Landslide risk evaluation method based on integrated deep learning recurrence period rainfall

Abstract

The invention relates to the technical field of slope risk evaluation, in particular to a landslide risk evaluation method based on integrated deep learning recurrence period rainfall, which comprises the following steps: acquiring landslide susceptibility evaluation factor data, and carrying out preprocessing and frequency ratio grading coefficient division; constructing a Stacking integration algorithm, training the Stacking integration algorithm by using the landslide susceptibility evaluation factor, and verifying the Stacking integration algorithm by using the susceptibility index; calculating the annual maximum daily rainfall of a given recurrence period by using Pearson III probability distribution, and constructing daily rainfall distribution diagrams under different recurrence periods; and replacing the annual average rainfall in the landslide susceptibility evaluation factor with the annual maximum daily rainfall of a given return period, predicting the trained Stacking integration algorithm, and outputting a landslide risk evaluation index. The problem of lack of a landslide risk evaluation system in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide hazard assessment, and in particular to a landslide hazard assessment method under return period rainfall based on integrated deep learning. Background Art

[0002] Landslide hazard refers to the possibility of landslide in a certain area under specific conditions and the risks it may bring. Landslide hazard is affected by a combination of factors, including geology, topography, climate, human activities, vegetation, etc. Through scientific assessment methods, landslide risks can be effectively identified and predicted, so that corresponding prevention and management measures can be taken to reduce the disaster losses caused by landslides.

[0003] When machine learning algorithms are applied to landslide susceptibility assessment, they are easily affected by data outliers and correlations between feature variables, resulting in underfitting results. Therefore, how to effectively combine machine learning models with ensemble learning algorithms and apply them to landslide susceptibility assessment remains an issue worthy of attention.

[0004] Patents with publication numbers CN116756700A and CN116522286A are mainly used for landslide displacement prediction and cannot be used for landslide hazard assessment. Currently, there are many comparative studies on different machine learning models, but there are few comparative analyses between integrated models and deep learning models as base learners. At the same time, most current hazard assessments use hazard indicators to directly train models, which may not be able to truly reflect reliable hazard assessment results. Summary of the invention

[0005] Aiming at the shortcomings of the existing methods, the present invention solves the problem of lack of landslide hazard assessment system in the prior art.

[0006] The technical solution adopted by the present invention is: a landslide hazard assessment method under return period rainfall based on integrated deep learning comprises the following steps:

[0007] Step 1: Collect landslide susceptibility evaluation factor data, and perform preprocessing and frequency ratio classification coefficient division;

[0008] As a preferred implementation of the present invention, the preprocessing includes: correlation analysis and multicollinearity test.

[0009] As a preferred embodiment of the present invention, the correlation analysis adopts the Pearson correlation coefficient method.

[0010] As a preferred embodiment of the present invention, the multicollinearity test index includes tolerance and variance inflation factor.

[0011] As a preferred embodiment of the present invention, the landslide susceptibility evaluation factors divided by frequency ratio classification coefficients include: slope, aspect, plane curvature, profile curvature, distance from the road, distance from the river, lithology, average annual rainfall, elevation, NDVI and land use.

[0012] Step 2: construct the Stacking ensemble algorithm, train the Stacking ensemble algorithm using the landslide susceptibility evaluation factor, and verify the Stacking ensemble algorithm using the susceptibility index;

[0013] As a preferred implementation of the present invention, the base learner of the Stacking ensemble algorithm uses LSTM, RNN, CNN and DNN integration, and the secondary learner adopts MLR; a grid search algorithm is used to optimize the hyperparameters of the number of neurons, batch size and number of iterations of the base learner.

[0014] Step 3: Calculate the annual maximum daily rainfall with a given return period using the Pearson III probability distribution, and construct a daily rainfall distribution map with different return periods;

[0015] As a preferred embodiment of the present invention, the annual maximum daily rainfall for a given return period is calculated using the Pearson III probability distribution, which includes:

[0016] Step 31, extracting the maximum rainfall value of each day in each year from the daily rainfall data of several years;

[0017] Step 32, calculating the shape, scale and location parameters of the Pearson III probability distribution;

[0018] Step 33: Calculate the annual maximum daily rainfall for a given return period using the probability of the rainfall return period. The formula is:

[0019] y=GAMMAINV(1-P r ,α,1 / β)+a0

[0020] Among them, P r is the probability of rainfall return period, α, β, a0 are the shape, scale and position of Pearson III distribution respectively.

