Dynamic assessment method and system of landslide disaster risk based on the progress of remediation projects

Abstract

The present invention discloses a method and system for dynamic assessment of landslide disaster risk based on the progress of remediation projects, which relates to the field of natural disaster management technology. By integrating multi-source heterogeneous data, real-time dynamic assessment, and machine learning models, the present invention achieves accurate and dynamic assessment of landslide disaster risk, improves assessment accuracy and stability, and enhances the timeliness and pertinence of decision-making. A feature perturbation stability mechanism is introduced into the SPY algorithm to construct a more robust and discriminative pseudo-negative sample screening strategy. Information entropy reduction is introduced to improve the acquisition function, increase the efficiency of hyperparameter optimization, and effectively train the stacked generalization integration model. Furthermore, a susceptibility correction coefficient is set to consider the impact of the remediation progress in the landslide disaster area, and the landslide susceptibility assessment results are corrected to ensure that the landslide susceptibility assessment results can promptly reflect the actual situation, providing strong technical support for the prevention, control, and management of landslide disasters.

Description

Technical Field

[0001] The present invention relates to the technical field of natural disaster treatment, and in particular to a method and system for dynamically assessing landslide disaster risks based on the progress of a remediation project. Background Art

[0002] Landslides are a common and highly destructive geological disaster, posing a serious threat to the safety of life and property, infrastructure stability, and regional sustainable development. Existing landslide risk assessment techniques primarily rely on natural environmental factors such as topography, geology, and meteorology, but generally overlook the critical role of village resilience in landslide disasters. This limits the applicability of assessment results in actual disaster response and can lead to inappropriate budget planning for prevention and control. There is an urgent need to develop a dynamic landslide risk assessment system that integrates natural environmental factors with social resilience factors to improve the applicability of landslide risk assessment results. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for dynamic assessment of landslide disaster risk based on the progress of remediation projects, so as to improve the above technical problems.

[0004] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0005] A dynamic landslide disaster risk assessment method based on the progress of remediation projects includes:

[0006] Construct a slope unit set and a landslide hazard susceptibility assessment dataset for the test area, and divide the positive sample set of landslide hazard risk points;

[0007] Use the SPY-FPS pseudo-negative sample screening method to generate a pseudo-negative sample training set for identification, and train the classifier to obtain a pseudo-negative sample generator for identification;

[0008] Input the global landslide hazard susceptibility assessment dataset into the identification pseudo negative sample generator to obtain the identification pseudo negative sample set;

[0009] The pseudo-negative sample set and the positive sample set of landslide hazard points are combined and input into the stacked generalization integration model to obtain the landslide susceptibility assessment result.

[0010] Obtain landslide disaster remediation progress data in real time, construct susceptibility correction coefficients and correct landslide disaster susceptibility assessment results;

[0011] Based on the revised landslide susceptibility assessment results, the risk of the tested area is divided.

[0012] Furthermore, a slope unit set of the tested area and a landslide susceptibility assessment dataset were constructed, and a positive sample set of landslide hazard points was divided, including:

[0013] The slope unit extraction method based on forward and reverse hydrological division and terrain curvature constraint is used to process the test area to obtain a slope unit set.

[0014] Collecting initial susceptibility assessment data corresponding to the slope unit set, and connecting the initial susceptibility assessment data and the slope unit set using a spatial join tool to obtain an initial dataset for global landslide susceptibility assessment; the initial susceptibility assessment data includes topographic data, geological data, meteorological data, and human activity data for each slope unit;

[0015] The initial data set of landslide hazard global susceptibility assessment was processed using the Spearman rank correlation coefficient to obtain the positive sample set of landslide hazard potential points and the data set of landslide hazard global susceptibility assessment.

[0016] Furthermore, the process of obtaining the slope unit set is as follows:

[0017] Collect high-precision DEM data of the area to be measured and input it into the geographic information system for depression filling processing to build a high-precision digital elevation model;

[0018] Calculate the flow direction and flow accumulation of each grid in the high-precision digital elevation model;

[0019] Filter each grid based on flow accumulation to obtain a filtered grid;

[0020] According to the flow direction of each filtered grid, the watershed tool is used to divide the watershed area and generate the positive division result;

[0021] Perform inversion processing on high-precision DEM data to obtain high-precision inverted DEM data; process high-precision inverted DEM data to generate reverse division results;

[0022] Use the Union tool to merge the boundaries and spatially unify the forward and reverse division results to generate a slope boundary framework.

[0023] Calculate the terrain profile curvature and plane curvature of the slope boundary frame and perform region elimination to obtain the eliminated slope boundary frame;

[0024] The area in the slope boundary frame after elimination is divided using the surface cutting tool to form a slope unit set.

