Landslide meteorological early warning method coupling machine learning and rainfall early warning index

Abstract

The invention relates to a landslide meteorological early warning method coupling machine learning and a rainfall early warning index, which comprises the following steps: taking a frequency ratio of screened landslide evaluation factors in different grading intervals as input, taking a landslide occurrence state as output, adopting a machine learning model to carry out landslide susceptibility modeling, and establishing a landslide susceptibility partition map; modeling a quantitative relation between the rainfall scene and the rainfall early warning index; calculating early effective rainfall and induced event rainfall in different return periods by adopting Gumbel distribution based on historical rainfall statistical data of the to-be-researched area, determining a minimum rainfall early warning index and a maximum rainfall early warning index, and obtaining a rainfall early warning index distribution diagram of the to-be-researched area; and carrying out coupling superposition on the landslide susceptibility partition map and the rainfall early warning index distribution map to obtain a landslide risk meteorological early warning map of the to-be-researched area. According to the method, a scientific basis is provided for regional scale landslide early warning, and the problem of splitting of a traditional model in data utilization and mechanism interpretation is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of geological disaster early warning, and specifically relates to a landslide meteorological early warning method that couples machine learning with a rainfall early warning index, which is used to achieve refined graded early warning of landslide disasters at a regional scale. Background Art

[0002] Landslide disasters are one of the common geological disasters around the world. Their occurrence is closely related to multiple factors such as topography, geology, and meteorology. They are characterized by suddenness and great destructiveness. Regional landslides induced by rainfall, in particular, pose a serious threat to people's lives and property.

[0003] Traditional landslide early warning methods primarily rely on physical or statistical models, but these models have significant limitations. Purely physical models (such as the infinite slope model) ignore complex geological factors, while machine learning models lack physical mechanisms to explain them. They are unable to quantify the direct impact of rainfall on landslides and fail to consider the coupled effects of rainfall accumulation and extreme events. Statistical models rely on historical landslide data to categorize landslide susceptibility without incorporating real-time rainfall dynamics, resulting in poor early warning effectiveness.

[0004] In recent years, machine learning models (such as support vector machines and neural networks) have been introduced into landslide susceptibility assessment, which can automatically extract features from multidimensional environmental factors. However, their prediction results lack an explanation of the physical mechanism of the rainfall dynamic process, resulting in a single warning threshold that is difficult to reflect the risk changes under different rainfall scenarios. In addition, the application of a single model often leads to limited warning accuracy.

[0005] Furthermore, traditional methods often use static thresholds to categorize warning levels, resulting in unreliable warning results under extreme rainfall or complex terrain conditions. For example, a fixed rainfall threshold (e.g., daily rainfall ≥100 mm) is pre-set as the landslide warning redline, and once exceeded, a fixed-level alert (e.g., yellow / red) is issued. This method fails to establish a dynamic correlation between rainfall parameters and landslide hazard, ignoring the dynamic rainfall process: the combined effects of rainfall intensity, duration, and cumulative rainfall are not considered. Furthermore, the same threshold is applied to different terrains, such as mountains and plains, disregarding differences in the geological environment.

[0006] Therefore, there is an urgent need for a comprehensive warning method that integrates the advantages of multiple models and takes into account both spatial prediction accuracy and physical process interpretability to improve the accuracy and practicality of landslide warnings under complex meteorological conditions. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention aims to provide a landslide meteorological warning method that couples machine learning with a rainfall warning index. By quantifying the weights of environmental factors, constructing a quantitative relationship between rainfall scenarios and rainfall warning indices, and implementing matrix-based warning rules, this method improves warning accuracy and spatial resolution. Through landslide susceptibility modeling and rainfall warning index analysis, this method provides a scientific basis for regional-scale landslide warnings. This not only addresses the disconnect between data utilization and mechanism interpretation in traditional models, but also provides more practical technical support for disaster prevention and mitigation.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] A landslide meteorological early warning method coupling machine learning with a rainfall early warning index, the early warning method comprising the following contents:

[0010] Step S1: Screening landslide evaluation factors through landslide development contribution, Pearson correlation coefficient, and multicollinearity test, using the frequency ratio of the screened landslide evaluation factors in different classification intervals as input and the landslide occurrence status as output, using a machine learning model to model landslide susceptibility. The machine learning model is trained to predict the landslide susceptibility of the study area and establish a landslide susceptibility zoning map;

