A regional landslide risk early warning method based on rainfall threshold

Abstract

This invention relates to a regional landslide hazard early warning method based on rainfall thresholds, belonging to the field of disaster early warning. The method includes the following steps: S1: Establishing a coupled landslide susceptibility prediction model based on information content and negative sample selection: Information Content-Random Forest (IV-RF) and Information Content-Backpropagation Neural Network (IV-BPNN); S2: Establishing a continuous probability rainfall threshold model based on multi-function fitting; S3: Providing regional landslide hazard early warning based on susceptibility and continuous probability rainfall thresholds. This invention couples IV with both RF and BPNN models to obtain two coupled models, and then optimizes the two models to obtain a more accurate landslide susceptibility value for the study area. Coupled with the landslide susceptibility value and the continuous probability value, a continuous probability landslide hazard value is obtained, reflecting the spatial distribution characteristics of rainfall and achieving a more accurate landslide hazard early warning classification.

Description

Technical Field

[0001] This invention belongs to the field of disaster early warning and relates to a method for early warning of regional landslide hazard based on rainfall threshold. Background Technology

[0002] Mountainous areas have complex geological and topographical conditions, leading to frequent geological disasters. Landslides, as one of the most common geological disasters, threaten people's lives, property, living environment, and production and daily life. To meet the needs of regional disaster prevention and mitigation and land resource planning, regional landslide hazard early warning is necessary. Regional landslide hazard is generally considered to be the coupled result of the spatial and temporal probabilities of landslide occurrence. Landslide susceptibility can represent spatial probability and is a prerequisite for landslide hazard early warning, determining its accuracy. Landslide hazard assessment not only considers internal factors but also the dynamic influence of external factors, such as rainfall and earthquakes.

[0003] In recent years, statistical analysis and machine learning models have yielded significant results in predicting regional landslide susceptibility. Statistical analysis models (such as information content models) are easy to operate, but they rely heavily on subjective experience and cannot reflect the complex nonlinear relationship between landslides and evaluation factors. Machine learning models (such as random forest models and BP neural network models) are blind when selecting non-sample landslides, and the accuracy and reliability of prediction cannot be guaranteed when using a single model for areas with complex geological and topographical conditions.

[0004] The landslide time probability calculated by existing traditional rainfall threshold models based on statistical analysis is discontinuous. It can only roughly and coarsely classify the time probability levels and cannot reflect the spatial distribution characteristics of rainfall. The accuracy and efficiency of traditional landslide hazard warning need to be improved.

[0005] Therefore, in order to fully leverage the role of statistical analysis models and machine learning models in landslide susceptibility prediction (spatial probability) and overcome the problem of discontinuous temporal probability calculated by traditional rainfall threshold models, spatial probability and temporal probability are coupled to achieve higher accuracy and efficiency in regional landslide hazard early warning. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a regional landslide hazard early warning method based on rainfall thresholds. First, a statistical analysis model and a machine learning model are coupled to improve the accuracy of regional landslide susceptibility prediction. Second, a continuous probability rainfall threshold model based on multi-function fitting is constructed to obtain continuous landslide temporal probability values. Finally, the spatial probability is multiplied by the temporal probability to obtain regional landslide hazard values ​​and classify them, thereby reflecting the spatial distribution characteristics of rainfall and improving the efficiency of landslide hazard early warning.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] (1) As Figure 1 As shown, this invention proposes a coupled landslide susceptibility prediction model based on information content and negative sample selection: Information Content-Random Forest (IV-RF) and Information Content-Backpropagation Neural Network (IV-BPNN). Based on the landslide-prone and low-risk zones initially delineated by the IV model, non-landslide points (0) are randomly selected and combined with landslide points (1) to form a "1-0 dataset." Next, the information content values ​​in the dataset are introduced as the input layers for RF and BPNN, establishing the coupled IV-RF and IV-BPNN models. The prediction accuracy of the single model and the coupled model is compared using the Receiver Operating Characteristic (ROC) curve and confusion matrix. Finally, the rationality of the susceptibility partitioning of different models is compared based on the FR model.

