees are obtained, and a target training data set is randomly sampled for each decision tree to obtain a training data subset corresponding to each decision tree. The training data subset is used to train the corresponding decision tree, and each trained decision tree is integrated to obtain a landslide recognition model.

[0016] In one embodiment, the landslide samples in the initial training data set are expanded, including:

[0017] For each landslide sample in the initial training dataset, perform the following operations:

[0018] Performing one or more of rotation processing, cropping processing, and scaling processing on the landslide sample;

[0019] And / or, a plurality of adjacent landslide samples corresponding to the landslide sample are determined from the initial training data set, and the landslide sample and each of the adjacent landslide samples are fused respectively to obtain a plurality of new landslide samples.

[0020] In one embodiment, before simulating the dynamic evolution process of landslide disasters in the target area according to the landslide disaster basic data set and the landslide identification result by using the landslide fine analysis model to obtain the landslide simulation result, the method further includes:

[0021] Obtain a differentiated model parameter library corresponding to the target area, which includes model parameters matching different geological types and meteorological conditions, including geomechanical parameters, rainfall triggering condition parameters, and groundwater dynamics parameters;

[0022] Based on the differentiated model parameter library, multiple landslide detailed analysis models are established according to different geological types and meteorological conditions in the target area.

[0023] In one embodiment, the dynamic evolution process of landslide disasters in the target area is simulated according to the basic data set of landslide disasters and the landslide identification results through the landslide fine analysis model to obtain the landslide simulation results, including:

[0024] Based on the landslide identification results, high-risk sub-areas for landslides are extracted from the target area. The high-risk sub-areas for landslides are sub-areas where the probability of landslide disasters is higher than a preset probability threshold.

[0025] Taking the scope of the high-risk sub-area for landslides as the boundary condition, a fine landslide analysis model that matches the geological type and meteorological conditions corresponding to the high-risk sub-area for landslides is used to simulate the dynamic evolution process of landslide hazards in the high-risk sub-area for landslides according to the basic data set of landslide hazards, and the landslide simulation results are obtained.

[0026] In one embodiment, the landslide identification result and the landslide simulation result are integrated to obtain the landslide hazard prediction result corresponding to the target area, including:

[0027] Based on the pre-configured weight values, the landslide identification results and the landslide simulation results are superimposed, and the landslide hazard level corresponding to the sub-area within the target area is determined according to the superimposed results.

[0028] In a second aspect, the present invention further provides a landslide disaster prediction device combining machine learning and physical mechanism, comprising:

[0029] Data acquisition module, used to obtain the basic data set of landslide hazards in the target area;

[0030] A data input module is used to input the basic data set of landslide hazards into a pre-built joint model, which includes a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanisms;

[0031] The landslide identification module is used to identify the landslide identification results in the target area according to the landslide hazard basic data set through the landslide identification model. The landslide identification results are used to describe the probability of landslide disasters occurring in each sub-area in the target area.

[0032] The landslide fine analysis module is used to simulate the dynamic evolution process of landslide disasters in the target area according to the landslide disaster basic data set and landslide identification results through the landslide fine analysis model, and obtain the landslide simulation results. The landslide simulation results are used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-area in the target area;

[0033] The landslide prediction module is used to fuse the landslide identification results and the landslide simulation results to obtain the landslide hazard prediction results corresponding to the target area; wherein the landslide hazard prediction results are used to describe the landslide hazard levels corresponding to the sub-areas within the target area.

[0034] In one embodiment, a landslide identification model building module is further included, which is used to:

[0035] Obtain historical landslide disaster basic data sets and historical landslide disaster event data;

[0036] Preprocess the historical landslide disaster dataset and historical landslide disaster event data;

[0037] Based on the preprocessed historical landslide disaster data set, data features that can express the triggering mechanism of landslide events are constructed, and the data features are used as model inputs, and the preprocessed historical landslide disaster event data is used as training labels to construct the initial training data set;

[0038] The landslide samples in the initial training data set are expanded to balance the ratio between landslide samples and non-landslide samples to obtain the target training data set;

[0039] A plurality of pre-built decision trees are obtained, and a target training data set is randomly sampled for each decision tree to obtain a training data subset corresponding to each decision tree. The training data subset is used to train the corresponding decision tree, and each trained decision tree is integrated to obtain a landslide recognition model.

[0040] In a third aspect, the present invention further provides an electronic device, comprising a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement any one of the methods provided in the first aspect.

[0041] In a fourth aspect, the present invention further provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement any one of the methods provided in the first aspect.

[0042] The present invention provides a landslide disaster prediction method and device combining machine learning with physical mechanism, firstly, a basic data set of landslide disasters in a target area is obtained; then, the basic data set of landslide disasters is input into a pre-built joint model, the joint model comprising a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanism; then, the landslide identification model is used to identify a landslide identification result in the target area according to the basic data set of landslide disasters, and the landslide identification result is used to describe the probability of landslide disasters occurring in a sub-area in the target area; then, the landslide fine analysis model is used to simulate the dynamic evolution process of landslide disasters in the target area according to the basic data set of landslide disasters and the landslide identification result, and obtain a landslide simulation result, and the landslide simulation result is used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-area in the target area; finally, the landslide identification result and the landslide simulation result are fused to obtain a landslide disaster prediction result corresponding to the target area; wherein the landslide disaster prediction result is used to describe the landslide disaster level corresponding to the sub-area in the target area. The above method preliminarily identifies the probability of landslide disasters occurring in sub-areas within the target area through a landslide identification model based on machine learning, and on this basis simulates the dynamic evolution process of landslide disasters in the target area through a landslide fine analysis model based on physical intelligence, and estimates the landslide hazard level in the target area by combining the landslide identification results output by the landslide identification model and the landslide simulation results output by the landslide fine analysis model. The joint model provided by the present invention combines the advantages of machine learning models and physical models, and demonstrates powerful capabilities in landslide disaster prediction. Through the joint model, the rapid screening function of machine learning can be used to significantly narrow the analysis scope, and the physical mechanism can be used to perform detailed analysis and dynamic simulation of high-risk sub-areas for landslides, thereby making up for the limitations of a single model, comprehensively improving the efficiency, accuracy and scientificity of landslide prediction, and providing reliable decision support for landslide disaster risk assessment, early warning and emergency response.

