A natural disaster risk early warning method based on sky-ground multi-source data fusion

Abstract

The application provides a natural disaster risk early warning method based on sky-ground multi-source data fusion, and relates to the technical field of natural disaster risk early warning. Through multi-source data feature construction, time and space alignment and data standardization are realized. A parallel double-channel architecture is adopted to combine data-driven and knowledge-driven for risk assessment. Based on deep learning assisted Kalman filtering, the fusion weight is dynamically adjusted. The multi-source heterogeneous information is effectively integrated and the evaluation accuracy is improved. The method has the advantages of efficiently fusing sky-ground multi-source heterogeneous data, realizing dynamic and accurate evaluation of disaster risk, and improving the robustness and interpretability of the early warning system.

Description

Technical Field

[0001] This application relates to the field of natural disaster risk early warning technology, and in particular to a natural disaster risk early warning method that integrates multi-source data from the sky and ground. Background Technology

[0002] In practical applications of natural disaster risk early warning, the acquisition and processing of multi-source data from the sky and ground face multiple obstacles. Environmental complexity manifests in the undulating terrain, sudden weather changes, and differences in land cover in disaster-affected areas. This makes satellite remote sensing images susceptible to cloud interference, drone patrols limited by visibility and wind speed conditions, and ground monitoring equipment prone to signal attenuation or failure in extreme environments. Spatiotemporal dynamism is reflected in the rapid and uncertain nature of disaster evolution. Different data sources have significant differences in acquisition frequency and coverage. For example, satellite remote sensing data has a long update cycle and wide coverage, while drone patrols can provide high-precision local information but are limited by flight time. Ground monitoring sensors have strong real-time performance but are sparsely distributed spatially, and historical disaster records are discrete and lagging, making it difficult to align data in time series and match data in spatial locations. Data heterogeneity further exacerbates the processing difficulty. Satellite remote sensing data presents macroscopic information such as surface temperature and vegetation cover in the form of raster images, while UAV patrol data focuses on microscopic details such as slope cracks and road blockages. Ground monitoring data contains continuous time-series parameters such as rainfall and soil moisture, while historical disaster records involve structured information such as risk zoning and building distribution. These data differ in format, unit, precision, and semantic expression, making them difficult to integrate and utilize directly.

[0003] Traditional early warning methods based on a single data source, such as relying solely on satellite remote sensing data to assess flood risk or issuing warnings based solely on ground rainfall thresholds, fail to fully utilize the complementary advantages of multi-source data, leading to distorted depictions of the dynamic evolution of disasters. Fixed threshold mechanisms are insufficiently adaptable to changing environmental conditions, such as being prone to false alarms in scenarios of sudden torrential rain following prolonged drought. While purely data-driven models can uncover potential correlations from historical data, their decision-making logic lacks transparency. When data is missing or sensors malfunction, the model output is prone to drastic fluctuations, making it difficult to gain user trust. In the multi-source risk information fusion stage, existing technologies generally employ static weight allocation or simple linear combination strategies, failing to dynamically adjust the fusion ratio based on the real-time quality of the data sources. For example, fixed weights are still assigned even when satellites are obscured by clouds or drone signals are interrupted, resulting in decreased accuracy and robustness of the fusion results. Furthermore, the lack of a unified risk quantification framework for key disaster-bearing scenarios such as residential areas, buildings, and roads and bridges makes it difficult to accurately connect early warning information with specific disaster prevention measures, affecting the timeliness and targeting of emergency responses.

[0004] The aforementioned problems severely restrict the practical effectiveness of natural disaster risk early warning systems, and there is an urgent need to overcome the technical bottlenecks in deep fusion of multi-source data, interpretability of risk assessment models, and adaptability of fusion mechanisms.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] The purpose of this application is to provide a natural disaster risk early warning method based on the fusion of multi-source data from the sky and ground. This method has the advantages of efficiently fusing heterogeneous multi-source data from the sky and ground, realizing dynamic and accurate assessment of disaster risks, and improving the robustness and interpretability of the early warning system.

[0007] The natural disaster risk early warning method provided in this application, which integrates multi-source data from the sky and ground, adopts the following technical solution: A method for early warning of natural disaster risks through the fusion of multi-source data from the sky and ground includes: Acquire multi-source data from the sky and ground, which includes at least satellite remote sensing data, UAV patrol data, ground monitoring data, and historical disaster records; Multi-source feature construction is performed on the aforementioned multi-source data from the sky and ground to generate multi-source feature input. The multi-source feature construction includes at least spatiotemporal alignment, missing data imputation, continuous feature standardization, and discrete feature fuzzification. A parallel dual-channel architecture is adopted to perform disaster risk assessment on the multi-source feature input. The parallel dual-channel architecture includes a data-driven channel and a knowledge-driven channel. The data-driven channel outputs a data-driven risk score based on a deep learning model. The knowledge-driven channel outputs a knowledge-driven risk score based on expert rule reasoning. Based on deep learning-assisted Kalman filtering, the data-driven risk score and the knowledge-driven risk score are adaptively fused to output the final disaster risk score. Based on the final disaster risk score, natural disaster risk early warning information is generated.

[0008] Optionally, the construction of multi-source features from the sky-ground multi-source data includes: Preliminary quality control and deviation correction are performed on the multi-source data from the sky and ground, and abnormal data is marked or removed. Time series alignment of multi-source data with different sampling frequencies is performed based on the least squares method, unifying the data sources to the same time scale. Spatial alignment of multi-source data with different spatial resolutions is performed based on spatial grid mapping, mapping satellite remote sensing data, UAV patrol data and ground monitoring data to a unified spatial grid cell; The missing data after alignment is filled using Lagrange interpolation. Continuous features are normalized by range and mapped to a preset interval; The discrete features are fuzzified, and the membership degree of each feature to different fuzzy sets is calculated using membership functions to generate a fuzzified discrete feature set.

