Mountain torrent disaster danger area dynamic demarcation method based on multi-source data coupling

Abstract

The invention discloses a mountain torrent disaster danger area dynamic demarcation method based on multi-source data coupling, and the method comprises the steps: obtaining meteorological radar rainfall forecast data, surface deformation monitoring data and high-precision DEM topographic data in real time, carrying out the data preprocessing, and generating a fused multi-source data set; constructing a distributed hydrological model, and calculating the production and convergence flow of the sub-basins; training a Seq2Seq-LSTM model on the basis of rainfall-water level-swelling speed time sequence data, predicting danger areas in a future set time period, and dynamically dividing high / medium / low risk areas by adopting an SOM-CLARA clustering algorithm and taking rainfall intensity, earth surface deformation rate and gradient as feature vectors; and determining the early warning level of each risk area in combination with the critical rainfall threshold and the spatial heterogeneity index. Through the multi-source data dynamic coupling and self-adaptive demarcation algorithm, the dangerous area range is dynamically adjusted, and the accuracy and timeliness of mountain torrent disaster risk prevention and control are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mountain torrent disaster risk zoning, and particularly relates to a mountain torrent disaster risk zone dynamic zoning method based on multi-source data coupling. BACKGROUND

[0002] Mountain torrent disaster is one of the main natural disasters in the world, especially in complex mountainous areas, which causes major casualties and property losses due to its strong suddenness, great destructive power and difficult prediction and early warning. Precise and dynamic delineation of mountain torrent disaster risk zones is the core basis of mountain torrent disaster monitoring, early warning and risk prevention and control, and is directly related to the timeliness and effectiveness of disaster prevention and mitigation measures. The traditional mountain torrent disaster risk zoning method has significant limitations in dealing with complex and variable mountain torrent disaster scenarios: (1) Insufficient data fusion: existing technologies generally rely on a single or limited data source (such as historical disaster records, sparse rainfall station data or static terrain data), and fail to effectively integrate meteorological radar monitoring data, ground deformation monitoring data, high-precision DEM and other multi-source heterogeneous data, resulting in a single information dimension and incomplete spatio-temporal coverage for zoning, which severely limits the accuracy and reliability of the risk zone delineation; (2) Static zoning lag: most traditional methods use static zoning methods based on historical experience or fixed thresholds (such as critical rainfall), without considering real-time rainfall, dynamic changes in soil moisture content and other key factors, which cannot respond to sudden mountain torrent events, resulting in slow or even rigid updates of the delineated risk zone range, and a serious lack of timeliness in early warning, making it difficult to meet the real-time dynamic needs of disaster emergency response; (3) Low model adaptability: traditional zoning methods (such as GIS-based spatial analysis and simple empirical formulas) fail to effectively combine advanced dynamic hydrological physical process simulation and machine learning algorithms, and have weak model generalization ability when dealing with small watersheds with significant spatial heterogeneity (such as terrain changes, vegetation coverage differences and soil type changes), making it difficult to adapt to dynamic changes in different regions, seasons or rainfall patterns; (4) Weak dynamic early warning capability: due to the static nature of the above data, models and zoning methods, existing methods do not integrate real-time rainfall prediction and dynamic risk level updating mechanisms, resulting in insufficient timeliness of early warning, low frequency of risk zone range and risk level updates, and difficulty in meeting the high requirements of mountain torrent disaster rapid response.

[0003] In summary, the core technical problems of existing mountain torrent disaster risk zoning technology can be summarized as "data silos, model rigidity and poor real-time performance", and there is an urgent need to develop a mountain torrent disaster risk zone dynamic zoning method that can deeply integrate multi-source dynamic data, have self-adaptation ability and achieve high timeliness updates. SUMMARY

[0004] The purpose of this invention is to provide a dynamic delineation method for flash flood hazard zones based on multi-source data coupling. By dynamically coupling multi-source data and using an adaptive delineation algorithm, the scope of the hazard zone is dynamically adjusted, solving the pain points of "data silos, rigid models, and poor real-time performance" in existing flash flood hazard zone delineation methods, and significantly improving the accuracy and timeliness of flash flood disaster risk prevention and control.

[0005] To achieve the above objectives, the present invention provides the following solution: A dynamic delineation method for flash flood hazard zones based on multi-source data coupling, comprising the following steps: S1, Multi-source data fusion processing: Real-time acquisition of meteorological radar rainfall forecast data, surface deformation monitoring data, and high-precision DEM topographic data, and data preprocessing to generate a fused multi-source dataset; S2, Dynamic Hydrological Model Construction and Correction: Based on multi-source datasets, a distributed hydrological model is constructed to calculate the runoff and runoff of sub-basins and obtain a real-time corrected watershed runoff distribution map; S3, Dynamic prediction and classification of danger zones: Based on the rainfall-water level-rise rate time series data, a Seq2Seq-LSTM model is trained, and the danger zones within a set time period are predicted based on the Seq2Seq-LSTM model. The SOM-CLARA clustering algorithm is used to dynamically divide high / medium / low risk zones with rainfall intensity, surface deformation rate and slope as feature vectors. S4, Adaptive Threshold Optimization and Early Warning Triggering: Based on the Bayesian optimization algorithm, the critical rainfall threshold is dynamically calculated, and the spatial heterogeneity index of each risk area is calculated. Combining the critical rainfall threshold and the spatial heterogeneity index, the early warning level of each risk area is determined.

