A landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold

Abstract

The application provides a landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold, and belongs to the technical field of data processing, and specifically comprises the following steps: taking satellite data as a background field and ground stations as control points, performing spatial residual correction based on historical reliability of the stations to generate a robust fusion rainfall field; based on the fusion rainfall field, short-term trigger variables and previous effective rainfall are extracted in a rolling manner, and historical daily rainfall is subjected to nonlinear physical infiltration reduction to peel off ineffective runoff; an offline three-dimensional threshold candidate library is constructed to adapt to three working conditions of a conventional steady state, long-time high-humidity cumulative infiltration and short-time heavy rainfall burst; in online early warning, the current working condition is diagnosed based on real-time rainfall characteristics, and a corresponding threshold model is adaptively selected; rainfall state parameters are used to form a three-dimensional evolution trajectory, which is dynamically compared with a multi-level nested threshold surface, and a graded early warning signal is output in real time according to the penetration degree. Through the scheme of the application, the early warning efficiency, accuracy and adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold. Background Technology

[0002] Currently, landslides are typical rainfall-induced geological hazards. Most existing regional landslide early warning technologies rely primarily on rainfall data, establishing empirical threshold relationships between rainfall indicators and landslide occurrence to achieve risk identification and early warning dissemination. However, existing landslide early warning technologies have the following shortcomings: Poor accuracy of multi-source rainfall fusion: Simple splicing or interpolation makes it difficult to balance the spatial continuity of satellite data and the local accuracy of ground station data, resulting in large deviations in the reconstruction of rainfall fields in mountainous areas.

[0003] Distortion in the physical representation of early rainfall: fixed number of days and fixed coefficients are accumulated without distinguishing between effective infiltration and ineffective runoff, which can easily lead to continuous false alarms after extreme rainstorms.

[0004] The single threshold model has poor adaptability: it does not distinguish between different disaster-causing conditions such as regular rainfall, long-term soaking, and short-term rainstorms, resulting in a slow response to long-duration processes and a delayed response to short-term rainstorms.

[0005] Lack of hourly dynamic rolling early warning: It relies heavily on daily or fixed-time static assessments, making it difficult to capture the trigger window and risk evolution process of short-term landslides.

[0006] It is evident that there is an urgent need for a landslide early warning system based on multi-source rainfall data fusion and three-dimensional thresholds, which offers high efficiency, accuracy, and adaptability. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a landslide early warning system based on multi-source rainfall data fusion and three-dimensional thresholds, which at least partially solves the problems of poor early warning efficiency, accuracy and adaptability in the prior art.

[0008] This invention provides a landslide early warning method based on multi-source rainfall data fusion and three-dimensional thresholding, comprising: Step 1: Real-time acquisition of satellite remote sensing rainfall data and ground meteorological station measured rainfall data in the study area to form multi-source rainfall data. Satellite remote sensing rainfall data is used as the spatial background field, and ground meteorological station measured rainfall data is used as high-precision true control points. Based on the historical reliability evaluation index of each meteorological station, spatial residual correction is performed on the satellite rainfall background field to generate a real-time robust fused rainfall field. Step 2: Based on the real-time robust fused rainfall field, for each landslide risk assessment unit, the short-term trigger variables and previous effective rainfall at the current moment are extracted in a rolling manner; wherein, the calculation of the previous effective rainfall includes: nonlinear physical infiltration reduction of historical daily rainfall to remove the ineffective runoff portion that exceeds the regional infiltration capacity; Step 3: Construct a three-dimensional rainfall threshold candidate library offline to adapt to different rainfall-induced disaster conditions. The candidate library includes a first threshold model adapted to the steady-state evolution of normal rainfall, a second threshold model adapted to the long-term high-humidity cumulative infiltration condition, and a third threshold model adapted to the short-term heavy rainfall sudden condition. Step 4, the online early warning stage, for each risk assessment unit at the current moment, based on the rainfall characteristic variables extracted in real time, diagnose the current rainfall-induced disaster conditions, and adaptively select the corresponding threshold model from the three-dimensional rainfall threshold candidate library according to the diagnosis results; Step 5: Based on the selected threshold model, construct a three-dimensional evolution trajectory from the real-time extracted rainfall state parameters, and dynamically compare it with the multi-level nested threshold surfaces under the model. Based on the degree to which the three-dimensional evolution trajectory penetrates different levels of threshold surfaces, output graded early warning signals in real time.

[0009] According to a specific implementation of an embodiment of the present invention, before step 1, the method further includes: The historical reliability evaluation index of ground weather stations was calculated using the leave-one-out cross-validation method. Within a historical time window, for each meteorological station, its own observations are removed, and the rainfall at that station is cross-estimated using observations from surrounding stations. The estimated values ​​are compared with the actual observations at that station, and the historical error variance for each station is calculated cumulatively. The historical error variance or its dimensionless index is used as the basis for evaluating the reliability of the station. The expression for the historical error variance is as follows: ; in, This represents the total number of historically valid rainfall samples. For the first Each site in history The actual observed rainfall at that moment; In order to be in Remove the first time The theoretical estimate is derived by cross-interpolating the location of a site using other nearby healthy sites after the site is identified.

[0010] According to a specific implementation of an embodiment of the present invention, the step of performing spatial residual correction on the satellite precipitation background field includes: For any target grid point, search for surrounding ground meteorological stations within its effective influence radius, and construct adaptive residual correction weights for each station based on the spatial distance of each meteorological station and its historical reliability evaluation index. Extract the satellite rainfall background value at the current location of each meteorological station and subtract it from the measured rainfall at that station to obtain the system residual at each station; The system residuals of each station are weighted and summed according to the adaptive residual correction weights, and the summation result is superimposed with the satellite rainfall background value of the target grid point to obtain the robust fused rainfall amount of the grid point.

[0011] According to a specific implementation of an embodiment of the present invention, the step of performing nonlinear physical infiltration reduction on historical daily rainfall includes: Obtain the daily saturation infiltration limit threshold of the corresponding geological zone of the study area. ; For any given day's rainfall When the rainfall does not exceed the daily saturation infiltration limit threshold, the entire daily rainfall is counted as effective infiltration rainfall. When the rainfall exceeds the daily saturation infiltration limit threshold, the portion within the daily saturation infiltration limit threshold is counted, and the excess portion is multiplied by a surface runoff residual coefficient less than 1 before being counted as effective infiltration rainfall. The expression is: ; in, This represents the residual coefficient of surface runoff.

[0012] According to a specific implementation of an embodiment of the present invention, the calculation of the effective rainfall in the early stage depends on the cumulative number of days and the attenuation coefficient, which are objectively calibrated through the following data-driven method: The cumulative number of days was determined by time-lag cross-correlation analysis between the historical landslide occurrence time series and the cumulative rainfall on different days before the disaster. The number of days corresponding to the convergence inflection point of the correlation coefficient curve was selected as the cumulative number of days. The attenuation coefficient is determined by optimizing the Fisher separability score of the effective rainfall in the early stage between the samples of the landslide occurrence day and the samples of the safe, non-disaster day.