[0021] As a preferred embodiment of the present invention, constructing a daily rainfall distribution map under different return periods includes: using the daily maximum rainfall under different return periods of each station in a county and adjacent counties, using the daily maximum rainfall and Kriging interpolation method to generate a daily maximum rainfall raster map of a county under the rainfall return period; using reclassification to obtain the daily maximum rainfall raster map.

[0022] Step 4: Replace the annual average rainfall in the landslide susceptibility evaluation factor with the annual maximum daily rainfall with a given return period, and use the trained Stacking ensemble algorithm to predict and output the landslide hazard evaluation index;

[0023] As a preferred embodiment of the present invention, the formula for landslide hazard evaluation index is:

[0024] probability=P r *I sus (3)

[0025] Among them, P r represents the probability of rainfall return period, I sus Represents the landslide hazard assessment index.

[0026] As a preferred embodiment of the present invention, a landslide hazard assessment system under a return period rainfall based on integrated deep learning includes: a memory for storing instructions executable by a processor; and a processor for executing instructions to implement a landslide hazard assessment method under a return period rainfall based on integrated deep learning.

[0027] Beneficial effects of the present invention:

[0028] 1. The present invention constructs a Stacking model based on a base learner and a two-layer learning to predict the landslide hazard evaluation index;

[0029] 2. When historical data are missing, the annual maximum daily rainfall for a given return period of 50 years or 100 years is calculated using the Pearson III probability distribution;

[0030] 3. Different from the traditional landslide hazard assessment method, an innovative proposal is made to use a prediction model based on landslide susceptibility assessment to predict the landslide hazard assessment data set to output a landslide hazard index to reflect a more realistic and reliable hazard assessment result. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the landslide hazard assessment method under return period rainfall based on integrated deep learning of the present invention;

[0032] Figure 2 This is the rainfall distribution map with a return period of 50 years;

[0033] Figure 3 This is the rainfall distribution map with a return period of 100 years;

[0034] Figure 4 Produce hazard maps for rainfall conditions with a 50-year return period;

[0035] Figure 5 Hazard mapping for rainfall conditions with a 100-year return period. DETAILED DESCRIPTION

[0036] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0037] like Figure 1 As shown in the figure, the landslide hazard assessment method under return period rainfall based on integrated deep learning includes the following steps:

[0038] Step 1: Select landslide susceptibility assessment factors and build a landslide susceptibility assessment factor database;

[0039] Specifically include:

[0040] A number of influencing factors related to landslide disasters are selected, the landslide susceptibility evaluation factors are preprocessed to obtain the preprocessed landslide susceptibility evaluation factors, and the landslide susceptibility evaluation factors are screened to obtain the screened landslide susceptibility evaluation factors.

[0041] The influencing factor data were converted into vector data, and the landslide list point data were connected to the influencing factor vector data. Through correlation analysis, it was obtained whether there was a high correlation between the influencing factors. The influencing factors were tested for multicollinearity, and the influencing factors with high correlation were appropriately eliminated. Finally, 11 influencing factors were selected, including slope, aspect, plane curvature, profile curvature, distance from the road, distance from the river, lithology, average annual rainfall, elevation, NDVI, and land use. After the evaluation factor data were acquired and spatial consistency was processed, a landslide susceptibility assessment factor database was constructed.

[0042] Taking 30m×30m as the basic unit, correlation analysis and multicollinearity analysis were conducted on all influencing factors to ensure the independence of factors and improve the accuracy of model prediction; the Pearson correlation coefficient method was used to test the correlation between factors;

[0043] In the correlation analysis and multicollinearity analysis, when the absolute value of the correlation coefficient is higher than 0.4, it is considered that there is a high correlation between the factors and they are eliminated; when the tolerance TOL is less than 0.1 or the variance inflation factor VIF is greater than 10, it is considered that the factor has strong collinearity and is eliminated.

[0044] Table 1 Factor correlation calculation results

[0045]

[0046] Table 2 Factor multicollinearity calculation results

[0047]

[0048]

[0049] After determining the influencing factors, the correlation of the grading factors of the influencing factors to the landslide is analyzed through the formula, that is, the relative frequency ratio of the landslide.

