A Method for Risk Assessment of Secondary Emergencies Based on Multi-Source Data Fusion
By constructing a two-layer grid map that integrates multi-source data, the problem of real disaster expansion and data error differentiation in the risk assessment of secondary sudden environmental events was solved, the accuracy and uncertainty of the risk index were quantified, and the reliability of the risk assessment was improved.
Patent Information
- Application Number
- CN202611497247.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively distinguish between actual disaster expansion and data errors in risk assessment of secondary environmental emergencies, leading to risk index deviations and distortions of regional boundaries, and failing to quantify the uncertainty of risk outcomes.
By constructing a two-layer grid graph based on multi-source data fusion, and combining spatial adjacency relationships and disaster mechanism correlations, cross-source consistency residuals, spatial continuity residuals, and mechanism constraint residuals are calculated to correct the data source credibility. Constraint inferences for missing or delayed data are also performed, and a comprehensive risk index and risk credibility interval are output.
It effectively suppresses the spread of errors caused by low-resolution, delayed update, or local anomaly data, ensures that continuous risk changes in the actual disaster expansion are preserved, reduces boundary shifts in high-risk areas and anomalies in local risk values, and provides a credible range and uncertainty range for risk results.
Smart Images

Figure CN123117833A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer data processing and risk assessment technology, specifically a method for risk assessment of secondary sudden environmental events based on multi-source data fusion. Background Technology
[0002] Risk assessment of secondary environmental emergencies typically uses the target area as the basic object, collecting data on meteorological disasters, topography, land use, population and environmentally sensitive targets, historical disaster records, and real-time environmental monitoring. A unified dataset is then formed through coordinate transformation, time alignment, and indicator normalization. Subsequently, indicators such as disaster hazard, environmental carrying capacity, disaster-prone conditions, and exposure levels are calculated according to administrative divisions or regular grids. Existing conventional methods often pre-set fixed weights for various indicators, or determine weights based on data completeness, expert scoring, and historical statistical results, and then obtain a regional risk index through linear weighting, hierarchical analysis, or comprehensive scoring. For data with inconsistent spatial resolution, resampling, neighborhood interpolation, or regional averaging are typically used to complete grid mapping; for data with inconsistent update frequencies, the most recent observation value is usually retained, or the mean is calculated according to a time window. Once the risk index reaches a preset level, the system outputs a high-risk grid and its spatial distribution to form risk zoning results.
[0003] Some existing schemes introduce time-series forecasting or spatial correlation processing on top of fixed weighting, using historical disaster processes and real-time monitoring sequences to determine the stage of a disaster, and supplementing missing data or smoothing anomalous results based on the risk status of adjacent grids. These schemes typically describe risk propagation using a single spatial adjacency relationship, treating geographically adjacent or closely spaced grids as correlation objects, and then updating grid risk values using neighborhood weighting, graph propagation, or iterative correction methods. For inconsistencies between meteorological, geographical, socioeconomic, and real-time monitoring data, they typically remove observations with large deviations by unifying anomaly conditions, or adjust the weight of a particular data source across the entire region based on overall error. Since the reliability of data sources is usually determined only by source category or global statistical results, the same reliability is applied to the same data source during the disaster incubation, triggering, expansion, and attenuation stages. Spatial continuity processing also struggles to distinguish between risk gradients formed by the expansion of actual disasters and anomalous gradients formed by observation errors.
[0004] The reliability of multi-source heterogeneous data exhibits significant spatial and disaster stage differences. Fixed weights, global weight correction, and ordinary neighborhood smoothing cannot establish a correspondence between data error sources and specific grid risk deviations. When low-resolution data, delayed update data, or locally anomalous data are mapped to spatial grids, their deviations may spread to adjacent grids through neighborhood interpolation and spatial propagation. Simultaneously, real disaster processes such as rainstorm confluence, slope erosion, and flood expansion themselves create continuous spatial variations. Existing solutions lack a processing mechanism that simultaneously utilizes cross-source consistency, spatial continuity, and disaster mechanism constraints to identify deviation sources and reverse-correct local reliability. Therefore, it is difficult to determine whether grid risk changes are due to real risk propagation or data error propagation, easily leading to shifts in the comprehensive risk index, distortion of high-risk area boundaries, and an inability to quantify the uncertainty of risk outcomes. Summary of the Invention
[0005] The purpose of this invention is to provide a method for risk assessment of secondary sudden environmental events based on multi-source data fusion, which can solve the problems in the background art mentioned above.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for risk assessment of secondary environmental emergencies based on multi-source data fusion includes: collecting meteorological disaster data, geographic environment data, socio-economic data, historical disaster data and real-time monitoring data of the target area, mapping them to a pre-divided spatial grid according to a unified spatiotemporal identifier, and forming a grid risk evidence set; Data quality description vectors are generated based on the observation time, spatial coverage, update delay, missing status, and historical deviation of each data point, and the reliability of the data source is determined in conjunction with the current disaster evolution stage of the grid. A two-layer grid diagram containing spatial adjacency relationships and disaster mechanism correlations is constructed. The grid risk evidence is initially fused, and cross-source consistency residuals, spatial continuity residuals, and mechanism constraint residuals are calculated. Based on the data sources and propagation paths corresponding to various residuals, the data source credibility in the corresponding grid is corrected, and the constraint inference of missing or delayed data is completed based on the corrected data source credibility. The comprehensive risk index and risk credibility interval of each spatial grid are output.
[0007] Preferably, the formation of the grid risk evidence set includes: extracting disaster type, intensity and movement trend from meteorological disaster data, elevation, slope, confluence path and land cover status from geographic environmental data, population exposure and distribution of environmentally sensitive targets from socioeconomic data, location, duration and impact range from historical disaster data, and environmental variable change sequence from real-time monitoring data. The coverage relationship with the spatial grid is determined according to the spatial resolution of the data. A coverage ratio index is established for data that spans multiple spatial grids. A time window index is established for data with different time scales within the same spatial grid. The data value, coverage ratio, time window position and data source identifier are encapsulated into risk evidence records in the grid risk evidence set.
[0008] Preferably, the determination of the disaster evolution stage includes: extracting the meteorological intensity change rate, the surface environmental variable change rate, the historical similar disaster trajectory matching degree, and the risk migration direction of adjacent spatial grids from a continuous time window to construct a grid stage state sequence; The grid stage state sequence is aligned with the preset incubation, triggering, expansion and decay stage state templates to obtain the matching probability of each stage; The matching probability is corrected based on the stage transition constraints of adjacent time windows, and the stage with the highest corrected probability is selected as the current disaster evolution stage. The data source credibility is generated based on the historical residual distribution, real-time completeness, and spatial coverage consistency of the data source at the corresponding stage, and organized into a dynamic credibility field according to the spatial grid and the disaster evolution stage.
[0009] Preferably, the construction of the double-layer grid map includes: establishing spatial adjacency edges based on the adjacency relationship between spatial grids, topographic elevation difference, confluence direction and disaster movement direction; and establishing disaster mechanism edges based on the constraint relationship between disaster risk, environmental carrying capacity, disaster-prone conditions and exposure degree. Connect risk evidence nodes within the same spatial grid and continuous time windows using time-related edges, and record the data source and time window identifier for each type of edge; The cross-source consistency residual is obtained from the difference in multi-source estimates of the same risk variable; the spatial continuity residual is obtained from the discontinuous changes in the risk states at both ends of the spatial adjacent edge; and the mechanism constraint residual is obtained from the deviation between the variables connected by the disaster mechanism edge.