[0009] a. Initial prediction

[0010] The analysis area was divided and numbered using evaluation units (referring to grid units or slope units). Evaluation factors were selected by collecting data and information on topography, geological conditions, meteorology, hydrology, and human engineering activities in areas related to landslide susceptibility through data collection, remote sensing interpretation, and field surveys. Evaluation factors with strong independence and non-collinearity were selected based on expert experience or statistical analysis methods and assigned values. The various evaluation factors within the evaluation unit were used as independent variables for landslides, and the landslide susceptibility of the evaluation unit was used as the dependent variable. The information content method was used to assign values ​​to the evaluation factors (see Equation 1). Historical landslide information included: landslide spatial location, area, volume, and outline range, and a landslide sample database was established.

[0011]

[0012] In the formula, N n S represents the area of ​​landslides occurring in the nth grade of the evaluation factor; N0 represents the total area of ​​landslides occurring; S n Sn represents the area of ​​the nth grade; S0 represents the total area of ​​the study area; N n / N0 represents the percentage of the area where landslides occurred in the classification; S n / S0 represents the percentage of area classified by grade; IV represents the amount of information, reflecting the degree of contribution of the landslide.

[0013] The information values ​​assigned to all evaluation factors in the evaluation units within the analysis area are summed to obtain the regional landslide susceptibility value. The natural breakpoint method is then used for classification, and the initial regional landslide susceptibility prediction results are finally obtained.

[0014] b. Negative sample selection

[0015] On the prepared basic evaluation factor layer, a suitable dataset needs to be selected as the input to the landslide susceptibility prediction model. First, based on existing landslide point records and relevant vector data obtained from field surveys, landslide rasters (positive samples, labeled 1) are obtained using ArcGIS's spatial connectivity tool. Furthermore, since the chosen susceptibility prediction model is a supervised machine learning model, non-landslide rasters (negative samples) need to be selected. To avoid selecting truly non-landslide rasters, water systems and residential areas must be avoided. Secondly, based on the landslide susceptibility values ​​calculated and categorized by the information content model, non-landslide rasters are randomly selected in extremely low or low susceptibility areas, while ensuring that these non-landslide rasters are outside the landslide point buffer zone and labeled 0. Figure 2 As shown. For example, from 100 landslide boundary vector maps, 500 30m*30m landslide grids can be obtained using ArcGIS's spatial join tool, such as... Figure 3 As shown, the grid size is determined by both the scale and the layer precision. Then, following the negative sample selection method described above, 500 non-landslide grids are selected to form a '0-1' dataset containing 1000 data points. 70% of the samples in the dataset are randomly selected as the machine learning training sample set, and 30% are used as the test sample set.

[0016] c. Model coupling

[0017] To combine the advantages of information content models and machine learning models, the two models are coupled. The specific steps are as follows: the information content values ​​of the '0-1' dataset are used as the input layer of the random forest model and the BP neural network model for model training; finally, the landslide susceptibility values ​​predicted by the model are classified.

[0018] d. Comparative analysis

[0019] After training the random forest and backpropagation (BP) neural network models using sample data, it is necessary to quantify whether the training accuracy meets the prediction requirements. Generally, ROC curves (Receiving Operator Characteristic curves) and the area under the AUC curve are used to evaluate the training effect of supervised machine learning models. Furthermore, in the field of machine learning, especially for supervised machine learning, confusion matrices can also be used to evaluate model performance.

[0020] e. Optimal Model Selection

[0021] To discuss the predictive effectiveness of the model, the ratio of landslide area proportion to frequency was used to measure the predictive performance. Generally, the lower the landslide susceptibility level, the smaller the ratio of landslide area proportion to frequency should be; as the susceptibility level increases, the ratio of landslide area proportion to frequency should gradually increase.

[0022] (2) Figure 4 As shown, a continuous probability rainfall threshold model based on multi-function fitting is proposed.

[0023] The cumulative percentage of landslides P based on different prior effective rainfall (EE) and different durations of rainfall (D) EE and P D , will P EE With P D The time probability P is obtained by multiplying the cumulative normal distribution function and the power function. A P-EE-D fit is then performed, where the effective rainfall and the number of days of rainfall at a given point are obtained to calculate the time probability P of a landslide at that point. A certain number of landslide points not included in the model are randomly selected and substituted into the established model to verify its accuracy.

