News event spatialization verification analysis method and system based on satellite remote sensing image

Abstract

The invention provides a news event spatialization verification analysis method and system based on a satellite remote sensing image, and relates to the technical field of satellites, and the method comprises the steps: extracting spatio-temporal information from a news text through natural language processing, obtaining a remote sensing image based on multi-source satellite collaborative observation, and extracting ground feature elements and change features through a deep learning model; and constructing an event feature association graph for graph reasoning, and calculating the credibility of the event in combination with multi-dimensional cross validation and transfer learning to realize intelligent verification of the authenticity of the news event. The automation degree and accuracy of news event verification can be effectively improved, and the manual verification cost is reduced.

Description

[0001] Spatial verification and analysis method and system for news events based on satellite remote sensing images Technical Field

[0002] The present invention relates to satellite technology, and in particular to a spatial verification and analysis method and system for news events based on satellite remote sensing images. Background Art

[0003] With the rapid development of the Internet and social media, the speed of news event information dissemination is accelerating. Traditional news event authenticity verification mainly relies on manual verification and text analysis, which is inefficient and difficult to guarantee accuracy. In recent years, with the advancement of remote sensing technology, satellite remote sensing images have shown great potential in the field of event verification. At present, some studies have begun to try to apply satellite remote sensing data to the authenticity verification of news events, and verify the authenticity of the event by comparing and analyzing the characteristics of ground object changes in remote sensing images.

[0004] The main problems of the existing technology: Existing satellite remote sensing verification methods often rely only on a single type of satellite data and lack the ability to coordinate observations of multi-source remote sensing data. This results in the acquisition of incomplete image information and makes it difficult to meet the complex and diverse verification needs of news events.

[0005] Current remote sensing image analysis mainly uses traditional image processing methods, which lacks intelligent analysis models for different types of events and cannot effectively extract key event-related features, affecting the accuracy of verification results.

[0006] Existing event verification methods generally lack knowledge accumulation and experience inheritance mechanisms, and fail to fully utilize the empirical knowledge in historical verification cases, resulting in low verification efficiency and difficulty in meeting the verification needs of new types of events. Summary of the invention

[0007] The embodiments of the present invention provide a method and system for spatial verification and analysis of news events based on satellite remote sensing images, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention, Provides spatial verification and analysis methods for news events based on satellite remote sensing images, including: Receive news event text information, and use a natural language processing model to extract event timestamps, geospatial coordinates, and event type identifiers from the news event text information; determine a target observation area based on the extracted geospatial coordinates, and calculate an optimal satellite observation time window based on the event timestamp; input the target observation area and the optimal satellite observation time window into a strategy correction model to generate a satellite cluster collaborative observation strategy, wherein the satellite cluster collaborative observation strategy defines a collaborative observation sequence of high-resolution satellites, high-phase satellites, and hyperspectral satellites; issue a scheduling instruction based on the satellite cluster collaborative observation strategy to activate the corresponding satellite cluster to perform a collaborative observation task; Receive multi-source remote sensing image data acquired during the satellite group's collaborative observation mission, perform adaptive correction processing on the multi-source remote sensing image data, and generate standardized image data with spatiotemporal registration; select corresponding target detection networks and scene segmentation networks from a deep learning model library according to the event type identifier, perform intelligent analysis on the standardized image data, and extract event-related ground object elements and spatiotemporal variation characteristics; input the extracted ground object elements and spatiotemporal variation characteristics into a graph neural network, integrate historical verification knowledge to construct an event feature association graph, and obtain feature association strength values ​​through graph reasoning calculations; The feature association strength value is cross-validated with the event timestamp and the geographic spatial coordinates in multiple dimensions to construct an event authenticity evaluation index; based on the event authenticity evaluation index and in combination with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score; when the event credibility score exceeds a preset benchmark credibility threshold, the authenticity of the news event is confirmed, and the satellite cluster collaborative observation strategy, the event feature association map, and the event authenticity evaluation index in this verification process are written into the verification knowledge base.

[0009] Determining a target observation area based on the extracted geospatial coordinates, and calculating an optimal satellite observation time window according to the event timestamp; inputting the target observation area and the optimal satellite observation time window into a strategy correction model, and generating a satellite group collaborative observation strategy includes: Receive the geospatial coordinates of a news event, use the geospatial coordinates of the event as a seed point, iteratively expand the target area based on an adaptive region growing algorithm, obtain the geographic information entropy value and the spatial correlation in each iterative calculation process, determine the boundary of the target observation area when the geographic information entropy value and the spatial correlation meet a preset expansion threshold, and assign an observation priority weight to each sub-area within the boundary of the target observation area, so as to generate a weighted target observation area set; Acquire the event timestamp and the weighted target observation area set, construct the atmospheric visibility function, the solar altitude angle function, and the satellite observation angle function as an observation suitability evaluation function in the time dimension, perform time optimization on the weighted target observation area set based on the observation suitability evaluation function, determine differentiated observation time windows according to the observation priority weights of each sub-area, and generate the optimal observation time series for each sub-area; A satellite observation capability evaluation matrix is ​​established based on the optimal observation time sequence of the sub-region, wherein each element of the satellite observation capability evaluation matrix represents the observation capability score of a single satellite for a specific sub-region within a specified time window, a multi-dimensional optimization objective function is constructed based on the observation capability score and satellite mission time, energy consumption, and data transmission delay, and an initial satellite group collaborative observation strategy is obtained through iterative optimization and solving, wherein the initial satellite group collaborative observation strategy includes the observation timing, attitude adjustment sequence, and data transmission path of each satellite; The real-time status of the satellites during the execution of the initial satellite cluster collaborative observation strategy is monitored to obtain satellite position deviation values, observation condition change values, and energy status change values. When any change value exceeds a preset fluctuation threshold, the deviation between any change value and the original observation parameter is input into a strategy correction model. The strategy correction model dynamically optimizes the observation timing of the satellite, the attitude adjustment sequence, and the data transmission path based on the observation priority weights of each sub-area to generate an updated satellite cluster collaborative observation strategy.

[0010] The extracted ground features and spatiotemporal change characteristics are input into the graph neural network, and the historical verification knowledge is integrated to construct the event feature association graph. The feature association strength values ​​obtained through graph reasoning calculation include: Receiving ground feature elements and spatiotemporal change characteristics, constructing the ground feature elements into feature vectors including category attributes, spatial attributes, and temporal attributes, constructing the spatiotemporal change characteristics into change vectors describing the evolution of ground features, and constructing an initial feature graph using the feature vectors and the change vectors as nodes; Based on the current event type, historical verification cases are extracted from the verification knowledge base to construct a knowledge subgraph, and the cosine similarity and spatial position distance of the feature vectors of each node in the initial feature graph and the nodes in the knowledge subgraph are calculated to generate a node similarity matrix; The node similarity matrix is ​​input into a knowledge transfer mapping model, and the knowledge transfer mapping model projects the verification experience in the knowledge subgraph to the corresponding node of the initial feature graph to generate an enhanced feature graph integrating the verification experience; The enhanced feature graph is input into a graph attention network, which aggregates and updates node features based on attention coefficients between nodes, calculates edge feature values ​​according to the updated node features, uses the edge feature values ​​to adjust the connection relationship between nodes, and outputs a feature association graph; for each pair of connected nodes in the feature association graph, the node features and edge features are input into a multilayer perceptron to calculate feature association strength values.

[0011] The feature association strength value is cross-validated with the event timestamp and the geographic space coordinates in multiple dimensions to construct an event authenticity evaluation index; based on the event authenticity evaluation index and in combination with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score, including: Calculating the rationality score of the event occurrence sequence based on the event timestamp, calculating the relevance score of the geographic location based on the geographic spatial coordinates, calculating the feature consistency score based on the feature association strength value, and fusing the rationality score of the sequence, the relevance score of the geographic location and the feature consistency score to construct an event authenticity evaluation index; Retrieving historical verification cases from a verification knowledge base, constructing the historical verification cases into a source domain feature set, and constructing the current event to be evaluated into a target domain feature set; inputting the source domain feature set and the target domain feature set into a feature encoder for feature mapping, and the feature encoder maps the features to a common feature space to generate source domain encoding features and target domain encoding features; The source domain encoding features and the target domain encoding features are input into a domain discriminator, the domain discriminator calculates the difference between the distribution of source domain features and target domain features, and optimizes the feature encoder based on the difference to make the distribution of the source domain features and the target domain features tend to be consistent; the optimized target domain encoding features are input into an authenticity evaluator, and the authenticity evaluator calculates the event credibility score based on the source domain verification experience.

