Multi-source monitoring analysis and decision-making method, device, equipment and medium

Abstract

The invention relates to the technical field of data analysis, can be applied to agricultural disaster monitoring, financial science and technology and other business scenes, and discloses a multi-source monitoring analysis and decision-making method, device, equipment and medium, and the method comprises the steps: obtaining multi-source monitoring data, and carrying out the time-space multi-scale fusion to generate time-space fusion data, extracting difference features of the target object before and after the event based on the fusion data to identify a first type of events, predicting a second type of events based on dynamic time sequence features by using a space-time diagram network model, and constructing a knowledge graph to perform causal reasoning on two types of event results to generate a reasoning enhancement result; and generating comprehensive abnormal analysis information and a response decision according to a reasoning enhancement result. According to the method, spatial-temporal features of multi-source data are fused, a causal reasoning mechanism is introduced, the identification and prediction results are subjected to correlation analysis, dynamic interpretation and response generation of complex events are achieved, and therefore the anomaly identification precision and the real-time performance and interpretability of risk decision making are improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method, apparatus, equipment and medium for analysis and decision-making based on multi-source monitoring. Background Technology

[0002] In the field of agricultural disaster monitoring and risk management, existing technologies have significant shortcomings in multi-source data fusion, spatiotemporal feature extraction, and disaster inference capabilities. Traditional agricultural disaster identification and prediction methods mostly rely on single-type remote sensing images or meteorological observation data, which have low spatial and temporal resolutions and are limited by cloud cover and data sampling cycles, making it impossible to continuously monitor the occurrence, development, and decline of disasters. Especially in the identification of lodging and waterlogging disasters, existing models are mostly built based on static remote sensing images or historical meteorological data, making it difficult to characterize the dynamic changes in crop growth status, resulting in delayed disaster identification and a high misjudgment rate. In addition, data fusion methods mostly remain at the level of surface feature stitching, lacking spatiotemporal correlation modeling and scale adaptive processing mechanisms between multi-source monitoring data, resulting in limited model generalization ability and significant differences in identification and prediction performance across different regions and crop types.

[0003] In the fintech sector, agricultural insurance risk assessment and claims decisions heavily rely on human experience. Existing risk assessment systems generally employ rule templates or static scorecard models, failing to respond in real-time to changes in disaster conditions and lacking an understanding of disaster propagation paths and causal relationships. Due to the lack of multi-source spatiotemporal information fusion and causal reasoning mechanisms, insurance institutions struggle to quantify the scope and extent of losses in a timely manner, resulting in long payout cycles, low capital allocation efficiency, and an inability to meet the risk-sharing and precise pricing needs of agricultural finance. Furthermore, the lack of a data loop between disaster identification and prediction stages leads to a disconnect between risk assessment results and disaster response decisions, hindering the development of a dynamic agricultural risk management system. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, device, and storage medium for analysis and decision-making based on multi-source monitoring. This invention aims to solve the technical problem that existing technologies lack the ability to fuse multi-source monitoring data across time and space at multiple scales and to perform causal reasoning, resulting in fragmented identification and prediction of abnormal events and an inability to achieve dynamic and interpretable comprehensive decision analysis.

[0005] To achieve the above objectives, the present invention provides an analysis and decision-making method for multi-source monitoring, comprising: Acquire multi-source monitoring data, and perform spatiotemporal multi-scale fusion processing on the multi-source monitoring data to generate spatiotemporal fused data; Based on the spatiotemporal fusion data, by extracting the difference features of the target object before and after the event, the first type of event is identified, and the first type of event identification result is generated. Based on the dynamic temporal characteristics in the spatiotemporal fusion data, a spatiotemporal graph network model is used to predict the second type of event and generate the prediction result of the second type of event. Construct a knowledge graph and use the knowledge graph to perform causal reasoning on the identification results of the first type of event and the prediction results of the second type of event to generate reasoning-enhanced results; Based on the enhanced reasoning results, comprehensive anomaly analysis information and response decisions are generated.

[0006] Furthermore, to achieve the above objectives, the present invention provides a multi-source monitoring analysis and decision-making device, comprising: The multi-source monitoring data fusion module is used to acquire multi-source monitoring data, perform spatiotemporal multi-scale fusion processing on the multi-source monitoring data, and generate spatiotemporal fused data. The event recognition module is used to identify a first type of event and generate a first type of event recognition result by extracting the difference features of the target object before and after the event based on the spatiotemporal fusion data. The event prediction module is used to predict the second type of event based on the dynamic temporal characteristics in the spatiotemporal fusion data and using a spatiotemporal graph network model, and generate the prediction result of the second type of event. The causal reasoning enhancement module is used to construct a knowledge graph and use the knowledge graph to perform causal reasoning on the identification results of the first type of event and the prediction results of the second type of event, and generate reasoning enhancement results. The analysis and response decision module is used to generate comprehensive anomaly analysis information and response decisions based on the reasoning enhancement results.

[0007] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a multi-source monitoring analysis and decision-making program stored in the memory and executable on the processor, wherein when the multi-source monitoring analysis and decision-making program is executed by the processor, it implements the steps of the multi-source monitoring analysis and decision-making method as described above.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a multi-source monitoring analysis and decision-making program, wherein when the multi-source monitoring analysis and decision-making program is executed by a processor, it implements the steps of the multi-source monitoring analysis and decision-making method as described above.

[0009] Beneficial Effects: This invention relates to the field of data analysis technology and can be applied to business scenarios such as agricultural disaster monitoring and fintech. It discloses a method, apparatus, equipment, and medium for multi-source monitoring analysis and decision-making, including: acquiring multi-source monitoring data and performing spatiotemporal multi-scale fusion processing to generate spatiotemporal fusion data; extracting the difference features of the target object before and after an event based on the spatiotemporal fusion data to identify a first type of event and generate a first type of event identification result; predicting a second type of event based on the dynamic temporal features in the spatiotemporal fusion data using a spatiotemporal graph network model to generate a second type of event prediction result; constructing a knowledge graph and using the knowledge graph to perform causal reasoning on the first type of event identification result and the second type of event prediction result to generate reasoning enhancement results; and generating comprehensive anomaly analysis information and response decisions based on the reasoning enhancement results. This invention, by fusing the spatiotemporal features of multi-source monitoring data and introducing a causal reasoning mechanism, links the event identification result and prediction result at the knowledge level, realizing dynamic analysis and interpretation of abnormal events. This enables the system to generate more targeted comprehensive response decisions in different scenarios, thereby improving the accuracy of anomaly identification and the real-time performance and interpretability of risk decisions. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for the analysis and decision-making method of multi-source monitoring in one embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the multi-source monitoring analysis and decision-making method of the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-source monitoring analysis and decision-making device of the present invention. Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0011] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0012] The multi-source monitoring analysis and decision-making method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can acquire multi-source monitoring data from the client and perform spatiotemporal multi-scale fusion processing to generate spatiotemporal fusion data. Based on the spatiotemporal fusion data, it extracts the difference features of the target object before and after the event to identify the first type of event and generate the first type of event identification result. Based on the dynamic temporal features in the spatiotemporal fusion data, it uses a spatiotemporal graph network model to predict the second type of event and generate the second type of event prediction result. It constructs a knowledge graph and uses the knowledge graph to perform causal reasoning on the first type of event identification result and the second type of event prediction result to generate reasoning enhancement result. Based on the reasoning enhancement result, it generates comprehensive anomaly analysis information and response decisions. This invention, by fusing the spatiotemporal features of multi-source monitoring data and introducing a causal reasoning mechanism, associates the event identification result and the prediction result at the knowledge level, realizing the dynamic analysis and interpretation of abnormal events. This enables the system to generate more targeted comprehensive response decisions in different scenarios, thereby improving the accuracy of anomaly identification and the real-time performance and interpretability of risk decisions. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster composed of multiple servers. The following is a detailed description of the invention through specific embodiments.

[0013] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the multi-source monitoring analysis and decision-making method provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0014] like Figure 2 As shown, the multi-source monitoring analysis and decision-making method proposed in this invention includes the following steps: S10, acquire multi-source monitoring data, perform spatiotemporal multi-scale fusion processing on the multi-source monitoring data, and generate spatiotemporal fusion data; In this embodiment, the process of acquiring multi-source monitoring data includes collecting and aggregating data from different sources, with different resolutions and update cycles. The types of multi-source monitoring data include remote sensing imagery, ground sensor monitoring records, meteorological observation data, and UAV aerial photography data. The acquisition phase requires establishing corresponding interfaces and format parsing mechanisms for each type of data source. For example, for remote sensing imagery data, the image file is read through the image acquisition interface with attached geospatial identifiers; for ground sensor data, the data stream reported by the sensor terminal is received through a real-time communication module; for meteorological observation data, the standardized data interface of the meteorological center is called; and for UAV imagery data, flight path planning and image upload synchronization are completed through a task scheduling system. After acquisition, data from different sources enters a unified data buffer, forming a traceable data queue using timestamps and geographic locations as index keys.

[0015] In the process of spatiotemporal multi-scale fusion processing of multi-source monitoring data, the first step is to perform temporal alignment, interpolating or resampling data from different sampling periods according to a unified time step. Next, spatial alignment is performed, mapping each data source to a unified spatial coordinate system through geographic coordinate projection transformation and gridded interpolation algorithms. Subsequently, scale transformation and fusion are performed at the feature level, including super-resolution reconstruction of low-resolution images, smoothing filtering of high-frequency noise signals, and pyramid-based hierarchical fusion of data at different spatial granularities. A weighted attention mechanism can be introduced during the fusion process, dynamically adjusting the fusion weights based on the confidence level and spatiotemporal correlation of the data sources to achieve coordinated and unified representation of multi-scale information.

[0016] The process of generating spatiotemporal fused data involves creating a dataset with a unified time step, spatial resolution, and feature dimensions after the alignment and fusion process described above. This dataset is organized into time-series segments, each containing geographic coordinates, a time stamp, and a fused feature vector. Data storage employs a hierarchical structure, supporting partitioned indexing and parallel access to accommodate subsequent event recognition and prediction model calls. The generated spatiotemporal fused data includes not only physical quantities (such as temperature, humidity, wind speed, and surface reflectivity) but also higher-order features extracted from the original data, such as vegetation index change rates, surface humidity gradients, or cloud cover ratios, thereby enhancing the availability and stability of subsequent calculations.