[0011] Step S2: The quantitative relationship between the modeled rainfall scenario and the rainfall warning index is:

[0012] R=a+bP e 2 +cP e +dlg(P a )

[0013] Where R represents the rainfall warning index, P e represents the induced event rainfall, P a represents the previous effective rainfall; a, b, c, d are constant parameters;

[0014] Based on the FSLAM model or TRIGRS model, different induced event rainfall P e With different previous effective rainfall P a After the combination, different rainfall scenarios were obtained. The percentage of grids with a damage probability greater than 0.5 under each rainfall scenario was counted. The percentage of grids with a damage probability greater than 0.5 was used as the rainfall warning index R. The R corresponding to each rainfall scenario was used to form an array to fit and determine the values ​​of a, b, c, and d.

[0015] Step S3: Using Gumbel distribution, based on the historical rainfall statistics of the study area, calculate the previous effective rainfall P with a return period of 2 to 100 years. a and induced rainfall event Pe , determine the maximum induced event rainfall P e The corresponding return period is P e = 0 and the previous effective rainfall P under a 2-year return period a Substitute the quantitative relationship in step S2 to determine the minimum rainfall warning index R min , P e The previous effective rainfall P under the maximum corresponding return period a and induced rainfall event P e Substitute the quantitative relationship in step S2 to determine the maximum rainfall warning index R max ; The range of rainfall warning index value in the study area is R min to R max During the period, the rainfall warning index distribution map is obtained;

[0016] Step S4: The landslide susceptibility zoning map obtained by the machine learning model in step S1 is coupled and superimposed with the rainfall warning index distribution map determined in step S3 to obtain a landslide hazard meteorological warning map for the area to be studied.

[0017] Furthermore, in step S3, within the range of the rainfall warning index value, the rainfall warning index is evenly divided into five levels, namely, very low, low, medium, high, and very high;

[0018] In step S1, the landslide susceptibility zone is divided into five susceptibility levels, namely, very low susceptibility zone, low susceptibility zone, medium susceptibility zone, high susceptibility zone and very high susceptibility zone;

[0019] The coupling and superposition of landslide susceptibility and rainfall warning indices were completed using the raster calculator tool in ArcGIS software. A 5×5 matrix grading rule for landslide susceptibility zones and rainfall warning indices was established, and the warning level corresponding to each combination in the matrix was defined.

[0020] Furthermore, the rainfall warning index changes dynamically with the rainfall conditions. When the rainfall warning index exceeds the threshold of each level, different levels of landslide hazard warnings are activated.

[0021] Furthermore, the machine learning model is an SVM model, a neural network model, etc.

[0022] The present invention also protects a computer-readable storage medium having a computer program stored thereon, which can implement the steps of the method described in the claims when the program is executed by a processor.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1) Using dual-model coupling technology, the weights of static geological factors are optimized through machine learning models, and the impact of rainfall infiltration is dynamically quantified in combination with physical models, effectively overcoming the defects of traditional methods in separating dynamic and static factors.

[0025] 2) Based on the Gumbel extreme value distribution theory, a three-dimensional dynamic rainfall warning threshold surface is constructed. a -P e Based on the combined influence of rainfall and landslide risk, a quantitative conversion between rainfall at different time scales and landslide risk was established, which provides an important scientific basis for further realizing landslide probability graded meteorological warning under different rainfall combination scenarios.

[0026] 3) By establishing a matrix-based grading rule for five-level landslide susceptibility zones and five-level dynamic rainfall warning indices, the corresponding warning level for each combination is defined. When both the susceptibility level and the rainfall warning index are high, the warning level is also high. This multi-dimensional spatial coupling criterion significantly improves the spatial resolution and risk assessment accuracy of regional landslide warnings, resolving the technical bottleneck of a single static model's slow response to rainfall dynamics.