[0009] Optionally, the data-driven channel outputs a data-driven risk score based on a deep learning model, including: Construct a Transformer time series prediction model based on a multi-head attention mechanism; The normalized continuous feature sequence is input into the Transformer time series prediction model, and long-distance temporal dependencies are captured in parallel through a multi-head attention mechanism to output a data-driven risk score. Historical disaster records are used as supervisory labels, and the parameters of the Transformer time series prediction model are optimized through a loss function.

[0010] Optionally, the knowledge-driven channel outputs a knowledge-driven risk score based on expert rule reasoning, including: Construct a Mamdani-type fuzzy inference system based on pre-defined expert rules, wherein the expert rules are in IF-THEN form; The fuzzy discrete feature set is matched with the expert rules, and the trigger strength of each rule is calculated. Max-min composite reasoning is used to prune and aggregate the conclusions of each rule according to the triggering intensity, generating a comprehensive output fuzzy set. The centroid method is used to defuzzify the comprehensive output fuzzy set and output a knowledge-driven risk score.

[0011] Optionally, the adaptive fusion of the data-driven risk score and the knowledge-driven risk score based on deep learning-assisted Kalman filtering includes: Construct a random walk state equation to describe the temporal evolution characteristics of disaster risk scores; The data-driven risk score and the knowledge-driven risk score are used as two observation inputs to construct an observation vector and an observation model. Based on fundamental rules and statistical methods, the basic values ​​of process noise covariance and observation noise covariance are estimated. Based on a deep learning network, the process noise covariance and the observation noise covariance are dynamically corrected to generate adaptive noise parameters. Based on the adaptive noise parameters, the Kalman gain is calculated, and the predicted and observed values ​​are weighted and fused according to the Kalman gain to output the final disaster risk score.

[0012] Optionally, the dynamic correction of the process noise covariance and the observation noise covariance based on a deep learning network includes: Construct a network input feature vector, wherein the feature vector includes at least two risk scores, the change magnitude of the two risk scores over time, an environmental change intensity index, and a data quality index; The feature vector is input into a multilayer perceptron network, and feature mapping is performed through a nonlinear activation function to output process noise correction factor and observation noise correction factor. The Softplus function is used to apply non-negative constraints to the process noise correction factor and the observation noise correction factor. The adaptive noise parameters are generated based on the base value and the correction factor.

[0013] Optionally, the deep learning-assisted Kalman filter further includes closed-loop optimization: The final disaster risk score is compared with the actual disaster risk score label to construct a fusion loss function; The parameters of the deep learning network are updated by using the backpropagation algorithm, enabling the deep learning network to learn to adaptively adjust noise parameters. It continuously receives actual disaster feedback information from subsequent moments, and performs incremental updates or periodic offline retraining based on the error between the real labels and the fusion results, forming a closed-loop adaptive mechanism.

[0014] Optionally, generating natural disaster risk early warning information based on the final disaster risk score includes: Based on the final disaster risk score, the disaster-bearing scenario units are ranked by risk. Based on the preset risk threshold range, determine the early warning level for each disaster-affected scenario unit; Based on the warning level and the attribute information of the disaster-bearing scenario unit, differentiated disaster prevention and mitigation emergency decision-making suggestions are generated.

[0015] Optionally, the multi-source air-ground data includes: The satellite remote sensing data includes at least surface temperature, vegetation cover, water distribution, and surface deformation data. The drone patrol data includes at least data on slope cracks, surface subsidence, water accumulation, road blockages, and changes in the surrounding environment of buildings. The ground monitoring data includes at least rainfall, wind speed, temperature and humidity, water level, soil moisture, and surface displacement data. The historical disaster record data includes at least risk zoning data and information on the spatial distribution of residential areas and buildings.

[0016] In summary, this application achieves spatiotemporal alignment and data standardization through multi-source data feature construction, adopts a parallel dual-channel architecture combining data-driven and knowledge-driven risk assessment, and dynamically adjusts the fusion weights based on deep learning-assisted Kalman filtering, effectively integrating multi-source heterogeneous information and improving assessment accuracy. It has the advantages of efficiently fusing multi-source heterogeneous data from the sky and ground, realizing dynamic and accurate assessment of disaster risks, and improving the robustness and interpretability of the early warning system. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the first embodiment of the natural disaster risk early warning method based on multi-source data fusion from the sky and ground in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] Traditional natural disaster risk early warning methods face challenges such as data heterogeneity, spatiotemporal dynamics, and data gaps when processing multi-source data from the sky and ground, making it difficult to fully utilize multi-source information and accurately depict the evolution of disasters. Purely data-driven models lack transparency in their early warning results and are prone to instability when data is missing or observation quality changes. Existing fusion methods typically employ fixed weights or static linear fusion approaches, failing to dynamically adjust fusion weights based on the observation quality and environmental changes of different data sources, thus affecting the accuracy and stability of early warning results.

[0020] This application provides a method for early warning of natural disaster risks through the fusion of multi-source data from the sky and ground, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the natural disaster risk early warning method based on multi-source data fusion from the sky and ground in this application.

[0021] Step S10: Obtain multi-source data from the sky and ground, which includes at least satellite remote sensing data, UAV patrol data, ground monitoring data, and historical disaster record data.

[0022] The multi-source data from the sky and ground includes: satellite remote sensing data, including at least surface temperature, vegetation cover, water distribution, and surface deformation data; UAV patrol data, including at least slope cracks, surface subsidence, water accumulation distribution, road blockage, and changes in the surrounding environment of buildings; ground monitoring data, including at least rainfall, wind speed, temperature and humidity, water level, soil moisture, and surface displacement data; and historical disaster record data, including at least risk zoning data and spatial distribution information of residential areas and buildings.

[0023] In practical implementation, multi-source data acquisition from the sky and ground, along with multi-source feature construction, includes: collecting natural disaster-related data through multi-source monitoring methods. Satellite remote sensing is used to acquire macro-environmental information at the regional scale, including data on surface temperature, vegetation cover, water distribution, and surface deformation; drone patrols are used to acquire local high-resolution observation data, such as slope cracks, surface subsidence, water accumulation, road blockages, and changes in the surrounding environment of buildings; ground monitoring equipment, including meteorological stations, hydrological stations, and geological monitoring equipment, is used to collect data on rainfall, wind speed, temperature and humidity, water level, soil moisture, and surface displacement. Simultaneously, regional scene information is constructed by combining historical disaster records, risk zoning data, and spatial distribution information of residential areas and buildings.

