Earthquake intensity mapping method based on multi-agent fusion social media evidence

Abstract

This seismic intensity mapping method, based on multi-agent fusion of social media evidence, belongs to the field of seismology. Addressing the shortcomings of existing early post-earthquake intensity maps—such as insufficient station constraints, delayed field investigations, and the difficulty in directly converting social media data into physically consistent intensity fields—this invention utilizes a multi-agent framework to transform multimodal social media data, including Weibo text, images, and videos, into social media intensity observations with location and confidence levels. Combined with strong-motion station data, it constructs instrumental intensity physical priors and fuses these data within the focal distance domain to generate dynamically updated seismic intensity maps and the distribution of facility damage and affected populations. This invention supports rapid post-earthquake identification of high-intensity areas, supports emergency dispatch and disaster assessment, and significantly improves the efficiency of earthquake emergency response.

Description

Technical Field

[0001] This invention relates to the field of seismology, specifically to a seismic intensity mapping method based on multi-agent fusion of social media evidence. Background Technology

[0002] Seismic intensity maps reflect the spatial distribution of earthquake impacts and disaster severity, serving as crucial evidence for post-earthquake emergency response, rescue resource allocation, traffic control assessment, field survey deployment, and disaster assessment. Traditional rapid intensity maps primarily rely on the epicenter, magnitude, empirical attenuation relationships, strong-motion station records, and subsequent field survey data. While these methods offer advantages such as clear physical meaning and stable results, they still face challenges in the early post-earthquake stages, including uneven spatial distribution of stations, insufficient constraints in mountainous or sparsely networked areas, and time lags in field surveys. This makes it difficult to accurately depict localized high-intensity areas and mitigate humanitarian impacts.

[0003] With the development of mobile internet and social media, the text, images, and video frames posted by the public on Weibo after an earthquake often contain information such as localized earthquake sensations, building damage, road blockages, bridge damage, landslides and collapses, disruption of lifelines, casualties, relocation and resettlement needs, and rescue requirements. This information is characterized by near real-time updates, strong locality, and rich human-centric impact, and can supplement early disaster information prior to station observations and on-site investigations. However, social media data suffers from problems such as high noise levels, numerous duplicate reposts, vague location descriptions, uneven spatial distribution, and significant differences in reliability. It cannot be directly used for interpolation of continuous seismic intensity fields; otherwise, population density, media attention, or platform activity may be mistakenly interpreted as the strength of the earthquake, resulting in spatial artifacts that are not constrained by physical reality.

[0004] In recent years, large language models and visual language models have been able to transform unstructured text, images, and video frames into structured disaster evidence. However, existing methods mostly remain at the level of event correlation classification, disaster information extraction, or point damage mapping, lacking a method to transform multimodal social media evidence into social media intensity observations that conform to the mechanism of seismic intensity and to constrain and integrate them with regional attenuation laws and station observations. Therefore, there is an urgent need for a dynamic seismic intensity mapping method that can simultaneously utilize multimodal social media evidence, station observations, and physical attenuation relationships to improve the ability to identify and conduct emergency screening of local high-intensity areas in the early post-earthquake period. Summary of the Invention

[0005] To address the technical problems of insufficient station constraints, delayed field investigations, and the difficulty in directly converting social media evidence into a physically consistent intensity field in existing post-early earthquake intensity mapping, this invention provides a seismic intensity mapping method based on multi-agent fusion of social media evidence, including:

[0006] Step 1: Obtain the earthquake occurrence time, epicenter location, magnitude, study area, regional intensity attenuation parameters, major axis azimuth of the regional empirical elliptical attenuation field, and strong motion station records of the target earthquake; determine the cumulative update time window; and obtain earthquake event configuration data.

[0007] Step 2: Based on the study area and cumulative time window in the earthquake event configuration data, construct a multi-agent evaluation framework consisting of a coordinator, a data acquisition agent, an evidence interpretation agent, a spatial positioning and aggregation agent, and a physical constraint fusion agent. The coordinator sends the corresponding input data source and processing boundary conditions to each agent.

[0008] Step 3: The data acquisition agent, based on the processing boundary conditions, collects social media text, images, videos, posting time, location metadata, and record identifiers within a preset time range after the target earthquake, according to the geographical dictionary of the study area and the search terms for earthquake events. Duplicate records are filtered out to obtain a multimodal social media record set.

[0009] Step 4: The evidence interpretation agent takes the multimodal social media record set as input and performs event filtering, geolocation, text evidence parsing, visual evidence parsing, multimodal evidence fusion, and usability checks one by one to obtain a structured social media evidence record set. Each record includes event relevance markers, geographic location, intensity availability, estimated intensity, information on people affected, information on facility damage, spatial granularity, confidence level, and quality markers.

[0010] Step 5: Spatial positioning and aggregation. The intelligent agent takes the structured social media evidence record set as input, selects records that simultaneously meet the requirements of being relevant to the target earthquake, having a locatable location, and having intensity-related physical effects as candidate records for social media intensity observation, and aggregates them according to geographical location and cumulative time window to obtain a set of social media intensity observation points. Each observation point includes geographical location, aggregated intensity value, and support weight.

[0011] Step 6: The physical constraint fusion agent takes the strong motion station records and regional intensity attenuation parameters as inputs, calculates the instrument intensity of the station based on the station peak acceleration and peak velocity, and combines the regional empirical elliptical attenuation field and the station residual Kriging correction to obtain the physical prior field of instrument intensity covering the entire study area.