[0025] Furthermore, the training process of the classifier includes:

[0026] Obtain a training dataset for global landslide susceptibility assessment, i.e., obtain a positive sample training set and an unlabeled sample training set;

[0027] Randomly extract several positive samples from the positive sample training set and use them as the embedded reference positive sample training set; use the unextracted positive samples in the positive sample training set as the new positive sample training set;

[0028] Merge the embedded reference positive sample training set and the unlabeled sample training set to obtain the recognition pseudo negative sample training set;

[0029] The pseudo-negative sample training set and the new positive sample training set are input into the classifier for training respectively to obtain the initially trained classifier;

[0030] Obtain the classification probability of the embedded reference positive sample training set through the initially trained classifier and set the confidence threshold;

[0031] Based on the confidence threshold and the unlabeled sample training set, the feature perturbation stability mechanism is used to generate a reliable negative sample training set, and the set is input into the initially trained classifier for secondary training to obtain a trained classifier, namely the identification pseudo negative sample generator.

[0032] Furthermore, the stacked generalized integrated model includes a base model and a meta-model connected in series; the base model includes a decision tree, a random forest, a support vector machine, and a gradient boosting tree processed in parallel; and the meta-model is a meta-learner;

[0033] The stacked generalization ensemble model is trained using a Bayesian hyperparameter optimization method guided by information entropy.

[0034] Furthermore, the Bayesian hyperparameter optimization method guided by information entropy includes:

[0035] Obtain a training set of pseudo-negative samples for identification and a training set of positive samples of landslide hazard potential points and input them into a stacked generalization ensemble model;

[0036] Define the hyperparameter combination of the stacked generalization ensemble model and construct the corresponding hyperparameter space; set the objective function;

[0037] Build a learnable agent model and set up an information entropy collection function; determine the optimal hyperparameter combination through the learnable agent model and the information entropy collection function;

[0038] Apply the optimal hyperparameter combination to the stacked generalized ensemble model to complete the training of the stacked generalized ensemble model.

[0039] Furthermore, the process of determining the optimal hyperparameter combination is:

[0040] Randomly select at least one hyperparameter combination and calculate the corresponding objective function value;

[0041] Based on the selected hyperparameter combination and objective function value, a learnable agent model is constructed using Gaussian process regression;

[0042] Set the information entropy acquisition function through the learnable agent model; calculate the information entropy acquisition function value of all hyperparameter combinations; select the hyperparameter combination whose information entropy acquisition function value meets the screening conditions as the evaluation point for the next iteration;

[0043] Apply the evaluation points to the stacked generalized ensemble model, calculate the corresponding objective function value, and update the learnable proxy model;

[0044] Repeatedly select hyperparameter combinations and calculate the corresponding objective function values, and update the learnable agent model until the current number of iterations reaches the maximum number of iterations or the objective function value is less than the target threshold, and the optimal hyperparameter combination is obtained.

[0045] Furthermore, the formula corresponding to the information entropy acquisition function is:

[0046] ;

[0047] ;

[0048] in, 、 denote the second original acquisition function and the first original acquisition function respectively, represents a hyperparameter combination, represents the reduction of information entropy, represents the optimal loss value, 、 represent the variance function and mean function respectively, represents the intermediate parameters, 、 They represent the standard normal distribution cumulative distribution function and the standard normal distribution probability density function, respectively. represents a logarithmic function with a constant as the base, 、 represent the variance in the previous iteration and the variance in the current iteration respectively.

[0049] Furthermore, the real-time acquisition of landslide disaster remediation progress data, the construction of a susceptibility correction coefficient, and the correction of the landslide disaster susceptibility assessment results include:

[0050] Obtain the locations and progress of landslide remediation in the villages in the area to be tested;

[0051] The landslide remediation points in the villages were input into the geographic information system for attribute connection to construct a landslide disaster remediation dataset;

[0052] The landslide hazard remediation dataset and the landslide hazard susceptibility assessment results were spatially connected to obtain the landslide-prone area remediation dataset.

[0053] According to the progress of the corresponding remediation of each slope remediation unit in the landslide prone area remediation concentration, the percentage of the project remediation progress is set as the corresponding susceptibility correction coefficient;

[0054] Based on the susceptibility correction coefficient, the landslide susceptibility assessment results corresponding to each slope remediation unit in the landslide-prone area remediation set are updated.

[0055] A dynamic landslide disaster risk assessment system based on the progress of remediation projects includes:

[0056] The slope unit construction module is used to construct the slope unit set of the test area and the landslide susceptibility assessment dataset for the entire area, and to divide the positive sample set of landslide hazard points;

[0057] The identification pseudo-negative sample generation training module is used to generate the identification pseudo-negative sample training set using the SPY-FPS pseudo-negative sample screening method, and train the classifier to obtain the identification pseudo-negative sample generator;

[0058] The identification pseudo-negative sample acquisition module is used to input the landslide hazard global susceptibility assessment dataset into the identification pseudo-negative sample generator to obtain the identification pseudo-negative sample set;

[0059] Landslide hazard susceptibility assessment module, including a stacked generalized integrated model training module and a landslide hazard susceptibility assessment and prediction module;

[0060] The stacked generalized ensemble model training module is used to input the landslide hazard susceptibility assessment training dataset into the stacked generalized ensemble model and train it using the Bayesian hyperparameter optimization method guided by information entropy;

[0061] The landslide susceptibility assessment and prediction module is used to combine the identified pseudo-negative sample set and the positive sample set of landslide hazard points and input them into the stacked generalization integration model to obtain the landslide susceptibility assessment results;

[0062] Landslide hazard susceptibility assessment result correction module, which is used to obtain landslide hazard remediation progress data in real time, construct susceptibility correction coefficients and correct landslide hazard susceptibility assessment results;

[0063] The landslide risk classification module is used to classify the risk of the tested area based on the revised landslide susceptibility assessment results.

