A Coupling Evaluation Method for Landslide Disaster Susceptibility and Risk

Abstract

In an embodiment of the present invention, a coupling evaluation method for landslide disaster susceptibility and risk is provided, belonging to the technical field of data processing, and specifically including: extracting landslide disaster factor characteristics from multi-source data of the target area and performing preprocessing; dividing the grid feature data into intervals and calculating the landslide event frequency of each interval; calculating the local Moran index of the cumulative rainfall within a preset time interval according to the rainfall grid data in the grid feature data to identify abnormal areas in the rainfall spatial distribution; performing a susceptibility-risk coupling evaluation based on Bayesian optimization according to the landslide event frequency and rainfall anomaly factors; based on the optimal weights of disaster-causing characteristics, inputting the real-time static disaster-causing characteristic interval index of each grid in the target area to obtain a susceptibility evaluation result map, and inputting the real-time feature vector of each grid into the risk evaluation model to obtain a risk evaluation result map. Through the solution of the present invention, the evaluation accuracy and adaptability are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of data processing, and in particular, to a coupling evaluation method for landslide disaster susceptibility and risk. Background Art

[0002] The disadvantages of current landslide risk assessment methods are as follows: (1) In the susceptibility assessment, "whether a landslide occurs" is used as a label. However, the occurrence of a landslide is closely related to external factors such as rainfall, resulting in the inability of this label to accurately represent the inherent susceptibility of the region, thus affecting the final risk assessment. (2) Areas where no landslide has occurred are usually regarded as negative samples. However, there may be potential high-risk areas where landslides have not been triggered yet, and the random sampling of negative samples may lead to unstable modeling results, affecting the accuracy of risk assessment.

[0003] It can be seen that there is an urgent need for a coupling evaluation method for landslide disaster susceptibility and risk with higher assessment accuracy and adaptability. Summary of the Invention

[0004] In view of this, the embodiments of the present invention provide a coupling evaluation method for landslide disaster susceptibility and risk, which at least partially solves the problems of poor assessment accuracy and adaptability in the prior art.

[0005] The embodiments of the present invention provide a coupling evaluation method for landslide disaster susceptibility and risk, including:

[0006] Step 1, extracting landslide disaster factor features from multi-source data of the target area and performing preprocessing to obtain raster feature data;

[0007] Step 2, dividing the raster feature data into intervals and calculating the landslide event frequency of each interval;

[0008] Step 3, calculating the local Moran index of the cumulative rainfall within a preset time interval based on the rainfall raster data in the raster feature data to identify abnormal areas in the rainfall spatial distribution, and obtaining a rainfall anomaly factor;

[0009] Step 4, performing a susceptibility-risk coupling evaluation based on Bayesian optimization according to the landslide event frequency and the rainfall anomaly factor to obtain the optimal weights of disaster-forming characteristics and a risk assessment model;

[0010] Step 5, based on the optimal weights of disaster-forming characteristics, inputting the real-time static disaster-forming characteristic interval index of each raster in the target area to obtain a susceptibility evaluation result map, and inputting the real-time feature vector of each raster into the risk assessment model to obtain a risk evaluation result map.

[0011] According to a specific implementation manner of the embodiments of the present invention, the specific content of the Step 1 includes:

[0012] Step 1.1: Divide the multi-source data of the target area into static factor features and dynamic factor features according to the data feature update period. Among them, the static factor features include geological structure elements, topographic and geomorphic features, deformation features, and human activity factors, and the dynamic factor feature is rainfall raster data.

[0013] Step 1.2: Unify the spatial reference and raster resolution of the static factor features and dynamic factor features to obtain raster feature data.

[0014] According to a specific implementation manner of an embodiment of the present invention, the specific steps of step 2 include:

[0015] Step 2.1: Perform intervalization processing on each feature in the static factor features according to the value type of each feature to obtain an interval set corresponding to each feature.

[0016] Step 2.2: Calculate the landslide event frequency of each interval in the interval set corresponding to each feature according to a preset formula, where the preset formula is

[0017] ;

[0018] Wherein, is the total number of pixels in the target area where the value of feature f is located in interval k, represents the counted number of landslide-occurring pixels, is the total number of positive samples, is the total number of pixels in the target area.

[0019] According to a specific implementation manner of an embodiment of the present invention, the specific steps of step 3 include:

[0020] Step 3.1: Calculate the local Moran's index corresponding to the rainfall amount of each pixel within a preset time interval according to the rainfall raster data.

[0021] Step 3.2: Perform statistical verification on the local Moran's index of all pixels according to a preset significance threshold to obtain a significance score.

