A landslide displacement prediction method, system, device and storage medium

Abstract

This invention discloses a landslide displacement prediction method, system, device, and storage medium, relating to the field of geological disaster monitoring and early warning technology. The method includes the following steps: collecting current rainfall events in the study area; extracting features from the current rainfall events to obtain multiple rainfall physical features that induce landslide displacement behavior. These rainfall physical features include a first feature reflecting the slope infiltration degree, a second feature reflecting the rainfall time series, and a third feature reflecting historical rainfall and landslide displacement data in the study area; inputting these multiple rainfall physical features into a pre-trained regression model, and outputting the landslide displacement caused by the current rainfall event. This invention does not rely on the periodicity and trend of the data, while simultaneously exploring the complex correlation between landslide displacement induction features and landslide displacement behavior. The regression model can learn multiple rainfall physical features simultaneously, thereby extracting more information from limited data and maintaining high prediction accuracy and interpretability.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and in particular to a landslide displacement prediction method, system, equipment and storage medium. Background Technology

[0002] Landslides, a typical type of geological hazard in mountainous areas, are numerous, widely distributed, and highly destructive, posing a serious threat to transportation infrastructure, engineering structures, and the safety of people's lives and property. The formation and evolution of landslides are accompanied by a wealth of observable information; effectively monitoring and identifying this information is a crucial measure for mitigating disaster risks. Among these, landslide displacement, as a direct dynamic parameter characterizing the instability process of rock and soil masses, is the most critical quantitative indicator for early landslide prediction, and its changing patterns have significant indicative significance for precursory disasters.

[0003] Landslide displacement prediction based on geological identification is the most universally applicable technique. It captures landslide hazards by analyzing landslide geomorphological features, slope structure, and geological environmental conditions, thus achieving effective landslide identification. However, this method is highly dependent on expert experience and often fails to meet the timeliness requirements of landslide prediction. In recent years, with the rapid development of machine learning and deep learning technologies, researchers have treated landslide displacement as time-series data, using models to mine its potential time dependencies and predict the displacement. Some scholars have further used methods such as variational mode decomposition (VMD) to decompose landslide displacement to extract trend and periodic terms. Regarding model selection, recurrent neural networks such as LSTM and GRU are widely used due to their strong ability to capture time dependencies.

[0004] Existing machine learning methods treat landslide displacement as a time-driven event, which has the advantage of using mature time series prediction models. However, short-term landslide displacement sequences often fail to exhibit significant periodicity or trend patterns, limiting the effectiveness of time series prediction models in capturing sequence structure information. Furthermore, time series prediction models tend to rely on the autocorrelation structure of the data itself, exhibiting limited ability to understand the causal and physical relationships between input variables, thus failing to learn deeper information and significantly reducing the accuracy of the final landslide displacement prediction. Summary of the Invention

[0005] Based on the shortcomings of the existing technology, the present invention provides a landslide displacement prediction method, system, device and storage medium, which solves the problem that existing short-term landslide displacement sequences often fail to show significant periodic or trend patterns, limiting the effective capture of sequence structure information by time series prediction models. At the same time, time series prediction models tend to rely on the autocorrelation structure of the data itself, have limited ability to understand the causal and physical relationships between input variables, and cannot learn deeper information, resulting in a significant reduction in the accuracy of the final landslide displacement prediction.

[0006] The present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a landslide displacement prediction method, comprising the following steps:

[0008] Collect current rainfall events in the study area; wherein, the rainfall event is the rainfall data of the study area over multiple consecutive days, and the interval without rainfall between two adjacent rainfall events is greater than a set threshold;

[0009] Feature extraction is performed on the current rainfall event to obtain multiple rainfall physical features that induce landslide displacement behavior. These rainfall physical features include a first feature reflecting the degree of slope infiltration, a second feature reflecting the rainfall time series, and a third feature reflecting historical rainfall and landslide displacement data in the study area.

[0010] Multiple rainfall physical features are input into a pre-trained regression model, which outputs the landslide displacement caused by the current rainfall event.