[0043] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 A schematic flow chart of a landslide disaster prediction method combining machine learning and physical mechanism provided in an embodiment of the present invention;

[0047] Figure 2 A technical framework diagram for constructing a joint model provided by an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of the structure of a landslide disaster prediction device combining machine learning and physical mechanism provided by an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] At present, the existing landslide prediction models mainly include machine learning models and physical models, but they each have their own shortcomings. Machine learning models rely on the quality and quantity of data, lack the physical explanation of landslide triggering mechanisms, and are difficult to deal with abnormal or unseen data situations; although physical models have clear mechanism explanations, they have limited requirements for high-precision parameters and numerical simulation capabilities for complex geological conditions, have high computational costs, and are highly dependent on input data.

[0052] Based on this, the present invention provides a landslide disaster prediction method and device that combines machine learning with physical mechanisms. The joint model combines the advantages of machine learning and physical mechanisms. It can not only quickly identify high-risk areas for landslides through machine learning, but also use physical models to deeply analyze key triggering mechanisms. It has both high efficiency and scientificity, greatly improves prediction accuracy and generalization ability, and provides more reliable technical support for landslide disaster prevention and control.

[0053] To facilitate understanding of this embodiment, a landslide disaster prediction method combining machine learning and physical mechanism disclosed in an embodiment of the present invention is first described in detail. Figure 1 The flowchart of a landslide disaster prediction method combining machine learning and physical mechanism is shown, and the method mainly includes the following steps S102 to S110:

[0054] Step S102, obtaining a basic data set of landslide hazards in the target area.

[0055] Among them, the basic data sets of landslide disasters include terrain data, geological data, meteorological data, human activity data, and remote sensing data.

[0056] Step S104, input the basic data set of landslide disasters into the pre-built joint model. The joint model includes a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanism; the input of the landslide identification model is the basic data set of landslide disasters, and the output is the landslide identification result; the input of the landslide fine analysis model is the basic data set of landslide disasters and the landslide identification result, and the output is the landslide simulation result.

[0057] Step S106, identifying the landslide identification results in the target area according to the landslide disaster basic data set through the landslide identification model. The landslide identification model based on machine learning can adopt a random forest model; the landslide identification results are used to describe the probability of landslide disasters occurring in sub-areas within the target area.

[0058] Step S108, through the landslide fine analysis model, according to the landslide disaster basic data set and the landslide identification result, the dynamic evolution process of the landslide disaster in the target area is simulated to obtain the landslide simulation result. Among them, the landslide simulation result is used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-area in the target area. In one example, the high-risk sub-area of ​​landslide in the target area can be extracted based on the landslide identification result. The high-risk sub-area of ​​landslide is a sub-area where the probability of landslide disaster is higher than the preset probability threshold. The range of the high-risk sub-area of ​​landslide is used as the boundary condition, and the corresponding landslide fine analysis model is used to simulate the dynamic evolution process of the landslide disaster in the high-risk sub-area of ​​landslide according to the landslide disaster basic data set, so as to output the landslide simulation result corresponding to the high-risk sub-area of ​​landslide.

[0059] Step S110, the landslide identification result and the landslide simulation result are merged to obtain the landslide hazard prediction result corresponding to the target area. The landslide hazard prediction result is used to describe the landslide hazard level corresponding to the sub-area in the target area. In one example, the landslide identification result and the landslide simulation result can be superimposed using a pre-configured weight value, and the landslide hazard level corresponding to the sub-area in the target area can be evaluated based on the superimposed result.

[0060] The landslide disaster prediction method combining machine learning and physical mechanism provided in the embodiment of the present invention preliminarily identifies the probability of landslide disasters occurring in sub-areas within the target area through a landslide recognition model based on machine learning, and on this basis simulates the dynamic evolution process of landslide disasters in the target area through a landslide fine analysis model based on physical intelligence, and estimates the landslide disaster level of the target area by combining the landslide recognition result output by the landslide recognition model and the landslide simulation result output by the landslide fine analysis model. The joint model provided by the present invention combines the advantages of machine learning models and physical models, and demonstrates powerful capabilities in landslide disaster prediction. Through the joint model, the rapid screening function of machine learning can be used to significantly narrow the analysis scope, and the physical mechanism can be used to perform detailed analysis and dynamic simulation of high-risk sub-areas for landslides, thereby making up for the limitations of a single model, comprehensively improving the efficiency, accuracy and scientificity of landslide prediction, and providing reliable decision support for landslide disaster risk assessment, early warning and emergency response.

[0061] For ease of understanding, the present invention first explains the process of constructing a joint model. Figure 2 A technical framework diagram for constructing a joint model is shown.

[0062] (1) The random forest machine learning model (i.e., landslide identification model) trained to identify high-risk landslide areas includes:

[0063] (1.1) Obtaining historical landslide disaster basic data sets and historical landslide disaster event data. The historical landslide disaster basic data sets include topographic data, geological data, meteorological data, human activity data, and remote sensing data at historical moments. In one example, based on the landslide historical record data, the area and time corresponding to the landslide disaster can be obtained; based on the corresponding area and time, the topographic data, geological data, meteorological data, human activity data, and remote sensing data of the area at the corresponding time can be obtained.

[0064] Topographic data are generated by remote sensing satellites (such as SRTM, ASTER) or surveying and mapping technology. The main parameters are: slope, which is the potential gravity driving force that affects surface sliding; slope direction, which determines the impact of rainfall infiltration and sunlight exposure on landslide stability; terrain undulation, which characterizes the degree of undulation of the regional surface and is often used to delineate high-risk terrain areas; gully density, which indicates the density of water flow convergence and reflects the strength of erosion.