[0022] Step 3.3: Identify rainfall anomaly pixels according to the significance score and calculate the corresponding rainfall anomaly factor

[0023] ;

[0024] Wherein, represents the local Moran's index of pixel of, represents the significance score of pixel of, represents the significance threshold, represents the high-high aggregation area of the pixel and its neighborhood, Represents the high-low aggregation area of pixels and the neighborhood.

[0025] According to a specific implementation manner of an embodiment of the present invention, step 4 specifically includes:

[0026] Step 4.1, based on the landslide event frequency, combine the frequencies of different types of features according to the set initial weight vector to obtain their corresponding weight coefficients, and synthesize the scores of each feature interval and the weight coefficients of the corresponding features on each pixel to obtain a susceptibility factor characterizing the vulnerability of the geological environment;

[0027] Step 4.2, randomly generate the date of each pixel, then form a negative sample set with the pixels that meet the preset conditions, and splice and combine it with the positive samples composed of all landslide events to form a risk model evaluation sample, where the preset condition is that the rainfall anomaly factor corresponding to the pixel is 0 and the susceptibility factor is less than the susceptibility threshold;

[0028] Step 4.3, construct a risk feature set according to the dynamic factor feature, rainfall anomaly factor and susceptibility factor;

[0029] Step 4.4, index and obtain the corresponding feature values from the risk feature set according to the coordinates and dates of each sample data in the risk model evaluation sample, and divide them into a training set and a test set according to a preset ratio;

[0030] Step 4.5, use the training set to train a random forest model, and calculate the accuracy index of the trained model on the test set;

[0031] Step 4.6, construct an objective function according to the accuracy rate and F1 score in the accuracy index, take minimizing the objective function value as the optimization goal, and iteratively solve the objective function through Bayesian optimization to obtain the optimal weight of the disaster-forming characteristics and the optimal susceptibility threshold;

[0032] Step 4.7, use the optimal susceptibility threshold as the new susceptibility threshold and execute steps 4.2 to 4.5 again to obtain a risk evaluation model.

[0033] According to a specific implementation manner of an embodiment of the present invention, the expression of the objective function is

[0034] ;

[0035] Wherein, represents the accuracy rate, represents the weight coefficient, represents the susceptibility threshold, represents the F1 score.

[0036] The coupling evaluation scheme for landslide disaster susceptibility and hazard in the embodiments of the present invention includes: Step 1, extracting the characteristics of landslide disaster factors from multi-source data of the target area and performing preprocessing to obtain raster feature data; Step 2, dividing the raster feature data into intervals and calculating the landslide event frequency of each interval; Step 3, calculating the local Moran index of the cumulative rainfall within a preset time interval based on the rainfall raster data in the raster feature data to identify abnormal areas in the rainfall spatial distribution, and obtaining rainfall anomaly factors; Step 4, performing susceptibility-hazard coupling evaluation based on Bayesian optimization according to the landslide event frequency and rainfall anomaly factors to obtain the optimal weights of disaster-forming characteristics and a hazard evaluation model; Step 5, based on the optimal weights of disaster-forming characteristics, inputting the real-time static disaster-forming characteristic interval index of each raster in the target area to obtain a susceptibility evaluation result map, and inputting the real-time feature vector of each raster into the hazard evaluation model to obtain a hazard evaluation result map.

[0037] The beneficial effects of the embodiments of the present invention are as follows: Through the solution of the present invention, firstly, based on historical landslide point samples, by statistically calculating the landslide event frequency in each characteristic interval and weighted overlay, and summarizing the weighted results of all spatial units to construct a preliminary susceptibility evaluation factor; secondly, combining susceptibility with dynamic disaster-inducing factors to establish a hazard assessment model; finally, through Bayesian optimization, adaptively adjusting the weights of susceptibility factors and negative sample constraint conditions to obtain the susceptibility-hazard evaluation result with the optimal accuracy in the test set. This solution can not only improve the accuracy and stability of landslide disaster assessment in terms of method, but also provide a more targeted scientific basis for geological disaster risk control and emergency decision-making. Brief Description of the Drawings

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0039] Figure 1 It is a schematic flowchart of a method for coupling evaluation of landslide disaster susceptibility and hazard provided by an embodiment of the present invention;

[0040] Figure 2 It is a schematic diagram of the result of abnormal detection of historical seven-day rainfall characteristics provided by an embodiment of the present invention;

[0041] Figure 3 It is a coupling evaluation framework diagram of susceptibility-hazard provided by an embodiment of the present invention;

[0042] Figure 4 It is a schematic diagram of the susceptibility evaluation result provided by an embodiment of the present invention;

[0043] Figure 5 A schematic diagram of the risk assessment result provided by the embodiment of the present invention. Detailed implementation manners

[0044] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0046] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present invention, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device and / or practice this method can be implemented using other structures and / or functions in addition to one or more of the aspects described herein.

