A multi-time scale risk trend prediction method and system for collapse disasters

Abstract

The application discloses a kind of multi-time scale risk trend prediction methods and systems for collapse disaster, it is related to disaster risk prediction technical field.A kind of multi-time scale risk trend prediction system for collapse disaster, including have: data acquisition module, stage determination module, scale adjustment module, prediction constraint module, scale prediction module, risk fusion module and event extraction module.The application outputs each scale risk trend sequence and uncertainty representation under the bidirectional constraint of space-time constraint propagation network in scale prediction module, then by dynamic weight distributor, stage identification information and uncertainty representation are combined to calculate the fusion weight that changes with time, cross-scale consistency constraint is applied, the fusion risk trend curve with confidence interval is obtained, the prediction result of each time scale is adaptively weighted and constrained according to uncertainty.

Description

Technical Field

[0001] This invention relates to the field of disaster risk prediction technology, and in particular to a method and system for predicting risk trends across multiple time scales for landslide disasters. Background Technology

[0002] Landslides are one of the main types of geological hazards in mountainous engineering construction and residential safety. Their occurrence is controlled by long-term geological environmental conditions such as topography, lithology, and geological structure, while also being significantly influenced by short-term triggering factors such as rainfall, snowmelt, and human engineering activities. With the improvement of monitoring and information technology, geological environmental survey data, meteorological and hydrological observation data, and deformation monitoring data such as displacement and dip angle can usually be obtained for landslide hazard points, providing a multi-source data foundation for risk trend analysis and early warning. In current engineering practice, common methods include static susceptibility zoning based on geological environmental factors, threshold early warning based on rainfall amount or intensity, and statistical regression or machine learning prediction based on a single time scale, used to determine landslide risk levels or trigger early warning signals.

[0003] However, in complex mountainous environments, traditional methods have several shortcomings: First, most methods only process triggering and response information on a single time scale, making it difficult to simultaneously depict short-term triggering processes at the minute to hour level, cumulative effects at the day to week level, and background susceptibility changes at the seasonal to flood season level, resulting in an insufficiently continuous and nuanced description of the risk evolution process; Second, even with the introduction of multi-source data or multi-model fusion, the results are weighted after each time scale or sub-model runs independently, lacking dynamic coordination and mutual constraints in the prediction process, and failing to reflect the impact of short-term anomalies on medium- and long-term trends and the reasonable boundaries of medium- and long-term background for short-term predictions; Third, existing early warnings mostly focus on judging whether a threshold is exceeded at a certain moment, making it difficult to provide information such as risk trend curves and key inflection points and risk cross-level time windows for a period of time in the future, which is not conducive to management departments arranging disaster prevention and mitigation measures in advance. Summary of the Invention

[0004] This invention constructs a method that, under complex geological environments and multi-source monitoring conditions, collaboratively utilizes geological environmental data, meteorological and hydrological data, and deformation response data across multiple time scales. Combined with real-time identification of the evolution stages of the landslide body, it establishes a two-way constraint relationship between subscale trend predictions. Based on this, it outputs a landslide risk trend curve with uncertainty characteristics and a prediction method for its key events, thereby overcoming the limitations of existing technologies that rely on single time scales, weak collaboration, and difficulty in characterizing risk evolution trends.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A multi-timescale risk trend prediction method for landslide disasters includes:

[0007] Acquire multi-source data of target potential hazards and perform time stamp unification, spatial correlation and quality identification to form a standard dataset; generate a multi-scale feature set based on the standard dataset;

[0008] The evolution stage identification model determines the evolution stage in real time using a multi-scale feature set and outputs stage identification information.

[0009] The time scale division parameters are dynamically determined based on the stage identification information, and the multi-scale feature set is adjusted accordingly;

[0010] Construct a spatiotemporal constraint propagation network, establish bidirectional constraints among trend predictors at each scale, and set constraint propagation strength parameters according to stage identification information;

[0011] The adjusted multi-scale feature sets are input into the corresponding subscale trend predictors. Under the bidirectional constraints of the spatiotemporal constraint propagation network, the risk trend sequences and their uncertainty representations at each scale are output.

[0012] The dynamic weight allocator determines the fusion weights based on stage identification information and uncertainty characterization and applies cross-scale consistency constraints, thereby obtaining the fusion risk trend curve and its confidence interval.

[0013] Extract key risk trend events from the integrated risk trend curve and output them.

[0014] As a preferred technical solution of the present invention, the generation of the multi-scale feature set includes: the multi-source data includes geological environment data, meteorological and hydrological data, and deformation response data; the multi-source data is written into data records containing a unified timestamp field and a spatial positioning field; the data records are processed by unified timestamp processing, spatial correlation processing, and a quality identification field is added to obtain a standard dataset; based on the standard dataset, under preset different time windows and sampling steps, the meteorological and hydrological data and deformation response data are accumulated, sliding statistics and change rate calculations are performed to form short-term features for characterizing minute to hourly changes, medium-term features for characterizing day to weekly changes, and long-term background scale features for characterizing seasonal to flood season changes; and the short-term features, medium-term features and long-term background scale features are combined to form a multi-scale feature set.

[0015] As a preferred technical solution of the present invention, the real-time determination of the evolution stage includes: forming an evolution feature sequence by arranging the short-term, medium-term, and long-term background scale features corresponding to the target hidden danger point in the multi-scale feature set in chronological order; extracting evolution characterization indicators including displacement, displacement velocity, displacement acceleration, and tilt angle change rate from the evolution feature sequence within a preset sliding time window; inputting the evolution characterization indicators into the evolution stage identification model for stage classification calculation to obtain the current evolution stage; the current evolution stage includes a stable creep stage, an accelerated creep stage, and a near-slip warning stage; encoding the current evolution stage as stage identification information and attaching a confidence level corresponding to the evolution stage.

[0016] As a preferred embodiment of the present invention, the adjustment of the multi-scale feature set includes: mapping the stage identification information to a set of time scale division parameters corresponding to each evolution stage, wherein the time scale division parameters include the time window length, sliding step size, and scale division boundary used to define the short-term, medium-term, and long-term background scales; when the stage identification information changes, selecting the time scale division parameters corresponding to the current evolution stage from the time scale division parameter set, re-accumulating, sliding statistics, and calculating the rate of change of the meteorological and hydrological data and deformation response data in the standard dataset under the updated time window and sliding step size to obtain updated short-term, medium-term, and long-term background scale features, and redistributing or discarding the features in the original multi-scale feature set based on the updated scale division boundary to form the adjusted multi-scale feature set.

[0017] As a preferred technical solution of the present invention, the output of the risk trend sequence and its uncertainty representation includes: taking the short-term, medium-term, and long-term background scale features in the adjusted multi-scale feature set as input features for the short-term trend predictor, medium-term trend predictor, and long-term background scale trend predictor, respectively; under the bidirectional constraint of the spatiotemporal constraint propagation network, each subscale trend predictor reads the stage identification information and selects the model parameter set corresponding to the current evolution stage; performs time series prediction operations on the input features to obtain the risk trend sequence of the corresponding time scale; and calculates the uncertainty representation of the risk trend sequence based on the prediction residual and sample variance.

[0018] As a preferred embodiment of the present invention, the bidirectional constraint includes: in the spatiotemporal constraint propagation network, a short-term trend predictor, a medium-term trend predictor, and a long-term background scale trend predictor are respectively used as constraint nodes; constraint channels are established between each constraint node, from the short-term trend predictor to the medium-term trend predictor and the long-term background scale trend predictor, as well as constraint channels from the medium-term trend predictor and the long-term background scale trend predictor to the short-term trend predictor; in each prediction step, the risk trend sequence and uncertainty characterization output by each subscale trend predictor are used as constraint information, and transmitted to the subscale trend predictors of other time scales through the corresponding constraint channels to update the input features, initial state, and prediction boundary conditions of the subscale trend predictors, and the time series prediction operation of the next prediction step is performed after the update.

[0019] As a preferred embodiment of the present invention, the constraint propagation strength parameters include: determining the current evolution stage based on stage identification information; selecting a set of constraint propagation strength parameters corresponding to the current evolution stage from a preset constraint propagation strength parameter table, including a first constraint weight coefficient for transmitting constraint information from the short-term trend predictor to the medium-term trend predictor and the long-term background scale trend predictor; a second constraint weight coefficient for transmitting constraint information from the medium-term trend predictor and the long-term background scale trend predictor to the short-term trend predictor; and a maximum adjustment range for limiting the input feature correction, initial state offset, and prediction boundary condition adjustment caused by the constraint information; when the spatiotemporal constraint propagation network performs constraint information transmission, the constraint information in each constraint channel is multiplied by the corresponding first constraint weight coefficient or second constraint weight coefficient, and the input feature correction, initial state offset, and prediction boundary condition adjustment are truncated according to the maximum adjustment range.