[0012] Step 7: The physical constraint fusion agent takes the social media intensity observation point set as input, selects high-confidence near-source social media intensity observations within each cumulative time window, performs weighted principal axis fitting based on the support weight and intensity excess weight of the observations, obtains the main effect direction updated with the cumulative time window, and constructs a finite equivalent social media source line; determines the combined weight of point source distance and finite equivalent social media source line distance based on the number of high-confidence near-source social media intensity observations, maps each social media intensity observation point to the equivalent source distance domain, performs bin aggregation according to the equivalent source distance, and performs one-dimensional ordinary kriging interpolation on the aggregated intensity of non-empty distance bins within the equivalent source distance domain to generate a social media evidence-driven source effect intensity field, where two-dimensional geospatial interpolation is not performed on the social media intensity observation points;

[0013] Step 8: Using the social media intensity observation point set, the instrumental intensity physical prior field, and the source effect intensity field as inputs, construct a fading function based on the spatial coverage of the social media intensity observation point set, and weight and fuse the instrumental intensity physical prior field and the source effect intensity field according to the effective fusion coefficient. Dynamically update the fusion function as the cumulative time window expands to obtain the source effect intensity correction field and fused intensity map under different cumulative time windows.

[0014] Furthermore, in step three, the geographical dictionary of the study area includes: the standard names of administrative districts, townships, administrative villages, and public facilities points of interest within the study area; the earthquake event search terms are generated by combining earthquake event keywords with geographical names from the geographical dictionary of the study area.

[0015] Furthermore, in step four, the evidence interpretation agent includes a text evidence parsing module, a visual evidence parsing module, a multimodal evidence fusion module, and an availability check module. The text evidence parsing module is used to identify event relevance, geographical location, intensity-related physical effects, information on people affected, and information on facility damage from social media text. The visual evidence parsing module is used to identify building damage, road blockage, bridge damage, landslides, and lifeline facility damage scenes from images or videos. The multimodal evidence fusion module is used to generate intensity availability, estimated intensity, types of people affected, types of facility damage, and confidence levels based on text and visual evidence. The availability check module is used to determine whether structured social media evidence records can be used for social media intensity observation, facility damage distribution, or distribution of people affected.

[0016] Furthermore, in step five, the aggregation intensity value is determined using a high quantile statistical method, and the calculation formula is as follows:

[0017]

[0018] in, The aggregation intensity value is the geographic location r. Let i be the estimated intensity value of the i-th available social media record. For high-quantile aggregation operators, Let r be the set of candidate records for geographic location r.

[0019] The weights can be calculated using the following formula:

[0020]

[0021] in, For the support weights of geographic location r, For cumulative time window The number of available records at position r supported within the internal memory. Extract the average confidence level for the model of geographic location r. For the spatial granularity reliability of geographic location r.

[0022] Among them, the location-level spatial granularity reliability adopts the average value of spatial granularity reliability weighted by the record confidence, so that records with high confidence and fine positioning granularity have a higher contribution to the support weight.

[0023] Furthermore, in step six, the instrument intensity at the station is calculated as follows:

[0024]

[0025]

[0026] in, Let be the peak acceleration of the m-th station, in cm / s². The peak velocity of the m-th station is expressed in cm / s. for Corresponding intensity, for Corresponding intensity; when and When both are not less than 6 degrees, the instrument intensity at the station shall be the average of the two values; otherwise, the average shall be used. ;

[0027] The physical prior field of instrument intensity is calculated using the following formula:

[0028]

[0029] in, For the instrument intensity physical prior field, For the target grid point, For the region's empirical elliptic decay field, This represents the robust median deviation of the residuals of station instrument intensity relative to the regional empirical elliptical attenuation field. The station residual correction field is obtained by Kriging interpolation from the station residuals. This means that the value is limited to the range of 1 to 12 degrees.

[0030] Further, in step seven, the weighted principal axis fitting based on the observed intensity to obtain the main effect direction updated with the cumulative time window and construct a finite equivalent social media source line specifically involves: projecting the received high-confidence near-source social media intensity observations onto a local kilometer coordinate system with the epicenter as the origin; performing weighted principal axis fitting based on the observation support weights and intensity excess information to obtain the main effect direction updated with the cumulative time window; and constructing a finite equivalent social media source line based on the main effect direction and the spatial coverage of the high-confidence near-source social media intensity observations. For the target grid point... Its cumulative time window The equivalent focal distance is calculated using the following formula:

[0031]

[0032] in, Let x be the equivalent source distance of the target grid point. The epicenter location, The shortest distance from the target grid point x to the finite equivalent social media seismic source line. Geometric maturity coefficient;

[0033] Geometric maturity coefficient Calculate using the following formula:

[0034]

[0035] in, For cumulative time window Number of high-confidence near-source earthquake intensity observations on social media. The threshold for the transition from a point source to a finite equivalent social media source line. The threshold for contributing sufficient activation is determined for the finite equivalent social media source line.

[0036] Furthermore, in step seven, bin aggregation is performed according to the equivalent source distance, and one-dimensional ordinary kriging interpolation is performed on the aggregated intensity of non-empty distance bins within the equivalent source distance domain. Specifically, the social media intensity observation points are divided into multiple distance bins according to the equivalent source distance; the representative source distance is calculated by weighting the support weights of each observation point within the bin, the bin aggregated intensity is calculated using the high quantile operator, and the bin support weights are summed to obtain the bin support weights; the representative source distance, aggregated intensity, and bin support weights of each distance bin are used as inputs, and ordinary kriging interpolation is performed within the one-dimensional equivalent source distance domain. Ordinary kriging uses a spherical variogram model.

[0037] Furthermore, in step eight, the formula for calculating the effective fusion coefficient is:

[0038]

[0039] in, For effective fusion coefficient, The fusion coefficient is... It is a gradually diminishing function;

[0040] The formula for calculating the fading function is:

[0041]

[0042]

[0043] in, For cumulative time window The maximum focal distance supported by internal social media intensity observations. The distance that gradually diminishes on the outer side;

[0044] The fusion intensity map is determined by the following formula:

[0045]

[0046] in To integrate the intensity map, For the physical priors of instrument intensity, The source effect intensity field is driven by social media evidence; the intensity changes of the intensity map relative to the physical prior of the instrument intensity constitute the source effect intensity correction field.