[0024] a. Rainfall data

[0025] 1) Effective rainfall in the early stage

[0026] The effective rainfall in the early stage, based on the cumulative rainfall, further considers water loss due to evaporation and runoff, meaning that the actual water content of the soil and rock infiltrated by rainfall is much lower than the cumulative recorded value.

[80] Rainfall infiltration increases the water content of the soil and rock, leading to higher pore water pressure and reduced shear strength, ultimately causing slope instability and landslides. Therefore, the effective rainfall in the preceding period better reflects the changes in soil and rock water content during rainfall. The expression for EE is as follows:

[0027]

[0028] Where R0 is the rainfall on the day of the landslide, k is the rainfall permeability coefficient, and R i This represents the rainfall on the i-th day before the landslide, where n is the number of days of rainfall considered. The rainfall permeability coefficient k is generally determined based on the conditions of the study area, with most studies using values ​​ranging from 0.72 to 0.86, and 0.84 being the most widely used.

[0029] 2) Number of days of continuous rainfall

[0030] The duration of rainfall is defined as the number of days from the start to the end of a rainfall event; it is generally considered that a rainfall event ends if the rainfall is less than 5 mm for two consecutive days.

[0031] bP EE and P D Rainfall probability

[0032] The cumulative percentage of landslides under different levels of prior effective rainfall (EE) is calculated (P). EE P was obtained by fitting the cumulative normal distribution function. EE-EE curve, as shown in the following formula.

[0033]

[0034] The cumulative percentage of landslides under different durations of rainfall (D) is calculated (P). D P was obtained by fitting a power function. D The -D curve is shown in the following formula.

[0035] P(D) = a + b × D c (4)

[0036] c. Continuous probability rainfall threshold model based on multi-function fitting

[0037] P EE With P D The time probability P is obtained by multiplying the two, and then P-EE-D fitting is performed using the product of the cumulative normal distribution function and the power function, as shown in the following formula.

[0038]

[0039] d. Accuracy verification of typical rainfall-induced landslide examples

[0040] To verify the accuracy of the continuous probability rainfall threshold model based on multi-function fitting, some landslide points that were not involved in the modeling were randomly selected and substituted into the established rainfall model. The higher the accuracy, the more accurate the model is.

[0041] (3) A regional landslide hazard early warning method based on susceptibility and continuous probability rainfall threshold

[0042] The continuous probability landslide hazard warning system classifies the continuous probability landslide hazard value into levels using either the natural discontinuity method or the equidistant division method. The continuous probability landslide hazard value is obtained by multiplying the landslide susceptibility value by the rainfall-induced continuous probability value fitted by a nonlinear equation, reflecting different combinations of landslide susceptibility and rainfall continuous probability values ​​(see equation). Compared to traditional landslide hazard warning systems, the continuous probability landslide hazard warning system is not a simple, qualitative matrix superposition, but rather a quantitative interval division result. After the landslide susceptibility is determined, the landslide hazard warning level increases with the increase of the rainfall continuous probability value, conforming to dynamic change characteristics. The equidistant division method is used to classify the continuous probability landslide hazard value into five levels, as shown in Table 1: extremely high hazard zone (1.0–0.8), high hazard zone (0.8–0.6), medium hazard zone (0.6–0.4), low hazard zone (0.4–0.2), and extremely low hazard zone (0.2–0.0).

[0043] H = P × S (6)

[0044] In the formula, H is the continuous probability landslide hazard value, P is the continuous probability value (time probability), and S is the landslide susceptibility value (spatial probability).

[0045] Table 1 Classification of Continuous Probability Landslide Hazard Values

[0046] Hazard classification Extremely high danger zone High-risk area Medium danger zone Low-risk area Extremely low danger zone

[0047] The beneficial effects of this invention are as follows: This invention couples the IV (Inductively Coupled Identifier) ​​with both RF (Radio Frequency) and BPNN (Background-Based Neural Network) models to obtain two coupled models, and then optimizes the selection between the two models to obtain a more accurate landslide susceptibility value for the study area. Furthermore, this invention uses nonlinear equations to fit rainfall parameters to obtain continuous probability values ​​to represent the temporal probability of landslides. Coupled with the landslide susceptibility value and the continuous probability value, a continuous probability landslide hazard value is obtained, thereby reflecting the spatial distribution characteristics of rainfall and achieving a more accurate landslide hazard early warning and classification.

