Graded early warning method and system for storm surge submerged area

Abstract

The invention discloses a grading early warning method and system for a storm surge submerging area, and relates to the technical field of data processing. Comprising the following steps: S1, multi-source data preprocessing: a standardized data set is obtained through multi-source heterogeneous data acquisition, space-time reference unification and quality grading processing, and a quality grading result is used for weight distribution of subsequent feature extraction; s2, dynamic feature extraction; s3, performing initial submerging prediction; s4, deviation self-adaptive correction is carried out; s5, grading threshold value adaptation is carried out; and S6, issuing and feeding back graded early warning. According to the grading early warning method, effective integration of multi-source data is realized through sub-steps of multi-source heterogeneous data acquisition, space-time reference unification and quality grading processing. The uniform space-time reference ensures the consistency of the data, the three-level quality grading standard defines the use priority of the data, the utilization rate and reliability of the data are improved, and a solid foundation is laid for subsequent feature extraction and model prediction.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a hierarchical early warning method and system for storm surge inundated areas. Background Art

[0002] Storm surges are common natural disasters in coastal areas. They are sudden and destructive, posing a serious threat to the safety of life and property of coastal residents, economic development and the ecological environment.

[0003] Traditional methods have a weak ability to integrate multi-source heterogeneous data. Data from different sources (such as radar echo data, tide station data, and typhoon path data) often suffer from inconsistent temporal and spatial benchmarks and uneven data quality, resulting in low data utilization and difficulty in fully leveraging the synergistic effects of multi-source data. Furthermore, the lack of clear grading standards for data quality has a negative impact on subsequent feature extraction and model prediction. Furthermore, in the feature extraction and model prediction stages, existing technologies fail to dynamically adjust the weights and time windows of real-time observation features and historical correlation features based on the timeliness of warnings. This results in insufficient reliance on real-time data in short-term warnings and insufficient utilization of historical data in long-term warnings, affecting forecast accuracy. Furthermore, the prediction model employed lacks feature learning for sensitive nearshore areas and has poor generalization capabilities for extreme weather cases, resulting in significant deviations between initial forecast results and actual conditions. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: A hierarchical early warning method for storm surge inundated areas, comprising the following steps: S1, multi-source data preprocessing: Through multi-source heterogeneous data collection, spatiotemporal benchmark unification and quality grading, a standardized data set is obtained. The quality grading results are used for weight allocation in subsequent feature extraction; S2, dynamic feature extraction: Based on the standardized data set, extract real-time observation features and historical correlation features, construct a feature time series matrix, and dynamically adjust the input weights and time window lengths of the two types of features according to the warning timeliness; S3, initial inundation prediction: The weighted features are input into a pre-trained deep learning early warning model. The model's multi-scale feature fusion module outputs a preliminary prediction of the storm surge inundation area. The preliminary prediction includes the spatiotemporal distribution data of the inundation range, depth, and duration. S4, adaptive deviation correction: Calling the historical storm surge case library, using the case similarity matching algorithm to calculate the deviation coefficient between the preliminary prediction results and historical cases, combined with the residual feedback of real-time monitoring data, the preliminary prediction results are dynamically corrected to obtain the corrected inundation data; S5, hierarchical threshold adaptation: combining the corrected inundation data with the preset regional disaster-bearing body characteristic parameters, the fuzzy analytic hierarchy process is used to determine the warning thresholds for different sub-regions and divide the warning levels; S6, graded warning release and feedback: Differentiated warning information is generated according to the warning level and released through the communication method adapted to the receiving object. The release process includes the threshold judgment logic adapted to the communication method, and real-time tracking of the warning information's reach and response data for model iterative optimization.

[0005] Preferably, step S1 includes the following sub-steps: S11, multi-source heterogeneous data collection: Classify and collect data by data type, including: Real-time observation data: collects reflectivity factor data of radar echoes (spatial resolution ≤ 1 km, temporal resolution ≤ 6 minutes), minute-level tide level and current velocity coupling data from tide stations, real-time tracking data of the longitude and latitude of the typhoon path and its intensity (wind speed, air pressure), and drone aerial photography of nearshore submerged boundaries with a resolution of 0.1 m. Historical basic data: Hourly records of storm surge surges over the past three decades (including peak tide levels and surge magnitudes), 1-meter-precision elevation data of coastal terrain (including three-dimensional information on artificial structures such as embankments and roads), and historical disaster loss records (including data correlating inundation depth with economic losses). Environmental data: Collecting data on the average monthly sea level change trend over the past five years, coastal engineering protection level and disaster resilience assessment data (such as seawall design standards and levee failure risk factors), and quarterly remote sensing data on vegetation cover at a resolution of 500m (for assessing surface runoff resistance). S12, unification of spatiotemporal benchmarks: All collected data are converted to a unified coordinate system (such as WGS-84) and timestamp (UTC time), the time synchronization errors of data from different sources are calibrated (error ≤ 30s), and the data in spatially overlapping areas are processed using a weighted fusion algorithm; S13, quality grading processing: perform three-level quality assessment on the data, specifically: Level A: Real-time data with data integrity ≥ 90%, timestamp error ≤ 30s, and normal sensor calibration status, marked as core input data; Level B: Historical data with a data missing rate of 10%-30% but complete key parameters (such as tide level and typhoon intensity) are marked as auxiliary input data after spatiotemporal interpolation (error ≤ 5%); Level C: Data that exceeds the effective time limit (satellite remote sensing exceeds 6 hours, drone imagery exceeds 2 hours) or the key parameter missing rate is greater than 30% is marked as reference constrained data.

[0006] Preferably, the dynamic allocation of input weights and time window settings of the two types of features in step S2 includes: S21, Feature Extraction and Matrix Construction: Extract real-time observation features (including radar echo intensity, tide level change rate, typhoon movement vector, etc.) and historical correlation features (including historical water increase coefficients during the same period, inundation patterns of typhoons with similar paths, etc.) from the standardized data set, and construct a 10×10 feature matrix based on the time series; S22, dynamic weight allocation: Short-term warning (0-3 hours): The weight of real-time observation features is ≥70%, of which the weight of radar echo and tide level data is not less than 50%; the weight of historical correlation features is ≤30%, and feature data of similar weather types in the same period of the past three years are selected; Medium-term warning (3-12 hours): The weight of real-time observation features and historical correlation features is allocated at a 1:1 ratio. The weight of typhoon path data in real-time features is increased to 40%, and historical features include water level increase data of typhoons with similar paths in the past five years. Long-term warning (12-24 hours): The weight of historical correlation features is ≥60%, focusing on the water surge curve data of similar typhoons in history; the weight of real-time observation features is ≤40%, mainly based on typhoon intensity and movement speed data; S23, time window adjustment: the short-term warning window is set to 1 hour, the medium-term to 3 hours, and the long-term to 6 hours, and the window sliding step is 30 minutes.

[0007] Preferably, the training process of the pre-trained deep learning early warning model in step S3 includes the following sub-steps: S31, sample dataset construction: Multi-source data of historical storm surge events is used as training samples. Input features include typhoon intensity (maximum wind speed, central pressure), initial tide level (including astronomical tide prediction value), terrain parameters (elevation, slope, levee height) and trend characteristics of the past 3 hours. Output labels are the vector boundary of the measured inundation range, inundation depth values ​​of different areas, and duration; S32, Model Architecture Design: Using an improved U-Net network architecture, we added an attention gating module to the encoding side to strengthen feature learning for sensitive nearshore areas (such as dam breaches and low-lying residential areas). We also introduced skip connection feature fusion weights at the decoding side (dynamically adjusted based on regional importance). S33, model training optimization: Combine the simulation results of numerical models (such as the ADCIRC model) as auxiliary supervision signals, migrate the simulated data of extreme weather cases (such as once-in-a-century storm surges) to the training set through transfer learning, and use the Adam optimizer for training.