[0047] It should also be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention schematically. The drawings only show the components related to the present invention, rather than being drawn according to the number, shape and size of the components in actual implementation. The type, quantity and proportion of each component in its actual implementation can be an arbitrary change, and the component layout type may also be more complex.

[0048] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0049] In the field of landslide disaster risk assessment, the evaluation of susceptibility and hazard is usually a relatively independent process. Generally speaking, landslide susceptibility mainly measures the potential for landslides inherent in the geological environment itself, focusing on the impact of internal geological environment factors such as geological structure, topography, and rock and soil structure on the gestation of landslides. It describes the vulnerability or potential instability of a region to landslide disasters under relatively static geological conditions. In other words, assuming the same external inducing factors, regions with higher susceptibility are more likely to experience landslides. The hazard assessment, on the other hand, is based on susceptibility and further incorporates or considers dynamic disaster-inducing factors, such as rainfall and ground surface deformation, to more comprehensively characterize the actual possibility of landslide disasters occurring. These dynamic inducing factors can significantly increase the likelihood of landslides during a certain period or under specific conditions, thus enabling a more realistic landslide hazard assessment. Therefore, the hazard assessment not only includes the inherent geological environmental vulnerability, i.e., susceptibility, but also incorporates the spatio-temporal distribution and intensity of external triggering factors, ultimately representing the probability or likelihood of landslide disasters occurring in the current or future period.

[0050] Existing technical solutions for landslide hazard assessment mostly tend to follow the following pattern: First, susceptibility assessment is carried out: By combining static disaster-forming factors such as topography and lithology, through model learning and evaluation, the susceptibility results are obtained, which represent the vulnerability of the geological environment. The susceptibility assessment is independent of the dynamic inducing process of landslide events and focuses on the stability of internal geological conditions. After obtaining the susceptibility results, a comprehensive hazard assessment is then conducted: By integrating the susceptibility assessment results with dynamic disaster-inducing factors, and using the susceptibility results as factors in the hazard assessment, the final assessment of landslide hazard is formed. This process comprehensively considers the vulnerability of the geological environment and the influence of external inducing factors, making the hazard assessment more comprehensive and accurate. However, in practical applications, there are still two prominent problems. First, there is uncertainty in the sampling of non-landslide samples. When using areas where "landslides have not occurred" as negative samples, it is often impossible to determine whether these areas are truly in a low-risk state. Some plots with high risk but not yet triggered disasters are also regarded as negative samples, resulting in deviations in the susceptibility and hazard assessment results. Second is the lack of true values in the susceptibility assessment. Usually, the vulnerability of the geological environment is calibrated by "whether a landslide actually occurs", but the occurrence of landslides is also affected by comprehensive dynamic disaster-inducing factors such as external rainfall, and does not necessarily truly reflect the potential instability of the region under static conditions. In other words, the "true values" of susceptibility samples become blurred due to external inducing factors, making it difficult to accurately distinguish the vulnerability of the geological environment itself from the triggering effects of external disaster-inducing factors. If the above problems are not solved, it will significantly affect the accuracy and applicability of landslide hazard assessment.

[0051] In summary, the current landslide hazard assessment methods lack an overall solution that can simultaneously address the lack of true values in susceptibility and the uncertainty in the sampling of non-landslide samples.

[0052] An embodiment of the present invention provides a coupling evaluation method for landslide disaster susceptibility and hazard, which can be applied to the landslide disaster assessment process in the geological disaster prevention and control scenario.

[0053] See Figure 1 , which is a schematic flowchart of a coupling evaluation method for landslide disaster susceptibility and hazard provided by an embodiment of the present invention. As Figure 1 shown, the method mainly includes the following steps:

[0054] Step 1, extract the landslide disaster factor characteristics from the multi-source data of the target area and perform preprocessing to obtain raster feature data;

[0055] Specifically, when implemented, the process of extracting the landslide disaster factor characteristics from the multi-source data of the target area and performing preprocessing to obtain raster feature data is as follows:

[0056] 1.1 Landslide disaster factor feature extraction and processing

[0057] For the coupling evaluation of landslide disaster susceptibility and hazard, the present invention preprocesses and extracts features from the multi-source data in the study area. According to the update period of the data features, the features are roughly divided into two categories: static features and dynamic features; among them, the static features are mainly used to evaluate the susceptibility of landslide disasters, while the dynamic features and a few static features, such as the annual average deformation rate, are mainly used for hazard evaluation.