[0065] The geological data comes from geological survey reports, geological radar detection, and drilling data. The main parameters are: rock and soil properties, cohesion (C), friction angle (φ) characterize the ability of rock and soil to resist sliding, porosity and permeability coefficient, which affect water penetration and the saturation state of the sliding surface; geological structure, fault distribution, fold belt and tectonic line location, which have a significant potential triggering effect on landslides; soil layer thickness, weak cover layer or thick loose soil layer is more likely to cause landslides;

[0066] Meteorological data comes from meteorological station observation data, numerical weather forecast data, and satellite precipitation products (such as GPM and TRMM). The main parameters are: rainfall characteristics, heavy rainfall (such as maximum rainfall in 24 hours) and cumulative rainfall; rainfall intensity, which affects the infiltration rate of rainwater and the formation of surface runoff; temperature, high temperature may increase the risk of soil expansion, while low temperature may cause landslides due to freeze-thaw effects; evaporation and water balance, long-term changes in soil moisture have a greater impact on slope stability.

[0067] The sources of human activity data are socio-economic statistics, land use maps, and engineering construction records. The main parameters are: land use type, farmland, construction areas, mining areas and other human activity areas may lead to slope damage and soil instability; transportation network, road construction and its ancillary facilities disturb the landslide area; engineering activities, large-scale excavation, slope reshaping and other activities that change the slope structure; population density, high population density may lead to more serious consequences of landslide disasters.

[0068] Remote sensing data comes from high-resolution remote sensing images, LiDAR data, and synthetic aperture radar (SAR). The main parameters are: vegetation coverage, vegetation has a certain effect on soil stability, and vegetation destruction may increase the risk of landslides; surface changes, by comparing remote sensing images of different periods to identify surface changes before and after landslides; landslide feature extraction, through synthetic aperture radar (SAR) images to monitor surface deformation and identify potential landslide areas; surface moisture, soil moisture data obtained by remote sensing is an important parameter for assessing the risk of rainfall-induced landslides.

[0069] The sources of landslide historical records are landslide event databases, literature and historical records. The main parameters are: time and location of occurrence, geographical location and triggering time of landslide, which are helpful for analyzing the temporal and spatial relationship between rainfall and landslide; landslide type, different types of landslides have different triggering mechanisms and impact ranges. Landslide scale, landslide volume, damage range and depth are used to assess the intensity of disasters; triggering conditions, specific causes such as rainfall, earthquakes or human activities; consequence description, including casualties, economic losses and environmental impacts, are used for disaster risk analysis.

[0070] (1.2) Preprocess the historical landslide disaster data set and historical landslide disaster event data. Preprocessing can include format conversion, time-space matching, data cleaning (or void elimination), data standardization (or dimensionless processing) and feature selection.

[0071] Format conversion: Based on landslide history records, terrain data, geological data, meteorological data, human activity data, and remote sensing data, data formats (such as raster or vector data) can be converted into model-compatible tabular forms, and interpolation or other methods can be used to process data missing areas to build a data set that is convenient for model training.

[0072] Time and space matching: Time and space matching is performed based on the landslide occurrence point to unify the resolution and coordinate system.

[0073] Data cleaning (or elimination of gaps): By checking the original information of each landslide disaster event, such as the site name, longitude, latitude, landslide disaster event start date, landslide disaster event end date and landslide disaster event intensity index, the information that is missing any field information or has obvious outliers will be eliminated. In other words, if a certain data is empty or missing, the sample will be deleted to complete the cleaning of the landslide disaster event data and avoid the data used for training containing missing data.

[0074] Data standardization (or dimensionless processing): Normalize the numerical features (such as slope and rainfall) in the historical landslide disaster data set, scale the data to the range of [0,1], and improve the convergence speed of the model; One-Hot Encoding is performed on the categorical variables such as geological types in the historical landslide disaster data set to convert them into numerical forms that can be used by the model. Preferably, maximum-minimum normalization processing can be used. All feature quantity data are subjected to maximum-minimum normalization processing and converted into dimensionless values ​​to eliminate the dimensional differences of each feature quantity data to form a landslide disaster and triggering condition sample set, wherein the feature quantity data includes terrain data, geological data, meteorological data, human activity data, remote sensing data and landslide historical record data; the maximum-minimum normalization processing process is as follows: let the original value of a certain feature quantity be x, and the minimum value in the data be x min , the maximum value is x max , then the normalized calculation formula is:

[0075] Feature selection: Select key features based on correlation or expert knowledge. Specifically, use correlation analysis (such as Pearson correlation coefficient) or feature importance analysis based on tree models to screen variables significantly related to landslide occurrence and remove redundant or low-correlation features; for high-dimensional data, methods such as principal component analysis (PCA) can be used to reduce the dimensionality, retain the main information, and reduce the computational cost.

[0076] (1.3) Based on the preprocessed historical landslide disaster data set, construct data features that can express the triggering mechanism of landslide events, use the data features as model input, and use the preprocessed historical landslide disaster event data as training labels to construct an initial training data set. For example, by combining topographic, meteorological and geological data, construct new features that can efficiently express the triggering mechanism of landslides, such as the slope-rainfall combination index, NDVI (Normalized Difference Vegetation Index) change rate, etc.

[0077] (1.4) The landslide samples in the initial training data set are expanded to balance the ratio between landslide samples and non-landslide samples to obtain the target training data set.

[0078] First, use landslide historical points to mark positive samples (i.e., landslide samples), non-landslide areas to mark negative samples (i.e., non-landslide samples), and combine landslide frequency ratio analysis to balance the number of positive and negative samples to avoid class imbalance. For landslide samples, expand them. For example, perform the following operations on each landslide sample in the initial training data set:

[0079] Operation 1: Perform one or more of the following operations on the landslide samples: rotation, cropping, and scaling. In practical applications, in view of the scarcity of landslide samples, oversampling technology is used to generate new samples, and data enhancement technology (such as rotation, cropping, and scaling) is used to expand the data volume of landslide samples.

[0080] Operation 2: Determine multiple neighboring landslide samples corresponding to the landslide sample from the initial training data set, and fuse the landslide sample and each neighboring landslide sample to obtain multiple new landslide samples. Among them, the neighboring landslide sample is the landslide sample whose Euclidean distance to the landslide sample is less than the preset distance threshold. For example, in order to solve the problem of unbalanced distribution of landslide event samples and non-landslide samples, SMOTE (synthetic minority oversampling technology) is used to balance the ratio of positive and negative samples.