[0013] According to a specific implementation of an embodiment of the present invention, the first threshold model adopts a variable-decoupled independent response structure, and its boundary equation is: ; The second threshold model employs a multi-parameter, strongly coupled nonlinear structure, and its boundary equations are as follows: ; The third threshold model adopts an extreme value mutation structure, and its boundary equation is: ; in, For rolling total rainfall, For the duration of rolling rainfall, This represents the effective rainfall in the preceding period. These are all core control parameters for each three-dimensional topological surface.

[0014] According to a specific implementation of an embodiment of the present invention, step 3 further includes: For each threshold model, the quantile regression algorithm is used to calculate the model parameter set under multiple preset probability quantiles, thereby generating a multi-level parallel nested warning threshold surface from low to high for each disaster-causing condition.

[0015] According to a specific implementation of the present invention, the rainfall characteristic variables include rolling average rainfall intensity and background saturation index; The rolling average rainfall intensity is calculated based on the current rolling total rainfall and the rolling rainfall duration. The background saturation index is calculated based on the ratio of the previous effective rainfall at the current moment to the historical maximum effective rainfall in the region.

[0016] According to a specific implementation of an embodiment of the present invention, step 4 specifically includes: When the rolling average rainfall intensity exceeds the preset heavy rainfall trigger threshold and the background saturation index is lower than the preset depth saturation threshold, it is diagnosed as a short-term heavy rainfall emergency and a third threshold model is selected. The heavy rainfall trigger threshold is determined by extracting the hourly average rainfall intensity sample set of historical disaster-free regular rainfall events in the study area and taking its preset high percentile. The depth saturation threshold is determined by extracting samples induced by long-term continuous rainfall in historical landslide events and taking its preset low percentile in the background saturation sample set at the time of the disaster. When the rolling average rainfall intensity does not exceed the preset heavy rainfall trigger threshold and the background saturation index is not lower than the preset depth saturation threshold, it is diagnosed as a long-term high humidity cumulative infiltration condition, and the second threshold model is selected. In other cases, the diagnosis is based on the steady-state evolution of normal rainfall, and the first threshold model is selected.

[0017] According to a specific implementation of an embodiment of the present invention, step 5 specifically includes: Substitute the current rolling rainfall duration and previous effective rainfall into the selected threshold model to dynamically solve the critical total rainfall limit corresponding to each warning level; The current rolling rainfall duration, previous effective rainfall, and rolling total rainfall are used to construct three-dimensional coordinate points, and the points of consecutive moments are connected to form a real-time rainfall evolution trajectory. Determine the situation where the real-time rainfall evolution trajectory crosses multiple nested threshold surfaces from bottom to top: for each threshold surface corresponding to a level that is crossed, output or upgrade the warning signal to the corresponding level. When the evolution trajectory falls back below the lower level threshold surface, the warning level is reduced or lifted accordingly.

[0018] The landslide early warning scheme based on multi-source rainfall data fusion and three-dimensional threshold in this embodiment of the invention includes: Step 1, acquiring satellite remote sensing rainfall data and ground meteorological station measured rainfall data of the study area in real time to form multi-source rainfall data, using satellite remote sensing rainfall data as the spatial background field, using ground meteorological station measured rainfall data as high-precision true value control points, and performing spatial residual correction on the satellite rainfall background field based on the historical reliability evaluation index of each meteorological station to generate a real-time robust fused rainfall field; Step 2, based on the real-time robust fused rainfall field, extracting short-term trigger variables and previous effective rainfall at the current moment for each landslide risk assessment unit; wherein, the calculation of the previous effective rainfall includes: performing nonlinear physical infiltration reduction on historical daily rainfall to remove the ineffective runoff portion exceeding the regional infiltration capacity; Step 3 3. Offline construction of a three-dimensional rainfall threshold candidate library adapted to different rainfall-induced disaster conditions. The candidate library includes a first threshold model adapted to the steady-state evolution of normal rainfall, a second threshold model adapted to the long-term high-humidity cumulative infiltration condition, and a third threshold model adapted to the short-term heavy rainfall sudden condition. 4. In the online early warning stage, for each risk assessment unit at the current moment, based on the rainfall feature variables extracted in real time, the current rainfall-induced disaster condition is diagnosed, and the corresponding threshold model is adaptively selected from the three-dimensional rainfall threshold candidate library according to the diagnosis results. 5. Based on the selected threshold model, the real-time extracted rainfall state parameters are used to construct a three-dimensional evolution trajectory, which is dynamically compared with the multi-level nested threshold surfaces under the model. According to the degree to which the three-dimensional evolution trajectory penetrates different levels of threshold surfaces, a graded early warning signal is output in real time.

[0019] The beneficial effects of the embodiments of the present invention are as follows: 1. This invention, through a robust fusion of "satellite background field + ground station residual correction", can take into account both the spatial continuity of satellite data and the local accuracy of ground station observations, thereby improving the real-time reconstruction accuracy of rainfall fields in complex mountainous areas and reducing the adverse effects of insufficient representativeness of single stations or local satellite bias on early warning results.

[0020] 2. This invention introduces infiltration limit constraints and runoff residual reduction in the early effective rainfall extraction stage, which can weaken the artificially high contribution of extreme rainstorms to the early moisture amount, reduce long-term continuous false alarms after disasters, and improve the physical rationality of the early warning results.

[0021] 3. This invention switches between three-dimensional threshold surfaces with different topological structures according to different rainfall-induced disaster conditions, making the early warning boundary more consistent with the actual disaster-causing mechanism. Therefore, compared with a single threshold model, it can simultaneously improve the adaptability to long-term cumulative conditions and short-term sudden conditions. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments 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.

[0023] Figure 1 A flowchart illustrating a landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the specific implementation process of a landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold provided in an embodiment of the present invention; Figure 3 A HICM three-dimensional rainfall threshold surface diagram provided in an embodiment of the present invention; Figure 4 A three-dimensional rainfall threshold surface plot provided for embodiments of the present invention; Figure 5 This invention provides an ERTM three-dimensional rainfall threshold surface plot. Detailed Implementation

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] 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. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0027] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0028] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0029] This invention provides a landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold, which can be applied to the landslide early warning process in geological disaster prevention and control scenarios.