[0024] Step S20: Perform multi-source feature construction on the sky-ground multi-source data to generate multi-source feature input. The multi-source feature construction includes at least spatiotemporal alignment, missing data imputation, continuous feature standardization, and discrete feature fuzzification.

[0025] It should be noted that the multi-source feature construction of the sky-ground multi-source data includes: performing preliminary quality control and deviation correction on the sky-ground multi-source data, and marking or removing abnormal data; performing time series alignment on multi-source data with different sampling frequencies based on the least squares method to unify the data sources to the same time scale; performing spatial alignment on multi-source data with different spatial resolutions based on spatial grid mapping to map satellite remote sensing data, UAV patrol data, and ground monitoring data to a unified spatial grid cell; filling in missing data after alignment based on the Lagrange interpolation method; performing range standardization on continuous features and mapping them to a preset interval; and performing fuzzification on discrete features, using membership functions to calculate their membership degrees to different fuzzy sets, and generating a fuzzy discrete feature set.

[0026] In practice, satellite remote sensing data, UAV patrol data, and ground monitoring data have significant differences in terms of collection cycle, update time, coverage, spatial resolution, and positioning method. Without unified preprocessing, the same risk evolution process may have inconsistent corresponding times in different data sources, and the same geographical location may have inconsistent corresponding ranges in different data sources. This can lead to the inability to establish accurate correspondences between multi-source features, affecting the accuracy of subsequent risk assessment results.

[0027] Therefore, for the aforementioned multi-source data, data preprocessing is first completed using timestamp alignment and spatial grid mapping methods. In the time dimension, interpolation is used to unify data with different sampling frequencies to the same time scale; in the spatial dimension, geographic information coordinate mapping is used to map satellite, UAV, and ground monitoring data to unified regional grid cells, thereby achieving spatial alignment of multi-source data.

[0028] The multi-source data reception and preliminary correction include: The system receives multi-source data from satellite remote sensing, UAV patrols, and ground monitoring equipment, and categorizes and organizes it according to data type, acquisition time, and spatial location. Preliminary quality control and deviation correction are performed on potential outliers, drift values, out-of-range values, and duplicate values ​​in the multi-source data. Data that cannot be directly corrected is marked as anomalies or removed. These processes improve the reliability and consistency of the input data, providing a foundation for subsequent time series alignment and missing data imputation.

[0029] Time series alignment based on least squares includes: Even after initial correction, the multi-source data still suffers from time synchronization issues due to differences in sampling periods and communication delays between different devices. To achieve a unified representation of the multi-source data at a consistent time, the least squares method is used to perform local linear fitting on multiple observations from high-frequency sampling devices near the target time. Taking two devices, Device 1 and Device 2, as an example, let the sampling period of Device 1 be... The sampling period of device 2 is And satisfy At the target time of device 1 Nearby, device 2 has been acquired. Each observation value is denoted as... Assuming that the data changes near the target time can be approximated by a linear model, then we have: in, To measure the noise vector, For the parameter to be estimated, Indicates that device 2 is at the target time Alignment estimate, Indicates local trends of change. The matrix is ​​composed of the time offsets of each observation time relative to the target time, as follows: , For the sampling time of device 2, This is the sampling time of device 1, which is the target time to which it needs to be aligned.

[0030] The parameters can be obtained through least squares estimation. and will As device 2 at time The alignment result is obtained. Repeating the above process for all relevant devices yields a multi-source data sequence for time synchronization. In the case of multiple devices, first select a unified reference timeline, and then align all other devices to this timeline.

[0031] Spatial alignment of spatial grid mapping includes: In the spatial dimension, a unified grid is first constructed based on the geographical extent of the target area. The study area is divided into multiple spatial grid units according to a preset grid resolution, and a unique identifier is assigned to each grid unit. Satellite remote sensing data is mapped to the corresponding grid units according to the pixel center coordinates or coverage area. UAV patrol data is mapped to the corresponding grid units according to the image georegistration results, patrol target location, or extracted anomaly area boundaries. Ground monitoring data is matched to the grid units according to the sensor's latitude and longitude coordinates. When multiple sensors in the same grid unit observe multiple similar observations, the mean, maximum, weighted average, or most recent value is aggregated to form the unified spatial characteristics of the grid unit at the current moment, thereby realizing the corresponding expression of data from different sources and with different spatial resolutions in a unified spatial unit.

[0032] Missing data imputation based on Lagrange interpolation includes: After completing time and spatial alignment, for the missing continuous feature within the same spatial grid cell at the target time, several known observations before and after the target time are selected as interpolation nodes, and the feature value at the target time is estimated using the Lagrange interpolation method to complete the missing data filling based on the time points before and after.

[0033] For any data sequence containing missing values, at its missing points Select nearby Known data nodes , For the time of the i-th known data point, For the i-th known data point at time The measured values ​​are (i=0,1,2,…,m). First, construct the Lagrange interpolation basis functions. : Then, based on these basis functions and the values ​​of the known nodes... Construct the Lagrange interpolation polynomial : Location of missing points Substitution The fill value at that point can then be calculated. This method can quickly restore the continuity of a data sequence while maintaining accuracy.

[0034] Multi-source feature standardization and fuzzification representation, including: Following the aforementioned correction-alignment-filling process, a high-quality, spatiotemporally consistent multi-source data matrix was obtained. Based on this, two types of feature sets were constructed: one is a continuous feature set containing physical quantities such as rainfall, wind speed abrupt change indices, and normalized difference vegetation index (NDVI). Second, it is a discrete set of features including geological categories, risk zones, and historical disaster types. To eliminate dimensional differences and improve model convergence speed, continuous features... Each feature in Range normalization (Min-MaxNormalization) is used to map the range to the [0,1] interval. The calculation formula is as follows: In the formula, These are the original continuous feature observations. and These represent the minimum and maximum values ​​of the feature within the historical time window. These are the normalized eigenvalues.