[0081] After determining the target training data set, the data can be divided into a training set and a test set at a ratio of 8:2 to ensure fairness in model evaluation. Specifically, labels are assigned to the sample set and the training set, the training set is assigned a Y label, and the test set is assigned an X label; in the embodiment of the present invention, the ratio of dividing the training set and the test set is 8:2, that is, if there are 10 sample sets, 8 of them are used as training sets, and 2 sample sets are used as test sets.

[0082] Furthermore, cross-validation (Cross-Validation) splits the data set for multiple training and validation, such as performing 5-fold cross-validation on the training set to evaluate the stability and generalization ability of the model performance.

[0083] (1.5) Obtain multiple pre-constructed decision trees, randomly sample the target training data set for each decision tree, obtain the training data subset corresponding to each decision tree, use the training data subset to train the corresponding decision tree, and integrate each trained decision tree to obtain a landslide identification model.

[0084] In a specific implementation, the landslide identification result output by the random forest model is a landslide occurrence probability distribution map, which can be used to identify high-risk sub-areas for landslides. Specifically, a random forest machine learning algorithm is selected, samples and features are randomly extracted, and multiple decision trees are constructed. Each tree generates a training data subset (in-bag sample) by random sampling for independent training, and the predicted average of all decision trees is calculated as the final result to obtain a random forest machine learning model, and the remaining data is used for out-of-bag evaluation. In addition, the prediction category of the high-risk sub-area for landslides can be determined by a majority voting-based method, and the prediction category includes a high-risk sub-area for landslides or a non-high-risk sub-area for landslides.

[0085] In one example, the model initialization parameter settings: the number of decision trees defaults to 100, which can be adjusted according to the amount of data; the maximum depth defaults to 10 to avoid overfitting, and the recommended value is adjusted according to the data dimension; the minimum number of split samples defaults to 5 to improve generalization ability. In one example, the model evaluation indicator selection: accuracy, precision, recall, F1 score and ROC curve are used to evaluate model performance. In one example, the optimization process: use grid search (GridSearchCV), random search or Bayesian optimization methods to adjust the key parameters of the machine learning model (such as n_estimators (the number of decision trees), max_depth (the maximum depth of the decision tree), max_features (the number of features considered at each split), etc.) to obtain the optimal configuration.

[0086] In one example, the classification threshold is set to adjust the model's sensitivity to identifying high-risk landslide sub-areas: Setting the threshold: The default classification threshold of the random forest is 0.5, that is, when the predicted probability is greater than or equal to 0.5, it is judged as a high-risk area. Threshold adjustment: Adjust the threshold according to the tolerance level of landslide risk: lowering the threshold, higher recall rate, suitable for reducing the false alarm rate; increasing the threshold, higher precision, suitable for reducing the false alarm rate. Select the optimal threshold: Find the optimal balance point by drawing the ROC curve to maximize the classification effect.

[0087] Furthermore, key features can be screened based on feature importance evaluation, random forest feature importance ranking or recursive feature elimination method to eliminate low-contribution or redundant features.

[0088] Furthermore, the prediction areas can be divided according to geological types and climatic conditions, and regional landslide identification models can be constructed to improve the regional prediction accuracy.

[0089] In an embodiment of the present invention, a random forest model is used with a landslide disaster and trigger condition sample set (i.e., slope-rainfall combination index, NDVI change rate, etc.) as input, and a high or low probability of landslide disaster loss (i.e., landslide identification result) as output, and the parameter values ​​of the number of decision trees, maximum depth, and minimum number of split samples of the random forest model are set. Since the problem involved in the present invention is a regression model, the following parameters need to be set to construct the random forest machine learning model in the embodiment of the present invention: n_estimators definition: the number of decision trees, default value: 100. bootstrap definition: whether to use sampling with replacement to construct a training subset, default value: True. max_features definition: the number of features considered each time the split is performed, default value: auto (the number of all features). max_depth definition: the maximum depth of the decision tree, default value: None (the tree will split until the leaf node). criterion definition: the indicator used to split the node, mse (mean square error, default): suitable for general regression tasks; mae (mean absolute error): more robust to outliers, but the computational cost is higher. min_samples_split definition: the minimum number of samples a node must contain before it can be further split, default value: 2. min_samples_leaf definition: the minimum number of samples in a leaf node, default value: 1. min_weight_fraction_leaf definition: the ratio of the minimum sample weight in a leaf node to the total sample weight, default value: 0. n_jobs definition: the number of threads used for parallel computing, default value: 1. Adjustment suggestion: set to -1 to use all available CPU resources to speed up the calculation. random_state definition: random seed, used for result reproduction, set to a fixed value of 42. oob_score definition: whether to use out-of-bag samples to calculate the prediction performance of the model, default value: False.

[0090] In summary, Random Forest is a machine learning method based on ensemble learning. It improves the stability and accuracy of the model while avoiding overfitting by constructing multiple decision trees and integrating their landslide identification results. Its main features include: Randomness: randomly extract samples from the original data set to construct each decision tree (Bagging); when each node splits, some features are randomly selected for splitting, so that the structure of each tree is different. Ensemble idea: For regression tasks, the average value of all decision tree predictions is used as the final landslide identification result.

[0091] (2) Establish a physical model (also known as a landslide fine analysis model) in the high-risk area for landslides for fine analysis. The landslide fine analysis model uses a physical model to conduct an in-depth analysis of the landslide mechanism:

[0092] According to the output of the landslide identification model, the areas with higher landslide probability values ​​are extracted as key analysis objects. The physical mechanism modeling is carried out using limit equilibrium analysis, rainfall infiltration simulation and finite element method for the key analysis objects, the landslide triggering conditions are analyzed, and the landslide hazard distribution of the landslide identification model is combined with the refined results of the physical model to form a multi-level prediction model.

[0093] The input of the landslide detailed analysis model includes: data preparation of geological related parameters, terrain related parameters, meteorological related parameters, surface cover related parameters, and parameters related to human activities, and the above data are processed into raster format, unified spatial resolution and coordinate system, and the data are normalized or standardized.

[0094] In one example, the stability analysis method was used as the core element and calculation method of the physical model, the finite element method was used as the numerical simulation tool, and the DEM (Digital Elevation Model) and geological profile were used to construct a three-dimensional geometric model of the landslide body. According to the results of the geological survey, each cell was assigned corresponding physical parameters (such as cohesion and friction angle), and dynamic data such as rainfall and earthquakes were input into the model to calculate the landslide risk distribution in real time to complete the model construction.