[0006] Furthermore, in S1, data preprocessing includes: For high-precision DEM topographic data, the mean-variable point algorithm is used to eliminate abnormal elevation points; Kriging spatial interpolation is used to generate continuous spatial precipitation and deformation fields from meteorological radar rainfall forecast data and surface deformation monitoring data.

[0007] Furthermore, the method of eliminating abnormal elevation points using the mean-changing point algorithm specifically includes: Set up a 30×30 grid sliding window, calculate the arithmetic mean μ and variance σ² of the elevations of all grid cells within the sliding window, and calculate the mean elevation anomaly. :

[0008] In the formula, Indicates the first i The elevation value of each grid, mThis represents the total number of grid cells within the sliding window. like If the elevation is greater than 3σ, it is determined to be an abnormal elevation point and is corrected by eight-neighborhood bilinear interpolation. The method of using Kriging space interpolation to generate continuous spatial precipitation and deformation fields specifically includes: Construct a Gaussian semivariogram model:

[0009] In the formula, The distance between two points is 1 h The semivariogram over time, where c0 represents the nugget effect, indicating measurement error or microscale variation, and c represents the sill value, indicating total spatial variation. h Indicates the spatial distance between two points. a Indicates range change; By optimizing the interpolation parameters a=500m and c0=0.2 through cross-validation, a continuous spatial precipitation field and deformation field are generated.

[0010] Furthermore, in S2, the construction and calibration of the dynamic hydrological model specifically includes: The watershed to be measured is divided into multiple sub-watersheds. A distributed hydrological model is constructed based on a multi-source dataset. The runoff and sinking discharge of each sub-watershed are calculated based on the distributed hydrological model.

[0011] In the formula, R( ) represents the real-time rainfall sequence; ET( ) represents evapotranspiration, calculated using the Penman-Monteith formula; G is the terrain slope matrix; K s ε represents the soil permeability coefficient, calibrated through field sampling; ε is the model residual. Among them, the real-time flow rate Q of the hydrological station is connected. obs The Kalman filter is used to dynamically correct the model residual ε:

[0012] In the formula, K k This represents the Kalman gain matrix, which is dynamically adjusted based on the observation noise covariance matrix. This represents the model residual at the current time. This represents the model residual at the previous time step. Q obs This indicates that the hydrological station is monitoring the flow rate in real time. Q model This indicates that the model predicts the flow.

[0013] Furthermore, in S3, the dynamic prediction and classification of hazardous areas specifically includes: S301, predicting the dangerous area in the next 1 hour by the Seq2Seq-LSTM model: input the rainfall-water level-rising velocity time series data [R(t-30:t), H(t-30:t), V(t-30:t)] in the past 30 minutes, time step 1 minute, the encoder uses double-layer LSTM to extract features, the decoder uses single-layer LSTM, and the output is the dangerous probability distribution in the next 60 minutes:

[0014] In the formula, Y t+1:t+6 , which represents the prediction result sequence from time step t +1 to t +6, covering 6 consecutive time steps; W represents the weight matrix of the output layer, which linearly transforms the hidden state h t to map the dimension of the hidden state to the output dimension; h t is the hidden state corresponding to time step t , which encodes historical information: contains all input information from the start of the sequence to time step t ; b is the bias vector of the output layer, which adds an offset to the result of linear transformation ; S302, construct a 10x10 self-organizing mapping grid, input feature vector contains [rainfall intensity, surface deformation rate, slope], use hierarchical sampling strategy to select representative samples, use SOM-CLARA clustering algorithm to optimize clustering center, dynamically divide high / medium / low risk area: Calculate the normalized distance , where d is the Euclidean distance from the sample point to the cluster center, d max is the maximum Euclidean distance from all sample points to their cluster centers; When 0≤ ≤0.3, it is judged as a high-risk area; When 0.3< ≤0.6, it is judged as a medium-risk area; When 0.6< , it is judged as a low-risk area.

[0015] Further, in the S4, adaptive threshold optimization and early warning triggering, specifically including: S401, dynamically calculate the critical rainfall threshold:

[0016] Where α1∈[0.8,1.2] and β1∈[0.5,1.5] are the watershed characteristic coefficients calibrated by Bayesian optimization, respectively. =2.3; This represents the dynamic critical rainfall threshold. α 1 represents the weighting coefficient, used to adjust the contribution ratio of the historical quantile term to the critical value. This represents the 90th percentile of historical data. In the historical sample, 90% of the data are less than or equal to this value. It is used to measure the high risk or extreme level in the historical distribution. β1 represents the weighting coefficient, used to adjust the contribution of the state adjustment term to the critical value. S ( t ) represents the current state variable, indicating the state of the research object at time t. t The current value, S max The largest state variable represents the state variable. S ( t The historical maximum or theoretical upper limit of ) Used to control the nonlinearity of the state adjustment term; S402, Calculate the spatial heterogeneity index for each risk zone:

[0017] In the formula, The standard deviation of rainfall intensity within the risk zone reflects the degree of data dispersion. The mean of rainfall intensity within the risk zone reflects the central tendency; CV represents the coefficient of variation, a standardized dispersion index. S403, when real-time rainfall R( )≥0.7R critical A blue alert is triggered when the spatial variation coefficient CV > 0.5; when R( )≥R critical A red alert is triggered when the area of ​​a high-risk zone exceeds 20%.