[0012] Calculating the rationality score of the event occurrence sequence based on the timestamp, calculating the relevance score of the geographic location based on the geographic space coordinates, calculating the feature consistency score based on the feature association strength value, and fusing the rationality score, the relevance score and the feature consistency score to construct an event authenticity evaluation index includes: A time reference sequence model is constructed based on the timestamp, the time reference sequence model uses a Gaussian kernel function with adaptive weights to calculate the time difference between the current event timestamp and the time point of the historical associated event, calculates the state transition probability matrix of the event sequence through a time Markov chain model, and generates a time series rationality score based on the time difference and the state transition probability matrix; The temporal rationality score is used as a constraint condition and combined with the geographic spatial coordinates to construct a spatial feature description model, wherein the spatial feature description model uses Hausdorff distance to calculate a spatial distance matrix between event locations and a set of geographic reference points, extracts spatial distribution features based on the spatial distance matrix, calculates the matching degree between the spatial distribution features and the spatial constraint rules in the geographic information system, and generates a geographic location relevance score; A feature verification model is constructed based on the temporal rationality score and the geographic location correlation score, the feature verification model normalizes the feature correlation strength value, uses a multi-layer perceptron to extract the deep spatiotemporal correlation pattern between features, and performs consistency verification between the deep spatiotemporal correlation pattern and a preset feature constraint rule set to generate a feature correlation consistency score; An adaptive weight allocation network including a feature extraction layer, an attention mechanism layer and a normalized output layer is constructed, and the rationality score of the time series, the correlation score of the geographic location and the feature correlation consistency score are input into the adaptive weight allocation network. The adaptive weight allocation network generates a dynamic fusion weight based on the event type feature vector, and the three scores are weightedly fused based on the dynamic fusion weight to generate an event authenticity evaluation index.

[0013] Inputting the source domain encoding features and the target domain encoding features into a domain discriminator, the domain discriminator calculates the difference between the source domain features and the target domain features, and optimizing the feature encoder based on the difference so that the source domain and the target domain features are distributed in a consistent manner; inputting the optimized target domain encoding features into an authenticity evaluator, and the authenticity evaluator calculates the event credibility score based on the source domain verification experience, including: Inputting the source domain coding feature and the target domain coding feature into a domain discriminator to generate a domain coding feature, the domain discriminator calculating a first-order statistic and a second-order statistic of the domain coding feature, constructing a feature distribution representation based on the first-order statistic and the second-order statistic to obtain a first-order statistic feature distribution representation and a second-order statistic feature distribution representation, and using Wasserstein distance to calculate the difference between the first-order statistic feature distribution representation and the second-order statistic feature distribution representation; Constructing a domain adversarial loss function including an adversarial loss term, a gradient penalty term and a consistency constraint term according to the difference, wherein the gradient penalty term limits the feature gradient norm, the consistency constraint term maintains feature semantic information, and updating the network parameters of the feature encoder through back propagation; The optimized target domain encoding features are input into the authenticity evaluator, which first calculates the semantic similarity between the target features and the source domain verification samples. The semantic similarity is based on the cosine metric function to calculate the distance between feature vectors, while considering the temporal correlation and spatial proximity of the features; An attention weight matrix is ​​constructed based on the semantic similarity, and the source domain verification samples are weighted and aggregated to generate a comprehensive evaluation feature of the target event, wherein the comprehensive evaluation feature includes event attribute features, spatiotemporal correlation features, and context features; The comprehensive evaluation features are input into a multi-layer perceptron, each layer of the multi-layer perceptron learns evaluation rules at different abstract levels respectively, and the final output layer maps the features into event credibility scores through a normalized exponential function.

[0014] Inputting the source domain coding feature and the target domain coding feature into a domain discriminator to generate a domain coding feature, the domain discriminator calculating the first-order statistics and the second-order statistics of the domain coding feature, constructing a feature distribution representation based on the first-order statistics and the second-order statistics to obtain a first-order statistic feature distribution representation and a second-order statistic feature distribution representation, and using the Wasserstein distance to calculate the difference between the first-order statistic feature distribution representation and the second-order statistic feature distribution representation includes: Inputting the source domain encoding feature and the target domain encoding feature into a feature fusion layer of a domain discriminator, wherein the feature fusion layer uses an attention mechanism to calculate a feature weight matrix, and performs weighted fusion on the source domain encoding feature and the target domain encoding feature based on the feature weight matrix to generate a domain encoding feature; Calculating a weighted mean vector and a median vector of the domain coding feature as first-order statistics, wherein the weighted mean vector is obtained by performing sample weighted averaging on each dimension of the domain coding feature, and the median vector is obtained by performing sorting statistics on each dimension of the domain coding feature; Calculating a covariance matrix and a correlation coefficient matrix of the domain coding features as second-order statistics, wherein the covariance matrix represents the relationship between different dimensions of the domain coding features, and the correlation coefficient matrix is ​​obtained by normalizing the covariance matrix; Combining the weighted mean vector and the median vector to construct a first-order statistical feature distribution representation, and combining the eigenvalues ​​and eigenvectors of the covariance matrix to construct a second-order statistical feature distribution representation; The Wasserstein distance represented by the characteristic distribution of the first-order statistics is calculated by using the Sinkhorn algorithm, and the Wasserstein distance represented by the characteristic distribution of the second-order statistics is calculated by using the matrix decomposition technique; The optimal weight coefficient is determined based on cross-validation, and the Wasserstein distance represented by the first-order statistic feature distribution and the Wasserstein distance represented by the second-order statistic feature distribution are weightedly combined to generate the inter-domain distribution difference.

[0015] According to a second aspect of the embodiments of the present invention, Provide a news event spatial verification and analysis system based on satellite remote sensing images, including: The first unit is used to receive news event text information, extract event timestamp, geospatial coordinates, and event type identifier from the news event text information using a natural language processing model; determine a target observation area based on the extracted geospatial coordinates, and calculate an optimal satellite observation time window according to the event timestamp; input the target observation area and the optimal satellite observation time window into a strategy correction model to generate a satellite cluster collaborative observation strategy, wherein the satellite cluster collaborative observation strategy defines a collaborative observation sequence of high-resolution satellites, high-phase satellites, and hyperspectral satellites; and issue a scheduling instruction according to the satellite cluster collaborative observation strategy to activate the corresponding satellite cluster to perform a collaborative observation task; The second unit is used to receive multi-source remote sensing image data acquired during the satellite group's collaborative observation mission, perform adaptive correction processing on the multi-source remote sensing image data, and generate standardized image data with spatiotemporal registration; select a corresponding target detection network and a scene segmentation network from a deep learning model library according to the event type identifier, perform intelligent analysis on the standardized image data, and extract event-related ground objects and spatiotemporal variation characteristics; input the extracted ground objects and spatiotemporal variation characteristics into a graph neural network, integrate historical verification knowledge to construct an event feature association graph, and obtain a feature association strength value through graph reasoning calculation; The third unit is used to perform multi-dimensional cross-validation on the feature association strength value, the event timestamp, and the geographic spatial coordinates to construct an event authenticity evaluation index; based on the event authenticity evaluation index and in combination with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score; when the event credibility score exceeds a preset benchmark credibility threshold, the authenticity of the news event is confirmed, and the satellite group collaborative observation strategy, the event feature association map, and the event authenticity evaluation index in this verification process are written into the verification knowledge base.

[0016] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0017] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0018] The beneficial effects of this application are as follows: By extracting key information of news events through natural language processing models and formulating satellite cluster collaborative observation strategies based on this information, intelligent scheduling and collaborative observation of multiple types of satellites are realized, the efficiency of spatial verification of news events is improved, and the verification process is made more systematic and automated.

[0019] A deep learning model is used to perform intelligent analysis of multi-source remote sensing image data, and a graph neural network is used to construct an event feature association graph. The precise identification and association analysis of event elements are achieved through the calculation of feature association strength values, greatly improving the accuracy and reliability of news event verification.