[0010] Preferably, the establishment of the time window index includes: forming a reference time axis based on the update time of real-time monitoring data, projecting the meteorological disaster data and the historical disaster data onto the reference time axis according to the event duration interval, and generating segmented retention intervals for the geographical environment data and the socio-economic data according to the validity period; For data with multiple observations between two adjacent update times, retain the observation records corresponding to the inflection points and extreme values of change, and represent them as a piecewise change sequence; For data whose update time lags behind the reference time axis, record the delay length and the previous valid observation position; A traceable time window index is generated based on the event duration interval, segmented retention interval, segmented change sequence, and delay length.
[0011] Preferably, the generation of the dynamic credibility field includes: establishing a credibility state table for each data source, which is jointly indexed by the disaster evolution stage and the spatial grid, and writing the residual statistics of the same stage, similar geographical conditions and similar observation coverage in the historical disaster trajectory into the corresponding credibility state; When the current spatial grid lacks sufficient historical samples, reference grids with the same risk propagation direction are selected according to spatial adjacency edges and disaster mechanism edges, and migration coefficients are generated based on the differences in stage matching probability, geographical conditions, and data coverage between the reference grids and the current spatial grid. The reliability status of the reference grid is weighted and migrated using the migration coefficient, and the data source reliability of the current spatial grid is updated together with the data integrity of the current time window.
[0012] Preferably, the correction of the data source credibility includes: establishing a residual attribution path based on the risk evidence node, edge type and time window identifier associated with the residual; allocating the cross-source consistent residual to each data source that provides the same risk variable; allocating the spatial continuity residual in reverse along the spatial adjacency edge to the upstream risk evidence node; and allocating the mechanism constraint residual along the disaster mechanism edge to the risk evidence node that violates the variable constraint. For each attribution residual obtained from the same data source within the same spatial grid, a directional consistency judgment is made. When multiple attribution residuals point to the same deviation direction, a joint correction is formed. When the attribution residuals are in opposite directions, the retained items are determined according to the temporal relationship of the edges and the direction of risk propagation, and the corresponding data source credibility is updated accordingly.
[0013] Preferably, the constraint inference of the missing or delayed data includes: determining the target spatial grid and target time window corresponding to the risk variable to be inferred, and extracting candidate risk evidence nodes from the two-layer grid diagram that have spatial adjacency edges, disaster mechanism edges or time correlation edges with the target spatial grid; A candidate constraint set is formed based on the data source credibility of the candidate risk evidence nodes, their consistency with the risk propagation direction of the target spatial grid, and their time distance. Multiple candidate inference values are generated using the candidate constraint set. The cross-source consistency residual, spatial continuity residual, and mechanism constraint residual after the introduction of each candidate inference value are calculated respectively. The candidate inference value with the smallest comprehensive residual that does not change the order of the confirmed disaster evolution stages is selected and written into the target time window. At the same time, the inference source path is recorded.
[0014] Preferably, obtaining the risk confidence interval includes: extracting the observation time interval of each risk evidence record based on the time window index, and extracting the remaining attribution residuals of each data source based on the residual attribution path; The observation time interval and the remaining attribution residual are jointly grouped according to the data source, spatial grid and disaster evolution stage to generate the uncertainty boundary of the corresponding risk evidence record; The uncertainty boundary is propagated along the spatial adjacency edge, disaster mechanism edge and temporal correlation edge in the two-layer grid diagram, and correlation merging is performed between propagation paths with common sources of risk evidence to obtain the joint uncertainty range of each risk variable; The upper and lower bounds of the comprehensive risk index are calculated based on the joint uncertainty range to form the risk confidence interval.
[0015] Preferably, the output of the comprehensive risk index includes: associating the corrected data source credibility with the inference source path of the candidate inference value, and forming risk evidence subsets according to actual observation evidence and constrained inference evidence respectively; For each subset of risk evidence, the grid characterization values of disaster risk, environmental carrying capacity, and disaster-prone conditions are calculated, and the credibility of each grid characterization value is weighted according to the dynamic credibility field. Compare the grid representation values corresponding to actual observation evidence with the grid representation values corresponding to constraint inference evidence, and form an inference occupancy coefficient based on the difference between the two. The contribution ratio of the constraint inference evidence to the comprehensive risk index is limited by the inference occupancy coefficient, and the comprehensive risk index, risk confidence interval, risk evidence composition and corresponding disaster evolution stage are written into the spatial grid risk result set.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By mapping meteorological disaster data, geographic environment data, socioeconomic data, historical disaster data, and real-time monitoring data to a spatial grid using a unified spatiotemporal identifier, and generating data quality descriptions based on observation time, spatial coverage, update delay, missing status, and historical bias, the reliability of the data source can be correlated with the specific grid and the current stage of disaster evolution. The dual-layer grid map simultaneously records spatial adjacency relationships and disaster mechanism correlations. The cross-source consistency residuals, spatial continuity residuals, and mechanism constraint residuals generated from the initial fusion results can be attributed along the corresponding data sources and propagation paths. Reliability correction is thus limited to the grids, time windows, and data sources where biases occurred. This processing avoids using uniform weighting across the entire region or ordinary spatial smoothing, preserving continuous risk changes resulting from the actual disaster expansion in the comprehensive risk index, and suppressing error diffusion caused by low-resolution, delayed update, or local anomaly data, thereby reducing boundary shifts in high-risk areas and amplification of local risk values.
[0017] 2. By establishing coverage ratio indexes, time window indexes, and data source identifiers, data with different spatial resolutions, update cycles, and validity periods can retain their original coverage relationships and temporal locations, ensuring a continuous data traceability path for risk evidence during fusion, correction, and inference. For missing or delayed data, candidate evidence is extracted using spatial adjacency edges, disaster mechanism edges, and temporal correlation edges. Inferred values are jointly screened using cross-source consistency residuals, spatial continuity residuals, and mechanism constraint residuals, avoiding excessive smoothing of risk states caused by directly using neighborhood means. A joint uncertainty range is formed based on the observation time interval, residual attribution residuals, and correlation propagation paths, providing a corresponding risk confidence interval for the comprehensive risk index of each spatial grid. Simultaneously, the contribution ratio of inferred evidence is constrained by the inference occupancy coefficient, ensuring that actual observation evidence and inferred evidence remain distinct in risk results, facilitating the use of a consistent data foundation for subsequent risk level determination and time series updates. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the overall process of a risk assessment method for secondary environmental emergencies based on multi-source data fusion. Figure 2 A unified diagram of the three types of residuals; Figure 3 Select a plot for candidate inferred values; Figure 4 A flowchart for identifying disaster evolution stages and constructing a dynamic credibility field; Figure 5 This is a risk confidence interval diagram; Figure 6 This is a graph showing the output of the comprehensive risk index. Detailed Implementation
[0019] In one embodiment, please refer to Figure 1 This paper presents a method for risk assessment of secondary environmental emergencies based on multi-source data fusion, including: collecting meteorological disaster data, geographical environment data, socio-economic data, historical disaster data, and real-time monitoring data of the target area, mapping them to pre-divided spatial grids according to a unified spatiotemporal identifier to form a grid risk evidence set; generating data quality description vectors based on the observation time, spatial coverage, update delay, missing status, and historical deviation of each data, and determining the data source credibility in conjunction with the current disaster evolution stage of the grid; constructing a two-layer grid diagram containing spatial adjacency relationships and disaster mechanism correlation relationships, initially fusing the grid risk evidence, and calculating cross-source consistency residuals, spatial continuity residuals, and mechanism constraint residuals; correcting the data source credibility in the corresponding grids according to the data sources and propagation paths corresponding to various residuals, and completing constraint inference of missing or delayed data based on the corrected data source credibility, outputting the comprehensive risk index and risk credibility interval of each spatial grid.