[0020] As a preferred embodiment of the present invention, the acquisition of the fusion risk trend curve and its confidence interval includes: a dynamic weight allocator selecting preset basic weight coefficients for different evolution stages based on stage identification information, and calculating the fusion weights that change over time in combination with the uncertainty characterization of each scale; using the fusion weights to perform weighted operations on the risk trend sequences of each scale to obtain the initial fusion risk trend sequence, and calculating the upper and lower boundaries of the confidence interval corresponding to the initial fusion risk trend sequence according to the predetermined fusion rules based on the uncertainty characterization of each scale; applying cross-scale consistency constraints to the initial fusion risk trend sequence and its confidence interval, so that the fusion risk value at the same time is limited to the value range or change rate threshold of the risk trend sequence of each scale; when the cross-scale consistency constraints are violated, the final fusion risk trend curve and its confidence interval are obtained by adjusting the corresponding fusion weights or truncating the adjusted fusion risk value and the confidence interval boundary.

[0021] As a preferred embodiment of the present invention, the extraction of key risk trend events includes: calculating the rate of change sequence and the rate of change of the rate of change sequence of the fused risk trend curve according to the time step; determining the trend inflection point and the peak value of the risk growth rate based on the change of the sign of the rate of change or the extreme point of the rate of change; finding the time point when the fused risk value first reaches or exceeds a preset risk threshold and the time point when it last falls below the risk threshold from the fused risk trend curve; determining the continuous time interval between the two time points as the risk cross-level time window; and encoding the trend inflection point, the peak value of the risk growth rate, and the risk cross-level time window as key risk trend events and outputting them.

[0022] A multi-timescale risk trend prediction system for landslide disasters includes:

[0023] Data acquisition module: Acquires multi-source data of target potential hazard points and performs time stamp unification, spatial correlation and quality identification to form a standard dataset; generates a multi-scale feature set based on the standard dataset;

[0024] Stage determination module: Determines the evolution stage in real time using an evolution stage identification model based on a multi-scale feature set, and outputs stage identification information;

[0025] Scale adjustment module: Dynamically determines time scale division parameters based on stage identifier information and adjusts multi-scale feature sets;

[0026] Prediction constraint module: Constructs a spatiotemporal constraint propagation network, establishes bidirectional constraints between trend predictors at each scale, and sets constraint propagation strength parameters according to stage identification information;

[0027] The segmented prediction module inputs the adjusted multi-scale feature set into the corresponding segmented trend predictor and outputs the risk trend sequence and its uncertainty characterization at each scale under the bidirectional constraints of the spatiotemporal constraint propagation network.

[0028] Risk fusion module: The dynamic weight allocator determines the fusion weights based on stage identification information and uncertainty characterization and applies cross-scale consistency constraints to obtain the fusion risk trend curve and its confidence interval.

[0029] Event Extraction Module: Extracts key risk trend events from the fused risk trend curve and outputs them.

[0030] The present invention has the following advantages:

[0031] This invention constructs a standard dataset by unifying the timestamps, spatially correlated, and quality-identified geological environment data, meteorological and hydrological data, and deformation response data. Based on this, a multi-scale feature set is generated, enabling collapse risk analysis to utilize both long-term background conditions and short-term triggering and deformation response information on a unified data benchmark, thereby improving the basic data quality and information completeness for risk trend prediction.

[0032] This invention establishes an evolution stage identification model to determine different evolution stages in real time for an evolution feature sequence composed of multi-scale feature sets, outputs stage identification information, and uses a scale adjustment module to dynamically determine the time window length, sliding step size, and scale division boundary based on the stage identification information, thereby adjusting the multi-scale feature set and realizing the adaptive change of time scale division with the collapse evolution stage.

[0033] This invention constructs a spatiotemporal constraint propagation network, using short-term trend predictors, medium-term trend predictors, and long-term background-scale trend predictors as constraint nodes, establishing bidirectional constraint channels between them, and selecting constraint weights and maximum adjustment amplitudes from the constraint propagation strength parameter table based on stage identification information. This enables each subscale trend predictor to transmit constraint information, correct input features, and initial states during the prediction process, achieving dynamic collaboration and process-level interaction in risk trend prediction at different time scales.

[0034] This invention outputs risk trend sequences and uncertainty representations at each scale in the multi-scale prediction module under the bidirectional constraints of the spatiotemporal constraint propagation network. Then, the dynamic weight allocator calculates the fusion weights that change over time by combining the stage identification information and uncertainty representations, and applies cross-scale consistency constraints to obtain a fusion risk trend curve with a confidence interval. This invention achieves adaptive weighting and constraint of the prediction results at each time scale according to uncertainty.

[0035] This invention identifies key events in risk trends by analyzing the rate of change and the rate of change of the fused risk trend curve. It transforms continuous risk trend information into several key events with clear time positions and duration intervals, providing management departments with specific key information nodes for risk evolution. Attached Figure Description

[0036] 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 in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0037] Figure 1This is a schematic diagram of the structure of a multi-timescale risk trend prediction system for landslide disasters, as used in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0039] Example 1: A multi-timescale risk trend prediction method for landslide disasters, comprising the following steps:

[0040] Step S1: Obtain multi-source data of target potential hazards and perform time stamp unification, spatial correlation and quality identification to form a standard dataset; generate a multi-scale feature set based on the standard dataset;

[0041] In this embodiment, the target hazard point is a landslide hazard point that has been included in the management list through geological surveys and disaster surveys in landslide-prone mountainous areas, such as a hazard point on a tourist highway slope. The multi-source data is used to characterize the geological background conditions, external triggering conditions, and actual deformation response behavior of the target hazard point. Specifically, it includes three categories: geological environment data, meteorological and hydrological data, and deformation response data. The multi-source data is written into data records containing a unified timestamp field and a spatial location field. The data records are then processed by unified timestamp processing, spatial correlation processing, and a quality identifier field is added to obtain a standard dataset.

[0042] The geological environment data includes basic information for characterizing the long-term stability and susceptibility of potential hazard points, including: topographic data, such as elevation, slope, aspect, slope height, slope length, topographic relief, and landform type within the target hazard point and its catchment area, derived from digital elevation models, remote sensing image interpretation results, and existing topographic survey data; and stratigraphic lithology and weathering data, including dominant lithology types, distribution of weak interlayers, degree of weathering, and range of rock mass mechanical parameters, derived from regional geological survey reports, engineering exploration data, and necessary indoor data. Test results; structural plane and tectonic data, including the number, attitude, spacing, and extension range of structural planes such as joints, bedding planes, and faults, as well as the distribution characteristics of tectonic fracture zones. The data are derived from field surveys, engineering geological profiles, and comprehensive interpretation results; human engineering activity data, including information on engineering disturbances that affect slope stability, such as road cutting, house construction, retaining wall construction, and drainage ditch excavation, such as the distance of the road from the slope toe, excavation height, form and integrity of protective structures, etc., derived from engineering design data, construction records, and supplementary on-site investigations.

[0043] The meteorological and hydrological data includes time-series information reflecting external triggering conditions, including: rainfall and rainfall intensity data, such as hourly rainfall, daily rainfall, total rainfall over a continuous process, maximum 1-hour rainfall, and maximum 3-hour rainfall, derived from observation records of automatic rain gauges and regional meteorological stations deployed near potential hazard points; cumulative rainfall and previous rainfall indicators, including cumulative rainfall at different time scales such as 3-day, 7-day, and 15-day periods, as well as cumulative rainfall during the flood season and the number of consecutive rainless days, obtained through time window statistics of the original rainfall sequence; temperature and freeze-thaw related data, including daily maximum temperature, daily minimum temperature, daily average temperature, number of times around 0°C, and number of freeze-thaw cycles, used to reflect the impact of seasonal freeze-thaw on slope strength, derived from meteorological stations or regional meteorological reanalysis data; and hydrological and saturation related indicators, including changes in water levels in nearby rivers or valleys, groundwater depth monitoring data, and soil and rock saturation indicators estimated based on rainfall and geological conditions, derived from hydrological station monitoring records or hydrogeological survey results.