[0017] In different implementations, the fusion process can be achieved through various technical paths. For example, in scenarios involving joint monitoring by satellite remote sensing and ground sensors, a distributed fusion engine can process data from different sources in parallel, and feature layer fusion can be achieved using multi-layer convolutional and temporal coding networks. In scenarios with only ground sensor and meteorological observation data, a Kalman filter-based spatiotemporal state estimation method can be used to obtain continuous fusion results by dynamically updating model parameters. In environments requiring high-time-response capabilities, edge computing nodes can be used to perform initial data fusion and cleaning, with only the intermediate fused features uploaded to the central node for unified processing, thereby reducing network latency and computational load.

[0018] In different application environments, the spatiotemporal fusion parameters can be adaptively adjusted. For situations with high data source noise, the filtering smoothing coefficient can be increased or the time window length can be shortened. For plots or regions with significant spatial heterogeneity, the grid size and fusion radius can be dynamically adjusted to ensure both local accuracy and overall consistency in the fusion results. After data fusion is completed, a data quality assessment module can be used to mark and compensate for abnormal missing data or fusion deviations to improve the usability of the spatiotemporal fusion data.

[0019] This embodiment achieves spatiotemporal multi-scale fusion of multi-source monitoring data, unifying data from different sources with varying temporal and spatial resolutions into a standardized representation, thus enabling information complementarity and dynamic correlation. This process effectively enhances the monitoring system's responsiveness to complex environmental changes, allowing subsequent event identification and prediction models to operate with more complete, continuous, and consistent data input, thereby improving the accuracy and stability of anomaly detection.

[0020] S20, based on the spatiotemporal fusion data, by extracting the difference features of the target object before and after the event, the first type of event is identified, and the first type of event identification result is generated; In this embodiment, the event recognition process based on spatiotemporal fusion data focuses on differential feature extraction. It identifies abnormal states by comparing the multidimensional changes of the target object before and after the event. The target object can be a geographical unit, crop plot, equipment node, or individual entity within the monitoring area, and its corresponding spatiotemporal fusion data contains multi-source feature vectors from continuous time slices. First, the comparison interval before and after the event needs to be defined in the time dimension. Typically, a sliding window mechanism is used to determine the comparison period to ensure the representativeness and stability of the difference calculation. Then, at the feature level, the changes in indicators across different time slices are compared, including numerical features such as spectral reflectance, surface temperature and humidity, wind speed, irradiance, and soil conductivity, as well as composite indicators generated based on these feature combinations, such as vegetation indices, energy balance indices, or dynamic moisture indices. By modeling the changing trends of these indicators in the time dimension, feature difference vectors can be generated to characterize the disturbance patterns caused by the event.

[0021] Multiple technical approaches can be employed in the differential feature extraction process. For example, differential convolution-based feature extraction units can perform differential convolution operations on feature maps of adjacent time slices to obtain regions of significant change; alternatively, time series analysis methods can be used to identify abrupt change points through trend decomposition and rate of change calculation. Changes in spatial dimensions can be detected through a neighborhood comparison mechanism, calculating the feature difference distribution between the target object and its surrounding objects to distinguish between local anomalies and global changes. For multi-source heterogeneous data, feature standardization and alignment modules can be introduced to map feature values ​​from different sources to a unified dimensional space, thereby avoiding biases caused by dimensional differences.

[0022] The process of identifying the first type of event occurs after the differential feature extraction is completed. The system uses an event recognition model to classify or cluster the differential features. The model structure can include convolutional neural networks, graph neural networks, or hybrid spatiotemporal attention networks. By learning feature change patterns corresponding to different event types during the training phase, the event recognition model can classify and output the input differential feature vectors during the inference phase. The output includes event type labels, confidence scores, and spatial location information. Based on the output of the recognition model, the system generates the first type of event recognition results and encapsulates them in a structured manner, forming a standardized record containing event identifier, location, time interval, and feature summary.

[0023] In one implementation, differential feature extraction is achieved through a bidirectional temporal convolutional network based on a sliding window. The network's two-branch structure inputs the preceding and following data segments of the target object in the time series to automatically learn temporal difference features. In another implementation, a change detection algorithm based on a dynamic threshold can be used. The time series changes of various indicators of the target object are compared with an adaptive threshold; when the rate of change exceeds the threshold, it is marked as a potential event trigger point. Alternatively, a frequency domain analysis-based method can be employed. By comparing the changes in spectral energy distribution before and after the event, regions with significant high-frequency fluctuations can be identified, thereby determining possible abnormal events.

[0024] The input and structure of the event recognition model can be adapted to different scenarios. In remote sensing monitoring environments, spatial feature extraction based on multi-channel convolutional networks can be used; in sensor network environments, topological input based on graph attention networks can be used to reflect changes in the relationships between different nodes. Regarding model parameter settings, feature transfer weights can be extracted from existing labeled datasets using transfer learning strategies to improve recognition performance in new scenarios.

[0025] This embodiment, through differential feature extraction and event recognition based on spatiotemporal fusion data, can accurately capture abnormal change characteristics of target objects in the temporal and spatial dimensions, achieving intelligent identification of sudden events in dynamic environments. This process enables the system to not only rely on static threshold judgments but also automatically learn event patterns from changing trends, thereby improving the sensitivity and accuracy of identification for complex events.

[0026] S30, Based on the dynamic temporal characteristics in the spatiotemporal fusion data, a spatiotemporal graph network model is used to predict the second type of event and generate the prediction result of the second type of event; In this embodiment, the core of event prediction based on dynamic temporal features in spatiotemporal fusion data lies in mining the implicit trends and spatial relationships within the time series. Dynamic temporal features refer to feature sequences that change continuously over time and have spatial dependence, originating from fused multi-source monitoring data. To extract these features, the spatiotemporal fusion data of continuous time slices is first temporally encoded. Time segments are divided using a sliding window, and feature indicators such as difference components, rate of change, and periodic fluctuations in the time dimension are extracted. Simultaneously, the feature correlation matrix between neighboring regions or nodes is calculated in the spatial dimension to reflect the spatial dependence structure during the dynamic evolution process. The input tensor generated after feature extraction contains time series features, spatial adjacency relationships, and node attribute information, forming a data format that can be input into a spatiotemporal graph network model.

[0027] Spatiotemporal graph network models utilize graph structures to represent spatial nodes and their connections. In graph convolution operations, features of each node and its neighboring nodes are aggregated to capture spatial dependencies. Simultaneously, a time-series modeling module is incorporated, using gated recurrent units or temporal convolutional networks to model temporal dependencies. The graph network is constructed based on the geographical proximity, meteorological similarity, or physical transmission paths of the monitoring area to determine the adjacency matrix. During the training phase, the model learns the spatiotemporal mapping relationship between node features and target events through supervised or semi-supervised learning strategies. During the inference phase, after inputting new dynamic temporal features, the model can output the probability of future event occurrences or the risk level.

[0028] The process of predicting the second type of event comprises two parts: model inference and result generation. In the model inference stage, the processed dynamic temporal feature tensor is input, and after multi-layer graph convolution and time series encoding, a latent space representation is generated. This latent representation is then output as a prediction result through classification or regression layers. In the result generation stage, the model output undergoes post-processing, including confidence filtering, spatial smoothing, and outlier suppression, to ensure the stability and spatial consistency of the prediction results. The final generated prediction result for the second type of event includes the event type, prediction time period, spatial range, and probability distribution, which can be used as input for subsequent inference enhancement and decision analysis.

[0029] In one implementation, the spatiotemporal graph network model is constructed based on a fixed adjacency matrix, suitable for monitoring nodes with relatively fixed geographical locations, such as farmland monitoring stations or urban meteorological grid nodes. The graph convolutional layers employ Chebyshev multinomial approximation to aggregate high-order neighborhood features, and the temporal dimension uses gated recurrent units for sequence modeling to improve the temporal sensitivity of predictions. In another implementation, the model's adjacency relationships can be dynamically updated, adaptively adjusting connection weights based on the feature similarity between nodes to adapt to the spatial diffusion characteristics of sudden or migratory events. A multi-channel input mechanism can also be employed, inputting different types of temporal features (e.g., temperature, humidity, wind speed sequences) into the graph network separately, performing weighted fusion in the fusion layer to capture the coupling relationships between multidimensional features.

[0030] During model optimization, supervised training can be performed using historical event label data, with cross-entropy loss or mean squared error as the optimization objective, combined with early stopping and regularization techniques to prevent overfitting. For scenarios with severe data sparsity or missing data, transfer learning or graph completion strategies can be introduced to transfer some parameters from training models in similar regions, thereby improving the generalization ability of predictions. In environments with limited computing resources, lightweight graph network structures or pruning strategies can be used to reduce the computational cost of the model, enabling rapid prediction of edge nodes.

[0031] This embodiment utilizes a spatiotemporal graph network model to analyze and predict dynamic temporal characteristics, capturing temporal evolution patterns while maintaining the integrity of the spatial structure, thus enabling early warning of future events. This process effectively overcomes the limitations of traditional static models in representing spatial dependencies and temporal dynamics, giving the system the ability to predict nonlinear evolution phenomena in complex environments, thereby improving the accuracy and foresight of predictions.

[0032] S40, construct a knowledge graph, and use the knowledge graph to perform causal reasoning on the identification results of the first type of event and the prediction results of the second type of event to generate reasoning-enhanced results; In this embodiment, the process of constructing a knowledge graph aims to express the entities, attributes, and relationships between multi-source monitoring data and event outcomes through a structured graph model. The knowledge graph consists of nodes and edges, where nodes represent entities, such as monitoring areas, meteorological elements, crop types, equipment nodes, or time slices; edges represent semantic relationships between entities, such as "impact," "trigger," "dependency," or "relatedness." During the construction process, semantic extraction and structured parsing are first performed on the spatiotemporal fusion data and its derived event identification and prediction results to extract key fields and their interaction features. For example, event type, occurrence time, and spatial location are extracted from the first type of event identification results, and risk probability, time interval, and prediction indicators are extracted from the second type of event prediction results. Subsequently, through entity alignment and relationship matching mechanisms, entities from different sources are mapped to a unified knowledge space, establishing a dual connection relationship at the spatiotemporal and semantic levels.

[0033] The generation of a knowledge graph typically involves three stages: semantic parsing, relation inference, and graph fusion. The semantic parsing stage uses natural language processing models to identify semantic labels for entities and attributes. The relation inference stage establishes connections through rule-based logical relationships and data-driven statistical correlations, such as using Pearson correlation coefficients or mutual information coefficients to determine the degree of coupling between two events. The graph fusion stage merges and disambiguates subgraphs from different sources, maintaining graph consistency through unique identifiers and confidence thresholds. The completed knowledge graph has a dynamic update mechanism, automatically and incrementally updating nodes and edges when new monitoring data or event results enter the system.