[0020] In this embodiment, the target area is determined according to the spatial boundary of the risk assessment task. The spatial boundary is represented by a unified coordinate reference. The spatial grid adopts a regular grid or an irregular grid formed according to the terrain zoning. The grid scale is determined by the minimum spatial resolution in the input data that can maintain a stable coverage relationship. The meteorological disaster data includes at least rainfall intensity, cumulative rainfall, wind field changes, air pressure changes and disaster movement paths. The geographic environment data includes at least elevation, slope, aspect, confluence direction, land cover and surface water system. The socio-economic data includes at least population exposure, production activity density and distribution of environmentally sensitive targets. The historical disaster data includes at least disaster type, occurrence location, start and end time, evolution process and impact range. The real-time monitoring data includes at least a continuous observation sequence that can reflect the state of water bodies, soil, slope and surface environment. All input data retain the source identifier, generation time, collection time, validity period and original spatial range.
[0021] Specifically, the unified spatiotemporal identifier consists of regional coding, grid coding, time window coding, data source coding, and variable coding. Regional coding is used to distinguish the spatial boundaries corresponding to the assessment task. Grid coding is used to locate the spatial position of data within the target area. Time window coding is used to mark the time interval in which data participates in risk fusion. Data source coding is used to trace meteorological disaster data, geographic environment data, socioeconomic data, historical disaster data, or real-time monitoring data. Variable coding is used to distinguish risk variables such as rainfall, slope, runoff, exposure, environmental status, and historical impacts. For data with isal coverage, the overlap range between the data and each spatial grid is calculated. For data with point locations, the data is written into the corresponding spatial grid according to the location. For data with linear propagation paths, the data is written into the spatial grid along the path according to the intersection of the path and the grid boundary. The coverage ratio and boundary truncation information generated by spatial mapping are saved simultaneously.
[0022] Spatial mapping weights are calculated according to the following formula: ; in, This represents the evidence weights that map the i-th original data point to the g-th spatial grid and the t-th time window. This represents the overlap area or equivalent overlap length between the original data spatial range and the g-th spatial grid. Represents the area or equivalent length of the original data space. This represents the overlap duration between the valid time interval of the original data and the t-th time window. This indicates the total duration of the valid time interval for the original data. This represents the joint determination value of spatial location and temporal validity. It is set to 1 if the coverage is satisfied and the data is within the validity period, and 0 otherwise. For example, if the area of a certain isometric rainfall data covers 0.6 of the target grid area and the validity period covers 0.5 of the current time window, the mapping weight is 0.3.
[0023] A data quality description vector is generated for each piece of risk evidence entering the spatial grid. The vector includes observation timeliness, spatial coverage, update delay, missing data marker, historical bias, source stability, and stage applicability. Observation timeliness is obtained by normalizing the distance between the acquisition time and the end of the current time window. Spatial coverage is obtained from the spatial portion of the spatial mapping weights. Update delay is obtained from the time difference between the data generation time and the available time. The missing data marker is used to distinguish between actual observations, partial missing data, complete missing data, and constrained inferences. Historical bias is extracted from the historical residual statistics of the same data source under similar geographical conditions and similar disaster stages. Source stability is determined based on data value jumps, timestamp continuity, and spatial coverage changes within a continuous time window. Stage applicability is used to represent the interpretability range of the current data source for the current disaster evolution stage.
[0024] The current disaster evolution stage of the grid is determined by the multivariate state changes within a continuous time window. Each spatial grid forms a stage state sequence composed of changes in meteorological intensity, changes in surface environmental variables, risk migration direction of adjacent grids, and similarity of historical trajectories. The stage state sequence is aligned with the templates for the incubation, triggering, expansion, and decay stages. In the template, the incubation stage corresponds to the continuous accumulation of external disaster-causing factors while the environmental state has not yet undergone a leap. The triggering stage corresponds to a sudden change in at least one key environmental variable in the same direction. The expansion stage corresponds to the continuous propagation of the risk state along the spatially adjacent direction. The decay stage corresponds to the weakening of external disaster-causing factors and the continuous decline of risk variables. The alignment process preserves the sequential constraints between stages to avoid a single abnormal observation directly causing a stage jump.
[0025] Data source reliability is determined separately for spatial grids and disaster evolution stages, and cannot be replaced by a fixed value for the entire region from the same data source. For spatial grids in the incubation stage, data that reflects the accumulation of external disaster-causing factors has high stage applicability. For spatial grids in the triggering stage, real-time monitoring data that reflects environmental state transitions has high stage applicability. For spatial grids in the expansion stage, data with complete spatial coverage and that can indicate the direction of propagation has high stage applicability. For spatial grids in the decay stage, data with continuous and stable characteristics has high stage applicability. The initial value of reliability is calculated by combining the data quality description vector, historical residual statistics, and stage matching probability, and is locally corrected in the subsequent residual attribution process.
[0026] The two-layer grid diagram consists of risk evidence nodes, spatial adjacency edges, disaster mechanism edges, and temporal correlation edges. Risk evidence nodes correspond to specific grids, specific time windows, specific data sources, and specific risk variables. Spatial adjacency edges connect spatial grids with adjacency relationships or clear propagation reachability relationships. Disaster mechanism edges connect risk variables with constraints between disaster hazard, environmental carrying capacity, disaster-inducing conditions, and exposure levels. Temporal correlation edges connect similar risk variables within adjacent time windows in the same spatial grid. Each edge records the direction, source, basis for establishment, effective time window, and correlation strength. Spatial adjacency edges are not based solely on distance, but rather combine elevation difference, confluence direction, slope consistency, and disaster movement direction to determine whether the risk state is allowed to propagate along the corresponding direction.
[0027] When initially fusing grid risk evidence, evidence records within the same grid and time window are first grouped according to risk variables. Then, initial estimates of each risk variable are calculated based on initial credibility and spatial mapping weights. Subsequently, intermediate representations of disaster risk, environmental carrying capacity, and disaster-prone conditions are formed based on the preset semantic mapping between meteorological disaster variables, geographical environment variables, socioeconomic variables, historical disaster variables, and real-time monitoring variables. The initial fusion retains the evidence source set corresponding to each intermediate representation and does not delete the source identifier after fusion, so as to locate the data source, spatial grid, and time window that caused the bias through residual location later.
[0028] The three types of residuals are calculated uniformly according to the following formula: ; in, This represents the combined residual of the g-th spatial grid at the t-th time window. This represents the dispersion of multiple sources of estimates for the same risk variable. This indicates that data source d represents the risk variable. The estimated value, Represents the set of spatially adjacent edges. This represents the direction conversion coefficient between adjacent grid cells. Represents the set of disaster mechanism edges. This represents the mechanistic deviation function between risk variable a and risk variable b. , and These represent the combination coefficients of cross-source, spatial, and mechanistic residuals, respectively.
[0029] Please refer to Figure 2 , Figure 2 The horizontal axis represents the time window number, ranging from 0 to 47, and the vertical axis represents the normalized residual value, ranging from approximately 0.046 to 0.824. The curve sets represent cross-source consistency residuals, spatial continuity residuals, and mechanistic constraint residuals, respectively. The right-hand line annotation set indicates the residual type and line type corresponding to each curve. Figure 2 The formula for the unified calculation of the three types of residuals in the text is used to demonstrate the residual change process of the same grid within a continuous time window, which is composed of differences in multi-source estimation, discontinuity in spatial propagation, and deviation of variable mechanism.