[0030] See Figure 1 This is a flowchart illustrating a landslide early warning method based on multi-source rainfall data fusion and three-dimensional thresholding, provided by an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method mainly includes the following steps: Step 1: Real-time acquisition of satellite remote sensing rainfall data and ground meteorological station measured rainfall data in the study area to form multi-source rainfall data. Satellite remote sensing rainfall data is used as the spatial background field, and ground meteorological station measured rainfall data is used as high-precision true control points. Based on the historical reliability evaluation index of each meteorological station, spatial residual correction is performed on the satellite rainfall background field to generate a real-time robust fused rainfall field. In practice, real-time data from provincial automatic weather stations, national weather stations, and satellite remote sensing rainfall data are acquired within the study area. Preferably, satellite data uses GPM-type continuous precipitation products, and ground station data uses hourly rainfall observation data. Specific sub-steps are as follows: Step 1.1: Establish a historical reliability evaluation mechanism for various surface meteorological stations. Within a set historical time window, the system uses leave-one-out cross-validation to evaluate the historical observation error variance of each station. Perform rolling calculations. Specifically, for the first... For each weather station, its own observations are deleted at each historical moment. Only cross-estimation is performed using data from nearby healthy stations. The estimated values ​​are then compared with the station's actual observations, and the historical error variance for that station is accumulated. The specific calculation formula is as follows: ; In the formula: This represents the total number of historically valid rainfall samples. For the first Each site in history The actual observed rainfall at that moment; In order to be in Remove the first time The theoretical estimate is derived by cross-interpolating the location of a site using other nearby healthy sites after the site is identified.

[0031] Subsequently, the variance of each station's error was further dimensionless, serving as a relative historical error index characterizing the station's reliability. : ; In the formula, This represents the maximum historical error variance among all stations within the study area at the current moment.

[0032] Step 1.2: Using satellite remote sensing rainfall data as the spatial background field and ground meteorological stations as high-precision true value control points, construct residual correction weights. For any grid point to be evaluated... The spatial distances between each station and surrounding surface stations are calculated, and the unnormalized weights are obtained by combining the relative historical error index of each station. : ; Subsequently, the unnormalized weights were normalized to obtain the final residual correction weights for each associated station for that grid point. : ; In the formula: The distance decay power exponent (preferably set to 2); The dimensionless spatial relative distance is calculated using the following formula: (in For the target grid to the th The actual Euclidean distance between the stations (The maximum effective interpolation influence radius set for the system). This represents the total number of ground weather stations within the effective influence radius of the target grid point.

[0033] Step 1.3: At the current moment Regarding the first The spatial location of each ground meteorological station Extract the corresponding satellite background rainfall value and subtract it from the measured rainfall at the station to obtain the local system residual of satellite rainfall at that station. : ; In the formula: For the first High-precision measured rainfall data from several ground weather stations; This is an estimated value of the satellite remote sensing rainfall background for the corresponding spatial location.

[0034] The residual accurately quantifies the local drift and inversion bias of the satellite remote sensing inversion algorithm at this specific spatial location, providing a high-precision numerical benchmark for subsequent spatial assimilation calibration.

[0035] Step 1.4: For any target grid point within the study area The weighted sum of the satellite background rainfall value at this grid point and the residuals of the surrounding ground station system is superimposed and corrected to obtain the robust fused rainfall at the current moment. This method preserves the spatial continuity of the satellite background field while utilizing highly reliable ground stations to provide constrained corrections for local biases. ; In the formula: The background value of rainfall as perceived by satellite for the target grid points.

[0036] Step 2: Based on the real-time robust fused rainfall field, for each landslide risk assessment unit, the short-term trigger variables and previous effective rainfall at the current moment are extracted in a rolling manner; wherein, the calculation of the previous effective rainfall includes: nonlinear physical infiltration reduction of historical daily rainfall to remove the ineffective runoff portion that exceeds the regional infiltration capacity; In practice, it relies on the real-time robust fusion of rainfall fields that are continuously output at high frequency in step 1. For each landslide risk assessment unit within the study area, the system is configured with a high-frequency sliding time window and a time step. The preferred time is 1 hour, depending on the current time. As the process progresses, the three-dimensional control variables driving the subsequent state machine are extracted and updated in real time: rolling total rainfall. Duration of rolling rainfall Previous effective rainfall The specific sub-steps are as follows: Step 2.1: Sliding and rolling extraction of short-term trigger variables. To overcome the system delay defect of traditional early warning systems that rely on single static assessments at fixed daily times, making them prone to missing sudden short-term heavy rainfall, this invention introduces a time continuity dimension. At any current time... The system backtracks to historical data and calculates the following two short-term trigger variables in real time: Rolling total rainfall : Calculate from the current time The cumulative rainfall (in mm) of the target risk unit within the 24-hour time window.

[0037] Rolling Rainfall Duration Calculate the effective duration (in hours) of actual rainfall within the above 24-hour window.

[0038] Through a sliding time window mechanism, the diurnal static value is upgraded to a time-varying value. The dynamic time series with high-frequency evolution lays the data foundation for capturing sudden changes in rainfall and achieving advanced rolling forecasts.

[0039] Step 2.2: Perform nonlinear physical infiltration reduction on previous rainfall. The system calculates the current time... The previous Daily rainfall First, determine whether the rainfall exceeds the daily saturation infiltration limit threshold of the corresponding geological zone of the study area. If it does not exceed the threshold, all rainfall on that day is counted as effective infiltration rainfall. If it exceeds the threshold, only the small contribution of the portion within the threshold and the portion above the threshold multiplied by the residual surface runoff coefficient is counted as effective infiltration rainfall. This process can separate the portion of extreme rainfall that is primarily converted into surface runoff. The calculation formula is as follows: ; In the formula: The regional daily saturated infiltration limit threshold (unit: mm / d) represents the maximum physical water volume that a specific geological zone can absorb in 24 hours, which can be set by inversion from regional hydrogeological parameters or historical runoff coefficients. Residual coefficient of surface runoff (range of values) (Preferred value: 0.05-0.15). The physical meaning of this coefficient is: when a single day's torrential rain exceeds the soil infiltration limit... At that time, the vast majority of excess rainfall is converted into surface runoff and lost, with only a very small portion ( (Proportion) continues to slowly seep down through the fissures in the rock and soil.

[0040] Step 2.3: To ensure the extracted hydrological memory variables are adaptable to the regional geology, the system uses previous effective rainfall data. The core hyperparameters in the calculation formula are established using a data-driven, objective calibration mechanism, eliminating the need for arbitrary, manual value assignment. Valid cumulative days Dynamic extraction: Through "historical disaster-rainfall time lag cross-correlation analysis," the Pearson correlation coefficient between the historical landslide occurrence time series and the cumulative rainfall on different days before the disaster was calculated. The number of days corresponding to the convergence inflection point of the correlation coefficient curve was selected as the threshold. This parameter precisely matches the hydrological memory duration of previous rainfall for different soil and rock layers.