[0058] (1) Static factor features

[0059] Denote the static feature set as , these features remain relatively stable on a relatively long time scale and can be used to measure the vulnerability of the regional geological environment. In the present invention, the static features mainly include: ① Geological structure elements: distance from the fault, lithology type, and soil type, which are formed into data layers that can be superimposed under a unified coordinate system through vectorization or rasterization; ② Topographic and geomorphic features: including elevation, slope, aspect, planar curvature, and profile curvature, which are extracted based on the digital elevation model (DEM); ③ Deformation features: Based on the InSAR deformation monitoring of historical years, the annual average deformation amount in the region can be calculated. In the present invention, the deformation features are regarded as hazard characteristic factors. Although they have a certain time significance, they can also be regarded as relatively static features in the hazard evaluation; ④ Other factors such as human activities: According to data such as road vector data, land use type, and optical remote sensing images in the study area, feature factors such as road network density, land use type, and NDVI are extracted.

[0060] (2) Dynamic factor features

[0061] The dynamic factor features of landslide disasters refer to the external disaster-inducing factors that change significantly with time t, denoted as , which have a more direct short-term impact on the occurrence of landslide disasters. The dynamic factors in the present invention mainly refer to rainfall-related factors, including the daily cumulative rainfall, the cumulative rainfall in the past seven days, and the maximum rainfall intensity on the current day, which are obtained by interpolating the observed data products of meteorological stations.

[0062] 1.2 Unifying Spatial Reference and Resolution

[0063] Since geological, remote sensing, meteorological, and deformation data often originate from different coordinate systems and resolutions, they need to be processed to ensure the consistency of subsequent analysis and the accuracy of comparison. First is the unification of spatial reference: project all vector and raster data into the same coordinate system EPSG:3857. At the same time, convert the coordinate values recorded in the landslide catalog events into XY in the corresponding coordinate system. Second is the unification of raster resolution: for data with inconsistent resolutions, select a unified grid size , downsample high-resolution data and upsample low-resolution data by interpolation to make them match the target resolution. Finally is the processing of invalid and outlier values: for missing data areas, interpolation or moderate filling can be used; for extreme outliers, elimination or clipping processing is performed.

[0064] Step 2: Divide the raster feature data into intervals and calculate the landslide event frequency of each interval;

[0065] Specifically, when implemented, the process of calculating the landslide event frequency within the disaster-forming feature interval is as follows:

[0066] 2.1 Division of Landslide Disaster-Forming Factor Feature Intervals

[0067] For the raster feature data obtained in the study area, interval processing is performed according to their value types. For continuous features such as slope and elevation, the natural breakpoint method is used for grading. This method obtains relatively reasonable segmentation points by minimizing the within-group variance and maximizing the between-group variance to form several intervals . For discrete features such as land use type and soil type, their respective possible values are regarded as discrete levels and no further continuous division is required. After processing, each feature f obtains a set of non-overlapping intervals in the study area, , where is the number of intervals of feature f.

[0068] 2.2 Calculation of Landslide Event Frequency within Feature Intervals

[0069] After completing the interval division, to measure the susceptibility of landslide events in each interval, the present invention introduces the "interval landslide event frequency" as a quantitative index, denoted as , and the definition is shown in formula (1):

[0070] ;

[0071] Among them, represents the number of landslide-occurring pixels counted within the interval ; is the total number of pixels where the value of feature f in the entire region is within interval k; is the total number of positive samples; is the total number of pixels in the entire region. The larger is, it indicates that within the interval

[0072] Step 3: Calculate the local Moran's I of the cumulative rainfall within the preset time interval based on the rainfall raster data in the raster feature data to identify the abnormal regions in the rainfall spatial distribution, and obtain the rainfall anomaly factor;

[0073] In specific implementation, after extracting the basic rainfall raster data in Step 1, the present invention identifies the abnormal regions in the rainfall spatial distribution by calculating the local Moran's I of the seven-day cumulative rainfall, and determines the degree of abnormality through statistical tests. The main process is as follows:

[0074] 3.1 Calculation of local Moran's I

[0075] The local Moran's I is used to measure the attribute correlation between unit i and its spatial neighborhood units, and is often used in clustering and outlier analysis. Let represent the historical seven-day rainfall of the i-th spatial unit during the target period, be the overall mean, be the element of the adjacency weight matrix, which is used to reflect the adjacency relationship between spatial units i and j, then the local Moran's I is defined as shown in formula (2):

[0076] ;

[0077] Among them, is the normalization term of the global variance, and its calculation method is shown in formula (3), where n represents the total number of spatial units.