[0095] In one example, the optimization process of the model includes: obtaining a differentiated model parameter library corresponding to the target area, which includes model parameters matching different geological types and meteorological conditions, including geomechanical parameters, rainfall triggering condition parameters and groundwater dynamics parameters; based on the differentiated model parameter library, multiple landslide fine analysis models are established for different geological types and meteorological conditions in the target area. In specific implementation, a differentiated model parameter library is established for different geological types (rock slopes, soil slopes, etc.) and meteorological conditions (high rainfall areas, arid areas); based on geographical regional divisions (such as watersheds, geological units), the input parameters of the landslide fine analysis model are spatially resolved and locally adjusted.

[0096] The inherent mechanism of landslide occurrence is quantitatively analyzed through the principles of mechanics and fluid dynamics, including landslide triggering conditions, instability process, movement characteristics and potential impact range. The analysis of landslide occurrence mechanism uses mechanical principles to deeply analyze the landslide process, including landslide triggering conditions and rainfall-induced landslides. The movement characteristics of landslides include sliding surface identification, determining the location, shape and range of potential sliding surfaces, usually determined by the Mohr-Coulomb yield criterion or field observations; dynamic motion simulation, using the finite element method (FEM) or discrete element method (DEM) to simulate the sliding process of the sliding body under gravity and external triggering conditions; simulating the velocity field and displacement field of the landslide and its changes in time and space. The prediction of the impact range of the landslide, including the diffusion of the sliding body, combines the fluid mechanics model to simulate the flow range of the landslide soil in the downhill direction.

[0097] Furthermore, the landslide detailed analysis model can be optimized as follows: considering the rainfall infiltration process, optimizing the landslide triggering mechanism simulation through the nonlinear relationship between dynamic water pressure changes and slope shear strength; using real landslide event data to calibrate key parameters in the physical model (such as permeability coefficient, shear strength) to improve model fitting; comparing the predicted results with the actual distribution, and dynamically adjusting the input parameters and triggering thresholds.

[0098] For the joint model of machine learning and physical mechanism, historical landslide event data can also be used to calibrate and verify the joint model to improve the reliability and accuracy of the model results. x Input into the landslide hazard prediction model to obtain the predicted value Test Y′ ; The predicted value Test Y′ Test with real values Y The landslide hazard model can be evaluated by comparison. According to the characteristics of the joint model, the embodiment of the present invention selects the mean absolute error and the root mean square error to evaluate the model accuracy, wherein the mean absolute error is used to measure the average absolute difference between the predicted value and the true value, indicating the average deviation degree between the predicted value and the true value, and its calculation formula is as follows:

[0099]

[0100] Where N is the total number of samples, y i is the predicted value, x i is the real value;

[0101] The root mean square error is the square root of the ratio of the square of the deviation between the predicted value and the actual value to the number of observations. The calculation formula is as follows:

[0102]

[0103] The evaluation indicators NAE and RMSE of the joint model judge the quality of the model. The closer the NAE and RMSE are to 0, the higher the accuracy of the neural network model. It should be noted that when training the model, the data of the training set Y is Train y , the test set X is Train x ; When validating the model, the test set X is Test x , that is, input the basic data set of historical landslide disasters into the joint model, obtain the predicted value of the landslide disaster level corresponding to the basic data set of historical landslide disasters, and compare the predicted value of the landslide disaster level with the actual value to verify the accuracy of the joint model.

[0104] The landslide prediction model of the joint machine learning and physical mechanism proposed in the embodiment of the present invention combines the advantages of machine learning models and physical models, and demonstrates powerful capabilities in landslide disaster prediction. The machine learning model can efficiently process massive, multi-dimensional input data (such as topography, geology, meteorology, human activities, etc.), quickly identify high-risk areas for landslides through training, and has the advantages of automation and real-time performance, but the analysis of the physical mechanism and dynamic evolution process of landslides is limited. The physical model is based on the mechanical and hydrological mechanisms of landslides, and can deeply analyze the triggering conditions, damage range and dynamic evolution characteristics of landslides, and provide high-precision process simulation and result interpretation, but the computational complexity is high and the quality requirements for input parameters are high. Through the joint model, the rapid screening function of machine learning can be used to significantly narrow the analysis scope, and the physical model can be used to perform detailed analysis and dynamic simulation of high-risk areas, making up for the limitations of a single model, comprehensively improving the efficiency, accuracy and scientificity of landslide prediction, and providing reliable decision support for landslide disaster risk assessment, early warning and emergency response.

[0105] Based on the joint model constructed above, the embodiment of the present invention further explains the landslide disaster prediction method combining machine learning and physical mechanism, including:

[0106] (I) Obtain the basic data set of landslide disasters in the target area. The basic data set of landslide disasters includes topographic data, geological data, meteorological data, human activity data, and remote sensing data.

[0107] Furthermore, the terrain data can be time-space matched to obtain raster data of terrain data; the geological data can be time-space matched to obtain raster data of geological data; the meteorological data can be time-space matched to obtain raster data of meteorological data; the human activity data can be time-space matched to obtain raster data of human activity data; the remote sensing data can be time-space matched to obtain raster data of remote sensing data; the above-mentioned raster data can be integrated to obtain characteristic quantity data of terrain data, geological data, meteorological data, human activity data, and remote sensing data.

[0108] (ii) Inputting the basic data set of landslide hazards into the pre-built joint model, which includes a landslide identification model based on machine learning and a landslide detailed analysis model based on physical mechanism.

[0109] (III) Using the landslide identification model, the landslide identification results in the target area are identified based on the landslide disaster basic data set. In one embodiment, the topographic, geological, meteorological and other parameter data of the target area are input, and after pre-processing, they are input into the trained landslide identification model to predict the classification probability of the high-risk sub-area for landslides, and the high-risk areas are marked according to the set threshold, and the distribution map or raster file of the high-risk sub-area for landslides is output.