[0047] Furthermore, the method also includes: using information on affected persons and damaged facilities from a structured social media evidence record set as input, generating the distribution of damaged facilities and the distribution of affected persons respectively, and then combining them with the intensity of fusion. Figure 1 The system outputs information on the following: personnel affected include one or more of the following: death, injury, missing persons, and relocation; facility damage information includes one or more of the following: buildings, transportation facilities, lifeline facilities, and public service facilities, where transportation facilities include roads and bridges, and lifeline facilities include power facilities, water supply facilities, and communication facilities; the distribution of facility damage and personnel affected is only used for early post-earthquake disaster analysis, emergency assessment, and priority ranking of on-site verification.

[0048] The beneficial effects of this invention are as follows: First, through multi-agent collaborative processing, local physical effects in microblog texts, images, and videos are transformed into traceable social media intensity observations with geographic location, confidence level, and spatial granularity information. Second, a stable instrumental intensity physical prior is constructed using regional empirical elliptic attenuation and station residual two-dimensional ordinary Kriging correction, and finite equivalent social media source lines are dynamically inferred from cumulative social media observations, performing Kriging interpolation on non-empty distance bins within the one-dimensional equivalent source distance domain. Third, the social media correction is restricted to the equivalent source distance range with observational support through a fading function, refining the near-source intensity distribution while maintaining the regional attenuation structure. Fourth, the fused intensity map can be dynamically updated with the cumulative time window, and the distribution of facility damage and personnel impact can be output for early post-earthquake emergency screening and on-site verification priority ranking. Attached Figure Description

[0049] Figure 1 This is a flowchart of the overall process of the earthquake intensity mapping method based on multi-agent fusion of social media evidence according to the present invention.

[0050] Figure 2 This is a flowchart illustrating the generation process of the regional empirical elliptical attenuation field, instrumental intensity physical prior, dynamic finite equivalent social media source line, social media evidence-driven source effect intensity field, source effect intensity correction field, and fused intensity map of this invention.

[0051] Figure 3 This is a graph showing the results of updating the source effect intensity field driven by social media evidence with the cumulative time window in the embodiment.

[0052] Figure 4 This is a graph showing the results of updating the source effect intensity correction field with the cumulative time window in the example;

[0053] Figure 5 This is the result of updating the intensity map with the cumulative time window in the embodiment;

[0054] Figure 6 This is a comparison chart of the intensity of different Qwen3 / Qwen3-VL models in the embodiments, where Figure 6 (a) is the Pearson correlation coefficient. Figure 6 (b) is the F1 value for the VII+ high-intensity seismic zone;

[0055] Figure 7 This is a map showing the distribution of facility damage and the distribution of affected personnel in the example. Figure 7 (a) Distribution of facility damage Figure 7 (b) is the distribution of people affected. Detailed Implementation

[0056] The technical solution of the present invention will be further described below with reference to embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention. In the following embodiments, process equipment or devices not specifically specified are all conventional equipment or devices in the art. Unless specifically specified, the technical means used in the embodiments of the present invention are all conventional means well known to those skilled in the art.

[0057] Example 1: The overall implementation process of the present invention is as follows Figure 1 As shown, the generation process of instrument intensity physical priors, social media evidence-driven source effect intensity field, source effect intensity correction field, and fused intensity map is as follows: Figure 2 As shown: A seismic intensity mapping method based on multi-agent fusion of social media evidence, including:

[0058] Step 1: Obtain the earthquake occurrence time, epicenter location, magnitude, study area, regional intensity attenuation parameters, and strong motion station records of the target earthquake; determine the cumulative update time window; and obtain earthquake event configuration data.

[0059] Step 2: Based on the study area and cumulative time window in the earthquake event configuration data, construct a multi-agent evaluation framework consisting of a coordinator, a data acquisition agent, an evidence interpretation agent, a spatial positioning and aggregation agent, and a physical constraint fusion agent. The coordinator sends the corresponding input data source and processing boundary conditions to each agent.

[0060] Step 3: The data acquisition agent, based on the processing boundary conditions, collects social media text, images, videos, posting time, location metadata, and record identifiers within a preset time range after the target earthquake, according to the geographical dictionary of the study area and the search terms for earthquake events. Duplicate records are filtered out to obtain a multimodal social media record set.

[0061] Step 4: The evidence interpretation agent takes the original multimodal social media record set as input and performs event filtering, geolocation, text evidence parsing, visual evidence parsing, multimodal evidence fusion and usability check one by one to obtain a structured social media evidence record set. Each record includes event relevance markers, geographic location, intensity availability, estimated intensity, information on people affected, information on facility damage, spatial granularity, confidence level and quality markers.

[0062] Step 5: Spatial positioning and aggregation. The intelligent agent takes the structured social media evidence record set as input, selects records that simultaneously meet the requirements of being relevant to the target earthquake, having a locatable location, and having intensity-related physical effects as candidate records for social media intensity observation, and aggregates them according to geographical location and cumulative time window to obtain a set of social media intensity observation points. Each observation point includes geographical location, aggregated intensity value, and support weight.

[0063] Step 6: The physical constraint fusion agent takes the strong motion station records and regional intensity attenuation parameters as input, calculates the instrument intensity of the station based on the station peak acceleration and peak velocity, and obtains the physical prior field of instrument intensity by combining the regional empirical elliptical attenuation field and the station residual Kriging correction.