[0008] Preferably, the deviation adaptive correction in step S4 includes the following sub-steps: S41, historical case screening: extract candidate cases from the historical case database with a spatial similarity of ≥80% (calculated by Hausdorff distance) with the current typhoon path, a pressure gradient difference of ≤5hPa, and the same occurrence season. The number of candidate cases to be screened is controlled at 5-10; S42, optimal case matching: Calculate the topographic difference coefficient (focusing on comparing levee height and coastline morphology) and tidal phase difference (matching degree of astronomical tide rise and fall) between the candidate case and the current prediction scenario, and determine the optimal matching case using the weighted Euclidean distance formula (topographic weight 0.6, tidal weight 0.4); S43, deviation calculation and correction: Deviation coefficient = (measured water increase of the best case - model simulated water increase) / measured water increase of the best case. The deviation coefficient is dynamically corrected based on the 5-minute residual of the real-time monitoring data (measured value - predicted value). The corrected deviation is used to adjust the inundation depth and range of the preliminary prediction results. The correction range is controlled within ±10%.

[0009] Preferably, determining the warning threshold by the fuzzy analytic hierarchy process in step S5 includes the following sub-steps: S51, indicator system construction: the corrected inundation data (inundation depth, duration) and regional hazard-prone body characteristic parameters (population density, economic vulnerability, emergency response capacity) are used as first-level indicators, and each first-level indicator is divided into 3-5 second-level indicators (e.g., inundation depth is divided into interval indicators such as <0.5m, 0.5-1.0m, etc.); S52, weight assignment and fuzzy evaluation: Use the 1-9 scale method to determine the weight of each indicator, construct a fuzzy judgment matrix through triangular fuzzy numbers, and calculate the comprehensive weight value of the indicators, among which the weight of the characteristic parameters of the disaster-prone body accounts for no less than 40%; S53, threshold interval division: Based on the comprehensive evaluation results, the maximum membership principle is adopted to divide the warning threshold intervals of different sub-regions. The difference in thresholds of the same warning level in different sub-regions is controlled within ±20%.

[0010] Preferably, the threshold judgment logic of the characteristic parameters of the regional disaster-prone body in step S5 is specifically as follows: S51, construct the disaster-prone body characteristic benchmark vector V0, including the benchmark population density (500 people / km2), benchmark economic vulnerability (10 million yuan / km2), and benchmark emergency response capacity (3 shelters / 10 km2 and average evacuation distance ≤ 2 km); S52, calculating the cosine similarity S between the characteristic vector V of the disaster-prone body in the current area and the reference vector V0; S53, performing sliding window (1-hour window) statistics on the real-time disaster-prone body data in the current area to generate a characteristic change rate curve R; S54, calculating the extreme value D of the first derivative of the rate of change curve R, reflecting the degree of mutation of the characteristic parameter; S55, set the vulnerability correlation coefficient K = 0.3 × population density weight + 0.4 × economic vulnerability weight + 0.3 × emergency capacity weight; S56: If the cosine similarity S is greater than or equal to 0.8, the extreme value of the derivative of the rate of change D is less than or equal to 5% / hour, the vulnerability correlation coefficient K is less than or equal to 0.4, and the population density mutation ΔP is less than 10%, then the current hazard-prone body characteristics are determined to be in a low-sensitivity state and the basic warning threshold is adopted; otherwise, execute S57; S57: If it is determined to be a highly sensitive state, the density clustering algorithm is used to divide the area into sub-units. S52-S55 are repeated for each sub-unit to eliminate abnormal fluctuation sub-units (such as sudden changes in population density caused by temporary large-scale gatherings). The weighted mean of the characteristic parameters of the remaining sub-units is taken as the basis for calculating the final threshold.

[0011] Preferably, the specific rules for classifying warning levels in step S5 are: A three-dimensional grading model of flooding depth, duration, and vulnerability index of the hazard-bearing body is established and stored in the early warning system database. A dynamic threshold verification mechanism is used for the triggering conditions of each level of warning, with a verification accuracy of ±5% of the measured value. This three-dimensional grading model is constructed in the following way: When the vulnerability index of the hazard-bearing body is less than 0.3, the low-risk warning sub-model is matched and the warning level is set as follows: when the flooding depth is less than 0.5m and the duration is less than 3h, a blue warning is triggered; when the flooding depth is 0.5m≤<1.0m and the duration is 3h≤<6h, a yellow warning preparatory level is triggered (internal prompt only); When the vulnerability index of the hazard-bearing body is 0.3≤<0.6, the medium-risk warning sub-model is matched, and the grading strategy is: when the flooding depth is less than 0.5m but the duration is ≥3h, or when the flooding depth is 0.5m≤<1.0m and the duration is less than 3h, a yellow warning is triggered; when the flooding depth is 1.0m≤<2.0m and the duration is 3h≤<6h, an orange warning preparatory level is triggered; When the vulnerability index of the disaster-prone body is ≥0.6, the high-risk warning enhancement strategy is activated, specifically: when the flooding depth is <1.0m but the duration is ≥6h, or when the flooding depth is 1.0m≤<2.0m and the duration is ≥3h, an orange warning is triggered; when the flooding depth is ≥2.0m or the duration is ≥12h, or when the vulnerability index is ≥0.8 and the real-time flooding depth is ≥0.5m, a red warning is directly triggered regardless of the duration; when the flooding depth is ≥1.5m and the vulnerability index is ≥0.7, the warning level is automatically upgraded by one level (such as the original orange warning is upgraded to a red warning).

[0012] Preferably, the threshold judgment logic for adapting the communication mode of the receiving object in step S6 is specifically: S61: Construct the communication object feature matrix M, which includes object type (general public / flood control command department / offshore operating vessel), historical response time (seconds), device coverage rate (%), and information complexity requirement (low / medium / high); S62: Calculate the matching degree P between different communication methods (SMS / electronic screen / encrypted data packet / Beidou voice, etc.) and the object feature matrix. The matching degree for the general public is: P = 0.4 × response timeliness compliance rate + 0.3 × device compatibility rate + 0.3 × information carrying adaptability; the matching degree for maritime vessels is: P = 0.6 × response timeliness compliance rate + 0.4 × device compatibility rate; S63, setting a communication priority threshold T, where T ≥ 0.85 for a red warning, T ≥ 0.75 for an orange warning, T ≥ 0.65 for a yellow warning, and T ≥ 0.55 for a blue warning; S64, if the matching degree P of a certain communication method is greater than or equal to the priority threshold T of the corresponding warning level, then the communication method is included in the candidate communication method set; S65: Perform secondary screening of the candidate communication mode set based on real-time channel quality (packet loss rate ≤ 5% is excellent, 5%-10% is medium, and >10% is poor). Communication modes with excellent or medium channel quality are retained. When the real-time channel quality is greater than 10% packet loss rate, the system automatically switches to the backup communication channel. The switching priority is: Beidou short message > maritime satellite > 5G emergency base station. S66: If there is only one communication method after screening, then this method is directly used for publishing; if there are multiple methods, they are sorted from high to low according to the matching degree P, and the first two are selected for collaborative publishing.