[0018] Furthermore, the method also includes: S5, Visualization and Route Planning: Risk heatmaps are generated using the Epanechnikov kernel function on a GIS platform and overlaid with satellite imagery; an improved A... The algorithm integrates water depth and road damage coefficients to construct a path cost function and determine evacuation routes.

[0019] Furthermore, S5, visualization and path planning, specifically includes: S501 uses the Epanechnikov kernel function to generate risk heatmaps on a GIS platform:

[0020] Wherein, the kernel function K(u) is the Epanechnikov kernel:

[0021] In the formula, The risk heatmap value represents the risk probability density at coordinates (x, y), with a higher value indicating higher risk; n represents the number of risk points; bandwidth B is adaptively adjusted, with B=50m in densely populated areas and B=200m in uninhabited areas; d i Represent the Euclidean distance from point (x,y) to the i-th risk point, and calculate the intermediate quantity; K( u ) represents the Epanechnikov kernel function, which assigns distance weights, with nearest neighbors contributing more and far neighbors contributing zero. ; S502, adopts improved A Algorithm, which integrates water depth and road damage coefficient to construct path cost function:

[0022] In the formula, F(n) represents the total cost of node j, and A The total cost used for priority sorting in the algorithm is G(j), where a smaller value indicates a higher priority. G(j) represents the actual cost from the starting point to node j, which is the cumulative cost of the already traveled path. H(j) is a heuristic function representing the estimated cost to the destination. λ represents the dynamic weight coefficient. D(j) represents the water depth cost at node j. R(j) represents the road damage coefficient at node j, with a value range of [0,1], where 0 represents intact and 1 represents completely damaged. It is based on satellite imagery or sensor data. α2 and β2 represent the weight coefficients of water depth and damage, respectively.

[0023] This invention also provides a dynamic delineation system for flash flood hazard zones based on multi-source data coupling, applied to the above-mentioned dynamic delineation method for flash flood hazard zones based on multi-source data coupling, comprising: The multi-source data fusion processing module is used to acquire meteorological radar rainfall forecast data, surface deformation monitoring data, and high-precision DEM topographic data in real time, and to perform data preprocessing to generate a fused multi-source dataset. The dynamic hydrological model construction and calibration module is used to build a distributed hydrological model based on multi-source datasets, calculate the runoff and runoff of sub-basins, and obtain a real-time calibrated watershed runoff distribution map. The dynamic prediction and classification module for dangerous areas is used to train a Seq2Seq-LSTM model based on rainfall-water level-rise rate time series data, predict dangerous areas within a set time period based on the Seq2Seq-LSTM model, and use the SOM-CLARA clustering algorithm to dynamically classify high / medium / low risk areas using rainfall intensity, surface deformation rate, and slope as feature vectors. The adaptive threshold optimization and early warning triggering module is used to dynamically calculate the critical rainfall threshold based on the Bayesian optimization algorithm, calculate the spatial heterogeneity index of each risk area, and determine the early warning level of each risk area by combining the critical rainfall threshold and the spatial heterogeneity index.

[0024] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the above-described method for dynamic delineation of flash flood hazard zones based on multi-source data coupling.

[0025] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention provides a dynamic delineation method for flash flood hazard zones based on multi-source data coupling. It acquires real-time meteorological radar rainfall forecast data, surface deformation monitoring data, and high-precision DEM data to generate a fused multi-source dataset. This method fully utilizes the coupling and correlation information between multiple data sources to capture the multi-dimensional dynamic changes of flash flood hazards in real time, significantly improving the timeliness and spatial resolution of predictions and reducing the limitations of single data sources. Based on a distributed hydrological model, it calculates the runoff and runoff of sub-basins, corrects the runoff distribution map of the basin in real time, and accurately captures heterogeneous features such as local topography, soil, and vegetation, avoiding the coarse assumptions of traditional lumped models and improving the accuracy of runoff distribution calculation. Seq2Seq-LSTM captures... The long-term dependency between rainfall, water level, and rate of rise can predict the evolution trend of danger zones in the future, providing a time window for early warning and emergency response. The SOM-CLARA clustering algorithm uses key features such as rainfall intensity, surface deformation rate, and slope to perform dynamic clustering, achieving a scientific division of high / medium / low-risk zones, avoiding the rigidity of static threshold division, and improving the dynamic adaptability and spatial heterogeneity of the classification. Bayesian optimization efficiently searches for the optimal critical rainfall threshold, dynamically balancing the risks of missed and false alarms, enhancing the robustness of early warnings under complex terrain and climate conditions. Combined with spatial heterogeneity indicators, it closely links the warning level with the actual disaster sensitivity of the region, improving the accuracy of early warnings and the efficiency of resource allocation.