[0050] The slope division formula is:

[0051] The slope is ≤10, the classification is defined as 1, and the classification coefficient is defined as 0.25 according to the frequency ratio;

[0052] 10<slope≤20, the classification is defined as 2, and the classification coefficient according to the frequency ratio is defined as 0.99;

[0053] 20<slope≤30, the classification is defined as 3, and the classification coefficient according to frequency ratio is defined as 0.75;

[0054] 30<slope≤40, the classification is defined as 4, and the classification coefficient according to the frequency ratio is defined as 0.50;

[0055] For slopes > 40, the classification is defined as 5, and the classification coefficient is defined as 0.01 according to the frequency ratio;

[0056] The slope division formula is:

[0057] The slope aspect is flat, the classification is defined as 1, and the classification coefficient is defined as 0.01 according to the frequency ratio;

[0058] The slope direction is north, the classification is defined as 2, and the classification coefficient is defined as 0.77 / 0.99 according to the frequency ratio;

[0059] The slope direction is northeast, the classification is defined as 3, and the classification coefficient is defined as 0.44 according to the frequency ratio;

[0060] The slope direction is east, the classification is defined as 4, and the classification coefficient based on the frequency ratio is defined as 0.33;

[0061] The slope direction is southeast, the classification is defined as 5, and the classification coefficient is defined as 0.11 according to the frequency ratio;

[0062] The slope direction is south, the classification is defined as 6, and the classification coefficient is defined as 0.55 according to the frequency ratio;

[0063] The slope direction is southwest, the classification is defined as 7, and the classification coefficient is defined as 0.88 according to the frequency ratio;

[0064] The slope direction is west, the classification is defined as 8, and the classification coefficient is defined as 0.66 according to the frequency ratio;

[0065] The slope direction is northwest, the classification is defined as 9, and the classification coefficient is defined as 0.22 according to the frequency ratio;

[0066] The plane curvature division formula is:

[0067] 0≤ plane curvature≤15, the classification is defined as 1, and the classification coefficient according to the frequency ratio is defined as 0.75;

[0068] 15<plane curvature≤30, the classification is defined as 2, and the classification coefficient is defined as 0.99 according to the frequency ratio;

[0069] 30<plane curvature≤45, the classification is defined as 3, and the classification coefficient according to frequency ratio is defined as 0.50;

[0070] 45<plane curvature≤60, the classification is defined as 4, and the classification coefficient is defined as 0.01 according to the frequency ratio;

[0071] When the plane curvature is greater than 60, the classification is defined as 5, and the classification coefficient according to the frequency ratio is defined as 0.25;

[0072] The formula for dividing the section curvature is:

[0073] 0≤profile curvature≤10, the classification is defined as 1, and the classification coefficient is defined as 0.99 according to the frequency ratio;

[0074] 10<profile curvature≤20, the classification is defined as 2, and the classification coefficient according to frequency ratio is defined as 0.66;

[0075] 20<profile curvature≤30, the classification is defined as 3, and the classification coefficient according to frequency ratio is defined as 0.33;

[0076] When the profile curvature is >30, the classification is defined as 4, and the classification coefficient is defined as 0.01 according to the frequency ratio;

[0077] The distance to the road is divided into:

[0078] When the road distance is ≤100m, the classification is defined as 1, and the classification coefficient is defined as 0.99 according to the frequency ratio;

[0079] 100m<road distance≤200m, the classification is defined as 2, and the classification coefficient is defined as 0.33 according to the frequency ratio; 200m<road distance≤500m, the classification is defined as 3, and the classification coefficient is defined as 0.01 according to the frequency ratio; 500m<road distance≤1000m, the classification is defined as 4, and the classification coefficient is defined as 0.17 according to the frequency ratio; 1000m<road distance≤2000m, the classification is defined as 5, and the classification coefficient is defined as 0.83 according to the frequency ratio; 2000m<road distance≤5000m, the classification is defined as 6, and the classification coefficient is defined as 0.67 according to the frequency ratio; road distance>5000m, the classification is defined as 7, and the classification coefficient is defined as 0.50 according to the frequency ratio;

[0080] The distance to the river is divided into:

[0081] When the river distance is ≤100 m, the classification is defined as 1, and the classification coefficient according to the frequency ratio is defined as 0.67;