[0035] For discrete features To address the loss of boundary information caused by traditional hard classification, fuzzy set theory is used to transform it into fuzzy linguistic variables. Taking wind speed levels as an example, the membership degree of wind speed levels belonging to different fuzzy sets (such as "low", "medium", and "high") is calculated using a triangular membership function. The calculation formula is as follows: In the formula, ... This represents the membership degree value, ranging from [0,1]. After the above processing, a normalized continuous feature set is finally generated. With fuzzy discrete feature set This provides standardized input for subsequent modeling. By using fuzzing, transitional risk semantic information can be preserved, avoiding the loss of boundary information caused by hard classification.

[0036] The above technical solution first performs preliminary quality control and bias correction on multi-source data from the sky and ground, effectively identifying and handling abnormal data and ensuring the reliability and accuracy of the original data. Next, time series alignment is performed based on least squares, and spatial alignment is performed based on spatial grid mapping, solving the heterogeneity problem of different data sources in time and space dimensions, enabling all data to be fused and analyzed within a unified spatiotemporal framework. Subsequently, missing data is filled using Lagrange interpolation, ensuring the integrity of the data input and avoiding evaluation bias caused by incomplete data. Furthermore, range standardization is applied to continuous features to eliminate the influence of different units and numerical ranges on model training, ensuring that each feature has equal importance in the model; fuzzification is applied to discrete features to better capture the uncertainty and ambiguity in the data, providing a more robust input for subsequent knowledge-driven channels. Overall, these refined multi-source feature construction steps significantly improve the quality, consistency, and applicability of multi-source feature input, providing a high-quality, standardized data foundation for subsequent disaster risk assessment, thereby effectively improving the accuracy and reliability of the final disaster risk warning.

[0037] Step S30: A parallel dual-channel architecture is used to perform disaster risk assessment on the multi-source feature input. The parallel dual-channel architecture includes a data-driven channel and a knowledge-driven channel. The data-driven channel outputs a data-driven risk score based on a deep learning model, and the knowledge-driven channel outputs a knowledge-driven risk score based on expert rule reasoning.

[0038] The data-driven channel outputs a data-driven risk score based on a deep learning model, which includes: constructing a Transformer time-series prediction model based on a multi-head attention mechanism; inputting a normalized continuous feature sequence into the Transformer time-series prediction model, capturing long-distance time dependencies in parallel through the multi-head attention mechanism, and outputting a data-driven risk score; using historical disaster records as supervision labels, and optimizing the parameters of the Transformer time-series prediction model through a loss function.

[0039] By introducing a Transformer-based time-series prediction model with a multi-head attention mechanism, this application effectively addresses the problem of traditional deep learning models struggling to efficiently capture long-distance temporal dependencies when processing complex natural disaster time-series data. The multi-head attention mechanism allows the model to focus on all time steps in the sequence in parallel across different representation subspaces, thus comprehensively capturing the complex correlations between near and far time points during disaster evolution and avoiding information forgetting. Simultaneously, using historical disaster records as supervisory labels for model parameter optimization ensures that the model can learn accurate risk patterns from real disaster data, improving the prediction accuracy and robustness of data-driven risk scores. This enables the data-driven channel to provide more accurate and reliable risk assessment results, laying a solid foundation for subsequent risk fusion and early warning.

[0040] The knowledge-driven channel outputs a knowledge-driven risk score based on expert rule reasoning, including: constructing a Mamdani-type fuzzy reasoning system based on pre-set expert rules, wherein the expert rules are in IF-THEN form; matching the fuzzy discrete feature set with the expert rules and calculating the trigger strength of each rule; using max-min composite reasoning, pruning and aggregating the conclusions of each rule according to the trigger strength to generate a comprehensive output fuzzy set; and using the centroid method to defuzzify the comprehensive output fuzzy set and output the knowledge-driven risk score.

[0041] The above technical solution effectively integrates the qualitative knowledge and experience of domain experts, matches the fuzzy discrete feature set with pre-set expert rules, and calculates the trigger strength of each rule, thereby overcoming the limitations of traditional hard rule systems in handling uncertainty and fuzzy information. Max-min composite reasoning is used to prune and aggregate rule conclusions, comprehensively reflecting the combined judgment of multiple expert rules and generating a more robust comprehensive output fuzzy set. Finally, defuzzification is performed using the centroid method, transforming fuzzy judgments into precise knowledge-driven risk scores, making risk assessment results smoother, more continuous, and consistent with expert intuition. This method not only improves the adaptability of the knowledge-driven channel to complex and uncertain disaster scenarios but also provides high-quality, expert-experience-based input for subsequent risk score fusion, significantly enhancing the accuracy and reliability of the entire early warning method.

[0042] In practice, a dual-channel parallel computing architecture is adopted, which uses deep learning to process time-varying data and expert rules to perform fuzzy inference.

[0043] The data-driven channels include: A Transformer time-series prediction model based on multi-head attention is constructed. This model can capture long-distance temporal dependencies using normalized continuous feature sequences from the most recent T time periods. Given the input, output the risk score of the i-th disaster-affected scenario at time k. The details are as follows.

[0044] Let the first Each disaster-bearing scenario unit is in time The recent The normalized continuous feature sequence for each time period is as follows: First, map the input sequence to a unified feature space and then overlay positional encoding: Then construct the query matrix, key matrix, and value matrix: In multi-head attention mechanisms, the first The output of each attention head is: All The results of each attention head are concatenated and linearly mapped to obtain the multi-head attention output: The temporal representation is then extracted via a feedforward network, and the hidden vector corresponding to the current time step is obtained. As a comprehensive temporal feature, the final output data disaster risk score is: in, Represents the input time-series feature matrix. Represents a real matrix with T rows and d columns. For the input mapping matrix, For position encoding, , , These are the query matrix, key matrix, and value matrix, respectively. , , is a trainable linear mapping matrix, used to map the input hidden representation to the query matrix, key matrix, and value matrix, respectively. It is a normalization function. The r-th attention head consists of the query matrix, the transpose of the key matrix, and the value matrix, respectively. The dimension of the key vector. This is a concatenation function used to join the outputs of multiple attention heads into a unified representation along the feature dimensions. This is the output mapping matrix after multi-head concatenation. This is the comprehensive temporal hiding representation for the current moment. For the Sigmoid function, and These are the output layer weight matrix and the bias term, respectively.