[0044] The deformation response data includes monitoring data used to characterize the actual response behavior of target hazard points under load, including: displacement monitoring data, including three-dimensional displacement, slope-direction displacement, and displacement velocity and acceleration over time obtained from GNSS monitoring points, total station reflecting prisms, and inclinometers deployed on slopes or unstable rock masses, sourced from automated monitoring systems or periodic manual monitoring; tilt monitoring data, including tilt values ​​and tilt change rates recorded by inclinometers deployed on unstable rock masses, slopes, or retaining structures, used to identify attitude change processes; crack monitoring data, including crack opening amounts and their change rates recorded by crack gauges and fissure gauges, used to reflect the degree of opening and closing of rock masses and structural surfaces; and remote sensing deformation data, including surface displacement fields or key monitoring point displacement time series obtained based on InSAR deformation inversion or comparison of multi-period oblique photography three-dimensional models, used to supplement spatially sparse ground monitoring data.

[0045] When the aforementioned multi-source data is written into a unified data record structure, the data record shall include at least the following: a unified timestamp field, a spatial location field, a data type field, and a quality identifier field.

[0046] When unifying the timestamps of the aforementioned data records, the original time records from different sampling frequencies and data sources are converted into a unified time reference. The data is then aligned or interpolated according to a preset time step to ensure that the meteorological and hydrological data and deformation response data are consistent with the observation period of the target hazard point on the time axis. When spatially associating the data records, the regional-scale geological environment data is linked with the point-scale monitoring data based on the hazard point number and monitoring point coordinates in the spatial positioning field. This unifies and merges multi-source data from the same target hazard point and its upstream catchment area, forming a standard data record set organized by target hazard point. After timestamp unification and spatial association are completed, a quality identifier field is added to each data record to remove obviously erroneous or incomplete data. Records with missing or abnormal data are marked and preprocessed as necessary, resulting in a standard dataset for subsequent feature construction.

[0047] Based on standard datasets, meteorological and hydrological data and deformation response data are accumulated, sliding statistics and rate of change are calculated under different preset time windows and sampling steps to form short-term features for describing minute to hourly changes, medium-term features for describing day to weekly changes, and long-term background features for describing seasonal to flood season changes. The short-term, medium-term and long-term background features are then combined to form a multi-scale feature set.

[0048] In this embodiment, the short-term scale features are used to reflect the rapid response behavior of the target hazard point on a time scale of minutes to hours, such as statistical indicators such as rainfall, rainfall intensity, displacement increment, displacement velocity, displacement acceleration, dip angle change rate, and crack opening increment calculated within 5-minute, 15-minute, and 1-hour sliding time windows; the medium-term scale features are used to reflect the relationship between external triggering and cumulative effects on a time scale of days to weeks, such as cumulative rainfall over 1 day, 3 days, 7 days, and 15 days, number of consecutive rainy days, number of consecutive rainless days, daily average displacement velocity, daily average dip angle change rate, and weekly InSAR displacement change; the long-term background scale features are used to reflect the background susceptibility and environmental conditions on a time scale of seasons to flood seasons, such as cumulative rainfall during the flood season, number of freeze-thaw cycles, soil and rock saturation level within a long time window, seasonal deformation amplitude, and static or slowly changing indicators related to topography, stratigraphy, structural plane combination, and human engineering activities.

[0049] By performing cumulative, sliding statistics, and rate of change calculations on the standard dataset according to the different time windows and sampling steps mentioned above, meteorological and hydrological data and deformation response data are transformed into short-term, medium-term, and long-term background scale features. Combined with the static features in geological environmental data that reflect long-term stability and susceptibility, a multi-scale feature set covering multiple time scales from minutes to the flood season is constructed.

[0050] Step S2: Determine the evolution stage in real time using the evolution stage identification model to identify the multi-scale feature set and output stage identification information;

[0051] In this embodiment, the evolution stage identification model is used to identify the current instability evolution stage of a target hazard point based on its multi-timescale characteristics, and converts the identification results into stage identification information for subsequent steps. The evolution stage identification model is pre-trained based on long-term monitoring data of typical landslide hazard points and the results of post-landslide events. The training samples include a multi-scale feature set composed of geological environmental data, meteorological and hydrological data, and deformation response data, as well as evolution stage labels given by combining field investigations and expert interpretations. Through comparative analysis of stable periods, slow deformation periods, accelerated deformation periods, and short-term rapid deformation periods before landslides in historical cases, a classification relationship between multi-scale feature patterns and evolution stages is established and solidified into parameters for the evolution stage identification model to be used online.

[0052] The real-time determination of the evolution stage includes: arranging the short-term, medium-term, and long-term background scale features corresponding to the target hidden danger point in the multi-scale feature set into an evolution feature sequence according to time order; within a preset sliding time window, extracting evolution characterization indicators including displacement, displacement velocity, displacement acceleration, and tilt angle change rate from the evolution feature sequence; inputting the evolution characterization indicators into the evolution stage identification model for stage classification calculation to obtain the current evolution stage; the current evolution stage includes a stable creep stage, an accelerated creep stage, and a near-slip warning stage; encoding the current evolution stage as stage identification information and attaching the confidence level corresponding to the evolution stage.

[0053] In this embodiment, the evolutionary feature sequence is a multidimensional time series formed by splicing short-term, medium-term, and long-term background scale features in chronological order based on a standard dataset. Each time step in the evolutionary feature sequence includes long-term background scale features reflecting the geological environment of the target hazard point, medium-term scale features reflecting the external triggering conditions in the recent period, and short-term scale features reflecting the response behavior in the current and recently passed periods. For example, for a hazard point on a tourist highway slope, a single time step in the evolutionary feature sequence simultaneously includes: long-term background indicators such as cumulative rainfall during the flood season, number of freeze-thaw cycles, slope, and lithology; medium-term indicators such as cumulative rainfall in the past 7 days and the past 3 days, number of consecutive rainy days, and average daily displacement velocity; and short-term indicators such as rainfall intensity, displacement velocity, displacement acceleration, and rate of change of dip angle within the most recent 1-hour sliding window.

[0054] The preset sliding time window is used to extract local time segments reflecting stage characteristics from the evolutionary feature sequence. The length of the sliding time window is set according to the response characteristics of the target hazard point. For example, for rock collapse hazard points that are sensitive to rainfall, the sliding time window length is set to 6 hours or 24 hours to take into account both the rapid response caused by short-term heavy rainfall and the daily cumulative effect. For hazard points controlled by seasonal freeze-thaw cycles, the sliding time window length is further superimposed with multi-day scale feature extraction to comprehensively reflect the freeze-thaw cycle process. The sliding time window advances step by step on the time axis according to the set step size, and at each window position, evolutionary characterization indicators are extracted from the evolutionary feature sequence.

[0055] The evolutionary characterization indices are characteristic quantities used to describe the deformation evolution state of potential hazard points, including displacement, displacement velocity, displacement acceleration, and dip angle change rate, calculated in conjunction with the specific type of deformation response data. Displacement reflects the overall deformation degree during the monitoring period, displacement velocity reflects the deformation development rate, displacement acceleration depicts the increasing or decreasing trend of the deformation rate, and dip angle change rate reflects the speed of change in the attitude of the slope or unstable rock mass. For potential hazard points with crack monitoring or remote sensing deformation monitoring, the evolutionary characterization indices also include crack opening increment, crack opening change rate, and multi-source displacement difference indices, used to enhance the stability of stage identification. For each sliding time window, while summarizing the above evolutionary characterization indices, the corresponding meteorological and hydrological indices and long-term background indices are retained, so that the evolutionary stage identification model considers the coupling relationship between external triggering conditions and internal response behavior when classifying stages.

[0056] The evolution stage identification model is a multi-classification model, which maps the input evolutionary characterization indicators and their corresponding multi-scale features to one of a finite number of predefined stages. In this embodiment, the evolution stages include a stable creep stage, an accelerated creep stage, and a landslide warning stage. The stable creep stage corresponds to a state with low displacement velocity, near-zero displacement acceleration, and a small rate of change of dip angle, i.e., a period of slow deformation and stable trend; the accelerated creep stage corresponds to a state with relatively increased displacement velocity, positive displacement acceleration that persists for a certain period of time, and an increased rate of change of dip angle without abrupt changes, reflecting that the slope is in an accelerated deformation process; the landslide warning stage corresponds to a state with a sharp increase in displacement velocity, significantly positive displacement acceleration that rapidly amplifies in a very short time, and a significant increase in the rate of change of dip angle and the rate of change of crack opening, reflecting a high-risk stage where the slope is close to instability and landslide.