[0041] Nonlinear attenuation coefficient One-dimensional Fisher separability decoupling optimization: The effective rainfall from previous periods, calculated under different attenuation coefficients, is statistically analyzed in historical landslide day samples and historical safe day samples to identify differences in their distribution. A Fisher separability score function is constructed using the one-dimensional sequence of historical rainfall and the corresponding binary landslide labels (1 for landslide days, 0 for safe days). Perform decoupling optimization: ; In the formula: and In the given Below, the calculated effective rainfall preceding all historical landslide dates. The mean and variance; and The effective rainfall prior to all safe, disaster-free days is respectively. The mean and variance of [value]. Select [value] that results in Fisher's score. Reaching the global maximum value As the optimal calibration value.

[0042] Step 2.4: Effective rainfall in the early stage The rolling extraction. Combined with the effective infiltration rainfall after nonlinear truncation in step 2.2 above. and the hyperparameters objectively calibrated in step 2.3 and The system calculates the effective rainfall in the previous period at the current time t. : ; In the formula, The actual cumulative wetting and softening effect of the geological body in its early stages was quantified in real time. Since ineffective surface runoff has already been removed in the preceding steps, the output of this step... It is no longer a purely mathematical accumulation of rainfall, but a physical state variable that closely approximates the actual soil moisture saturation. This provides background humidity-driven data free from "extreme value contamination" for subsequent state machine adaptive diversion, eliminating the pain point of long-term continuous false alarms after extreme rainstorms from the source.

[0043] Step 3: Construct a three-dimensional rainfall threshold candidate library offline to adapt to different rainfall-induced disaster conditions. The candidate library includes a first threshold model adapted to the steady-state evolution of normal rainfall, a second threshold model adapted to the long-term high-humidity cumulative infiltration condition, and a third threshold model adapted to the short-term heavy rainfall sudden condition. In practice, based on the parallel sample set containing multi-source rainfall and real hydrological memory output from steps 1 and 2, the system extracts the effective rainfall data from the previous period. Rolling total rainfall and duration of rainfall As the core three-dimensional control variable.

[0044] To overcome the physical misalignment caused by the traditional early warning system's single static threshold surface for all weather conditions, this invention starts from the underlying physical mechanism of landslides and constructs three-dimensional rainfall evolution models with different spatial topological boundary characteristics for three typical disaster-causing meteorological conditions, forming a comprehensive candidate model library covering all conditions. The specific sub-steps are as follows: Step 3.1: Construct a threshold model HICM adapted to the steady-state evolution of conventional rainfall. This model is mainly for conventional rainfall processes without extreme short-duration pulses and where the soil is not extremely saturated in the early stages. Under this condition, landslide incubation is a linear and gradual process of pore water pressure accumulation. To match this disaster-causing physical mechanism, the system uses a variable-decoupled independent response structure to construct the threshold surface under this condition, and its boundary equation is expressed as: ; The structure appears in three-dimensional space as a smoothly transitioning topological surface, such as... Figure 3 As shown. Its physical nature determines the duration of rainfall. Compared with previous humidity The effects on slope stability are independent physical processes, and neither produces nonlinear amplification of the other's disturbances. This structure acts as a stabilizing compensation, ensuring extremely high stability of the system's warning boundary under normal meteorological conditions, thereby effectively filtering out false alarm interference caused by minor rainfall fluctuations.

[0045] Step 3.2: Construct a threshold model MRHM adapted to deep cumulative infiltration conditions under prolonged continuous rainfall and high background humidity. This model mainly targets high-risk situations where the deep sliding surface of the soil is softened by long-term soaking after prolonged continuous rain. Under these high background humidity conditions, the shear strength of the soil decreases sharply, and even a small amount of rainfall can easily induce deep cluster landslides. To accurately capture this mechanism, the system uses a multi-parameter strongly coupled nonlinear structure to construct the threshold surface under this condition, and its boundary equation is expressed as: ; This structure appears in three-dimensional space as a significantly nonlinear concave topological surface, such as... Figure 4 As shown. Its underlying technical logic lies in the creative introduction of a synergistic amplification mechanism between rainfall duration and preceding humidity. When preceding humidity is extremely high, this characteristic is amplified in the equation, significantly lowering the current required critical trigger rainfall. This mathematically maps the hydrogeological pattern of prolonged rainfall inevitably leading to disasters and minor rainfall causing major disasters, eliminating the blind spot in conventional linear models that are insensitive to previous high humidity conditions.

[0046] Step 3.3: Constructing a threshold model ERTM adapted to short-duration extreme heavy rainfall conditions. This model is specifically designed for extreme sudden rainfall events with extremely short durations and extremely high instantaneous rainfall intensity. Such heavy rainfall rapidly forms a dominant flow in the shallow layer of the slope, causing a sudden surge in pore water pressure, which can easily induce shallow debris flows or clusters of shallow landslides. To achieve an extremely rapid response, the system uses an extreme value catastrophe structure that includes the natural base to construct the threshold surface under this condition. Its boundary equation is expressed as: ; In three-dimensional space, this structure manifests as a warning surface that rapidly narrows and rises sharply over short-duration regions, such as... Figure 5 As shown, by utilizing the extreme sensitivity of the natural base e, this topological surface can generate a jump-like response to sudden high-frequency rainfall pulses. When extreme short-duration rainstorms occur in the rainfall evolution trajectory, the system can penetrate the tightened warning surface at the fastest speed, eliminating the time lag defect of traditional early warning models that must wait for the rainfall to accumulate slowly before issuing an alarm when facing sudden thunderstorms.

[0047] In the formulas of steps 3.1 to 3.3 above, These are all core control parameters for each three-dimensional topological surface. During the offline system construction phase, these parameters were all fixed offline by extracting a large sample set of historical disasters and rainfall data from the study area and using nonlinear least squares method or heuristic optimization algorithm for fitting and calibration. Step 3.4: Generation of Dynamic Multi-Level Nested Surfaces Based on Quantile Regression. After completing the offline fitting of the three models in Steps 3.1, 3.2, and 3.3 above, the system uses the quantile regression algorithm to calculate the model parameter set for each structure under specific probability distributions such as 15%, 30%, 60%, and 85%. This generates a multi-level, parallel nested warning surface library for each disaster-causing condition, providing a complete three-dimensional spatial scale for subsequent dynamic tiered early warning tracking.

[0048] Step 4, the online early warning stage, for each risk assessment unit at the current moment, based on the rainfall characteristic variables extracted in real time, diagnose the current rainfall-induced disaster conditions, and adaptively select the corresponding threshold model from the three-dimensional rainfall threshold candidate library according to the diagnosis results; In practical implementation, during the rolling online early warning process, the system constructs a meteorological characteristic state machine architecture to achieve dynamic routing at the underlying level of the early warning system. For each time step... Based on the rolling time series variables output in step S2, the system evaluates the physical evolution characteristics of the current rainfall in real time and selects a model. The specific sub-steps are as follows: Step 4.1: For each time step, the system calculates the current rolling average rainfall intensity in real time. and the saturation index characterizing the soil wetting background. Its mathematical formula is: ; ; In the formula, The historical limit of effective rainfall in the region can be objectively set by extracting the maximum value of previous effective rainfall from the historical long-term meteorological sequence (such as the last 10 years) of the target study area, and used to normalize the absolute value of previous rainfall to the range of zero to one. It characterizes the transient excitation kinetic energy of the current rainfall, while This quantifies the cumulative softening degree of the geological body in the early stage. Together, they constitute a two-dimensional meteorological evolution phase space, providing a quantitative benchmark for the system to diagnose the current rainfall conditions.