[0030] from Figure 2 As can be seen, the spatial continuity residuals remain high around time windows 20 to 30, reaching approximately 0.808 around time window 26. This indicates a significant discontinuity between the risk state changes and permissible propagation directions between adjacent grids during this period, which may be related to spatial gradient anomalies caused by storm runoff, slope aspect constraints, or monitoring delays. The cross-source consistency residuals show a spike around time window 30, with a peak value of approximately 0.824, indicating a concentrated divergence of the same risk variable across different data sources. This divergence is more like a local evidence bias rather than a global risk increase.
[0031] The mechanistic constraint residuals were generally lower than the other two types of residuals, with a peak value of about 0.160. This indicates that the mechanistic constraints between disaster risk, carrying capacity, disaster-prone conditions and exposure degree are generally consistent, with only slight deviations in local time windows.
[0032] Cross-source consistency residuals are calculated based on the differences in multi-source estimates of the same risk variable within the same spatial grid and time window. Spatial continuity residuals are calculated based on the permissible changes in risk status and propagation direction at both ends of spatial adjacent edges. Mechanism-constrained residuals are calculated based on the constraint deviations between disaster hazard, environmental carrying capacity, disaster-inducing conditions, and exposure levels. After residual calculation, a residual record is formed containing residual type, associated nodes, associated edges, direction symbols, and time window identifiers. The direction symbol is used to indicate whether a certain piece of evidence makes the fusion result too high or too low. The associated edges are used to restrict the residuals to propagate back along the path consistent with their cause, avoiding the distribution of local anomalies to all data sources.
[0033] When correcting the data source credibility in the corresponding grid based on the data source and propagation path corresponding to various residuals, cross-source consistent residuals are traced back to the evidence records that provide the same risk variable; spatial continuity residuals are traced back to the risk evidence nodes upstream and downstream of the spatial adjacent edges; mechanistic constraint residuals are traced back to the variable nodes that participate in the constraint relationship; merged attribution is performed on data records with common sources; joint attribution is performed on data records with different sources but forming the same direction of deviation within the same time window; and priority retention items are determined for residuals with opposite directions based on the time sequence, propagation direction, and evidence timeliness. The correction only applies to the credibility status of the corresponding data source in the corresponding grid and the corresponding time window.
[0034] The revised credibility is calculated using the following formula: ; in, Indicates the credibility of the revised version. This represents the set of residual paths attributed to the data source d, the spatial grid g, and the time window t. This represents the attribution proportion of the p-th residual path. This represents the directed residual value corresponding to the residual path. This represents the response coefficient of the data source to residual correction. It represents the consistency measure that is repeatedly verified by independent evidence within a continuous time window. Represents the consistency recovery coefficient. This means limiting the calculation results to between 0 and 1. For example, when local rainfall data is continuously higher than normal and the propagation direction is inconsistent with that of adjacent grids, the reliability of the rainfall data in the corresponding grid and time window will be reduced only.
[0035] Constraint inferences for missing or delayed data are performed based on the corrected confidence level. For the target spatial grid and target time window, spatially adjacent nodes, disaster mechanism-related nodes, and nodes of the preceding and following time windows are extracted from the two-layer grid diagram. Nodes with confidence levels lower than the median state of the current candidate set and inconsistent propagation directions are removed. For the retained nodes, spatial propagation candidate values, mechanism constraint candidate values, and time continuation candidate values are generated respectively. Spatial propagation candidate values are obtained based on the relationship between the state and direction transformation of the upstream grid, mechanism constraint candidate values are obtained based on the allowable range between related risk variables, and time continuation candidate values are obtained based on the segmented change trend of the preceding and following time windows. Multiple candidate values are not directly averaged, but rather the value consistent with the existing evidence system is selected through residual back substitution.
[0036] The selection of candidate inferred values is performed according to the following formula: ; in, This represents the inferred values for the target spatial grid and the target time window. This represents the candidate set formed by spatial propagation, mechanistic constraints, and temporal continuity. This represents the cross-source consistency residual generated after the candidate value is written. Represents spatial continuity residuals. Represents the mechanism-constrained residuals. This indicates the deviation from the order of disaster evolution stages. , , and This represents the combination coefficient of the corresponding constraint. For example, if a candidate value is close to the neighborhood mean but would cause the mesh to jump directly from the incubation stage to the expansion stage, the deviation in the stage order would prevent it from being selected.
[0037] Please refer to Figure 3 , Figure 3The horizontal axis represents the candidate number, and the current exported graph shows candidates 1 to 18 ranked by the objective function value; the vertical axis represents the objective value and the values of each component. Different texture bars represent the objective value, cross-source components, spatial components, mechanistic components, and stage components, respectively. The legend further distinguishes between selected candidates and control candidates. Figure 3 The corresponding candidate inference value selection formula in the text is used to show the screening criteria formed by cross-source consistency, spatial continuity, mechanistic constraints and stage order after each candidate value is written into the temporary risk evidence copy.
[0038] from Figure 3 As can be seen, the target values of the leading candidates are generally lower, indicating that these candidates are more consistent with the existing evidence system after back-substitution of multiple constraints. The stage component maintains a stable bar height of close to 0.05 among multiple candidates, indicating that these candidates have not triggered significant stage order jumps. The cross-source component and spatial component show more significant differences among different candidates, indicating that the main distinction between candidate values comes from differences in multi-source estimation and the propagation relationship between adjacent grids. The mechanism component is higher in some candidates, suggesting that although the candidate may be close to the trend of the neighborhood, it may cause the combination of risk variables to deviate from the disaster mechanism constraint.
[0039] When outputting the comprehensive risk index and risk confidence interval for each spatial grid, the comprehensive risk index is jointly formed by the grid characterization values of disaster hazard, environmental carrying capacity, and disaster-inducing conditions. Actual observation evidence and constrained inference evidence are retained and fused according to different contribution upper limits. The risk confidence interval is jointly determined by observation timeliness, residual attribution, spatial propagation correlation, and inferred source path. The output results include at least the spatial grid code, time window code, current disaster evolution stage, lower bound of comprehensive risk index, central value of comprehensive risk index, upper bound of comprehensive risk index, main evidence sources, inferred evidence occupancy, and residual correction records. The output results can be used for subsequent high-risk area screening, disaster chain analysis, and risk early warning.
[0040] In a preferred embodiment, the formation of the grid risk evidence set includes: extracting disaster type, intensity, and movement trend from meteorological disaster data; elevation, slope, confluence path, and land cover status from geographic environmental data; population exposure and distribution of environmentally sensitive targets from socioeconomic data; occurrence location, duration, and impact range from historical disaster data; and environmental variable change sequences from real-time monitoring data; determining the coverage relationship with the spatial grid according to the spatial resolution of the data; establishing a coverage ratio index for data spanning multiple spatial grids; establishing a time window index for data with different time scales within the same spatial grid; and encapsulating the data value, coverage ratio, time window position, and data source identifier into risk evidence records in the grid risk evidence set.
[0041] Specifically, disaster types are recorded using a unified disaster type code, intensity is represented by a dimensionless interval that maintains the monotonic relationship of the original physical quantities, movement trend is represented by the location of the disaster center, the boundary of influence, and the main direction of movement at adjacent times, elevation, slope, and confluence path maintain the original spatial topological relationship, land cover status is coded according to categories related to disaster-inducing conditions such as surface infiltration, bareness, vegetation cover, and construction cover, population exposure and distribution of environmentally sensitive targets form grid values based on the coverage quantity and spatial aggregation state of each spatial grid, historical disaster occurrence location, duration, and impact range are converted into event records with start and end times and spatial boundaries, and real-time monitoring data are formed into a continuous sequence according to the collection order and anomaly markers are retained instead of being directly deleted before entering the grid.