[0027] 4) The method of the present invention fully considers the dynamic physical process of "rainfall-seepage-rock and soil response" in real landslide warning, can characterize the dynamic evolution of slope stability under the cumulative action of rainfall, and realize the dynamic update of the rainfall warning index. It improves the accuracy of warning, especially in areas with abnormal rainfall patterns or geologically sensitive areas, and avoids the shortcomings of the existing technology that simplifies landslide warning into a static mapping of "rainfall-hazard", resulting in threshold rigidity and causing catastrophic misjudgment.

[0028] 5) The method of the present invention can adjust the warning level in real time and quantitatively based on rainfall intensity, duration, etc., and can dynamically calculate and update the current slope stability state. When the soil moisture and groundwater conditions are in dynamic changes (such as the saturation period after rain), it can adapt to the risk mutations brought about by heavy rainfall that far exceeds historical records, realizing data-driven dynamic association. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a surface diagram of the dynamic rainfall warning threshold based on the FSLAM model in Example 2 of the present invention.

[0030] Figure 2 This is a schematic diagram of the rainfall warning index classification based on the FSLAM model in Example 2 of the present invention.

[0031] Figure 3 The landslide hazard warning classification diagram of the coupled landslide susceptibility and rainfall warning index in Example 2 of the present invention is as follows: (a) very low hazard; (b) low hazard; (c) medium hazard; (d) high hazard; (e) very high hazard. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] The landslide meteorological early warning method of the present invention, which couples machine learning with a rainfall early warning index, comprises the following steps:

[0034] Step S1: Screening landslide evaluation factors through landslide development contribution, Pearson correlation coefficient, and multicollinearity test, using the frequency ratio of the screened landslide evaluation factors in different classification intervals as input and the landslide occurrence status as output, using a machine learning model to model landslide susceptibility. The machine learning model is trained to predict the landslide susceptibility of the study area and establish a landslide susceptibility zoning map;

[0035] Step S2: The quantitative relationship between the modeled rainfall scenario and the rainfall warning index is:

[0036] R=a+bP e 2 +cP e +dlg(P a )

[0037] Where R represents the rainfall warning index, P e represents the induced event rainfall, P a represents the previous effective rainfall; a, b, c, d are constant parameters;

[0038] Based on the FSLAM model or TRIGRS model, different induced event rainfall P e With different previous effective rainfall P a After the combination, different rainfall scenarios were obtained. The proportion of grids with a damage probability greater than 0.5 under each rainfall scenario (as a proportion of the entire study area) was counted. The proportion of grids with a damage probability greater than 0.5 was used as the rainfall warning index R. The R corresponding to each rainfall scenario was used to form an array to determine the values ​​of a, b, c, and d.

[0039] Step S3: Using Gumbel distribution, based on the historical rainfall statistics of the study area, calculate the previous effective rainfall P with a return period of 2 to 100 years. a and induced rainfall event P e , determine the maximum induced event rainfall P e The corresponding return period is P e = 0 and the previous effective rainfall P under a 2-year return period a Substitute the quantitative relationship in step S2 to determine the minimum rainfall warning index R min , Pe The previous effective rainfall P under the maximum corresponding return period a and induced rainfall event P e Substitute the quantitative relationship in step S2 to determine the maximum rainfall warning index R max ; The range of rainfall warning index value in the study area is R min to R max During the forecast period, the rainfall warning index distribution map was obtained.

[0040] Step S4: The landslide susceptibility zoning map obtained by the machine learning model in step S1 is coupled and superimposed with the rainfall warning index distribution map determined in step S3 to obtain a landslide hazard meteorological warning map for the area to be studied.

[0041] The present invention adopts the Gumbel distribution method to estimate the extreme value of rainfall in the future specific return period of the study area, which can be achieved by existing technology. Based on historical rainfall statistical data, the previous effective rainfall Pa and induced event rainfall Pe under different return periods are calculated, which can provide a rainfall standard reference for landslide early warning. Specifically, the daily temperature, rainfall and evapotranspiration data of the study area in recent decades are obtained. The Gumbel distribution calculation code developed based on the Matlab program is used to input the historical rainfall statistical data of the study area to determine the previous effective rainfall P under different rainfall return periods. a and induced rainfall event P e .

[0042] In the present invention, when determining the parameters a, b, c, and d, the FSLAM model or the TRIGRS model can be used to obtain the rainfall warning index under different rainfall scenarios for fitting and determination. e With different previous effective rainfall P a As independent variables, multiple rainfall scenarios are obtained after combination as input rainfall, which is input into the above-mentioned FSLAM model or TRIGRS model.