[0064] The beneficial effects of the present invention are:

[0065] The present invention integrates multi-source heterogeneous data, real-time dynamic evaluation and machine learning models to achieve accurate and dynamic assessment of landslide disaster risks, improve assessment accuracy and stability, and enhance the timeliness and pertinence of decision-making; introduces the "feature perturbation stability" mechanism into the SPY (Spy Sample-Based Semi-Supervised Learning Algorithm) method to construct a more robust and discriminative pseudo-negative sample screening strategy, which can effectively identify reliable negative samples in data scenarios with extremely unbalanced positive and negative samples; introduces information entropy reduction, improves the acquisition function, enhances the efficiency of hyperparameter optimization, and effectively trains the stacked generalization integration model; considers the impact of the progress of remediation in the landslide disaster area, sets a susceptibility correction coefficient, and corrects the landslide disaster susceptibility evaluation results to ensure that the landslide disaster susceptibility evaluation results can promptly reflect the actual situation, providing strong technical support for the prevention, control and management of landslide disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without making any creative efforts.

[0067] Figure 1 A flow chart of a method in an embodiment of the present invention;

[0068] Figure 2 1 is a system flow chart in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0070] See also Figure 1 This embodiment provides a method for dynamically assessing landslide disaster risk based on the progress of a remediation project, which includes:

[0071] S1. Construct a slope unit set and a landslide hazard susceptibility assessment dataset for the test area, and divide the positive sample set of landslide hazard potential points;

[0072] Said S1 comprises:

[0073] S1-1. The slope unit extraction method based on forward and reverse hydrological division and terrain curvature constraint is used to process the test area to obtain a slope unit set.

[0074] The slope unit extraction method with forward and reverse hydrological division and terrain curvature constraint is implemented in a geographic information system. Thus, the S1-1 includes:

[0075] S1-1-1. Collect high-precision DEM data of the area to be measured and input it into the geographic information system for depression filling processing to construct a high-precision digital elevation model;

[0076] S1-1-2. Calculate the flow direction and flow accumulation for each grid in the high-precision digital elevation model.

[0077] S1-1-3, determine whether the accumulated flow of each grid exceeds the flow threshold; if so, the grid is regarded as the main water system and retained; otherwise, the grid is discarded; and the filtered grid is obtained;

[0078] S1-1-4. Based on the main water systems and the flow direction of each filtered grid, use the Watershed tool to divide the watershed area and generate the positive preliminary slope boundary, i.e. the positive division result;

[0079] S1-1-5. Invert the high-precision DEM data to obtain high-precision inverted DEM data. Use the same method as S1-1-2 to S1-1-4 to process the high-precision inverted DEM data to generate an inverted preliminary slope boundary, i.e., an inverted partitioning result, which can enhance the accuracy of landform structure recognition.

[0080] S1-1-6. Use the Union tool to merge the boundaries and spatially unify the forward and reverse division results to generate a slope boundary framework.

[0081] S1-1-7. Calculate the terrain profile curvature and plane curvature of the slope boundary frame; eliminate the area boundaries where the terrain profile curvature and plane curvature do not exceed the curvature threshold condition to obtain the eliminated slope boundary frame; the curvature threshold condition includes the terrain profile curvature threshold and the plane curvature threshold.

[0082] S1-1-8. Use the Split tool to divide the area in the slope boundary frame after elimination to obtain the division results; use each division result as a slope unit to form a slope unit set. The expression of the slope unit set is .in, 、 、 Represent the first slope unit, the second slope unit and the last slope unit respectively. represents the slope unit set, Indicates the number of slope units.

[0083] S1-2. Collect the initial susceptibility assessment data corresponding to the slope unit set, and connect the initial susceptibility assessment data and the corresponding slope units through the spatial join tool (SpatialJoin) to obtain the initial dataset for the global landslide susceptibility assessment.

[0084] The initial susceptibility assessment data includes topographic data, geological data, meteorological data and human activity data of each slope unit; topographic data includes elevation, slope, aspect, curvature and river channel; geological data includes soil type, rock structure, fault distribution and seismic parameters; meteorological data includes previous accumulated precipitation, precipitation intensity, daily temperature difference, and groundwater level; human activity data includes land use type, road network, and underground engineering area.

[0085] Said S1-2 includes:

[0086] S1-2-1. Format the initial susceptibility assessment data, interpolate missing values, process outliers, and encode data to obtain susceptibility assessment data;

[0087] S1-2-2. Use the Spatial Join tool to dynamically connect the slope units in the slope unit set with the susceptibility assessment data to obtain the initial dataset for the global landslide susceptibility assessment. .in, represents the initial dataset for global landslide susceptibility assessment, 、 、 They represent the first initial data of global landslide susceptibility assessment, the second initial data of global landslide susceptibility assessment and the last initial data of global landslide susceptibility assessment, respectively. It represents the initial data volume for the global landslide susceptibility assessment.