[0042] The risk evidence record consists of data values, variable codes, spatial grid codes, time window positions, coverage ratios, data source identifiers, quality description vectors, and original record pointers. The original record pointers point to the record positions before the data enters the spatial grid. When data values are normalized, resampled, or constrained inferences are performed, the risk evidence record simultaneously adds processing type and previous record pointers, so that each transformation can be traced back to the original data. Coverage ratios and time window positions are stored as independent fields to avoid the inability to distinguish whether numerical changes come from the data itself or from spatial and temporal mappings after multiple fusions.
[0043] In a preferred embodiment, please refer to Figure 4 The determination of the disaster evolution stage includes: extracting the meteorological intensity change rate, surface environmental variable change rate, historical similar disaster trajectory matching degree, and risk migration direction of adjacent spatial grids from a continuous time window to construct a grid stage state sequence; aligning the grid stage state sequence with preset incubation, triggering, expansion, and decay stage state templates to obtain the matching probability of each stage; correcting the matching probability according to the stage transition constraints of adjacent time windows, and selecting the stage with the highest corrected probability as the current disaster evolution stage; generating the data source credibility based on the historical residual distribution, real-time completeness, and spatial coverage consistency of the data source in the corresponding stage, and organizing it into a dynamic credibility field according to the spatial grid and disaster evolution stage.
[0044] The grid-stage state sequence is formed using a fixed-length sliding time window. Each time window stores the direction of change in meteorological intensity, the cumulative amount of meteorological intensity, the direction of change in surface environmental variables, the abrupt change in surface environmental variables, the inflow and outflow of risk from adjacent spatial grids, and the historical trajectory matching degree. The rate of change is obtained by dividing the difference between adjacent valid observations by the time interval. When encountering missing observations, the calculation is not performed directly across the missing interval. Instead, the length of the missing interval is written into the quality description vector and the constraint inference results are called. The historical similar disaster trajectory matching degree is calculated after filtering historical trajectories based on disaster type, seasonal conditions, geographical conditions, and prior risk status, so as to avoid mixing historical processes with different disaster-causing mechanisms into the same stage template.
[0045] The dynamic credibility field is organized using data source, spatial grid, time window, and disaster evolution stage dimensions. Each credibility state is associated with the historical residual distribution, current integrity, current coverage consistency, and most recent correction record of the data source. When a spatial grid enters a new disaster evolution stage, it does not directly inherit the credibility of the previous stage. Instead, it reads the basic value from the historical state of the same data source in the new stage and updates it in combination with the current data quality description vector. When the disaster process propagates between adjacent grids, only credibility states with the same propagation direction and similar geographical conditions are allowed to provide migration references.
[0046] In a preferred embodiment, the construction of the two-layer grid map includes: establishing spatial adjacency edges based on the adjacency relationship between spatial grids, topographic elevation difference, confluence direction, and disaster movement direction; establishing disaster mechanism edges based on the constraint relationship between disaster hazard, environmental carrying capacity, disaster-inducing conditions, and exposure degree; connecting risk evidence nodes of the same spatial grid within a continuous time window through time-related edges, and recording data sources and time window identifiers for various edges; the cross-source consistency residual is obtained from the multi-source estimation difference of the same risk variable, the spatial continuity residual is obtained from the discontinuous change of risk state at both ends of the spatial adjacency edge, and the mechanism constraint residual is obtained from the deviation between the variables connected by the disaster mechanism edge.
[0047] The direction of spatial adjacency edges is determined according to the type of risk variable. For variables related to rainfall movement, directed edges are established according to the direction of disaster movement. For variables related to runoff, directed edges are established according to the direction of elevation descent and runoff path. For variables related to slope erosion and soil loss, directed edges are established according to slope aspect and downslope direction. For variables related to environmental exposure, bidirectional edges are established when there is no clear direction of propagation. The effective time window of the edge is jointly determined by the corresponding meteorological process, surface condition, and historical propagation time lag. If adjacent grids are adjacent but the elevation, slope aspect, or runoff relationship does not allow the current risk variable to propagate directly, then no spatial adjacency edge is established for the corresponding variable.
[0048] Disaster mechanism edges connect risk variables with conditional constraints within the same spatial grid. Triggering conditions are recorded between disaster hazard nodes and disaster-inducing condition nodes, carrying capacity constraints are recorded between disaster-inducing condition nodes and environmental carrying capacity nodes, impact contact relationships are recorded between disaster hazard nodes and exposure degree nodes, and risk amplification or buffering relationships are recorded between environmental carrying capacity nodes and comprehensive risk nodes. Each disaster mechanism edge stores the allowed input interval, allowed change direction, and time lag range. Mechanism edges do not directly provide fixed risk results, but are used to check whether the multi-source fusion results violate known variable relationships.
[0049] In a preferred embodiment, the establishment of the time window index includes: forming a reference time axis based on the update time of real-time monitoring data; projecting meteorological disaster data and historical disaster data onto the reference time axis according to the event duration interval; generating segmented retention intervals for geographical environment data and socio-economic data according to their validity period; for data with multiple observations between two adjacent update times, retaining the observation records corresponding to the inflection points and extreme values of change, and representing them as segmented change sequences; for data whose update time lags behind the reference time axis, recording the delay length and the previous valid observation position; and generating a traceable time window index based on the event duration interval, segmented retention interval, segmented change sequence, and delay length.
[0050] The baseline timeline is established based on the stable update times in the real-time monitoring data. When different real-time monitoring variables have different update frequencies, the common update times that can cover the main environmental state changes are selected as the baseline nodes. Higher frequency data are retained as subsequences between adjacent baseline nodes, and lower frequency data are attached to multiple baseline nodes according to their actual effective range. When a new baseline timeline is added, only the subsequent time window is added, and the historical time window that has already been risk fusion is not rewritten. If the collection time of late data falls into the historical time window, the local recalculation of the corresponding time window is triggered by delay recording.
[0051] Data whose update time lags behind the baseline time axis forms a delay record. The delay record includes the data source identifier, variable code, expected time window, actual arrival time window, delay length, previous valid observation position, and missing interval. The previous valid observation position is used to generate candidate values for time continuation. The delay length is entered into the data quality description vector. After the late data arrives, it is written back to the corresponding time window according to the original acquisition time, and the affected time correlation edge, residual record, and confidence status are recalculated. The traceable time window index saves the version relationship before and after the recalculation.
[0052] In a preferred embodiment, the generation of the dynamic credibility field includes: establishing a credibility state table for each data source, jointly indexed by the disaster evolution stage and the spatial grid; writing the residual statistics of the same stage, similar geographical conditions, and similar observation coverage in the historical disaster trajectory into the corresponding credibility state; when the current spatial grid lacks sufficient historical samples, selecting reference grids with the same risk propagation direction according to spatial adjacency edges and disaster mechanism edges, and generating migration coefficients based on the stage matching probability difference, geographical condition difference, and data coverage difference between the reference grid and the current spatial grid; using the migration coefficients to perform weighted migration of the credibility state of the reference grid, and updating the data source credibility of the current spatial grid together with the data integrity of the current time window.
[0053] The credibility status table uses data source coding, spatial grid coding, and disaster evolution stage coding as the main indexes. Each status stores the median status of historical residuals, residual direction distribution, continuous anomaly length, coverage stability, most recent update time, and applicable historical trajectory set. Historical residual statistics are extracted only from historical trajectories with consistent disaster types, similar geographical conditions, and similar spatial coverage. The same data source establishes statuses separately for different disaster evolution stages. Spatial propagation residuals formed in the expansion stage are not directly used for credibility initialization in the incubation stage.