[0078] ;

[0079] When , it indicates that and the surrounding are both positive or negative, representing high-high aggregation (HH) or low-low aggregation (LL), reflecting that pixel i and its neighborhood are similar in numerical distribution and are both relatively high or low; when When it is [condition], it indicates that pixel i has the opposite characteristics to its neighborhood, such as high-low aggregation (HL) or low-high aggregation (LH), suggesting potential outliers or mutations. The larger the value, the more significant the degree that i and its neighborhood are both high values or both low values in the rainfall distribution. The more negative the value, the greater the contrast with the neighborhood.

[0080] 3.2 Statistical significance test

[0081] After calculating , it is necessary to conduct a statistical test on the calculated results. For each cell i, its corresponding Z-score can be estimated to express the significance level, and the calculation method is shown in formula (4):

[0082] ;

[0083] Where and are the expectation and variance of the local Moran's I index under the assumption of random distribution respectively. When exceeds a pre-set critical value , it can be determined that there is significant spatial aggregation or outliers statistically.

[0084] 3.3 Generation of rainfall anomaly feature raster

[0085] This invention focuses on the anomaly of rainfall in space. In particular, areas with high-value aggregation and high-low aggregation areas that form a strong contrast with surrounding cells are more likely to represent concentrated or extreme rainfall anomalies. Therefore, this invention takes the HH and HL types, and cells as rainfall anomaly cells, and the remaining types or insignificant cells are regarded as non-rainfall anomaly cells.

[0086] For each pixel i that has been determined to be a rainfall anomaly, its Z-score can be directly used as a measure of the anomaly degree. The larger the value of the anomaly degree, the higher the anomaly degree of the rainfall characteristic value of this cell. The calculation method of its corresponding rainfall anomaly index is shown in formula (5).

[0087] ;

[0088] Step 4: Conduct a coupling evaluation of susceptibility and hazard based on Bayesian optimization according to the landslide event frequency and rainfall anomaly factor to obtain the optimal weights of disaster-forming characteristics and the hazard evaluation model;

[0089] In specific implementation, the present invention has completed the characteristic interval division and the calculation of the interval landslide event probability in step 2. In order to further improve the prediction accuracy of the hazard model, this step uses Bayesian optimization to link and adjust the susceptibility weight and the constraint strategy of negative samples, and finally obtains the optimal parameters through iteration, thereby obtaining the optimal susceptibility-hazard coupling evaluation result.

[0090] 4.1 Generation of susceptibility factors

[0091] After calculating the landslide event frequency corresponding to the disaster-pregnant characteristic value interval in step 2.2, according to the set initial weight vector to combine the frequencies of different features together, where F represents the total number of static disaster-prone features, Represents the weight coefficient of feature f in the overall susceptibility calculation. On the other hand, the scores of each feature interval and the corresponding feature weights are combined. , we can get the vulnerability index that represents the vulnerability of the geological environment , as shown in formula (6):

[0092] ;

[0093] in, represents the weight of feature f, Indicate the pixel The partition index of feature f, that is, which value interval or category it belongs to; is the probability of landslide event in this interval.

[0094] 4.2 Hazard Assessment Model Training

[0095] (1) Dangerous Negative Sample Constrained Sampling

[0096] In the hazard assessment model, negative samples usually need to be selected from units where landslides have never occurred. If the negative samples are not selected properly, they may include areas with high landslide risks but have not yet triggered disasters, thereby interfering with model learning. To this end, the present invention constrains the sampling of dangerous negative samples by using rainfall anomaly characteristic factors and susceptibility factors, and only includes the pixel in the optional negative sample set when the following conditions are met: ① It is outside the rainfall anomaly range, that is, the rainfall characteristic anomaly calculated according to step 3 is 0; ② The susceptibility index is low, that is, the internal geological environment is relatively stable, and the susceptibility index is less than the set threshold. .

[0097] Before determining the spatial location of the negative sample points, first randomly generate the date , and then screen according to the above conditions to obtain the negative sample candidate set that meets the conditions, as shown in formula (7).

[0098] ;

[0099] Among them represents the set of negative sample candidates that meet the conditions randomly selected on the date , is the rainfall anomaly degree, is the susceptibility index. Finally, the required number of negative samples is randomly selected from the candidate set, denoted as , which is spliced with the positive sample points to obtain the set of dangerous sample points. Then, according to the set number of negative sample samplings, the corresponding number of negative samples is randomly sampled from the candidate set. Finally, it is spliced with the positive samples to obtain the landslide dangerous sample point set.