[0110] (IV) Through the landslide fine analysis model, the dynamic evolution process of landslide disasters in the target area is simulated according to the landslide disaster basic data set and the landslide identification results to obtain the landslide simulation results. In an optional implementation, based on the landslide identification results, a high-risk sub-area for landslides is extracted from the target area; with the range of the high-risk sub-area for landslides as the boundary condition, through the landslide fine analysis model that matches the geological type and meteorological conditions corresponding to the high-risk sub-area for landslides, the dynamic evolution process of landslide disasters in the high-risk sub-area for landslides is simulated according to the landslide disaster basic data set to obtain the landslide simulation results.

[0111] Specifically: First, the results of machine learning and physical models are integrated, and the joint probability distribution map of high-risk landslide sub-areas is generated using Bayesian inference or multi-model integration technology. Then, spatial correction is performed on the high-risk landslide sub-areas, and the scope of the high-risk landslide sub-areas is optimized by combining terrain analysis and domain expert knowledge. Finally, a physical model is established in the high-risk landslide sub-areas for fine analysis, and the characteristic quantity data of the target area is input into the landslide fine analysis model to obtain the landslide hazard level of the target area, so that the future landslide hazard situation can be accurately predicted.

[0112] Among them, the input data of the landslide detailed analysis model include: the spatial scope of the high-risk area (identified by the machine learning model). Topographic data: high-resolution digital elevation model (DEM), slope, slope aspect. Geological data: lithological characteristics, fault distribution, soil type and physical and mechanical parameters (such as shear strength, porosity). Hydrological data: groundwater level distribution, rainfall infiltration rate, surface runoff characteristics. Meteorological data: rainfall intensity and distribution, rainfall duration, temperature and humidity changes. Human activity data: engineering excavation, road construction, farmland irrigation and other information. Physical model construction: simulate the process of landslide occurrence, including the movement path, speed and energy of the sliding body; commonly used models: discrete element method (DEM), finite element method (FEM), coupled hydrological-mechanical model.

[0113] Input and parameterization of landslide fine analysis model: Geomechanical parameters: including density, friction angle, cohesion, pore water pressure, etc. of rock and soil. Rainfall triggering conditions: design different rainfall scenarios (such as short-term heavy rainfall, long-term slow rainfall), calculate rainfall penetration depth and saturation time. Groundwater dynamic parameters: simulate the rise of groundwater level caused by rainfall and its impact on shear stress and shear strength.

[0114] Verify the predictive capability of the landslide fine analysis model: by comparing the process of simulating historical landslide events with actual observation results, calculate error indicators (such as the error between the simulated landslide volume and the actual volume).

[0115] Dynamic simulation of high-risk landslide areas: In the identified high-risk areas, the triggering and evolution process of landslides under different rainfall scenarios is simulated based on the physical model, and the following results are output: the movement path and velocity field distribution of the sliding body; the impact range of the sliding body (landslide destruction area and accumulation area); simulate the start time of the landslide, the instability time of the sliding body and the duration of the entire landslide process, and output the time range of the landslide and the potential disaster impact.

[0116] (V) The landslide identification results and the landslide simulation results are integrated to obtain the landslide hazard prediction results corresponding to the target area. In an optional embodiment, the landslide identification results and the landslide simulation results are superimposed based on the pre-configured weight values, and the landslide hazard level corresponding to the sub-area in the target area is determined according to the superposition results. Exemplarily, the landslide hazard level is evaluated by combining the triggering conditions, damage range, dynamic evolution characteristics of the sliding body, and socioeconomic data of the affected area (such as population density, distribution of important infrastructure).

[0117] The embodiment of the present invention combines the results of machine learning and physical models to form the probability distribution of high-risk sub-areas for landslides, simulates the dynamic evolution of the physical model in the high-risk sub-areas for landslides, outputs the landslide evolution process and landslide time range, and obtains the landslide hazard level in the target area. In summary, the embodiment of the present invention combines the efficiency of machine learning with the accuracy of physical models to improve the credibility of landslide risk assessment. The application process of the joint model is summarized as follows:

[0118] Preliminary identification of high-risk areas: Identify high-risk sub-areas for landslides, use random forest or other machine learning models to predict the probability of landslides based on input parameters (topography, geology, meteorology, etc.), and generate a preliminary probability distribution map of high-risk sub-areas for landslides; the probability threshold corresponding to the high-risk sub-areas for landslides is verified by the training results. For example, areas with a probability higher than 0.7 are considered to be high-risk sub-areas for landslides.

[0119] Refined analysis and regional division: The results are gridded, and the prediction results of the high-risk sub-areas of landslides are discretized into grid cells. The value of each cell represents the probability of landslide occurrence at that location.

[0120] Dynamic simulation of landslide evolution: In the high-risk sub-areas of landslides identified by machine learning, a physical refined model is further constructed to simulate the mechanism of landslide occurrence and possible landslide characteristics (such as triggering conditions and damage range). Specifically, the physical model simulates the dynamic evolution process of landslides under different rainfall scenarios, including the start time of the sliding body, movement path, velocity field, accumulation range, etc.; combined with time series data, the time range and dynamic evolution characteristics of the landslide are output; the area range that may be affected by the landslide is predicted, and the detailed distribution of the landslide damage area and accumulation area is output.

[0121] Fusion of detailed analysis results of the physical model: The simulation results of the physical model are superimposed on the preliminary results of the high-risk area to introduce the instability analysis data of the physical model (such as safety factor, shear strength distribution, etc.); through weight distribution, the deterministic analysis results of the physical model are superimposed on the probability distribution of machine learning, and the comprehensive results generate a high-risk probability distribution map of the target area, providing basic data for subsequent dynamic evolution simulation.

[0122] The embodiment of the present invention obtains basic parameter data of landslide disasters in the target area, the basic parameter data includes terrain data, geological data, meteorological data, human activity data, remote sensing data and landslide historical record data; based on the basic parameter data, the characteristic quantity data of the target area is determined; the characteristic quantity data of the target area is input into the landslide disaster prediction model to obtain the landslide hazard level of the target area, so that the future landslide disaster situation can be accurately predicted, and universality is achieved.