[0013] A hierarchical early warning system for storm surge inundation areas, used to implement a hierarchical early warning method for storm surge inundation areas, the system comprising: Data collection and preprocessing module: used to execute step S1, including a multi-source data collection unit, a spatiotemporal reference unification unit and a quality grading unit. The multi-source data collection unit collects real-time observation data, historical basic data and environmental related data according to data type classification. The spatiotemporal reference unification unit converts the collected data into a unified coordinate system and timestamp and processes spatially overlapping data. The quality grading unit performs a three-level quality assessment on the data. Feature extraction module: used to execute step S2, including a feature extraction unit, a matrix construction unit, a weight allocation unit and a time window adjustment unit. The feature extraction unit extracts real-time observation features and historical correlation features from the standardized data set. The matrix construction unit constructs a feature time series matrix according to the time series. The weight allocation unit dynamically adjusts the input weights of the two types of features according to the warning timeliness. The time window adjustment unit sets the time window and sliding step size corresponding to different warning timeliness. Initial prediction module: used to execute step S3, including a model storage unit and a prediction execution unit. The model storage unit stores a pre-trained deep learning warning model, which adopts an improved U-Net network architecture and is trained and optimized with a sample data set. The prediction execution unit inputs weighted features into the model and outputs preliminary prediction results through a multi-scale feature fusion module. Deviation correction module: used to execute step S4, including a case library unit, a similarity calculation unit and a deviation correction unit. The case library unit stores a historical storm surge case library. The similarity calculation unit screens candidate cases and determines the optimal matching case through a case similarity matching algorithm. The deviation correction unit calculates the deviation coefficient and dynamically corrects the preliminary prediction results in combination with residual feedback of real-time monitoring data. A hierarchical threshold module is used to execute step S5 and includes a parameter storage unit, a threshold calculation unit, and a level division unit. The parameter storage unit stores characteristic parameters of the regional hazard-prone body. The threshold calculation unit determines the warning thresholds of different sub-regions using the fuzzy analytic hierarchy process combined with the corrected inundation data. The level division unit divides warning levels according to the three-dimensional hierarchical model and includes threshold judgment logic for the characteristic parameters of the regional hazard-prone body. Warning release and feedback module: used to execute step S6, including an information generation unit, a communication adaptation unit and a feedback tracking unit. The information generation unit generates differentiated warning information according to the warning level. The communication adaptation unit releases the warning information by adapting the communication method of the receiving object and includes the threshold judgment logic for the communication method adaptation. The feedback tracking unit tracks the reach rate and response data of the warning information in real time for model iterative optimization.

[0014] The present invention discloses a hierarchical early warning method and system for storm surge inundation areas, which has the following beneficial effects: (1) This invention achieves effective integration of multi-source data through the sub-steps of multi-source heterogeneous data collection, spatiotemporal benchmark unification, and quality grading. The unified spatiotemporal benchmark ensures data consistency, and the three-level quality grading standard clarifies the priority of data use, improves data utilization and reliability, and lays a solid foundation for subsequent feature extraction and model prediction. (2) The present invention dynamically adjusts the input weights and time window lengths of real-time observation features and historical correlation features according to the warning timeliness through the dynamic feature extraction process, making short-term warnings more dependent on real-time data and long-term warnings more focused on historical data, thereby enhancing the pertinence of feature extraction. The pre-trained deep learning warning model adopts an improved U-Net network architecture, adds an attention gating module, strengthens feature learning of nearshore sensitive areas, and combines numerical model simulation results with transfer learning to enhance the model's generalization ability in extreme weather cases, effectively improving the accuracy of initial forecast results; (3) This invention combines the corrected inundation data with the characteristic parameters of the regional hazard-bearing body through the hierarchical threshold adaptation process, and uses the fuzzy analytic hierarchy process to determine the warning thresholds for different sub-regions and classify them into warning levels. The construction of the three-dimensional classification model takes into account factors such as inundation depth, duration, and hazard-bearing body vulnerability index, making the classification of warning levels more scientific and reasonable, and accurately reflecting the risk status of different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 This is a schematic block diagram of the hierarchical early warning method of the present invention; Figure 2 This is a schematic block diagram of the hierarchical early warning system of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0018] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0019] The present invention provides Figure 1The hierarchical warning method for storm surge inundation areas shown in the figure includes the following steps: S1, multi-source data preprocessing: through multi-source heterogeneous data collection, spatiotemporal benchmark unification and quality grading processing, a standardized data set is obtained, wherein the quality grading result is used for weight allocation of subsequent feature extraction; S2, dynamic feature extraction: based on the standardized data set, real-time observation features and historical correlation features are extracted, and a feature time series matrix is ​​constructed, and the input weights and time window lengths of the two types of features are dynamically adjusted according to the warning timeliness; S3, initial inundation prediction: the weighted features are input into the pre-trained deep learning warning model, and the preliminary prediction results of the storm surge inundation area are output through the multi-scale feature fusion module of the model. The preliminary prediction results include the spatiotemporal distribution data of the inundation range, depth and duration; S4, bias adaptive correction: calling the historical storm surge case library, Using a case similarity matching algorithm, the deviation coefficient between the initial prediction results and historical cases is calculated. Combined with residual feedback from real-time monitoring data, the initial prediction results are dynamically corrected to obtain corrected inundation data. S5, hierarchical threshold adaptation: Combining the corrected inundation data with pre-set regional hazard-bearing body characteristic parameters, the fuzzy analytic hierarchy process is used to determine warning thresholds for different subregions and categorize warning levels. S6, hierarchical warning issuance and feedback: Differentiated warning information is generated based on the warning level and issued via a communication method adapted to the recipient. The issuance process includes threshold judgment logic for communication method adaptation and real-time tracking of the warning information's reach and response data for model iterative optimization. Multi-source data preprocessing addresses the challenges of traditional technologies, such as the difficulty in integrating heterogeneous multi-source data, inconsistent spatiotemporal benchmarks, and uneven data quality that impact subsequent analysis. This approach enables the acquisition of standardized datasets, providing a reliable data foundation for accurate feature extraction. Dynamic feature extraction overcomes the previous issue of inaccurate warnings, which often resulted from the inability to flexibly adjust feature weights and time windows based on warning timeliness. This allows the construction of a reasonable feature time series matrix. The use of pre-trained deep learning early warning models for initial flooding predictions improves the poor accuracy of traditional forecasting models in predicting storm surge inundation areas, and outputs preliminary forecast results that include the spatiotemporal distribution of flooding range, depth, and duration. Adaptive deviation correction solves the problem of large deviations between preliminary forecast results and actual conditions and the difficulty in effective correction. Corrected flooding data is obtained through historical case comparison and real-time monitoring feedback. Graded threshold adaptation solves the problem of unreasonable warning threshold settings and inadequate consideration of regional characteristics. It realizes the determination of warning thresholds and the classification of warning levels based on the characteristic parameters of regional disaster-prone bodies. Graded warning release and feedback solves the problems of lack of pertinence in warning information release, low reach and response rates, and difficulty in iterative optimization of models. Information is released through adaptive communication methods and feedback data is tracked for model improvement.