[0011] Preferably, the first feature includes the total rainfall of a single rainfall event, the duration of rainfall, the maximum rainfall intensity within the event, and the effective rainfall; the second feature includes a rainfall event index; and the third feature includes the average rainfall over the 5 days prior to the event, the average rainfall over the 15 days prior to the event, the average rainfall over the 30 days prior to the event, the slip over the 3 days prior to the event, the slip over the 15 days prior to the event, and the slip over the 30 days prior to the event.

[0012] The total rainfall of a single rainfall event is the sum of the cumulative rainfall from the start to the end of the event;

[0013] The duration of rainfall is the continuous rainfall time from the start to the end of the event;

[0014] The maximum rainfall intensity during the event is the maximum hourly rainfall intensity during the event.

[0015] The effective rainfall amount is the daily rainfall P over the previous N days, calculated using an exponential decay coefficient. Weighted summation;

[0016] The rainfall event index is a number that sorts each independent rainfall event by time within the same monitoring period;

[0017] The slippage in the three days prior to the event refers to the cumulative displacement over the three days preceding the event.

[0018] The slippage in the 15 days prior to the event refers to the cumulative displacement over the 15 days preceding the event.

[0019] The slippage in the 30 days prior to the event refers to the cumulative displacement over the 30 days preceding the event.

[0020] Preferably, the pre-training of the regression model includes the following steps:

[0021] Historical rainfall data and corresponding historical slip volume were obtained for the study area. The historical rainfall data were divided according to the intervals between periods of no rainfall to obtain multiple historical rainfall events. The intervals between periods must be greater than a set threshold.

[0022] For each historical rainfall event, feature extraction is performed to obtain multiple corresponding rainfall physical features;

[0023] For each historical event, the corresponding multiple rainfall physical characteristics are used as inputs and the corresponding slip is used as outputs to pre-train the regression model.

[0024] Preferably, the set threshold is determined by local climate conditions.

[0025] Preferably, the regression model is the XGBOOST model.

[0026] Secondly, the present invention provides a landslide displacement prediction system, comprising:

[0027] The data acquisition module is used to collect current rainfall events in the study area; wherein, the rainfall event is the rainfall data of the study area over multiple consecutive days, and the interval between two adjacent rainfall events without rainfall is greater than a set threshold.

[0028] The extraction module is used to extract features from the current rainfall event to obtain multiple rainfall physical features that induce landslide displacement behavior. The rainfall physical features include a first feature reflecting the degree of slope infiltration, a second feature reflecting the rainfall time series, and a third feature reflecting the historical rainfall and landslide displacement data of the study area.

[0029] The prediction module is used to input multiple rainfall physical features into a pre-trained regression model and output the landslide displacement caused by the current rainfall event.

[0030] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described landslide displacement prediction method.

[0031] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described landslide displacement prediction method.

[0032] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0033] This invention first collects current rainfall events in the study area. Rainfall events are defined as continuous rainfall data over multiple days, with the interval between two adjacent rainfall events exceeding a set threshold. This invention reconstructs landslide displacement prediction from a time-driven autoregressive model to a rainfall event-driven regression model. It does not rely on the periodicity or trend of the data, but rather on the mining and establishment of multiple rainfall physical features that induce landslide displacement behavior. Specifically, these include a first feature reflecting the degree of slope infiltration, a second feature reflecting the rainfall time series, and a third feature reflecting historical rainfall and landslide displacement data in the study area. This invention uncovers the complex correlation between landslide displacement induction features and landslide displacement behavior, increasing the model's ability to understand the causal and physical relationships between input variables. Finally, multiple rainfall physical features are input into a pre-trained regression model, outputting the landslide displacement caused by the current rainfall event. The regression model can learn multiple rainfall physical features simultaneously, thereby extracting more information from limited data and maintaining high prediction accuracy and interpretability. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of a landslide displacement prediction method according to the present invention;

[0036] Figure 2 This is a schematic diagram of rainfall data at a monitoring point in the study area of ​​this invention;

[0037] Figure 3 This is a schematic diagram of the cumulative slip data at a monitoring point in the study area of ​​this invention;

[0038] Figure 4 This is a schematic diagram of the daily slip data at a monitoring point in the study area of ​​this invention;

[0039] Figure 5 This is a schematic diagram of rainfall events after the historical data of a monitoring point in the study area of ​​this invention has been divided.