[0048] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0049] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0050] Figure 1 Flowchart for research on regional landslide susceptibility prediction models;

[0051] Figure 2 A diagram illustrating the selection of negative samples;

[0052] Figure 3 For grid cells;

[0053] Figure 4 This is a schematic diagram of a continuous probability rainfall threshold model based on multi-function fitting.

[0054] Figure 5 For the study area;

[0055] Figure 6 This is a landslide susceptibility prediction map based on an information content model.

[0056] Figure 7 This is a regional landslide susceptibility prediction map based on BPNN and RF.

[0057] Figure 8 This is a regional landslide susceptibility prediction map based on IV-BPNN and IV-RF.

[0058] Figure 9 Comparison of ROC curves for the four models;

[0059] Figure 10 A comparison of the confusion matrices of the four models;

[0060] Figure 11 The frequency ratio of each model at different levels;

[0061] Figure 12 Marginal distribution and cumulative distribution;

[0062] Figure 13 The fitting is based on the cumulative frequency proportion of previous effective rainfall and the number of days of rainfall;

[0063] Figure 14 The continuous probability rainfall threshold is based on multi-function fitting;

[0064] Figure 15 This is to verify the accuracy of the continuous probability rainfall threshold based on multi-function fitting. Detailed Implementation

[0065] The following specific examples illustrate the implementation 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. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0066] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0067] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0068] The predicted area is Fuling District of Chongqing Municipality, geographically located in the central part of Chongqing, covering an area of ​​2942.4 km². 2 It is 74.4 km wide from east to west and 70.9 km long from north to south, spanning latitudes of 29°20′~30°00′ and longitudes of 106°57′~107°44′. Administratively, it borders Nanchuan District and Wulong District to the south, Fengdu County to the east, and Banan District to the west.

[0069] (1) Prediction of regional landslide susceptibility

[0070] a. Initial prediction

[0071] First, the study area was divided into 30m × 30m evaluation units, resulting in 3.269 million evaluation units. Data on the region's topography, geological conditions, meteorology, hydrology, and human engineering activities were systematically collected and analyzed. Based on remote sensing interpretation and field verification, 495 landslide samples were obtained. Figure 5 GIS software was used to compile the acquired field and indoor data, establishing a database of regional geological environmental conditions and landslide hazard distribution. Data source information is shown in Table 2. Evaluation factors were optimized using principal component analysis, correlation analysis, and multicollinearity analysis, selecting those with strong independence and noncollinearity, ultimately choosing 15 evaluation factors. Based on the relationship between landslide points and evaluation factors, an information content model was used to classify and assign values ​​to the evaluation factors; the classification and assignment results are shown in Table 3. The information content method was used to conduct an initial landslide susceptibility prediction for the study area. The susceptibility values ​​were classified using the natural breakpoint method; the initial prediction results are shown in Table 3. Figure 6 As shown.

[0072] Table 2 Data Source Information

[0073]

[0074] Table 3. Evaluation Factor Classification Results

[0075]

[0076]

[0077]

[0078]

[0079] b. Negative sample selection

[0080] Based on existing landslide point logging and relevant vector data obtained from field surveys, 8675 landslide rasters were obtained using ArcGIS spatial connectivity tools. Firstly, waterways and residential areas were avoided. Secondly, based on the landslide susceptibility values ​​calculated and categorized using the IV model, non-landslide rasters were randomly selected in extremely low or low susceptibility areas, while ensuring that these non-landslide rasters were located outside a 1000m buffer zone around the landslide points. This resulted in 8675 non-landslide rasters, which, along with the known 8675 landslide rasters, formed a landslide-non-landslide dataset. 70% of the landslide-non-landslide dataset was randomly selected as the training set, and the remaining 30% was used as the test set.