[0043] This paper proposes the concept of rainfall warning index for the first time. The proportion of grids with a probability of damage greater than 0.5 under each rainfall scenario is used as the rainfall warning index, which is represented by the dependent variable R. Grids in this state are considered to be more dangerous. In the quantitative relationship, a, b, c, and d are constant parameters. These parameters can quantitatively reflect the quantitative relationship between rainfall scenarios and rainfall warning index. By combining multiple groups of rainfall scenarios (P e With P a ) and the corresponding rainfall warning index, the fitting coefficient can be determined and the specific function expression can be obtained.

[0044] In this invention, when there is no rainfall in the study area (i.e., P e = 0, the probability of damage is the lowest) when the corresponding minimum rainfall warning index Rmin The study area is under extreme rainfall that occurs once in a century (P e At this time, the maximum rainfall warning index R is obtained according to the quantitative relationship under the condition of maximum max The rainfall warning index value range R in the study area is obtained. min to R max During the forecast period, the rainfall warning index was divided into five levels: very low, low, medium, high and very high.

[0045] The overlay of the susceptibility and rainfall warning indices was performed using the Raster Calculator tool in ArcGIS software. A 5×5 matrix of susceptibility zones and rainfall warning indices was established, and the corresponding warning level for each combination in the matrix was defined. The rainfall warning index can dynamically adjust the warning level based on changes in rainfall conditions. When the rainfall warning index exceeds a specific threshold (each rainfall warning index level has a corresponding threshold), a landslide hazard warning of different levels is triggered.

[0046] Furthermore, step S1 includes the following contents:

[0047] S101: Integrate the topography (slope, curvature), geology (lithology, soil thickness), land use and other environmental factors and historical landslide data of the study area;

[0048] S102: The frequency ratio method is used to quantitatively evaluate the contribution of each environmental factor to landslide development. If the frequency ratio FR is greater than 1, it means that the factor classification interval has a high correlation with landslide occurrence; if FR is less than 1, it means that the correlation is low, so the environmental factors with FR greater than 1 are retained.

[0049] S103: Through factor correlation analysis and multicollinearity test, highly correlated factors are eliminated and the model input variables are optimized. The factors after screening are called evaluation factors.

[0050] S104: Adjust the parameters of the machine learning model, and use the screened input variables and the corresponding output variables (landslide occurrence status (occurrence status refers to the susceptibility index. By applying the trained machine learning model to each grid unit in the entire area, the landslide susceptibility index (0-1) can be output, and the risk level (such as extremely high, high, medium, low, and extremely low) can be divided, and then a susceptibility zoning map can be generated) for training and verification of the machine learning model.

[0051] This example uses an SVM model for landslide susceptibility modeling. Landslide surfaces within the study area are captured through remote sensing image interpretation, and non-landslide points are randomly generated in QGIS software, forming an equal set of landslide / non-landslide samples for training and validating the machine learning model. During model training and validation, the frequency ratios of these samples at different grading intervals of the selected landslide evaluation factors serve as input variables for training the machine learning model (i.e., the input variables are the frequency ratios calculated for each graded evaluation factor after screening). The susceptibility index of known landslide units is set to 1, and the susceptibility index of non-landslide units is set to 0.

[0052] The environmental factors include all environmental data that may affect landslides, such as elevation, slope, slope direction, plan curvature, profile curvature, stratum lithology, soil type, and distance from faults. The evaluation factors are modeling variables that can be directly input into the model and are screened from the environmental factors to quantitatively characterize the susceptibility to landslides.

[0053] S105: Use the trained machine learning model to predict the landslide susceptibility of the study area and classify it into different landslide susceptibility levels. Use the natural breakpoint classification method to divide the susceptibility areas into five levels (very low, low, medium, high, and very high). For the SVM model, the landslide susceptibility index ranges corresponding to the five susceptibility levels are 0-0.115, 0.115-0.277, 0.277-0.480, 0.480-0.717, and 0.717-1.

[0054] Further: The quantitative relationship determination process is:

[0055] S201: Calculate the destruction probability of the grid under different combined rainfall scenarios based on the FSLAM model, define the rainfall warning index, and analyze the physical impact of rainfall on landslide hazard.