[0054] The migration coefficient is determined by the stage matching probability difference, geographical condition difference, and data coverage difference. The smaller the stage matching probability difference and the more consistent the propagation direction, the higher the proportion of the corresponding reference state. The geographical condition difference is calculated based on elevation, slope, confluence relationship, and land cover category. The data coverage difference is calculated based on spatial coverage ratio, update frequency, and missing state. The migration coefficients of multiple reference grids are normalized and used to combine the confidence state. The migration result is only used as the basic confidence of the current spatial grid and does not cover the actual integrity of the current time window and the current residual record.
[0055] The data completeness of the current time window is determined by the number of expected observations, the number of actual observations, the length of consecutive missing data, and the effective coverage ratio. When the completeness update is combined with the confidence status after migration, the proportion of the current state is increased for data with sufficient actual observations and stable residual direction, and the proportion of the current state is decreased for data with consecutive missing data or abrupt changes in spatial coverage. The updated data source confidence is written into the dynamic confidence field and the reference grid, migration coefficient, and completeness record are saved so that the source of confidence can be identified in subsequent residual attribution.
[0056] In a preferred embodiment, the correction of the data source credibility includes: establishing residual attribution paths based on the risk evidence nodes, edge types, and time window identifiers associated with the residuals; allocating cross-source consistent residuals to each data source providing the same risk variable; allocating spatial continuity residuals in reverse along spatial adjacent edges to upstream risk evidence nodes; and allocating mechanism constraint residuals along disaster mechanism edges to risk evidence nodes that violate variable constraints; performing directional consistency judgment on various attribution residuals obtained from the same data source within the same spatial grid; forming a joint correction amount when multiple attribution residuals point to the same deviation direction; determining the retention item according to the temporal relationship of the edges and the risk propagation direction when the attribution residuals are opposite in direction; and updating the corresponding data source credibility accordingly.
[0057] The residual attribution path starts from the residual record and connects in sequence the risk variable node that generates the residual, the risk evidence record that participates in the fusion, and the original data source. Cross-source consistent residuals are allocated according to the direction and magnitude of the deviation of the estimated value of each data source from the reference center. Data sources whose deviation direction is consistent with the direction of the composite residual are given a higher attribution ratio. When tracing back along the spatial adjacent edge, the spatial continuity residual checks the timeliness, propagation direction, and effective time window of the upstream node. When allocating the mechanism constraint residual along the disaster mechanism edge, it checks whether each variable exceeds the corresponding allowable input interval.
[0058] When determining the directional consistency of various attribution residuals obtained from the same data source within the same spatial grid, the residuals are arranged in order of time window, and the continuous length and source independence of the same-direction residuals are calculated. If the cross-source consistency residuals, spatial continuity residuals, and mechanistic constraint residuals are generated by different evidence paths and are in the same direction, a joint correction quantity is formed. The joint correction quantity is combined according to the attribution ratio of the residual path and the timeliness of the evidence. If multiple residuals are formed by repeated propagation of the same original data, the common source merging is performed first, and then the joint correction is performed.
[0059] When the attribution residuals are in opposite directions, the temporal sequence is used to determine whether the early bias has been corrected by subsequent actual observations, the risk propagation direction is used to determine whether the current anomaly is located upstream or downstream, and the evidence validity is used to determine whether the data on which the conflicting residuals depend is still valid. The retained items are selected first from residuals that are consistent with the current propagation direction, are relatively new in time, and have independent data sources to support them. The unretained residuals are not deleted directly, but are recorded as conflicting states and re-evaluated when new evidence appears in subsequent time windows.
[0060] In a preferred embodiment, the constraint inference of missing or delayed data includes: determining the target spatial grid and target time window corresponding to the risk variable to be inferred; extracting candidate risk evidence nodes from the two-layer grid diagram that have spatial adjacency edges, disaster mechanism edges, or time-related edges with the target spatial grid; forming a candidate constraint set based on the data source credibility of the candidate risk evidence nodes, the consistency of the risk propagation direction with the target spatial grid, and the time distance; generating multiple candidate inference values using the candidate constraint set; calculating the cross-source consistency residual, spatial continuity residual, and mechanism constraint residual after introducing each candidate inference value; selecting the candidate inference value with the smallest comprehensive residual that does not change the order of the confirmed disaster evolution stages and writing it into the target time window; and recording the inference source path.
[0061] Candidate risk evidence nodes are extracted in groups according to edge type. Spatially adjacent candidate nodes must be located in the upstream or same-direction neighborhood where the risk variable is allowed to propagate. Disaster mechanism candidate nodes must have a clear constraint relationship with the risk variable to be inferred. Temporally related candidate nodes must belong to the same spatial grid and be located at the nearest valid position before or after the target time window. Candidate nodes also retain the corrected confidence, actual observation or inference identifier, temporal distance, spatial distance and source independence. Candidate evidence with the same source and only different paths to the target node are merged first.
[0062] After each candidate inferred value is written to a temporary copy of the target time window, the cross-source consistency residuals of the relevant risk variables, the spatial continuity residuals between the target grid and adjacent grids, the mechanistic constraint residuals between the target risk variable and the mechanistic associated variable, and the stage order deviation are recalculated. The candidate inferred value with the smallest comprehensive residual and the stage order is written to the formal risk evidence record. The formal record is marked as constrained inference and saves the candidate source, screening conditions, residual comparison results, and inference time. When the actual observation arrives late, the inferred value is replaced by the actual observation and a local recalculation is triggered.
[0063] In a preferred embodiment, obtaining the risk confidence interval includes: extracting the observation time interval of each risk evidence record based on the time window index, and extracting the residual attribution residual of each data source based on the residual attribution path; jointly grouping the observation time interval and residual attribution residual according to the data source, spatial grid, and disaster evolution stage to generate the uncertainty boundary of the corresponding risk evidence record; propagating the uncertainty boundary along the spatial adjacency edge, disaster mechanism edge, and temporal correlation edge in the two-layer grid diagram, and performing correlation merging between propagation paths with common risk evidence sources to obtain the joint uncertainty range of each risk variable; calculating the upper and lower bounds of the comprehensive risk index based on the joint uncertainty range to form the risk confidence interval.
[0064] The observation time interval is determined based on the data acquisition time, validity period, update delay, and current time window position. Data that is within the validity period and is continuously updated forms a narrower time interval, while data that is close to the validity period boundary or has a delay forms a wider time interval. The residual attribution residual is the residual part that cannot be explained by known evidence after confidence correction. The residual attribution residual is jointly grouped according to data source, spatial grid, and disaster evolution stage to prevent residuals from different stages and different propagation conditions from canceling each other out.
[0065] When performing correlation merging between propagation paths with common sources of risk evidence, common sources are identified based on the original record pointers. Boundaries formed from the same original rainfall data and propagated through multiple adjacent grids are not repeatedly accumulated. Boundaries from independent real-time monitoring data and independent historical trajectories are retained respectively. The correlation merging result forms the joint uncertainty range of each risk variable. The lower and upper bounds of the comprehensive risk index are calculated by the minimum and maximum combinations of each risk variable satisfying the mechanism constraints within the joint uncertainty range, respectively.
[0066] The confidence interval for risk is formed according to the following formula: ; in, This represents the lower bound of the risk for the g-th spatial grid at the t-th time window. This represents the upper bound of the corresponding risk, and y represents the risk variable vector composed of disaster hazard, environmental carrying capacity, disaster-prone conditions, and exposure level. This represents the feasible set jointly defined by the combined uncertainty range of each risk variable and the constraints of the disaster mechanism. This represents a comprehensive risk calculation function. For example, when the upper bound of a certain variable conflicts with the mechanism constraint, that value will not be included in the feasible set, and the risk upper bound is determined by the combination of variables that satisfy all constraints.