[0020] Through multi-dimensional cross-validation and transfer learning methods, a scientific event authenticity assessment system was established, and the key information in the verification process was stored in the knowledge base, forming a sustainably optimized verification mechanism, making the spatial verification of news events more objective and credible. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention is a flowchart of a method for spatial verification and analysis of news events based on satellite remote sensing images according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an event authenticity assessment system according to an embodiment of the present invention; Figure 3 This is a schematic diagram for comparing comprehensive evaluation indicators of different methods in the embodiments of the present invention; Figure 4 It is a structural schematic diagram of a news event spatial verification and analysis system based on satellite remote sensing images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0024] Figure 1 FIG. 1 is a flow chart of a method for spatial verification and analysis of news events based on satellite remote sensing images according to an embodiment of the present invention. Figure 1 As shown, the method includes: Receive news event text information, and use a natural language processing model to extract event timestamps, geospatial coordinates, and event type identifiers from the news event text information; determine a target observation area based on the extracted geospatial coordinates, and calculate an optimal satellite observation time window based on the event timestamp; input the target observation area and the optimal satellite observation time window into a strategy correction model to generate a satellite cluster collaborative observation strategy, wherein the satellite cluster collaborative observation strategy defines a collaborative observation sequence of high-resolution satellites, high-phase satellites, and hyperspectral satellites; issue a scheduling instruction based on the satellite cluster collaborative observation strategy to activate the corresponding satellite cluster to perform a collaborative observation task; Receive multi-source remote sensing image data acquired during the satellite group's collaborative observation mission, perform adaptive correction processing on the multi-source remote sensing image data, and generate standardized image data with spatiotemporal registration; select corresponding target detection networks and scene segmentation networks from a deep learning model library according to the event type identifier, perform intelligent analysis on the standardized image data, and extract event-related ground object elements and spatiotemporal variation characteristics; input the extracted ground object elements and spatiotemporal variation characteristics into a graph neural network, integrate historical verification knowledge to construct an event feature association graph, and obtain feature association strength values ​​through graph reasoning calculations; The feature association strength value is cross-validated with the event timestamp and the geographic spatial coordinates in multiple dimensions to construct an event authenticity evaluation index; based on the event authenticity evaluation index and in combination with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score; when the event credibility score exceeds a preset benchmark credibility threshold, the authenticity of the news event is confirmed, and the satellite cluster collaborative observation strategy, the event feature association map, and the event authenticity evaluation index in this verification process are written into the verification knowledge base.

[0025] In an optional implementation, determining a target observation area based on the extracted geospatial coordinates, calculating an optimal satellite observation time window according to the event timestamp; inputting the target observation area and the optimal satellite observation time window into a strategy correction model, and generating a satellite cluster collaborative observation strategy includes: Receive the geospatial coordinates of a news event, use the geospatial coordinates of the event as a seed point, iteratively expand the target area based on an adaptive region growing algorithm, obtain the geographic information entropy value and the spatial correlation in each iterative calculation process, determine the boundary of the target observation area when the geographic information entropy value and the spatial correlation meet a preset expansion threshold, and assign an observation priority weight to each sub-area within the boundary of the target observation area, so as to generate a weighted target observation area set; Acquire the event timestamp and the weighted target observation area set, construct the atmospheric visibility function, the solar altitude angle function, and the satellite observation angle function as an observation suitability evaluation function in the time dimension, perform time optimization on the weighted target observation area set based on the observation suitability evaluation function, determine differentiated observation time windows according to the observation priority weights of each sub-area, and generate the optimal observation time series for each sub-area; A satellite observation capability evaluation matrix is ​​established based on the optimal observation time sequence of the sub-region, wherein each element of the satellite observation capability evaluation matrix represents the observation capability score of a single satellite for a specific sub-region within a specified time window, a multi-dimensional optimization objective function is constructed based on the observation capability score and satellite mission time, energy consumption, and data transmission delay, and an initial satellite group collaborative observation strategy is obtained through iterative optimization and solving, wherein the initial satellite group collaborative observation strategy includes the observation timing, attitude adjustment sequence, and data transmission path of each satellite; The real-time status of the satellites during the execution of the initial satellite cluster collaborative observation strategy is monitored to obtain satellite position deviation values, observation condition change values, and energy status change values. When any change value exceeds a preset fluctuation threshold, the deviation between any change value and the original observation parameter is input into a strategy correction model. The strategy correction model dynamically optimizes the observation timing of the satellite, the attitude adjustment sequence, and the data transmission path based on the observation priority weights of each sub-area to generate an updated satellite cluster collaborative observation strategy.

[0026] The process of determining the target observation area based on geographic spatial coordinates first receives the latitude and longitude coordinate information of the event and uses the coordinates as the initial seed point. The adaptive region growing algorithm is used to expand the region, and the geographic information entropy and spatial correlation are calculated during the expansion process. The geographic information entropy is measured by analyzing the complexity of geographical elements such as terrain features and distribution of objects in the target area, and the spatial correlation is quantified based on the spatial dependence between geographical elements. When these two indicators reach the preset threshold (for example, the geographic information entropy is greater than 0.8 and the spatial correlation is greater than 0.7), the boundary range of the target observation area can be determined.

[0027] The determined target area is divided into sub-regions, and observation weights are assigned to each sub-region based on the regional importance assessment. The assessment factors include population density, economic activity intensity, environmental sensitivity, etc., and the weight value ranges from 0 to 1. For example, the weight of economically intensive areas can be set to 0.9, the weight of environmentally sensitive areas can be set to 0.8, and the weight of general areas can be set to 0.5.

[0028] The calculation of the optimal observation time window first constructs an observation suitability evaluation system. The atmospheric visibility function takes into account factors such as atmospheric transparency and cloud cover. The highest score is when visibility is greater than 10 kilometers and cloud cover is less than 30%. The solar altitude angle function evaluates lighting conditions based on the solar incidence angle, and the angle is most suitable for observation in the range of 45-75 degrees. The satellite observation angle function evaluates the imaging quality based on the angle between the satellite and the target, and the highest score is obtained when observing vertically.

[0029] The above three evaluation functions are combined to form the observation suitability index of the time dimension, and the weight of each sub-region is combined for optimization calculation. High-weight regions are preferentially arranged in the time window with the best observation conditions, and low-weight regions can be arranged in the suboptimal time window, and finally the optimal observation time series of each sub-region is generated.

[0030] The construction of the satellite observation capability evaluation matrix needs to consider the satellite's spatial resolution, spectral resolution, temporal resolution and other technical indicators. For example, the observation capability score of a satellite for a specific sub-region can be expressed as follows: 10 points for a spatial resolution of 2 meters, 8 points for a spectral channel number greater than 4, and 7 points for a revisit period of less than 3 days. The final score is obtained by weighted average of the scores.

[0031] In the process of strategy optimization, the mission time is calculated based on the satellite transit time, the energy consumption is evaluated based on the attitude maneuver and data transmission power consumption, and the data transmission delay takes into account the visible time and data volume of the ground station. The optimal solution that meets the multi-dimensional constraints is obtained through iterative calculation to form the initial observation strategy.

[0032] The strategy correction process monitors satellite status changes in real time. When the position deviation exceeds 100 meters, or the observation conditions change and cause the score to drop by 20%, or the remaining energy is less than 30%, the strategy correction is triggered. The correction model is based on the deviation value of the real-time status and the original parameters, combined with the sub-region weights to re-optimize the calculation to ensure that the observation tasks in high-weight areas are prioritized.

[0033] The optimal observation area and time window are determined by adaptive regional growth and multidimensional evaluation system, which improves the scientificity and accuracy of the observation plan and realizes accurate coverage and efficient observation of the target area. Based on real-time status monitoring and dynamic strategy correction mechanism, the adaptability of the satellite observation system to uncertain factors is enhanced, the reliable execution and continuity of the observation task are guaranteed, and the stability of the system operation is improved. The regional differentiated observation strategy and multidimensional constraint optimization method are adopted to realize the reasonable scheduling and efficient utilization of satellite resources, reduce the system operation cost, and improve the observation efficiency and timeliness of data acquisition.

[0034] In an optional implementation, the extracted ground features and spatiotemporal change features are input into a graph neural network, historical verification knowledge is integrated to construct an event feature association graph, and the feature association strength values ​​obtained by graph reasoning calculation include: Receiving ground feature elements and spatiotemporal change characteristics, constructing the ground feature elements into feature vectors including category attributes, spatial attributes, and temporal attributes, constructing the spatiotemporal change characteristics into change vectors describing the evolution of ground features, and constructing an initial feature graph using the feature vectors and the change vectors as nodes; Based on the current event type, historical verification cases are extracted from the verification knowledge base to construct a knowledge subgraph, and the cosine similarity and spatial position distance of the feature vectors of each node in the initial feature graph and the nodes in the knowledge subgraph are calculated to generate a node similarity matrix; The node similarity matrix is ​​input into a knowledge transfer mapping model, and the knowledge transfer mapping model projects the verification experience in the knowledge subgraph to the corresponding node of the initial feature graph to generate an enhanced feature graph integrating the verification experience; The enhanced feature graph is input into a graph attention network, which aggregates and updates node features based on attention coefficients between nodes, calculates edge feature values ​​according to the updated node features, uses the edge feature values ​​to adjust the connection relationship between nodes, and outputs a feature association graph; for each pair of connected nodes in the feature association graph, the node features and edge features are input into a multilayer perceptron to calculate feature association strength values.

[0035] This paper introduces the processing methods of land features and spatiotemporal change characteristics. For the input land features, it is necessary to construct a feature vector containing three dimensions: category, space, and time series. Category attributes include information such as the type identification and physical characteristics of the land features; spatial attributes include geographic coordinates, area range, shape characteristics, etc.; and time series attributes record the state information of the land features at different time points. For example, for a certain building, its feature vector may contain information such as: the building type is residential, the location coordinates are 116 degrees east longitude and 40 degrees north latitude, the building area is 800 square meters, and the construction time is 2010. The spatiotemporal change characteristics describe the evolution process of the land features, such as the area expansion rate, the height growth trend, etc. These change characteristics are also converted into vector form.