[0026] In summary, this invention, through multi-source data-driven approaches, deep integration of deep learning and traditional hydrological models, and the introduction of adaptive optimization techniques, achieves dynamic, accurate, and intelligent delineation and graded early warning of flash flood hazard zones. It has achieved good results in terms of real-time performance, spatial resolution, scientific rigor of risk classification, and reliability of early warning, providing efficient technical support for disaster prevention and mitigation decision-making and significantly improving the accuracy and timeliness of flash flood risk prevention and control. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of a dynamic delineation method for flash flood hazard zones based on multi-source data coupling, according to an embodiment of the present invention. Detailed Implementation

[0029] The embodiments of the present invention are described in detail below. These embodiments are intended to explain the present invention and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in the art or according to the product instructions. Materials or instruments whose manufacturers are not specified are all conventional products that can be obtained commercially.

[0030] This invention aims to address the pain points of traditional flash flood hazard zone demarcation—namely, "data silos, rigid models, and poor real-time performance"—through dynamic coupling of multi-source data and an adaptive demarcation algorithm. Specific objectives include: Multimodal data fusion: Integrating multi-source data such as meteorological, hydrological, and topographic data to construct a spatiotemporally coupled dynamic boundary model; Adaptive threshold optimization: Based on real-time rainfall forecasts and distributed hydrological models, dynamically adjust the threshold for delineating danger zones; High-precision dynamic early warning: Combining deep learning algorithms, it achieves minute-level updates of the danger zone range.

[0031] Example 1

[0032] like Figure 1 As shown in the figure, the dynamic delineation method for flash flood hazard zones based on multi-source data coupling provided by this invention includes the following steps: S1, Multi-source data fusion processing: Real-time acquisition of high spatiotemporal resolution meteorological radar rainfall forecast data (time resolution ≤ 5 minutes), surface deformation monitoring data reflecting geological stability and soil moisture dynamics (accuracy ± 1 cm), and high-precision DEM topographic data (resolution ≤ 10 m) that finely depicts topographic features, and data preprocessing to generate a fused multi-source dataset; making full use of the coupling and correlation information between multi-source data to more comprehensively depict the comprehensive environment of flash flood formation.

[0033] Data preprocessing includes: (1) For high-precision DEM topographic data, the mean-variable point algorithm is used to eliminate abnormal elevation points: Set up a 30×30 grid sliding window, calculate the arithmetic mean μ and variance σ² of the elevations of all grid cells within the sliding window, and calculate the mean elevation anomaly. :

[0034] In the formula, Indicates the first i The elevation value of each grid, m This represents the total number of grid cells within the sliding window. like If the elevation is greater than 3σ, it is determined to be an abnormal elevation point and is corrected by eight-neighborhood bilinear interpolation. (2) Using meteorological radar rainfall forecast data and surface deformation monitoring data, kriging spatial interpolation is used to generate continuous spatial rainfall and deformation fields: Construct a Gaussian semivariogram model:

[0035] In the formula, The distance between two points is 1 h The semivariogram (representing spatial variation) is given by c0, which represents the nugget effect (representing measurement error or microscale variation, with units consistent with the data, such as rainfall / mm), and c represents the sill value (total spatial variation, c0+c is the maximum value of the semivariogram). h Indicates the spatial distance between two points (unit: meters). a This indicates the range of spatial autocorrelation (e.g., 500m, meaning that spatial autocorrelation disappears beyond this distance). By optimizing the interpolation parameters a=500m and c0=0.2 through cross-validation, a continuous spatial precipitation field and deformation field are generated.

[0036] Through step S1, we can obtain DEM raster data with anomaly correction, spatially continuous rainfall intensity distribution map (10m resolution), and surface deformation rate distribution map (0.5cm / h accuracy), which eliminates the influence of topographic data anomalies on hydrological simulation, improves the spatial matching accuracy of multi-source data, and reduces interpolation error by 62%.

[0037] S2, Dynamic Hydrological Model Construction and Correction: A distributed hydrological model is constructed based on multi-source datasets to calculate the runoff and runoff in sub-basins, obtaining a real-time corrected watershed runoff distribution map; specifically including: The watershed to be measured is divided into multiple sub-watersheds. A distributed hydrological model is constructed based on a multi-source dataset. The runoff and sinking discharge of each sub-watershed are calculated based on the distributed hydrological model.

[0038] In the formula, R( ) represents the real-time rainfall sequence; ET( ) represents evapotranspiration, calculated using the Penman-Monteith formula; G is the terrain slope matrix; K s ε represents the soil permeability coefficient, calibrated through field sampling; ε is the model residual. Among them, the real-time flow rate Q of the hydrological station is connected. obs The Kalman filter is used to dynamically correct the model residual ε:

[0039] In the formula, K k This represents the Kalman gain matrix, which is dynamically adjusted based on the observation noise covariance matrix. This represents the model residual at the current time. This represents the model residual at the previous time step. Q obs This indicates the real-time flow rate observed at the hydrological station (unit: m³ / s). Q model This represents the model's predicted flow rate (unit: m³ / s).