[0034] In the process of causal reasoning using knowledge graphs, the system performs causal path analysis based on the topological structure of event identification and prediction results within the graph. Causal reasoning differs from correlation analysis in that it focuses on determining the causal direction and influence paths between events. The system can employ causal graph models or structural equation models to verify the directionality of candidate paths. The reasoning process includes candidate path generation, path scoring, and causal chain determination. The candidate path generation stage traverses the set of nodes in the knowledge graph related to the target event; the path scoring stage calculates path credibility based on chronological order, correlation strength, and node importance; and finally, a set of causal chains is formed through threshold filtering.

[0035] The process of generating enhanced inference results involves combining the output of the causal inference with the original event outcome to form an enhanced result set. The enhanced results include causal chains, key influencing factors, event triggering patterns, and probability correction values. This result retains the correspondence between events and causal paths in its data structure to support subsequent response decision analysis and visualization.

[0036] In one implementation, the knowledge graph construction relies on a graph database framework, using a graph query language to define and update entity relationships, making it suitable for large-scale heterogeneous data fusion scenarios. In another implementation, the knowledge graph's relationship inference module employs a graph representation method based on embedding learning. By training a graph embedding model (such as TransE or RotatE), nodes and edges are mapped to a vector space, and relationship prediction is performed using vector similarity, thereby enhancing causal path recognition capabilities. Furthermore, a streaming update mechanism can be used in environments requiring high timeliness, inserting graph nodes in real-time as event data streams arrive and dynamically updating causal edge weights.

[0037] Different implementation methods can employ different algorithmic paths for causal reasoning. In scenarios with abundant data, structural equation modeling can be used to determine the causal direction by fitting functional relationships between nodes. In scenarios with scarce data, Bayesian network-based reasoning can be used to estimate causal dependencies through conditional probability calculations. To enhance the interpretability of the reasoning, visual descriptions and textual explanations of the path can be generated in the reasoning output, enabling users to intuitively understand the logic and direction of the causal chain.

[0038] This embodiment constructs a knowledge graph and performs causal reasoning based on it, transforming event identification and prediction results from independent analytical outcomes into a reasoning system with interconnected logic and an interpretable structure. This process enhances the system's ability to understand the causal relationships between events, making the results not only predictive but also interpretable and traceable, thus providing a solid data logic foundation for subsequent intelligent responses and strategy generation.

[0039] S50, Based on the reasoning enhancement results, generate comprehensive anomaly analysis information and response decisions.

[0040] In this embodiment, generating comprehensive anomaly analysis information and response decisions based on the reasoning enhancement results is a key step in achieving a closed-loop intelligent decision-making system. This process takes the reasoning enhancement results as input, comprehensively considering event identification results, event prediction results, and causal chain information inferred from the knowledge graph to perform multi-dimensional anomaly assessment of the monitored area or target object. First, it extracts anomaly-related indicator information and causal relationship weights from the reasoning enhancement results, and calculates a comprehensive anomaly score for each monitoring unit. This score can be generated using a multi-factor weighted model, including parameters such as event intensity, spatial influence range, causal chain path length, and node importance. A hierarchical weight allocation mechanism can be introduced during the calculation process to increase the weight of direct causal factors, thereby enhancing the model's sensitivity to dominant factors.

[0041] After obtaining the comprehensive anomaly score, the system uses spatial clustering and semantic aggregation methods to classify monitoring areas with similar anomaly characteristics into warning areas. The clustering algorithm can employ density-based spatial clustering or graph partitioning methods, combining geographical proximity and anomaly feature similarity to determine the region affiliation. The resulting warning areas not only reflect spatial concentration but also temporal synchronicity, providing structured input for subsequent response decisions.

[0042] The generation of response decisions relies on the causal reasoning chain within the enhanced reasoning results. Based on the anomaly level of different warning areas, and considering the direction and degree of influence of key factors in the causal chain, the system generates targeted response strategies. For example, when the dominant factor in the causal chain is a meteorological parameter, the response strategy may include adjusting agricultural irrigation plans, issuing meteorological disaster warnings, or optimizing insurance claims processes. When the dominant factor is equipment malfunction, it triggers equipment maintenance and reconfiguration instructions.

[0043] Ultimately, the system integrates comprehensive anomaly scores, early warning area division results, and response strategies into a visual interactive interface, generating comprehensive anomaly analysis information and response decisions. The interface displays the spatial distribution of anomalies and risk level gradients through a geographic information visualization module, and supports users in querying causal paths, viewing anomaly indicators, and triggering response commands, achieving integrated data analysis and decision execution.

[0044] In one implementation, the comprehensive anomaly score is calculated using a weighted graph aggregation method. The influence values ​​of nodes in the reasoning enhancement results are used as input weights. By traversing the causal paths of the knowledge graph, the transmission effects between nodes are accumulated to reflect the intensity of anomalies under the superposition of multiple factors. In another implementation, an anomaly aggregation model based on a neural attention mechanism can be used to perform feature learning on the reasoning enhancement results to automatically identify the most influential causal relationships and generate a comprehensive anomaly score.

[0045] The algorithm for delineating warning areas can be adjusted according to different scenarios. For example, density-based clustering methods (such as DBSCAN) can be used in environments with dense geospatial data, while graph-based regional clustering methods (such as Spectral Clustering) can be used in environments with sparse monitoring nodes or topological dependencies. The response strategy generation module can combine an expert knowledge base and use a rule engine to map causal inference results to corresponding response templates to achieve automated strategy recommendation; alternatively, a reinforcement learning model can be used to dynamically adjust strategy priorities and thresholds based on historical response performance.

[0046] In terms of presenting and interacting with decision results, the system can adopt a multi-layered visualization structure. The first layer is a spatial risk distribution map, used to display the anomaly levels of different regions; the second layer is a causal path visualization, used to present the logical relationship between key factors and events; the third layer is a response strategy panel, which allows users to filter and trigger corresponding execution instructions by region, time, or event type.

[0047] This embodiment generates comprehensive anomaly analysis information and response decisions based on enhanced reasoning results, enabling the system to achieve closed-loop intelligent processing from data analysis to strategy execution. This process not only integrates multi-source data and causal reasoning results but also constructs a dynamically interpretable response system by quantifying anomaly characteristics and causal relationships. Therefore, the system achieves synergistic linkage in anomaly identification, risk classification, and decision execution, significantly improving response speed, decision accuracy, and event controllability.

[0048] This invention relates to the field of data analysis technology and can be applied to business scenarios such as agricultural disaster monitoring and fintech. It discloses a method, apparatus, equipment, and medium for multi-source monitoring analysis and decision-making, comprising: acquiring multi-source monitoring data and performing spatiotemporal multi-scale fusion processing to generate spatiotemporal fusion data; extracting the difference features of the target object before and after an event based on the spatiotemporal fusion data to identify a first type of event and generate a first type of event identification result; predicting a second type of event based on the dynamic temporal features in the spatiotemporal fusion data using a spatiotemporal graph network model to generate a second type of event prediction result; constructing a knowledge graph and using the knowledge graph to perform causal reasoning on the first type of event identification result and the second type of event prediction result to generate reasoning enhancement result; and generating comprehensive anomaly analysis information and response decisions based on the reasoning enhancement result. This invention, by fusing the spatiotemporal features of multi-source monitoring data and introducing a causal reasoning mechanism, links the event identification result and prediction result at the knowledge level, realizing dynamic analysis and interpretation of abnormal events. This enables the system to generate more targeted comprehensive response decisions in different scenarios, thereby improving the accuracy of anomaly identification and the real-time performance and interpretability of risk decisions.

[0049] In one embodiment, step S10 includes: S101 acquires optical remote sensing data, synthetic aperture radar data, and ground sensor data to form multi-source monitoring data; S102, Detect cloud and fog obscured areas in the optical remote sensing data and identify pixels affected by cloud and fog obscuration; S103, Based on spectral features and temporal variation models, repair the pixels affected by cloud and fog obstruction, and generate repaired optical remote sensing data; S104, the repaired optical remote sensing data, the synthetic aperture radar data and the ground sensor data are spatiotemporally registered to form registered multi-source monitoring data; S105, The registered multi-source monitoring data is input into a spatiotemporal multi-scale fusion network to extract the fusion features of spatial and temporal dimensions; S106, perform super-resolution reconstruction on the fusion features to generate spatiotemporal fusion data.

[0050] In this embodiment, the process of acquiring optical remote sensing data, synthetic aperture radar data, and ground sensor data to form multi-source monitoring data is organized and executed with unified acquisition, quality screening, and indexing, taking into account heterogeneous sources, resolutions, and sampling periods. Optical remote sensing data is read through an image interface to access multispectral or hyperspectral bands, while simultaneously loading imaging geometry and time stamps. Synthetic aperture radar data mainly consists of amplitude maps and coherence information, including dual-polarization or multi-polarization channels. Ground sensor data includes continuous monitoring records of temperature, humidity, soil moisture content, wind speed, and rainfall. After acquisition, a spatiotemporal keyed storage is established using timestamps and geographic coordinates as the primary index key, providing a searchable data view for subsequent alignment and fusion. Quality screening removes and corrects missing frames, mismatched time stamps, outliers, saturated pixels, and sensor drift, generating a data availability mask and confidence weights. The confidence weights directly participate in weight allocation during subsequent weighted fusion.

[0051] The process of detecting cloud-obscured areas and identifying pixels affected by cloud obscuration in optical remote sensing data is based on a multi-discriminator detection framework constructed using spectral, texture, and temporal consistency criteria. The spectral discriminator identifies highly albedo clouds and thick cloud boundaries based on the combined ratio of blue, shortwave infrared, and near-infrared wavelengths. The texture discriminator characterizes thin clouds and cloud shadows using gradient sparsity and local contrast distribution. The temporal discriminator detects sudden high albedo and anomalous darkening using a stability metric from continuous observations. The three discrimination results are fused with confidence to generate a cloud-obscured area mask, while simultaneously outputting the set of pixels affected by cloud obscuration and a confidence surface to guide subsequent restoration areas and intensities.