[0040] Figure 6 This is a schematic diagram of the feature extraction process of the present invention;

[0041] Figure 7 This is a schematic diagram of the importance of SHAP features for test samples at a monitoring point in the study area of ​​this invention.

[0042] Figure 8 This is a schematic diagram of the beeswarm feature contribution of a test sample at a monitoring point in the study area of ​​this invention.

[0043] Figure 9 A plan view showing the deployment of monitoring for typical landslides in the study area;

[0044] Figure 10 A schematic diagram of LSTM prediction results for a test set collected from typical landslides in the study area;

[0045] Figure 11 A schematic diagram showing the prediction results of landslide displacement prediction methods on a test set collected for typical landslides in the study area;

[0046] Figure 12 A standardized schematic diagram of the prediction results of the test set collected for typical landslides in the study area using the landslide displacement prediction method;

[0047] Figure 13 Contribution rate diagram of different features of test samples of typical landslides in the study area;

[0048] Figure 14 This is a graph showing the SHAP values ​​of test samples from typical landslides in the study area. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] To address the shortcomings of existing landslide displacement prediction methods, such as strong dependence on time series data and weak interpretability, this embodiment proposes a landslide displacement prediction method, specifically a landslide displacement prediction method based on the XGBOOST machine learning model, used to predict landslide displacement. In this method, an event-driven, regression model prediction framework is proposed. Through event-driven mechanisms, the learned data becomes more physically meaningful. This method can predict the displacement at a specified future time step. Feature engineering assists the model in extracting physical information from the data for prediction, resulting in high interpretability and prediction accuracy. The method includes a model pre-training process, specifically comprising the following steps:

[0052] S1: Data Acquisition.

[0053] First, it is necessary to obtain historical slip volume and historical rainfall data for a monitoring point in the study area, and ensure that the timestamps of the two are aligned, such as... Figures 2-4 As shown. Data is retrieved at least once a day.

[0054] S2: Rainfall event classification and data processing.

[0055] Historical rainfall data is divided into N independent rainfall events based on a threshold for the interval between rainfall events. The interval between two adjacent rainfall events must be greater than a set threshold. In this embodiment, the set threshold is determined by local climate conditions. For example, in the Northwest region, "48 consecutive hours without rain" is a suitable threshold for dividing rainfall intervals.

[0056] Rainfall events can last for one day or multiple days. Furthermore, to ensure real-time forecasting, a rainfall event is defined as "the rainfall process from the start of the event to the forecast date," even if the event has not yet ended, it is considered a complete input sequence to drive the model for displacement prediction in advance. Threshold settings are determined by local climate conditions; for example, in the Northwest region, "48 consecutive hours without rain" is a suitable threshold for defining rainfall intervals.

[0057] After segmentation, rainfall data with a value of 0.2 mm were discarded (0.2 mm represents the sensor's error value); that is, rainfall > 0.2 mm was considered a valid rainfall event. Figure 5 The result is based on the division of rainfall time.

[0058] S3: Feature extraction of rainfall events.

[0059] This method introduces multiple rainfall physical features corresponding to landslide displacement to improve the model's expressive power. If historical rainfall data is directly input into a machine learning model, the model struggles to capture deep-seated physical information; therefore, event segmentation and feature extraction of the rainfall data are necessary. The features extracted by this method are content that traditional time series prediction models cannot automatically acquire. By employing a regression model, this method allows for the manual design and introduction of more features closely related to landslide displacement, thereby enhancing the interpretability of the prediction results. In simpler terms, it involves manually uncovering potential information to assist the model in learning and improving prediction performance.