[0088] Initial data for global landslide susceptibility assessment Includes the corresponding slope cell susceptibility assessment data and its pixel raster.

[0089] S1-3. Use the Spearman rank correlation coefficient to process the initial data set of the global landslide hazard susceptibility assessment to obtain the positive sample set of landslide hazard potential points and the global landslide hazard susceptibility assessment data set.

[0090] Said S1-3 includes:

[0091] S1-3-1. Take the slope unit containing the hidden danger point as the positive sample of the landslide hidden danger point, and obtain the positive sample set of the landslide hidden danger point.

[0092] The Spatial Join tool is called to connect the slope unit set and the initial data of global landslide susceptibility assessment to obtain the connected initial data of global landslide susceptibility assessment; the attribute field "Landslide (judgment character)" corresponding to each slope unit in the connected initial data of global landslide susceptibility assessment is created and the judgment attribute information is assigned, that is, the slope unit containing the hidden danger point is assigned a value of 1 (representing the "positive sample set"), and the unit without the hidden danger point is assigned a value of 0 (representing the "unlabeled sample set").

[0093] S1-3-2, according to the formula:

[0094] ;

[0095] Calculate the judgment characters of the initial data of the landslide hazard global susceptibility assessment after connection . Represents the initial data of landslide hazard susceptibility assessment in the whole area after connection The data. Represents the positive sample set of landslide hazard points. Indicates the Positive samples of landslide hazard risk points.

[0096] S1-3-3. Use the Spearman rank correlation coefficient to optimize the initial data set for the global landslide susceptibility evaluation and obtain the global landslide susceptibility evaluation data set.

[0097] The initial data of the global susceptibility assessment of landslide disasters were used as susceptibility assessment factors, and the correlation coefficients between the susceptibility assessment factors were calculated using the Spearman rank correlation coefficient analysis method. The susceptibility assessment factors with collinearity (VIF (correlation coefficient) ≥ 10) in the initial data set of the global susceptibility assessment of landslide disasters were eliminated to obtain the global susceptibility assessment data set of landslide disasters. . 、 、 They represent the first landslide hazard global susceptibility evaluation data, the second landslide hazard global susceptibility evaluation data, and the last landslide hazard global susceptibility evaluation data, respectively. Represents the global landslide hazard susceptibility assessment dataset.

[0098] The landslide hazard global susceptibility assessment dataset includes the positive sample set and unlabeled sample sets .

[0099] S2. Use the SPY-FPS pseudo-negative sample screening method to generate a pseudo-negative sample training set for identification, and train the classifier to obtain a pseudo-negative sample generator for identification;

[0100] The training process of the classifier includes:

[0101] S2-1. Obtaining positive sample training set (labeled positive) and unlabeled sample training set (no label);

[0102] Obtain a positive sample training set of landslide hazard potential points; define the slope units that intersect with the positive sample training set of landslide hazard potential points as positive samples (labeled as positive), and define the slope units that do not contain the positive sample training set of landslide hazard potential points as unlabeled sample sets (no label);

[0103] S2-2. Randomly extract positive sample training set Several positive samples in , and used as the embedded reference positive sample training set ; The positive sample training set The unextracted positive samples in are used as new positive sample training sets . 20% of the positive sample training set can be randomly selected ,Right now .in, Indicates absolute value.

[0104] S2-3. Merge the embedded reference positive sample training set and unlabeled sample training set , get the identification pseudo negative sample training set ;

[0105] S2-4, respectively identify the pseudo negative sample training set and new positive sample training set Input the data into the classifier for training to obtain an initially trained classifier. The classifier can be a CNN classifier or a naive Bayes classifier, and its output is the classification probability. The classification probability includes the probability of being classified as a positive sample or a negative sample.

[0106] S2-5. Obtain the embedded reference positive sample training set through the initially trained classifier The classification probability of the embedded reference positive sample prediction probability set is obtained Based on the probability distribution characteristics of the embedded reference positive sample prediction probability set, combined with its skewness and quantile characteristics, the confidence threshold is set based on the analysis results. For example, choose The lower 5% quantile of is used as the confidence threshold.

[0107] S2-6, based on confidence threshold , unlabeled sample training set , a reliable negative sample training set is generated by using the feature perturbation stability mechanism, and is input into the initially trained classifier for secondary training to obtain a trained classifier, namely, the identification pseudo negative sample generator.

[0108] The S2-6 includes:

[0109] S2-6-1. Prediction probability set and confidence threshold based on unlabeled samples , for the unlabeled sample training set Screening to obtain the preliminary pseudo negative sample training set ; Keep the prediction probability of unlabeled samples not less than the confidence threshold Unlabeled samples with a value less than the confidence threshold are removed Unlabeled samples are obtained to obtain the initial pseudo negative sample training set .

[0110] S2-6-2, introduce the feature perturbation stability mechanism to the initial pseudo-negative sample training set Each sample in is subjected to feature perturbation to generate R perturbation samples, namely , construct a more robust and discriminative pseudo-negative sample screening strategy, which can effectively identify reliable negative samples in data scenarios with extremely unbalanced positive and negative samples. Represents the initial pseudo negative sample training set No. samples, Indicates the perturbation samples, represents the perturbation function (characteristic perturbation stability mechanism).