[0052] The restoration of pixels affected by cloud and fog based on spectral features and temporal variation models employs a joint restoration strategy of temporal interpolation, spatial guidance, and cross-modal constraints. Temporal interpolation constructs local trends by comparing previous and subsequent clean observations and extrapolates reflectance or exponential values ​​at the time of occlusion. Spatial guidance uses unoccluded neighboring pixels as priors and recovers details and textures using guided filtering or local regression. Cross-modal constraints utilize the sensitivity of synthetic aperture radar to changes in surface structure and water content as structural priors to suppress blurring and drift. The three-way estimation is weighted by uncertainty to obtain the restored optical remote sensing data, while simultaneously updating the availability mask, enabling differentiated and reliable management of the restored area and the original clean area in subsequent registration and fusion.

[0053] The process of spatiotemporally registering restored optical remote sensing data, synthetic aperture radar data, and ground sensor data to form registered multi-source monitoring data is divided into two steps: time synchronization and spatial alignment. Time synchronization involves resampling and aligning data from different sampling periods to construct a time-series stack with a unified time step. Spatial alignment addresses the geometric differences between optical and radar imaging by iteratively estimating geometric transformations using sensor geometric models, terrain elevations, and ground control points, followed by sub-pixel fine registration to reduce parallax. Ground sensor data is projected onto a unified raster through spatial interpolation or station-to-pixel mapping, corresponding one-to-one with pixel-level remote sensing features. After alignment, registered multi-source monitoring data is generated, including a unified coordinate system, a unified pixel grid, and a unified time axis, along with confidence and quality labels for each source.

[0054] The goal of inputting the registered multi-source monitoring data into a spatiotemporal multi-scale fusion network and extracting fusion features in both spatial and temporal dimensions is to simultaneously preserve local texture, regional structure, and intertemporal evolution. The network front end uses multi-resolution pyramid convolution to extract spatial representations of different receptive fields; the temporal channel uses one-dimensional temporal convolution or gated units to encode short-term changes and medium-term trends; the cross-source channel introduces weighted attention, adaptively allocating the contribution ratios of optical, radar, and ground-based sources, with weights modulated by the aforementioned confidence and quality labels; graph structure units can aggregate neighborhood information on the pixel adjacency graph to enhance spatial consistency. The fusion output forms fusion features in both spatial and temporal dimensions at the pixel scale. These features include texture sub-representations, structure sub-representations, temporal change sub-representations, and source contribution vectors. The source contribution vectors are used for subsequent reconstruction and uncertainty control.

[0055] The process of super-resolution reconstruction of fused features to generate spatiotemporal fusion data focuses on spatial detail restoration and temporal continuity maintenance. The reconstruction network achieves detail gain with a residual-dense structure and enhances resolution with multi-scale upsampling modules; edge consistency branches constrain the coherence of fine-grained boundaries and elongated textures; temporal consistency loss suppresses intertemporal jumps and flicker; and source contribution vectors are used as modulation during reconstruction to reduce the gain of noise-dominated sources. The output spatiotemporal fusion data possesses uniform spatial resolution, uniform time step, and uniform feature dimensions, along with uncertainty estimation and quality masks, facilitating direct retrieval for subsequent identification and prediction. The data is organized using block-level indexing and columnar storage, supporting large-area parallel access and incremental updates.

[0056] This embodiment acquires multi-source monitoring data from optical remote sensing data, synthetic aperture radar data, and ground sensor data. It detects and identifies pixels affected by cloud and fog obscuration in cloud-covered areas and performs repair based on spectral features and temporal variation models. After temporal synchronization and spatial alignment, it forms registered multi-source monitoring data. Furthermore, a spatiotemporal multi-scale fusion network extracts fusion features of spatial and temporal dimensions and completes super-resolution reconstruction, ultimately generating spatiotemporal fused data. This achieves a unified representation of heterogeneous sources, resolutions, and sampling periods. The effects of clouds and fog are explicitly addressed through detection and repair; geometric differences and temporal misalignments are eliminated through alignment and resampling; redundant and complementary information between sources is integrated through attention weighting; spatial details are restored through reconstruction; and temporal continuity is ensured through temporal coding. This provides a high signal-to-noise ratio, strong consistency, and high-resolution input foundation for subsequent event recognition and prediction, reducing the probability of false positives and false negatives and improving stability and generalization capabilities under complex meteorological conditions and multi-plot scenarios.

[0057] In one embodiment, step S20 above includes: S201, extract the first multimodal feature set of the target object before the event occurs and the second multimodal feature set after the event occurs from the spatiotemporal fusion data; S202, determine the difference between the first multimodal feature set and the second multimodal feature set, and generate a difference feature set of the target object; S203, input the difference feature set into the event recognition model to identify the target object region where the event occurred; S204, perform event severity analysis on the target object area where the event occurred, and generate a first type of event identification result containing event type and severity level.

[0058] In this embodiment, based on spatiotemporal fusion data, the differential features of the target object before and after the event are extracted. First, the target object and time interval are determined. The target object can be a raster cell, a plot unit, a device node, or other uniquely locatable monitoring entity. Paired windows are established around the potential event time, and a symmetrical sliding method is used to obtain the observation sequence before the event and the observation sequence after the event. To ensure comparability, uniform cleaning and scale normalization are first performed on the two sequences, including outlier removal, temporal interpolation, radiometric calibration consistency, spatial neighborhood smoothing, and source weight correction, so that features from different sources and at different time periods are brought to the same dimension and noise level.

[0059] The first multimodal feature set comes from the observation sequence before the event, and the second multimodal feature set comes from the observation sequence after the event. Both feature sets are organized using multi-source channels, including reflectivity and vegetation index family from optical channels, amplitude and coherence features from synthetic aperture radar channels, moisture and meteorological parameters from ground sensing channels, and texture and morphological descriptors derived from these basic variables. To enhance spatiotemporal representation, statistical and dynamic quantities are extracted within two time windows. The statistical quantities cover time mean, steady-state quantile, and robust dispersion, while the dynamic quantities cover monotonic trend strength, short-term fluctuation strength, and periodic residual energy. Multi-scale neighborhood calculation is used for the spatial texture of optical and radar data to generate direction-sensitive boundary and structural components; neighborhood aggregation after site-to-cell mapping is used for ground observations to form ground constraint components consistent with the raster. The first and second multimodal feature sets constructed in this way possess pairable channel structures and statistical properties.

[0060] The differential feature set is generated based on a channel-by-channel comparison of two feature sets. The comparison strategy simultaneously considers intensity difference, ratio difference, and directional consistency, and incorporates robust centering to make the differences insensitive to occasional noise. For optical indicators, channel-level differences and ratios after brightness and chromaticity decomposition are used to mitigate illumination and aerosol disturbances; for radar indicators, joint amplitude difference and coherence attenuation difference reflect structural damage and moisture changes; for ground observations, the rate of change and threshold crossing count are calculated at a unified time step to characterize abrupt changes and persistence. The differential feature set also includes two normalized coordinates: spatial comparison and group comparison. Spatial comparison compares the target object with its neighboring baselines to avoid interference from regional overall drift; group comparison compares the target object with the concurrent distribution of similar objects to highlight individual anomalies. All differential channels are weighted by source confidence weight and time window weight to form a structured differential feature vector, and outputs a quality mask and uncertainty estimate for use in the downstream identification stage.

[0061] The event recognition model receives a set of differential features and outputs the target object region where the event occurred. The model can employ a spatiotemporal attention aggregation structure, grouping differential channels by source and scale, and then focusing on high-contribution channels and key time periods through multi-head attention units; alternatively, it can use a pairwise encoding structure, performing metric learning on the embeddings before and after the event to ensure linear separability of the difference in the embedding space. To improve spatial consistency, a pixel adjacency graph or plot topology is introduced to aggregate differential evidence of nearby entities in graph convolutional units, suppressing isolated noise. During the inference phase, an event occurrence probability and category assignment are generated for each entity, and post-processing is performed using a quality mask, including small patch denoising, boundary refinement, and temporal consistency constraints, ultimately forming the target object region where the event occurred.

[0062] Event severity analysis is performed on the target area where the event occurred, aiming to provide a joint determination of type and severity. First, a hierarchical feature cluster is constructed, incorporating three types of evidence: amplitude evidence, duration evidence, and spatial spread evidence. Amplitude evidence reflects the strength of damage and the degree of functional degradation; duration evidence comes from cross-window stability measures, distinguishing between transient disturbances and persistent damage; spatial spread evidence characterizes the diffusion trend through indicators such as regional connectivity and boundary roughness. The classifier employs a hierarchical determination, first distinguishing the type, then determining the severity level within each type to avoid cross-type confusion; cost-sensitive learning is introduced when necessary to reduce false positives for severity levels. The final output is the first type of event identification result, including target object identifier, type label, severity level, time interval, spatial range, and a summary of differential features, while maintaining a consistent mapping with the differential channels for easy subsequent interpretation and tracing.

[0063] This embodiment constructs a first multimodal feature set and a second multimodal feature set on spatiotemporal fusion data, and generates a difference feature set by combining intensity difference, proportion difference, and control normalization. Then, the difference evidence is aggregated and post-processed within spatial topology and temporal windows. The identification result shifts from single-source threshold judgment to multi-source collaborative change pattern judgment. Type differentiation and grade classification are based on the synthesis of amplitude, persistence, and spread evidence. As a result, abnormal changes in the target object are more completely quantified and located, noise interference and overall drift are systematically weakened, and spatial consistency and temporal stability are simultaneously constrained. The first type of event identification results achieve synergistic improvement in accuracy, robustness, and interpretability, providing highly reliable input for subsequent prediction, causal inference, and response decision-making.

[0064] In one embodiment, step S30 above includes: S301, Extract multivariate time series data related to the second type of event from the spatiotemporal fusion data; S302, based on geospatial relationships, divides the monitoring area into multiple spatial nodes and constructs a spatiotemporal graph structure describing the relationship between spatial nodes; S303, the multivariate time series data and the spatiotemporal graph structure are input into the spatiotemporal graph network model for processing to generate the probability and anomaly distribution of the second type of event in the future time period as the prediction result of the second type of event.

[0065] In this embodiment, prediction of the second type of event is carried out based on the dynamic temporal features in spatiotemporal fusion data. First, multivariate temporal data is organized on a unified time axis and a unified spatial grid. The multivariate temporal data comes from the fused multi-source observation channels and includes continuous sequences such as optical reflectance and vegetation index family, radar amplitude and coherence, soil moisture content and surface temperature, rainfall and wind field elements. To avoid dimensional differences and scale bias, robust standardization and outlier suppression are performed on all channels first. Continuous and reliable sequence slices are generated through time interpolation and quality mask constraints. The dynamic temporal features not only retain the original channel values, but also introduce derived quantities such as difference components, rate of change, persistence and periodic residuals to establish a unified representation of abrupt changes, gradual changes and seasonal fluctuations. At the same time, uncertainty weights are retained for each time slice, which are subsequently used for adaptive weighting and attention modulation.