[0060] Reference Figure 6 For each rainfall event, features were extracted. The extracted rainfall physical features include a first feature reflecting the slope infiltration degree, a second feature reflecting the rainfall time series, and a third feature reflecting historical rainfall and landslide displacement data in the study area. The first feature includes the total rainfall (Rainfall), rainfall duration (Rain Duration), maximum rainfall intensity (Max Rain Intensity Event), and effective rainfall (Antecedent Effective Rainfall, AER) of a single rainfall event. The second feature includes the rainfall event index (Event ID). The third feature includes the average rainfall over the 5 days before the event (Rain 5d mean), the average rainfall over the 15 days before the event (Rain 15d mean), the average rainfall over the 30 days before the event (Rain 30d mean), the landslide displacement over the 3 days before the event (Disp 3d), the landslide displacement over the 15 days before the event (Disp 15d), and the landslide displacement over the 30 days before the event (Disp 30d).

[0061] The total rainfall of a single rainfall event is the sum of the cumulative rainfall from the start to the end of the event, usually obtained by integrating rain gauge data. The total rainfall of a single rainfall event reflects the overall rainfall situation on the slope, and the unit is mm.

[0062] Rainfall duration refers to the continuous rainfall time from the start to the end of an event. Rainfall duration controls the infiltration depth and degree of the slope. For the same amount of total rainfall, if it is distributed over a longer period, the slope will infiltrate more fully, the potential infiltration depth will be greater, and the likelihood of slope displacement will be higher. The unit is h or d.

[0063] A rainfall event index is defined as a time-ordered number assigned to each independent rainfall event within the same monitoring period. Physically, it represents coarse-grained "time sequence" information, containing a certain sequential relationship between rainfall events.

[0064] Effective rainfall in the preceding period was calculated by applying an exponential decay coefficient to the daily rainfall P over the previous N days. Weighted summation. The effective rainfall in the preceding period (AER) approximates the slope's "soil moisture content." It reflects the lagged response of rainfall to the soil moisture content. A high AER indicates a higher soil moisture content, making subsequent rainfall more likely to trigger slope displacement and instability. The unit is mm. The formula is as follows.

[0065] ;

[0066] In the formula, E To effectively accumulate rainfall, The first before the landslide occurred i Daily rainfall of the day n The number of days before the landslide (1≤ i ≤ n ), The coefficient is the exponential decay coefficient, which in this embodiment is the evaporation decay coefficient, and is taken as 0.84.

[0067] The average rainfall over the 5 days prior to the event is the average of the daily rainfall over the 5 days preceding the event. This characteristic represents short-term cumulative rainfall and reflects recent precipitation conditions. The unit is mm / d.

[0068] The average rainfall over the 15 days prior to the event is the average of the daily rainfall over the 15 days preceding the event. This characteristic represents the mesoscale cumulative rainfall, reflecting the mid-term rainfall situation. The unit is mm / d.

[0069] The average rainfall over the 30 days prior to the event is the average of the daily rainfall over the 30 days preceding the event. This characteristic represents long-term cumulative rainfall, reflecting medium- to long-term precipitation patterns. The unit is mm / d.

[0070] The landslide volume in the three days prior to the event represents the cumulative displacement of the landslide monitoring points over the three days preceding the event. This characteristic indicates recent landslide volume and reflects the recent slope slippage. The unit is mm.

[0071] The landslide volume in the 15 days prior to the event refers to the cumulative displacement of the landslide monitoring points over the 15 days before the event. This characteristic is the mid-term landslide volume, reflecting the mid-term slope landslide situation. The unit is mm.

[0072] The landslide volume in the 30 days prior to the event represents the cumulative displacement of the landslide monitoring points over the 30 days preceding the event. This characteristic is the long-term landslide volume, reflecting the medium- to long-term slope slippage. The unit is mm.

[0073] The maximum rainfall intensity within the event is the maximum hourly rainfall intensity within the event window. The peak rainfall within the event indicates that the peak rainfall will lead to slope runoff and localized scouring, resulting in relatively weak infiltration capacity and often triggering shallow displacement. The unit is mm / h.

[0074] S4: Dividing the slide window.