[0082] 100m<river distance≤200m, the classification is defined as 2, and the classification coefficient is defined as 0.01 according to the frequency ratio; 200m<river distance≤500m, the classification is defined as 3, and the classification coefficient is defined as 0.99 according to the frequency ratio; 500m<river distance≤1000m, the classification is defined as 4, and the classification coefficient is defined as 0.50 according to the frequency ratio;

[0083] 1000m<river distance≤2000m, the classification is defined as 5, and the classification coefficient is defined as 0.83 according to the frequency ratio;

[0084] 2000m<river distance≤5000m, the classification is defined as 6, and the classification coefficient is defined as 0.33 according to the frequency ratio; road distance>5000m, the classification is defined as 7, and the classification coefficient is defined as 0.17 according to the frequency ratio;

[0085] The lithology classification formula is:

[0086] There are five different types of lithology in this typical area, which are divided in order:

[0087] The classification of lithology 1 is defined as 1, and the classification factor according to frequency ratio is defined as 0.25;

[0088] The classification of lithology 2 is defined as 2, and the classification factor according to the frequency ratio is defined as 0.01;

[0089] The classification of lithology 3 is defined as 3, and the classification coefficient according to the frequency ratio is defined as 0.99;

[0090] The classification of lithology 4 is defined as 4, and the classification factor according to frequency ratio is defined as 0.50;

[0091] The classification of lithology 5 is defined as 5, and the classification factor according to frequency ratio is defined as 0.75;

[0092] The annual average rainfall classification formula is: 1051mm<rainfall≤1059mm, the classification is defined as 1, and the classification coefficient is defined as 0.99 according to the frequency ratio; 1059mm<rainfall≤1064mm, the classification is defined as 2, and the classification coefficient is defined as 0.01 according to the frequency ratio; 1064mm<rainfall≤1069mm, the classification is defined as 3, and the classification coefficient is defined as 0.66 according to the frequency ratio; 1069mm<rainfall≤1075.44mm, the classification is defined as 4, and the classification coefficient is defined as 0.33 according to the frequency ratio;

[0093] The elevation division formula is:

[0094] For elevation ≤400m, the classification is defined as 1, and the classification coefficient is defined as 0.75 according to the frequency ratio;

[0095] 400m<elevation≤700m, the classification is defined as 2, and the classification coefficient is defined as 0.99 according to the frequency ratio;

[0096] 700m<elevation≤1100m, the classification is defined as 3, and the classification coefficient is defined as 0.50 according to the frequency ratio;

[0097] 1100m<elevation≤1500m, the classification is defined as 4, and the classification coefficient is defined as 0.25 according to the frequency ratio;

[0098] 1500m<elevation≤2123m, the classification is defined as 5, and the classification coefficient is defined as 0.01 according to the frequency ratio;

[0099] The NDVI classification formula is:

[0100] 0<NDVI≤0.2, the classification is defined as 1, and the classification coefficient according to the frequency ratio is defined as 0.01;

[0101] 0.2<NDVI≤0.4, the classification is defined as 2, and the classification coefficient according to the frequency ratio is defined as 0.50;

[0102] 0.4<NDVI≤0.6, the classification is defined as 3, and the classification coefficient according to the frequency ratio is defined as 0.25;

[0103] 0.6<NDVI≤0.8, the classification is defined as 4, and the classification coefficient according to the frequency ratio is defined as 0.99;

[0104] 0.8<NDVI≤1, the classification is defined as 5, and the classification coefficient according to frequency ratio is defined as 0.75;

[0105] The land use classification formula is:

[0106] The classification of land use 1 is defined as 1, and the classification coefficient according to frequency ratio is defined as 0.99;

[0107] The classification of land use 2 is defined as 2, and the classification coefficient according to frequency ratio is defined as 0.75;

[0108] The classification of land use 3 is defined as 3, and the classification coefficient according to frequency ratio is defined as 0.50;

[0109] The classification of land use 4 is defined as 4, and the classification coefficient according to frequency ratio is defined as 0.25;

[0110] The classification of land use 5 is defined as 5, and the classification coefficient according to frequency ratio is defined as 0.01;

[0111] In Excel, the frequency ratio classification definition coefficients were used to replace the classification definition of each factor to construct the initial landslide susceptibility data set;

[0112] Step 2: Use the Stacking ensemble algorithm to build a typical regional landslide susceptibility evaluation model, use slope, slope aspect, plane curvature, profile curvature, distance from the road, distance from the river, lithology, annual average rainfall, elevation, NDVI, and land use to train the Stacking ensemble algorithm, and use the susceptibility index to verify the Stacking ensemble algorithm;

[0113] Among them, the susceptibility index is a value between 0 and 1, which is marked in advance in the database of landslide susceptibility evaluation factors.