[0056] S202: Select different induced rainfall events P e and different antecedent effective rainfall P a As the independent variable, multiple rainfall scenarios are combined as input rainfall, and the rainfall warning index calculated from the input rainfall scenario is used as the dependent variable R. A quantitative relationship model between Pa-Pe and the rainfall warning index is constructed to generate a dynamic warning threshold surface.

[0057] S203: Using nonlinear surface fitting to perform regression analysis on the three-dimensional spatial data points, a quantitative relationship between the rainfall scenario and the rainfall warning index is obtained, which can systematically quantify the coupling relationship between the rainfall parameters and the rainfall warning index.

[0058] By coupling and superimposing the landslide susceptibility zoning with the five rainfall warning index levels, a meteorological warning map for landslide hazard in the study area was obtained. The coupled warning judgment matrix is ​​shown in Table 1. Spatially targeted warnings are achieved by matching warning levels with real-time rainfall data.

[0059] Table 1 is the meteorological early warning judgment matrix of rainfall-induced landslide in the present invention.

[0060]

[0061] Example 1:

[0062] The landslide meteorological warning method of this embodiment is coupled with machine learning and rainfall warning index. The specific steps are the same as above, wherein:

[0063] The daily temperature, rainfall, and evapotranspiration data for the study area over the past few decades were collected and entered into the EasyBal software to calculate the monthly effective infiltration, which is the portion of rainfall that can penetrate into the soil in a particular month. The ratio of effective rainfall infiltration to the number of days in the month is the previous effective rainfall P. a The data basis of extreme value analysis.

[0064] The Gumbel distribution calculation code developed based on Matlab program was used to input the historical rainfall statistical data of the study area, including the monthly effective rainfall infiltration and the maximum two-day rainfall extreme value in the study area in recent decades, to calculate the P of the study area under different return periods of 2 to 100 years. a and P e value.

[0065] The Gumbel distribution is widely used in hydrology and meteorology to analyze extreme events, such as the statistical properties of rainfall extremes. The probability density function (PDF) f(x) and cumulative distribution function (CDF) F(x) of the Gumbel distribution are shown below:

[0066]

[0067] Among them, -∞<x<∞, β is the scale parameter, and μ is the location parameter.

[0068] Input a variety of rainfall scenarios into the FSLAM model, and calculate the proportion of grids with damage probability > 0.5 as the rainfall warning index. Based on the FSLAM model, combined with different P a and P e The combined scenarios were used to calculate the rainfall warning index for each rainfall scenario, and a quantitative relationship model was fitted to obtain the relationship between rainfall conditions and the rainfall warning index. The rainfall warning index was divided into five levels: very low, low, medium, high, and very high.

[0069] In the calculation process of the FSLAM model, static topographic and geological parameters such as elevation, soil type and land use are calibrated and kept within a fixed threshold range. The model calculation results are mainly affected by the previous effective rainfall P. a and induced rainfall event P e Therefore, the rainfall warning index level of regional landslides only changes with P a and P e The combined rainfall scenarios vary.

[0070] Select different triggering events rainfall P e and different antecedent effective rainfall P a As the independent variable, multiple rainfall scenarios were obtained as input rainfall after combination. The rainfall warning index calculated from the input rainfall scenario was then used as the dependent variable. A three-dimensional spatial relationship between the induced event rainfall, the previous effective rainfall, and the rainfall warning index was constructed, i.e., a quantitative relationship model between rainfall scenarios and rainfall warning index:

[0071] R=a+bP e 2 +cP e +dlg(P a )

[0072] Where R represents the rainfall warning index, P e represents the induced event rainfall, P a represents the previous effective rainfall; a, b, c, and d are constant parameters.

[0073] Through the quantitative conversion of rainfall and landslide hazards at different time scales, a dynamic rainfall warning threshold surface based on the FSLAM model was constructed, which provides an important scientific basis for further realizing landslide probability graded meteorological warning under different rainfall combination scenarios.