[0040] Through step S2, a real-time corrected watershed runoff distribution map (time resolution 5 minutes) and a three-dimensional spatiotemporal distribution model of soil moisture content (vertical stratification ≥ 3 layers) are obtained; the model residuals are reduced by 45% compared with traditional methods, and the runoff prediction response time is shortened to within 8 minutes.

[0041] S3, Dynamic Prediction and Classification of Hazardous Areas: A Seq2Seq-LSTM model is trained based on rainfall-water level-rise rate time-series data. This model predicts hazardous areas within a given future time period. The SOM-CLARA clustering algorithm is then used, employing rainfall intensity, surface deformation rate, and slope as feature vectors to dynamically classify areas into high, medium, and low-risk zones. Specifically, this includes: S301 predicts the danger zone for the next hour using a Seq2Seq-LSTM model: Input is the rainfall-water level-rate of rise time series data for the past 30 minutes [R(t-30:t), H(t-30:t), V(t-30:t)], with a time step of 1 minute. The encoder uses a two-layer LSTM to extract features, and the decoder uses a single-layer LSTM. The output is the danger probability distribution for the next 60 minutes.

[0042] In the formula, Y t+1:t+6 Indicates from time step t +1 to t The prediction result sequence of +6 covers 6 consecutive time steps; W represents the weight matrix of the output layer, which is applied to the hidden state. ht Perform a linear transformation to map the dimensions of the hidden state to the output dimension; ht It is a time step t The corresponding hidden states encode historical information: they contain information from the beginning of the sequence to the time step. t All input information, b It is the bias vector of the output layer, and its function is to provide bias for the linear transformation. Add an offset to the result; S302: Construct a 10×10 self-organizing map grid. The input feature vector includes [rainfall intensity, surface deformation rate, slope]. A stratified sampling strategy is used to select representative samples, with an 80% sampling rate in high-density areas and a mandatory 5% retention rate in edge areas. The SOM-CLARA clustering algorithm is used to optimize cluster centers and dynamically divide high / medium / low-risk areas. Calculate normalized distance ,in, d The distance is the Euclidean distance from the sample point to the cluster center. d max The maximum Euclidean distance from all sample points to their cluster centers; When 0≤ A value ≤0.3 indicates a high-risk area; When 0.3 < A value ≤0.6 indicates a medium-risk area; When 0.6 < It has been identified as a low-risk area.

[0043] Through step S3, a probability prediction map of flash flood hazard areas for the next hour (spatial resolution 10m) and vector boundary files of high / medium / low risk areas (GeoJSON format) are obtained; the disaster hit rate is increased to 93%, and the adaptability to small watershed terrain is improved by 35%.

[0044] S4, Adaptive Threshold Optimization and Early Warning Triggering: Based on a Bayesian optimization algorithm, the critical rainfall threshold is dynamically calculated, and the spatial heterogeneity index for each risk area is calculated. Combining the critical rainfall threshold and the spatial heterogeneity index, the early warning level for each risk area is determined. Specifically, this includes: S401, Dynamically calculate the critical rainfall threshold:

[0045] Where α1∈[0.8,1.2] and β1∈[0.5,1.5] are the watershed characteristic coefficients calibrated by Bayesian optimization, respectively. =2.3; This represents the dynamic critical rainfall threshold (unit: mm / h). α 1 represents the weighting coefficient, used to adjust the contribution ratio of the historical quantile term to the critical value. This represents the 90th percentile of historical data. In the historical sample, 90% of the data are less than or equal to this value. It is used to measure the high risk or extreme level in the historical distribution. β1 represents the weighting coefficient, used to adjust the contribution of the state adjustment term to the critical value. S ( t ) represents the current state variable, indicating the state of the research object at time t. t The current value, S max The largest state variable represents the state variable. S ( t The historical maximum or theoretical upper limit of ) Used to control the nonlinearity of the state adjustment term; S402, Calculate the spatial heterogeneity index for each risk zone:

[0046] In the formula, The standard deviation of rainfall intensity within the risk area (unit: mm / h) reflects the degree of data dispersion. The mean rainfall intensity within the risk area (unit: mm / h) reflects the central tendency; CV represents the coefficient of variation (dimensionless), a standardized dispersion index. S403, when real-time rainfall R( )≥0.7R critical A blue alert is triggered when the spatial variation coefficient CV > 0.5; when R( )≥R critical A red alert is triggered when the area of ​​a high-risk zone exceeds 20%.

[0047] The optimal evacuation route planning scheme is obtained by dynamically updating the early warning level distribution map in step S4 (update cycle ≤ 1 minute); this reduces the false alarm rate of early warnings and improves the accuracy of emergency resource deployment.