[0114] In ArcGIS software, the “Random Point Selection” function was used to select non-landslide disaster points, with the ratio of the number to the number of landslide disaster points being 1:1. The “Extract Data to Points” function was used to extract the influencing factor data of non-landslide disaster points and landslide disaster points, and the data were exported to an Excel spreadsheet. The integrated deep learning model was used to train and verify the data.

[0115] A town in a certain county was selected as the research object. The number of landslides that occurred in the town up to 2015 was counted, totaling 95. According to the area of ​​landslide points and the average scale of landslides, the landslide sample grids were determined, totaling 4665.

[0116] A buffer zone of 500 m was set around the landslide site, and non-landslide disaster points equal in number to the landslide disaster points were randomly selected outside the buffer zone; the output value of the landslide sample was set to 1, and the output value of the non-landslide sample was set to 0.

[0117] The base learner uses the long short-term memory neural network algorithm (LSTM), recurrent neural network algorithm (RNN), convolutional neural network algorithm (CNN), and deep neural network algorithm (DNN); the secondary learner, that is, the two-layer algorithm, uses the multivariate linear regression algorithm (MLR), and Stacking integration is implemented in the python software based on the tensorflow environment.

[0118] All four models are directly called from the sklearn library in Python.

[0119] The preprocessed data set is divided into a fitting set and a test set in a ratio of 80% and 20%, and the training set is divided into five parts. Each base learner uses four-fold cross validation, and three parts of the data in the fitting set are used to train the base learner. The trained base learner is used to predict the validation set data and the test set data. This cycle is repeated four times to obtain four prediction results. Then, the prediction results of the test set are averaged to obtain the prediction results of each base learner.

[0120] Among them, the grid search algorithm is used to optimize the model hyperparameters. The specific search range is shown in Table 3;

[0121] Table 3. Deep learning model hyperparameter settings

[0122]

[0123] Use the Stacking ensemble algorithm in the ensemble learning algorithm to explore and evaluate the landslide susceptibility, and compare the performance of the Stacking ensemble algorithm with other single deep learning algorithms (LSTM, RNN, CNN, DNN) in landslide susceptibility evaluation modeling;

[0124] The stacking model was selected using the susceptibility index.

[0125] The evaluation indicators are used to evaluate and verify the Stacking model;

[0126] The AUC value is used to statistically analyze the prediction accuracy of the four deep learning models and the Stacking ensemble learning model. The results are shown in Table 4.

[0127] Table 4 Model accuracy calculation results

[0128]

[0129] Judging from the model calculation results, the accuracy of the Stacking algorithm (0.75) is higher than that of other single deep learning algorithms, so the Stacking model with the best evaluation ability is used for landslide hazard assessment.

[0130] Step 3: Use the Pearson III probability distribution to calculate the annual maximum daily rainfall for a given return period of 50 or 100 years, and construct a daily rainfall distribution map under different return periods;

[0131] Among them, due to the incompleteness of historical data, for example, there may be only thirty years of data in 50 years, but it is necessary to analyze the annual maximum daily rainfall in a given return period of 50 years.

[0132] Specifically include:

[0133] Step 31, based on the daily rainfall data of several years (a certain time period), extract the rainfall value of the largest day in a year, extract corresponding samples according to the time period, and sort the samples in descending order;

[0134] Step 32, using the maximum rainfall value of the day in the time period, calculate the three parameters of the Pearson III probability distribution;

[0135] The formula is as follows:

[0136]

[0137] Where, parameters α, β, a0 are the shape, scale and location parameters of Pearson III distribution, α>0, β>0, x is the maximum daily rainfall value, and f(x) is the return period corresponding to the maximum daily rainfall.

[0138]

[0139] Among them, C v is the variation parameter, σ is the standard deviation of rainfall in the time period, E(x) is the average rainfall in the time period; C s is the skewness coefficient, obtained from the Pearson curve;

[0140] Step 33: Input the probability of rainfall return period P r The annual maximum daily rainfall y for a given return period is calculated using the formula:

[0141] y=GAMMAINV(1-P r ,α,1 / β)+a0 (2)

[0142] Among them, P r It indicates the probability of a 50-year or 100-year rainfall return period.