[0036] Construct the initial feature graph. The above feature vectors and change vectors are used as nodes in the graph, and connections are established between nodes based on spatial proximity and change correlation. For example, an edge connection is established between the nodes of two adjacent buildings, indicating that they may be related.

[0037] Extract historical cases from the verification knowledge base. For the event type of current concern, select similar historical verification cases, extract the features and change patterns of the objects in them to build a knowledge subgraph. Calculate the similarity between each node in the initial feature graph and the knowledge subgraph node, including the cosine similarity of the feature vector and the distance of the actual geographical location, and generate a similarity matrix between nodes.

[0038] In the knowledge transfer mapping phase, the similarity matrix is ​​input into the pre-trained mapping model. The model can project the verified experience in the knowledge subgraph to the corresponding node of the initial feature graph. For example, if the change pattern of a certain type of building in the historical case is of great value to the event judgment, this experience will be transferred to the current building node with similar features.

[0039] In the graph attention network processing stage, the network aggregates features based on the attention coefficients between nodes. The attention coefficients reflect the importance weights between different nodes. The network updates node features through multiple rounds of iterations and calculates edge feature values ​​based on them. The edge feature values ​​are used to adjust the connection relationship between nodes, and finally output a feature association graph that reflects the association relationship between ground features.

[0040] For each pair of connected nodes in the feature association graph, their node features and edge features are input into the multilayer perceptron network to calculate the feature association strength value that represents the degree of association. The larger the strength value, the closer the association between the two ground features.

[0041] By constructing multidimensional feature vectors and change vectors, the temporal and spatial characteristics of land features are fully expressed, the completeness and accuracy of feature expression are improved, and a reliable data basis is provided for subsequent analysis. The integration of historical verification knowledge for feature association analysis makes full use of existing experience, improves the reliability and persuasiveness of the analysis results, and avoids the one-sidedness that may arise from relying solely on current data. The use of graph neural networks for feature association analysis can automatically learn the complex relationships between nodes, highlight important features through the attention mechanism, and improve the accuracy and interpretability of association analysis.

[0042] In an optional implementation, the feature association strength value is cross-validated with the event timestamp and the geographic space coordinates in multiple dimensions to construct an event authenticity evaluation index; based on the event authenticity evaluation index, combined with historical cases in the verification knowledge base, the event credibility score is calculated using a transfer learning method, including: Calculating the rationality score of the event occurrence sequence based on the event timestamp, calculating the relevance score of the geographic location based on the geographic spatial coordinates, calculating the feature consistency score based on the feature association strength value, and fusing the rationality score of the sequence, the relevance score of the geographic location and the feature consistency score to construct an event authenticity evaluation index; Retrieving historical verification cases from a verification knowledge base, constructing the historical verification cases into a source domain feature set, and constructing the current event to be evaluated into a target domain feature set; inputting the source domain feature set and the target domain feature set into a feature encoder for feature mapping, and the feature encoder maps the features to a common feature space to generate source domain encoding features and target domain encoding features; The source domain encoding features and the target domain encoding features are input into a domain discriminator, the domain discriminator calculates the difference between the distribution of source domain features and target domain features, and optimizes the feature encoder based on the difference to make the distribution of the source domain features and the target domain features tend to be consistent; the optimized target domain encoding features are input into an authenticity evaluator, and the authenticity evaluator calculates the event credibility score based on the source domain verification experience.

[0043] This embodiment provides a method for event authenticity assessment and credibility calculation, which achieves accurate assessment of events through multi-dimensional cross-validation and transfer learning.

[0044] Calculate the plausibility score of the event sequence based on the event timestamp. In this step, the system analyzes whether the time of the event is in a logical order. For example, if an event claims to have occurred before a historical event, but the timestamp shows that it actually occurred after, then the temporal plausibility score will be low. The system may use a time window to evaluate the plausibility of events. For example, given a 30-day window, if the event occurs within 15 days before or after the expected time, a higher plausibility score will be given.

[0045] The relevance score for the location is calculated based on the geospatial coordinates. The system checks whether the location where the event occurred matches the location claimed. For example, if an event claims to have occurred in a desert area, but the geo-coordinates show it was located in the ocean, the geo-relevance score will be low. The system may use Geographic Information System (GIS) data to verify the plausibility of the location, such as checking whether the coordinates fall within the expected geographic area.

[0046] A feature consistency score is calculated based on the feature association strength values. This step analyzes whether there are contradictions or inconsistencies between the various features of the event. For example, if an event claims to be a large-scale rally, but there is little related social media activity, the feature consistency score will be low. The system may use predefined feature weights to calculate the overall consistency, such as giving a weight of 0.3 to social media activity, a weight of 0.5 to photos of the scene, and a weight of 0.2 to eyewitness testimony.

[0047] The temporal rationality score, geographic location relevance score and feature consistency score are integrated to construct an event authenticity evaluation index. This integration process may adopt a weighted average method, for example, a weight of 0.3 is given to temporal rationality, a weight of 0.3 is given to geographic location relevance, and a weight of 0.4 is given to feature consistency, thereby obtaining a comprehensive event authenticity evaluation index.

[0048] The system retrieves historical verification cases from the verification knowledge base and constructs these cases as the source domain feature set, while constructing the current event to be evaluated as the target domain feature set. For example, if the current event is about a natural disaster, the system will retrieve similar natural disaster events in the past from the knowledge base as the source domain feature set.

[0049] The source domain feature set and the target domain feature set are input into the feature encoder for feature mapping. The feature encoder may use a deep neural network structure, such as a multi-layer perceptron or a convolutional neural network, to map the input features to a common feature space and generate source domain encoding features and target domain encoding features. This process can be understood as converting features from different fields into a common representation.

[0050] The source domain encoded features and the target domain encoded features are fed into a domain discriminator. The role of the domain discriminator is to calculate the difference between the source domain features and the target domain features. This may be achieved by calculating the distance between the two distributions, such as using KL divergence or Wasserstein distance. Based on the calculated difference, the system will optimize the feature encoder with the goal of making the source domain features and target domain features more consistent. This optimization process may involve adjusting the parameters of the feature encoder so that it can generate more similar source domain and target domain encoded features.

[0051] The optimized target domain encoded features are fed into the authenticity evaluator. The authenticity evaluator calculates a credibility score for the event based on verified experience from the source domain. This process may involve using a machine learning model, such as a support vector machine or random forest, which is trained on the source domain data and then applied to the target domain data. For example, if certain combinations of features are often associated with high credibility in the source domain, then when similar combinations of features are observed in the target domain, the system will also give a high credibility score.

[0052] This method improves the accuracy and comprehensiveness of event authenticity assessment through multi-dimensional cross-validation, comprehensively considers multiple aspects such as time, geographic location and feature consistency, and effectively reduces the bias and errors that may be caused by single-dimensional evaluation. The transfer learning method is used to achieve effective migration from historical experience to new event evaluation, overcome the evaluation difficulties that may arise in traditional methods when facing new or rare events, and improve the system's adaptability and evaluation accuracy for various types of events. Through the optimization process of feature encoders and domain discriminators, the alignment of feature distributions in the source domain and target domain is achieved, effectively reducing the impact of domain differences on the evaluation results and improving the reliability and generalization ability of the evaluation results.

[0053] In an optional implementation, the rationality score of the event occurrence sequence is calculated based on the timestamp, the relevance score of the geographic location is calculated based on the geographic spatial coordinates, and the feature consistency score is calculated based on the feature association strength value. The rationality score, the relevance score and the feature consistency score are integrated to construct an event authenticity evaluation index, including: A time reference sequence model is constructed based on the timestamp, the time reference sequence model uses a Gaussian kernel function with adaptive weights to calculate the time difference between the current event timestamp and the time point of the historical associated event, calculates the state transition probability matrix of the event sequence through a time Markov chain model, and generates a time series rationality score based on the time difference and the state transition probability matrix; The temporal rationality score is used as a constraint condition and combined with the geographic spatial coordinates to construct a spatial feature description model, wherein the spatial feature description model uses Hausdorff distance to calculate a spatial distance matrix between event locations and a set of geographic reference points, extracts spatial distribution features based on the spatial distance matrix, calculates the matching degree between the spatial distribution features and the spatial constraint rules in the geographic information system, and generates a geographic location relevance score; A feature verification model is constructed based on the temporal rationality score and the geographic location correlation score, the feature verification model normalizes the feature correlation strength value, uses a multi-layer perceptron to extract the deep spatiotemporal correlation pattern between features, and performs consistency verification between the deep spatiotemporal correlation pattern and a preset feature constraint rule set to generate a feature correlation consistency score; An adaptive weight allocation network including a feature extraction layer, an attention mechanism layer and a normalized output layer is constructed, and the rationality score of the time series, the correlation score of the geographic location and the feature correlation consistency score are input into the adaptive weight allocation network. The adaptive weight allocation network generates a dynamic fusion weight based on the event type feature vector, and the three scores are weightedly fused based on the dynamic fusion weight to generate an event authenticity evaluation index.