[0081] c. Model Prediction

[0082] 1) Single-model susceptibility prediction

[0083] Using BPNN and RF models for landslide susceptibility prediction requires training the model with a training set and validating it with a test set. Finally, all raster data from the study area are input into the trained model for prediction. This paper uses the built-in software packages in Matlab to construct the BPNN and RF models. The prediction results are as follows: Figure 7 As shown.

[0084] 2) Coupled model susceptibility prediction

[0085] The informative values ​​of the landslide-non-landslide dataset were used as input layers for both the RF and BPNN models for training. Finally, the predicted landslide susceptibility values ​​were categorized. This coupling approach improves the targeting and regularity of data processing, while also enhancing the accuracy of the RF and BPNN predictions. The prediction results of the coupled model are shown below. Figure 8 As shown.

[0086] 3) Comparative Analysis

[0087] like Figure 9As shown, the AUC values ​​of the BPNN, RF, IV-BPNN coupled model, and IV-RF coupled model are 0.87, 0.88, 0.92, and 0.96, respectively. The ROC curves of all four models are close to the upper left corner, indicating that the four models trained using 14 evaluation factors have high accuracy and can predict the spatial probability of landslides relatively accurately. Secondly, coupling BPNN and RF machine learning models with information content significantly improves the prediction accuracy of a single machine learning model. Compared to network models, the effect of coupling tree models with information content is more significant than that of coupling network models.

[0088] Depend on Figure 10 It can be seen that the prediction performance of the RF and BPNN models is around 85%, indicating that each model has good prediction performance, but the prediction performance of the RF model is slightly higher than that of the BPNN model; the prediction performance of the IV-RF and IV-BPNN models is around 95%, indicating that the information content model can significantly improve the prediction performance of the tree model and the network model, and the prediction performance of the IV-RF model is higher than that of the IV-BPNN model.

[0089] Depend on Figure 11 It can be seen that the landslide area in the extremely high susceptibility and high susceptibility zones predicted by the RF, BPNN, IV-RF, and IV-BPNN models accounts for 86.72%, 87.08%, 92.75%, and 89.14% of the total landslide area, respectively. Meanwhile, the area proportions of the extremely high susceptibility and high susceptibility zones are 15.89%, 31.52%, 14.17%, and 21.08%, respectively (generally, the extremely high and high susceptibility zones should be as small as possible, and the proportion of landslide area in these zones should be as large as possible). This indicates that the landslide susceptibility classification predicted by the IV-RF model is more consistent with the spatial distribution of landslide disasters in Fuling District and has higher prediction accuracy. Furthermore, it also shows that the information-rich model significantly improves and optimizes the accuracy of single network and tree models.

[0090] d. Optimal model selection

[0091] Based on a reasonable comparison of ROC, confusion matrix, and partitioning, the susceptibility prediction results of the IV-RF model were selected as the basis for subsequent risk warnings.

[0092] (2) Continuous probability rainfall threshold model based on multi-model coupling

[0093] a. Rainfall data

[0094] Based on the geographical location of Fuling District and the distribution of surrounding rain gauge stations, rainfall data from 11 rain gauge stations were selected, including: Fuling, Changshou, Fengdu, Fenglai, Yutaishan, Muhe, Shuanghekou, Damu, Hetu, Wulingshan, and Longtan, as shown in Table 4. The rainfall data were downloaded from the China Meteorological Center website (http: / / data.cma.cn / ). Considering the spatially uneven distribution of rainfall, the rainfall corresponding to all rainfall-type landslides in the study area within the number of rainfall durations (D) was interpolated using the Splines method in ArcGIS software. The interpolation results were overlaid with the landslide points, and the rainfall at the rainfall-type landslide points was extracted using MATLAB. The rainfall permeability coefficient of the previous effective rainfall (EE) was determined to be 0.8 through correlation analysis.