[0075] Machine learning regression models require input features X and prediction labels Y to train various parameters and weights of the model. The extracted rainfall physical features are represented as X, and Y as the slip. The slip value is selected as the slip value N days after the first rainfall (including the day of the first rainfall). For example, if the slip value for the next 3 days of the current rainfall event is selected for training, then the model learns the non-linear relationship between the features of the rainfall event and the slip value over the next 3 days, meaning the predicted result is the slip value over the next 3 days. The slip value window should not be too small (because a slip process is not yet fully completed) nor too large (it may cover the next slip event). The window length is determined by engineering requirements and should balance response integrity and event independence; 3 days, 5 days, and 7 days are recommended choices.

[0076] The slip is differentially calculated to determine the daily slip increment.

[0077] S5: Construct the input dataset.

[0078] Based on the multiple rainfall physical features from step S3 and the slip window data from step S4, alignment is performed in the table file to ensure that the features corresponding to each rainfall event match the slip window data, thus obtaining the dataset. This dataset can then be successfully input into the model and recognized in the next step. Overall, the model input data includes X and Y, where X represents the input features and Y represents the label.

[0079] S6: XGBOOST model construction and training.

[0080] XGBOOST is widely used and relatively mature, so it was chosen as the regression model to be used.

[0081] (1) Divide the dataset into training and test sets in a ratio of 8:2 or 7:3. Choose the former when the data sample size is small and choose the latter when the data sample size is sufficient.

[0082] (2) Landslide data has a long tail effect (uneven data distribution), and rainfall data is also unevenly distributed throughout the year. Logarithmic standardization is used to standardize the data when inputting the data (log1p function). log1p will not have a negative impact on the prediction results.

[0083] (3) The XGBOOST parameter uses XGBRegressor(objective="reg:squarederror") and is optimized by RandomizedSearchCV.

[0084] To ensure prediction accuracy, key control parameters (number of trees, tree depth, and learning rate, etc.) of the XGBoost algorithm are selected, as shown in Table 1.

[0085] Table 1 Key control parameters of the XGBoost algorithm

[0086]

[0087] The search iterations are set to 30, and the cross-validation scheme is 5-fold cross-validation. The evaluation metric is negative RMSE (Neg Root Mean Squared Error), which is equivalent to minimizing the RMSE. After training, the optimal hyperparameter combination and the corresponding fitted model are automatically returned.

[0088] (4) Performance evaluation: The generalization ability of the model is quantified by calculating the coefficient of determination R² and RMSE.

[0089] (5) Interpretability analysis: Use shap.Explainer to build a tree model interpreter for the final model and the entire feature set.

[0090] Calculate the SHAP (SHapley Additive exPlanations) value for each test sample to obtain the marginal contribution of the feature to displacement prediction. Plot Beeswarm and Bar graphs: Beeswarm shows the distribution and direction of single-sample SHAP; the bar graph displays the global mean SHAP, giving the top 15 most important features, such as... Figure 7 and Figure 8 As shown. In short, bar charts show the importance of features, while SHAP charts show the contribution of each quantity to the result.

[0091] (6) Model saving and reuse.

[0092] Save the best XGBoost model as a *.pkl file for later deployment or reloading of predictions.

[0093] The innovation of this method lies in: reconstructing landslide displacement prediction into a "rainfall event-driven + regression modeling" task, overcoming the limitations of traditional time series methods; achieving high interpretability and nonlinear expression by inputting structured rainfall event features closely related to landslide displacement into XGBoost regression; and using methods such as SHAP for feature contribution analysis, providing technical assurance for model transparency.

[0094] Reference Figure 1 The landslide displacement prediction method based on the XGBOOST machine learning model proposed in this invention includes the following steps:

[0095] Step 1: Collect current rainfall events in the study area.

[0096] The second step is to extract features from the current rainfall event to obtain multiple rainfall physical features that induce landslide displacement behavior.

[0097] Step 3: Input multiple rainfall physical features into the pre-trained regression model and output the landslide displacement caused by the current rainfall event.