[0064] Step 7: The physical constraint fusion agent takes the social media intensity observation point set as input, selects high-confidence near-source social media intensity observations within each cumulative time window, performs weighted principal axis fitting based on the support weight and intensity excess weight of the observations, obtains the main effect direction updated with the cumulative time window, and constructs a finite equivalent social media source line; determines the combined weight of point source distance and finite equivalent social media source line distance based on the number of high-confidence near-source social media intensity observations, maps each social media intensity observation point to the equivalent source distance domain, performs bin aggregation according to the equivalent source distance, and performs one-dimensional ordinary kriging interpolation on the aggregated intensity of non-empty distance bins within the equivalent source distance domain to generate a social media evidence-driven source effect intensity field;

[0065] Step 8: Using the social media intensity observation point set, the instrument intensity physical prior field, and the source effect intensity field as inputs, construct a fading function based on the spatial coverage of the social media intensity observation point set, and weight and fuse the instrument intensity physical prior field and the source effect intensity field according to the effective fusion coefficient. Dynamically update the fusion function as the cumulative time window expands to obtain the source effect intensity correction field and fused intensity map under different cumulative time windows.

[0066] Furthermore, the method also includes step nine: using the information on people affected and the information on facility damage from the structured social media evidence record set as input, generating the distribution of facility damage and the distribution of people affected respectively, and then combining them with the intensity of fusion. Figure 1 And output it.

[0067] Specifically, the main inventive points of this invention are: First, transforming multimodal social media records into traceable macro-intensity evidence, personnel impact evidence, and facility impact evidence through a multi-agent process; Second, aggregating social media records that conform to the physical effects of seismic intensity into social media intensity observations with location, intensity, confidence level, and spatial granularity reliability; Third, constructing instrumental intensity physical priors and dynamically inferring finite equivalent social media source lines from cumulative social media observations, mapping social media intensity observations to a one-dimensional equivalent source distance domain for binning aggregation and ordinary kriging interpolation, so that social media evidence only corrects the instrumental intensity physical priors within the equivalent source distance range supported by observations; Fourth, generating a fused intensity map updated with the cumulative time window, and further generating the distribution of facility damage and personnel impact.

[0068] The social media data in step three consists of Weibo records within a preset time range after the earthquake. In this example, the time range is 0 to 72 hours after the earthquake. The data types include text, images, and videos. The geographical dictionary includes counties, townships, villages, and other points of interest.

[0069] Furthermore, in step three, the geographical dictionary of the study area includes: the standard names of administrative districts, townships, administrative villages, and public facilities points of interest within the study area; the earthquake event search terms are generated by combining earthquake event keywords with geographical names from the geographical dictionary of the study area.

[0070] Specifically, in step four, a textual language model is used to perform text event screening, geolocation, intensity evidence extraction, and impact analysis, while a visual language model is used to perform image and video frame analysis. When textual and visual evidence are inconsistent, the multimodal evidence fusion module generates the final structured evidence results based on local physical effects, the clarity of evidence location, and confidence level. For isolated extreme estimates lacking physical damage evidence, confidence level and quality labeling are used to reduce the likelihood of them entering the social media intensity observation set.

[0071] The intensity availability determination in step four requires that the record simultaneously meet the following criteria: relevant to the target earthquake, geographically locatable, containing local physical effects that conform to the intensity description in GB / T 17742-2020, and reaching the minimum confidence level.

[0072] Furthermore, in step four, the evidence interpretation agent includes a text evidence parsing module, a visual evidence parsing module, a multimodal evidence fusion module, and an availability check module. The text evidence parsing module is used to identify event relevance, geographical location, intensity-related physical effects, information on people affected, and information on facility damage from social media text. The visual evidence parsing module is used to identify building damage, road blockage, bridge damage, landslides, and lifeline facility damage scenes from image or video frames. The multimodal evidence fusion module is used to generate intensity availability, estimated intensity, types of people affected, types of facility damage, and confidence levels based on text and visual evidence. The availability check module is used to determine whether structured social media evidence records can be used for social media intensity observation, facility damage distribution, or distribution of people affected.

[0073] Furthermore, in step seven, the high-confidence near-source social media intensity observations received within each cumulative time window are projected onto a local kilometer coordinate system with the epicenter as the origin; a weighted covariance matrix is ​​constructed based on the support weights of each observation and the excess intensity weights relative to the instrument intensity physical prior field; the eigenvector corresponding to the largest eigenvalue is determined as the main effect direction; and a finite equivalent social media source line updated with the cumulative time window is constructed along this direction.

[0074] Furthermore, in step seven, the social media intensity observation points are divided into multiple distance bins according to the equivalent source distance, and empty bins are not included in subsequent calculations; the representative source distance is calculated by weighting the support weights of the observation points within the bin, the bin aggregate intensity is calculated by the high quantile operator, and the bin support weights are summed by summing the support weights of the observation points; using the representative source distance and aggregate intensity of each non-empty distance bin as input, ordinary Kriging interpolation is performed in the one-dimensional equivalent source distance domain to generate a social media evidence-driven source effect intensity field.

[0075] Furthermore, in step five, the aggregation intensity value is determined using a high quantile statistical method, and the calculation formula is as follows:

[0076]

[0077] in, The aggregation intensity value is the geographic location r. Let i be the estimated intensity value of the i-th available social media record. For high-quantile aggregation operators, Let r be the set of candidate records for geographic location r.

[0078] The weights can be calculated using the following formula:

[0079]

[0080] in, For the support weights of geographic location r, For cumulative time window The number of available records at position r supported within the internal memory. Extract the average confidence level for the model of geographic location r. For the spatial granularity reliability of geographic location r.

[0081] Among them, the location-level spatial granularity reliability adopts the average value of spatial granularity reliability weighted by the record confidence, so that records with high confidence and fine positioning granularity have a higher contribution to the support weight.