[0049] Step 4.2: Load the critical benchmark for flow splitting based on objective quantile calibration. 1) Critical threshold for triggering heavy rainfall The calibration involves extracting long-term, disaster-free, routine rainfall events from the target study area to construct a sample set of historical hourly average rainfall intensity. Sort the values ​​in the set in ascending order from smallest to largest, and directly extract the first value. Percentile as a threshold for triggering heavy rainfall: ; In the formula, This is a statistical percentile extraction function; This represents the total number of historical rainfall events. The extreme weather baseline set for the system is preferably set as follows: .

[0050] 2) Critical value of soil depth saturation Calibration. Extract all parallel samples of landslides induced by prolonged continuous rainfall from the historical landslide disaster sample set. Calculate the soil background saturation of these samples at the moment of disaster triggering, and construct a disaster-induced saturation sample set. Sort the set in ascending order and extract its first element. Percentile as the critical value for deep saturation: ; In the formula, This represents the total number of historically accumulated landslide events. To ensure a low percentage point for triggering the minimum threshold, the preferred setting is... .

[0051] Step 4.3: Perform adaptive model selection and route splitting based on operational condition diagnosis. The system diagnoses the current rainfall physical conditions in real time based on the relationship between the current rainfall data and the critical benchmark in Step 4.2 above, and selects the optimal three-dimensional topology threshold model to perform early warning calculations accordingly. When satisfied and When the conditions are met, the system diagnoses the current operating condition as a shallow, sudden-triggered condition dominated by a short-duration, strong convective drop. At this time, the system directly routes and distributes the warning data stream to the extreme value-sensitive transition topology model (ERTM) constructed in step S3 for independent calculation, utilizing its sharp rise characteristics to cope with the sudden pulse.

[0052] When satisfied and When the conditions are met, the system diagnoses the current operating condition as a deep cumulative infiltration condition caused by prolonged soaking due to continuous rain. At this time, the system routes and distributes the early warning data stream to the collaborative amplified concave topology threshold model (MRHM) for calculation, using extremely high background humidity as the physical amplifier to pull down the trigger boundary.

[0053] If the current meteorological coordinates do not meet the above extreme conditions, the system diagnoses it as a normal rainfall steady-state evolution condition. The warning data stream is matched to the Independent Compensated Smooth Topology Threshold Model (HICM) for calculation by default to ensure the stability of the global warning under normal climate conditions.

[0054] The core technical logic of this step lies in changing the rigid pattern of traditional early warning systems that forcibly apply a single fixed model to all complex weather conditions. The meteorological characteristic state machine constructed in this invention is essentially an intelligent condition diagnosis and routing distribution center. Based on the evolution of real-time rainfall characteristics, the system accurately identifies the current disaster-causing physical conditions and adaptively selects the most suitable early warning model for calculation, achieving a high degree of consistency between the underlying topology of the early warning algorithm and the actual physical disaster-causing mechanism of rainfall.

[0055] Step 5: Based on the selected threshold model, construct a three-dimensional evolution trajectory from the real-time extracted rainfall state parameters, and dynamically compare it with the multi-level nested threshold surfaces under the model. Based on the degree to which the three-dimensional evolution trajectory penetrates different levels of threshold surfaces, output graded early warning signals in real time.

[0056] In practice, based on the routing and diversion results of the meteorological characteristic state machine under the current rainfall conditions, the system performs dynamic spatial comparison and hierarchical response output of the rainfall evolution trajectory. The specific sub-steps are as follows: Step 5.1: Instantiation and Solving of Dynamic Critical Limit Surface and Construction of 3D Evolution Trajectory. Based on the activated 3D model and its corresponding multi-level parameter set, the system calculates the current rolling rainfall duration. Compared with the previous effective rainfall Substituting into the equation, the critical total rainfall limit corresponding to each warning level is dynamically solved. Simultaneously, the system incorporates the measured real-time rainfall status, i.e., the rainfall duration... Previous effective rainfall and rolling total rainfall The three-dimensional coordinate points are connected end to end according to the time step to fit and generate a real rainfall evolution trajectory that grows and spreads continuously over time in three-dimensional space.

[0057] Step 5.2: Spatial Approach Collision Detection and Tiered Early Warning Signal Response. The system calculates in real time the relative position and spatial crossing status of the above-mentioned rainfall evolution trajectory and the dynamic critical threshold surfaces at each level. As the actual rainfall evolution trajectory penetrates the bottom-level safe space from bottom to top over time, and successively breaks through the historical disaster probability quantile surfaces of 15%, 30%, 60%, and 85%, the system outputs progressively advancing landslide early warning signals in blue, yellow, orange, and red tiers. If subsequent rainfall weakens or stops, causing the real-time evolution trajectory to fall below the low quantile threshold surface, the system dynamically lowers the warning level in the reverse order, achieving adaptive alarm cancellation.

[0058] The key technical point of this invention is: 1. Using satellite precipitation as the spatial background field and ground stations as true control points, this method introduces the historical error variance of stations based on leave-one-out cross-validation to describe station reliability. Furthermore, it adaptively corrects the satellite precipitation residuals using distance attenuation relationships, thereby achieving robust spatial fusion of multi-source precipitation. This approach differs from conventional interpolation, simple stitching, or methods that rely solely on a single data source.

[0059] 2. In the early stage of effective rainfall extraction, the daily saturated infiltration limit threshold and the residual coefficient of surface runoff are first used to physically reduce the extreme rainstorms. Then, the cumulative number of days and the attenuation coefficient are determined through time-lag correlation analysis and Fisher separability optimization, so that the early effective rainfall has both physical significance and regional adaptability.

[0060] 3. Multiple three-dimensional threshold models are pre-constructed instead of using only a single fixed threshold structure. Different models correspond to three operating conditions: normal steady-state evolution, deep cumulative infiltration, and sudden heavy rainfall.

[0061] 4. Introduce a state machine with working condition characteristics, using rolling average rainfall intensity and background saturation index as the core criteria. Identify the real-time rainfall process based on the critical value obtained by objective calibration, and route it to the most matching threshold model.