[0074] Furthermore, the Fast Shallow Landslide Assessment Model (FSLAM) is a physics-based deterministic model used to rapidly assess shallow landslide hazard in a region. The FSLAM model consists of two main components: a geotechnical model and a hydrological model. The geotechnical model is used to calculate slope stability and the failure probability of a grid cell, which can be used to assess landslide hazard under different rainfall conditions.

[0075] A quantitative relationship between rainfall scenarios and rainfall warning indices was constructed for critical classification of rainfall warnings. Regional landslide hazard shows a clear upward trend with increasing rainfall, but this growth is not unlimited. Instead, it is constrained by upper and lower limits of rainfall. The proportion of grid cells in the study area that are in a dangerous state has a natural upper and lower limit.

[0076] When there is no rainfall in the study area, the entire area is in the most stable state, and the regional landslide risk is at its lowest value. The minimum rainfall warning index R is set according to the calculation results of the corresponding rainfall scenario. min .

[0077] When the most dangerous rainfall scenario is under extreme rainfall conditions exceeding 1 in 100 years, the landslide risk level in the region reaches the highest value, and the maximum rainfall warning index R is set. max .

[0078] The rainfall warning index value range of the study area is R min to R max The rainfall warning index is divided into five levels: very low, low, medium, high and very high.

[0079] Based on the above upper and lower limits, a schematic diagram of the rainfall warning index classification based on the FSLAM model is drawn.

[0080] Example 2

[0081] This example applies the landslide meteorological warning method coupled with machine learning and rainfall warning index to the southwest of Wenzhou City, Zhejiang Province. The total area of ​​the study area is approximately 2067.7 km 2 .

[0082] S1: Data Collection

[0083] Collect basic data such as topographic factors, basic geological factors, hydrogeological factors, and vegetation cover factors of the study area. At the same time, obtain rainfall data and rainfall intensity data from the meteorological station.

[0084] S2: Landslide Susceptibility Modeling

[0085] The distribution data of landslide hazards in the study area in different intervals of these factors, such as number and area, were statistically analyzed. The contribution of each environmental factor to landslide development was quantitatively evaluated using the frequency ratio method to obtain the input variables for landslide susceptibility modeling.

[0086] Specifically, the frequency ratio (FR) is a statistical analysis method that quantifies the spatial correlation and influence of different factors on landslide occurrence. It can also be used to assess the relative impact of each factor classification on landslide occurrence. It is based on the relationship between historical landslides and landslide causative factors. By calculating the ratio of the landslide area within a certain factor classification interval to the proportion of the area of ​​this factor classification interval to the total area of ​​the study area, it reflects the contribution of this factor classification interval to landslide occurrence. The frequency ratio FR quantifies the impact of a certain factor classification interval on landslide occurrence by comparing the proportion of landslide area within a certain factor classification interval with its proportion of the entire area. FR>1 indicates a strong positive correlation; FR<1 indicates a weak correlation.

[0087] If the FR value is greater than 1, it means that the factor interval has a high correlation with landslide occurrence; if the FR value is less than 1, it means that the correlation is low. The calculation formula of FR is as follows:

[0088]

[0089] Among them, N i is the landslide area of ​​the classification interval i, N is the total landslide area, S i is the distribution area of ​​classification interval i, and S is the total area of ​​the study area.

[0090] Furthermore, the factors were subjected to correlation analysis and multicollinearity test.

[0091] In landslide susceptibility modeling, highly correlated spatial data can lead to redundant information, increasing model complexity and potentially reducing predictive accuracy and stability. Redundant factors with high Pearson correlation coefficients (i.e., strong correlations between factors) were removed, and the remaining factors were retested for multicollinearity. Appropriate landslide assessment factors were then selected for subsequent landslide susceptibility modeling.

[0092] The SVM machine learning model was selected for landslide susceptibility modeling. The SVM model parameters were adjusted, and the frequency ratios of each graded evaluation factor, calculated after screening, were used as input variables and the corresponding output variables (landslide occurrence status) for machine learning model training and validation. Landslides were used in a 7:3 ratio for model training and testing to determine the final SVM model parameters.

[0093] A landslide susceptibility map was created for the study area using the trained SVM model. Natural breaks were used to categorize landslide susceptibility into five levels: very low susceptibility, low susceptibility, moderate susceptibility, high susceptibility, and very high susceptibility. For the SVM model, the corresponding landslide susceptibility index ranges for the five susceptibility levels were 0–0.115, 0.115–0.277, 0.277–0.480, 0.480–0.717, and 0.717–1, respectively.