[0054] Evaluate the rationality of event timing. By constructing a time reference sequence model, the model uses a Gaussian kernel function with adaptive weights to process time information. In specific implementation, for the timestamp of the current event, extract the time point information of historical related events and calculate the time difference. For example, for a commercial activity event that occurred in a business district in Beijing, its timestamp was 3 o'clock in the afternoon of a certain day. By querying historical data, it was found that the business activities of the business district were usually concentrated between 2 o'clock and 5 o'clock in the afternoon. The time difference calculated by the Gaussian kernel function was 0.85, indicating that the time point has a high rationality. Then, the state transition probability matrix was constructed using the time Markov chain model to analyze the temporal law of the event sequence. Based on the time difference and state transition probability, the final temporal rationality score was 0.9.

[0055] Conduct geographic location relevance assessment. Take the temporal rationality score obtained above as a constraint condition, and build a spatial feature description model in combination with the event's geographic spatial coordinate information. This model uses the Hausdorff distance to calculate the spatial distance matrix between the event location and the geographic reference point. Taking the above-mentioned commercial activities as an example, the spatial distance matrix is ​​obtained by calculating the distance between the location of the event and the main commercial facilities in the business district. According to the matrix, spatial distribution characteristics such as location density and degree of aggregation are extracted. The matching degree of these characteristics is calculated with the spatial constraint rules in the geographic information system, such as considering factors such as building distribution and road network, and finally a geographic location relevance score of 0.85 is generated.

[0056] Construct a feature verification model to evaluate the consistency of feature association. The model first normalizes the feature association strength value and unifies the feature values ​​of different dimensions into the range of zero to one. Then use a multi-layer perceptron to extract the deep spatiotemporal association pattern between features. In this example, the features of business activities such as traffic, consumption level, and surrounding facilities are extracted, and the association patterns of these features with time and location are analyzed by a multi-layer perceptron. The extracted association pattern is verified for consistency with the preset feature constraint rule set, and a feature association consistency score of 0.88 is generated.

[0057] An adaptive weight distribution network is constructed for score fusion. The network includes a feature extraction layer, an attention mechanism layer, and a normalized output layer. The above three scores are input into the network, and dynamic fusion weights are generated based on the event type feature vector. For commercial event events, the weights of temporal rationality, geographic location relevance, and feature consistency are calculated based on their feature vectors to be 0.3, 0.35, and 0.35, respectively. The final event authenticity evaluation index obtained through weighted fusion is 0.87.

[0058] Figure 2 This is a schematic diagram of the authenticity assessment analysis of an event in an embodiment of the present invention: This figure shows a complete event authenticity assessment and analysis platform. The top of the interface displays the title "Event Authenticity Assessment and Analysis" and the event ID (EV-20250407-001), and the main part below consists of three parallel analysis cards and a comprehensive assessment area at the bottom. The card on the left shows the temporal rationality analysis, which visually presents the five key time points from 08:15 to 14:05 and their Markov transition probabilities (0.87, 0.75, 0.62, and 0.91) through a timeline graph. The Gaussian kernel time difference is 0.23, and the final temporal rationality score is 0.79; the middle card shows the geographic location correlation analysis, and the map marks the current event location with multiple reference points and their Hausdorff distances (42.8) and spatial constraint areas. The spatial constraint matching degree is 0.83, and the geographic location correlation score is 0.83; the card on the right shows the feature association consistency analysis, which presents feature normalization (7 feature bar charts), deep association patterns (multi-layer perceptron network structure diagram) and consistency verification (5 constraint rule tables, with consistency values ​​ranging from 0.68 to 0.92) in the form of tabs, and the feature association consistency score is 0.82. The comprehensive evaluation area at the bottom shows the final score of 0.82 in a pie chart, and lists three sub-indicators (time series 0.79, geography 0.83, and features 0.82) and their corresponding weights (0.32, 0.36, and 0.32). The entire interface uses progress bars, charts, and data visualization to fully display the multi-dimensional evaluation system for event authenticity based on an adaptive weight distribution network, providing users with an intuitive and scientific basis for determining event authenticity.

[0059] By analyzing the time series features with adaptive weighted Gaussian kernel functions and Markov chain models, we can achieve an accurate assessment of the time series of events, improve the accuracy of the judgment of the rationality of the time series, and make the assessment results more objective and reliable. By using the spatial feature description model and Hausdorff distance calculation method, combined with the constraint rules of the geographic information system, we can achieve a comprehensive assessment of the correlation between the geographical locations of events, and improve the reliability and adaptability of spatial correlation analysis. Based on the fusion method of multi-layer perceptron and adaptive weight allocation network, we can achieve deep mining and dynamic weight allocation of multi-dimensional features, improve the accuracy of feature correlation analysis, and make the authenticity assessment results more comprehensive and reliable.

[0060] In an optional implementation, the source domain encoding feature and the target domain encoding feature are input into a domain discriminator, the domain discriminator calculates the difference between the source domain feature and the target domain feature distribution, and optimizes the feature encoder based on the difference so that the source domain and the target domain feature distribution tend to be consistent; the optimized target domain encoding feature is input into an authenticity evaluator, and the authenticity evaluator calculates the event credibility score based on the source domain verification experience, including: Inputting the source domain coding feature and the target domain coding feature into a domain discriminator to generate a domain coding feature, the domain discriminator calculating a first-order statistic and a second-order statistic of the domain coding feature, constructing a feature distribution representation based on the first-order statistic and the second-order statistic to obtain a first-order statistic feature distribution representation and a second-order statistic feature distribution representation, and using Wasserstein distance to calculate the difference between the first-order statistic feature distribution representation and the second-order statistic feature distribution representation; Constructing a domain adversarial loss function including an adversarial loss term, a gradient penalty term and a consistency constraint term according to the difference, wherein the gradient penalty term limits the feature gradient norm, the consistency constraint term maintains feature semantic information, and updating the network parameters of the feature encoder through back propagation; The optimized target domain encoding features are input into the authenticity evaluator, which first calculates the semantic similarity between the target features and the source domain verification samples. The semantic similarity is based on the cosine metric function to calculate the distance between feature vectors, while considering the temporal correlation and spatial proximity of the features; An attention weight matrix is ​​constructed based on the semantic similarity, and the source domain verification samples are weighted and aggregated to generate a comprehensive evaluation feature of the target event, wherein the comprehensive evaluation feature includes event attribute features, spatiotemporal correlation features, and context features; The comprehensive evaluation features are input into a multi-layer perceptron, each layer of the multi-layer perceptron learns evaluation rules at different abstract levels respectively, and the final output layer maps the features into event credibility scores through a normalized exponential function.

[0061] The feature encoder encodes the source domain data and the target domain data respectively to generate the corresponding feature vectors. The source domain data comes from the verified real event data set, and the target domain data comes from the event data to be verified. The feature encoding adopts a deep neural network structure, which includes multiple convolutional layers and fully connected layers to extract the deep semantic features of the data.

[0062] The domain discriminator receives the encoded features of the source domain and the target domain as input. The domain discriminator first calculates the first-order statistics of the feature, that is, the mean of the feature vector; at the same time, it calculates the second-order statistics, that is, the covariance matrix of the feature vector. Based on these statistics, a feature distribution representation is constructed. In specific implementation, the feature vector can be divided into multiple subspaces, local statistics are calculated separately, and then they are combined to form a global distribution representation. For example, for a 128-dimensional feature vector, it can be divided into 8 16-dimensional subspaces.

[0063] The domain discriminator uses the Wasserstein distance to measure the difference between the source domain and the target domain. The greater the difference, the more significant the difference in the feature distribution of the two domains. Based on the calculated difference, a domain adversarial loss function is constructed. The loss function consists of three parts: the adversarial loss term is used to minimize the difference between domains; the gradient penalty term ensures the stability of the model by limiting the feature gradient norm; and the consistency constraint term keeps the semantic information of the feature intact. The parameters of the feature encoder are optimized through the back-propagation algorithm so that the feature distribution of the source domain and the target domain gradually tend to be consistent.

[0064] After domain adversarial training, the optimized target domain encoding features are input into the authenticity evaluator. The evaluator first calculates the semantic similarity between the target features and the source domain verification samples. The similarity calculation is based on the cosine metric function, taking into account both the temporal correlation and spatial proximity of the features. For example, for temporal correlation, the time window can be set to 24 hours, and only samples within this window can be considered; for spatial proximity, the geographical distance threshold can be set to 10 kilometers.