[0098] Example 2

[0099] This embodiment focuses on rainfall-induced landslide displacement prediction modeling, and the Qinling Mountains provide the typical geological background required for this study. Its terrain is dramatically undulating, with well-developed fault structures. Long-term intense weathering and gravity deposition have resulted in a widespread overburden of loose, gravelly, silty clay on fractured bedrock, forming numerous "depositional layer-bedrock" sliding surfaces. Furthermore, the region is influenced by the summer southeast monsoon, resulting in concentrated and intense annual rainfall, which easily creates low-shear surfaces at the bedrock-depositional layer interface or within the depositional layer, triggering numerous landslides. The combination of these multiple factors makes the Qinling Mountains one of the most densely populated areas of depositional landslides in my country, providing an ideal natural experimental field for studying the "rainfall-slip" mechanism.

[0100] Against this backdrop, this embodiment selects a typical rainfall-induced landslide located in the Qinling-Bashan Mountains as the research object. For example... Figure 9 As shown, the landslide is located in a low-to-medium mountain valley, which is V-shaped and runs approximately east-west. The Han River flows through the valley. The elevation difference on the northern slope is greater than 200 meters, with a slope aspect of 180°-260°. The slope aspect of this particular slope is 190°, with an angle of 85° with the strata, making it a sloping slope. The slope is generally higher in the north and lower in the south, with a straight slope shape. Terraces, consisting of short steps, are located on the slope.

[0101] Rapid uplift and intense river downcutting in the region resulted in deposits several to tens of meters thick on both sides of the canyon where the landslide occurred, making it highly susceptible to slip surface development along the basement interface after rainfall infiltration. Furthermore, the region receives an average annual rainfall of approximately 800 mm, with 70% concentrated between June and September. Short-duration torrential rains and persistent plum rains coexist, and the high forest cover and strong surface-soil water retention capacity further contribute to this. Effective rainfall in the preceding period is a crucial controlling factor in triggering the landslide. Considering factors such as stratigraphic structure, fault control, river valley lateral erosion, and rainfall concentration, the selected target landslide fully embodies the common characteristics of rainfall-induced deposit landslides in the Qinling Mountains region, serving as an ideal example for validating the "continuous rainfall characteristics—landslide displacement response" prediction framework. Rainfall data was collected from the target landslide to obtain a dataset. This dataset was then divided into training and test sets at an 8:2 or 7:3 ratio, and the corresponding rainfall event characteristics were obtained.

[0102] Taking a 15-day prediction step size as an example, this embodiment uses an LSTM multi-step prediction model (sliding prediction) and compares it with the method of this invention. (Refer to...) Figure 10In the results of time series methods, if the series trend is irregular, there is no basis to predict the occurrence of later displacements; therefore, only a slow upward trend is predicted. (Refer to...) Figure 11 and Figure 12 For regression prediction methods, the characteristics of rainfall events are the basis for determining whether slip has occurred.

[0103] In addition to intuitive predictive results, SHAP analysis provides interpretability of slippage. (See reference...) Figure 13 and Figure 14 Total rainfall is the most contributing factor, followed by rainfall duration, rainfall event index, previous effective rainfall, and maximum rainfall within the event (for SHAP plots, landslide displacement values ​​are always positive; regardless of whether the contribution value is positive or negative, it represents an increase in landslide). Combining the SHAP plot, it can be seen that high rainfall values ​​and long rainfall durations contribute the most to large landslides. Secondly, the rainfall event indexes also contribute significantly. These indexes imply potential time-series relationships; theoretically, earlier rainfall event indices contain less historical information, and as the event progresses, subsequent rainfall event indices should contribute more to the results, consistent with expectations. It can also be seen that smaller indices appear in the middle, while larger indices appear on the sides of the plot, indicating a larger contribution. Anterior effective rainfall also makes a significant contribution, characterizing to some extent the sustained impact of early rainfall on the soil. Furthermore, as shown in the figure, the SHAP value of anterior effective rainfall is unevenly distributed, indicating that it does not directly contribute to landslides (it merely increases the susceptibility of slopes). In the final prediction results, its influence is determined by combining it with other characteristics; for example, the influence of a characteristic is only reflected when a rainfall event sufficient to cause a landslide occurs, which aligns with objective facts. In addition, the contribution of anterior landslide volume to the results is mainly focused on small landslides, as detailed in the following paragraph. The contribution of anterior effective rainfall is relatively regular, with large rainfall amounts showing a positive contribution and small rainfall amounts showing a negative contribution.