[0057] During training, the evolutionary stage identification model is trained by labeling and optimizing parameters on a large number of historical monitoring sequences and their corresponding actual evolutionary processes. In real-time judgment, for each sliding time window, the model reads the evolutionary characterization vector within that window, combines long-term background scale features and mid-term scale features to perform stage classification, outputs the current evolutionary stage, and assigns a confidence value to that stage. The confidence value is a quantitative indicator reflecting the reliability of the model's stage judgment, normalized from the results of classification probability or distance metric calculations. In this embodiment, the confidence value is limited to a range of 0 to 1; a value closer to 1 indicates a more stable judgment of the current evolutionary stage by the model.

[0058] After obtaining the current evolution stage, the evolution stage is encoded as stage identification information. The stage identification information includes at least the stage category code and the corresponding confidence level, and may be supplemented with auxiliary information to describe the duration of the stage, such as the length of time the stage has lasted, the statistical characteristics of displacement velocity and rainfall intensity within the stage, etc.

[0059] Step S3: Dynamically determine the time scale division parameters based on the stage identification information, and adjust the multi-scale feature set;

[0060] In this embodiment, the stage identification information is used to indicate the real-time status of the target potential hazard point during the collapse evolution process. The time scale division parameters are used to describe the division method of short-term, medium-term, and long-term background scales on the time axis, and are a set of control parameters used when dividing the continuous time series in the standard dataset into several time scale levels.

[0061] The adjustment of the multi-scale feature set includes: mapping stage identification information to a set of time scale division parameters corresponding to each evolution stage. The time scale division parameters include the time window length, sliding step size, and scale division boundary used to define short-term, medium-term, and long-term background scales. When the stage identification information changes, the time scale division parameters corresponding to the current evolution stage are selected from the time scale division parameter set. The meteorological and hydrological data and deformation response data in the standard dataset are recalculated using the updated time window and sliding step size to obtain updated short-term, medium-term, and long-term background scale features. Based on the updated scale division boundary, the features in the original multi-scale feature set are redistributed or discarded to form the adjusted multi-scale feature set.

[0062] In this embodiment, the time scale division parameter set is pre-set based on a large amount of historical monitoring data and the evolution characteristics of typical collapse cases before the method is deployed. For the stable creep stage, the accelerated creep stage, and the imminent collapse warning stage, corresponding time scale division parameter subsets are constructed. Each subset includes at least: a combination of time window length and sliding step size for the short-term imminent scale, a combination of time window length and sliding step size for the intermediate scale, a combination of time window length and sliding step size for the long-term background scale, and a scale division boundary used to distinguish the boundaries between the short-term, intermediate, and long-term background scales on the time axis.

[0063] For example, in the timescale parameter subset for the stable creep stage, the short-term ad hoc scale has a relatively long time window and a relatively large sliding step size to smooth out short-term random disturbances; the medium-term scale has a time window covering the past few days to several weeks to highlight the cumulative effects of low and medium frequencies; and the long-term background scale has a time window covering the entire flood season or season to maintain the characterization of background susceptibility. In the timescale parameter subset for the accelerated creep stage, the short-term ad hoc scale has a shorter time window and a tighter sliding step size to achieve higher temporal resolution for changes in displacement velocity and acceleration over short periods; the medium-term scale emphasizes the linkage between rainfall and deformation over the past few days; and the long-term background scale still maintains seasonal or flood season-level statistics, but its weight is relatively reduced in subsequent fusion. In the time scale division parameter subset of the imminent landslide warning stage, the short imminent scale time window is further shortened to the level of several minutes to one hour, and the sliding step size is close to the original sampling frequency of the monitoring data to capture the rapid changes in displacement and tilt angle; the medium scale focuses on the evolution process several hours to several days before the imminent landslide; the long-term background scale is mainly used to constrain the overall risk level, rather than dominating the short-term prediction results.

[0064] The scale division boundary is used to give the segmentation position of short-term scale, medium-term scale and long-term background scale on the time axis. For example, the segmentation is defined as "minute to hour" as short-term scale, "day to week" as medium-term scale and "season to flood season" as long-term background scale.

[0065] When the stage identification information changes, such as switching from a stable creep stage to an accelerated creep stage, or from an accelerated creep stage to a slippage warning stage, the method retrieves the subset of time scale division parameters corresponding to the current evolution stage from the time scale division parameter set. To avoid frequent time scale switching caused by short-term fluctuations in stage identification, a stage duration threshold or confidence threshold is introduced when selecting time scale division parameters. The time scale division parameters are only updated when the duration of the current evolution stage exceeds a predetermined threshold or the confidence of the current stage is higher than a predetermined threshold.

[0066] After selecting the time scale division parameters, the meteorological and hydrological data and deformation response data in the standard dataset are recalculated using cumulative, sliding statistics, and rate of change under the updated time window and sliding step size. During this process, the meteorological and hydrological data still include rainfall, rainfall intensity, cumulative rainfall, number of consecutive rainy days, freeze-thaw cycle indicators, and water level and saturation-related indicators. The deformation response data still includes displacement, displacement velocity, displacement acceleration, dip angle change rate, crack opening amount and its change rate, etc., but their statistical methods and time aggregation scales are adjusted according to the new time scale division parameters. For example, in the accelerated creep stage, a shorter sliding time window and a denser sliding step size are used for rainfall and displacement increments in the most recent few hours to highlight short-term trends; in the stable creep stage, a longer sliding time window and a wider sliding step size are used for similar data to mitigate the impact of occasional short-term fluctuations.

[0067] After obtaining updated short-term, medium-term, and long-term background scale features, the features in the original multi-scale feature set are redistributed or discarded based on the updated scale boundaries. Feature redistribution refers to merging features originally belonging to a certain time scale to other time scales according to the new scale boundaries. For example, displacement increments originally calculated on a daily scale are reclassified as short-term or medium-term scale features after the update. Feature discarding refers to removing features that contribute little to risk trend judgment at the current evolution stage and are highly redundant with other features from the multi-scale feature set to reduce redundant dimensions in subsequent prediction calculations and improve the efficiency and stability of trend prediction. During feature redistribution and discarding, static features reflecting long-term stability and susceptibility in geological environmental data remain unchanged and only participate in subsequent calculations as a basic part of the long-term background scale features.

[0068] Step S4: Construct a spatiotemporal constraint propagation network, establish bidirectional constraints between each subscale trend predictor, and set constraint propagation strength parameters according to stage identification information;

[0069] In this embodiment, a spatiotemporal constraint propagation network is used to establish constraint relationships between risk trend prediction results at different time scales, ensuring consistency and physical plausibility in the temporal evolution of risk trend sequences at short-term, medium-term, and long-term background scales. Structurally, the spatiotemporal constraint propagation network consists of several constraint nodes and constraint channels. Each constraint node corresponds to a subscale trend predictor, and each constraint channel is used to transmit constraint information between two constraint nodes. The subscale trend predictor is a time series prediction model built for a specific time scale, including a short-term trend predictor, a medium-term trend predictor, and a long-term background scale trend predictor. These predictors use short-term, medium-term, and long-term background scale features as inputs to predict risk trends at their respective time scales.

[0070] In practical applications, taking a potential hazard point on a tourist highway slope as an example, the short-term trend predictor predicts the risk trend sequence for the next few hours based on minute- to hourly rainfall intensity, displacement velocity, displacement acceleration, and slope change rate characteristics; the medium-term trend predictor predicts the risk trend sequence for the next few days based on the cumulative rainfall, number of consecutive rainy days, daily average displacement velocity, and InSAR daily-scale displacement change characteristics for the past few days to weeks; and the long-term background scale trend predictor predicts the risk background change trend over the entire flood season or seasonal scale based on long-term background scale characteristics such as cumulative rainfall during the flood season, number of freeze-thaw cycles, lithology, slope, structural surface combination, and human engineering activities. The spatiotemporal constraint propagation network transmits constraint information between the three types of trend predictors through constraint channels, ensuring that the risk trend prediction results at the three time scales remain consistent in terms of numerical magnitude, direction of change, and rate of change.

[0071] The bidirectional constraints include: in the spatiotemporal constraint propagation network, short-term trend predictors, medium-term trend predictors, and long-term background scale trend predictors are respectively used as constraint nodes; constraint channels are established between each constraint node, from the short-term trend predictor to the medium-term trend predictor and the long-term background scale trend predictor, as well as constraint channels from the medium-term trend predictor and the long-term background scale trend predictor to the short-term trend predictor; in each prediction step, the risk trend sequence and uncertainty characterization output by each subscale trend predictor are used as constraint information, and transmitted to the subscale trend predictors of other time scales through the corresponding constraint channels to update the input features, initial state, and prediction boundary conditions of the subscale trend predictors, and after the update, the time series prediction operation of the next prediction step is performed.