[0062] The landslide early warning method based on multi-source rainfall data fusion and three-dimensional thresholding provided in this embodiment, through a robust fusion approach of "satellite background field + ground station residual correction," can balance the spatial continuity of satellite data and the local accuracy of ground station observations, thereby improving the real-time reconstruction accuracy of rainfall fields in complex mountainous areas and reducing the adverse effects of insufficient representativeness of single stations or local satellite bias on early warning results. Introducing infiltration limit constraints and runoff residual reduction in the early effective rainfall extraction stage can weaken the artificially inflated contribution of extreme rainstorms to early-stage moisture levels, reduce long-term continuous false alarms after disasters, and improve the physical rationality of early warning results. Switching between three-dimensional threshold surfaces with different topological structures according to different rainfall-induced disaster conditions makes the early warning boundary more closely match the actual disaster-induced mechanism. Therefore, compared to a single threshold model, it can simultaneously improve adaptability to both long-term cumulative and short-term sudden disaster conditions.

[0063] The method of the present invention will be further described below with reference to a specific embodiment: 1) Implementation area and basic data This embodiment uses a county as the study area and employs historical rainfall and landslide data from January 1, 2022 to December 31, 2024 to fully implement the present invention. The county's terrain is predominantly mountainous and hilly with well-developed valleys. It receives abundant rainfall, but the rainfall distribution is uneven in time and space, making it a typical area prone to rainfall-induced landslides. A total of 216 landslide events clearly triggered by rainfall were recorded in the study area between 2022 and 2024. The available basic rainfall data for the study area includes: hourly rainfall data from 46 provincial automatic weather stations within the county, hourly rainfall data from one national weather station, and GPM IMERG FinalRun V07B satellite remote sensing rainfall data, wherein the satellite remote sensing rainfall data has a temporal resolution of 0.5 hours and a spatial resolution of 0.1° × 0.1°.

[0064] In this embodiment, instead of establishing independent early warning models for provincial automatic station data, national meteorological station data, and satellite remote sensing rainfall data, the three types of data are uniformly incorporated into the multi-source robust fusion process of this invention to form a unique fused rainfall field. Based on this fused rainfall field, all subsequent parameter calibrations, threshold modeling, and online early warnings are completed.

[0065] To achieve online early warning, this embodiment divides the study area of ​​a certain county into... The rules for early warning units are as follows. Based on the area of ​​a county of approximately 4950 km², approximately 4950 early warning units can be formed. The center coordinates of each early warning unit serve as a unified location identifier for spatial fusion calculations, variable extraction, and early warning output. This grid scale can balance the local characteristics of heavy rainfall in the mountainous areas of the county with the computational efficiency of operational early warning. The area, topography, and sample size of the county can serve as the realistic basis for the above-mentioned division scale.

[0066] 2) Preprocessing of multi-source rainfall data The three types of rainfall data were processed to a unified time scale. For data from 46 provincial automatic weather stations and 1 national weather station, hourly rainfall observation sequences were directly read; for GPM satellite remote sensing rainfall data, adjacent 0.5-hour steps were accumulated to obtain an hourly rainfall sequence consistent with that of ground weather stations.

[0067] A standardized database of 216 landslide events was created, forming a historical landslide sample library. Each sample includes at least the landslide occurrence time, landslide point coordinates, and the corresponding early warning unit number. For subsequent parameter calibration and effectiveness verification, this embodiment divides the 216 landslide events into a fitting set and a validation set in a 7:3 ratio. The fitting set is used to fit the key parameters and threshold surfaces involved in this invention, while the validation set is used to verify the online early warning process. The 216 landslide events and the 7:3 sample organization are well-suited to the data conditions of a specific county.

[0068] 3) Site reliability calculation and spatially robust fusion precipitation field construction In this embodiment, GPM satellite rainfall data is used as the spatial continuous background rainfall field for the entire study area, and 46 provincial automatic weather stations and 1 national meteorological station are used as high-precision true control points.

[0069] For each ground meteorological station, the system calculates the historical error variance of the station using leave-one-out cross-validation within a historical window. Specifically, at any given historical moment, the station to be evaluated is temporarily removed, and the rainfall at that station's location is cross-estimated using only other healthy stations. The estimated value is then subtracted from the actual observed value at that moment to obtain the cross-error for that time. The squared errors of all valid samples within the historical window are accumulated to obtain the historical error variance of that station. Finally, the historical error variances of all stations are dimensionless to obtain the relative historical error index for each station. The smaller the relative historical error index, the more stable and reliable the station's historical observations are.

[0070] For any grid point within the study area to be warned, the system searches for all ground stations within its maximum effective influence radius, calculates the spatial distance between the grid point and each station, and constructs an adaptive residual correction weight for the grid point with respect to each station, based on the relative historical error index of the corresponding stations. Preferably, the distance attenuation power exponent can be set to 2. The closer the station is and the smaller its historical error, the higher its residual correction weight for that grid point.

[0071] At any given moment, the system extracts the GPM background rainfall value corresponding to each ground station location and subtracts it from the measured rainfall at that station to obtain the system residual for that station location. Then, the system residuals of each station are propagated to the grid points to be warned according to the aforementioned adaptive weights and superimposed with the GPM background value of that grid point to obtain the robust fused rainfall for that grid point at the current moment. Thus, a robust fused rainfall field updated hourly can be formed over the entire county.

[0072] In this embodiment, the system performs the above-mentioned fusion process hourly for all times from January 1, 2022 to December 31, 2024, to form an hourly robust fused rainfall database covering the entire county, which is used for offline modeling and online early warning.

[0073] 4) Extraction of three-dimensional control variables After constructing a robust and integrated rainfall field, the system continuously extracts three-dimensional control variables for each risk assessment unit with a time step of 1 hour, and continuously integrates the total rainfall. Duration of rainfall and previous effective rainfall .

[0074] in, This refers to the cumulative value of all merged rainfall over the past 24 hours from the current moment. This refers to the duration of rainfall within the aforementioned 24-hour window. Using this sliding window method, the system can provide the short-term triggered rainfall status of the current risk unit at any given time.

[0075] For the effective rainfall in the early stage This embodiment does not use the method of "directly adding the daily rainfall of the previous N days according to a fixed attenuation coefficient," but instead first performs physical infiltration reduction on the daily rainfall for each day. The system compares the cumulative daily rainfall of the nth day prior to the current moment with the regional daily saturated infiltration limit threshold: when the daily rainfall does not exceed the daily saturated infiltration limit threshold, it is considered that all the rainfall of that day can enter the soil and is counted as effective infiltrated rainfall; when the daily rainfall exceeds the daily saturated infiltration limit threshold, only the portion within the threshold is counted, and the excess portion is reduced according to the surface runoff residual coefficient before being counted. Preferably, the surface runoff residual coefficient ranges from 0.05 to 0.15. Through this processing, a large amount of ineffective rainfall that is converted into surface runoff can be removed, reducing the pollution of the preceding wet background by extreme rainstorms.

[0076] 5) Objective calibration of effective rainfall parameters in the early stage In this embodiment, to avoid the subjective setting of the cumulative number of days in the early stage... and attenuation coefficient The parameters were objectively calibrated using historical samples from a certain county.