[0066] The monitoring area is divided into a set of spatial nodes based on geospatial relationships. Spatial nodes can be mapped to fixed pixels, farmland plots, monitoring station networks, or raster blocks. The division follows boundary consistency and surface homogeneity, achieving spatial aggregation through topographic zoning, land cover consistency, and management unit boundaries, ensuring low intra-node differences and high inter-node differences. A spatiotemporal graph structure is constructed based on the adjacency relationships between spatial nodes. Static adjacency relationships are derived from Wang's adjacency or shared boundary topology, while dynamic adjacency relationships are updated over time. Edge weights are generated by integrating factors such as distance decay, topographic connectivity, runoff channels, wind direction dominance, and observational correlation. The range of edge weights is normalized and bound to uncertainty weights to reflect the strength of time-varying correlations. The time dimension is achieved by attaching a time series index and sliding window edges to each node, encoding time slices and node topology together into a three-dimensional graph data volume, forming an input object that can be directly entered into the graph model.

[0067] Multivariate temporal series data and spatiotemporal graph structure are jointly input into the spatiotemporal graph network model. The spatial branch uses spectral domain or spatial domain graph convolution to extract neighborhood dependencies, employing multi-order aggregation to cover the influence of nearest and second-nearest neighbors, avoiding information loss due to relying solely on a single neighbor. The temporal branch uses dilated temporal convolution or gated recurrent units to encode short-term oscillations and medium-term trends, and mitigates gradient decay through residual connections. Cross-source fusion is achieved collaboratively through channel attention and edge attention. Channel attention assigns adaptive weights to observed channels based on channel uncertainty and historical contribution, while edge attention adjusts the message propagation amplitude according to the strength of the collaborative change between edge weights and the most recent time period, thus stably aggregating effective information in a time-varying context. To address observation gaps and local anomalies, mask-aware normalization and mask-gated units are introduced, enabling the model to automatically weaken affected channels and edges when observations are missing, reducing misleading propagation. During training and inference, a consistent gain path between the data and topology is maintained to avoid bias caused by a single data source dominating certain periods.

[0068] The model's front end outputs the latent space representation of node time series, while the rear end generates the probability of occurrence and anomaly distribution of the second type of event in future time periods through classification or regression heads. The probability output employs multi-temporal horizon parallel prediction, covering short- to medium-term windows with a fixed prediction interval; the anomaly distribution output provides a node-level or grid-level anomaly intensity field in space. To improve spatiotemporal consistency and calibration reliability, the post-processing stage introduces temporal consistency constraints and spatial smoothing constraints to suppress sporadic spikes and isolated pixels. Simultaneously, temperature scaling or piecewise monotonic mapping is used to calibrate the probability output, ensuring stable hit rates and false alarm rates under different thresholds. The final generated prediction results for the second type of event contain both the probability of occurrence and the anomaly distribution in future time periods, maintaining a one-to-one correspondence with the input node and time indices, facilitating seamless integration with subsequent causal inference and response decision modules.

[0069] This embodiment constructs multivariate time-series data on a unified grid and time axis, explicitly expressing time-varying correlations between regions using a spatiotemporal graph structure. Spatial aggregation, temporal encoding, and cross-source attention are then synergistically applied to dynamic time-series features. The prediction process can simultaneously capture local disturbances and cross-regional linkages, reducing the distortion of results by missing data and noise, and significantly improving the stability and calibration of future occurrence probabilities and anomaly distributions. Against the backdrop of complex meteorological conditions and rapid surface changes, the output results are more temporally coherent and spatially consistent, providing high-confidence forward-looking information support for subsequent causal inference and response decisions.

[0070] In one embodiment, step S40 above includes: S401, construct a knowledge graph that includes event types, environmental factors and geographic entities; S402, Establish the mapping connection between entity nodes and relation edges in the knowledge graph based on the semantic association model; S403, map the first type of event identification result and the second type of event prediction result to event instance nodes in the knowledge graph; S404 analyzes the causal path between event instance nodes and environmental factor nodes through a causal reasoning model, and generates a causal reasoning chain. S405, Based on the causal reasoning chain, an explanatory description of the cause and evolution of the event is generated, forming an enhanced reasoning result.

[0071] In this embodiment, when constructing a knowledge graph that includes event types, environmental factors, and geographic entities, the conceptual system and relational framework for monitoring operations are first determined. Event types cover categories such as landslides and waterlogging, retaining sub-class levels. Environmental factors encompass continuous and discrete quantities such as precipitation, wind field, soil moisture content, and surface temperature. Geographic entities include pixels, land parcels, watersheds, administrative units, and topographic units. During the data access phase, spatiotemporal fusion data, the results of the first type of event identification, and the results of the second type of event prediction are uniformly stored on the disk, organized by time index, spatial index, and source index. Subsequently, entity extraction, attribute extraction, and relation identification are performed. Entity extraction generates nodes based on place names, land parcel identifiers, sensor sites, time slices, and event tags. Attribute extraction attaches numerical and state quantities to nodes. Relationship identification generates edges based on triggering, influence, adjacency, inclusion, and dependency relationships between events and environmental factors and geographic entities. To avoid duplicate names and multiple sources, entity alignment and disambiguation processes are introduced. Spatial overlap, temporal overlap, and contextual similarity are used to complete the merging, and source weights and timestamps are recorded in the graph to ensure version traceability and incremental updates. To improve global consistency, a lightweight ontology vocabulary and relation constraints are established to limit relation directions, domains, and value ranges, avoiding invalid edges and self-loops that pollute the structure. Graph data is finally stored in graph storage, which introduces node labels, edge types, attribute dictionaries, and quality identifiers to facilitate subsequent batch retrieval and batch processing inference.

[0072] When establishing the mapping connection between entity nodes and relation edges based on the semantic association model, a vectorized representation is generated for each type of node to carry semantic and structural information. The node vector consists of a name representation, a context representation, and a structural representation: the name representation comes from word segmentation and subword embedding, the context representation comes from attribute aggregation of the node's neighborhood, and the structural representation comes from random walks or subgraph encoding. The edge vector is jointly determined by the endpoint vector and the relation type, and the edge decision and confidence assessment are completed through a scoring function or matching function. To reduce the impact of noise input, source weights and time decay are introduced, and the edge threshold is dynamically adjusted by the stable occurrence frequency within the time window, cross-source consistency, and geographical consistency. Low-confidence candidate edges are retained as weak connections, awaiting subsequent evidence accumulation or manual verification. Through this mapping connection process, a dense and interpretable semantic skeleton can be formed in the graph, providing a basic carrier for subsequent path search and causal analysis.

[0073] When mapping the identification results of the first type of event and the prediction results of the second type of event to event instance nodes, the identification and prediction results are first standardized into unified instance records, including event category, occurrence or prediction time period, spatial range, intensity index, confidence level, and data source. The instantiation process creates event instance nodes within the knowledge graph and establishes location relationships with corresponding geographic entities, observation relationships with environmental factors, and temporal relationships with similar or successive events. For multiple outputs of the same spatial unit within adjacent time periods, a merging strategy is implemented to form inter-period instances, recording their duration and evolution direction to reduce path breaks caused by fragmented instances. For similar instances from different sources, fusion is performed based on spatial overlap, temporal overlap, and intensity compatibility, retaining multi-source contribution coefficients and conflict markers to facilitate subsequent weighted evaluation and anomaly auditing.

[0074] When analyzing causal paths between event instance nodes and environmental factor nodes using a causal reasoning model, candidate subgraphs are first defined in the graph, requiring that the temporal sequence satisfy the condition of cause preceding effect, and the spatial relationship satisfy adjacency, inclusion, or hydrothermal transport connectivity. Within the candidate subgraphs, path generation, path filtering, and direction determination are performed. In the path generation phase, a limited-depth multi-hop search is conducted between event instances and environmental factors, retaining links that meet temporal and spatial constraints. In the path filtering phase, each link is scored based on evidence from temporal consistency (whether the lag relationship is stable), strength consistency (whether the change in the dependent variable is in the same direction as the event strength or is explainable), multi-source consistency (common support for the same link from different sources), and background common sense consistency (whether it violates ontology constraints). In the direction determination phase, the direction of the edges is verified, and alternative paths are excluded when necessary using mediator and suppressor variables. After filtering and determination, a set of causal reasoning chains is obtained, each chain containing a node sequence, an edge type sequence, evidence weights, and uncertainty quantification.

[0075] When generating explanatory descriptions of the causes and evolution of events based on causal reasoning chains, the chain structure is first transformed into human-readable template-filling units, including starting factors, key intermediaries, target events, lag time windows, contribution weights, and spatial coverage. To avoid redundancy and repetition, the chains are sorted and deredundant according to the strength of evidence, retaining representative links and providing concise descriptions, along with quantitative summaries such as contribution percentage ranges, time lag ranges, and the spatial scope involved. For the same event instance, multi-granular outputs for management and decision-making are generated: an overview description is used to quickly present the dominant factors and main paths, while detailed descriptions are used to trace the source, check boundary conditions, and compare alternative paths; all descriptions and chain objects maintain traceable links to ensure smooth subsequent queries, verification, and version iterations. Explanatory descriptions and event instances are written back into the graph, forming self-enhancing knowledge assets that provide reusable evidence for subsequent iterative learning, parameter calibration, and strategy optimization.

[0076] This embodiment constructs a knowledge graph covering event types, environmental factors, and geographical entities. It establishes a stable entity-relationship mapping using semantic associations, and then instantiates and embeds the identification results of the first type of event and the prediction results of the second type of event into the graph structure. Causal reasoning can identify highly reliable causal chains while satisfying temporal and spatial constraints, and presents the causes and evolutionary patterns with a traceable and interpretable description. Thus, isolated identification and prediction outputs are linked into an interpretable reasoning network, dominant and mediating factors are supported by quantitative evidence, path uncertainty is explicitly characterized, and the directionality for subsequent early warning and response is enhanced. The overall decision-making chain is simultaneously improved in terms of consistency, interpretability, and reusability.