[0020] Among them, step S1 includes the following sub-steps: S11, multi-source heterogeneous data acquisition: classified acquisition according to data type, specifically including: real-time observation data: acquisition of reflectivity factor data of radar echo (spatial resolution ≤1km, time resolution ≤6 minutes), minute-level tide level and flow velocity coupling data of tide stations, real-time tracking data of latitude and longitude and intensity (wind speed, air pressure) of typhoon paths, and 0.1m resolution flood boundary images of nearshore aerial photography taken by drones; historical basic data: acquisition of hourly records of storm surge water increase processes in the past three decades (including the highest tide level and water increase amplitude), 1m precision elevation data of coastal terrain (including three-dimensional information of artificial structures such as embankments and roads), historical disaster loss ledgers (including correlation data of flooding depth and economic losses); environmental correlation data: acquisition of monthly average sea level change trend data in the past five years, protection level and disaster resistance capacity assessment data of coastal projects (such as seawall design standards and embankment failure risk factors), and 500m resolution vegetation coverage quarterly remote sensing data (for evaluating surface runoff resistance); S12, spatiotemporal benchmark statistics 1. Convert all collected data to a unified coordinate system (such as the WGS-84 coordinate system) and timestamp (UTC time), calibrate the time synchronization errors of data from different sources (error ≤ 30s), and use a weighted fusion algorithm to process data in spatially overlapping areas; S13, quality grading processing: Perform a three-level quality assessment on the data, specifically: Level A: Real-time data with data integrity ≥ 90%, timestamp error ≤ 30s, and normal sensor calibration status, marked as core input data; Level B: Historical data with a data missing rate of 10%-30% but complete key parameters (such as tide level and typhoon intensity), marked as auxiliary input data after spatiotemporal interpolation (error ≤ 5%); Level C: Data that exceeds the effective time limit (satellite remote sensing exceeds 6 hours, drone imagery exceeds 2 hours) or has a key parameter missing rate of > 30% is marked as reference constraint data. Through multi-source heterogeneous data collection, the problems of incomplete data collection and unclear classification are solved, and real-time observations, historical basis, and environmental related data are comprehensively collected to provide a rich data source for subsequent analysis. Unified spatiotemporal benchmarks address the confusion and inability to collaboratively analyze data from different sources. All data are converted to a unified coordinate system and timestamp, with error correction and handling of spatially overlapping data. Quality grading addresses the lack of clear data quality standards and the resulting inefficiencies in data utilization. A three-level quality assessment clarifies core inputs, auxiliary inputs, and reference constraint data, improving data utilization efficiency.

[0021] Specifically, the input weights and time window settings for the two types of features dynamically allocated in step S2 include: S21, feature extraction and matrix construction: extracting real-time observation features (including radar echo intensity, tide level change rate, typhoon moving vector, etc.) and historical correlation features (including historical water increase coefficients in the same period, inundation patterns of typhoons with similar paths, etc.) from the standardized data set, and constructing a 10×10 dimension feature matrix according to the time series; S22, dynamic weight allocation: short-term warning (0-3 hours): real-time observation feature weight ≥70%, of which radar echo and tide level data weight account for no less than 50%; historical correlation feature weight ≤30%, selecting feature data of similar weather types in the same period in the past three years; medium-term warning (3-12 ... The weighting of historical correlation features is allocated 1:1. The weight of typhoon path data in real-time features is increased to 40%, and historical features incorporate water-increase patterns from typhoons with similar paths over the past five years. For long-term warnings (12-24 hours), historical correlation features are weighted ≥ 60%, focusing on water-increase curves from similar typhoons. Real-time observation features are weighted ≤ 40%, primarily based on typhoon intensity and speed data. S23, time window adjustment: The short-term warning window is set to 1 hour, the medium-term to 3 hours, and the long-term to 6 hours, with a sliding step of 30 minutes. Feature extraction and matrix construction address the issues of incomplete and lack of systematic feature extraction. Real-time observation and historical correlation features are comprehensively extracted from standardized datasets and a feature matrix is ​​constructed. Dynamic weight allocation addresses the issue of irrational weighting of features with different timeliness during the warning process. The weights of real-time observation and historical correlation features are appropriately allocated based on the different needs of short-term, medium-term, and long-term warnings. Time window adjustment addresses the issue of time window settings failing to adapt to changes in warning timeliness. Appropriate time windows and sliding steps are set based on different warning timelinesses, ensuring that warnings are more tailored to actual conditions.

[0022] Among them, the training process of the pre-trained deep learning warning model in step S3 includes the following sub-steps: S31, sample data set construction: multi-source data of historical storm surge events are used as training samples, and the input features include typhoon intensity (maximum wind speed, central air pressure), initial tide level (including astronomical tide prediction value), terrain parameters (elevation, slope, embankment height) and trend characteristics of the past 3 hours. The output label is the vector boundary of the measured inundation range, the inundation depth value and duration of different areas; S32, model architecture design: adopting the improved U-Net network architecture (the improved U-Net network retains the traditional "encoding-decoding" symmetric structure, the left encoding end is responsible for feature extraction, and the right decoding end is responsible for feature extraction). Responsible for feature reconstruction, fusing features of the same scale through jump connections in the middle. An attention gating module is added to the encoding end to strengthen the feature learning of nearshore sensitive areas (such as dams and low-lying areas) through weight adjustment; a dynamic weight allocation mechanism is introduced in the jump connection at the decoding end, and the fusion weight is adjusted in combination with the regional importance to make the features more in line with the prediction needs of different regions. The overall architecture improves the prediction accuracy of key areas flooded by storm surges and adapts to the needs of graded warnings), and adds an attention gating module at the encoding end (the attention gating module is used for the improved U-Net network encoding end, and its function is to strengthen the feature learning of nearshore sensitive areas. This module receives the feature map of the current convolution block at the encoding end and the feature map of the corresponding previous layer at the decoding end. Feature map, map the two to the same channel dimension through 1×1 convolution, and generate an attention weight map after activation function processing. In the weight map, the weight values ​​corresponding to the positions of near-shore sensitive areas (such as dams and low-lying areas) are higher, and the weight values ​​of non-sensitive areas are lower. Multiplying this weight map by the current feature map on the encoding end element by element can enhance the expression of sensitive area features, suppress irrelevant area features, and make the network more focused on feature extraction of key warning areas, thereby improving prediction targeting), strengthen feature learning of near-shore sensitive areas (such as dam breach points and low-lying residential areas), and introduce jump connection feature fusion weights at the decoding end (dynamically adjusted according to regional importance); S33, model training optimization: combined with numerical mode The simulation results of the ADCIRC model are used as auxiliary supervision signals. Through transfer learning, simulated data of extreme weather cases (such as a 100-year storm surge) are transferred to the training set. The Adam optimizer is used in the training process of the improved U-Net network model. Its function is to accelerate model convergence and improve training stability by dynamically adjusting the update step size of model parameters. This optimizer combines the advantages of momentum gradient descent and adaptive learning rate. It records the first-order moment estimate (momentum) and second-order moment estimate (the basis for adaptive learning rate adjustment) of the parameter update, and corrects these two estimates through a bias correction mechanism, thereby dynamically determining the update amplitude of each parameter.In the present invention, the Adam optimizer is used to train the model, which can make the model converge to a better state faster, reduce the oscillation during the training process, ensure that the model is more stable and efficient when learning the relevant features of the storm surge inundated area, and improve the predictive performance of the model. The sample data set is constructed to solve the problems of unrepresentative model training samples and unreasonable feature selection. The historical storm surge multi-source data is used as samples to select reasonable input features and output labels to provide high-quality data for model training. The model architecture design solves the problems of insufficient feature learning of nearshore sensitive areas and unreasonable feature fusion of traditional models. An improved U-Net network architecture is adopted, and an attention gate module is added to strengthen the feature learning of nearshore sensitive areas. The feature fusion weights of the jump connection at the decoding end are dynamically adjusted. Model training optimization solves the problems of poor generalization ability of model training and difficulty in dealing with extreme weather cases. The numerical model simulation results are combined as auxiliary supervision signals, and the extreme weather case data is migrated to the training set through transfer learning. The Adam optimizer is used to improve the training effect.