[0045] The calculation process is described as follows: The model first... Different types of physical quantities are projected into a unified mathematical space, enabling them to be compared and computed. Simultaneously, location codes are added to the data at each time step, allowing the model to explicitly know the position of each data point on the time axis. This step transforms the original, heterogeneous set of physical quantities into an internal representation that the model can understand, incorporating spatiotemporal information.

[0046] Next, temporal correlation mining is performed, which uses a multi-head attention mechanism to analyze the entire historical sequence in parallel and adaptively. The model sets up multiple independent "analysis heads", each working independently and evaluating the correlation importance between the current moment and every moment in history according to its own preferred analytical perspective.

[0047] After obtaining all the important historical correlations, the model aggregates this information, which is scattered across different time points and features, to form a comprehensive result, and finally outputs a disaster risk score. .

[0048] Finally, historical disaster records were used as monitoring labels to optimize the parameters of the data-driven early warning model. Let be the reference risk score label for the i-th disaster-affected scenario at time k, which is normalized from factors such as disaster level, impact range, degree of loss, and damage status of the affected structures. The loss function is expressed as: The loss function is minimized using the backpropagation algorithm, and the network parameters are iteratively updated. This allows the model to gradually learn the mapping relationship between multi-source temporal features and the true disaster risk score, thereby improving the disaster risk score. The accuracy of the prediction.

[0049] Knowledge-driven pathways include: Construct Mamdani-type fuzzy inference. This is achieved by pre-setting an experience-based "IF-THEN" fuzzy rule base and applying it to the fuzzified discrete feature set of the input. Perform rule matching and maximum-minimum composition reasoning.

[0050] Rule base: Stores a series of expert rules in IF-THEN format.

[0051] Rule matching: Fuzzy features of the current input It will match the premise (IF part) of all rules and calculate the trigger strength of each rule.

[0052] Max-Min Composite Reasoning: Each triggered rule has its conclusion (THEN part) defined as a fuzzy set "trimmed" based on its trigger strength. Higher trigger strength results in a larger range and higher membership degree of the retained conclusion fuzzy set after trimming. The union of all trimmed conclusion fuzzy sets forms a unified, comprehensive output fuzzy set. The shape of this set represents a summary of all triggered expert opinions.

[0053] Subsequently, the centroid method was used to defuzzify the fuzzy output surface to obtain an accurate semantic disaster risk score. The calculation formula is as follows: In the formula, z is a variable in the risk assessment domain, with a value range of [0,1]. This is the risk membership function, used to describe the degree to which the risk value z belongs to the fuzzy concept of high risk. This step realizes the quantitative mapping of qualitative expert experience.

[0054] Step S40: Based on deep learning-assisted Kalman filtering, adaptively fuse the data-driven risk score and the knowledge-driven risk score to output the final disaster risk score.

[0055] The adaptive fusion of the data-driven risk score and the knowledge-driven risk score based on deep learning-assisted Kalman filtering includes: constructing a random walk state equation to describe the temporal evolution characteristics of the disaster risk score; using the data-driven risk score and the knowledge-driven risk score as two observation inputs to construct an observation vector and an observation model; estimating the basic values ​​of the process noise covariance and the observation noise covariance based on basic rules and statistical methods; dynamically correcting the process noise covariance and the observation noise covariance based on a deep learning network to generate adaptive noise parameters; calculating the Kalman gain based on the adaptive noise parameters, and weighting and fusing the predicted and observed values ​​according to the Kalman gain to output the final disaster risk score.

[0056] Through the above technical solution, this embodiment introduces a deep learning-assisted Kalman filtering mechanism when fusing data-driven and knowledge-driven risk scores. Specifically, by constructing a random walk state equation and an observation model, the two risk scores are used as observation inputs, and the noise covariance is initially estimated by combining basic rules and statistical methods. More importantly, this embodiment uses a deep learning network to dynamically correct these noise covariances, enabling them to adaptively reflect the uncertainties of the current environment, data quality, and disaster evolution characteristics. This dynamically adjusted noise parameter allows the Kalman gain to be optimized in real time, resulting in a more accurate and robust weighted fusion between predicted and observed values. Compared to traditional methods of fixed or simple statistical estimation of noise parameters, this solution significantly improves the accuracy of the final disaster risk score and its adaptability to complex and variable disaster scenarios, effectively solving the limitations of traditional Kalman filtering in noise parameter estimation and ensuring the reliability of early warning information.

[0057] It should be noted that the dynamic correction of the process noise covariance and the observation noise covariance based on the deep learning network includes: constructing a network input feature vector, wherein the feature vector includes at least two risk scores, the variation magnitude of the two risk scores over time, an environmental change intensity index, and a data quality index; inputting the feature vector into a multilayer perceptron network, performing feature mapping through a nonlinear activation function, and outputting process noise correction factors and observation noise correction factors; applying a Softplus function to impose non-negative constraints on the process noise correction factors and the observation noise correction factors; and generating the adaptive noise parameters based on the base values ​​and the correction factors.

[0058] Through the above technical solution, this embodiment effectively solves the problem of inaccurate estimation and dynamic adjustment of noise parameters in traditional Kalman filtering. By constructing a rich feature vector containing two risk scores, their variation amplitudes, environmental change intensity indicators, and data quality indicators, and inputting it into a multilayer perceptron network for nonlinear mapping, the network can learn the complex dynamic relationship between noise parameters and the current system state and environmental factors. Furthermore, the Softplus function is used to impose non-negativity constraints on the output correction factors, ensuring the physical rationality of the noise parameters. Finally, based on the base values ​​and these adaptive correction factors, real-time and accurate adaptive noise parameters can be generated. This dynamic correction mechanism allows Kalman filtering to flexibly adjust the degree of trust in data-driven and knowledge-driven risk scores according to actual conditions, thereby significantly improving the accuracy and robustness of the fusion results. This enables the final disaster risk score to more accurately reflect the evolution of actual disaster risks, providing a more reliable decision-making basis for natural disaster risk early warning.