[0067] Please refer to Figure 5 , Figure 5 The horizontal axis represents the time window number, ranging from 0 to 47, and the vertical axis represents the normalized risk value. The gray band represents the confidence interval between the lower and upper risk bounds, the black main line represents the central risk, and the dashed line represents the interval width; the legend and line labels together explain the meaning of each line. Figure 5 The corresponding formula for forming the confidence interval of risk in the text is used to show the range of uncertainty of the comprehensive risk result in different time windows after combining the observation time interval, the remaining attribution residual, and the correlation of the propagation path.
[0068] from Figure 5As can be seen, the central risk reached approximately 0.512 around time window 20, and then gradually declined, indicating that after the grid entered a high-risk state in the middle time window, the risk intensity showed a trend of easing. The upper risk bound reached approximately 0.769 around time window 25, the lower risk bound reached approximately 0.367 around time window 20, and the interval width reached approximately 0.477 around time window 25. This indicates that although the central risk was already high at this stage, the uncertainty was also amplified, which may be related to late data, differences in inference paths, or incomplete elimination of residual attribution.
[0069] Figure 5 The key to engineering interpretation lies in distinguishing between an increase in the central value and an increase in the upper bound. If the central risk increases and the bandwidth narrows, it indicates that the high-risk judgment has stronger evidentiary consistency; if the upper bound increases but the central value does not increase synchronously, the priority of observation and verification should be maintained, rather than directly expanding the scope of treatment.
[0070] In a preferred embodiment, the output of the comprehensive risk index includes: associating the corrected data source credibility with the inference source path of the candidate inference values, forming risk evidence subsets according to actual observation evidence and constrained inference evidence; calculating grid characterization values of disaster hazard, environmental carrying capacity, and disaster-prone conditions for each risk evidence subset, and weighting each grid characterization value according to the dynamic credibility field; comparing the grid characterization values corresponding to actual observation evidence with the grid characterization values corresponding to constrained inference evidence, and forming an inference occupancy coefficient based on the difference between the two; using the inference occupancy coefficient to limit the contribution ratio of constrained inference evidence to the comprehensive risk index, and writing the comprehensive risk index, risk evidence composition, and corresponding disaster evolution stage into the spatial grid risk result set.
[0071] The subset of actual observation evidence includes data that can be traced back to the original collection records and has not undergone missing inference. The subset of constraint inference evidence includes data generated by spatial propagation, mechanistic constraints, or temporal continuity. After the corrected data source credibility is associated with the inference source path, each piece of inference evidence inherits the combination state of credibility on the source path and is supplemented with inference type, number of candidates, and comprehensive residual. Actual observation evidence and constraint inference evidence are respectively fused with similar variables to avoid inference data being mistakenly identified as independent actual observations after entering the fusion.
[0072] The inference occupancy coefficient is determined based on the difference between the actual observation evidence grid representation value and the constrained inference evidence grid representation value, the proportion of constrained inference evidence in all evidence, and the independence of the inference source path. When the actual observation evidence is sufficient and the difference between the two is large, the contribution of constrained inference evidence is limited. When the actual observation evidence is missing and multiple independent inference paths give consistent results, constrained inference evidence is allowed to participate in the comprehensive risk calculation within a set upper limit. The comprehensive risk index and the risk confidence interval are output together. The risk evidence composition is used to explain whether the central value mainly comes from actual observation or constrained inference.
[0073] In this embodiment, the comprehensive risk index can be calculated according to the following formula: ; in, This represents the overall risk index. , and These represent the grid characterization values of disaster risk, environmental carrying capacity, and disaster-inducing conditions formed by actual observational evidence. , and These represent the corresponding grid representation values formed by the constraint inference evidence. This represents the occupancy coefficient of actual observational evidence. G represents the constraint coefficient for the contribution of inferential evidence, and G represents the comprehensive function that maintains the monotonic relationship of risk variables, such as when the actual observational evidence is complete. Take values close to 1; the constraint inference part is only used to supplement missing variables.
[0074] Please refer to Figure 6 , Figure 6 The horizontal axis represents the time window number, ranging from 0 to 47, and the vertical axis represents the normalized risk value, ranging from approximately 0.056 to 0.554. The line annotation set lists the final risk curves for several key spatial grids, as well as the observed principal values and inferred contribution curves for representative grids; the low-level reference line and the high-level reference line are used to help determine the boundaries of daily concerns and key actions. Figure 6 The corresponding comprehensive risk index output formula in the text is used to demonstrate the joint contribution of actual observational evidence and constraint-inferred evidence to the risk outcome.
[0075] from Figure 6As can be seen, the final risk of key grids such as G1720, G1620, and G1820 reaches a high level around time window 33 to 34, with a peak value of approximately 0.554, and then generally falls back to around 0.49 to 0.50, indicating that the risk process underwent a temporal change from accumulation to expansion and gradual release. The observed principal value curve of G1620 is below the main interval of the final risk but maintains the same direction, indicating that actual observational evidence is the main source of the central risk; the inferred contribution curve of G1620 is generally low, with a peak value of approximately 0.14, indicating that constraint inference is mainly used to supplement missing or delayed windows and has not replaced actual observational evidence as the dominant factor.
[0076] Taking the regional assessment of the risk of rainstorms, floods, landslides and soil erosion caused by typhoons as an example, the typhoon movement path, wind field and rainfall process in meteorological disaster data are mapped to a continuous spatial grid. The elevation, slope, confluence path and land cover status in geographic environmental data form stable spatial constraints. The population exposure and distribution of environmentally sensitive targets in socio-economic data form exposure information. Historical disaster data provide stage templates and residual statistics under similar paths. Real-time monitoring data provide the change sequence of surface environmental variables. When rainfall in the region is still accumulating and the slope state has not changed, the stage determination remains in the incubation state. When rainfall continues and there is a sudden change in the surface state in the same direction, it switches to the trigger state.
[0077] When there is a delay in the real-time monitoring data of the downstream grid, candidate evidence is extracted from the upstream confluence-related grid, the preceding time window of the same grid, and risk variables that have mechanistic constraints with the surface state. Spatial propagation candidate values, temporal continuity candidate values, and mechanistic constraint candidate values are generated. After each candidate value is written to a temporary copy, the three types of residuals and the stage order deviation are recalculated. Candidate values that meet the requirements of confluence direction, mechanistic allowable interval, and do not cause stage jumps are written to the target time window. The inferred source path and candidate comparison records are saved synchronously. After the delayed actual observation arrives, the inferred value is replaced and the risk results are locally updated.
[0078] After the risk status moves from the triggering stage to the expansion stage, the spatial adjacent edges in the double-layer grid map open up the corresponding propagation paths according to the actual rainfall movement direction, confluence direction, and slope aspect. The spatial continuity residual adopts a comparison method that changes with the edge transformation relationship in the allowed propagation direction, without taking the equality of adjacent grid values as a continuity condition. The gradient formed by the real risk from upstream to downstream is retained, and the residuals formed by the reverse confluence direction or the abnormal gradients that do not meet the terrain constraints are back to the corresponding evidence sources. The risk confidence interval simultaneously reflects the actual observation timeliness, the remaining residuals, and the difference in the inferred path.