[0111] S2-6-3. Input each perturbation sample into the initial classifier respectively, count the number of times the perturbation sample is judged as a negative label, and calculate the corresponding robustness score; the initial classifier can use a naive Bayes classifier.

[0112] S2-6-4. Arrange the perturbation samples in descending order according to the robustness score, select the perturbation samples corresponding to the first K robustness scores, and obtain the final pseudo-negative sample training set. .

[0113] S2-6-5, the final pseudo negative sample training set and positive sample training set Merge to obtain a reliable negative sample training set;

[0114] S2-6-6. Input the reliable negative sample training set into the initially trained classifier for secondary training to obtain a trained classifier, i.e., an identification pseudo negative sample generator.

[0115] S3. Input the global landslide susceptibility assessment dataset into the identification pseudo negative sample generator to obtain the identification pseudo negative sample set.

[0116] The landslide hazard global susceptibility assessment dataset is input into the pseudo negative sample generator to obtain the corresponding classification probability; the probability of the classification is greater than the confidence threshold. The global landslide susceptibility evaluation data is used to obtain the identification pseudo negative sample set.

[0117] S4. The identified pseudo-negative sample set and the positive sample set of landslide hazard potential points are merged and input into the stacked generalization integration model to obtain the landslide susceptibility evaluation result.

[0118] This embodiment uses a stacking generalization integration model (Stacking Integration Model) to evaluate susceptibility; the stacking generalization integration model includes a base model and a meta-model connected in series; the base model includes a decision tree, a random forest, a support vector machine, and a gradient boosting tree processed in parallel; the meta-model is a meta-learner.

[0119] The stacked generalization ensemble model is trained using a Bayesian hyperparameter optimization method guided by information entropy. Thus, the Bayesian hyperparameter optimization method guided by information entropy includes:

[0120] S4-1. Obtain a training set of pseudo-negative samples and a training set of positive samples of landslide hazard points and input them into a stacked generalization ensemble model;

[0121] S4-2. Define the hyperparameter combination of the stacked generalization ensemble model , construct the corresponding hyperparameter space ; Set the objective function, the initial value of the current number of iterations, and the maximum number of iterations; the formula corresponding to the objective function is:

[0122] ;

[0123] in, represents the objective function, Represents the AUC indicator of the stacked generalization ensemble model. Objective function The purpose is to maximize the AUC index during the iteration process.

[0124] Hyperparameter space Including several sets of hyperparameter combinations , each set of hyperparameter combinations This includes all adjustable parameters of the base model and meta-model. The initial value of the current iteration number is 0.

[0125] S4-3. Construct a learnable agent model and set an information entropy acquisition function; determine the optimal hyperparameter combination through the learnable agent model and the information entropy acquisition function;

[0126] The S4-3 includes:

[0127] S4-3-1. Randomly select at least one hyperparameter combination and calculate the corresponding objective function value;

[0128] S4-3-2. Based on the selected hyperparameter combination and objective function value, a learnable agent model is constructed using Gaussian process regression. ;in, represents the kernel function, represents the selected hyperparameter combination, represents the mean function, represents the Gaussian function;

[0129] S4-3-3. Set the information entropy acquisition function through the learnable agent model; calculate the information entropy acquisition function values for all hyperparameter combinations; select the hyperparameter combination whose information entropy acquisition function value meets the screening conditions as the evaluation point for the next iteration;

[0130] The information entropy acquisition function is an expected improved acquisition function based on information gain.

[0131] The S4-3-3 includes:

[0132] A1. Use the selected hyperparameter combination and objective function value as historical data ;in, Indicates the number of hyperparameter combinations selected, 、 、 Respectively represent the first hyperparameter combination, the second hyperparameter combination, and the last hyperparameter combination in the selected hyperparameter combination, 、 、 Respectively 、 、 The corresponding objective function value.

[0133] A2. Calculate the similarity between pairs of historical data, the backward variance of the current iteration, and the backward mean through a learnable proxy model.

[0134] The kernel function in the learnable agent model is used to calculate the similarity between each two historical data, that is, ; 、 Respectively represent Historical data, Historical data.

[0135] Based on each similarity, the variance and mean of the current iteration are calculated through a learnable proxy model.

[0136] A3. Calculate the optimal loss value (optimal target loss value) for the current iteration and construct an information entropy acquisition function based on the variance and mean.

[0137] Current iteration The optimal loss value The corresponding formula is:

[0138] ;

[0139] First original acquisition function The corresponding formula is:

[0140] ;

[0141] in, Indicates the selected hyperparameter combination corresponding to the optimal loss value of the current iteration, represents the minimum function, 、 represent the variance function and mean function respectively, represents the intermediate parameters, 、 They represent the standard normal distribution cumulative distribution function and the standard normal distribution probability density function respectively.