[0048] S5, Visualization and Route Planning: Risk heatmaps are generated using the Epanechnikov kernel function on a GIS platform and overlaid with satellite imagery; an improved A... The algorithm integrates water depth and road damage coefficients to construct a path cost function and determine evacuation routes; specifically, it includes: S501 uses the Epanechnikov kernel function to generate risk heatmaps on a GIS platform:

[0049] Wherein, the kernel function K(u) is the Epanechnikov kernel:

[0050] In the formula, The risk heatmap value represents the risk probability density at coordinates (x, y), with a higher value indicating higher risk; n represents the number of risk points; bandwidth B is adaptively adjusted, with B=50m in densely populated areas and B=200m in uninhabited areas; d i Represent the Euclidean distance from point (x,y) to the i-th risk point, and calculate the intermediate quantity; K( u ) represents the Epanechnikov kernel function, which assigns distance weights, with nearest neighbors contributing more and far neighbors contributing zero. ; S502, adopts improved A Algorithm, which integrates water depth and road damage coefficient to construct path cost function:

[0051] In the formula, F(n) represents the total cost of node j, and A The total cost used for priority sorting in the algorithm is G(j), where a smaller value indicates a higher priority. G(j) represents the actual cost from the starting point to node j, which is the cumulative cost of the already traveled path. H(j) is a heuristic function representing the estimated cost to the destination. λ represents the dynamic weight coefficient. D(j) represents the water depth cost at node j. R(j) represents the road damage coefficient at node j, with a value range of [0,1], where 0 represents intact and 1 represents completely damaged. It is based on satellite imagery or sensor data. α2 and β2 represent the weight coefficients of water depth and damage, respectively.

[0052] Step S5 obtains a real-time risk heat map, overlays it with a satellite image base map, and generates an optimal evacuation route navigation file (including a GPS coordinate point sequence); this reduces visualization rendering latency and improves route safety.

[0053] Table 1

[0054] As shown in Table 1, the method described in this invention reduces the error of DEM anomaly point correction, improves the hit rate of hazard area prediction, and enables faster early warning response compared to the traditional static model.

[0055] This invention proposes a method for delineating flash flood hazard zones based on dynamic coupling of multi-source data. By integrating meteorological radar rainfall forecasts, surface deformation monitoring, and high-precision DEM topographic data, and combining distributed hydrological models with deep learning algorithms (Seq2Seq-LSTM, SOM-CLARA clustering), it achieves minute-level dynamic updates and adaptive early warnings for hazard zones, solving the pain points of "data silos, rigid models, and poor real-time performance" in traditional methods.

[0056] This invention achieves accurate, real-time, and adaptive delineation and early warning of flash flood hazard zones through multi-source data fusion, hydrological-machine learning hybrid modeling, and dynamic optimization algorithms. It is significantly superior to traditional methods in terms of data dimensionality, model accuracy, response speed, and resource efficiency, providing a breakthrough technical solution for flash flood disaster prevention and control.

[0057] This invention, through dynamic coupling of multi-source heterogeneous data, hydrological-machine learning hybrid modeling, and adaptive evolution of risk thresholds, completely solves the shortcomings of traditional flash flood demarcation, such as data silos, static models, and early warning delays, and has the following technical advantages: Dynamic breakthrough: The frequency of danger zone demarcation updates has been increased from hourly to minutely, improving the accuracy of responding to sudden rainstorm events; Significantly improved accuracy: Compared with single-source methods, the spatial matching degree (F1-score) of the multi-source coupled model is effectively improved; Resource optimization and allocation: By dynamically delineating redundant early warning areas, the investment of emergency resources is reduced; Enhanced generalization ability: The model's applicability in complex terrains has been improved, and the accuracy of disaster prediction in small watersheds in mountainous areas has been increased.

[0058] Example 2 This invention also provides a dynamic delineation system for flash flood hazard zones based on multi-source data coupling, applied to the above-mentioned dynamic delineation method for flash flood hazard zones based on multi-source data coupling, comprising: The multi-source data fusion processing module is used to acquire meteorological radar rainfall forecast data, surface deformation monitoring data, and high-precision DEM topographic data in real time, and to perform data preprocessing to generate a fused multi-source dataset. The dynamic hydrological model construction and calibration module is used to build a distributed hydrological model based on multi-source datasets, calculate the runoff and runoff of sub-basins, and obtain a real-time calibrated watershed runoff distribution map. The dynamic prediction and classification module for dangerous areas is used to train a Seq2Seq-LSTM model based on rainfall-water level-rise rate time series data, predict dangerous areas within a set time period based on the Seq2Seq-LSTM model, and use the SOM-CLARA clustering algorithm to dynamically classify high / medium / low risk areas using rainfall intensity, surface deformation rate, and slope as feature vectors. The adaptive threshold optimization and early warning triggering module is used to dynamically calculate the critical rainfall threshold based on the Bayesian optimization algorithm, calculate the spatial heterogeneity index of each risk area, and determine the early warning level of each risk area by combining the critical rainfall threshold and the spatial heterogeneity index.

[0059] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the above-described method for dynamic delineation of flash flood hazard zones based on multi-source data coupling.

[0060] Matters not covered in this invention are common knowledge.