[0104] The contribution of the initial slip volume varies across different time windows. The initial slip volume on the 3rd day is greater than that on the 15th and 30th days, indicating that the more recent slip event has a greater impact on the problem. Furthermore, the SHAP plot shows that smaller initial slip volumes contribute more, possibly due to the following reasons: smaller slips mean the slope still maintains a steep geometry and has already developed through-cracks and softening, making subsequent rainfall more likely to trigger large-scale instability; larger slips may have already removed unstable soil, reduced the slope angle, and resulted in denser residual soil, making seepage difficult to concentrate. However, the actual situation depends on the residual geometry, seepage conditions, and the degree of material softening. Nevertheless, the model at this point can already capture deeper information and aligns with reality. The initial slip volumes on the 30th day and 15th day, due to their longer time windows, do not accurately reflect the true impact of this slip event.

[0105] While the antecedent rainfall amounts in different time windows may exhibit some collinearity with the effective antecedent rainfall, their actual performance is not entirely identical. The 5-day antecedent rainfall reflects the cumulative moisture level on a short timescale, and since it occurs close to the time of slip, its characteristic contribution is strong. Furthermore, the SHAP plot shows that a larger SHAP value contributes more positively to the model, while a smaller value contributes more negatively (small slip), consistent with reality. The 15-day antecedent rainfall reflects the soil wetting effect of medium-term rainfall, with a lower characteristic contribution, but the SHAP plot shows strong regularity. The 30-day antecedent rainfall reflects the impact of long-term rainfall on the soil. Long-term rainfall is unrelated to single large rainfall events; therefore, the contribution of single large rainfall events in its SHAP plot distribution is not regular. However, compared to the characteristics of the effective antecedent rainfall, its contribution is smaller because the time decay effect of distant rainfall events is not considered.

[0106] The contribution of prior rainfall to landslides is influenced by local evaporation. Preliminary analysis shows that the characteristics that contribute significantly to landslides in this region are total rainfall, duration of rainfall, rainfall time index, prior effective rainfall, landslide amount in the first 3 days, effective rainfall in the first 5 days, and maximum rainfall during the event.

[0107] Example 3

[0108] Based on the same concept, the present invention also provides a landslide displacement prediction system, including an acquisition module, an extraction module and a prediction module.

[0109] The data acquisition module is used to collect current rainfall events in the study area. A rainfall event is the rainfall data of the study area over several consecutive days, and the interval between two adjacent rainfall events without rainfall is greater than a set threshold.

[0110] The extraction module is used to extract features from the current rainfall event to obtain multiple rainfall physical features that induce landslide displacement behavior. The rainfall physical features include a first feature reflecting the degree of slope infiltration, a second feature reflecting the rainfall time series, and a third feature reflecting the historical rainfall and landslide displacement data of the study area.

[0111] The prediction module is used to input multiple rainfall physical features into a pre-trained regression model and output the landslide displacement caused by the current rainfall event.

[0112] Example 3

[0113] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described landslide displacement prediction method.

[0114] Example 4

[0115] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described landslide displacement prediction method.

[0116] Considering the limitations of time series forecasting methods in predicting landslide displacement, this invention proposes a landslide displacement regression model prediction method based on "rainfall event segmentation" and "feature extraction" to predict landslide displacement. This method can predict the displacement at a specified time step in the future. It uses feature engineering to assist the model in extracting physical information from the data for prediction, and has high interpretability and prediction accuracy.

[0117] This invention is the first to propose a method for predicting landslide displacement using a regression model (previously time series models were used). This method avoids a series of problems associated with time series models and is more scientific and interpretable.

[0118] Disadvantages of time series forecasting methods: (1) Data must be timestamp data, such as the total rainfall of a single event, which cannot be input. (2) The model can hardly learn deeper information, and can only learn the trend change information of the curve, which is not interpretable. (3) The prediction step size is mostly 1 step, and there will be obvious errors if it exceeds 1 step.