[0059] In practice, different disaster-bearing scenario units are all processed according to the same fusion process. The following section describes the process for any disaster-bearing scenario unit at any given time. Taking the integration process as an example, other disaster-bearing scenario units can be processed according to the same process.

[0060] A state recursive model is constructed by Kalman filtering. The output of the pattern recognition model and the output of fuzzy inference are used as multi-source observations. The Kalman gain is adaptively calculated based on the observation noise covariance to achieve online optimal fusion estimation of the final disaster risk score.

[0061] Model building, including: The state equation for the random walk model is: This formula means that the input value will not change drastically without any reason, but it is allowed to change gradually with changes in the environment. The allowable range of change was controlled. By introducing a random walk state equation, the temporal evolution of the state was modeled, achieving a unified description of temporal continuity and uncertainty changes without explicitly characterizing complex disaster mechanisms. This provides a reasonable prior model for Kalman filtering prediction and fusion.

[0062] Let k be the actual disaster risk score at time k. The noise is process noise and follows a mean of 0 and a covariance of . The multidimensional normal distribution. Let K be the process noise covariance, representing the degree of natural variation at each time step.

[0063] Initialization, including: During the initial run, initialization operations are performed on the state estimates, state estimate error covariance, and process noise variance, as follows: (1) Initialize the state estimate: in and These are the disaster risk scores output by the pattern recognition model and the fuzzy inference model at the initial time, respectively.

[0064] (2) Initial state estimation error covariance: in This is a preset constant used to represent the uncertainty of the initial stage state.

[0065] (3) Initialize the noise baseline: in and The preset constants are set based on empirical values ​​and adjusted according to actual data during operation, starting from these values. This step is performed only once at startup to provide initial conditions for subsequent Kalman recursive fusion.

[0066] Construct the observation equations, including: The data at time k drives the model output. With fuzzy inference output Treating it as two observations, construct the observation vector: in The actual observed value at time K. ， .

[0067] The observation model is: in H= For the observation matrix, The observation error (noise) refers to the fact that the disaster risk scores calculated by the data-driven model and the expert rules cannot be completely accurate. These inaccuracies are represented by noise, with an average value of 0, and the magnitude of the fluctuation is determined by the noise covariance matrix. The decision is as follows: and This is the noise variance between the two observations (a reliability parameter for disaster risk score observations); the larger the value, the lower the reliability. Furthermore, the errors of the two observations do not affect each other.

[0068] Basic noise estimation, including: To avoid deep learning networks directly outputting noise parameters from scratch, this method first constructs basic estimates of process noise and observation noise using rule-based and statistical methods.

[0069] (1) Basic process noise estimation in For the initial process noise, here is a constant representing the noise intensity under steady-state conditions. This is an adjustment coefficient, a constant here, representing the impact of environmental changes on noise intensity, set according to actual needs. This is an indicator of the intensity of environmental change at time k, representing the change in the environment over time. It is calculated based on real-time external environmental data (such as wind speed, rainfall, etc.), and the specific expression is as follows: in For the first The observed value of a physical quantity at time k. For the first The observed value of a physical quantity at time k-1. The first The maximum and minimum values ​​of a physical quantity are determined based on historical data or a set maximum and minimum range.

[0070] (2) Basic observation noise estimation For the two risk score observations, their initial observation noise is constructed respectively: in Quality metrics for data-driven channels can be represented by data missing rate, noise level, etc. The quality indicators for knowledge-driven channels can be composed of rule conflict degree, rule coverage, etc.

[0071] Adjustment coefficient , It is adjusted during operation based on environmental changes and actual data, and can be obtained through data-based learning algorithms.

[0072] In this process, the observation vector is used to characterize the specific output results obtained from different evaluation channels at the current moment, while the observation model is used to characterize the mapping relationship between the above output and the real state and its uncertainty, providing a basis for subsequent Kalman fusion.

[0073] Deep learning-based noise parameter correction includes: To further enhance the adaptive capability of fusion in complex scenarios, a deep learning parameter estimation network is introduced to dynamically correct the process noise covariance and observation noise covariance.

[0074] (1) Network input construction Construct the first Network input feature vector at time step: in: This indicates the magnitude of change in the risk scores of the two paths over time.

[0075] (2) Deep learning network structure A multilayer perceptron (MLP) architecture is adopted, including an input layer, two hidden layers, and an output layer. The hidden layers map the input features using a non-linear activation function, and the calculation process is as follows: First hidden layer: Second hidden layer: in , This is the weight matrix. , For bias vectors, It is a non-linear activation function; Network output noise correction factor: in, This represents a nonlinear mapping function composed of a multilayer sensing mechanism. These are the parameters of the process noise network. Their input is the feature vector. After feature extraction from the hidden layer, the output is a noise correction factor.

[0076] Observation noise correction factor: in and This represents a nonlinear mapping function composed of a multilayer sensing mechanism. To observe the parameters of the noisy network, its input is a feature vector. The noise correction factors for the first and second observations are output respectively.

[0077] To ensure that the noise parameter remains positive, the Softplus function is used for constraint: in To prevent numerical degradation, a minimal constant is used.

[0078] (3) Construct the final adaptive noise parameters The process noise covariance is updated as follows: The observed noise covariance is updated as follows: in and These are the basic noise estimates obtained based on rules and statistical methods; and This refers to the corrected noise parameters that actually participate in the Kalman recursion, as output by the deep learning network. This method enables the deep learning network to adaptively adjust the filtering parameters.

[0079] Perform Kalman fusion recursion, including: Based on the fusion results from the previous time step, the state at the current time step is predicted: in Indicates the time at which the result is obtained. Before the observed value, a predicted value for the state at time k is made based on all information up to time k-1 (observation data). This represents the final estimated state at the previous time step. F is the state transition matrix; since a random walk model is used, F=1 is chosen here.

[0080] in To account for the uncertainty in the prediction at time k, This represents the uncertainty of the final estimate at the previous moment.