[0077] In one embodiment, after step S40 above, the method further includes: S406, Extract key causal relationship factors from the enhanced reasoning results; S407, the key causal relationship factors are fed back to the event recognition model and the spatiotemporal graph network model; S408, the key causal relationship factors are used to adjust the feature weights of the event recognition model and the attention mechanism of the spatiotemporal graph network model.

[0078] In this embodiment, the process of extracting key causal factors from the inference enhancement results focuses on identifying and aggregating the dominant paths, key mediators, and inhibitory effects within the causal chain. First, the causal chain set is constrained and filtered according to chronological order and spatial connectivity, retaining chains that satisfy the cause-effect sequence while removing chains that conflict with geographical boundaries or management units. Then, based on the evidence weights, frequency of occurrence, and cross-source consistency labels inherent in the inference enhancement results, event factors and environmental factors within the chains are aggregated to generate a factor list and annotations of the interaction directions between factors. To facilitate subsequent integration with model input, the factor list is organized in a mappable, structured form, encompassing source identifiers, interaction directions, time lag intervals, spatial coverage, and confidence labels, while maintaining a traceable index to the original chain. For factors with conflicting sources or divergent confidence levels, weak constraint markers are set, used only for lightweight modulation during the feedback process to avoid introducing excessive bias.

[0079] The process of feeding back key causal factors to the event recognition model and the spatiotemporal graph network model is built on a neutral factor adaptation interface. This interface maps factors to the model domain, including channel mapping, spatial mapping, and temporal mapping. Channel mapping establishes a connection between factors and multimodal feature sets, allowing observation channels such as vegetation, humidity, structural texture, and wind field to be pointed to by corresponding factors. Spatial mapping connects the scope of factors to pixel, plot, and node sets. Temporal mapping aligns lag intervals with the model's sliding window or sequence slices. After adaptation, two types of modulation instructions are generated: one for the event recognition model and one for the spatiotemporal graph network model, each with an intensity level and effective range. Before entering the model, the modulation instructions undergo consistency checks. If there is a conflict with the input quality indicator or data missing indicator, the intensity level is reduced or the effect is temporarily suspended to ensure stable feedback.

[0080] When adjusting the feature weights of the event recognition model using key causal factors, a combination of grouped weighting and intra-layer modulation is employed to avoid abrupt changes in the overall distribution. Grouped weighting generates weight vectors on the input side according to feature groups, increasing the weights of channels strongly correlated with the factors and decreasing the weights of channels unrelated to or conflicting with the factors. Intra-layer modulation applies gentle biases to the scale and offset parameters of the normalization and transformation units, allowing highly correlated evidence to achieve higher responses in deeper representations. Spatial consistency is maintained through adjacency aggregation units, smoothly transitioning weight modulation within adjacent pixels or the same plot. Temporal consistency is maintained through gradual variation constraints within a window, ensuring that weights change slowly along the sequence and avoiding instantaneous jumps. For factors with weak constraint markers, weak modulation is applied only in edge regions and low-confidence channels to preserve the model's adaptive capabilities.

[0081] When adjusting the attention mechanism of a spatiotemporal graph network model using key causal factors, biases are injected at three levels: node attention, edge attention, and temporal attention. Node attention maps the spatial coverage of factors to node-level priority, increasing the message reception intensity of nodes within the factor's influence range. Edge attention maps the direction of factor action to edge bias, strengthening edge connections aligned with water and heat transfer, wind direction, or watershed runoff, while weakening edges opposite to the causal direction or with insufficient evidence. Temporal attention maps lag intervals to temporal weights, increasing the weight of time slices matching factor lags and suppressing time slices inconsistent with the causal order. These three types of attention biases are additively or multiplicatively fused with the original attention distribution and modulated by input quality labels and uncertainty weights to amplify causal guidance when evidence is sufficient and revert to the model's own distribution when evidence is insufficient. After receiving the biases, the model performs an offline validation process, comparing node-level risk ranking, consistency indicators, and allergy indicators before and after injection. If abnormal drift occurs, it automatically rolls back to the previous validated bias version and records a snapshot of the differences.

[0082] This embodiment extracts key causal factors from the reasoning enhancement results and implements bounded feedback. The event recognition model and the spatiotemporal graph network model receive directional guidance oriented towards causes and paths. The weight allocation of input channels, spatial adjacency, and time windows more closely reflects the true causal and propagation relationships. Smooth modulation of weights and attention reduces sensitivity to noisy evidence, while offline verification and version registration ensure the adjustment process is controllable and traceable. As a result, the consistency and stability of the recognition and prediction results are simultaneously enhanced, the generalization ability across regions and time periods is improved, and false positives and false negatives decrease as the causal guidance converges. This provides more reliable and interpretable upstream input for subsequent anomaly analysis and response decisions.

[0083] In one embodiment, step S50 above includes: S501, Determine the comprehensive anomaly score of the monitoring area based on the inference enhancement results; S502, cluster analysis is performed on the monitoring area based on geographical proximity and similarity of abnormal characteristics to divide the warning area; S503, Based on the anomaly level of the warning area and the causal reasoning chain in the reasoning enhancement result, generate response strategies for different warning areas; S504, integrate the comprehensive anomaly score, early warning area division results and response strategy into the visual interactive interface to generate comprehensive anomaly analysis information and response decisions.

[0084] In this embodiment, when determining the comprehensive anomaly score of the monitoring area based on the inference enhancement results, an anomaly signature vector is constructed for each spatial unit. This vector includes a severity measure of the first type of event, the probability of occurrence of the second type of event, the contribution strength of key factors in the causal chain, duration indicators, spatial spread indicators, and uncertainty weights. First, robust standardization and dimensionless processing are performed on each dimension. Then, uncertainty is used as a modulation quantity to complete weighted aggregation, forming a comprehensive anomaly score with good discriminative power. To avoid spikes caused by sporadic noise, the scores are spatially smoothed using neighborhood consistency constraints, and temporally varied constraints are introduced to suppress jumps. For dimensions with missing measurements, continuity is restored using quality masks and interpolation strategies, and confidence decay is reflected in the scores. After score generation, calibration is performed to stabilize the hit rate and false alarm rate corresponding to different thresholds. Monotonic mapping or temperature scaling calibration procedures can be used to improve the usability and interpretability of the scores.

[0085] When delineating warning areas based on geographical proximity and anomaly feature similarity, spatial adjacency relationships are first established, derived from shared boundaries, topographic connectivity, river network runoff channels, or road frameworks. Then, a feature similarity metric is constructed using a comprehensive anomaly score and anomaly signature vector, taking into account magnitude, morphology, and persistence. Clustering employs a two-stage process: the spatial agglomeration stage merges connected components on the adjacency graph to ensure the geographical integrity of the region; the feature merging stage further subdivides or merges within each spatial candidate based on a similarity threshold, ensuring consistent anomaly phenotypes within the region. For patches that are too small or fragmented, a nearest neighbor merging strategy is used to avoid excessive fragmentation; for regions crossing administrative or natural boundaries, boundary constraints are retained to avoid unreasonable merging. Each warning area outputs its center location, boundary polygon, internal score distribution, and representative anomaly signature, which serve as input for subsequent strategy generation.

[0086] When generating response strategies based on the anomaly levels and causal reasoning chains of early warning areas, each area is first assigned a level label. The level mapping is based on the segmentation of the score distribution and the empirical quantiles of historical priors. Then, templates matching the causal chain are retrieved from the strategy template library. These templates are organized in a structure of "triggering factor—target object—action set—time limit—resource," covering various actions such as agronomic treatment, monitoring reinforcement, equipment maintenance, emergency allocation, and financial services. For areas containing multiple dominant factors, multiple templates are combined to provide action sets and priorities in parallel. The priority is determined based on the strength of causal contribution, time lag window, and resource accessibility. To suppress strategy oscillations, an activation threshold and minimum duration are introduced; low-confidence evidence triggers only observation enhancement or prudent actions, while high-confidence evidence triggers expansion actions and resource allocation. The strategy objects are ultimately solidified into an executable list, including execution windows, responsible units, resource allocation, and linkage conditions, along with a backtracking index for subsequent evaluation and iteration.

[0087] When integrating comprehensive anomaly scores, warning zones, and response strategies into a visual interactive interface, a three-tiered view system is constructed. The spatial view displays score distribution and level zoning using a base map overlaid with heatmaps and region boundaries, supporting zooming, panning, and multi-level switching. The diagnostic view displays key factors, path evidence, and lag windows along a causal chain, supporting chain expansion and evidence weight queries. The execution view displays action sets, resource lists, and timelines centered on a strategy list, supporting one-click deployment and status feedback. Objects in the three layers remain interconnected; selecting any object simultaneously highlights it in other views. All objects have version tags and timelines for easy review and auditing. The interface also provides export and subscription capabilities to meet cross-departmental collaboration and external delivery needs.

[0088] Example Description: In an agricultural disaster monitoring scenario, the system receives multi-source observation data covering the farmland monitoring area, including multispectral remote sensing images, meteorological station measured data, land surface temperature remote sensing products, soil moisture sensor data, and crop growth survey data. Through spatial registration and temporal alignment, a spatiotemporal fusion dataset is formed. This dataset covers farmland grid cells spatially and multiple temporal monitoring cycles temporally, thus constructing a foundation for agricultural ecological monitoring with continuous spatial representation and dynamic temporal characteristics.

[0089] When identifying disaster types, the system extracts multimodal difference features of the target plot before and after the disaster based on spatiotemporal fusion data to identify disaster events such as drought, waterlogging, and pests and diseases. Specifically, it extracts multimodal feature sets such as vegetation index (NDVI, EVI), land surface temperature (LST), soil moisture content, and leaf area index (LAI) from the fusion data before the disaster as baseline features for the normal growth stage; then it extracts the corresponding feature sets after the disaster and calculates feature difference mapping. The system classifies and spatially aggregates the difference mapping using a trained disaster identification model to identify abnormal areas. After identifying abnormal areas, the model further analyzes the severity of the disaster, such as assessing the drought level based on the decrease in NDVI and the deviation of soil moisture, and assessing the waterlogging level based on changes in water depth and duration, thus outputting disaster identification results that include disaster type, intensity level, and spatial range.