[0077] First, regarding the determination of the effective cumulative days, the time series of landslide events were extracted from 216 events, and the cumulative rainfall under different cumulative days before the disaster was calculated. Time-lag cross-correlation analysis was then performed with the landslide time series. As the cumulative days increased, the correlation first strengthened and then stabilized. When the correlation curve reached a convergence inflection point, the corresponding cumulative days were selected as the effective cumulative days in this embodiment. This approach ensures that the cumulative days are no longer manually specified based on experience, but are automatically provided by the historical disaster-rainfall relationship of a county.

[0078] Secondly, regarding the determination of the attenuation coefficient, the system calculates the effective rainfall in the preceding period for each historical sample within the candidate attenuation coefficient range, and separately calculates the mean and variance of the effective rainfall in the corresponding preceding period for samples from landslide-occurring days and samples from safe, non-landslide-occurring days, constructing a Fisher separability score function. The attenuation coefficient that maximizes the Fisher score is selected as the optimal attenuation coefficient for a certain county. This allows the constructed effective rainfall in the preceding period to have higher distinguishability between "landslide days" and "non-landslide days".

[0079] After completing the above parameter calibration, the system will weight and accumulate the effective infiltration rainfall of each previous day according to the determined cumulative number of days and attenuation coefficient to obtain the effective rainfall of the previous period for each risk assessment unit and each hourly step.

[0080] 6) Offline construction of a three-dimensional threshold candidate library for a certain county After completing the construction of the hourly fused rainfall field and the three-dimensional control variable database, the system establishes a three-dimensional threshold candidate library adapted to different disaster-causing conditions based on the fitted set samples. The three-dimensional control variables include: the effective rainfall in the early stage, the total rolling rainfall in 24 hours, and the duration of the rolling rainfall in 24 hours.

[0081] In this embodiment, three types of threshold models are established. On each type of topological surface, the system further employs quantile regression to perform multi-level fitting. Preferably, four quantile levels of 15%, 30%, 60%, and 85% are used to fit the four-level risk threshold surfaces, forming a nested surface library from low to high, providing criteria for subsequent blue, yellow, orange, and red four-level early warnings.

[0082] The first type is the Independent Compensated Smooth Topological Threshold Model (HICM), adapted to the steady-state evolution of landslides under normal rainfall conditions. This type of model assumes that the duration of rainfall and the impact of the preceding wet background on landslide instability are relatively independent, and is suitable for steady-state risk assessment under normal continuous rainfall conditions in a certain county.

[0083] ; The second category is the collaboratively amplified concave topological threshold model (MRHM), adapted to prolonged periods of continuous rainfall and high background humidity. This type of model emphasizes that when the effective rainfall in the previous period is high, it has an amplifying and weakening effect on the current triggering rainfall threshold, making it suitable for describing landslide triggering scenarios in a county under continuous rainfall and deep soaking conditions.

[0084] ; The third category is the extreme value-sensitive transition topology threshold model ERTM, which is adapted to short-duration heavy rainfall events. This type of model is more sensitive to high rainfall pulses within a short duration and is suitable for describing localized short-term sudden disasters caused by rainstorms in a county.

[0085] ; 7) Establishment of a state machine for judging working conditions in a certain county In this embodiment, before online early warning, a state machine for identifying operating conditions needs to be established using historical data from a certain county. Specifically, the system defines two operating condition identification indicators for each time step: rolling average rainfall intensity. and background saturation index The rolling average rainfall intensity is calculated from the current 24-hour rolling total rainfall and the current 24-hour rolling rainfall duration, and is used to describe the intensity of the current rainfall pulse; the background saturation index is given by the normalized result of the current previous effective rainfall relative to the regional historical limit effective rainfall, and is used to describe the soil wetting background.

[0086] Based on historical data from a certain county, the system first extracts a sample set of hourly average rainfall intensity from long-term disaster-free routine rainfall events, and takes the 95th percentile as the threshold for triggering heavy rainfall. Secondly, a background saturation sample set at the time of the disaster was extracted from historical landslide samples induced by long-term continuous rainfall, and its 30th percentile was taken as the depth saturation critical value. The two critical values ​​obtained in this way are used to characterize the boundary between "sudden heavy rainfall conditions" and "high humidity background conditions," respectively.

[0087] During online operation, if the current rolling average rainfall intensity reaches or exceeds the heavy rainfall trigger threshold, and the current background saturation index is lower than the depth saturation threshold, the system determines the current operating condition as a shallow sudden trigger condition dominated by short-term strong convective descent, and automatically routes to an extreme value sensitive transition topology. If the current rolling average rainfall intensity is lower than the heavy rainfall trigger threshold, and the current background saturation index reaches or exceeds the depth saturation threshold, the system determines the current operating condition as a deep cumulative infiltration condition dominated by long-term soaking, and routes to a cooperatively amplified concave topology. In other cases, the system defaults to a normal rainfall steady-state evolution condition and routes to an independent compensation smooth topology.

[0088] 8) Online early warning operation process in a certain county In this embodiment, after the system enters online operation, it executes a complete early warning process every hour. Taking a certain current time t as an example, the system first reads the latest rainfall data of 46 provincial automatic weather stations, 1 national meteorological station, and the corresponding time period of the GPM product in a certain county at that time; then, it calculates the robust fused rainfall field of the entire study area at the current time according to method 3).

[0089] Next, for each risk assessment unit in a county, the system recalculates the 24-hour rolling total rainfall, 24-hour rolling rainfall duration, and previous effective rainfall at time t; then it calculates the rolling average rainfall intensity and background saturation index; and then calls the working condition state machine established in 8) to determine the type of three-dimensional threshold model that should be attached to the unit.

[0090] Subsequently, the system combines the three-dimensional variables of the unit at the current moment into a three-dimensional coordinate point, and connects this point with the corresponding points at the previous moment, etc., to form the real-time rainfall evolution trajectory of the unit over time. Based on the currently mounted topological surface model and quantile threshold surface library, the system dynamically solves the critical total rainfall surface corresponding to each warning level, and then determines whether the current evolution trajectory crosses the 15%, 30%, 60%, 85% quantile threshold surfaces.

[0091] When the evolution trajectory of a risk assessment unit only exceeds the lowest risk threshold, the system outputs a blue alert; when it continues to exceed the next threshold, the system upgrades it to a yellow alert; when it continues to exceed even higher thresholds, the system outputs orange and then red alerts in sequence. If subsequent rainfall weakens, causing the real-time evolution trajectory to fall back below the lower threshold, the system can lower the alert level in the reverse order.