[0065] Based on the calculated semantic similarity, an attention weight matrix is ​​constructed. The weight matrix is ​​used to perform weighted aggregation on the source domain verification samples to generate comprehensive evaluation features of the target event. The comprehensive evaluation features include three aspects: event attribute features describe the basic features of the event; spatiotemporal correlation features describe the spatiotemporal laws of the event; and context features reflect the relationship between the event and other related events.

[0066] The comprehensive evaluation features are input into the multi-layer perceptron for credibility evaluation. The multi-layer perceptron contains multiple hidden layers, and each layer learns evaluation rules at different abstract levels. For example, the first layer learns the evaluation rules of basic attribute features, the second layer learns the evaluation rules of spatiotemporal features, and the third layer learns the evaluation rules of contextual features. The final output layer maps the features to event credibility scores between 0 and 1 through a normalized exponential function.

[0067] Figure 3 This is a schematic diagram of the comparison of comprehensive evaluation indicators of different methods in the embodiments of the present invention: This multi-indicator radar chart shows the performance comparison of four different technical methods on five key evaluation indicators: accuracy, F1 score, precision, recall rate and feature matching. In the figure, the cross lines represent the "present technical solution", which has achieved the best results in all five dimensions, with values ​​of 0.95 accuracy, 0.97 F1 score, 0.94 precision, 0.94 recall rate and 0.96 feature matching, forming the outermost polygon; the triangular lines represent the "DANN method", which is the second-best solution, with values ​​of 0.85 accuracy, 0.86 F1 score, 0.83 precision, 0.85 recall rate and 0.84 feature matching; the circular lines represent the "first-order statistics only" method, with values ​​of 0.80 accuracy, 0.79 F1 score, 0.77 precision, 0.76 recall rate and 0.81 feature matching; the square lines represent the "traditional GAN ​​method", which has the weakest comprehensive performance among the four methods, with values ​​of 0.75 accuracy, 0.74 F1 score, 0.77 precision, 0.73 recall rate and 0.71 feature matching. This technical solution maintains a high level of over 0.94 in all indicators, especially in F1 score and feature matching, which is generally 15-20 percentage points higher than other methods, demonstrating its excellent performance and robustness in authenticity assessment tasks.

[0068] Through domain adversarial training, the feature distribution of the source domain and the target domain is aligned, which effectively reduces the negative impact of domain shift and improves the generalization performance of the model in the target domain. At the same time, the multi-level statistical feature distribution representation can more comprehensively characterize the data distribution characteristics and improve the domain adaptation effect. The semantic similarity calculation method that introduces time correlation and spatial proximity constraints can more accurately identify similar events and avoid interference from irrelevant events that are far away or have a large time span. The introduction of the attention mechanism realizes the adaptive weighting of source domain verification samples and highlights the contribution of important samples. The hierarchical evaluation mechanism of the multi-layer perceptron can learn evaluation rules from different levels of abstraction to achieve multi-angle and multi-level evaluation of event credibility. Comprehensive consideration of event attributes, spatiotemporal associations, and contextual features ensures the comprehensiveness and reliability of the evaluation results.

[0069] In an optional implementation, the source domain coding feature and the target domain coding feature are input into a domain discriminator to generate a domain coding feature, the domain discriminator calculates a first-order statistic and a second-order statistic of the domain coding feature, constructs a feature distribution representation based on the first-order statistic and the second-order statistic, obtains a first-order statistic feature distribution representation and a second-order statistic feature distribution representation, and uses the Wasserstein distance to calculate the difference between the first-order statistic feature distribution representation and the second-order statistic feature distribution representation, including: Inputting the source domain encoding feature and the target domain encoding feature into a feature fusion layer of a domain discriminator, wherein the feature fusion layer uses an attention mechanism to calculate a feature weight matrix, and performs weighted fusion on the source domain encoding feature and the target domain encoding feature based on the feature weight matrix to generate a domain encoding feature; Calculating a weighted mean vector and a median vector of the domain coding feature as first-order statistics, wherein the weighted mean vector is obtained by performing sample weighted averaging on each dimension of the domain coding feature, and the median vector is obtained by performing sorting statistics on each dimension of the domain coding feature; Calculating a covariance matrix and a correlation coefficient matrix of the domain coding features as second-order statistics, wherein the covariance matrix represents the relationship between different dimensions of the domain coding features, and the correlation coefficient matrix is ​​obtained by normalizing the covariance matrix; Combining the weighted mean vector and the median vector to construct a first-order statistical feature distribution representation, and combining the eigenvalues ​​and eigenvectors of the covariance matrix to construct a second-order statistical feature distribution representation; The Wasserstein distance represented by the characteristic distribution of the first-order statistics is calculated by using the Sinkhorn algorithm, and the Wasserstein distance represented by the characteristic distribution of the second-order statistics is calculated by using the matrix decomposition technique; The optimal weight coefficient is determined based on cross-validation, and the Wasserstein distance represented by the first-order statistic feature distribution and the Wasserstein distance represented by the second-order statistic feature distribution are weightedly combined to generate the inter-domain distribution difference.

[0070] In the process of domain adaptation, the source domain encoding features and the target domain encoding features are first obtained. These two features are processed by the feature fusion layer, which uses the attention mechanism to calculate the feature weights. Specifically, the attention score is calculated for each feature vector, and the attention score reflects the importance of the feature. For example, for a 128-dimensional feature vector, the corresponding 128 weight values ​​are calculated, and the weight value range is between 0 and 1. These weight values ​​are combined into a feature weight matrix for subsequent feature fusion.

[0071] When fusion is performed, the feature vectors of the source domain and the target domain are multiplied by the corresponding weight values, and the weighted sum is performed to obtain the fused domain encoding features. To illustrate with a specific example, assuming that the source domain feature vector is [0.3, 0.5, 0.7], the target domain feature vector is [0.4, 0.6, 0.8], and the corresponding weight values ​​are [0.4, 0.3, 0.3], then the fused feature vector is [0.35, 0.55, 0.75].

[0072] Calculate the first-order statistics of the domain encoding feature. First, calculate the weighted mean vector, and perform weighted average on each dimension of the feature. At the same time, calculate the median vector, sort the values ​​of each dimension of the feature, and take the value in the middle as the median.

[0073] Calculate the second-order statistics. When calculating the covariance matrix, analyze the correlation between different dimensions of the features. For the feature value sequence of two dimensions, the covariance is obtained by calculating the mean of their deviation products. The covariance matrix is ​​standardized to obtain the correlation coefficient matrix, in which the values ​​range from -1 to 1.

[0074] When constructing the feature distribution representation, the weighted mean vector and the median vector are concatenated to form the first-order statistic feature distribution representation. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors, which are combined to form the second-order statistic feature distribution representation.

[0075] When calculating the distribution difference, the Sinkhorn iterative algorithm is used to calculate the Wasserstein distance of the first-order statistic. Through iterative optimization, the optimal transmission scheme between the two distributions is found. For the second-order statistics, the matrix decomposition method is used to calculate the Wasserstein distance. Finally, the weight coefficient is determined by cross-validation, and the two distances are weighted and combined to obtain the final inter-domain distribution difference.

[0076] The feature weight matrix is ​​calculated through the attention mechanism, which realizes the adaptive fusion of source domain and target domain features and improves the accuracy and robustness of feature representation. The importance of different features is taken into account in the feature fusion process, making the fusion result more reasonable. At the same time, the first-order statistics and second-order statistics are used to construct the feature distribution representation, which comprehensively characterizes the characteristics of data distribution. The weighted mean and median reflect the central tendency of the data, and the covariance matrix and correlation coefficient matrix reflect the relationship between the features, making the distribution representation more complete. The Wasserstein distance is used to measure the distribution difference, and the optimal weight coefficient is determined by cross-validation, making the calculation of the inter-domain difference more accurate and reliable. This method can effectively capture the distribution difference between the source domain and the target domain, providing an important basis for subsequent domain adaptation tasks.

[0077] The transformation of satellite news topics is the first step in the production process. The transformation process requires in-depth analysis and decomposition of the topics, breaking down the news topics into multiple specific sub-topics. Taking the report on the changes in the urban ecological environment as an example, it can be broken down into multiple sub-topics such as urban expansion, vegetation cover changes, and water body changes. There is an inherent connection between the sub-topics, and through analysis of different dimensions, a complete news reporting framework is jointly constructed.

[0078] A detailed data demand analysis is required for the decomposed sub-topics. Data demand includes multiple dimensions such as time scale, spatial scale, data type and observation indicators. The time scale determines the required observation period and time resolution; the spatial scale clarifies the observation area and spatial resolution requirements; the data type includes different sensor data such as optical and radar; and the observation indicators involve professional parameters such as vegetation index and surface temperature. For example, when monitoring the progress of large-scale infrastructure construction, it is necessary to select high-resolution optical satellite data with a spatial resolution better than 2 meters. A combination of 0.5-meter resolution panchromatic and 2-meter resolution multispectral satellite data can be used.