[0123] Based on the above embodiments, the present invention provides a landslide disaster prediction device combining machine learning and physical mechanism, see Figure 3 The schematic diagram of the structure of a landslide hazard prediction device combining machine learning and physical mechanism is shown in FIG. The device mainly includes the following parts:

[0124] The data acquisition module 302 is used to acquire the basic data set of landslide hazards in the target area;

[0125] A data input module 304 is used to input the basic data set of landslide hazards into a pre-built joint model, wherein the joint model includes a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanisms;

[0126] The landslide identification module 306 is used to identify the landslide identification results in the target area according to the landslide disaster basic data set through the landslide identification model, and the landslide identification results are used to describe the probability of landslide disasters occurring in each sub-area in the target area;

[0127] The landslide fine analysis module 308 is used to simulate the dynamic evolution process of landslide disasters in the target area according to the landslide disaster basic data set and the landslide identification result through the landslide fine analysis model to obtain the landslide simulation result, and the landslide simulation result is used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-area in the target area;

[0128] The landslide prediction module 310 is used to fuse the landslide identification result and the landslide simulation result to obtain the landslide hazard prediction result corresponding to the target area; wherein the landslide hazard prediction result is used to describe the landslide hazard level corresponding to the sub-area within the target area.

[0129] The landslide disaster prediction method combining machine learning and physical mechanism provided in the embodiment of the present invention preliminarily identifies the probability of landslide disasters occurring in sub-areas within the target area through a landslide recognition model based on machine learning, and on this basis simulates the dynamic evolution process of landslide disasters in the target area through a landslide fine analysis model based on physical intelligence, and estimates the landslide disaster level of the target area by combining the landslide recognition result output by the landslide recognition model and the landslide simulation result output by the landslide fine analysis model. The joint model provided by the present invention combines the advantages of machine learning models and physical models, and demonstrates powerful capabilities in landslide disaster prediction. Through the joint model, the rapid screening function of machine learning can be used to significantly narrow the analysis scope, and the physical mechanism can be used to perform detailed analysis and dynamic simulation of high-risk sub-areas for landslides, thereby making up for the limitations of a single model, comprehensively improving the efficiency, accuracy and scientificity of landslide prediction, and providing reliable decision support for landslide disaster risk assessment, early warning and emergency response.

[0130] In one embodiment, a landslide identification model building module is further included, which is used to:

[0131] Obtain historical landslide disaster basic data sets and historical landslide disaster event data;

[0132] Preprocess the historical landslide disaster dataset and historical landslide disaster event data;

[0133] Based on the preprocessed historical landslide disaster data set, data features that can express the triggering mechanism of landslide events are constructed, and the data features are used as model inputs, and the preprocessed historical landslide disaster event data is used as training labels to construct the initial training data set;

[0134] The landslide samples in the initial training data set are expanded to balance the ratio between landslide samples and non-landslide samples to obtain the target training data set;

[0135] A plurality of pre-built decision trees are obtained, and a target training data set is randomly sampled for each decision tree to obtain a training data subset corresponding to each decision tree. The training data subset is used to train the corresponding decision tree, and each trained decision tree is integrated to obtain a landslide recognition model.

[0136] In one embodiment, the landslide identification model building module is specifically used to:

[0137] For each landslide sample in the initial training dataset, perform the following operations:

[0138] Performing one or more of rotation processing, cropping processing, and scaling processing on the landslide sample;

[0139] And / or, a plurality of adjacent landslide samples corresponding to the landslide sample are determined from the initial training data set, and the landslide sample and each of the adjacent landslide samples are fused respectively to obtain a plurality of new landslide samples.

[0140] In one embodiment, a landslide fine analysis model building module is also included, which is used to:

[0141] Obtain a differentiated model parameter library corresponding to the target area, which includes model parameters matching different geological types and meteorological conditions, including geomechanical parameters, rainfall triggering condition parameters, and groundwater dynamics parameters;

[0142] Based on the differentiated model parameter library, multiple landslide detailed analysis models are established according to different geological types and meteorological conditions in the target area.

[0143] In one embodiment, the landslide detailed analysis module 308 is specifically used to:

[0144] Based on the landslide identification results, high-risk sub-areas for landslides are extracted from the target area. The high-risk sub-areas for landslides are sub-areas where the probability of landslide disasters is higher than a preset probability threshold.

[0145] Taking the scope of the high-risk sub-area for landslides as the boundary condition, a fine landslide analysis model that matches the geological type and meteorological conditions corresponding to the high-risk sub-area for landslides is used to simulate the dynamic evolution process of landslide hazards in the high-risk sub-area for landslides according to the basic data set of landslide hazards, and the landslide simulation results are obtained.

[0146] In one embodiment, the landslide prediction module 310 is specifically used to:

[0147] Based on the pre-configured weight values, the landslide identification results and the landslide simulation results are superimposed, and the landslide hazard level corresponding to the sub-area within the target area is determined according to the superimposed results.

[0148] The device provided in the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0149] An embodiment of the present invention provides an electronic device. Specifically, the electronic device includes a processor and a storage device. The storage device stores a computer program, and when the computer program is executed by the processor, it executes the method described in any one of the above-mentioned implementation methods.

[0150] Figure 4 A structural diagram of an electronic device provided in an embodiment of the present invention, the electronic device 100 includes: a processor 40, a memory 41, a bus 42 and a communication interface 43, wherein the processor 40, the communication interface 43 and the memory 41 are connected via the bus 42; the processor 40 is used to execute an executable module stored in the memory 41, such as a computer program.

[0151] The memory 41 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 43 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0152] The bus 42 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0153] Among them, the memory 41 is used to store programs, and the processor 40 executes the program after receiving the execution instruction. The method executed by the device for flow process definition disclosed in any embodiment of the above-mentioned embodiment of the present invention can be applied to the processor 40 or implemented by the processor 40.

[0154] The processor 40 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 40. The above processor 40 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present invention can be directly embodied as a hardware decoding processor to execute, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 41, and the processor 40 reads the information in the memory 41 and completes the steps of the above method in combination with its hardware.

[0155] The computer program product of the readable storage medium provided in the embodiment of the present invention includes a computer-readable storage medium storing program code, and the instructions included in the program code can be used to execute the methods described in the previous method embodiments. The specific implementation can be referred to the previous method embodiments, which will not be repeated here.