[0023] Among them, the adaptive deviation correction in step S4 includes the following sub-steps: S41, historical case screening: extracting from the historical case library the spatial similarity with the current typhoon path ≥80% (calculated by Hausdorff distance, Hausdorff distance is used to measure the similarity between two point sets, and is used in the present invention to calculate the spatial similarity between the current typhoon path and the historical typhoon path. The calculation method is: for two point sets A and B, first calculate the distance from each point in A to the nearest point in B, take the maximum value of these distances, and record it as h(A, B); then calculate the distance from each point in B to the nearest point in A, take the maximum value and record it as h(B, A); Hausdorff distance is the larger value of h(A, B) and h(B, A). In step S41 of the present invention, the spatial similarity between the current typhoon path and the historical typhoon path is calculated by Hausdorff distance, so as to screen out candidate cases with similar paths, provide reliable historical case references for subsequent deviation correction, and improve the accuracy of similar case matching.), pressure gradient difference ≤5h For candidate cases with the same Pa and occurrence season, the number of screening is controlled at 5-10; S42, optimal case matching: calculate the terrain difference coefficient (focus on comparing the levee height and coastline morphology) and tidal phase difference (matching degree of the rise and fall phases of the astronomical tide) between the candidate case and the current prediction scenario, and determine the optimal matching case through the weighted Euclidean distance formula (topography weight 0.6, tidal weight 0.4); S43, deviation calculation and correction: deviation coefficient = (measured water increase of the optimal case - model simulated water increase) / measured water increase of the optimal case, and dynamically correct the deviation coefficient based on the 5-minute residual of the real-time monitoring data (measured value - predicted value). The corrected deviation is used to adjust the flooding depth and range of the preliminary prediction results, and the correction amplitude is controlled within ±10%. Through historical case screening, the problem of accurately screening similar cases from a large number of historical cases is solved. By setting strict screening conditions, such as the spatial similarity of typhoon paths, pressure gradient differences and occurrence seasons, a suitable number of candidate cases are screened out. Optimal Case Matching solves the problem of determining the best-matching case when considering only a single factor. It comprehensively considers factors such as terrain difference coefficient and tidal phase difference, and determines the optimal matching case using a weighted Euclidean distance formula. Deviation Calculation and Correction addresses the issues of inaccurate deviation coefficient calculation and untimely correction. By dynamically correcting the deviation coefficient based on residuals from real-time monitoring data, the system accurately adjusts the initial forecast of flooding depth and range, while controlling the magnitude of the correction to ensure stable results.

[0024] Among them, the fuzzy analytic hierarchy process in step S5 to determine the warning threshold includes the following sub-steps: S51, indicator system construction: using the corrected inundation data (inundation depth, duration) and the characteristic parameters of the regional hazard-prone body (population density, economic vulnerability, and emergency response capacity) as first-level indicators, each first-level indicator is divided into 3-5 second-level indicators (for example, the inundation depth is divided into interval indicators such as <0.5m and 0.5-1.0m); S52, weight assignment and fuzzy evaluation: using the 1-9 scaling method to determine the weight of each indicator, constructing a fuzzy judgment matrix using triangular fuzzy numbers, and calculating the comprehensive weight value of the indicator, where the weight of the characteristic parameters of the hazard-prone body accounts for no less than 40%; S53, threshold interval division: based on the comprehensive evaluation results, using the maximum membership principle to divide the warning threshold intervals into different sub-regions. The difference in thresholds for the same warning level in different sub-regions is controlled within ±20%. Through the construction of the indicator system, the problems of incomplete indicators and unclear levels in determining the warning threshold are solved. The corrected inundation data and the characteristic parameters of the regional hazard-prone body are used as first-level indicators, and reasonable second-level indicators are assigned to them, forming a complete indicator system. Weight assignment and fuzzy evaluation address the subjective nature of indicator weight determination and unscientific evaluation. Weights are determined using a 1-9 scale and a fuzzy judgment matrix constructed using triangular fuzzy numbers. Comprehensive indicator weights are calculated scientifically, ensuring a reasonable weighting of characteristic parameters of the hazard-prone body. Threshold interval division addresses the issues of irrational early warning threshold interval division and insufficient reflection of regional differences. Based on comprehensive evaluation results, the maximum membership principle is used to divide early warning threshold intervals into different subregions, controlling for differences in thresholds within the same early warning level across different subregions.

[0025] Specifically, the threshold judgment logic of the characteristic parameters of the regional hazard-prone body in step S5 is as follows: S51, constructing a hazard-prone body characteristic benchmark vector V0, including the benchmark population density (500 people / square kilometer), the benchmark economic vulnerability (10 million yuan / square kilometer), and the benchmark emergency response capacity (3 shelters / 10 square kilometers and the average evacuation distance ≤ 2 kilometers); S52, calculating the cosine similarity S between the current regional hazard-prone body characteristic vector V and the benchmark vector V0; S53, performing sliding window (1-hour window) statistics on the real-time hazard-prone body data of the current region to generate a characteristic change rate curve R; S 54. Calculate the extreme value D of the first derivative of the rate of change curve R to reflect the degree of mutation of the characteristic parameter; S55. Set the vulnerability correlation coefficient K = 0.3 × population density weight + 0.4 × economic vulnerability weight + 0.3 × emergency capacity weight; S56. If the cosine similarity S ≥ 0.8, and the extreme value D of the rate of change derivative ≤ 5% / hour, and the vulnerability correlation coefficient K ≤ 0.4, and the population density mutation ΔP < 10%, then the current disaster-prone body characteristics are judged to be in a low-sensitivity state and the basic warning threshold is adopted; otherwise, execute S57; S57. If it is judged to be in a high-sensitivity state, density clustering is adopted. The algorithm (density clustering algorithm is used to group data according to the density of data points. The core idea is to divide closely connected points into the same cluster. It is suitable for identifying cluster structures with complex shapes. The main process is: by setting the radius and the minimum number of included points, the core point is identified; then the core point and all points within its radius are divided into the same cluster, and it is iterated continuously until all connectable points are assigned to the corresponding cluster, and the remaining unassigned points are noise points. In step S57 of the present invention, the density clustering algorithm is used to divide the area into sub-units, and areas with similar features can be divided into the same sub-units. The region is divided into subunits, and steps S52-S55 are repeated for each subunit to eliminate subunits with abnormal fluctuations (e.g., sudden population density changes caused by temporary large-scale gatherings). The weighted mean of the characteristic parameters of the remaining subunits is used as the basis for the final threshold calculation. By constructing a baseline vector for the hazard-prone body characteristics, the lack of a benchmark for determining the hazard-prone body's characteristic status is addressed, and indicators such as baseline population density, economic vulnerability, and emergency response capacity are clarified. Calculating cosine similarity addresses the difficulty of measuring the degree of difference between the current regional hazard-prone body characteristics and the baseline, quantifying the difference. Using sliding window statistics and calculating derivative extremes addresses the difficulty of effectively monitoring the trend and degree of sudden changes in the hazard-prone body characteristic parameters. A characteristic change rate curve is generated, and the extreme values ​​of the first-order derivative are calculated to reflect the sudden changes. Setting a vulnerability correlation coefficient addresses the difficulty of comprehensively measuring the vulnerability of the hazard-prone body. The vulnerability correlation coefficient is calculated through a reasonable weight distribution.The threshold judgment logic solves the problem of being unable to accurately determine the characteristic status of the disaster-prone body and reasonably determine the basis for warning thresholds. The characteristic status of the disaster-prone body is determined based on a series of conditions, and a reasonable algorithm is used to process highly sensitive states and determine the basis for the final threshold calculation.