[0079] The selection of data sources needs to be evaluated through feasibility analysis. The evaluation content includes the completeness of data coverage, whether the temporal resolution meets the needs of dynamic monitoring, whether the spatial resolution is sufficient to identify target features, the influence of interference factors such as cloud cover, and the cost of data acquisition and processing. Only after a comprehensive evaluation to ensure that the data source meets the reporting needs can we proceed to the next stage.

[0080] In the data acquisition and preprocessing stage, the data must first be downloaded according to the determined time range and spatial range. The downloaded data must be quality checked, and data with excessive cloud cover and unqualified quality must be eliminated, and a standardized data management directory must be established. Subsequently, radiation correction and atmospheric correction are performed to eliminate the sensor system errors and the influence of the atmosphere on the reflection characteristics of the ground objects, and the data is converted into surface reflectivity. In addition, geometric correction and registration are required to eliminate the geometric distortion caused by terrain undulations and satellite attitude changes, unify data of different phases into the same coordinate system, and realize accurate registration of multi-source data. Finally, through data mosaicking and cropping, a data set with complete coverage of the study area is generated.

[0081] Satellite data analysis and visualization are the core links of news production. Through feature extraction and classification, the types of objects and change information in the target area can be identified. Change detection is performed based on multi-temporal images, the area and change rate of the changed area are calculated, and the change trends and laws are analyzed. In terms of visual expression, it is necessary to select appropriate forms of expression according to the content of the news, such as rolling shutter contrast, dynamic images, etc., and design reasonable color schemes and legends, and add necessary geographical elements and thematic information. Finally, satellite data is integrated with multi-source data such as ground data and statistical information to form a comprehensive visual news product.

[0082] Figure 4 FIG. 1 is a schematic diagram of the structure of a news event spatial verification and analysis system based on satellite remote sensing images according to an embodiment of the present invention. Figure 4 As shown, the system comprises: The first unit is used to receive news event text information, extract event timestamp, geospatial coordinates, and event type identifier from the news event text information using a natural language processing model; determine a target observation area based on the extracted geospatial coordinates, and calculate an optimal satellite observation time window according to the event timestamp; input the target observation area and the optimal satellite observation time window into a strategy correction model to generate a satellite cluster collaborative observation strategy, wherein the satellite cluster collaborative observation strategy defines a collaborative observation sequence of high-resolution satellites, high-phase satellites, and hyperspectral satellites; and issue a scheduling instruction according to the satellite cluster collaborative observation strategy to activate the corresponding satellite cluster to perform a collaborative observation task; The second unit is used to receive multi-source remote sensing image data acquired during the satellite group's collaborative observation mission, perform adaptive correction processing on the multi-source remote sensing image data, and generate standardized image data with spatiotemporal registration; select a corresponding target detection network and a scene segmentation network from a deep learning model library according to the event type identifier, perform intelligent analysis on the standardized image data, and extract event-related ground objects and spatiotemporal variation characteristics; input the extracted ground objects and spatiotemporal variation characteristics into a graph neural network, integrate historical verification knowledge to construct an event feature association graph, and obtain a feature association strength value through graph reasoning calculation; The third unit is used to perform multi-dimensional cross-validation on the feature association strength value, the event timestamp, and the geographic spatial coordinates to construct an event authenticity evaluation index; based on the event authenticity evaluation index and in combination with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score; when the event credibility score exceeds a preset benchmark credibility threshold, the authenticity of the news event is confirmed, and the satellite group collaborative observation strategy, the event feature association map, and the event authenticity evaluation index in this verification process are written into the verification knowledge base.

[0083] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0084] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0085] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spatial verification and analysis method for news events based on satellite remote sensing images, characterized in that: include: Receiving news event text information, and extracting event timestamps, geospatial coordinates, and event type identifiers from the news event text information using a natural language processing model; Determine the target observation area based on the extracted geospatial coordinates, and calculate the optimal satellite observation time window according to the event timestamp; input the target observation area and the optimal satellite observation time window into the strategy correction model to generate a satellite cluster collaborative observation strategy, wherein the satellite cluster collaborative observation strategy defines a collaborative observation sequence of high-resolution satellites, high-phase satellites, and hyperspectral satellites; issue a scheduling instruction according to the satellite cluster collaborative observation strategy to activate the corresponding satellite cluster to perform the collaborative observation task; Receive multi-source remote sensing image data acquired during the satellite group's collaborative observation mission, perform adaptive correction processing on the multi-source remote sensing image data, and generate standardized image data with spatiotemporal registration; select a corresponding target detection network and scene segmentation network from a deep learning model library according to the event type identifier, perform intelligent analysis on the standardized image data, and extract event-related ground features and spatiotemporal change characteristics; The extracted ground features and spatiotemporal change characteristics are input into the graph neural network, and the event feature association graph is constructed by integrating historical verification knowledge. The feature association strength value is obtained through graph reasoning calculation. The feature association strength value is cross-validated with the event timestamp and the geographic space coordinates in multiple dimensions to construct an event authenticity evaluation index; based on the event authenticity evaluation index and in combination with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score; When the event credibility score exceeds the preset benchmark credibility threshold, the authenticity of the news event is confirmed, and the satellite group collaborative observation strategy, the event feature association map, and the event authenticity evaluation index in this verification process are written into the verification knowledge base.

2. The method according to claim 1, characterized in that determining a target observation area based on the extracted geospatial coordinates, and calculating an optimal satellite observation time window according to the event timestamp; Inputting the target observation area and the optimal satellite observation time window into the strategy correction model to generate a satellite cluster collaborative observation strategy includes: Receive the geospatial coordinates of a news event, use the geospatial coordinates of the event as a seed point, iteratively expand the target area based on an adaptive region growing algorithm, obtain the geographic information entropy value and the spatial correlation in each iterative calculation process, determine the boundary of the target observation area when the geographic information entropy value and the spatial correlation meet a preset expansion threshold, and assign an observation priority weight to each sub-area within the boundary of the target observation area, so as to generate a weighted target observation area set; Acquire the event timestamp and the weighted target observation area set, construct the atmospheric visibility function, the solar altitude angle function, and the satellite observation angle function as an observation suitability evaluation function in the time dimension, perform time optimization on the weighted target observation area set based on the observation suitability evaluation function, determine differentiated observation time windows according to the observation priority weights of each sub-area, and generate the optimal observation time series for each sub-area; A satellite observation capability evaluation matrix is ​​established based on the optimal observation time sequence of the sub-region, wherein each element of the satellite observation capability evaluation matrix represents the observation capability score of a single satellite for a specific sub-region within a specified time window, a multi-dimensional optimization objective function is constructed based on the observation capability score and satellite mission time, energy consumption, and data transmission delay, and an initial satellite group collaborative observation strategy is obtained through iterative optimization and solving, wherein the initial satellite group collaborative observation strategy includes the observation timing, attitude adjustment sequence, and data transmission path of each satellite; The real-time status of the satellites during the execution of the initial satellite cluster collaborative observation strategy is monitored to obtain satellite position deviation values, observation condition change values, and energy status change values. When any change value exceeds a preset fluctuation threshold, the deviation between any change value and the original observation parameter is input into a strategy correction model. The strategy correction model dynamically optimizes the observation timing of the satellite, the attitude adjustment sequence, and the data transmission path based on the observation priority weights of each sub-area to generate an updated satellite cluster collaborative observation strategy.

3. The method according to claim 1, characterized in that: The extracted ground features and spatiotemporal change characteristics are input into the graph neural network, and the historical verification knowledge is integrated to construct the event feature association graph. The feature association strength values ​​obtained through graph reasoning calculation include: Receiving ground feature elements and spatiotemporal change features, constructing the ground feature elements into feature vectors including category attributes, spatial attributes, and temporal attributes, constructing the spatiotemporal change features into change vectors describing the evolution of ground features, and constructing an initial feature graph using the feature vectors and the change vectors as nodes; Based on the current event type, historical verification cases are extracted from the verification knowledge base to construct a knowledge subgraph, and the cosine similarity and spatial position distance of the feature vectors of each node in the initial feature graph and the nodes in the knowledge subgraph are calculated to generate a node similarity matrix; The node similarity matrix is ​​input into a knowledge transfer mapping model, and the knowledge transfer mapping model projects the verification experience in the knowledge subgraph to the corresponding node of the initial feature graph to generate an enhanced feature graph integrating the verification experience; The enhanced feature graph is input into a graph attention network, which aggregates and updates node features based on attention coefficients between nodes, calculates edge feature values ​​according to the updated node features, uses the edge feature values ​​to adjust the connection relationship between nodes, and outputs a feature association graph; for each pair of connected nodes in the feature association graph, the node features and edge features are input into a multilayer perceptron to calculate feature association strength values.