[0143] The classification formula for the annual maximum daily rainfall y with a given return period of 50 years is: 111.56mm≤rainfall≤112.25mm, the classification is defined as 1, and the classification coefficient is defined as 0.01 according to the frequency ratio; 112.25mm<rainfall≤112.79mm, the classification is defined as 2, and the classification coefficient is defined as 0.66 according to the frequency ratio; 112.79mm<rainfall≤113.32mm, the classification is defined as 3, and the classification coefficient is defined as 0.99 according to the frequency ratio; 113.32mm<rainfall≤114.08mm, the classification is defined as 4, and the classification coefficient is defined as 0.33 according to the frequency ratio;

[0144] The classification formula for the annual maximum daily rainfall y with a given return period of 100 years is: 117.51mm≤rainfall≤118.25mm, the classification is defined as 1, and the classification coefficient is defined as 0.01 according to the frequency ratio; 118.25mm<rainfall≤118.75mm, the classification is defined as 2, and the classification coefficient is defined as 0.66 according to the frequency ratio; 118.75mm<rainfall≤119.25mm, the classification is defined as 3, and the classification coefficient is defined as 0.99 according to the frequency ratio; 119.25mm<rainfall≤120.03mm, the classification is defined as 4, and the classification coefficient is defined as 0.33 according to the frequency ratio;

[0145] Different from only obtaining the rainfall in the return period of the town, in this embodiment, formula (2) is first used to obtain the daily maximum rainfall at each station in the county and its surrounding six counties under different return periods. The daily maximum rainfall data of the seven stations and the "Kriging interpolation method" in ArcGIS are used to generate a raster map of the daily maximum rainfall in the county under the rainfall return period. The raster map of the daily maximum rainfall in the town under different rainfall return periods is clipped using the raster clipping function in ArcGIS. The reclassification function in ArcGIS is used to obtain a raster map of the daily maximum rainfall that can reflect the classification situation. The clipping result is as follows: Figure 2 and Figure 3 shown.

[0146] Existing studies often only use the rainfall data of the station closest to the study area to produce the daily maximum rainfall grid map of the study area, ignoring the rainfall conditions in the adjacent areas, which will make the results lack authenticity; in this embodiment, the daily maximum rainfall data of seven stations in the county and its surrounding counties are used to further ensure the reliability of the daily maximum rainfall grid map under different rainfall return periods.

[0147] Step 4: Reconstruct the landslide susceptibility assessment factor database based on the annual maximum daily rainfall data with a given return period, and complete the typical landslide hazard assessment by combining the Stacking ensemble algorithm trained in step 3; the output is the landslide hazard assessment index I sus ;

[0148] Specifically include:

[0149] Use "Extract Multi-Values ​​to Points" to extract the maximum daily rainfall under the rainfall return period corresponding to the grid point of Gongping Town, and export it to Excel;

[0150] The 10 factors other than the annual average rainfall factor are not changed. The present invention uses the annual maximum daily rainfall factor (hazard) of a given return period to replace the annual average rainfall factor (susceptibility), and obtains rainfall data with different return periods containing 11 factors to reconstruct the landslide susceptibility assessment factor database. The trained Stacking integrated algorithm is used to output the landslide hazard evaluation index I sus .

[0151] The 11 factors in the two new datasets include slope, aspect, plan curvature, profile curvature, distance to roads, distance to rivers, lithology, annual maximum daily rainfall with a given return period of 50 or 100 years, elevation, NDVI, and land use.

[0152] The two-layer algorithm training set obtained in the landslide susceptibility assessment model is used as the fitting set in the landslide hazard assessment model, and the landslide hazard assessment data set is used as the test set in the landslide hazard assessment model. The multivariate linear regression model is used for prediction, and the result is the landslide hazard assessment index.

[0153] The probability of a landslide occurring under a rainfall condition with a return period of 50 or 100 years is calculated as follows:

[0154] probability=P r *I sus (3)

[0155] Among them, probability represents the probability of landslide under the condition of rainfall with a return period of 50 or 100 years, P r represents the probability of a 50-year or 100-year rainfall return period, I sus Represents the landslide hazard assessment index.