[0082] Specifically, for the geographic location r, within the cumulative time window The set of available records within is denoted as:

[0083]

[0084] in, For the first The location of each record is obtained after geolocation. Indicates the correlation of the target earthquake. Indicates the availability of intensity. This indicates the time the record was published.

[0085] Fine-grained locations such as buildings, facilities, and roads are assigned higher spatial granularity reliability, while township and county-level locations are assigned lower spatial granularity reliability. Support weights are used for subsequent binning support weights and representative source distance calculations for source distance binning, and can be normalized before participating in the above calculations; the aggregated intensity after binning is interpolated using Kriging in the one-dimensional equivalent source distance domain.

[0086] Specifically, in step five, Q98 is used as the location-level aggregation high quantile operator to enhance the response to high-intensity local evidence in emergency screening, while reducing the impact of isolated extreme estimates on the results through repeated filtering and confidence screening.

[0087] Furthermore, in step six, the instrument intensity at the station is calculated as follows:

[0088]

[0089]

[0090] in, Let be the peak acceleration of the m-th station, in cm / s². The peak velocity of the m-th station is expressed in cm / s. for Corresponding intensity, for Corresponding intensity; when and When both are not less than 6 degrees, the instrument intensity at the station shall be the average of the two values; otherwise, the average shall be used. ;

[0091] The physical prior field of instrument intensity is calculated using the following formula:

[0092]

[0093] in, For the instrument intensity physical prior field, For the target grid point, For the region's empirical elliptic decay field, This represents the robust median deviation of the residuals of station instrument intensity relative to the regional empirical elliptical attenuation field. The station residual correction field is obtained by Kriging interpolation from the station residuals. This means that the value is limited to the range of 1 to 12 degrees.

[0094] Specifically, step six calculates the instrument intensity of the station based on the peak ground acceleration (PGA) and peak velocity (PGV) recorded by the strong motion station; generates a regional empirical elliptical intensity decay field using a regional empirical elliptical intensity decay model; calculates the residual between the station instrument intensity and the regional empirical elliptical decay field, and obtains the station residual correction field through median deviation correction and ordinary Kriging interpolation; and superimposes the regional empirical elliptical decay field, the median deviation of the station residual, and the station residual correction field to obtain the physical prior of the instrument intensity.

[0095] The regional empirical elliptical intensity attenuation field is represented along its major and minor axes as follows:

[0096]

[0097]

[0098] in, The magnitude is [missing information]. and These are the distances between the major and minor axes, respectively. , , , , , , , This represents the regional empirical attenuation parameter. For each grid point x, the regional empirical elliptical attenuation field is obtained based on the elliptical isoseismal conditions. .

[0099] For station m, the original residuals, median bias, and corrected residuals are as follows:

[0100]

[0101] The station residual correction field is obtained by using ordinary Kriging interpolation to determine the debiased residual:

[0102]

[0103] in, s The number of available strong earthquake stations to participate in residual correction. For the first The Kriging weights of a strong seismic station at target grid point x.

[0104] Further, in step seven, a weighted principal axis fitting is performed based on the support weights and intensity excess weights of the observations to obtain the main effect direction updated with the cumulative time window and to construct a finite equivalent social media source line. Specifically, the received high-confidence near-source social media intensity observations are projected onto a local kilometer coordinate system with the epicenter as the origin. A weighted covariance matrix is ​​constructed based on the support weights and intensity excess weights. The eigenvector corresponding to the largest eigenvalue is determined as the main effect direction, and a finite equivalent social media source line updated with the cumulative time window is constructed along the main effect direction. The equivalent source distance is calculated using the following formula:

[0105]

[0106] in, Let x be the equivalent source distance of the target grid point. The epicenter location, The shortest distance from the target grid point x to the finite equivalent social media seismic source line. Geometric maturity coefficient;

[0107] Geometric maturity coefficient Calculate using the following formula:

[0108]

[0109] in, For cumulative time window Number of high-confidence near-source earthquake intensity observations on social media. The threshold for initiating the transition from a point source to a finite equivalent social media source line. The threshold for contributing sufficient activation is determined for the finite equivalent social media source line.

[0110] Specifically, in step seven, geometric maturity... Number of high-confidence near-source intensity observations on social media Control, transition threshold from point source to finite equivalent social media source line and Preferably 5 and 20; when When the distance to the seismic source is ≤5, the distance to the seismic source is approximately the point source distance; when When the distance is ≥20, the focal distance is mainly based on the finite equivalent social media focal line distance.

[0111] This step divides the social media intensity observations into bins according to the equivalent focal distance. Empty bins are not included in the calculation. The preferred bin aggregation operator is Q98. The representative focal distance is calculated by weighting the support weights of each observation point in the bin. The support weights are summed to obtain the support weight of the bin. The representative focal distance and aggregated intensity of each non-empty bin are used as inputs to perform ordinary Kriging interpolation in the one-dimensional equivalent focal distance domain to generate a social media evidence-driven focal effect intensity field.

[0112] when ≤ When, the equivalent seismic source is approximately a point source; when ≥ At that time, the equivalent focal distance mainly reflects the geometry of the finite equivalent social media focal line. For the r-th social media intensity observation point, its spatial location is denoted as... Its cumulative time window The equivalent focal distance is denoted as And calculate using the following formula:

[0113]

[0114] Each social media intensity observation point is divided into multiple source distance bins according to its equivalent source distance, and a high quantile operator is used to aggregate the social media intensity observations in each source distance bin.

[0115] For the target grid point x, based on its equivalent focal distance... The bins representing the distances to the seismic sources represent the distances to the seismic sources. To assess the proximity between them, a social media evidence-driven intensity field of the source effect was constructed using one-dimensional ordinary kriging interpolation.