[0072] In the above definition, the risk trend sequence is a discrete-time risk value sequence given by a certain scale trend predictor within a predetermined prediction period. For example, a short-scale trend predictor gives an hourly risk index sequence within a 6-hour prediction period, and a medium-scale trend predictor gives a daily risk index sequence within a 7-day prediction period. Prediction uncertainty characterization is used to quantitatively describe the degree of uncertainty of the prediction results. In this embodiment, it is constructed using indicators such as prediction residual statistics, sample variance, or prediction confidence interval width to indicate the reliability of prediction results at different time scales for subsequent constraint propagation.

[0073] Two-way constraints refer to a spatiotemporal constraint propagation network that simultaneously includes constraint channels transmitting constraint information from short-term to medium- and long-term scales, and constraint channels transmitting constraint information from medium- and long-term scales to short-term scales. Specifically, the constraint channels from the short-term trend predictor to the medium-term and long-term background scale trend predictors are used to propagate rapidly rising risk trends observed in the short term to the medium- and long-term scales. For example, during a near-slippage warning phase, if the short-term trend predictor forecasts a significant increase in the risk index within the next few hours, the constraint channels influence the medium-term trend predictor's assessment of the risk level for the next few days, causing the medium-term prediction result to rise accordingly in the short term. Conversely, the constraint channels from the medium-term and long-term background scale trend predictors to the short-term trend predictor are used to feed back reasonable risk ranges or background trends from the medium- and long-term scales to the short-term scale. During stable creep phases or when the background risk level is low, these channels constrain occasional outliers in the short-term predictions, preventing short-term noise from excessively amplifying the short-term risk trend.

[0074] In each prediction step, the spatiotemporal constraint propagation network calculates and transmits constraint information in a predetermined order. Specifically, firstly, the short-term trend predictor, medium-term trend predictor, and long-term background scale trend predictor calculate their respective risk trend values ​​and corresponding uncertainty representations for the next prediction step based on the input characteristics and internal state at the current moment. Subsequently, the risk trend sequence fragments and uncertainty representations output by the three types of subscale trend predictors at the same moment are combined into a constraint information package and input to the corresponding target subscale trend predictor according to the constraint channel direction. After receiving the constraint information, each subscale trend predictor corrects its own input characteristics, initial state, and prediction boundary conditions according to the preset constraint propagation rules. For example, when the short-term trend predictor predicts that the risk value in the next hour is significantly higher than the upper limit of the flood season risk level predicted by the long-term background scale, the medium-term trend predictor, after receiving the short-term constraint information, raises the risk benchmark for the next few days to a certain extent; at the same time, after receiving the constraint information from the long-term background scale, the short-term trend predictor limits its own predicted extreme peak values ​​to ensure that they do not exceed the reasonable risk range of the hidden danger point constrained by the long-term background scale.

[0075] The correction of input features includes weight adjustments or shifts to some feature values. For example, introducing upward trend information from short-term timescales to the "recent risk baseline" feature in the medium-term timescale, and introducing stable level information from the long-term background timescale to the "long-term risk background" feature in the short-term timescale. The correction of the initial state includes shifting the hidden states, memory unit states, or state variables within the subscale trend predictor, so that the prediction model reflects the constraints of other timescales when starting the next prediction step. The adjustment of prediction boundary conditions includes updating constraints such as the upper and lower limits of risk prediction and the risk change rate threshold, so that the time series prediction operation is performed under the new constraints. Through the joint correction of the above input features, initial state, and prediction boundary conditions, each prediction step generates a new risk trend prediction result under the bidirectional constraints of multi-timescale information.

[0076] The constraint propagation strength parameters include: determining the current evolution stage based on stage identification information; selecting a set of constraint propagation strength parameters corresponding to the current evolution stage from a preset constraint propagation strength parameter table; including a first constraint weight coefficient for transmitting constraint information from the short-term trend predictor to the medium-term trend predictor and the long-term background scale trend predictor; a second constraint weight coefficient for transmitting constraint information from the medium-term trend predictor and the long-term background scale trend predictor to the short-term trend predictor; and a maximum adjustment range for limiting the input feature correction, initial state offset, and prediction boundary condition adjustment caused by the constraint information; when the spatiotemporal constraint propagation network performs constraint information transmission, the constraint information in each constraint channel is multiplied by the corresponding first constraint weight coefficient or second constraint weight coefficient, and the input feature correction, initial state offset, and prediction boundary condition adjustment are truncated according to the maximum adjustment range.

[0077] In this embodiment, the constraint propagation strength parameter table is pre-constructed during the method deployment phase, and statistical analysis and parameter calibration are performed based on a large amount of monitoring data of historical landslide hazard points and the results of evolution stage identification. For the stable creep stage, the first constraint weight coefficient given in the constraint propagation strength parameter table has a smaller value, while the second constraint weight coefficient has a relatively larger value, thereby enhancing the constraint effect of the medium-term and long-term background scales on the short-term scale and suppressing abnormal fluctuations in risk prediction caused by short-term noise. For the accelerated creep stage, the first constraint weight coefficient and the second constraint weight coefficient have relatively close values, so that the short-term scale and the medium- and long-term scales maintain a balance in mutual constraints. For the landslide warning stage, the first constraint weight coefficient has a larger value, while the second constraint weight coefficient has a relatively smaller value, so that the rapid upward risk trend output by the short-term trend predictor is fully reflected in the medium-term and long-term background scales, while the stable trend of the long-term background does not exert excessive suppression on the sharp rise in the short-term scale.

[0078] The maximum adjustment range is used to control the upper bound of the correction intensity caused by constraint propagation, ensuring that the input feature correction, initial state offset, and prediction boundary condition adjustment are within an acceptable range in each prediction step. By pre-setting different maximum adjustment ranges for different evolution stages, the correction intensity caused by constraint propagation in any single prediction step is limited during the stable creep stage, the correction range is moderately relaxed during the accelerated creep stage, and the correction range related to short-term critical scales is further relaxed during the near-slip warning stage, so that the constraint propagation intensity matches the risk level of the evolution stage. When the spatiotemporal constraint propagation network performs constraint information transmission, it first multiplies the constraint information in each constraint channel by the corresponding first constraint weight coefficient or second constraint weight coefficient, then compares the calculated input feature correction, initial state offset, and prediction boundary condition adjustment with the preset maximum adjustment range, and truncates the portion exceeding the maximum adjustment range to form the final effective correction value.

[0079] Step S5: Input the adjusted multi-scale feature set into the corresponding subscale trend predictor, and under the bidirectional constraint of the spatiotemporal constraint propagation network, output the risk trend sequence and its uncertainty characterization at each scale.

[0080] The output of the risk trend sequence and its uncertainty representation includes: taking the short-term, medium-term, and long-term background scale features from the adjusted multi-scale feature set as input features for the short-term, medium-term, and long-term background scale trend predictors, respectively; under the bidirectional constraint of the spatiotemporal constraint propagation network, each subscale trend predictor reads the stage identification information and selects the model parameter set corresponding to the current evolution stage; performs time series prediction operations on the input features to obtain the risk trend sequence at the corresponding time scale; and calculates the uncertainty representation of the risk trend sequence based on the prediction residuals and sample variance.

[0081] In this embodiment, the segmented trend predictor is used to predict the change in risk level of a target hazard point over time within a given time scale. The short-term trend predictor, medium-term trend predictor, and long-term background trend predictor correspond to time scales of minutes to hours, days to weeks, and seasons to flood seasons, respectively, and are used to output risk trend sequences matching their respective time scales. The risk trend sequence refers to a set of discrete risk values ​​arranged at fixed time steps within a preset prediction period, used to characterize the evolution trajectory of the risk of the target hazard point over time, such as a future 6-hour risk index sequence arranged by hour, or a future 7-day risk index sequence arranged by day. Uncertainty characterization refers to a set of quantitative indicators used to describe the reliability of each predicted value in the risk trend sequence, such as prediction error variance, prediction confidence interval width, or uncertainty score, used to provide a basis for weight allocation and consistency constraints during subsequent cross-scale fusion.