[0090] To predict crop growth degradation trends, the system utilizes dynamic temporal features from fused data to predict future disaster risks through a spatiotemporal graph network model. The model extracts temporal patterns related to crop growth degradation from multi-dimensional time series, including historical meteorological sequences, surface humidity, evapotranspiration, and vegetation index trends. Based on geographical proximity and topographic channels, the monitoring area is divided into spatial nodes, and a spatiotemporal dependency graph between these nodes is constructed. By capturing the environmental coupling relationships between spatially adjacent nodes, the model predicts the probability of future drought index anomalies or crop growth anomalies, outputting a future disaster risk distribution map, thus generating early warning results for potential disasters.

[0091] The system further constructs an agricultural disaster knowledge graph, including meteorological factor nodes (rainfall, temperature, wind speed), soil factor nodes (moisture content, organic matter, salinity), crop physiological nodes (photosynthetic efficiency, canopy temperature, growth index), and geographical entity nodes (irrigation canals, terrain slope, field boundaries, etc.). Relationships between entity nodes are established based on a semantic association model, such as "insufficient rainfall → soil drying" and "persistent high temperatures → increased evapotranspiration → soil moisture deficit → leaf wilting." Disaster identification results and growth degradation prediction results are mapped to disaster instance nodes. A causal reasoning model analyzes the paths between disaster nodes and environmental factor nodes, generating a causal reasoning chain containing causal chains and propagation paths. For example, in drought stress, the reasoning chain can be represented as "lack of rainfall → increased evapotranspiration → decreased soil moisture content → limited crop photosynthesis → decreased growth index," forming an interpretable and enhanced reasoning result.

[0092] The system extracts key causal factors from the enhanced inference results, such as "rainfall below the threshold for two consecutive weeks" or "surface temperature exceeding 35 degrees Celsius for more than five consecutive days," and feeds these factors back to the disaster identification model and the spatiotemporal map network model. Through this feedback mechanism, the system adjusts the distribution of feature weights in the identification model, for example, by increasing the weight of surface temperature change features or increasing the attention coefficient of nodes related to evapotranspiration anomalies in the spatiotemporal map network. This enables the model to automatically focus on key factors during future monitoring, improving its sensitivity and generalization ability to complex disaster scenarios.

[0093] Finally, the system generates comprehensive anomaly analysis information and response decisions based on the enhanced reasoning results. First, the system calculates a comprehensive anomaly score for the monitored area based on disaster identification results and causal reasoning chains. This score comprehensively reflects the degree of crop physiological degradation, the intensity of environmental stress, and the risk of spatial propagation. Subsequently, the system performs cluster analysis based on geographical proximity and anomaly characteristic similarity, dividing farmland areas into different warning zones, such as "mild drought zones," "severe waterlogging zones," and "potential disease propagation zones." For each warning zone, the system automatically generates response strategies based on the anomaly level and causal chain. For example, in drought zones, it recommends water-saving irrigation and mulching for moisture retention; in waterlogged zones, it recommends drainage and foliar application of physiological regulators; and in potential disease propagation zones, it recommends optimizing pesticide spraying coverage and adjusting agricultural timing. All generated anomaly scores, warning zoning results, and response strategies are presented in a unified visual interface. Users can intuitively view disaster distribution, risk levels, and treatment recommendations through a map view and can update the strategy execution status based on real-time feedback.

[0094] In fintech scenarios, the system receives input from various financial data sources, including real-time transaction records, account balance change information, market data, user behavior logs, credit application records, and external economic indicator data. The system first performs structured parsing, time synchronization, and outlier cleaning on this heterogeneous data to form a multi-dimensional spatiotemporal fusion dataset. In this dataset, the time dimension corresponds to transaction time series and market fluctuation cycles, while the spatial dimension corresponds to the distribution of account groups across different institutional nodes, geographical branches, or financial networks, thus constructing a multimodal data foundation that simultaneously reflects transaction behavior, capital flows, and risk propagation.

[0095] Based on this spatiotemporal fusion data, the system identifies potential risk events by extracting the differences in characteristics between target account groups before and after abnormal transactions. Specifically, it extracts multimodal feature sets from the fusion data, including transaction frequency distribution, average inter-account transfer amount, device fingerprint stability, and geographical location concentration before the event, and extracts corresponding feature sets after the event. The system calculates the feature difference vector between the two, reflecting the abrupt change patterns in transaction behavior and asset status. The event recognition model inputs the difference features into a multi-layer deep network structure for classification and aggregation, identifying high-risk target groups involved in abnormal transactions, account hijacking, or abnormal fund migration. The model further determines the risk type and severity level based on the offset of transaction amount distribution, device consistency disruption rate, and account linkage degree, generating recognition results containing risk event type and level labels.

[0096] Building upon this foundation, the system leverages the dynamic temporal characteristics of the fused data to predict potential systemic risk events using a spatiotemporal graph network model. The spatiotemporal graph network model first extracts price linkage sequences between market sectors, fund flow sequences between institutions, and behavioral covariance matrices of account groups from the fused data, constructing a spatiotemporal dependency graph reflecting the structure of the financial system. Each node represents an institution or account cluster, and the edge weights are determined by transaction frequency, fund flow intensity, and historical correlation. The model uses temporal convolution and graph convolution units to capture the propagation patterns of trading activities in time and space, outputting the probability of abnormal transactions and the distribution of risk clusters in future periods. The prediction results not only indicate the institutional nodes or customer groups where risks may occur but also reveal the direction of risk diffusion and its temporal lag.

[0097] The system further constructs a knowledge graph encompassing risk event types, market factors, and financial entities. This knowledge graph includes market nodes (indices, interest rates, exchange rates), behavioral nodes (transaction patterns, account clustering, flow paths), environmental nodes (policy changes, macroeconomic indicators), and financial institution nodes (banks, payment institutions, trust companies, etc.). Based on a semantic association model, it establishes relationship mappings between nodes, such as "sharp exchange rate fluctuations → surge in cross-border transactions → increased frequency of abnormal transfers → triggering of anti-money laundering alerts." The system maps risk event identification results and risk prediction results to risk instance nodes in the knowledge graph. Through a causal reasoning model, it analyzes the relationship paths between instance nodes and market environment nodes, generating causal reasoning chains, such as "rising macroeconomic interest rates → tight high-leverage funding chains → decreased liquidity → rising delinquency rates → expanded risk exposure." The enhanced reasoning results not only explain the sources of risk but also provide the evolutionary patterns and key propagation pathways of events.

[0098] After generating enhanced inference results, the system extracts key causal factors, such as "increased reliance on short-term financing," "abnormal cross-border capital outflows," and "transaction volume fluctuations exceeding historical thresholds," and feeds these back to the risk identification model and the spatiotemporal graph network model. Through this feedback mechanism, the identification model adjusts the weight structure of its input features, enhancing its sensitivity to behavioral anomalies; the spatiotemporal graph network model reconfigures the node attention distribution based on the weights of key factors, strengthening its ability to capture macro-level interconnected risks. This causal feedback forms an adaptive optimization closed loop, enabling the system to continuously learn the evolutionary patterns of risk and dynamically adjust its detection sensitivity and prediction confidence interval.

[0099] The system ultimately generates comprehensive anomaly analysis information and response decisions based on the enhanced inference results. First, it calculates a comprehensive risk score for each institutional node or account group based on the inference results. This score integrates the severity of identified risks, the predicted probability of risks, and the causal strength index in the inference chain. Then, the system performs cluster analysis based on the similarity of fund flow network topology and risk characteristics, aggregating high-risk nodes into risk groups, such as "cross-border high-frequency abnormal transfer groups," "leveraged institutional groups," and "suspected money laundering path chains." For each group, the system generates customized response strategies based on the risk level and causal inference chain, such as freezing high-risk account clusters, restricting abnormal fund flow paths, dynamically adjusting credit limits, or triggering internal audit processes. Finally, the comprehensive risk score, risk cluster structure, and response strategies are integrated into a visual interactive interface, forming an auditable and traceable risk situation map. This interface displays the risk heat map distribution, node linkage relationships, and strategy execution status of the financial system in real time, supporting collaborative decision-making and dynamic intervention by regulatory agencies, risk control departments, and business operation units.

[0100] This embodiment transforms the enhanced reasoning results into a quantifiable comprehensive anomaly score and completes calibration, unifying the expression of anomaly intensity, persistence, and spread, while suppressing sporadic noise in space and time. By dividing warning areas under the dual constraints of geographical proximity and anomaly phenotype, spatial integrity and phenotypic consistency are simultaneously ensured, avoiding excessive fragmentation or unreasonable merging. By generating response strategies based on anomaly levels and causal chains, the handling actions are targeted to the dominant factors and lag windows, and resource allocation has a sequence and an effective threshold. Through the integrated presentation and linkage of scores, regions, and strategies in a visual interactive interface, diagnosis, decision-making, and execution form a continuous closed loop.

[0101] In one embodiment, a multi-source monitoring analysis and decision-making apparatus is provided, which corresponds one-to-one with the multi-source monitoring analysis and decision-making methods described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-source monitoring analysis and decision-making device of the present invention. The modules include a multi-source monitoring data fusion module 10, an event identification module 20, an event prediction module 30, a causal reasoning enhancement module 40, and an analysis and response decision-making module 50. Detailed descriptions of each functional module are as follows: The multi-source monitoring data fusion module 10 is used to acquire multi-source monitoring data, perform spatiotemporal multi-scale fusion processing on the multi-source monitoring data, and generate spatiotemporal fused data. Event recognition module 20 is used to identify a first type of event and generate a first type of event recognition result by extracting the difference features of the target object before and after the event based on the spatiotemporal fusion data. Event prediction module 30 is used to predict the second type of event based on the dynamic temporal characteristics in the spatiotemporal fusion data and using a spatiotemporal graph network model, and generate the prediction result of the second type of event. The causal reasoning enhancement module 40 is used to construct a knowledge graph and use the knowledge graph to perform causal reasoning on the identification results of the first type of event and the prediction results of the second type of event to generate reasoning enhancement results; The analysis and response decision module 50 is used to generate comprehensive anomaly analysis information and response decisions based on the reasoning enhancement results.

[0102] In one embodiment, the multi-source monitoring data fusion module 10 is specifically used for: Acquire optical remote sensing data, synthetic aperture radar data, and ground sensor data to form multi-source monitoring data; The optical remote sensing data is used to detect areas obscured by clouds and fog, and the pixels affected by cloud and fog obscuration are identified. Based on spectral features and temporal variation models, the pixels affected by cloud and fog obstruction are repaired, and repaired optical remote sensing data is generated. The repaired optical remote sensing data, the synthetic aperture radar data, and the ground sensor data are spatiotemporally registered to form registered multi-source monitoring data. The registered multi-source monitoring data is input into a spatiotemporal multi-scale fusion network to extract the fusion features of spatial and temporal dimensions; Super-resolution reconstruction is performed on the fused features to generate spatiotemporal fused data.