[0142] In order to improve the efficiency of hyperparameter optimization, the information entropy reduction is introduced The formula corresponding to the reduction in information entropy is:

[0143] ;

[0144] in, 、 They represent the uncertainty of the hyperparameter space in the previous iteration and the uncertainty of the hyperparameter space in the current iteration respectively.

[0145] 、 The posterior entropy of Gaussian process regression is used to calculate, thus, The corresponding formula is:

[0146] ;

[0147] ;

[0148] in, represents a natural constant, represents a logarithmic function with a constant as the base, Indicates uncertainty, 、 represent the variance in the previous iteration and the variance in the current iteration respectively.

[0149] Thus, the information entropy acquisition function The corresponding formula is:

[0150] ;

[0151] ;

[0152] in, represents the second original acquisition function.

[0153] A4. Calculate the information entropy acquisition function values corresponding to all hyperparameter combinations in the hyperparameter space;

[0154] A5. Select the hyperparameter combination whose information entropy acquisition function value meets the screening conditions as the evaluation point for the next iteration and merge it into the historical data to complete the update of the historical data. The screening conditions are:

[0155] ;

[0156] in, Represents the maximum function.

[0157] S4-3-4, apply the evaluation point to the stacked generalization ensemble model, calculate the corresponding objective function value, and update the learnable proxy model; the current number of iterations is increased by 1, that is ;

[0158] S4-3-5. Repeat S4-3-3 through S4-3-4 until the current iteration count reaches the maximum or the objective function value falls below the target threshold. The optimal hyperparameter combination is obtained and applied to the stacked generalized ensemble model, completing training of the stacked generalized ensemble model. The target threshold is 0.95. When the objective function value exceeds 0.95, the stacked generalized ensemble model is packaged and used to process the slope susceptibility assessment dataset.

[0159] S4-4. Apply the optimal hyperparameter combination to the stacked generalization ensemble model to complete the training of the stacked generalization ensemble model.

[0160] S5. Obtain landslide disaster remediation progress data in real time, construct a susceptibility correction coefficient and correct the landslide disaster susceptibility assessment results.

[0161] The S5 includes:

[0162] S5-1. Obtain the location and progress of landslide remediation in the villages in the area to be measured, i.e. the longitude and latitude of the landslide remediation points. and the percentage of progress of the project . Indicates longitude, Indicates latitude.

[0163] S5-2. Input the village landslide remediation points into the geographic information system for attribute connection and construct a landslide disaster remediation dataset . 、 Respectively represent The longitude and latitude of the landslide remediation points in each village, Indicates the The progress of landslide remediation at each village, Indicates the total number of landslide remediation points in the village.

[0164] S5-3, using the same method as S1-3 to analyze the landslide disaster remediation dataset and landslide susceptibility assessment results Perform spatial connection to obtain the landslide prone area remediation cluster where engineering remediation has been carried out. . 、 、 They represent the landslide susceptibility assessment results corresponding to the first slope unit, the second slope unit, and the last slope unit, respectively. 、 、 They respectively represent the first slope remediation unit, the second slope remediation unit and the last slope remediation unit in the landslide-prone area remediation concentration where engineering remediation has been carried out.

[0165] S5-4. According to the progress of the corresponding remediation of each slope remediation unit in the landslide-prone area remediation center, set the percentage of the engineering remediation progress as the corresponding susceptibility correction coefficient; where the value range of the susceptibility correction coefficient is [0,1].

[0166] S5-5. Based on the susceptibility correction coefficient, update the landslide susceptibility assessment results corresponding to each slope remediation unit in the landslide-prone area remediation set to obtain the revised landslide susceptibility assessment results. The corresponding formula is:

[0167] ;

[0168] in, Indicates that landslide-prone areas are under concentrated control The susceptibility correction coefficient corresponding to each slope regulation unit is: 、 They represent the concentrated control of landslide-prone areas The landslide hazard susceptibility assessment results corresponding to the slope remediation units and the revised landslide hazard susceptibility assessment results are Represents a constant.

[0169] S6. Based on the revised landslide susceptibility assessment results, the tested area is divided into risk categories. Based on the revised landslide susceptibility assessment results, the tested area is divided using the natural breakpoint method, resulting in corresponding risk classification results, as shown in Table 1. The risk classification results include high risk, relatively high risk, medium risk, relatively low risk, and low risk.

[0170] Table 1

[0171]

[0172] like Figure 2 As shown in FIG, a dynamic landslide disaster risk assessment system based on the progress of the remediation project includes:

[0173] The slope unit construction module is used to construct the slope unit set of the test area and the landslide susceptibility assessment dataset for the entire area, and to divide the positive sample set of landslide hazard points;

[0174] The identification pseudo-negative sample generation training module is used to generate the identification pseudo-negative sample training set using the SPY-FPS pseudo-negative sample screening method, and train the classifier to obtain the identification pseudo-negative sample generator;

[0175] The identification pseudo-negative sample acquisition module is used to input the landslide hazard global susceptibility assessment dataset into the identification pseudo-negative sample generator to obtain the identification pseudo-negative sample set;

[0176] Landslide hazard susceptibility assessment module, including a stacked generalized integrated model training module and a landslide hazard susceptibility assessment and prediction module;

[0177] The stacked generalized ensemble model training module is used to input the landslide hazard susceptibility assessment training dataset into the stacked generalized ensemble model and train it using the Bayesian hyperparameter optimization method guided by information entropy;

[0178] The landslide susceptibility assessment and prediction module is used to combine the identified pseudo-negative sample set and the positive sample set of landslide hazard points and input them into the stacked generalization integration model to obtain the landslide susceptibility assessment results;

[0179] Landslide hazard susceptibility assessment result correction module, which is used to obtain landslide hazard remediation progress data in real time, construct susceptibility correction coefficients and correct landslide hazard susceptibility assessment results;

[0180] The landslide risk classification module is used to classify the risk of the tested area based on the revised landslide susceptibility assessment results.