[0081] Update parameters: This formula is the Kalman gain formula, used to determine the weights of observed and predicted values. The larger the observed value (actual value), the higher its weight.

[0082] This formula is a state update formula. Based on the current observation value and the predicted value of the state at time k based on all information at time k-1, the final estimated value at time k is calculated, and the predicted value is corrected.

[0083] This formula is the covariance update formula, which corrects the uncertainty of the prediction based on the amount of information provided by the observations, resulting in the final uncertainty. Each correction reduces the uncertainty, meaning the observed data becomes increasingly reliable.

[0084] The final disaster risk score is output, including: Since the Kalman filter outputs a continuous state estimate, the final output is guaranteed to be in the [0,1] interval by the following formula: .

[0085] It should be noted that the deep learning-assisted Kalman filtering also includes closed-loop optimization: comparing the final disaster risk score with the real disaster risk score label to construct a fusion loss function; updating the parameters of the deep learning network through the backpropagation algorithm, so that the deep learning network learns to adaptively adjust the noise parameters; continuously receiving actual disaster feedback information at subsequent times, and performing incremental updates or periodic offline retraining based on the error between the real label and the fusion result to form a closed-loop adaptive mechanism.

[0086] In practice, during the training phase, the noise correction factor output by the deep learning network is substituted into the Kalman filter recursion process to obtain the fusion risk score. and compare it with the actual disaster risk score label. The loss function is constructed together for training as follows: The aforementioned loss function measures the deviation between the fusion result and the actual disaster state. By minimizing this loss, the deep learning parameter estimation network progressively learns to adaptively adjust the process noise covariance and observation noise covariance under different scenarios, thereby improving the Kalman fusion's ability to dynamically weight the two risk scores.

[0087] During training, the parameters of the deep learning network are adjusted using the backpropagation algorithm. and Update the process noise estimation network. and observation noise estimation network Based on the current dual-channel risk score, historical risk score trends, and data quality indicators, the network can output a noise correction factor that better reflects the actual evolution of disasters. After iterative training, the network can achieve online adaptive adjustment of filtering parameters in complex time-varying environments.

[0088] During the online operation phase, the trained deep learning network no longer outputs a disaster risk score independently. Instead, it estimates noise correction parameters in real time based on the input features at the current moment and feeds them into the Kalman filter recursive formula to output the final fused disaster risk score. Simultaneously, it continuously receives actual disaster feedback information from subsequent moments and uses the error between the true label and the fused result for subsequent incremental updates or periodic offline retraining, thus forming a closed-loop adaptive mechanism of "risk assessment—parameter correction—fusion output—feedback optimization".

[0089] Through the above technical solution, this embodiment constructs a closed-loop adaptive mechanism, enabling the deep learning network to continuously learn and optimize its adaptive adjustment capability to Kalman filter noise parameters. By comparing the final disaster risk score with the actual disaster risk score label and constructing a fusion loss function, the system can quantify the prediction error. Subsequently, the backpropagation algorithm is used to update the parameters of the deep learning network, allowing the network to continuously correct its internal parameters based on actual feedback, thereby more accurately and dynamically adjusting the process noise covariance and observation noise covariance in the Kalman filter. Furthermore, by continuously receiving actual disaster feedback information at subsequent times and performing incremental updates or periodic offline retraining based on the error between the actual label and the fusion result, the system is ensured to adapt to constantly changing disaster patterns, environmental conditions, and data characteristics. This closed-loop optimization mechanism significantly improves the accuracy, robustness, and timeliness of disaster risk early warning, avoiding the problem of model performance deterioration over time, and enabling the final disaster risk score to more accurately reflect the actual risk situation, thus providing a more reliable decision-making basis for natural disaster risk early warning.

[0090] Step S50: Generate natural disaster risk early warning information based on the final disaster risk score.

[0091] The step of generating natural disaster risk early warning information based on the final disaster risk score includes: ranking the disaster-bearing scenario units according to the final disaster risk score; determining the early warning level of each disaster-bearing scenario unit according to a preset risk threshold range; and generating differentiated disaster prevention and mitigation emergency decision-making suggestions based on the early warning level and the attribute information of the disaster-bearing scenario unit.

[0092] When generating differentiated disaster prevention and mitigation emergency decision-making recommendations, the warning level and the attribute information of the disaster-bearing scenario unit are comprehensively utilized. The attribute information of the disaster-bearing scenario unit may include, but is not limited to, population density, building type, distribution of critical infrastructure, land use type, historical disaster types, local emergency response capabilities, topography, and vegetation cover. This attribute information provides specific background and constraints for the decision-making recommendations. For example, for a densely populated residential area at a "red warning" level with attribute information indicating high population density, the system may recommend immediately activating evacuation plans, designating evacuation routes, opening refuge areas, and issuing emergency evacuation notices to residents. For an area also at a "red warning" level but with attribute information indicating important industrial facilities, recommendations may include immediately halting production, cutting off hazardous sources, strengthening equipment inspections, and activating internal emergency plans. This differentiated recommendation is typically achieved through a pre-set rule base or expert system, which matches and generates the most relevant and effective disaster prevention and mitigation measures based on the warning level and the specific attributes of the disaster-bearing scenario unit.

[0093] Through the above technical solution, this embodiment further refines the process of generating early warning information based on the final disaster risk score. First, by ranking the disaster-affected scenario units by risk, the highest-risk areas or objects can be clearly identified, providing a clear priority basis for resource allocation and emergency response. Second, by determining the early warning level of each disaster-affected scenario unit through preset risk threshold ranges, the continuous risk scores are transformed into easily understandable and communicable discrete levels, greatly simplifying the interpretation and dissemination of early warning information. More importantly, by combining the early warning level with the attribute information of the disaster-affected scenario unit, differentiated disaster prevention and mitigation emergency decision-making suggestions can be generated. This means that early warning information is no longer general, but provides specific and actionable guidance for specific areas, specific groups, or specific facilities. For example, evacuation routes may be suggested for densely populated areas, and equipment shutdown measures may be suggested for industrial areas. This refined and personalized early warning information significantly improves the practicality and guidance of early warnings, making disaster prevention and mitigation work more targeted and efficient, thereby effectively reducing disaster losses and protecting people's lives and property.