[0103] In one embodiment, the event recognition module 20 is specifically used for: Extract the first multimodal feature set of the target object before the event occurs and the second multimodal feature set after the event occurs from the spatiotemporal fusion data; Determine the differences between the first multimodal feature set and the second multimodal feature set, and generate a difference feature set for the target object; The difference feature set is input into the event recognition model to identify the target object region where the event occurred; The severity of the event is analyzed in the target area where the event occurred, and a first-class event identification result containing the event type and severity level is generated.

[0104] In one embodiment, the event prediction module 30 is specifically used for: Extract multivariate time-series data related to the second type of event from the spatiotemporal fusion data; Based on geospatial relationships, the monitoring area is divided into multiple spatial nodes, and a spatiotemporal graph structure describing the relationships between spatial nodes is constructed. The multivariate time series data and the spatiotemporal graph structure are input into the spatiotemporal graph network model for processing, and the probability and anomaly distribution of the second type of event in the future time period are generated as the prediction result of the second type of event.

[0105] In one embodiment, the causal reasoning enhancement module 40 is specifically used for: Construct a knowledge graph that includes event types, environmental factors, and geographic entities; A mapping connection between entity nodes and relation edges in the knowledge graph is established based on a semantic association model; The first type of event identification results and the second type of event prediction results are mapped to event instance nodes in the knowledge graph; By analyzing the causal path between event instance nodes and environmental factor nodes using a causal reasoning model, a causal reasoning chain is generated. Based on the aforementioned causal reasoning chain, an explanatory description of the causes and evolution of events is generated, forming an enhanced reasoning result.

[0106] In one embodiment, the causal reasoning enhancement module 40 is specifically used for: Extract key causal factors from the enhanced reasoning results; The key causal factors are fed back to the event recognition model and the spatiotemporal graph network model; The key causal relationship factors are used to adjust the feature weights of the event recognition model and the attention mechanism of the spatiotemporal graph network model.

[0107] In one embodiment, the analysis and response decision module 50 is specifically used for: The comprehensive anomaly score of the monitoring area is determined based on the enhanced reasoning results. Cluster analysis of the monitoring area is performed based on geographical proximity and similarity of abnormal characteristics to divide the warning area; Based on the anomaly level of the warning area and the causal reasoning chain in the enhanced reasoning results, a response strategy is generated for different warning areas. The comprehensive anomaly score, early warning area division results, and response strategies are integrated into a visual interactive interface to generate comprehensive anomaly analysis information and response decisions.

[0108] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a multi-source monitoring analysis and decision-making method on the server side.

[0109] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of a multi-source monitoring analysis and decision-making method.

[0110] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire multi-source monitoring data, and perform spatiotemporal multi-scale fusion processing on the multi-source monitoring data to generate spatiotemporal fused data; Based on the spatiotemporal fusion data, by extracting the difference features of the target object before and after the event, the first type of event is identified, and the first type of event identification result is generated. Based on the dynamic temporal characteristics in the spatiotemporal fusion data, a spatiotemporal graph network model is used to predict the second type of event and generate the prediction result of the second type of event. Construct a knowledge graph and use the knowledge graph to perform causal reasoning on the identification results of the first type of event and the prediction results of the second type of event to generate reasoning-enhanced results; Based on the enhanced reasoning results, comprehensive anomaly analysis information and response decisions are generated.

[0111] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire multi-source monitoring data, and perform spatiotemporal multi-scale fusion processing on the multi-source monitoring data to generate spatiotemporal fused data; Based on the spatiotemporal fusion data, by extracting the difference features of the target object before and after the event, the first type of event is identified, and the first type of event identification result is generated. Based on the dynamic temporal characteristics in the spatiotemporal fusion data, a spatiotemporal graph network model is used to predict the second type of event and generate the prediction result of the second type of event. Construct a knowledge graph and use the knowledge graph to perform causal reasoning on the identification results of the first type of event and the prediction results of the second type of event to generate reasoning-enhanced results; Based on the enhanced reasoning results, comprehensive anomaly analysis information and response decisions are generated.

[0112] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0113] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0115] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method of analysis and decision making for multi-source monitoring, characterized in that, The method comprises the following steps: Obtaining multi-source monitoring data, performing spatio-temporal multi-scale fusion processing on the multi-source monitoring data, and generating spatio-temporal fusion data; Based on the spatio-temporal fusion data, the difference features of the target object before and after the event are extracted to identify the first type of event, and the first type of event identification result is generated; Based on the dynamic time sequence features in the spatio-temporal fusion data, a spatio-temporal graph network model is used to predict the second type of event, and a second type of event prediction result is generated; Constructing a knowledge graph, and using the knowledge graph to perform causal reasoning on the first type of event identification result and the second type of event prediction result, and generating a reasoning enhancement result; According to the reasoning enhancement result, comprehensive abnormal analysis information and response decision are generated.

2. The multi-source monitored analytics and decision method of claim 1, wherein, Obtaining multi-source monitoring data, performing spatio-temporal multi-scale fusion processing on the multi-source monitoring data, and generating spatio-temporal fusion data, comprising: Obtaining optical remote sensing data, synthetic aperture radar data and ground sensor data to form multi-source monitoring data; Detecting cloud and fog shadowing areas on the optical remote sensing data to identify pixels affected by cloud and fog shadowing; Based on spectral features and time sequence change models, the pixels affected by cloud and fog shadowing are repaired to generate repaired optical remote sensing data; The repaired optical remote sensing data, the synthetic aperture radar data and the ground sensor data are spatio-temporally registered to form registered multi-source monitoring data; The registered multi-source monitoring data is input into a spatio-temporal multi-scale fusion network to extract fusion features in spatial and temporal dimensions; The fusion features are super-resolution reconstructed to generate spatio-temporal fusion data.

3. The multi-source monitored analytics and decision method of claim 1, wherein, Based on the spatio-temporal fusion data, the difference features of the target object before and after the event are extracted to identify the first type of event, and the first type of event identification result is generated, comprising: From the spatio-temporal fusion data, a first multi-modal feature set before the event occurs and a second multi-modal feature set after the event occurs are extracted from the target object; Determine the difference between the first multi-modal feature set and the second multi-modal feature set to generate a difference feature set of the target object; The difference feature set is input into an event identification model to identify the target object area where the event occurs; The target object area where the event occurs is analyzed for event severity to generate a first type of event identification result containing event type and severity level.

4. The multi-source monitored analytics and decision method of claim 1, wherein, Based on the dynamic time sequence features in the spatio-temporal fusion data, a spatio-temporal graph network model is used to predict the second type of event, and a second type of event prediction result is generated, comprising: From the spatio-temporal fusion data, multi-variable time sequence data related to the second type of event is extracted; Based on the geographical spatial relationship, the monitoring area is divided into multiple spatial nodes, and a spatio-temporal graph structure describing the association between the spatial nodes is constructed; The multi-variable time sequence data and the spatio-temporal graph structure are input into a spatio-temporal graph network model for processing, and the probability of occurrence of the second type of event and the abnormal distribution in the future period are generated as the second type of event prediction result.

5. The multi-source monitored analytics and decision method of claim 1, wherein, Constructing a knowledge graph, and using the knowledge graph to perform causal reasoning on the first type of event identification result and the second type of event prediction result, and generating a reasoning enhancement result, comprising: constructing a knowledge graph comprising event types, environmental factors, and geographical entities; establishing mapping connections between entity nodes and relationship edges in the knowledge graph based on a semantic association model; mapping the first type of event recognition results and the second type of event prediction results into event instance nodes in the knowledge graph; generating causal reasoning chains by analyzing causal paths between event instance nodes and environmental factor nodes through a causal reasoning model; generating explanatory descriptions of event causes and evolution rules based on the causal reasoning chains to form reasoning enhancement results.

6. The multi-source monitored analytics and decision method of claim 1, wherein, After constructing a knowledge graph and using the knowledge graph to perform causal reasoning on the first type of event recognition results and the second type of event prediction results to generate reasoning enhancement results, the method further includes: extracting key causal relationship factors from the reasoning enhancement results; feeding the key causal relationship factors back to the event recognition model and the spatio-temporal graph network model; adjusting feature weights of the event recognition model and attention mechanisms of the spatio-temporal graph network model using the key causal relationship factors.

7. The multi-source monitored analytics and decision method of claim 1, wherein, Based on the reasoning enhancement results, generating comprehensive anomaly analysis information and response decisions includes: determining a comprehensive anomaly score of a monitoring area based on the reasoning enhancement results; performing clustering analysis on monitoring areas based on geographical proximity and anomaly characteristic similarity to divide the monitoring areas into warning areas; generating response strategies for different warning areas based on anomaly levels of the warning areas and causal reasoning chains in the reasoning enhancement results; integrating the comprehensive anomaly score, the warning area division results, and the response strategies into a visual interactive interface to generate comprehensive anomaly analysis information and response decisions.

8. An analysis and decision device for multi-source monitoring, characterized by The multi-source monitoring analysis and decision-making device includes: a multi-source monitoring data fusion module configured to obtain multi-source monitoring data, perform spatio-temporal multi-scale fusion processing on the multi-source monitoring data, and generate spatio-temporal fusion data; an event recognition module configured to recognize a first type of event by extracting difference features of a target object before and after an event based on the spatio-temporal fusion data, and generate first type of event recognition results; an event prediction module configured to predict a second type of event using a spatio-temporal graph network model based on dynamic time series features in the spatio-temporal fusion data, and generate second type of event prediction results; a causal reasoning enhancement module configured to construct a knowledge graph, and use the knowledge graph to perform causal reasoning on the first type of event recognition results and the second type of event prediction results to generate reasoning enhancement results; an analysis and response decision-making module configured to generate comprehensive anomaly analysis information and response decisions based on the reasoning enhancement results.

9. A computer device, comprising: The computer device includes a memory, a processor, and a multi-source monitoring analysis and decision-making program stored on the memory and executable on the processor, and the multi-source monitoring analysis and decision-making program, when executed by the processor, implements the steps of the multi-source monitoring analysis and decision-making method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a multi-source monitoring analysis and decision-making program, and the multi-source monitoring analysis and decision-making program, when executed by a processor, implements the steps of the multi-source monitoring analysis and decision-making method of any one of claims 1-7.

Images