[0116]

[0117] in, This represents a one-dimensional ordinary kriging interpolation operator within the equivalent source distance domain, used to calculate the source effect intensity at the equivalent source distance corresponding to the target grid point based on the representative source distance and aggregate intensity of each non-empty distance bin; bin support weights are used for bin aggregation and representative source distance calculation.

[0118] Furthermore, in step eight, the formula for calculating the effective fusion coefficient is:

[0119]

[0120] in, For effective fusion coefficient, The fusion coefficient is... It is a gradually diminishing function;

[0121] The formula for calculating the fading function is:

[0122]

[0123]

[0124] in, For cumulative time window The maximum focal distance supported by internal social media intensity observations. The distance that gradually diminishes on the outer side;

[0125] The fusion intensity map is determined by the following formula:

[0126]

[0127] in To integrate the intensity map, For the physical priors of instrument intensity, The source effect intensity field is driven by social media evidence; the intensity changes of the intensity map relative to the physical prior of the instrument intensity constitute the source effect intensity correction field.

[0128] Specifically, as the cumulative time window gradually expands from the early post-earthquake period to a preset time range, the finite equivalent social media source line, the social media evidence-driven source effect intensity field, the source effect intensity correction field, and the fused intensity map are dynamically updated. The intensity change of the fused intensity map relative to the instrumental intensity physical prior constitutes the source effect intensity correction field, which is used to represent the intensity correction applied to the instrumental intensity physical prior by social media evidence during the fusion process considering the source effect.

[0129] The main fusion coefficient in this step Preferably 0.50; outer fading distance The preferred range is 35km.

[0130] Furthermore, the information on people affected includes one or more of the following: death, injury, missing persons, and relocation; the information on facilities damaged includes one or more of the following: buildings, transportation facilities, lifeline facilities, and public service facilities, where transportation facilities include roads, bridges, and tunnels, and lifeline facilities include power facilities, water supply facilities, and communication facilities; the distribution of facilities damaged and the distribution of people affected are only used for early post-earthquake disaster analysis, emergency assessment, and priority ranking of on-site verification, and are not used as the basis for official casualty or loss statistics.

[0131] Example 2: This example uses the 2022 Luding, Sichuan Ms 6.8 earthquake as an example to illustrate the implementation process of the present invention.

[0132] Step 1: Set the target earthquake parameters: The earthquake occurred at 12:52:18 on September 5, 2022, with the epicenter located at 102.08°E and 29.59°N, and the magnitude was Ms 6.8. The study area ranged from 101.0°E to 103.0°E and from 28.5°N to 30.5°N. The grid spacing was 0.01°, the cumulative update time window was 6 hours, and the azimuth of the major axis of the elliptical attenuation field was set according to the regional empirical attenuation model.

[0133] Step 2: Constructing a multi-agent evaluation framework coordinator. Based on the target seismic parameters and the study area, the data acquisition agent, evidence interpretation agent, spatial positioning and aggregation agent, and physical constraint fusion agent are all configured. The coordinator distributes the corresponding input data sources and processing boundary conditions to each agent. Each agent executes subsequent steps sequentially according to the scheduling rules.

[0134] Step 3: Collect Weibo data of the target earthquake: Within 72 hours after the earthquake, collect Weibo records containing keywords of the target earthquake and place names of the study area. After deduplication, 13,157 unique records are obtained. In the multimodal preparation stage, 7,836 local image files or copied images and 7,086 video frames are associated.

[0135] Step 4: Evidence interpretation using Qwen3 and Qwen3-VL models: Qwen3 is used to screen events, analyze evidence locations, extract intensity evidence, and analyze impact in text records; Qwen3-VL is used to analyze visible damage and impact evidence in images and video frames; and the final structured evidence is generated through multimodal evidence fusion.

[0136] Step 5: Geographically locate and aggregate social media intensity for available local physical effect records, with the minimum evidence confidence level set to 0.5 and Q98 used as the location-level aggregation operator; assign higher support weights to records with finer spatial granularity and sufficient repeated support.

[0137] Step 6: Construct a regional elliptical attenuation field using the intensity attenuation relationship in the Sichuan Basin, and construct a physical prior for instrument intensity using residual Kriging from strong motion stations; this physical prior serves as a stable baseline for subsequent social media evidence correction.

[0138] Step 7: Dynamically infer finite equivalent social media source lines from high-confidence near-source social media intensity observations. Within the equivalent source distance domain, perform bin aggregation with a 1km bin width, and construct a social media evidence-driven source effect intensity field using a one-dimensional ordinary kriging model with a spherical variogram. The transition threshold from point sources to finite equivalent social media source lines is determined. and Take values ​​of 5 and 20 respectively, and the outermost diminishing distance. Taking 35km as an example, the main fusion coefficient Take 0.50.

[0139] Step 8: Generate a fused intensity map from 6 hours to 72 hours post-earthquake and perform time convergence analysis:

[0140] Using the instrumental intensity physical prior constructed in step six as a fixed physical baseline, the cumulative time window is progressively expanded at 6-hour intervals. The finite equivalent social media source line and the social media evidence-driven source effect intensity field constructed in step seven are updated sequentially using post-earthquake accumulated social media evidence. Based on the fading function and the principal fusion coefficient, the social media evidence-driven source effect intensity field under each cumulative time window is fused with the instrumental intensity physical prior to generate the corresponding source effect intensity correction field and fused intensity map.

[0141] like Figure 3 As shown, with the continuous expansion of the cumulative time window, the intensity field of the source effect driven by social media evidence gradually changed from a relatively simple near-source distribution in the early post-earthquake period to a more stable directional high-intensity pattern; for example... Figure 4 As shown, the source effect intensity correction field reflects the positive or negative intensity correction imposed by social media evidence relative to the instrumental intensity physical prior; such as Figure 5 As shown, the fusion intensity map, while maintaining the overall attenuation structure of the instrument intensity physical prior, gradually refines the high-intensity zone near the epicenter as social media evidence accumulates.