[0026] Among them, the specific rules for dividing the warning level in step S5 are: establishing a three-dimensional grading model of flooding depth-duration-hazard-bearing body vulnerability index, storing it in the early warning system database, and the triggering conditions of each level of warning adopt a dynamic threshold verification mechanism with a verification accuracy of ±5% of the measured value. The three-dimensional grading model is constructed in the following way: when the vulnerability index of the hazard-bearing body is less than 0.3, the low-risk warning sub-model is matched, and the warning level is set as follows: when the flooding depth is less than 0.5m and the duration is less than 3h, a blue warning is triggered; when 0.5m≤flooding depth<1.0m and 3h≤duration<6h, a yellow warning preparatory level is triggered (internal prompt only); when 0.3≤hazard-bearing body vulnerability index<0.6, the medium-risk warning sub-model is matched, and the grading strategy is: when the flooding depth is less than 0.5m but the duration is ≥3h, or when 0.5m≤flooding depth<1.0m and the duration is less than 3h, a yellow warning is triggered. When the inundation depth is 1.0m ≤ < 2.0m and the duration is 3h ≤ < 6h, an orange alert is triggered. When the vulnerability index of the hazard-bearing body is ≥0.6, the high-risk alert enhancement strategy is activated. Specifically, an orange alert is triggered when the inundation depth is <1.0m but the duration is ≥6h, or when the inundation depth is 1.0m ≤ <2.0m and the duration is ≥3h. A red alert is triggered regardless of the duration when the inundation depth is ≥2.0m or the duration is ≥12h, or when the vulnerability index is ≥0.8 and the real-time inundation depth is ≥0.5m. When the inundation depth is ≥1.5m and the vulnerability index is ≥0.7, the alert level is automatically upgraded by one level (for example, the original orange alert is upgraded to a red alert). The establishment of a three-dimensional grading model addresses the problem of a single warning level classification standard that fails to comprehensively consider multiple factors. A comprehensive grading model is constructed based on inundation depth, duration, and the vulnerability index of the hazard-bearing body. A dynamic threshold verification mechanism addresses the problem of inflexible warning triggering conditions and their inability to adapt to actual changes. Dynamic threshold verification ensures verification accuracy, making the warning triggering more responsive to actual conditions. Different risk warning sub-models and strategies solve the problem of inaccurate warning strategies in different risk areas. Different risk levels are divided according to the vulnerability index of the disaster-prone body, and corresponding warning sub-models and grading strategies are matched to improve the targeted nature of warnings.

[0027] Among them, the threshold judgment logic of the communication mode adapted to the receiving object in step S6 is specifically as follows: S61, constructing a communication object feature matrix M, including object type (general public / flood control command department / offshore operating vessel), historical response time (seconds), equipment coverage rate (%), and information complexity requirement (low / medium / high); S62, calculating the matching degree P of different communication modes (SMS / electronic screen / encrypted data packet / Beidou voice, etc.) and the object feature matrix, the matching degree of the general public: P=0.4×response timeliness compliance rate+0.3×equipment compatibility rate+0.3×information carrying adaptability; the matching degree of offshore vessels: P=0.6×response timeliness compliance rate+0.4×equipment compatibility rate; S63, setting the communication priority threshold T, where T≥0.85 for red warning, T≥0.75 for orange warning, T≥0.65 for yellow warning, and T≥0.55 for blue warning; S64, if If the matching degree P of a certain communication mode is ≥ the priority threshold T of the corresponding warning level, it will be included in the candidate communication mode set; S65, the candidate communication mode set is secondary screened according to the real-time channel quality (packet loss rate ≤ 5% is excellent, 5%-10% is medium, and >10% is poor), and the communication modes with excellent or medium channel quality are retained. When the real-time channel quality is greater than the packet loss rate of 10%, it will automatically switch to the backup communication channel. The switching priority is: Beidou short message > maritime satellite > 5G emergency base station; S66, if there is only one communication mode after screening, it will be directly used for release; if there are multiple modes, they will be sorted from high to low according to the matching degree P, and the top two will be selected for collaborative release (such as SMS + community electronic screen, encrypted data packet + emergency command system pop-up window). By constructing a communication object feature matrix, the problem of incomplete understanding of the characteristics of the communication objects is solved, and factors such as object type, historical response time, equipment coverage, and information complexity requirements are fully considered. Calculating the matching degree of communication methods solves the problem of being unable to accurately evaluate the compatibility of communication methods and objects. According to different object types, reasonable weights such as response timeliness compliance rate, equipment compatibility rate, and information carrying adaptability are set to calculate the matching degree. Setting communication priority thresholds solves the problem of unclear priority for warning information release. According to different warning levels, corresponding communication priority thresholds are set. Screening of candidate communication methods solves the problem of unreasonable selection of communication methods and unguaranteed channel quality. Screening is performed based on matching degree and real-time channel quality to ensure the effectiveness of information release. Communication method release strategy solves the problem of being unable to make a reasonable choice when multiple communication methods are available. According to the screening results, a reasonable determination of single or collaborative release communication methods is made to improve the effectiveness of warning information release.