[0181] In summary, the present invention realizes accurate and dynamic assessment of landslide disaster risks by integrating multi-source heterogeneous data, real-time dynamic assessment and machine learning models, improves assessment accuracy and stability, and enhances the timeliness and pertinence of decision-making; introduces the "feature perturbation stability" mechanism into the SPY (Spy Sample-Based Semi-Supervised Learning Algorithm) method to construct a more robust and discriminative pseudo-negative sample screening strategy, so that reliable negative samples can be effectively identified in data scenarios with extremely unbalanced positive and negative samples; introduces information entropy reduction, improves the acquisition function, improves the efficiency of hyperparameter optimization, and effectively trains the stacked generalization integration model; considers the influence of the progress of remediation in the landslide disaster area, sets a susceptibility correction coefficient, and corrects the landslide disaster susceptibility evaluation results to ensure that the landslide disaster susceptibility evaluation results can reflect the actual situation in a timely manner, providing strong technical support for the prevention, control and management of landslide disasters.

[0182] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A dynamic landslide disaster risk assessment method based on the progress of remediation projects, characterized by: include: Construct a slope unit set and a landslide hazard susceptibility assessment dataset for the test area, and divide the positive sample set of landslide hazard risk points; Use the SPY-FPS pseudo-negative sample screening method to generate a pseudo-negative sample training set for identification, and train the classifier to obtain a pseudo-negative sample generator for identification; Input the global landslide hazard susceptibility assessment dataset into the identification pseudo negative sample generator to obtain the identification pseudo negative sample set; The pseudo-negative sample set and the positive sample set of landslide hazard points are combined and input into the stacked generalization integration model to obtain the landslide susceptibility assessment result. Obtain landslide disaster remediation progress data in real time, construct susceptibility correction coefficients and correct landslide disaster susceptibility assessment results; Based on the revised landslide susceptibility assessment results, the risk of the tested area is divided; The training process of the classifier includes: Obtain a training dataset for global landslide susceptibility assessment, i.e., obtain a positive sample training set and an unlabeled sample training set; Randomly extract several positive samples from the positive sample training set and use them as the embedded reference positive sample training set; use the unextracted positive samples in the positive sample training set as the new positive sample training set; Merge the embedded reference positive sample training set and the unlabeled sample training set to obtain the recognition pseudo negative sample training set; The pseudo-negative sample training set and the new positive sample training set are input into the classifier for training respectively to obtain the initially trained classifier; Obtain the classification probability of the embedded reference positive sample training set through the initially trained classifier and set the confidence threshold; Based on the confidence threshold and the unlabeled sample training set, the feature perturbation stability mechanism is used to generate a reliable negative sample training set, which is then input into the initially trained classifier for secondary training to obtain the trained classifier, i.e., the identification pseudo-negative sample generator; The stacked generalized integrated model includes a base model and a meta-model connected in series; the base model includes a decision tree, a random forest, a support vector machine, and a gradient boosting tree processed in parallel; the meta-model is a meta-learner; and the stacked generalized integrated model is trained using a Bayesian hyperparameter optimization method guided by information entropy; Bayesian hyperparameter optimization methods based on information entropy guidance include: Obtain a training set of pseudo-negative samples for identification and a training set of positive samples of landslide hazard potential points and input them into a stacked generalization ensemble model; Define the hyperparameter combination of the stacked generalization ensemble model and construct the corresponding hyperparameter space; set the objective function; Build a learnable agent model and set up an information entropy collection function; determine the optimal hyperparameter combination through the learnable agent model and the information entropy collection function; Apply the optimal hyperparameter combination to the stacked generalized ensemble model to complete the training of the stacked generalized ensemble model.

2. A method for dynamic landslide disaster risk assessment based on the progress of a remediation project according to claim 1, characterized in that: Construct a slope unit set and a landslide susceptibility assessment dataset for the entire region to be tested, and divide the positive sample set of landslide hazard points into: The slope unit extraction method based on forward and reverse hydrological division and terrain curvature constraint is used to process the test area to obtain a slope unit set. Collecting initial susceptibility assessment data corresponding to the slope unit set, and connecting the initial susceptibility assessment data and the slope unit set using a spatial join tool to obtain an initial dataset for global landslide susceptibility assessment; the initial susceptibility assessment data includes topographic data, geological data, meteorological data, and human activity data for each slope unit; The Spearman rank correlation coefficient was used to process the initial data set of landslide hazard global susceptibility assessment to obtain the positive sample set of landslide hazard potential points and the data set of landslide hazard global susceptibility assessment.