[0156] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0157] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A landslide hazard prediction method combining machine learning and physical mechanism, characterized in that: include: Obtain basic data sets of landslide hazards in the target area; Inputting the basic data set of landslide hazards into a pre-built joint model, wherein the joint model includes a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanism; Using the landslide identification model, identifying a landslide identification result in the target area according to the landslide disaster basic data set, wherein the landslide identification result is used to describe the probability of a landslide disaster occurring in a sub-area within the target area; By means of the landslide fine analysis model, the dynamic evolution process of the landslide disaster in the target area is simulated according to the landslide disaster basic data set and the landslide identification result to obtain a landslide simulation result, wherein the landslide simulation result is used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-area in the target area; The landslide identification result and the landslide simulation result are integrated to obtain a landslide hazard prediction result corresponding to the target area; wherein the landslide hazard prediction result is used to describe the landslide hazard level corresponding to the sub-area within the target area.

2. The landslide disaster prediction method combining machine learning and physical mechanism according to claim 1 is characterized in that: Before identifying the landslide identification result in the target area according to the landslide hazard basic data set by using the landslide identification model, the method further includes: Obtain historical landslide disaster basic data sets and historical landslide disaster event data; Preprocessing the historical landslide disaster data set and the historical landslide disaster event data; Based on the preprocessed historical landslide disaster data set, data features that can express the triggering mechanism of landslide events are constructed, and the data features are used as model inputs, and the preprocessed historical landslide disaster event data is used as training labels to construct an initial training data set; Expanding the landslide samples in the initial training data set to balance the ratio between the landslide samples and the non-landslide samples to obtain a target training data set; A plurality of pre-constructed decision trees are obtained, and the target training data set is randomly sampled for each decision tree to obtain a training data subset corresponding to each decision tree, and the corresponding decision tree is trained using the training data subset, and each of the trained decision trees is integrated to obtain a landslide identification model.

3. The landslide disaster prediction method combining machine learning and physical mechanism according to claim 2 is characterized in that: The landslide samples in the initial training data set are expanded, including: The following operations are performed for each landslide sample in the initial training data set: Performing one or more of rotation processing, cropping processing, and scaling processing on the landslide sample; And / or, a plurality of adjacent landslide samples corresponding to the landslide sample are determined from the initial training data set, and the landslide sample and each of the adjacent landslide samples are respectively fused to obtain a plurality of new landslide samples.

4. The landslide disaster prediction method combining machine learning and physical mechanism according to claim 1 is characterized in that: Before simulating the dynamic evolution process of landslide disasters in the target area by using the landslide fine analysis model according to the landslide disaster basic data set and the landslide identification result to obtain the landslide simulation result, the method further includes: Acquire a differentiated model parameter library corresponding to the target area, wherein the differentiated model parameter library includes model parameters matching different geological types and meteorological conditions, and the model parameters include geomechanical parameters, rainfall triggering condition parameters, and groundwater dynamics parameters; Based on the differentiated model parameter library, a plurality of landslide detailed analysis models are established for the different geological types and meteorological conditions in the target area.

5. The landslide disaster prediction method combining machine learning and physical mechanism according to claim 4 is characterized in that: The landslide fine analysis model is used to simulate the dynamic evolution process of the landslide disaster in the target area according to the landslide disaster basic data set and the landslide identification result to obtain the landslide simulation result, including: Based on the landslide identification result, extracting a high-risk sub-region for landslide from the target region, wherein the high-risk sub-region for landslide is a sub-region where the probability of landslide disaster is higher than a preset probability threshold; Taking the range of the high-risk sub-area for landslides as the boundary condition, the landslide fine analysis model that matches the geological type and meteorological conditions corresponding to the high-risk sub-area for landslides is used to simulate the dynamic evolution process of landslide disasters in the high-risk sub-area for landslides according to the basic data set for landslide disasters, so as to obtain the landslide simulation result.

6. The landslide disaster prediction method combining machine learning and physical mechanism according to claim 5 is characterized in that: The landslide identification result and the landslide simulation result are integrated to obtain the landslide hazard prediction result corresponding to the target area, including: Based on a pre-configured weight value, the landslide identification result and the landslide simulation result are superimposed, and the landslide hazard level corresponding to the sub-area in the target area is determined according to the superimposed result.

7. A landslide disaster prediction device combining machine learning and physical mechanism, characterized in that: include: Data acquisition module, used to obtain the basic data set of landslide hazards in the target area; A data input module, used for inputting the basic data set of landslide disasters into a pre-built joint model, wherein the joint model includes a landslide identification model based on machine learning and a landslide fine analysis model based on physical mechanism; A landslide identification module, used to identify a landslide identification result in the target area according to the landslide hazard basic data set through the landslide identification model, wherein the landslide identification result is used to describe the probability of a landslide hazard occurring in each sub-area in the target area; A landslide fine analysis module, used to simulate the dynamic evolution process of landslide disasters in the target area according to the landslide disaster basic data set and the landslide identification result through the landslide fine analysis model, and obtain a landslide simulation result, wherein the landslide simulation result is used to describe one or more of the sliding body movement path, sliding body velocity field distribution, and sliding body influence range corresponding to the sub-area in the target area; The landslide prediction module is used to fuse the landslide identification result and the landslide simulation result to obtain the landslide hazard prediction result corresponding to the target area; wherein the landslide hazard prediction result is used to describe the landslide hazard level corresponding to the sub-area in the target area.

8. The landslide disaster prediction device combining machine learning and physical mechanism according to claim 7 is characterized in that: Also included are landslide identification model building modules for: Obtain historical landslide disaster basic data sets and historical landslide disaster event data; Preprocessing the historical landslide disaster data set and the historical landslide disaster event data; Based on the preprocessed historical landslide disaster data set, data features that can express the triggering mechanism of landslide events are constructed, and the data features are used as model inputs, and the preprocessed historical landslide disaster event data is used as training labels to construct an initial training data set; Expanding the landslide samples in the initial training data set to balance the ratio between the landslide samples and the non-landslide samples to obtain a target training data set; A plurality of pre-constructed decision trees are obtained, and the target training data set is randomly sampled for each decision tree to obtain a training data subset corresponding to each decision tree, and the corresponding decision tree is trained using the training data subset, and each of the trained decision trees is integrated to obtain a landslide identification model.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 6.

Images