[0092] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A landslide early warning method based on multi-source rainfall data fusion and three-dimensional threshold, characterized in that, include: Step 1: Real-time acquisition of satellite remote sensing rainfall data and ground meteorological station measured rainfall data in the study area to form multi-source rainfall data. Satellite remote sensing rainfall data is used as the spatial background field, and ground meteorological station measured rainfall data is used as high-precision true control points. Based on the historical reliability evaluation index of each meteorological station, spatial residual correction is performed on the satellite rainfall background field to generate a real-time robust fused rainfall field. Step 2: Based on the real-time robust fused rainfall field, for each landslide risk assessment unit, the short-term trigger variables and previous effective rainfall at the current moment are extracted in a rolling manner; wherein, the calculation of the previous effective rainfall includes: nonlinear physical infiltration reduction of historical daily rainfall to remove the ineffective runoff portion that exceeds the regional infiltration capacity; Step 3: Construct a three-dimensional rainfall threshold candidate library offline to adapt to different rainfall-induced disaster conditions. The candidate library includes a first threshold model adapted to the steady-state evolution of normal rainfall, a second threshold model adapted to the long-term high-humidity cumulative infiltration condition, and a third threshold model adapted to the short-term heavy rainfall sudden condition. Step 4, the online early warning stage, for each risk assessment unit at the current moment, based on the rainfall characteristic variables extracted in real time, diagnose the current rainfall-induced disaster conditions, and adaptively select the corresponding threshold model from the three-dimensional rainfall threshold candidate library according to the diagnosis results; Step 5: Based on the selected threshold model, construct a three-dimensional evolution trajectory from the real-time extracted rainfall state parameters, and dynamically compare it with the multi-level nested threshold surfaces under the model. Based on the degree to which the three-dimensional evolution trajectory penetrates different levels of threshold surfaces, output graded early warning signals in real time.

2. The method according to claim 1, characterized in that, Before step 1, the method further includes: The historical reliability evaluation index of ground weather stations was calculated using the leave-one-out cross-validation method. Within a historical time window, for each meteorological station, its own observations are removed, and the rainfall at that station is cross-estimated using observations from surrounding stations. The estimated values ​​are compared with the actual observations at that station, and the historical error variance for each station is calculated cumulatively. The historical error variance or its dimensionless index is used as the basis for evaluating the reliability of the station. The expression for the historical error variance is as follows: in, This represents the total number of historically valid rainfall samples. For the first Each site in history The actual observed rainfall at that moment; In order to be in Remove the first time The theoretical estimate is derived by cross-interpolating the location of a site using other nearby healthy sites after the site is identified.

3. The method according to claim 2, characterized in that, The steps for spatial residual correction of the satellite precipitation background field include: For any target grid point, search for surrounding ground meteorological stations within its effective influence radius, and construct adaptive residual correction weights for each station based on the spatial distance of each meteorological station and its historical reliability evaluation index. Extract the satellite rainfall background value at the current location of each meteorological station and subtract it from the measured rainfall at that station to obtain the system residual at each station; The system residuals of each station are weighted and summed according to the adaptive residual correction weights, and the summation result is superimposed with the satellite rainfall background value of the target grid point to obtain the robust fused rainfall amount of the grid point.

4. The method according to claim 1, characterized in that, The step of performing nonlinear physical infiltration reduction on historical daily rainfall includes: Obtain the daily saturation infiltration limit threshold of the corresponding geological zone of the study area. ; For any given day's rainfall When the rainfall does not exceed the daily saturation infiltration limit threshold, the entire daily rainfall is counted as effective infiltration rainfall. When the rainfall exceeds the daily saturation infiltration limit threshold, the portion within the daily saturation infiltration limit threshold is counted, and the excess portion is multiplied by a surface runoff residual coefficient less than 1 before being counted as effective infiltration rainfall. The expression is: in, This represents the residual coefficient of surface runoff.

5. The method according to claim 1, characterized in that, The calculation of effective rainfall in the early stage relies on the cumulative number of days and the attenuation coefficient, which are objectively calibrated through the following data-driven method: The cumulative number of days was determined by time-lag cross-correlation analysis between the historical landslide occurrence time series and the cumulative rainfall on different days before the disaster. The number of days corresponding to the convergence inflection point of the correlation coefficient curve was selected as the cumulative number of days. The attenuation coefficient is determined by optimizing the Fisher separability score of the effective rainfall in the early stage between the samples of the landslide occurrence day and the samples of the safe, non-disaster day.

6. The method according to claim 1, characterized in that, The first threshold model adopts a variable-decoupled independent response structure, and its boundary equation is: ; The second threshold model employs a multi-parameter, strongly coupled nonlinear structure, and its boundary equations are as follows: ; The third threshold model adopts an extreme value mutation structure, and its boundary equation is: ; in, For rolling total rainfall, For the duration of rolling rainfall, This represents the effective rainfall in the preceding period. These are all core control parameters for each three-dimensional topological surface.

7. The method according to claim 1, characterized in that, Step 3 also includes: For each threshold model, the quantile regression algorithm is used to calculate the model parameter set under multiple preset probability quantiles, thereby generating a multi-level parallel nested warning threshold surface from low to high for each disaster-causing condition.

8. The method according to claim 1, characterized in that, The rainfall characteristic variables include rolling average rainfall intensity and background saturation index; The rolling average rainfall intensity is calculated based on the current rolling total rainfall and the rolling rainfall duration. The background saturation index is calculated based on the ratio of the previous effective rainfall at the current moment to the historical maximum effective rainfall in the region.

9. The method according to claim 8, characterized in that, Step 4 specifically includes: When the rolling average rainfall intensity exceeds the preset heavy rainfall trigger threshold and the background saturation index is lower than the preset depth saturation threshold, it is diagnosed as a short-term heavy rainfall emergency and a third threshold model is selected. The heavy rainfall trigger threshold is determined by extracting the hourly average rainfall intensity sample set of historical disaster-free regular rainfall events in the study area and taking its preset high percentile. The depth saturation threshold is determined by extracting samples induced by long-term continuous rainfall in historical landslide events and taking its preset low percentile in the background saturation sample set at the time of the disaster. When the rolling average rainfall intensity does not exceed the preset heavy rainfall trigger threshold and the background saturation index is not lower than the preset depth saturation threshold, it is diagnosed as a long-term high humidity cumulative infiltration condition, and the second threshold model is selected. In other cases, the diagnosis is based on the steady-state evolution of normal rainfall, and the first threshold model is selected.

10. The method according to claim 1, characterized in that, Step 5 specifically includes: Substitute the current rolling rainfall duration and previous effective rainfall into the selected threshold model to dynamically solve the critical total rainfall limit corresponding to each warning level; The current rolling rainfall duration, previous effective rainfall, and rolling total rainfall are used to construct three-dimensional coordinate points, and the points of consecutive moments are connected to form a real-time rainfall evolution trajectory. Determine the situation where the real-time rainfall evolution trajectory crosses multiple nested threshold surfaces from bottom to top: for each threshold surface corresponding to a level that is crossed, output or upgrade the warning signal to the corresponding level. When the evolution trajectory falls back below the lower level threshold surface, the warning level is reduced or lifted accordingly.

Images