[0100] (2) Obtaining the characteristic values of dangerous samples

[0101] According to the dynamic features extracted in step 1, the rainfall anomaly factor obtained in step 3, and the susceptibility results obtained in step 4.1, a set of dangerous characteristics is constructed, which mainly includes static susceptibility factors and deformation factors, and dynamic rainfall factors and anomaly factors, as shown in Equation (8):

[0102] ;

[0103] Among them, is the susceptibility factor; is the deformation factor; is the cumulative rainfall of the day; is the cumulative rainfall of the past seven days; is the maximum rainfall intensity of the day; is the rainfall anomaly factor.

[0104] According to the coordinates and the date in the sample point set, the corresponding characteristic values are indexed and obtained from the characteristic raster file. This process can be expressed by Equation (9):

[0105] ;

[0106] (3) Training of the dangerousness evaluation model and calculation of accuracy

[0107] After constructing the dangerous samples, they are divided according to a fixed ratio of the training set to the test set. Assume the sample point set is , among which is the sample label, landslide is the positive sample, and non-landslide is the negative sample. The training set and the test set are randomly divided, with the training set accounting for 70% and the test set accounting for 30%. The division process can be expressed by Equations (10 - 11):

[0108] ;

[0109] The random forest model is selected for training. After the training is completed, the accuracy metrics of the model on the test set are calculated, including accuracy, precision, recall, and F1-score, as shown in formulas (12-15):

[0110] ;

[0111] where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives.

[0112] 4.3 Bayesian Optimization for Solving Optimal Parameters

[0113] (1) Optimization Objective and Hyperparameter Definition

[0114] In the present invention, the parameters adjusted by the Bayesian optimization method mainly include two parts: the static feature weight parameter in the susceptibility evaluation and the susceptibility constraint threshold in the sampling of dangerous negative samples . The optimization objective function is the sum of the accuracy and F1-score of the danger evaluation model on the test set, as shown in formula (16). The present invention aims to minimize .

[0115] ;

[0116] (2) Bayesian Optimization Solving Process

[0117] Initial Sampling: In the search space, several groups of initial candidate points are selected, and all operations in 4.1~4.2 are performed on each group, including generating susceptibility factors, constraining negative samples, training the danger model, calculating ACC and F1, to obtain the corresponding ; Constructing or Updating the Surrogate Model: Using the Gaussian process as the surrogate model, fitting the current sample relationship to obtain the posterior distribution; Selecting New Candidate Points: Determining the next batch of parameter combinations that are most likely to improve according to the acquisition function, and evaluating them again; Iterating until Convergence: When the surrogate model converges or reaches the maximum number of iterations, the optimal solution is output, that is, the parameter configuration that can maximize the comprehensive index of ACC+F1 for danger evaluation. After the Bayesian optimization is completed, the following can be obtained: Optimal Susceptibility Factor: The

[0118] generated by weighting the landslide occurrence frequencies in each feature interval by ; Optimal Danger Model: Combined with the threshold ; After performing negative sample constrained sampling, a hazard assessment model is trained with a random forest and achieves the highest accuracy on the test set.

[0119] Step 5: Based on the optimal weights of disaster-forming characteristics, input the real-time static disaster-forming characteristic interval index of each grid in the target area to obtain the susceptibility assessment result map, and input the real-time feature vector of each grid into the hazard assessment model to obtain the hazard assessment result map.

[0120] In specific implementation, based on the optimal weights of disaster-forming characteristics, input the real-time static disaster-forming characteristic interval index of each grid in the target area to obtain the susceptibility assessment result map, and by inputting the feature vectors of all spatial units in the area into the trained hazard assessment model, the hazard assessment results of all spatial units in the area can be obtained.

[0121] The coupling assessment method for landslide disaster susceptibility and hazard provided in this embodiment avoids the deviation of directly using the areas where landslides have not occurred as negative samples in traditional methods by combining susceptibility and dynamic disaster-inducing factors to perform constrained sampling on hazard negative samples, improves the accuracy of negative sample selection, effectively prevents the omission of potential high-risk areas, and thus significantly improves the stability and accuracy of the hazard assessment model; the susceptibility-hazard coupling learning method based on Bayesian optimization can adaptively adjust the weights of susceptibility factors and negative sample constraint conditions, optimizes the landslide hazard assessment process, maximizes the accuracy and F1 score, and overcomes the problem of insufficient integration of susceptibility factors and dynamic disaster-inducing factors in traditional landslide hazard assessment methods; using the local Moran index to detect spatial anomalies in rainfall characteristics can effectively identify and quantify rainfall anomaly areas, improves the contribution of rainfall inducement to the hazard assessment model, and makes the hazard assessment results closer to the actual probability of disaster occurrence.