[0061] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for dynamic delineation of flash flood hazard zones based on multi-source data coupling, characterized in that, Includes the following steps: S1, Multi-source data fusion processing: Real-time acquisition of meteorological radar rainfall forecast data, surface deformation monitoring data, and high-precision DEM topographic data, and data preprocessing to generate a fused multi-source dataset; S2, Dynamic Hydrological Model Construction and Correction: Based on multi-source datasets, a distributed hydrological model is constructed to calculate the runoff and runoff of sub-basins and obtain a real-time corrected watershed runoff distribution map; S3, Dynamic prediction and classification of danger zones: Based on the rainfall-water level-rise rate time series data, a Seq2Seq-LSTM model is trained, and the danger zones within a set time period are predicted based on the Seq2Seq-LSTM model. The SOM-CLARA clustering algorithm is used to dynamically divide high / medium / low risk zones with rainfall intensity, surface deformation rate and slope as feature vectors. S4, Adaptive Threshold Optimization and Early Warning Triggering: Based on the Bayesian optimization algorithm, the critical rainfall threshold is dynamically calculated, and the spatial heterogeneity index of each risk area is calculated. Combining the critical rainfall threshold and the spatial heterogeneity index, the early warning level of each risk area is determined.

2. The method for dynamic delineation of flash flood hazard zones based on multi-source data coupling according to claim 1, characterized in that, In step S1, data preprocessing includes: For high-precision DEM topographic data, the mean-variable point algorithm is used to eliminate abnormal elevation points; Kriging spatial interpolation is used to generate continuous spatial precipitation and deformation fields from meteorological radar rainfall forecast data and surface deformation monitoring data.

3. The method for dynamic delineation of flash flood hazard zones based on multi-source data coupling according to claim 2, characterized in that, The method of eliminating abnormal elevation points using the mean-changing point algorithm specifically includes: Set up a 30×30 grid sliding window, calculate the arithmetic mean μ and variance σ² of the elevations of all grid cells within the sliding window, and calculate the mean elevation anomaly. : ； In the formula, Indicates the first i The elevation value of each grid, m This represents the total number of grid cells within the sliding window. like If the elevation is greater than 3σ, it is determined to be an abnormal elevation point, and eight-neighborhood bilinear interpolation is used for correction. The method of using Kriging space interpolation to generate continuous spatial precipitation and deformation fields specifically includes: Construct a Gaussian semivariogram model: ； In the formula, The distance between two points is 1 h The semivariogram over time, where c0 represents the nugget effect, indicating measurement error or microscale variation, and c represents the sill value, indicating total spatial variation. h Indicates the spatial distance between two points. a Indicates range change; Optimize interpolation parameters through cross-validation a =500m, c0=0.2, generating a continuous spatial precipitation field and deformation field.

4. The method for dynamic delineation of flash flood hazard zones based on multi-source data coupling according to claim 1, characterized in that, In S2, the construction and calibration of the dynamic hydrological model specifically includes: The watershed to be measured is divided into multiple sub-watersheds. A distributed hydrological model is constructed based on a multi-source dataset. The runoff and sinking discharge of each sub-watershed are calculated based on the distributed hydrological model. ； In the formula, R( ) represents the real-time rainfall sequence; ET( ) represents evapotranspiration, calculated using the Penman-Monteith formula; G is the terrain slope matrix; K s ε represents the soil permeability coefficient, calibrated through field sampling; ε is the model residual. Among them, the real-time flow rate Q of the hydrological station is connected. obs The residual ε of the Kalman filter dynamic correction model is calculated using the following formula: ； In the formula, K k This represents the Kalman gain matrix, which is dynamically adjusted based on the observation noise covariance matrix. This represents the model residual at the current time. This represents the model residual at the previous time step. Q obs This indicates that the hydrological station is monitoring the flow rate in real time. Q model This indicates that the model predicts the flow.

5. The method for dynamic delineation of flash flood hazard zones based on multi-source data coupling according to claim 1, characterized in that, In S3, the dynamic prediction and classification of hazardous areas specifically includes: S301 predicts the danger zone for the next hour using a Seq2Seq-LSTM model: Input is the rainfall-water level-rate of rise time series data for the past 30 minutes [R(t-30:t), H(t-30:t), V(t-30:t)], with a time step of 1 minute. The encoder uses a two-layer LSTM to extract features, and the decoder uses a single-layer LSTM. The output is the danger probability distribution for the next 60 minutes. ； In the formula, Y t+1:t+6 Indicates from time step t +1 to t The prediction result sequence of +6 covers 6 consecutive time steps; W represents the weight matrix of the output layer, which is applied to the hidden state. h t Perform a linear transformation to map the dimensions of the hidden state to the output dimension; h t It is a time step t The corresponding hidden states encode historical information: they contain information from the beginning of the sequence to the time step. t All input information; b It is the bias vector of the output layer, and its function is to provide bias for the linear transformation. Add an offset to the result; S302: Construct a 10×10 self-organizing map grid. The input feature vector includes [rainfall intensity, surface deformation rate, slope]. A stratified sampling strategy is used to select representative samples. The SOM-CLARA clustering algorithm is used to optimize the cluster centers and dynamically divide high / medium / low risk areas. Calculate normalized distance in, d The distance is the Euclidean distance from the sample point to the cluster center. d max The maximum Euclidean distance from all sample points to their cluster centers; When 0≤ A value ≤0.3 indicates a high-risk area; When 0.3 < A value ≤0.6 indicates a medium-risk area; When 0.6 < It has been identified as a low-risk area.