[0156] In this example, the probability of a 50-year or 100-year rainfall return period is r The probability is 2% or 1%, and the landslide hazard evaluation index is multiplied by this probability to obtain the landslide hazard index under different return periods of rainfall. The results are input into ArcGIS, and the landslide hazard mapping is completed through the point-to-raster function.

[0157] Figure 2 and Figure 3 The following are the rainfall distribution maps for the 50-year return period and the 100-year return period, showing the rainfall division and Figure 4 and Figure 5 This is strongly correlated with the dangerous area division in the present invention, which further illustrates the rationality of the dangerous area division in the present invention.

[0158] The frequency ratio is used to evaluate the hazard partitioning results based on the Stacking ensemble deep learning algorithm. Figure 4 and Figure 5 The results show that under the rainfall condition of 50-year return period, 88% of the landslide hazard points are distributed in the extremely high, high and medium hazard areas of Stacking, and only 1% of the landslide hazard points fall in the extremely low susceptibility area; under the rainfall condition of 100-year return period, 83% of the landslide hazard points are distributed in the extremely high, high and medium hazard areas of Stacking, and only 1% of the landslide hazard points fall in the extremely low susceptibility area, indicating that there is a high consistency between the zoning results and the actual landslides.

[0159] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A landslide hazard assessment method under return period rainfall based on integrated deep learning, characterized in that: The following steps are involved: Step 1: Collect landslide susceptibility evaluation factor data, and perform preprocessing and frequency ratio classification coefficient division; Step 2: construct the Stacking ensemble algorithm, train the Stacking ensemble algorithm using the landslide susceptibility evaluation factor, and verify the Stacking ensemble algorithm using the susceptibility index; Step 3: Calculate the annual maximum daily rainfall with a given return period using the Pearson III probability distribution, and construct a daily rainfall distribution map with different return periods; Step 4: Replace the annual average rainfall in the landslide susceptibility evaluation factor with the annual maximum daily rainfall with a given return period, and use the trained Stacking ensemble algorithm for prediction to output the landslide hazard evaluation index.

2. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 1 is characterized in that: The base learner of the Stacking ensemble algorithm uses LSTM, RNN, CNN and DNN integration, and the secondary learner uses MLR. The grid search algorithm is used to optimize the hyperparameters of the number of neurons, batch size and number of iterations of the base learner.

3. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 1 is characterized in that: The calculation of the annual maximum daily rainfall for a given return period using the Pearson III probability distribution includes: Step 31, extracting the maximum rainfall value of each day in each year from the daily rainfall data of several years; Step 32, calculating the shape, scale and location parameters of the Pearson III probability distribution; Step 33: Calculate the annual maximum daily rainfall for a given return period using the probability of the rainfall return period. The formula is: y=GAMMAINV(1-P r ,a,1 / b)+a0 Among them, P r is the probability of rainfall return period, α, β, a0 are the shape, scale and position of Pearson III distribution respectively.

4. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 3 is characterized in that: Constructing the daily rainfall distribution map under different return periods includes: using the daily maximum rainfall under different return periods of each station in a county and adjacent counties, generating a daily maximum rainfall raster map of a county under the rainfall return period using the daily maximum rainfall and Kriging interpolation method; and using reclassification to obtain the daily maximum rainfall raster map.

5. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 3 is characterized in that: The formula for landslide hazard assessment index is: probability=P r *I sus (3) Among them, P r represents the probability of rainfall return period, I sus Represents the landslide hazard assessment index.

6. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 1 is characterized in that: Preprocessing includes: correlation analysis and multicollinearity test.

7. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 1 is characterized in that: Correlation analysis was performed using the Pearson correlation coefficient method.

8. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 1 is characterized in that: Multicollinearity test indicators include tolerance and variance inflation factor.

9. The landslide hazard assessment method under return period rainfall based on integrated deep learning according to claim 1 is characterized in that: The landslide susceptibility evaluation factors divided by frequency ratio classification coefficients include: slope, aspect, plan curvature, profile curvature, distance from roads, distance from rivers, lithology, average annual rainfall, elevation, NDVI and land use.

10. A landslide hazard assessment system based on integrated deep learning under return period rainfall, characterized in that: include: a memory for storing instructions executable by a processor; A processor, used to execute instructions to implement the landslide hazard assessment method under return period rainfall based on integrated deep learning as described in any one of claims 1-9.

Images