[0122] The method of the present invention will be further described below with a specific embodiment. The specific implementation process of the present invention will be described using the landslide event catalog of Province A from 2022 to 2024, together with the disaster factor characteristic data and multi-period rainfall data:

[0123] (1) Data preprocessing

[0124] A total of 456 landslide events that occurred between 2019 and 2024 were collected. First, the longitude and latitude coordinates recorded in the catalog were converted to coordinates in the EPSG:3857 coordinate system. For the existing data, through data conversion, all features were processed into raster file format, and then the spatial reference of all raster data was unified to the EPSG:3857 coordinate system, and the grid resolution was uniformly set to 300m×300m. The static feature set includes NDVI, elevation, slope, aspect, soil type, land use type, distance to water system, distance to road, distance to fault, human density, and annual average deformation rate. After extraction, 11 TIF raster files were obtained; after calculation, the dynamic feature set mainly includes the daily rainfall, the cumulative rainfall in the past seven days, and the maximum rainfall intensity on the current day. An independent raster layer is generated for each time period, and the update period of the dynamic features is one day.

[0125] (2)Calculation of the frequency of landslide events within the disaster-forming feature interval

[0126] For continuous features such as slope and curvature, each feature is divided into 5 value intervals using the natural breakpoint method; for discrete features such as soil type and land use type, their original categories are retained. The number of landslide events in different feature intervals within five years is counted, and the occurrence frequency of landslide events in different feature intervals is statistically calculated according to formula (1).

[0127] (3)Spatial anomaly detection of dynamic rainfall features

[0128] In the present invention, anomaly detection is performed on the feature of the cumulative rainfall in the past seven days. The raster of the cumulative rainfall in the past seven days of a certain day is input, and for each pixel in the region, the local Moran index is calculated according to formula (2), where the adjacent weight takes values in the 8-neighborhood mode, that is, when pixel i is within the 8-neighborhood of pixel j, , otherwise it is 0; its global mean and variance are statistically obtained from the rainfall feature data of that day. When calculating statistical significance, the significance threshold is set to 1.96, corresponding to a 95% confidence level. When , it is determined that there is significant aggregation or outlier in space for this pixel. Finally, according to formula (5), all pixels in the region are assigned an anomaly degree to obtain a new rainfall anomaly feature. The anomaly detection result of the cumulative rainfall in the past seven days on September 13, 2023 is as Figure 2 shown.

[0129] (4)Coupled evaluation of landslide disaster susceptibility and hazard

[0130] The coupled learning framework for landslide disaster susceptibility and hazard of the present invention is as Figure 3 shown. Considered as adjustable parameters, their value ranges can be set to [0, 1]. The susceptibility index under the current weight combination is calculated through formulas. . When sampling dangerous negative samples, first assign dates to the negative samples. In the present invention, the ratio of positive samples to negative samples is 1, so the dates of negative samples can simply be set as duplicates of the dates of positive samples. Then, for each date t existing in a negative sample, first filter the candidate set for sampling: eliminate pixels and retain the pixels with normal rainfall; let . If , then it is not included in the negative samples. Randomly select points corresponding to the quantity of date t from the remaining non-landslide pixels after double filtering as negative samples. After the negative sample sampling is completed, they are spliced and combined with the positive samples composed of all landslide events to form a risk model evaluation sample. Then, according to the coordinates of each sample point and date t, obtain the eigenvalue of the corresponding pixel, and the feature set is shown in formula (8). Finally, assign labels to the sample columns, assign 1 to positive samples and 0 to negative samples.

[0131] Divide the samples into a training set and a test set according to a ratio of 7:3. Select the random forest model, set the number of trees to 100, and the maximum tree depth to 4. Input the training set for fitting to obtain a training model, then calculate the accuracy index of the model on the test set, calculate the objective function value according to formula (16), take minimizing this value as the optimization objective, and continuously iterate through Bayesian optimization to finally obtain the optimal parameters, including the optimal weights of disaster-forming characteristics and the optimal susceptibility constraint threshold . Through the optimal , input the static disaster-forming characteristic interval index of all pixels in the province to obtain a susceptibility evaluation result map, as shown in Figure 4 .