[0082] In the specific implementation, the adjusted multi-scale feature set is classified according to time scale and then input into the corresponding sub-scale trend predictor. Short-term features include minute to hourly rainfall, rainfall intensity, short-term cumulative rainfall, displacement increment, displacement velocity, displacement acceleration, dip angle change rate, and crack opening increment, calculated under the update time window and sliding step size. These features are used to characterize the triggered loading and deformation response of the target hazard point on a short time scale. Medium-term features include cumulative rainfall over the past few days to several weeks, number of consecutive rainy days, number of consecutive rainless days, daily average displacement velocity, daily average dip angle change rate, and InSAR daily-scale displacement change. These features are used to characterize the cumulative effect and mid-frequency deformation pattern on a day to week time scale. Long-term background scale features include cumulative rainfall during the flood season, number of freeze-thaw cycles, lithology, slope, slope height, structural surface combination, human engineering activity intensity index, and long-term deformation amplitude. These features are used to characterize the risk background level and long-term evolution trend on a seasonal to flood season time scale.

[0083] After receiving short-term characteristics, the short-term trend predictor, under the bidirectional constraints of the spatiotemporal constraint propagation network, selects a set of model parameters corresponding to the current evolution stage from its internal parameter library. Here, the model parameter set is a set of parameters trained and solidified separately according to different evolution stages based on historical monitoring data before method deployment. For example, for the stable creep stage, the short-term trend predictor uses a parameter set that limits short-term random fluctuations, resulting in a smoother response to short-term rainfall and minor deformation disturbances; for the accelerated creep stage, it uses a parameter set more sensitive to changes in displacement velocity and acceleration to highlight the continuous upward trend of risk in the short term; for the imminent slippage warning stage, it uses a parameter set highly sensitive to rapid changes in a short period, ensuring that the prediction results show a significant increase in risk before approaching instability. After parameter selection, the short-term trend predictor uses time series prediction methods to progressively predict risk values ​​for the next few hours, forming a risk trend sequence corresponding to the short-term time scale.

[0084] After receiving the mesoscale features, the mesoscale trend predictor also reads the stage identification information and selects a set of model parameters that matches the current evolution stage. The mesoscale trend predictor primarily focuses on risk changes over the next few days to weeks. In the stable creep stage, parameter settings tend to reflect the slow coupling relationship between rainfall and displacement; in the accelerated creep stage, parameter settings emphasize the linkage between recent rainfall and accelerated deformation; and in the imminent landslide warning stage, parameter settings enable the mesoscale prediction results to respond to the sharp uplift trend at the short-term imminent scale. After selecting the parameters, the mesoscale trend predictor performs time series predictions based on the mesoscale features, outputting a risk trend sequence arranged by daily or multi-day steps.

[0085] After receiving long-term background scale characteristics, the long-term background scale trend predictor focuses on using geological environmental data and seasonal meteorological and hydrological characteristics to predict the risk background for the future flood season or the entire season. During the stable creep stage, the risk trend sequence output by the long-term background scale trend predictor typically exhibits a slow change pattern, mainly determined by factors such as lithology, slope, structural surface combination, intensity of human engineering activities, and cumulative rainfall during the flood season. During the accelerated creep stage and the imminent landslide warning stage, the long-term background scale trend predictor has a limited response at the parameter level to significant short-term risk increases, serving more as a background constraint to limit the reasonable range of medium-term and short-term risk predictions.

[0086] Under the bidirectional constraints of the spatiotemporal constraint propagation network, each subscale trend predictor reads constraint information from other subscale trend predictors before performing time series prediction operations at each prediction step, and adjusts its own input features, internal initial state, and prediction boundary conditions. After absorbing constraint information and correcting its internal state, the subscale trend predictor performs time series prediction operations for the current prediction step, outputs the risk prediction value corresponding to this time step, and updates its internal state, providing initial conditions for the prediction of the next time step. After multiple iterations, the short-term trend predictor, the medium-term trend predictor, and the long-term background scale trend predictor each form a risk trend sequence covering their respective prediction durations.

[0087] After the risk trend sequence is generated, to quantify the uncertainty of the prediction results, uncertainty characterization calculations are performed on the risk trend sequences output by each subscale trend predictor. In this embodiment, the uncertainty characterization is constructed based on historical prediction residuals and sample variance. Specifically, during the method deployment phase, historical monitoring data and known real risk evolution processes are used to statistically analyze the prediction errors of prediction models at different evolution stages for each time scale, obtaining the prediction error distribution characteristics corresponding to each time scale, time step, and evolution stage. During the method operation, the current prediction results are combined with historical error statistics to assign uncertainty indicators to each prediction time step. For example, the variance of historical residuals is used as a measure of uncertainty of the risk prediction value at that time step, or a confidence interval is formed by constructing upper and lower bounds of risk values ​​containing preset confidence levels based on the error distribution.

[0088] The short-term trend predictor outputs risk trend sequences that include not only hourly or higher time-resolution risk predictions, but also uncertainty indicators corresponding to each prediction, such as variance estimates or confidence intervals. The medium-term and long-term background-scale trend predictors also output risk trend sequences and their uncertainty characteristics in the same way. For the application scenario of potential hazard points on tourist highway slopes, the short-term trend predictor outputs hourly risk sequences and corresponding confidence intervals for the next 6 hours, the medium-term trend predictor outputs daily risk sequences and corresponding confidence intervals for the next 7 days, and the long-term background-scale trend predictor outputs risk background sequences divided by week or ten-day periods for the remaining flood season, along with corresponding uncertainty indicators.

[0089] Step S6: The dynamic weight allocator determines the fusion weights based on the stage identification information and uncertainty characterization and applies cross-scale consistency constraints to obtain the fusion risk trend curve and its confidence interval.

[0090] In this embodiment, the dynamic weight allocator is a weight allocation unit constructed based on rules and data statistics. Its input includes stage identification information and risk trend sequences and their uncertainty characteristics at each time scale. The output is the fusion weight assigned to the risk trend sequences at the short-term, medium-term, and long-term background scales at each fusion time step. Cross-scale consistency constraints are used to restrict the relationship between risk trend values ​​at different time scales at the same moment during the fusion operation.

[0091] The acquisition of the fusion risk trend curve and its confidence interval includes: a dynamic weight allocator selecting preset basic weight coefficients for different evolution stages based on stage identification information, and calculating the fusion weights that change over time by combining the uncertainty characterization of each scale; using the fusion weights to perform weighted operations on the risk trend sequences of each scale to obtain the initial fusion risk trend sequence, and calculating the upper and lower boundaries of the confidence interval corresponding to the initial fusion risk trend sequence according to the predetermined fusion rules based on the uncertainty characterization of each scale; applying cross-scale consistency constraints to the initial fusion risk trend sequence and its confidence interval, so that the fusion risk value at the same time is limited to the value range or change rate threshold of the risk trend sequence of each scale; when the cross-scale consistency constraints are violated, the final fusion risk trend curve and its confidence interval are obtained by adjusting the corresponding fusion weights or truncating the adjusted fusion risk value and the confidence interval boundary.

[0092] In its implementation, the dynamic weight allocator first selects pre-defined base weight coefficients for different evolution stages based on stage identification information. These base weight coefficients are pre-calibrated before method deployment based on historical monitoring data of typical landslide hazard points and the results of evolution stage division. For example, for the stable creep stage, the base weight coefficients have a larger weight for the long-term background scale, a medium weight for the medium-term scale, and a smaller weight for the short-term critical scale, to reflect the dominant role of the long-term background state and the medium-term cumulative effect on the overall risk level in this stage; for the accelerated creep stage, the base weight coefficients have higher weights for the short-term critical scale and the medium-term scale, and a relatively lower weight for the long-term background scale, to highlight the impact of short-term rainfall processes and recent accelerated deformation on risk; for the landslide warning stage, the base weight coefficients have the highest weight for the short-term critical scale, a medium weight for the medium-term scale, and the lowest weight for the long-term background scale, to ensure that rapid deformation signals in a short period of time are fully reflected in the fusion results.

[0093] After selecting the basic weight coefficients, the dynamic weight allocator further incorporates the uncertainty characteristics of each time scale to calculate the fusion weights that change over time. For a given fusion time step, if the uncertainty characteristic of the short-term risk trend sequence is low at that time step, meaning the prediction result is relatively stable, then the weights allocated to the short-term scale at that time step are increased relative to the basic weight coefficients. Conversely, if the uncertainty characteristic of the medium-term or long-term background scale is high at certain time steps, then the weights allocated to these time scales at the corresponding time steps are decreased relative to the basic weight coefficients.