[0142] like Figure 6 As shown, using the instrument intensity physical prior as a static baseline, the temporal convergence of the Qwen3 / Qwen3-VL models for earthquakes 4B, 8B, and 32B is compared under various cumulative time windows within 72 hours post-earthquake. Figure 6 (a) represents the Pearson correlation coefficient between the fused intensity map and the official intensity reference map. Figure 6 (b) represents the F1 value in the VII+ high-intensity seismic zone. Figure 6 (c) represents the F1 value of the VIII+ high-intensity zone. The results show that the Qwen3 / Qwen3-VL model of 32B improves the evaluation indicators more rapidly in the early post-earthquake period and begins to form a relatively stable spatial pattern of the main intensity around 36 hours after the earthquake; thereafter, the accumulated social media evidence is mainly used to refine the local intensity map and improve its stability.

[0143] Within the cumulative time window of 72 hours post-earthquake, the Pearson correlation coefficient between the fused intensity map generated by the Qwen3 / Qwen3-VL model of 32B and the official intensity reference map was 0.73, the F1 values ​​for the high intensity zones of VII+ and VIII+ were 0.90 and 0.88, respectively, and the coefficient of determination R² for distance attenuation analysis was 0.726.

[0144] Step Nine: Generate Facility Damage Distribution and Personnel Affected Distribution: Generate impact analysis results based on structured facility damage information and personnel affected information. For example... Figure 7 As shown, Figure 7 (a) Distribution of facility damage, including buildings, transportation facilities, lifeline facilities, public service facilities and other facilities; Figure 7 (b) Distribution of affected persons, including death, injury, missing persons and relocation.

[0145] As can be seen from the above embodiments, the present invention can transform noisy multimodal microblog records into traceable, geolocated social media intensity observations, dynamically infer finite equivalent social media source lines from cumulative observations, and form a source effect intensity field in the one-dimensional equivalent source distance domain through ordinary kriging; under the constraints of instrument intensity physical priors and fading functions, a fused intensity map updated with the cumulative time window is generated.

Claims

1. A method for earthquake intensity mapping based on multi-agent fusion social media evidence, characterized in that, include: Step 1: Obtain the earthquake occurrence time, epicenter location, magnitude, study area, regional intensity attenuation parameters, major axis azimuth of the regional empirical elliptical attenuation field, and strong motion station records of the target earthquake; determine the cumulative update time window; and obtain earthquake event configuration data. Step 2: Based on the study area and cumulative time window in the earthquake event configuration data, construct a multi-agent evaluation framework consisting of a coordinator, a data acquisition agent, an evidence interpretation agent, a spatial positioning and aggregation agent, and a physical constraint fusion agent. The coordinator sends the corresponding input data source and processing boundary conditions to each agent. Step 3: The data acquisition agent, based on the processing boundary conditions, collects social media text, images, videos, posting time, location metadata, and record identifiers within a preset time range after the target earthquake, according to the geographical dictionary of the study area and the search terms for earthquake events. Duplicate records are filtered out to obtain a multimodal social media record set. Step 4: The evidence interpretation agent takes the multimodal social media record set as input and performs event filtering, geolocation, text evidence parsing, visual evidence parsing, multimodal evidence fusion, and usability checks one by one to obtain a structured social media evidence record set. Each record includes event relevance markers, geographic location, intensity availability, estimated intensity, information on people affected, information on facility damage, spatial granularity, confidence level, and quality markers. Step 5: Spatial positioning and aggregation. The intelligent agent takes the structured social media evidence record set as input, selects records that simultaneously meet the requirements of being relevant to the target earthquake, having a locatable location, and having intensity-related physical effects as candidate records for social media intensity observation, and aggregates them according to geographical location and cumulative time window to obtain a set of social media intensity observation points. Each observation point includes geographical location, aggregated intensity value, and support weight. Step 6: The physical constraint fusion agent takes the strong motion station records and regional intensity attenuation parameters as inputs, calculates the instrument intensity of the station based on the station peak acceleration and peak velocity, and combines the regional empirical elliptical attenuation field and the station residual Kriging correction to obtain the physical prior field of instrument intensity covering the entire study area. Step 7: The physical constraint fusion agent takes the social media intensity observation point set as input, selects high-confidence near-source social media intensity observations within each cumulative time window, performs weighted principal axis fitting based on the observed intensity, obtains the main effect direction updated with the cumulative time window, and constructs a finite equivalent social media source line; The combined weights of point source distance and finite equivalent social media source line distance are determined, and each social media intensity observation point is mapped to the equivalent source distance domain. The bins are aggregated according to the equivalent source distance, and the aggregated intensity of the non-empty distance bins is interpolated by a Vickridge interpolation within the equivalent source distance domain to generate a social media evidence-driven source effect intensity field. Step 8: Using the social media intensity observation point set, the instrumental intensity physical prior field, and the source effect intensity field as inputs, construct a fading function based on the spatial coverage of the social media intensity observation point set, and weight and fuse the instrumental intensity physical prior field and the source effect intensity field according to the effective fusion coefficient. Dynamically update the fusion function as the cumulative time window expands to obtain the source effect intensity correction field and fused intensity map under different cumulative time windows.

2. The earthquake intensity mapping method based on multi-agent fusion of social media evidence as described in claim 1, characterized in that, In step three, the geographical dictionary of the study area includes: the standard names of administrative districts, townships, administrative villages and public facilities points of interest within the study area; the earthquake event search terms are generated by combining earthquake event keywords with geographical names in the geographical dictionary of the study area.