[0132] For each pixel i, input its corresponding feature vector into the trained risk evaluation model, including static eigenvalues and dynamic eigenvalues of a certain period, and the landslide disaster risk index of the corresponding period can be obtained. When the ranges of all feature spaces correspond to the spatial positions of grid cells, output the risk indices of all pixels to obtain a regional risk evaluation map, as shown in Figure 5 .

[0133] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof.

[0134] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A coupled evaluation method for landslide disaster susceptibility and risk, characterized in that Including: Step 1: Extract landslide disaster factor features from multi-source data of the target area and perform preprocessing to obtain raster feature data; Step 2: Divide the raster feature data into intervals and calculate the landslide event frequency of each interval; Step 3: Calculate the local Moran's index of the cumulative rainfall within a preset time interval based on the rainfall raster data in the raster feature data to identify abnormal areas in the rainfall spatial distribution, and obtain rainfall anomaly factors; Step 4: Perform a coupling evaluation of susceptibility and hazard based on Bayesian optimization according to the landslide event frequency and rainfall anomaly factors to obtain the optimal weights of disaster-forming characteristics and a hazard evaluation model; The specific content of Step 4 includes: Step 4.1: Based on the landslide event frequency, combine the frequencies of different types of features according to the set initial weight vector to obtain their corresponding weight coefficients, and synthesize the scores of each feature interval and the weight coefficients of the corresponding features on each pixel to obtain a susceptibility factor characterizing the vulnerability of the geological environment; Step 4.2: Randomly generate the date of each pixel, then form a negative sample set of pixels that meet the preset conditions, and splice and combine it with the positive samples composed of all landslide events to form a hazard model evaluation sample, where the preset condition is that the rainfall anomaly factor corresponding to the pixel is 0 and the susceptibility factor is less than the susceptibility threshold; Step 4.3: Construct a hazard feature set according to the dynamic factor features, rainfall anomaly factors, and susceptibility factors, where the dynamic factor features are rainfall raster data; Step 4.4: Index and obtain the corresponding feature values from the hazard feature set according to the coordinates and dates of each sample data in the hazard model evaluation sample, and divide them into a training set and a test set according to a preset ratio; Step 4.5: Use the training set to train a random forest model and calculate the accuracy index of the trained model on the test set; Step 4.6: Construct an objective function according to the accuracy rate and F1 score in the accuracy index, take minimizing the objective function value as the optimization goal, and iteratively solve the objective function through Bayesian optimization to obtain the optimal weights of disaster-forming characteristics and the optimal susceptibility threshold; Step 4.7: Take the optimal susceptibility threshold as the new susceptibility threshold and execute Steps 4.2 to 4.5 again to obtain a hazard evaluation model; Step 5: Based on the optimal weights of disaster-forming characteristics, input the real-time static disaster-forming feature interval index of each raster in the target area to obtain a susceptibility evaluation result map, and input the real-time feature vector of each raster into the hazard evaluation model to obtain a hazard evaluation result map.

2. The method according to claim 1, characterized in that, The specific content of Step 1 includes: Step 1.1: Divide the multi-source data of the target area into static factor features and dynamic factor features according to the data feature update cycle, where the static factor features include geological structure elements, topographic and geomorphic features, deformation features, and human activity factors; Step 1.2: Unify the spatial reference and raster resolution of the static factor features and dynamic factor features to obtain raster feature data.

3. The method according to claim 2, characterized in that, The specific content of Step 2 includes: Step 2.1: Perform interval processing respectively according to the value types of each feature in the static factor features to obtain an interval set corresponding to each feature; Step 2.2, calculate the landslide event frequency of each interval in the interval set corresponding to each feature according to a preset formula, where the preset formula is ； Among them, is the total number of pixels where the value of feature f in the target area is in the interval k, represents the number of landslide-occurring pixels counted, is the total number of positive samples, is the total number of pixels in the target area.

4. The method according to claim 3, characterized in that, The specific steps of Step 3 include: Step 3.1, calculate the local Moran's index corresponding to the rainfall of each pixel within a preset time interval according to the rainfall raster data; Step 3.2, perform a statistical test on the local Moran's indices of all pixels according to a preset significance threshold to obtain a significance score; Step 3.3, identify rainfall anomaly pixels according to the significance score and calculate the corresponding rainfall anomaly factors ； Among them, represents the local Moran's index of the pixel , represents the significance score of the pixel , represents the significance threshold, represents the high-high aggregation area of the pixel and its neighborhood, represents the high-low aggregation area of the pixel and its neighborhood.

5. The method according to claim 4, characterized in that, The expression of the objective function is ； Among them, represents the accuracy rate, represents the weight coefficient, represents the easy-occurrence threshold, represents the F1 score.

Images