[0095] Table 4 Information on 11 Rainfall Gauge Stations

[0096] Station Number 57520 57522 57523 A8108 A7585 A8718 A8104 A7520 A8101 A7474 A7510 Station Name longevity Fuling Fengdu big wood Phoenix Comes Hetu Yutaishan Wuling Mountain Harmony Shuanghekou Longtan districts and counties longevity Fuling Fengdu Fuling Wulong Nanchuan Fuling Fuling Fuling Banan Fuling Longitude (°) 107.07 107.42 107.68 107.67 107.31 107.03 107.45 107.53 107.47 106.90 107.10 Latitude (°) 29.83 29.75 29.87 29.62 29.40 29.31 29.72 29.52 29.87 29.61 29.45 Elevation (m) 378 274 218 980 616 660 660 1210 180 220.7 705

[0097] Before constructing different rainfall threshold models, it is necessary to verify the landslide samples and remove samples with incomplete time records of landslide points or those caused by other external factors. Therefore, 418 rainfall-type landslide points from 1980 to 2015 were selected to construct a continuous probability rainfall threshold model. 23 rainfall-type landslide points from 2016 to 2020 were used to verify the model accuracy. In addition, four typical rainfall-type landslide events from 2020 to 2021 were selected to verify the risk warning.

[0098] bP EE and P D Rainfall probability

[0099] The effective rainfall (EE) and duration of rainfall (D) preceding 418 rainfall-induced landslides between 1980 and 2015 were statistically analyzed. The effective rainfall ranged from 10 mm to 130 mm in the study area, and the duration of rainfall was less than 8 days. Figure 12 The graph shows the distribution of D and EE values ​​(grey dots) corresponding to 418 rainfall-induced landslides in Fuling District on a linear coordinate system, as well as the marginal distribution of D and EE (histogram) and their cumulative frequency percentage (red curve). The graph shows that the cumulative frequency percentage of the number of days with rainfall (P...)... D The distribution conforms to the power function distribution characteristics, while the cumulative frequency ratio of previous effective rainfall (P) EE If P conforms to the characteristics of a cumulative normal distribution function, then P EE and P D Equations 3 and 4 can be used for fitting, and the fitting results are as follows: Figure 13 As shown. The fitted equations are shown in Equations 7 and 8. Based on the fitting results, the cumulative frequency proportion P EE and P D Fitted R2 The values ​​of 0.95 and 0.97 respectively indicate that the effective rainfall in the preceding period had a significant impact on P. EE The number of consecutive days of rainfall affects P D It has strong explanatory power, and the mean squared errors of 0.08 and 0.04 respectively indicate that the fitting results are very ideal.

[0100]

[0101] P(D) = 1.31 - 1.08 × D -0.6 (8)

[0102] c. Continuous probability rainfall threshold model based on multi-function fitting

[0103] The cumulative frequency percentage of days with continuous rainfall, P D The cumulative frequency percentage of previous effective rainfall (P) EE Multiplying these yields the time probability P. Since the fitted data conforms to the characteristics of a power function and a cumulative normal distribution function, a two-dimensional nonlinear fitting is performed using MATLAB to obtain the continuous probability values ​​of landslides in Fuling District under different combinations of previous effective rainfall and different durations of rainfall. The fitting results are as follows: Figure 14 As shown in Equation 9. The fitted R... 2 The result of 0.980 indicates that the effective rainfall in the early stage and the number of days of rainfall have a strong explanatory power for P, and the mean square error of 0.03 indicates that the fitting result is very ideal.

[0104]

[0105] d. Accuracy verification of typical rainfall-induced landslide examples

[0106] Substituting 23 rainfall-induced landslide sites from 2016 to 2020 into the above rainfall model, the temporal probabilities corresponding to the landslide sites were obtained, such as... Figure 15 As shown, there are 5 landslide points with a time probability in the range of 80%–100%, 7 landslide points with a time probability in the range of 60%–80%, 6 landslide points with a time probability in the range of 40%–60%, and 5 landslide points with a time probability in the range of 20%–40%.

[0107] (3) Regional landslide hazard early warning based on susceptibility and continuous probability rainfall threshold

[0108] The continuous probability value of landslides based on multi-function fitting is multiplied by the landslide susceptibility value to obtain the continuous probability landslide hazard value in Fuling District.

[0109] Four typical rainfall-induced landslide events from 2020 to 2021 were selected for hazard warning verification. The landslide hazard calculation results are shown in Table 5. This demonstrates that the method can predict the spatiotemporal distribution of landslides with high accuracy and better reflect the continuous probabilistic hazard value of landslides.