[0094] S3: Estimation of rainfall extremes

[0095] Daily temperature, rainfall, and evapotranspiration data for Wenzhou from 1981 to 2016 were collected and entered into EasyBal software to calculate monthly effective infiltration.

[0096] Calculate the previous effective rainfall P under different rainfall return periods a and induced rainfall event P eBased on the Gumbel distribution calculation code developed by Matlab program, the historical rainfall statistical data of the study area (i.e., the monthly effective rainfall infiltration and the maximum two-day rainfall extreme value from 1981 to 2015) were input to calculate the P of the study area under different return periods of 2 to 100 years. a and P e Among them, under the 20-year return period, P a 5.63 mm / d, P e is 347.47 mm; under a 50-year return period, P a 7.30mm / d, P e is 415.13 mm; under a 100-year return period, P a 8.55mm / d, P e It is 465.83mm.

[0097] S4: Rainfall Warning Index Analysis

[0098] 49 rainfall scenario combinations were selected (P a Changing in the range of 0~1mm / d, P e The calculation results of the rainfall warning index (ranging from 0 to 600 mm) are used to count the proportion of grids with a damage probability greater than 0.5 under each rainfall scenario. Grids in this state are considered to be more dangerous and are defined as the rainfall warning index.

[0099] Select 6 different triggering events rainfall P e With 6 different antecedent effective rainfall P a As the independent variable, 36 rainfall scenarios were obtained as input rainfall after combination, and the rainfall warning index calculated from the input rainfall scenario was used as the dependent variable R. The three-dimensional spatial relationship between the induced event rainfall, the previous effective rainfall and the rainfall warning index was constructed, that is, the quantitative relationship model between the rainfall scenario and the rainfall warning index, as shown in the following example: Figure 1 shown.

[0100] The quantitative relationship between rainfall scenarios and rainfall warning index is obtained as follows:

[0101] R = 0.41577-1.31369 × 10 -7 ×(P e ) 2 +1.85306×10 -4 ×P e +0.0123lg(P a )

[0102] The results show that the fitting accuracy index R 2=0.922, indicating that the nonlinear surface equation is well fitted. Among them, the obtained parameters a are 0.41577 and b are -1.31369×10 -7 , c is 1.85306×10 -4 , d is 0.0123.

[0103] When there is no rainfall in the study area, the corresponding minimum rainfall warning index R min =0.359. The maximum rainfall warning index R corresponding to the extreme rainfall conditions of once in a hundred years in the study area max =0.508. Therefore, the rainfall warning index value range of the study area is between 0.359 and 0.508.

[0104] The rainfall warning index is divided into five levels: extremely low (0.359-0.389), low (0.389-0.419), medium (0.419-0.448), high (0.448-0.478) and extremely high (0.478-0.508).

[0105] The specific formula is as follows:

[0106] 0.359≤0.41577-1.31369×10 -7 ×(P e ) 2 +1.85306×10 -4 ×P e +0.0123In(P a )<0.389

[0107] 0.389≤0.41577-1.31369×10 -7 ×(P e ) 2 +1.85306×10 -4 ×P e +0.0123In(P a )<0.419

[0108] 0.419≤0.41577-1.31369×10 -7 ×(P e ) 2 +1.85306×10 -4 ×P e +0.0123In(P a )<0.448

[0109] 0.448≤0.41577-1.31369×10 -7 ×(P e ) 2+1.85306×10 -4 ×P e +0.0123In(P a )<0.478

[0110] 0.478≤0.41577-1.31369×10 -7 ×(P e ) 2 +1.85306×10 -4 ×P e +0.0123In(P a )<0.508 Based on the above upper and lower limits, the rainfall warning index classification diagram is drawn as follows Figure 2 shown.

[0111] S5: Landslide Weather Warning

[0112] The landslide susceptibility zoning (five levels) of the machine learning model was coupled and superimposed with the FSLAM rainfall warning index (five levels), and a warning judgment matrix with 25 combinations was constructed to obtain the meteorological warning map of landslide hazard in the study area, as shown in the following figure: Figure 3 shown.