[0028] The present invention provides Figure 2The hierarchical early warning system for storm surge inundation areas shown in the figure comprises: a data acquisition and preprocessing module: for executing step S1, comprising a multi-source data acquisition unit, a time-space benchmark unification unit and a quality grading unit, wherein the multi-source data acquisition unit collects real-time observation data, historical basic data and environmental related data according to data type classification, the time-space benchmark unification unit converts the collected data into a unified coordinate system and timestamp and processes spatially overlapping data, and the quality grading unit performs a three-level quality assessment on the data; a feature extraction module: for executing step S2, comprising a feature extraction unit, a matrix construction unit, a weight allocation unit and a time window adjustment unit, wherein the feature extraction unit extracts real-time observation data from the standardized data set. Features are associated with historical features. The matrix construction unit constructs a feature time series matrix according to the time series. The weight allocation unit dynamically adjusts the input weights of the two types of features according to the warning timeliness. The time window adjustment unit sets the time window and sliding step corresponding to different warning timeliness. The initial prediction module is used to execute step S3, including a model storage unit and a prediction execution unit. The model storage unit stores a pre-trained deep learning warning model. The model adopts an improved U-Net network architecture and is trained and optimized with a sample data set. The prediction execution unit inputs the weighted features into the model and outputs the preliminary prediction results through the multi-scale feature fusion module. The deviation correction module is used to execute step S4, including a case library unit. , similarity calculation unit and deviation correction unit, the case library unit stores the historical storm surge case library, the similarity calculation unit screens candidate cases and determines the optimal matching case through the case similarity matching algorithm, the deviation correction unit calculates the deviation coefficient and dynamically corrects the preliminary prediction results in combination with the residual feedback of the real-time monitoring data; the graded threshold module: used to execute step S5, including a parameter storage unit, a threshold calculation unit and a grade division unit, the parameter storage unit stores the characteristic parameters of the regional disaster-bearing body, the threshold calculation unit determines the warning threshold of different sub-regions through the fuzzy hierarchical analysis method combined with the corrected inundation data, the grade division unit divides the warning level according to the three-dimensional grading model, and includes the regional disaster-bearing body characteristics The threshold judgment logic of the parameters; the early warning release and feedback module: used to execute step S6, including an information generation unit, a communication adaptation unit and a feedback tracking unit. The information generation unit generates differentiated early warning information according to the early warning level. The communication adaptation unit releases the early warning information by adapting the communication method of the receiving object and includes the threshold judgment logic of the communication method adaptation. The feedback tracking unit tracks the reach rate and response data of the early warning information in real time for model iterative optimization. Among them, the data acquisition and preprocessing module, through the collaborative work of the multi-source data acquisition unit, the spatiotemporal benchmark unification unit and the quality grading unit, solves the problems in data collection, integration and quality assessment, and realizes the acquisition of high-quality standardized data sets.The feature extraction module, utilizing a feature extraction unit, a matrix construction unit, a weight assignment unit, and a time window adjustment unit, addresses issues related to feature extraction, matrix construction, and irrational weight and time window settings, thereby constructing a feature time series matrix that meets the timeliness requirements of early warnings. The initial prediction module, leveraging a model storage unit and a prediction execution unit, addresses model storage and prediction execution issues, utilizing high-quality models to output reliable preliminary predictions. The deviation correction module, utilizing a case library unit, a similarity calculation unit, and a deviation correction unit, addresses historical case screening, matching, and deviation correction, improving prediction accuracy. The grading threshold module, leveraging a parameter storage unit, a threshold calculation unit, and a grading unit, addresses parameter storage, threshold calculation, and grading, achieving scientific and rational early warning grading. The early warning release and feedback module, leveraging an information generation unit, a communication adapter unit, and a feedback tracking unit, addresses issues related to early warning information generation, release, and feedback tracking, improving the effectiveness of early warning releases and the model's iterative optimization capabilities. Through the coordinated operation of these modules, the entire system comprehensively addresses a series of issues associated with graded early warnings for storm surge inundation areas, enhancing the accuracy, relevance, and effectiveness of early warnings.

[0029] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A hierarchical early warning method for storm surge inundated areas, characterized in that: The following steps are involved: S1, multi-source data preprocessing: Through multi-source heterogeneous data collection, spatiotemporal benchmark unification and quality grading, a standardized data set is obtained. The quality grading results are used for weight allocation in subsequent feature extraction; S2, dynamic feature extraction: Based on the standardized data set, extract real-time observation features and historical correlation features, construct a feature time series matrix, dynamically adjust the input weights and time window lengths of the two types of features according to the warning timeliness, and dynamically allocate the input weights and time window settings of the two types of features, including: S21, Feature Extraction and Matrix Construction: Extract real-time observation features and historical correlation features from the standardized data set, and construct a 10×10 dimension feature matrix according to the time series; S22, dynamic weight allocation: Short-term warning: The weight of real-time observation features is ≥70%, of which radar echo and tide level data account for no less than 50%; the weight of historical correlation features is ≤30%, and feature data of similar weather types in the same period of the past three years are selected; Medium-term warning: The weight of real-time observation features and historical correlation features is allocated at a ratio of 1:

1. The weight of typhoon path data in real-time features is increased to 40%, and historical features include water level increase data of typhoons with similar paths in the past five years; Long-term warning: The weight of historical correlation features is ≥60%, focusing on the water surge curve data of similar typhoons in history; the weight of real-time observation features is ≤40%, mainly based on typhoon intensity and movement speed data; S23, time window adjustment: the short-term warning window is set to 1 hour, the medium-term to 3 hours, and the long-term to 6 hours, with a window sliding step of 30 minutes; S3, initial flooding prediction: The weighted features are input into the pre-trained deep learning warning model, and the multi-scale feature fusion module outputs the preliminary prediction results; S4, bias adaptive correction: call the historical storm surge case library, use the deviation coefficient and combine it with real-time monitoring data to correct the initial prediction and obtain the corrected inundation data; S5, hierarchical threshold adaptation: combining the corrected inundation data with the preset regional disaster-bearing body characteristic parameters, the fuzzy analytic hierarchy process is used to determine the warning thresholds for different sub-regions and divide the warning levels; S6, graded warning release and feedback: Differentiated warning information is generated according to the warning level and released through the communication method adapted to the receiving object. The release process includes the threshold judgment logic adapted to the communication method, and real-time tracking of the warning information's reach and response data for model iterative optimization.

2. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: Step S1 includes the following sub-steps: S11, multi-source heterogeneous data collection: classified collection by data type; S12, unification of spatiotemporal benchmarks: convert all collected data to a unified coordinate system and timestamp, calibrate the time synchronization errors of data from different sources, and use a weighted fusion algorithm to process data in spatially overlapping areas; S13, quality grading processing: perform three-level quality assessment on the data.

3. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: Step S3 specifically includes: inputting the weighted features into a pre-trained deep learning early warning model, and outputting a preliminary prediction result of the storm surge inundation area through the model's multi-scale feature fusion module. The preliminary prediction result includes the spatiotemporal distribution data of the inundation range, depth, and duration. The training process of the pre-trained deep learning early warning model includes the following sub-steps: S31, sample dataset construction: Multi-source data of historical storm surge events are used as training samples. Input features include typhoon intensity, initial tide level, terrain parameters, and trend characteristics of the past 3 hours. Output labels are the vector boundaries of the measured inundation range, inundation depth values ​​of different areas, and duration. S32, model architecture design: adopting an improved U-Net network architecture, adding an attention gating module on the encoding side to strengthen feature learning of near-shore sensitive areas, and introducing feature fusion weights of skip connections on the decoding side; S33, model training optimization: Combine the simulation results of the numerical model as an auxiliary supervision signal, migrate the simulated data of extreme weather cases to the training set through transfer learning, and use the Adam optimizer for training.

4. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: Step S4 specifically involves calling the historical storm surge case library, calculating the deviation coefficient between the preliminary prediction result and the historical case through the case similarity matching algorithm, and dynamically correcting the preliminary prediction result in combination with the residual feedback of the real-time monitoring data to obtain the corrected inundation data. The deviation adaptive correction includes the following sub-steps: S41, historical case screening: extract candidate cases from the historical case database with a spatial similarity of ≥80% with the current typhoon path, a pressure gradient difference of ≤5hPa, and the same occurrence season. The number of candidate cases to be screened is controlled at 5-10. S42, optimal case matching: Calculate the terrain difference coefficient and tidal phase difference between the candidate case and the current prediction scenario, and determine the optimal matching case using the weighted Euclidean distance formula; S43, deviation calculation and correction: Deviation coefficient = (measured water increase of the best case - model simulated water increase) / measured water increase of the best case. The deviation coefficient is dynamically corrected based on the 5-minute residual of the real-time monitoring data. The corrected deviation is used to adjust the inundation depth and range of the preliminary forecast results.

5. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: In step S5, the fuzzy analytic hierarchy process for determining the warning threshold includes the following sub-steps: S51, indicator system construction: the corrected inundation data and regional hazard-bearing body characteristic parameters are used as first-level indicators, and each first-level indicator is divided into 3-5 second-level indicators; S52, weight assignment and fuzzy evaluation: Use the 1-9 scale method to determine the weight of each indicator, construct a fuzzy judgment matrix through triangular fuzzy numbers, and calculate the comprehensive weight value of the indicators, among which the weight of the characteristic parameters of the disaster-prone body accounts for no less than 40%; S53, threshold interval division: Based on the comprehensive evaluation results, the maximum membership principle is adopted to divide the warning threshold intervals of different sub-regions. The difference in thresholds of the same warning level in different sub-regions is controlled within ±20%.

6. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: The specific logic for determining the threshold value of the characteristic parameters of the regional disaster-prone body in step S5 is: S51, construct the disaster-prone body characteristic benchmark vector V0, including benchmark population density, benchmark economic vulnerability, and benchmark emergency response capacity; S52, calculating the cosine similarity S between the characteristic vector V of the disaster-prone body in the current area and the reference vector V0; S53, performing sliding window statistics on the real-time disaster-prone body data in the current area to generate a characteristic change rate curve R; S54, calculating the extreme value D of the first derivative of the rate of change curve R, reflecting the degree of mutation of the characteristic parameter; S55, set the vulnerability correlation coefficient K = 0.3 × population density weight + 0.4 × economic vulnerability weight + 0.3 × emergency capacity weight; S56: If the cosine similarity S is greater than or equal to 0.8, the extreme value of the derivative of the rate of change D is less than or equal to 5% / hour, the vulnerability correlation coefficient K is less than or equal to 0.4, and the population density mutation ΔP is less than 10%, then the current hazard-prone body characteristics are determined to be in a low-sensitivity state and the basic warning threshold is adopted; otherwise, execute S57; S57: If it is determined to be a highly sensitive state, the density clustering algorithm is used to divide the area into sub-units, and S52-S55 are repeated for each sub-unit to eliminate abnormal fluctuation sub-units. The weighted mean of the characteristic parameters of the remaining sub-units is taken as the basis for calculating the final threshold.

7. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: The specific rules for classifying warning levels in step S5 are: A three-dimensional grading model of flooding depth, duration, and vulnerability index of the hazard-bearing body is established and stored in the early warning system database. A dynamic threshold verification mechanism is used for the triggering conditions of each level of warning, with a verification accuracy of ±5% of the measured value. This three-dimensional grading model is constructed in the following way: When the vulnerability index of the hazard-bearing body is less than 0.3, the low-risk warning sub-model is matched, and the warning level is set as follows: when the flooding depth is less than 0.5m and the duration is less than 3h, a blue warning is triggered; When the flooding depth is 0.5m≤<1.0m and the duration is 3h≤<6h, the yellow warning preparatory level is triggered; When the vulnerability index of the hazard-bearing body is 0.3≤<0.6, the medium-risk warning sub-model is matched, and the grading strategy is: when the flooding depth is less than 0.5m but the duration is ≥3h, or when the flooding depth is 0.5m≤<1.0m and the duration is less than 3h, a yellow warning is triggered; When the flooding depth is 1.0m≤<2.0m and the duration is 3h≤<6h, the orange warning preparatory level is triggered; When the vulnerability index of the disaster-prone body is ≥0.6, the high-risk warning enhancement strategy is activated, specifically: when the flooding depth is <1.0m but the duration is ≥6h, or when the flooding depth is 1.0m≤<2.0m and the duration is ≥3h, an orange warning is triggered; when the flooding depth is ≥2.0m or the duration is ≥12h, or when the vulnerability index is ≥0.8 and the real-time flooding depth is ≥0.5m, a red warning is directly triggered regardless of the duration; when the flooding depth is ≥1.5m and the vulnerability index is ≥0.7, the warning level is automatically raised by one level.

8. The hierarchical early warning method for storm surge inundation areas according to claim 1, characterized in that: The threshold judgment logic for adapting the communication mode of the receiving object in step S6 is specifically as follows: S61, constructing a communication object feature matrix M, including object type, historical response time, device coverage, and information complexity requirements; S62, calculate the matching degree P of different communication modes and the object feature matrix. The matching degree for the general public is: P = 0.4 × response time efficiency compliance rate + 0.3 × device compatibility rate + 0.3 × information carrying adaptability; the matching degree for maritime vessels is: P = 0.6 × response time efficiency compliance rate + 0.4 × device compatibility rate; S63, setting a communication priority threshold T, where T ≥ 0.85 for a red warning, T ≥ 0.75 for an orange warning, T ≥ 0.65 for a yellow warning, and T ≥ 0.55 for a blue warning; S64, if the matching degree P of a certain communication method is greater than or equal to the priority threshold T of the corresponding warning level, then the communication method is included in the candidate communication method set; S65, performing a secondary screening of the candidate communication mode set according to the real-time channel quality, and retaining the communication modes with excellent or medium channel quality; S66: If there is only one communication method after screening, then this method is directly used for publishing; if there are multiple methods, they are sorted from high to low according to the matching degree P, and the first two are selected for collaborative publishing.

9. A hierarchical early warning system for storm surge inundation areas, characterized in that: The system is used to implement the hierarchical early warning method for storm surge inundation areas according to any one of claims 1 to 8, comprising: Data collection and preprocessing module: used to execute step S1, including a multi-source data collection unit, a spatiotemporal reference unification unit and a quality grading unit. The multi-source data collection unit collects real-time observation data, historical basic data and environmental related data according to data type classification. The spatiotemporal reference unification unit converts the collected data into a unified coordinate system and timestamp and processes spatially overlapping data. The quality grading unit performs a three-level quality assessment on the data. Feature extraction module: used to execute step S2, including a feature extraction unit, a matrix construction unit, a weight allocation unit and a time window adjustment unit. The feature extraction unit extracts real-time observation features and historical correlation features from the standardized data set. The matrix construction unit constructs a feature time series matrix according to the time series. The weight allocation unit dynamically adjusts the input weights of the two types of features according to the warning timeliness. The time window adjustment unit sets the time window and sliding step size corresponding to different warning timeliness. Initial prediction module: used to execute step S3, including a model storage unit and a prediction execution unit. The model storage unit stores a pre-trained deep learning warning model, which adopts an improved U-Net network architecture and is trained and optimized with a sample data set. The prediction execution unit inputs weighted features into the model and outputs preliminary prediction results through a multi-scale feature fusion module. Deviation correction module: used to execute step S4, including a case library unit, a similarity calculation unit and a deviation correction unit. The case library unit stores a historical storm surge case library. The similarity calculation unit screens candidate cases and determines the optimal matching case through a case similarity matching algorithm. The deviation correction unit calculates the deviation coefficient and dynamically corrects the preliminary prediction results in combination with residual feedback of real-time monitoring data. A hierarchical threshold module is used to execute step S5 and includes a parameter storage unit, a threshold calculation unit, and a level division unit. The parameter storage unit stores characteristic parameters of the regional hazard-prone body. The threshold calculation unit determines the warning thresholds of different sub-regions using the fuzzy analytic hierarchy process combined with the corrected inundation data. The level division unit divides warning levels according to the three-dimensional hierarchical model and includes threshold judgment logic for the characteristic parameters of the regional hazard-prone body. Warning release and feedback module: used to execute step S6, including an information generation unit, a communication adaptation unit and a feedback tracking unit. The information generation unit generates differentiated warning information according to the warning level. The communication adaptation unit releases the warning information by adapting the communication method of the receiving object and includes the threshold judgment logic for the communication method adaptation. The feedback tracking unit tracks the reach rate and response data of the warning information in real time for model iterative optimization.

Images