[0094] After obtaining the time-varying fusion weights, they are matched with the risk trend sequences at each time scale. For a given fusion time step, the short-term, medium-term, and long-term background scales each provide their own risk trend values, and through the aforementioned time axis alignment and interpolation processing, the three types of risk trend values ​​correspond to the same fusion moment. The dynamic weight allocator assigns specific fusion weights to each time scale at this fusion time step, and uses the fusion weights to perform a weighted calculation on the three types of risk trend values ​​to obtain the initial fusion risk value corresponding to this fusion time step. After traversing all fusion time steps along the time axis, the initial fusion risk values ​​constitute the initial fusion risk trend sequence.

[0095] Based on the uncertainty characteristics at each time scale, the upper and lower boundaries of the confidence interval corresponding to the initial fusion risk trend sequence are calculated according to a predetermined fusion rule. In this embodiment, the predetermined fusion rule comprehensively considers the statistical characteristics of prediction errors at each time scale and the current fusion weight allocation when constructing the confidence interval. For example, at a certain fusion time step, if the short-term scale weight is large and the short-term scale uncertainty characteristic value is small, the dependence of the upper and lower boundaries of the fusion confidence interval on the short-term scale confidence interval increases; conversely, if the uncertainty characteristic value of the medium-term scale is relatively large at a certain fusion time step, the dependence of the fusion confidence interval on the medium-term scale confidence interval decreases within that time step.

[0096] After obtaining the initial fusion risk trend sequence and its confidence interval, cross-scale consistency constraints are applied. In these constraints, the fusion risk value at the same fusion time step must satisfy the constraint relationship with the risk trend values ​​at each time scale. One type of constraint is the range constraint, which requires the fusion risk value to be within a reasonable range of the risk trend values ​​at each time scale, avoiding the fusion risk value deviating far from the prediction results at all time scales. Another type of constraint is the rate of change constraint, which requires the rate of change of the fusion risk trend to not exceed the rate of change threshold given in the risk trend sequence at each time scale, preventing abrupt changes in the fusion curve that do not conform to the laws of physical evolution.

[0097] When the initial fusion risk trend sequence and its confidence interval are detected to violate cross-scale consistency constraints at a certain time step or time period, two types of methods are used for correction. One method involves adjusting the corresponding fusion weights: reducing the fusion weights of time scales that violate the constraints at that time step and increasing the fusion weights of time scales that comply with the constraints at that time step, and recalculating the fusion risk value and the upper and lower boundaries of the confidence interval. The other method involves truncating the fusion risk value and its upper and lower boundaries of the confidence interval while keeping the weights unchanged, ensuring that they do not exceed a pre-set value range or rate of change threshold. After correction using one or both of these methods, the final fusion risk trend curve and its confidence interval that satisfy the cross-scale consistency constraints are obtained.

[0098] Step S7: Extract key risk trend events from the fused risk trend curve and output them.

[0099] In this embodiment, key risk trend events refer to time-related information reflecting important stage transitions, changes in risk growth rate, and risk level crossing processes during risk evolution. These include three types of events: trend inflection points, peak risk growth rates, and risk level crossing time windows. A trend inflection point refers to the position where the integrated risk trend curve changes from an upward trend to a downward trend or from a slow rise to a rapid rise. A peak risk growth rate refers to the point in time when the risk change rate reaches its local maximum. A risk level crossing time window refers to the continuous time interval between when the integrated risk value falls below a preset risk threshold, reaches or exceeds that threshold, and then falls below that threshold again. The preset risk threshold is set based on the risk classification standards and engineering management requirements of the target hazard point. For example, by combining historical disaster data and expert experience, the risk index is divided into multiple levels: general risk, relatively high risk, and high risk, and a corresponding threshold is configured for each level.

[0100] The extraction of key risk trend events includes: calculating the rate of change sequence and the rate of change of the rate of change sequence of the fused risk trend curve according to the time step; determining the trend inflection point and the peak of the risk growth rate based on the change of the sign of the rate of change or the extreme point of the rate of change; finding the time point when the fused risk value first reaches or exceeds the preset risk threshold and the time point when it last falls below the risk threshold from the fused risk trend curve; determining the continuous time interval between the two time points as the risk cross-level time window; and encoding the trend inflection point, the peak of the risk growth rate, and the risk cross-level time window as key risk trend events and outputting them.

[0101] In the specific implementation, the fused risk trend curve is first discretized at a uniform time step to obtain a fused risk value sequence arranged by hour or day. The time step is consistent with the time resolution of the aforementioned short-term and medium-term risk trend sequences. For example, in the application scenario of potential hazards on tourist highway slopes, an hourly time step is selected to accurately depict the risk changes caused by rainfall and deformation response. Subsequently, the difference between adjacent time steps is calculated on the discretized fused risk value sequence to obtain a risk change rate sequence, which is used to characterize the increase or decrease trend and rate of change of risk values ​​over time. Based on this, the difference between adjacent time steps is calculated again on the risk change rate sequence to obtain a change rate of rate sequence, which is used to characterize the increase or decrease of the rate of change of risk itself.

[0102] After obtaining the risk change rate sequence and the rate of change of change sequence, trend inflection points and risk growth rate peaks are determined based on changes in the sign of the change rate or extreme points of the change rate. Specifically, when the risk change rate changes from positive to negative or from negative to positive, the corresponding time step represents a shift in the risk trend from rising to falling or from falling to rising, and the time point corresponding to this position is marked as the trend inflection point. When the rate of change of change changes from positive to negative and the risk change rate is positive, the corresponding time step represents a shift in the risk growth rate from accelerating to decelerating, and the time point corresponding to this position is marked as the risk growth rate peak. In the early warning stage of a risk crisis, trend inflection points and risk growth rate peaks typically occur in the early stages of a rapid increase in risk and near the peak. These events provide direct evidence for judging key acceleration and deceleration phases in the risk evolution process.

[0103] After identifying the trend inflection point and the peak of the risk growth rate, the time point when the fused risk value first reaches or exceeds a preset risk threshold and the time point when it last falls below that threshold are found from the fused risk trend curve. To accommodate different management needs, the preset risk threshold is configured with multiple levels, such as a general risk threshold, a higher risk threshold, and a high risk threshold. For any given risk threshold, the time series corresponding to the fused risk trend curve is traversed, and the time point when the first fused risk value reaches or exceeds the threshold is recorded as the start time of the risk level transition; continuing to traverse forward, the last time point before the risk value falls below the threshold again is recorded as the end time of the risk level transition. The continuous time interval between the above start time and end time is defined as the risk level transition time window corresponding to that risk threshold. The risk level transition time window is used to describe the time range during which a target hazard point remains in a high-risk state above a certain risk level.

[0104] After identifying trend inflection points, peak risk growth rates, and risk-crossing time windows, the above information is encoded as key risk trend events and output. The encoded content includes at least: event type (trend inflection point, peak risk growth rate, or risk-crossing time window), associated time information (single point in time or time interval), associated risk value information (e.g., risk value at the inflection point, risk growth rate at the peak, or maximum risk value within the risk-crossing time window), associated risk level information (e.g., whether it exceeds the thresholds for general risk, higher risk, or high risk), and uncertainty indicators related to event identification (e.g., the event time positioning error range or the confidence interval for the associated risk value).

[0105] Example 2: A multi-timescale risk trend prediction system for landslide disasters (see [link]). Figure 1 As shown, it includes the following modules:

[0106] Data acquisition module: Acquires multi-source data of target potential hazard points and performs time stamp unification, spatial correlation and quality identification to form a standard dataset; generates a multi-scale feature set based on the standard dataset;

[0107] Stage determination module: Determines the evolution stage in real time using an evolution stage identification model based on a multi-scale feature set, and outputs stage identification information;

[0108] Scale adjustment module: Dynamically determines time scale division parameters based on stage identifier information and adjusts multi-scale feature sets;

[0109] Prediction constraint module: Constructs a spatiotemporal constraint propagation network, establishes bidirectional constraints between trend predictors at each scale, and sets constraint propagation strength parameters according to stage identification information;

[0110] The segmented prediction module inputs the adjusted multi-scale feature set into the corresponding segmented trend predictor and outputs the risk trend sequence and its uncertainty characterization at each scale under the bidirectional constraints of the spatiotemporal constraint propagation network.

[0111] Risk fusion module: The dynamic weight allocator determines the fusion weights based on stage identification information and uncertainty characterization and applies cross-scale consistency constraints to obtain the fusion risk trend curve and its confidence interval.

[0112] Event Extraction Module: Extracts key risk trend events from the fused risk trend curve and outputs them.