[0113] The overlay of the susceptibility and rainfall warning indices was performed using the Raster Calculator tool in ArcGIS software. A 5×5 matrix grading rule was established, combining the five-level geological susceptibility zones and the five-level dynamic rainfall warning indices. The corresponding warning level for each combination was defined. When both the susceptibility level and the rainfall warning index were high, the warning level was also high.

[0114] The method of the present invention optimizes the weights of static environmental factors through the SVM model, combines the rainfall warning index classification divided by the FSLAM model, establishes a rainfall-induced landslide meteorological warning judgment matrix, and thus constructs a comprehensive meteorological warning model.

[0115] Using a Gumbel distribution to quantify rainfall extremes, and generating a dynamic threshold surface based on nonlinear fitting, this approach enables refined assessment of regional landslide risk. The resulting quantitative relationship between rainfall scenarios and rainfall warning indices has universal applicability. Demonstration experiments have demonstrated that the system achieves an 82% warning accuracy in typhoon and rainstorm scenarios, a significant improvement over traditional methods. This approach not only improves the accuracy of landslide warnings but also provides a scientific basis for geological disaster prevention and control.

[0116] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A landslide meteorological warning method coupling machine learning and rainfall warning index, characterized in that: The early warning method includes the following contents: Step S1: Screening landslide evaluation factors through landslide development contribution, Pearson correlation coefficient, and multicollinearity test, using the frequency ratio of the screened landslide evaluation factors in different classification intervals as input and the landslide occurrence status as output, using a machine learning model to model landslide susceptibility. The machine learning model is trained to predict the landslide susceptibility of the study area and establish a landslide susceptibility zoning map; Step S2: The quantitative relationship between the modeled rainfall scenario and the rainfall warning index is: R=a+bP e 2 +cP e +dlg(P a ) Where R represents the rainfall warning index, P e represents the induced event rainfall, P a represents the previous effective rainfall; a, b, c, d are constant parameters; Based on the FSLAM model or TRIGRS model, different induced event rainfall P e With different previous effective rainfall P a After the combination, different rainfall scenarios were obtained. The percentage of grids with a damage probability greater than 0.5 under each rainfall scenario was counted. The percentage of grids with a damage probability greater than 0.5 was used as the rainfall warning index R. The R corresponding to each rainfall scenario was used to form an array to fit and determine the values ​​of a, b, c, and d. Step S3: Using Gumbel distribution, based on the historical rainfall statistics of the study area, calculate the previous effective rainfall P with a return period of 2 to 100 years. a and induced rainfall event P e , determine the maximum induced event rainfall P e The corresponding return period is P e = 0 and the previous effective rainfall P under a 2-year return period a Substitute the quantitative relationship in step S2 to determine the minimum rainfall warning index R min , P e The previous effective rainfall P under the maximum corresponding return period a and induced rainfall event P e Substitute the quantitative relationship in step S2 to determine the maximum rainfall warning index R max ; The range of rainfall warning index value in the study area is R min to R max During the period, the rainfall warning index distribution map is obtained; Step S4: The landslide susceptibility zoning map obtained by the machine learning model in step S1 is coupled and superimposed with the rainfall warning index distribution map determined in step S3 to obtain a landslide hazard meteorological warning map for the area to be studied.

2. The method according to claim 1, characterized in that In step S3, within the range of the rainfall warning index, the rainfall warning index is evenly divided into five levels, namely, very low, low, medium, high, and very high; In step S1, the landslide susceptibility zone is divided into five susceptibility levels, namely, very low susceptibility zone, low susceptibility zone, medium susceptibility zone, high susceptibility zone and very high susceptibility zone; The coupling and superposition of landslide susceptibility and rainfall warning indices were completed using the raster calculator tool in ArcGIS software. A 5×5 matrix grading rule for landslide susceptibility zones and rainfall warning indices was established, and the warning level corresponding to each combination in the matrix was defined.

3. The method according to claim 1, characterized in that The rainfall warning index changes dynamically with the rainfall conditions. When the rainfall warning index exceeds the threshold of each level, different levels of landslide hazard warnings are activated.

4. The method according to claim 1, wherein The machine learning model is an SVM model or a neural network model.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 4 can be implemented.

Images