4. The method according to claim 1, characterized in that Perform multi-dimensional cross-validation on the feature association strength value, the event timestamp, and the geographic space coordinates to construct an event authenticity assessment index; Based on the event authenticity evaluation indicators and combined with historical cases in the verification knowledge base, the event credibility score is calculated using the transfer learning method, including: Calculating the rationality score of the event occurrence sequence based on the event timestamp, calculating the relevance score of the geographic location based on the geographic spatial coordinates, calculating the feature consistency score based on the feature association strength value, and fusing the rationality score of the sequence, the relevance score of the geographic location and the feature consistency score to construct an event authenticity evaluation index; Retrieving historical verification cases from a verification knowledge base, constructing the historical verification cases into a source domain feature set, and constructing the current event to be evaluated into a target domain feature set; inputting the source domain feature set and the target domain feature set into a feature encoder for feature mapping, and the feature encoder maps the features to a common feature space to generate source domain encoding features and target domain encoding features; The source domain encoding features and the target domain encoding features are input into a domain discriminator, the domain discriminator calculates the difference between the distribution of source domain features and target domain features, and optimizes the feature encoder based on the difference to make the distribution of the source domain features and the target domain features tend to be consistent; the optimized target domain encoding features are input into an authenticity evaluator, and the authenticity evaluator calculates the event credibility score based on the source domain verification experience.

5. The method according to claim 4, characterized in that Calculating the rationality score of the event occurrence sequence based on the timestamp, calculating the relevance score of the geographic location based on the geographic space coordinates, calculating the feature consistency score based on the feature association strength value, and fusing the rationality score, the relevance score and the feature consistency score to construct an event authenticity evaluation index includes: A time reference sequence model is constructed based on the timestamp, the time reference sequence model uses a Gaussian kernel function with adaptive weights to calculate the time difference between the current event timestamp and the time point of the historical associated event, calculates the state transition probability matrix of the event sequence through a time Markov chain model, and generates a time series rationality score based on the time difference and the state transition probability matrix; The temporal rationality score is used as a constraint condition and combined with the geographic spatial coordinates to construct a spatial feature description model, wherein the spatial feature description model uses Hausdorff distance to calculate a spatial distance matrix between event locations and a set of geographic reference points, extracts spatial distribution features based on the spatial distance matrix, calculates the matching degree between the spatial distribution features and the spatial constraint rules in the geographic information system, and generates a geographic location relevance score; A feature verification model is constructed based on the temporal rationality score and the geographic location correlation score, the feature verification model normalizes the feature correlation strength value, uses a multi-layer perceptron to extract the deep spatiotemporal correlation pattern between features, and performs consistency verification between the deep spatiotemporal correlation pattern and a preset feature constraint rule set to generate a feature correlation consistency score; An adaptive weight allocation network including a feature extraction layer, an attention mechanism layer and a normalized output layer is constructed, and the rationality score of the time series, the correlation score of the geographic location and the feature correlation consistency score are input into the adaptive weight allocation network. The adaptive weight allocation network generates a dynamic fusion weight based on the event type feature vector, and the three scores are weightedly fused based on the dynamic fusion weight to generate an event authenticity evaluation index.

6. The method according to claim 4, characterized in that Inputting the source domain encoding features and the target domain encoding features into a domain discriminator, the domain discriminator calculating the difference between the source domain features and the target domain features, and optimizing the feature encoder based on the difference so that the source domain features and the target domain features are distributed in a consistent manner; The optimized target domain encoding features are input into the authenticity evaluator, and the authenticity evaluator calculates the event credibility score based on the source domain verification experience, including: Inputting the source domain coding feature and the target domain coding feature into a domain discriminator to generate a domain coding feature, the domain discriminator calculating a first-order statistic and a second-order statistic of the domain coding feature, constructing a feature distribution representation based on the first-order statistic and the second-order statistic to obtain a first-order statistic feature distribution representation and a second-order statistic feature distribution representation, and using Wasserstein distance to calculate the difference between the first-order statistic feature distribution representation and the second-order statistic feature distribution representation; Constructing a domain adversarial loss function including an adversarial loss term, a gradient penalty term and a consistency constraint term according to the difference, wherein the gradient penalty term limits the feature gradient norm, the consistency constraint term maintains feature semantic information, and updating the network parameters of the feature encoder through back propagation; The optimized target domain encoding features are input into the authenticity evaluator, which first calculates the semantic similarity between the target features and the source domain verification samples. The semantic similarity is based on the cosine metric function to calculate the distance between feature vectors, while considering the temporal correlation and spatial proximity of the features; An attention weight matrix is ​​constructed based on the semantic similarity, and the source domain verification samples are weighted and aggregated to generate a comprehensive evaluation feature of the target event, wherein the comprehensive evaluation feature includes event attribute features, spatiotemporal correlation features, and context features; The comprehensive evaluation features are input into a multi-layer perceptron, each layer of the multi-layer perceptron learns evaluation rules at different abstract levels respectively, and the final output layer maps the features into event credibility scores through a normalized exponential function.

7. The method according to claim 6, characterized in that Inputting the source domain coding feature and the target domain coding feature into a domain discriminator to generate a domain coding feature, the domain discriminator calculating the first-order statistics and the second-order statistics of the domain coding feature, constructing a feature distribution representation based on the first-order statistics and the second-order statistics to obtain a first-order statistic feature distribution representation and a second-order statistic feature distribution representation, and using the Wasserstein distance to calculate the difference between the first-order statistic feature distribution representation and the second-order statistic feature distribution representation includes: Inputting the source domain encoding feature and the target domain encoding feature into a feature fusion layer of a domain discriminator, wherein the feature fusion layer uses an attention mechanism to calculate a feature weight matrix, and performs weighted fusion on the source domain encoding feature and the target domain encoding feature based on the feature weight matrix to generate a domain encoding feature; Calculating a weighted mean vector and a median vector of the domain coding feature as first-order statistics, wherein the weighted mean vector is obtained by performing sample weighted averaging on each dimension of the domain coding feature, and the median vector is obtained by performing sorting statistics on each dimension of the domain coding feature; Calculating a covariance matrix and a correlation coefficient matrix of the domain coding features as second-order statistics, wherein the covariance matrix represents the relationship between different dimensions of the domain coding features, and the correlation coefficient matrix is ​​obtained by normalizing the covariance matrix; Combining the weighted mean vector and the median vector to construct a first-order statistical feature distribution representation, and combining the eigenvalues ​​and eigenvectors of the covariance matrix to construct a second-order statistical feature distribution representation; The Wasserstein distance represented by the characteristic distribution of the first-order statistics is calculated by using the Sinkhorn algorithm, and the Wasserstein distance represented by the characteristic distribution of the second-order statistics is calculated by using the matrix decomposition technique; The optimal weight coefficient is determined based on cross-validation, and the Wasserstein distance represented by the first-order statistic feature distribution and the Wasserstein distance represented by the second-order statistic feature distribution are weightedly combined to generate the inter-domain distribution difference.

8. A news event spatial verification and analysis system based on satellite remote sensing images, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to receive news event text information and extract event timestamps, geospatial coordinates, and event type identifiers from the news event text information using a natural language processing model; Determine the target observation area based on the extracted geospatial coordinates, and calculate the optimal satellite observation time window according to the event timestamp; input the target observation area and the optimal satellite observation time window into the strategy correction model to generate a satellite cluster collaborative observation strategy, wherein the satellite cluster collaborative observation strategy defines a collaborative observation sequence of high-resolution satellites, high-phase satellites, and hyperspectral satellites; issue a scheduling instruction according to the satellite cluster collaborative observation strategy to activate the corresponding satellite cluster to perform the collaborative observation task; The second unit is used to receive multi-source remote sensing image data acquired during the satellite group's collaborative observation mission, perform adaptive correction processing on the multi-source remote sensing image data, and generate standardized image data with spatiotemporal registration; select a corresponding target detection network and scene segmentation network from a deep learning model library according to the event type identifier, perform intelligent analysis on the standardized image data, and extract event-related ground features and spatiotemporal change characteristics; The extracted ground features and spatiotemporal change characteristics are input into the graph neural network, and the event feature association graph is constructed by integrating historical verification knowledge. The feature association strength value is obtained through graph reasoning calculation. The third unit is used to perform multi-dimensional cross-validation on the feature association strength value, the event timestamp, and the geographic space coordinates to construct an event authenticity evaluation index; Based on the event authenticity evaluation index and combined with historical cases in the verification knowledge base, a transfer learning method is used to calculate the event credibility score; When the event credibility score exceeds the preset benchmark credibility threshold, the authenticity of the news event is confirmed, and the satellite group collaborative observation strategy, the event feature association map, and the event authenticity evaluation index in this verification process are written into the verification knowledge base.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images