6. The method for dynamic delineation of flash flood hazard zones based on multi-source data coupling according to claim 1, characterized in that, In step S4, the adaptive threshold optimization and early warning triggering specifically include: S401, Dynamically calculate the critical rainfall threshold: ； Where α1∈[0.8,1.2] and β1∈[0.5,1.5] are the watershed characteristic coefficients calibrated by Bayesian optimization, respectively. =2.3; This represents the dynamic critical rainfall threshold. α 1 represents the weighting coefficient, used to adjust the contribution ratio of the historical quantile term to the critical value. This represents the 90th percentile of historical data. In the historical sample, 90% of the data are less than or equal to this value. It is used to measure the high risk or extreme level in the historical distribution. β1 represents the weighting coefficient, used to adjust the contribution of the state adjustment term to the critical value. S ( t ) represents the current state variable, indicating the state of the research object at time t. t The current value, S max The largest state variable represents the state variable. S ( t The historical maximum or theoretical upper limit of ) Used to control the nonlinearity of the state adjustment term; S402, Calculate the spatial heterogeneity index for each risk zone: ； In the formula, The standard deviation of rainfall intensity within the risk zone reflects the degree of data dispersion. The mean of rainfall intensity within the risk zone reflects the central tendency; CV represents the coefficient of variation, a standardized dispersion index. S403, when real-time rainfall R( )≥0.7R critical A blue alert is triggered when the spatial variation coefficient CV > 0.5; when R( )≥R critical A red alert is triggered when the area of ​​a high-risk zone exceeds 20%.

7. The method for dynamic delineation of flash flood hazard zones based on multi-source data coupling according to claim 1, the method further includes: S5, Visualization and Route Planning: Risk heatmaps are generated using the Epanechnikov kernel function on a GIS platform and overlaid with satellite imagery; an improved A... The algorithm integrates water depth and road damage coefficients to construct a path cost function and determine evacuation routes.

8. The dynamic delineation method for flash flood hazard zones based on multi-source data coupling according to claim 7, wherein step S5, visualization and path planning, specifically includes: S501 uses the Epanechnikov kernel function to generate risk heatmaps on a GIS platform: ； Wherein, the kernel function K(u) is the Epanechnikov kernel: ； In the formula, The risk heatmap value represents the risk probability density at coordinates (x, y), with a higher value indicating higher risk; n represents the number of risk points; bandwidth B is adaptively adjusted, with B=50m in densely populated areas and B=200m in uninhabited areas; d i Represent the Euclidean distance from point (x,y) to the i-th risk point, and calculate the intermediate quantity; K( u ) represents the Epanechnikov kernel function, which assigns distance weights, with nearest neighbors contributing more and far neighbors contributing zero. ; S502, adopts improved A Algorithm, which integrates water depth and road damage coefficient to construct path cost function: ； In the formula, F(n) represents the total cost of node j, and A The total cost used for priority sorting in the algorithm is G(j), where a smaller value indicates a higher priority. G(j) represents the actual cost from the starting point to node j, which is the cumulative cost of the already traveled path. H(j) is a heuristic function representing the estimated cost to the destination. λ represents the dynamic weight coefficient. D(j) represents the water depth cost at node j. R(j) represents the road damage coefficient at node j, with a value range of [0,1], where 0 represents intact and 1 represents completely damaged. It is based on satellite imagery or sensor data. α2 and β2 represent the weight coefficients of water depth and damage, respectively.

9. A dynamic delineation system for flash flood hazard zones based on multi-source data coupling, applied to the dynamic delineation method for flash flood hazard zones based on multi-source data coupling as described in any one of claims 1-8, characterized in that, include: The multi-source data fusion processing module is used to acquire meteorological radar rainfall forecast data, surface deformation monitoring data, and high-precision DEM topographic data in real time, and to perform data preprocessing to generate a fused multi-source dataset. The dynamic hydrological model construction and calibration module is used to build a distributed hydrological model based on multi-source datasets, calculate the runoff and runoff of sub-basins, and obtain a real-time calibrated watershed runoff distribution map. The dynamic prediction and classification module for dangerous areas is used to train a Seq2Seq-LSTM model based on rainfall-water level-rise rate time series data, predict dangerous areas within a set time period based on the Seq2Seq-LSTM model, and use the SOM-CLARA clustering algorithm to dynamically classify high / medium / low risk areas using rainfall intensity, surface deformation rate, and slope as feature vectors. The adaptive threshold optimization and early warning triggering module is used to dynamically calculate the critical rainfall threshold based on the Bayesian optimization algorithm, calculate the spatial heterogeneity index of each risk area, and determine the early warning level of each risk area by combining the critical rainfall threshold and the spatial heterogeneity index.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed, implements the dynamic delineation method for flash flood hazard zones based on multi-source data coupling as described in any one of claims 1 to 8.

Images