Claims
1. A secondary sudden environmental event risk assessment method based on multi-source data fusion, characterized in that, include: Collect meteorological disaster data, geographic environment data, socio-economic data, historical disaster data and real-time monitoring data of the target area, and map them to a pre-divided spatial grid according to a unified spatiotemporal identifier to form a grid risk evidence set; Data quality description vectors are generated based on the observation time, spatial coverage, update delay, missing status, and historical deviation of each data point, and the reliability of the data source is determined in conjunction with the current disaster evolution stage of the grid. A two-layer grid diagram containing spatial adjacency relationships and disaster mechanism correlations is constructed. The grid risk evidence is initially fused, and cross-source consistency residuals, spatial continuity residuals, and mechanism constraint residuals are calculated. Based on the data sources and propagation paths corresponding to various residuals, the data source credibility in the corresponding grid is corrected, and the constraint inference of missing or delayed data is completed based on the corrected data source credibility. The comprehensive risk index and risk credibility interval of each spatial grid are output. 2.The secondary burst environment event risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The formation of the grid risk evidence set includes: extracting disaster type, intensity and movement trend from meteorological disaster data, elevation, slope, confluence path and land cover status from geographic environmental data, population exposure and distribution of environmentally sensitive targets from socioeconomic data, location, duration and impact range from historical disaster data, and environmental variable change sequence from real-time monitoring data. The coverage relationship with the spatial grid is determined according to the spatial resolution of the data. A coverage ratio index is established for data that spans multiple spatial grids. A time window index is established for data with different time scales within the same spatial grid. The data value, coverage ratio, time window position and data source identifier are encapsulated into risk evidence records in the grid risk evidence set. 3.The secondary burst environment event risk assessment method based on multi-source data fusion according to claim 1, characterized in that, The determination of the disaster evolution stage includes: extracting the rate of change of meteorological intensity, the rate of change of surface environmental variables, the matching degree of historical similar disaster trajectories, and the risk migration direction of adjacent spatial grids from a continuous time window, and constructing a grid stage state sequence; The grid stage state sequence is aligned with the preset incubation, triggering, expansion and decay stage state templates to obtain the matching probability of each stage; The matching probability is corrected based on the stage transition constraints of adjacent time windows, and the stage with the highest corrected probability is selected as the current disaster evolution stage. The data source credibility is generated based on the historical residual distribution, real-time completeness, and spatial coverage consistency of the data source at the corresponding stage, and organized into a dynamic credibility field according to the spatial grid and the disaster evolution stage.
4. The secondary burst environment event risk assessment method based on multi-source data fusion according to claim 3, characterized in that, The construction of the double-layer grid map includes: establishing spatial adjacency edges based on the adjacency relationship between spatial grids, topographic elevation difference, confluence direction and disaster movement direction; and establishing disaster mechanism edges based on the constraint relationship between disaster risk, environmental carrying capacity, disaster-prone conditions and exposure degree. Connect risk evidence nodes within the same spatial grid and continuous time windows using time-related edges, and record the data source and time window identifier for each type of edge; The cross-source consistency residual is obtained from the difference in multi-source estimates of the same risk variable; the spatial continuity residual is obtained from the discontinuous changes in the risk states at both ends of the spatial adjacent edge; and the mechanism constraint residual is obtained from the deviation between the variables connected by the disaster mechanism edge.
5. The secondary burst environment event risk assessment method based on multi-source data fusion according to claim 2, characterized in that, The establishment of the time window index includes: forming a reference time axis based on the update time of real-time monitoring data; projecting the meteorological disaster data and the historical disaster data onto the reference time axis according to the event duration interval; and generating segmented retention intervals for the geographical environment data and the socio-economic data according to the validity period. For data with multiple observations between two adjacent update times, retain the observation records corresponding to the inflection points and extreme values of change, and represent them as a piecewise change sequence; For data whose update time lags behind the reference time axis, record the delay length and the previous valid observation position; A traceable time window index is generated based on the event duration interval, segmented retention interval, segmented change sequence, and delay length.
6. The method for risk assessment of secondary environmental emergencies based on multi-source data fusion according to claim 3, characterized in that, The generation of the dynamic credibility field includes: establishing a credibility state table for each data source, which is jointly indexed by the disaster evolution stage and the spatial grid, and writing the residual statistics of the same stage, similar geographical conditions and similar observation coverage in the historical disaster trajectory into the corresponding credibility state; When the current spatial grid lacks sufficient historical samples, reference grids with the same risk propagation direction are selected according to spatial adjacency edges and disaster mechanism edges, and migration coefficients are generated based on the differences in stage matching probability, geographical conditions, and data coverage between the reference grids and the current spatial grid. The reliability status of the reference grid is weighted and migrated using the migration coefficient, and the data source reliability of the current spatial grid is updated together with the data integrity of the current time window.
7. The method for risk assessment of secondary environmental emergencies based on multi-source data fusion according to claim 4, characterized in that, The correction of the data source credibility includes: establishing a residual attribution path based on the risk evidence node, edge type and time window identifier associated with the residual; allocating the cross-source consistent residual to each data source that provides the same risk variable; allocating the spatial continuity residual in reverse along the spatial adjacency edge to the upstream risk evidence node; and allocating the mechanism constraint residual along the disaster mechanism edge to the risk evidence node that violates the variable constraint. For each attribution residual obtained from the same data source within the same spatial grid, a directional consistency judgment is made. When multiple attribution residuals point to the same deviation direction, a joint correction is formed. When the attribution residuals are in opposite directions, the retained items are determined according to the temporal relationship of the edges and the direction of risk propagation, and the corresponding data source credibility is updated accordingly.
8. The method for risk assessment of secondary environmental emergencies based on multi-source data fusion according to claim 4, characterized in that, The constraint inference of missing or delayed data includes: determining the target spatial grid and target time window corresponding to the risk variable to be inferred, and extracting candidate risk evidence nodes from the two-layer grid diagram that have spatial adjacency edges, disaster mechanism edges or time correlation edges with the target spatial grid; A candidate constraint set is formed based on the data source credibility of the candidate risk evidence nodes, their consistency with the risk propagation direction of the target spatial grid, and their time distance. Multiple candidate inference values are generated using the candidate constraint set. The cross-source consistency residual, spatial continuity residual, and mechanism constraint residual after the introduction of each candidate inference value are calculated respectively. The candidate inference value with the smallest comprehensive residual that does not change the order of the confirmed disaster evolution stages is selected and written into the target time window. At the same time, the inference source path is recorded.
9. The method for risk assessment of secondary environmental emergencies based on multi-source data fusion according to claim 8, characterized in that, Obtaining the risk confidence interval includes: extracting the observation time interval of each risk evidence record based on the time window index, and extracting the remaining attribution residuals of each data source based on the residual attribution path; The observation time interval and the remaining attribution residual are jointly grouped according to the data source, spatial grid and disaster evolution stage to generate the uncertainty boundary of the corresponding risk evidence record; The uncertainty boundary is propagated along the spatial adjacency edge, disaster mechanism edge and temporal correlation edge in the two-layer grid diagram, and correlation merging is performed between propagation paths with common sources of risk evidence to obtain the joint uncertainty range of each risk variable; The upper and lower bounds of the comprehensive risk index are calculated based on the joint uncertainty range to form the risk confidence interval.
10. The method for risk assessment of secondary environmental emergencies based on multi-source data fusion according to claim 9, characterized in that, The output of the comprehensive risk index includes: associating the corrected data source credibility with the inference source path of the candidate inference value, and forming risk evidence subsets according to actual observation evidence and constrained inference evidence respectively; For each subset of risk evidence, the grid characterization values of disaster risk, environmental carrying capacity, and disaster-prone conditions are calculated, and the credibility of each grid characterization value is weighted according to the dynamic credibility field. Compare the grid representation values corresponding to actual observation evidence with the grid representation values corresponding to constraint inference evidence, and form an inference occupancy coefficient based on the difference between the two. The contribution ratio of the constraint inference evidence to the comprehensive risk index is limited by the inference occupancy coefficient, and the comprehensive risk index, risk confidence interval, risk evidence composition and corresponding disaster evolution stage are written into the spatial grid risk result set.