3. A method for dynamic landslide disaster risk assessment based on the progress of a remediation project according to claim 2, characterized in that: The process of obtaining the slope unit set is as follows: Collect high-precision DEM data of the area to be measured and input it into the geographic information system for depression filling processing to build a high-precision digital elevation model; Calculate the flow direction and flow accumulation of each grid in the high-precision digital elevation model; Filter each grid based on flow accumulation to obtain a filtered grid; According to the flow direction of each filtered grid, the watershed tool is used to divide the watershed area and generate the forward division result; Perform inversion processing on high-precision DEM data to obtain high-precision inverted DEM data; process high-precision inverted DEM data to generate reverse division results; Use the union tool to merge the boundaries and spatially unify the forward and reverse partitioning results to generate the slope boundary framework; Calculate the terrain profile curvature and plane curvature of the slope boundary frame and perform region elimination to obtain the eliminated slope boundary frame; The area in the slope boundary frame after elimination is divided using the surface cutting tool to form a slope unit set.

4. The method for dynamic landslide disaster risk assessment based on the progress of a remediation project according to claim 1, characterized in that: The process of determining the optimal hyperparameter combination is: Randomly select at least one hyperparameter combination and calculate the corresponding objective function value; Based on the selected hyperparameter combination and objective function value, a learnable agent model is constructed using Gaussian process regression; Set the information entropy acquisition function through the learnable agent model; calculate the information entropy acquisition function value for all hyperparameter combinations; Select the hyperparameter combination whose information entropy acquisition function value meets the screening conditions as the evaluation point for the next iteration; Apply the evaluation points to the stacked generalized ensemble model, calculate the corresponding objective function value, and update the learnable proxy model; Repeatedly select hyperparameter combinations and calculate the corresponding objective function values, and update the learnable agent model until the current number of iterations reaches the maximum number of iterations or the objective function value is less than the target threshold, and the optimal hyperparameter combination is obtained.

5. A method for dynamic landslide disaster risk assessment based on the progress of a remediation project according to claim 4, characterized in that: The formula corresponding to the information entropy acquisition function is: ； ； in, 、 denote the second original acquisition function and the first original acquisition function respectively, represents a hyperparameter combination, represents the reduction of information entropy, represents the optimal loss value, 、 represent the variance function and mean function respectively, represents the intermediate parameters, 、 They represent the standard normal distribution cumulative distribution function and the standard normal distribution probability density function, respectively. represents a logarithmic function with a constant as the base, 、 They represent the variance in the previous iteration and the variance in the current iteration respectively.

6. The method for dynamic landslide disaster risk assessment based on the progress of a remediation project according to claim 1, characterized in that: The real-time acquisition of landslide disaster remediation progress data, the construction of a susceptibility correction coefficient, and the correction of landslide disaster susceptibility assessment results include: Obtain the locations and progress of landslide remediation in the villages in the area to be tested; The landslide remediation points in the villages were input into the geographic information system for attribute connection to construct a landslide disaster remediation dataset; The landslide hazard remediation dataset and the landslide hazard susceptibility assessment results were spatially connected to obtain the landslide-prone area remediation dataset. According to the progress of the corresponding remediation of each slope remediation unit in the landslide prone area remediation concentration, the percentage of the project remediation progress is set as the corresponding susceptibility correction coefficient; Based on the susceptibility correction coefficient, the landslide susceptibility assessment results corresponding to each slope remediation unit in the landslide-prone area remediation set are updated.

7. A landslide hazard risk dynamic assessment system based on the progress of a remediation project, used to implement the landslide hazard risk dynamic assessment method based on the progress of a remediation project according to any one of claims 1 to 6, characterized in that: include: The slope unit construction module is used to construct the slope unit set of the test area and the landslide susceptibility assessment dataset for the entire area, and to divide the positive sample set of landslide hazard points; The identification pseudo-negative sample generation training module is used to generate the identification pseudo-negative sample training set using the SPY-FPS pseudo-negative sample screening method, and train the classifier to obtain the identification pseudo-negative sample generator; The identification pseudo-negative sample acquisition module is used to input the landslide hazard global susceptibility assessment dataset into the identification pseudo-negative sample generator to obtain the identification pseudo-negative sample set; Landslide hazard susceptibility assessment module, including a stacked generalized integrated model training module and a landslide hazard susceptibility assessment and prediction module; The stacked generalized ensemble model training module is used to input the landslide hazard susceptibility assessment training dataset into the stacked generalized ensemble model and train it using the Bayesian hyperparameter optimization method guided by information entropy; The landslide susceptibility assessment and prediction module is used to combine the identified pseudo-negative sample set and the positive sample set of landslide hazard points and input them into the stacked generalization integration model to obtain the landslide susceptibility assessment results; Landslide hazard susceptibility assessment result correction module, which is used to obtain landslide hazard remediation progress data in real time, construct susceptibility correction coefficients and correct landslide hazard susceptibility assessment results; The landslide risk classification module is used to classify the risk of the tested area based on the revised landslide susceptibility assessment results.

Images