[0110] Table 5. Calculation of the probability of consecutive landslide hazard values ​​for four typical rainfall-induced landslide events.

[0111] Landslide Name time EE D Susceptibility value Continuous probability Continuous probability hazard value Risk level a. Liangshuipu landslide 2020 / 6 / 29 115.2 6 0.898 0.936 0.840 Extremely high b. Decline in rankings 2018 / 8 / 01 96.5 7 0.875 0.930 0.814 Extremely high c. Tongziyuan landslide 2020 / 6 / 27 104.1 6 0.966 0.919 0.888 Extremely high d. Wangye Temple landslide 2016 / 6 / 02 103.2 3 0.891 0.733 0.653 high

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early warning of regional landslide hazard based on rainfall threshold, characterized in that: The method includes the following steps: S1: Establish a coupled landslide susceptibility prediction model based on information content and negative sample selection: Information Content-Random Forest IV-RF and Information Content-BP Neural Network IV-BPNN. S2: Establish a continuous probability rainfall threshold model based on multi-function fitting; S3: Regional landslide hazard warning based on susceptibility and continuous probability rainfall threshold; S1 specifically involves establishing a coupled landslide susceptibility prediction model based on information content and negative sample selection: Information Content-Random Forest IV-RF and Information Content-Backpropagation Neural Network IV-BPNN. This includes: based on the initially defined extremely low and low susceptibility zones of the IV model, randomly selecting non-landslide point 0, and combining it with landslide point 1 to form a "1-0 dataset"; secondly, introducing the information content values ​​from the dataset as the input layer for RF and BPNN to establish the IV-RF and IV-BPNN coupled models; comparing the prediction accuracy of the single model and the coupled model using the Receiver Operational Characteristic (ROC) curve and confusion matrix; and then comparing the rationality of the susceptibility partitioning of different models based on the FR model. S11: Initial Prediction The analysis area was divided and numbered using evaluation units. Evaluation factors were collected through data collection, remote sensing interpretation, and field surveys, gathering data and information on topography, geological conditions, meteorology, hydrology, and human engineering activities related to landslide susceptibility. Based on expert experience or statistical analysis methods, highly independent and non-collinear evaluation factors were selected and assigned values. Various evaluation factors within each evaluation unit were used as independent variables for landslides, and the landslide susceptibility of the evaluation unit was used as the dependent variable. The information content method was used to assign values ​​to the evaluation factors, as shown in Equation 1). Historical landslide information included: landslide spatial location, area, volume, and outline. A landslide sample database was established. In the formula, N n S represents the area of ​​landslides occurring in the nth grade of the evaluation factor; N0 represents the total area of ​​landslides occurring; S n Sn represents the area of ​​the nth grade; S0 represents the total area of ​​the study area; N n / N0 represents the percentage of the area where landslides occurred in the classification; S n / S0 represents the percentage of area classified by grade; IV represents the amount of information, reflecting the degree of contribution of the landslide. The information values ​​assigned to all evaluation factors in the evaluation units within the analysis area are summed to obtain the regional landslide susceptibility value. The natural breakpoint method is then used for classification to obtain the initial regional landslide susceptibility prediction results. S12: Negative Sample Selection On the prepared basic evaluation factor layer, a suitable dataset needs to be selected as the input to the landslide susceptibility prediction model. First, based on the existing landslide point cataloging and relevant vector data obtained from field surveys, landslide rasters (positive samples) are obtained using ArcGIS's spatial connectivity tool, labeled as 1. Non-landslide rasters (negative samples) are then selected. To avoid selecting non-landslide rasters that are actually non-landslide rasters, water systems and residential areas are avoided. Second, based on the landslide susceptibility values ​​calculated and graded by the information content model, non-landslide rasters are randomly selected in extremely low or low susceptibility areas, while ensuring that the non-landslide rasters are outside the landslide point buffer zone, labeled as 0. The raster size is determined by both the scale and the layer precision. Then, according to the above negative sample selection method, 500 non-landslide rasters are selected to form a '0-1' dataset containing 1000 data points. 70% of the samples in the dataset are randomly selected as the machine learning training sample set, and 30% of the samples are used as the test sample set. S13: Model Coupling The two models are coupled, and the information values ​​of the '0-1' dataset are used as the input layer of the random forest model and the BP neural network model for model training; finally, the landslide susceptibility values ​​predicted by the model are classified. S14: Comparative Analysis After training the random forest model and the BP neural network model using sample data, we quantify whether the accuracy of the trained model can meet the prediction requirements; we use ROC curves and the area under the AUC curve to evaluate the training effect of the supervised machine learning model. S15: Optimal Model Selection To discuss the predictive performance of the model, the ratio of landslide area proportion to frequency was used to measure the predictive performance. For the ratio of landslide area proportion to frequency, the lower the landslide susceptibility level, the smaller the ratio of landslide area proportion to frequency, and the higher the susceptibility level, the larger the ratio of landslide area proportion to frequency.