[0119] Regression models require timestamp alignment; this method aligns with the "nth rainfall event." The model delves into the nonlinear relationships between various features within this rainfall event, and the contribution of each feature to the results can be clearly seen through SHOP plots and importance plots, demonstrating high interpretability. Furthermore, from a practical perspective, rainfall-induced landslide displacement is more closely related to rainfall events than to time series data. Simply put, without rain, slopes do not move over time, especially in winter when rainfall is low and the soil and rock moisture content is low.

[0120] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0121] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A landslide displacement prediction method, characterized in that, Includes the following steps: Collect current rainfall events in the study area; wherein, the rainfall event is the rainfall data of one day or multiple consecutive days in the study area, and the interval without rainfall between two adjacent rainfall events is greater than a set threshold; Feature extraction is performed on the current rainfall event to obtain multiple rainfall physical features that induce landslide displacement behavior. These rainfall physical features include a first feature reflecting the degree of slope infiltration, a second feature reflecting the rainfall time series, and a third feature reflecting historical rainfall and landslide displacement data in the study area. Multiple rainfall physical features are input into a pre-trained regression model, which outputs the landslide displacement caused by the current rainfall event. The first feature includes the total rainfall of a single rainfall event, the duration of rainfall, the maximum rainfall intensity within the event, and the effective rainfall. The second feature includes a rainfall event index. The third feature includes the average rainfall over the 5 days before the event, the average rainfall over the 15 days before the event, the average rainfall over the 30 days before the event, the slip over the 3 days before the event, the slip over the 15 days before the event, and the slip over the 30 days before the event. The total rainfall of a single rainfall event is the sum of the cumulative rainfall from the start to the end of the event; The duration of rainfall is the continuous rainfall time from the start to the end of the event; The maximum rainfall intensity during the event is the maximum hourly rainfall intensity during the event. The effective rainfall amount is the daily rainfall P over the previous N days, calculated using an exponential decay coefficient. Weighted summation; The rainfall event index is a number that sorts each independent rainfall event by time within the same monitoring period; The slippage in the three days prior to the event refers to the cumulative displacement over the three days preceding the event. The slippage in the 15 days prior to the event refers to the cumulative displacement over the 15 days preceding the event. The slippage in the 30 days prior to the event refers to the cumulative displacement over the 30 days preceding the event.

2. The landslide displacement prediction method as described in claim 1, characterized in that, The pre-training of the regression model includes the following steps: Historical rainfall data and corresponding historical slip volume were obtained for the study area. The historical rainfall data were divided according to the intervals between periods of no rainfall to obtain multiple historical rainfall events. The intervals between periods must be greater than a set threshold. For each historical rainfall event, feature extraction is performed to obtain multiple corresponding rainfall physical features; For each historical event, the corresponding multiple rainfall physical characteristics are used as inputs and the corresponding slip is used as outputs to pre-train the regression model.

3. The landslide displacement prediction method as described in claim 1, characterized in that, The set threshold is determined based on local climate conditions.

4. The landslide displacement prediction method as described in claim 1, characterized in that, The regression model used is the XGBOOST model.

5. A prediction system based on the landslide displacement prediction method of claim 1, characterized in that, include: The data acquisition module is used to collect current rainfall events in the study area; wherein, the rainfall event is the rainfall data of the study area over multiple consecutive days, and the interval between two adjacent rainfall events without rainfall is greater than a set threshold. The extraction module is used to extract features from the current rainfall event to obtain multiple rainfall physical features that induce landslide displacement behavior. The rainfall physical features include a first feature reflecting the degree of slope infiltration, a second feature reflecting the rainfall time series, and a third feature reflecting the historical rainfall and landslide displacement data of the study area. The prediction module is used to input multiple rainfall physical features into a pre-trained regression model and output the landslide displacement caused by the current rainfall event.

6. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the landslide displacement prediction method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the landslide displacement prediction method according to any one of claims 1-4.

Images