[0094] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0095] In addition, for technical details not described in detail in this embodiment, please refer to the method for natural disaster risk early warning by fusion of multi-source data from the sky and ground provided in any embodiment of this application, which will not be repeated here.

[0096] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0097] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for early warning of natural disaster risks through multi-source data fusion from the sky and ground, characterized in that, include: Acquire multi-source data from the sky and ground, which includes at least satellite remote sensing data, UAV patrol data, ground monitoring data, and historical disaster records; Multi-source feature construction is performed on the aforementioned multi-source data from the sky and ground to generate multi-source feature input. The multi-source feature construction includes at least spatiotemporal alignment, missing data imputation, continuous feature standardization, and discrete feature fuzzification. A parallel dual-channel architecture is adopted to perform disaster risk assessment on the multi-source feature input. The parallel dual-channel architecture includes a data-driven channel and a knowledge-driven channel. The data-driven channel outputs a data-driven risk score based on a deep learning model. The knowledge-driven channel outputs a knowledge-driven risk score based on expert rule reasoning. Based on deep learning-assisted Kalman filtering, the data-driven risk score and the knowledge-driven risk score are adaptively fused to output the final disaster risk score. Based on the final disaster risk score, natural disaster risk early warning information is generated.

2. The method according to claim 1, characterized in that, The construction of multi-source features from the sky-ground multi-source data includes: Preliminary quality control and deviation correction are performed on the multi-source data from the sky and ground, and abnormal data is marked or removed. Time series alignment of multi-source data with different sampling frequencies is performed based on the least squares method, unifying the data sources to the same time scale. Spatial alignment of multi-source data with different spatial resolutions is performed based on spatial grid mapping, mapping satellite remote sensing data, UAV patrol data and ground monitoring data to a unified spatial grid cell; The missing data after alignment is filled using Lagrange interpolation. Continuous features are normalized by range and mapped to a preset interval; The discrete features are fuzzified, and the membership degree of each feature to different fuzzy sets is calculated using membership functions to generate a fuzzified discrete feature set.

3. The method according to claim 1, characterized in that, The data-driven channel outputs a data-driven risk score based on a deep learning model, including: Construct a Transformer time series prediction model based on a multi-head attention mechanism; The normalized continuous feature sequence is input into the Transformer time series prediction model, and long-distance temporal dependencies are captured in parallel through a multi-head attention mechanism to output a data-driven risk score. Historical disaster records are used as supervisory labels, and the parameters of the Transformer time series prediction model are optimized through a loss function.

4. The method according to claim 1, characterized in that, The knowledge-driven channel outputs a knowledge-driven risk score based on expert rule reasoning, including: Construct a Mamdani-type fuzzy inference system based on pre-defined expert rules, wherein the expert rules are in IF-THEN form; The fuzzy discrete feature set is matched with the expert rules, and the trigger strength of each rule is calculated. Max-min composite reasoning is used to prune and aggregate the conclusions of each rule according to the triggering intensity, generating a comprehensive output fuzzy set. The centroid method is used to defuzzify the comprehensive output fuzzy set and output a knowledge-driven risk score.

5. The method according to claim 1, characterized in that, The adaptive fusion of the data-driven risk score and the knowledge-driven risk score based on deep learning-assisted Kalman filtering includes: Construct a random walk state equation to describe the temporal evolution characteristics of disaster risk scores; The data-driven risk score and the knowledge-driven risk score are used as two observation inputs to construct an observation vector and an observation model. Based on fundamental rules and statistical methods, the basic values ​​of process noise covariance and observation noise covariance are estimated. Based on a deep learning network, the process noise covariance and the observation noise covariance are dynamically corrected to generate adaptive noise parameters. Based on the adaptive noise parameters, the Kalman gain is calculated, and the predicted and observed values ​​are weighted and fused according to the Kalman gain to output the final disaster risk score.

6. The method according to claim 5, characterized in that, The dynamic correction of the process noise covariance and the observation noise covariance based on a deep learning network includes: Construct a network input feature vector, wherein the feature vector includes at least two risk scores, the change magnitude of the two risk scores over time, an environmental change intensity index, and a data quality index; The feature vector is input into a multilayer perceptron network, and feature mapping is performed through a nonlinear activation function to output process noise correction factor and observation noise correction factor. The Softplus function is used to apply non-negative constraints to the process noise correction factor and the observation noise correction factor. The adaptive noise parameters are generated based on the base value and the correction factor.

7. The method according to claim 1, characterized in that, The deep learning-assisted Kalman filter also includes closed-loop optimization: The final disaster risk score is compared with the actual disaster risk score label to construct a fusion loss function; The parameters of the deep learning network are updated by using the backpropagation algorithm, enabling the deep learning network to learn to adaptively adjust noise parameters. It continuously receives actual disaster feedback information from subsequent moments, and performs incremental updates or periodic offline retraining based on the error between the real labels and the fusion results, forming a closed-loop adaptive mechanism.

8. The method according to claim 1, characterized in that, The process of generating natural disaster risk early warning information based on the final disaster risk score includes: Based on the final disaster risk score, the disaster-bearing scenario units are ranked by risk. Based on the preset risk threshold range, determine the early warning level for each disaster-affected scenario unit; Based on the warning level and the attribute information of the disaster-bearing scenario unit, differentiated disaster prevention and mitigation emergency decision-making suggestions are generated.

9. The method according to claim 1, characterized in that, The multi-source data from the sky and ground includes: The satellite remote sensing data includes at least surface temperature, vegetation cover, water distribution, and surface deformation data. The drone patrol data includes at least data on slope cracks, surface subsidence, water accumulation, road blockages, and changes in the surrounding environment of buildings. The ground monitoring data includes at least rainfall, wind speed, temperature and humidity, water level, soil moisture, and surface displacement data. The historical disaster record data includes at least risk zoning data and information on the spatial distribution of residential areas and buildings.

Images