2. The method for early warning of regional landslide hazard based on rainfall threshold according to claim 1, characterized in that: Specifically, S2 refers to the cumulative percentage of landslides P under different prior effective rainfall amounts EE and different rainfall durations D. EE and P D , will P EE With P D The time probability P is obtained by multiplying the cumulative normal distribution function and the power function. P-EE-D fitting is performed in the form of the product of the cumulative normal distribution function and the power function. When the effective rainfall in the previous period and the number of days of rainfall at a certain point are obtained, the time probability P of landslide at that point is calculated. A certain number of landslide points that were not involved in the modeling are randomly selected and substituted into the established model to verify the accuracy of the model. S21: Rainfall Data 1) Effective rainfall in the early stage The expression for the effective rainfall EE in the preceding period is as follows: Where R0 is the rainfall on the day of the landslide, k is the rainfall permeability coefficient, and R i This represents the rainfall on the i-th day before the landslide, where n is the number of days the rainfall is considered; the rainfall permeability coefficient k ranges from 0.72 to 0.

86. 2) Number of days of continuous rainfall The duration of rainfall is defined as the number of days from the start to the end of a rainfall event; if the rainfall is less than 5 mm for two consecutive days, then a rainfall event ends. S22:P EE and P D Rainfall probability The cumulative percentage of landslides (P) under different levels of prior effective rainfall (EE) is calculated. EE P was obtained by fitting the cumulative normal distribution function. EE -EE curve, as shown in the following formula; The cumulative percentage of landslides under different durations of rainfall (D) is calculated. D P was obtained by fitting a power function. D -D curve: P(D)=a+b×D c (4) S23: Continuous Probabilistic Rainfall Threshold Model Based on Multifunction Fitting P EE With P D The time probability P is obtained by multiplying the products, and then P-EE-D fitting is performed using the product of the cumulative normal distribution function and the power function, as shown in the following formula: S24: Accuracy Verification of Typical Rainfall-Induced Landslide Examples To verify the accuracy of the continuous probability rainfall threshold model based on multi-function fitting, some landslide points that were not involved in the modeling were randomly selected and substituted into the established rainfall model. The higher the accuracy, the more accurate the model is.

3. The method for early warning of regional landslide hazard based on rainfall threshold according to claim 2, characterized in that: Specifically, S3 refers to: the continuous probability landslide hazard warning method is to classify the continuous probability landslide hazard value into levels using the natural breakpoint method or the equal interval division method; The continuous probability landslide hazard value is obtained by multiplying the landslide susceptibility value with the continuous probability value of rainfall-type landslides fitted by the nonlinear equation. It can reflect different combinations of landslide susceptibility value and continuous probability value of rainfall, as shown in equation (6). The continuous probability landslide hazard value is divided into 5 levels using the equal interval division method, including extremely high risk area (H = 1.0 to 0.8), high risk area (H = 0.8 to 0.6), medium risk area (H = 0.6 to 0.4), low risk area (H = 0.4 to 0.2), and extremely low risk area (H = 0.2 to 0.0). H = P × S (6) In the formula, H is the continuous probability landslide hazard value, P is the continuous time probability value, and S is the landslide susceptibility value.

Images