3. The earthquake intensity mapping method based on multi-agent fusion of social media evidence as described in claim 1, characterized in that, In step four, the evidence interpretation agent includes a text evidence parsing module, a visual evidence parsing module, a multimodal evidence fusion module, and an availability check module. The text evidence parsing module is used to identify event relevance, geographical location, intensity-related physical effects, information on people affected, and information on facility damage from social media text. The visual evidence parsing module is used to identify building damage, road blockage, bridge damage, landslides, and lifeline facility damage scenes from images or videos. The multimodal evidence fusion module is used to generate intensity availability, estimated intensity, types of people affected, types of facility damage, and confidence levels based on text and visual evidence. The availability check module is used to determine whether structured social media evidence records can be used for social media intensity observation, facility damage distribution, or distribution of people affected.

4. The earthquake intensity mapping method based on multi-agent fusion of social media evidence according to claim 1, characterized in that, In step five, the aggregation intensity value is determined using the high quantile statistical method, and the calculation formula is as follows: in, The aggregation intensity value is the geographic location r. Let i be the estimated intensity value of the i-th available social media record. For high-quantile aggregation operators, Let r be the set of candidate records for geographic location r. The weights can be calculated using the following formula: in, For the support weights of geographic location r, For cumulative time window The number of available records at position r within the internal support. Extract the average confidence level for the model of geographic location r. For the spatial granularity reliability of geographic location r.

5. The earthquake intensity mapping method based on multi-agent fusion of social media evidence according to claim 1, characterized in that, In step six, the instrument intensity at the station is calculated as follows: in, Let m be the peak acceleration of the m-th station, in cm / s². The peak velocity of the m-th station is expressed in cm / s. for Corresponding intensity, for Corresponding intensity; when and When both are not less than 6 degrees, the instrument intensity at the station shall be the average of the two values; otherwise, the average shall be used. ; The physical prior field of instrument intensity is calculated using the following formula: in, For the instrument intensity physical prior field, For the target grid point, For the region's empirical elliptic decay field, This represents the robust median deviation of the residuals of the station instrument intensity relative to the regional empirical elliptical attenuation field. The station residual correction field is obtained by Kriging interpolation from the station residuals. This means that the value is limited to the range of 1 to 12 degrees.

6. The earthquake intensity mapping method based on multi-agent fusion of social media evidence according to claim 1, characterized in that, In step seven, the weighted principal axis fitting based on the observed intensity to obtain the main effect direction updated with the cumulative time window and to construct a finite equivalent social media source line specifically involves: projecting the received high-confidence near-source social media intensity observations onto a local kilometer coordinate system with the epicenter as the origin; constructing a weighted covariance matrix of coordinate deviations based on the support weights and intensity excess weights of the high-confidence near-source social media intensity observation points; determining the eigenvector corresponding to the largest eigenvalue of the weighted covariance matrix as the main effect direction; and constructing a finite equivalent social media source line updated with the cumulative time window along the main effect direction; the equivalent source distance is calculated using the following formula: in, Let x be the equivalent source distance of the target grid point. The epicenter location, The shortest distance from the target grid point x to the finite equivalent social media seismic source line. Geometric maturity coefficient; Geometric maturity coefficient Calculate using the following formula: in, For cumulative time window Number of high-confidence near-source earthquake intensity observations on social media. The threshold for initiating the transition from a point source to a finite equivalent social media source line. The threshold for contributing sufficient activation is determined for the finite equivalent social media source line.

7. The earthquake intensity mapping method based on multi-agent fusion of social media evidence according to claim 1, characterized in that, In step seven, bin aggregation is performed based on equivalent source distance, and one-dimensional ordinary kriging interpolation is performed on the aggregated intensity of non-empty distance bins within the equivalent source distance domain. Specifically, the social media intensity observation points are divided into multiple distance bins based on equivalent source distance, skipping empty bins; the representative source distance is calculated by weighting the support weights of each observation point within a bin, the bin aggregated intensity is calculated using a high quantile operator, and the bin support weights are summed to obtain the bin support weights; using the representative source distance, aggregated intensity, and bin support weights of each non-empty distance bin as input, ordinary kriging interpolation is performed within the one-dimensional equivalent source distance domain, with the ordinary kriging using a spherical variogram model.

8. The earthquake intensity mapping method based on multi-agent fusion of social media evidence according to claim 1, characterized in that, In step eight, the formula for calculating the effective fusion coefficient is: in, For effective fusion coefficient, The fusion coefficient is... It is a gradually diminishing function; The formula for calculating the fading function is: in, For cumulative time window The maximum focal distance supported by internal social media intensity observations. The distance that gradually diminishes on the outer side; The fusion intensity map is determined by the following formula: in, To integrate the intensity map, For the physical priors of instrument intensity, The source effect intensity field is driven by social media evidence; the intensity changes of the intensity map relative to the physical prior of the instrument intensity constitute the source effect intensity correction field.

9. The earthquake intensity mapping method based on multi-agent fusion of social media evidence according to claim 1, characterized in that, The method also includes: taking the information on people affected and the information on facility damage from the structured social media evidence record set as input, generating the distribution of facility damage and the distribution of people affected, and outputting them together with the fused intensity map; the information on people affected includes one or more of death, injury, missing persons, and relocation; the information on facility damage includes one or more of buildings, transportation facilities, lifeline facilities, and public service facilities, of which transportation facilities include roads, bridges, and tunnels, and lifeline facilities include power facilities, water supply facilities, and communication facilities; the distribution of facility damage and the distribution of people affected are only used for early post-earthquake disaster analysis, emergency assessment, and priority ranking of on-site verification, and are not used as the basis for official casualty or loss statistics.

Images