[0113] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-timescale risk trend prediction method for landslide disasters, characterized in that, include: Acquire multi-source data of target potential hazards and perform time stamp unification, spatial correlation and quality identification to form a standard dataset; Generate multi-scale feature sets based on standard datasets; The evolution stage identification model determines the evolution stage in real time using a multi-scale feature set and outputs stage identification information. The time scale division parameters are dynamically determined based on the stage identification information, and the multi-scale feature set is adjusted accordingly; Construct a spatiotemporal constraint propagation network, establish bidirectional constraints among trend predictors at each scale, and set constraint propagation strength parameters according to stage identification information; The adjusted multi-scale feature sets are input into the corresponding subscale trend predictors. Under the bidirectional constraints of the spatiotemporal constraint propagation network, the risk trend sequences and their uncertainty representations at each scale are output. The dynamic weight allocator determines the fusion weights based on stage identification information and uncertainty characterization and applies cross-scale consistency constraints, thereby obtaining the fusion risk trend curve and its confidence interval. The acquisition of the fusion risk trend curve and its confidence interval includes: a dynamic weight allocator selecting preset basic weight coefficients for different evolution stages based on stage identification information, and calculating the fusion weights that change over time by combining the uncertainty characterization of each scale; using the fusion weights to perform weighted operations on the risk trend sequences of each scale to obtain the initial fusion risk trend sequence, and calculating the upper and lower boundaries of the confidence interval corresponding to the initial fusion risk trend sequence according to the predetermined fusion rules based on the uncertainty characterization of each scale; applying cross-scale consistency constraints to the initial fusion risk trend sequence and its confidence interval, wherein the cross-scale consistency constraints include limiting the fusion risk value at the same time to the value range or change rate threshold of the risk trend sequence of each scale, and when the cross-scale consistency constraints are violated, adjusting the corresponding fusion weights or truncating the adjusted fusion risk value and the confidence interval boundary to obtain the final fusion risk trend curve and its confidence interval; Extract key risk trend events from the integrated risk trend curve and output them.

2. The method for predicting risk trends of landslide disasters at multiple time scales according to claim 1, characterized in that, The generation of the multi-scale feature set includes: the multi-source data including geological environment data, meteorological and hydrological data, and deformation response data; writing the multi-source data into data records containing a unified timestamp field and a spatial positioning field; performing unified timestamp processing and spatial correlation processing on the data records and adding a quality identifier field to obtain a standard dataset; based on the standard dataset, performing cumulative, sliding statistics, and rate of change calculations on the meteorological and hydrological data and deformation response data under preset different time windows and sampling steps to form short-term features for characterizing minute- to hourly changes, medium-term features for characterizing day- to weekly changes, and long-term background scale features for characterizing seasonal to flood season changes; and combining the short-term features, medium-term features, and long-term background scale features to form a multi-scale feature set.

3. The method for predicting risk trends of landslide disasters at multiple time scales according to claim 2, characterized in that, The real-time determination of the evolution stage includes: arranging the short-term, medium-term, and long-term background scale features corresponding to the target hidden danger point in the multi-scale feature set into an evolution feature sequence according to time order; within a preset sliding time window, extracting evolution characterization indicators including displacement, displacement velocity, displacement acceleration, and tilt angle change rate from the evolution feature sequence; inputting the evolution characterization indicators into the evolution stage identification model for stage classification calculation to obtain the current evolution stage; the current evolution stage includes a stable creep stage, an accelerated creep stage, and a near-slip warning stage; encoding the current evolution stage as stage identification information and attaching the confidence level corresponding to the evolution stage.

4. The method for predicting risk trends of landslide disasters at multiple time scales according to claim 2, characterized in that, The adjustment of the multi-scale feature set includes: mapping stage identification information to a set of time scale division parameters corresponding to each evolution stage. The time scale division parameters include the time window length, sliding step size, and scale division boundary used to define short-term, medium-term, and long-term background scales. When the stage identification information changes, the time scale division parameters corresponding to the current evolution stage are selected from the time scale division parameter set. The meteorological and hydrological data and deformation response data in the standard dataset are recalculated using the updated time window and sliding step size to obtain updated short-term, medium-term, and long-term background scale features. Based on the updated scale division boundary, the features in the original multi-scale feature set are redistributed or discarded to form the adjusted multi-scale feature set.

5. The method for predicting risk trends of landslide disasters at multiple time scales according to claim 4, characterized in that, The output of the risk trend sequence and its uncertainty representation includes: taking the short-term, medium-term, and long-term background scale features from the adjusted multi-scale feature set as input features for the short-term, medium-term, and long-term background scale trend predictors, respectively; under the bidirectional constraint of the spatiotemporal constraint propagation network, each subscale trend predictor reads the stage identification information and selects the model parameter set corresponding to the current evolution stage; performs time series prediction operations on the input features to obtain the risk trend sequence at the corresponding time scale; and calculates the uncertainty representation of the risk trend sequence based on the prediction residuals and sample variance.

6. The method for predicting risk trends of landslide disasters at multiple time scales according to claim 5, characterized in that, The bidirectional constraints include: in the spatiotemporal constraint propagation network, short-term trend predictors, medium-term trend predictors, and long-term background scale trend predictors are respectively used as constraint nodes; constraint channels are established between each constraint node, from the short-term trend predictor to the medium-term trend predictor and the long-term background scale trend predictor, as well as constraint channels from the medium-term trend predictor and the long-term background scale trend predictor to the short-term trend predictor; in each prediction step, the risk trend sequence and uncertainty characterization output by each subscale trend predictor are used as constraint information, and transmitted to the subscale trend predictors of other time scales through the corresponding constraint channels to update the input features, initial state, and prediction boundary conditions of the subscale trend predictors, and after the update, the time series prediction operation of the next prediction step is performed.

7. A multi-timescale risk trend prediction method for landslide disasters according to claim 6, characterized in that, The constraint propagation strength parameters include: determining the current evolution stage based on stage identification information; selecting a set of constraint propagation strength parameters corresponding to the current evolution stage from a preset constraint propagation strength parameter table; including a first constraint weight coefficient for transmitting constraint information from the short-term trend predictor to the medium-term trend predictor and the long-term background scale trend predictor; a second constraint weight coefficient for transmitting constraint information from the medium-term trend predictor and the long-term background scale trend predictor to the short-term trend predictor; and a maximum adjustment range for limiting the input feature correction, initial state offset, and prediction boundary condition adjustment caused by the constraint information; when the spatiotemporal constraint propagation network performs constraint information transmission, the constraint information in each constraint channel is multiplied by the corresponding first constraint weight coefficient or second constraint weight coefficient, and the input feature correction, initial state offset, and prediction boundary condition adjustment are truncated according to the maximum adjustment range.

8. The method for predicting risk trends of landslide disasters at multiple time scales according to claim 1, characterized in that, The extraction of key risk trend events includes: calculating the rate of change sequence and the rate of change of the rate of change sequence of the fused risk trend curve according to the time step; determining the trend inflection point and the peak of the risk growth rate based on the change of the sign of the rate of change or the extreme point of the rate of change; finding the time point when the fused risk value first reaches or exceeds the preset risk threshold and the time point when it last falls below the risk threshold from the fused risk trend curve; determining the continuous time interval between the two time points as the risk cross-level time window; and encoding the trend inflection point, the peak of the risk growth rate, and the risk cross-level time window as key risk trend events and outputting them.

9. A multi-timescale risk trend prediction system for landslide disasters, characterized in that, The system applies a multi-timescale risk trend prediction method for landslide disasters as described in any one of claims 1 to 8, including: Data acquisition module: Acquires multi-source data of target potential hazard points and performs time stamp unification, spatial correlation and quality identification to form a standard dataset; generates a multi-scale feature set based on the standard dataset; Stage determination module: Determines the evolution stage in real time using an evolution stage identification model based on a multi-scale feature set, and outputs stage identification information; Scale adjustment module: Dynamically determines time scale division parameters based on stage identifier information and adjusts multi-scale feature sets; Prediction constraint module: Constructs a spatiotemporal constraint propagation network, establishes bidirectional constraints between trend predictors at each scale, and sets constraint propagation strength parameters according to stage identification information; The segmented prediction module inputs the adjusted multi-scale feature set into the corresponding segmented trend predictor and outputs the risk trend sequence and its uncertainty characterization at each scale under the bidirectional constraints of the spatiotemporal constraint propagation network. Risk fusion module: The dynamic weight allocator determines the fusion weights based on stage identification information and uncertainty characterization and applies cross-scale consistency constraints to obtain the fusion risk trend curve and its confidence interval. Event Extraction Module: Extracts key risk trend events from the fused risk trend curve and outputs them.

Images