A multi-modal geospatial data large model situation analysis method and system

Abstract

The present application relates to the technical field of geospatial data analysis, and discloses a multi-modal geospatial data large model situation analysis method and system, the method comprising: processing multi-modal geospatial data to extract hierarchical causal variables; applying a causal discovery algorithm to construct a hierarchical causal world model containing causal confidence; driving the model to perform dynamic simulation to generate a future state prediction with uncertainty; calculating the probability distance between the prediction and the newly observed real state to quantify cognitive dissonance; when the cognitive dissonance exceeds the preset threshold, a large language model agent generates a correction hypothesis, which is used to update the world model and re-simulate, and output the corrected situation analysis result. By constructing a hierarchical causal world model and introducing a cognitive dissonance-driven self-correction mechanism, the present application improves the interpretability, reliability and long-term adaptive ability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geospatial data analysis, in particular to a multi-modal geospatial data large model situation analysis method and system. BACKGROUND

[0002] With the rapid development of remote sensing satellites, the Internet of Things, and geographic information systems, a large amount of multi-modal geospatial data is accumulating at an unprecedented speed, providing a data foundation for deep understanding and managing complex dynamic systems on the Earth's surface, such as regional ecological environment, agricultural production, and urban operation. How to effectively mine knowledge from these data and establish analysis models that can explain and predict the evolution of the system is an important technical challenge in this field.

[0003] In the field of geospatial situation analysis, existing technologies usually use models based on machine learning or deep learning. These methods use large-scale historical data sets, such as time-series satellite images or sensor readings, to train complex nonlinear models, such as long short-term memory networks (LSTM) or convolutional neural networks (CNN). The model learns statistical laws and correlation patterns existing in the data during the training process, and then uses the learned patterns to predict the future state of the system. For example, a model can be trained to predict the regional drought level in the next week based on the soil moisture and vegetation index data of the past few weeks.

[0004] Although existing technologies can make predictions based on historical data to some extent, there are still some deficiencies:

[0005] The analysis models constructed by existing technologies generally lack explainability and quantification of uncertainty. The fundamental reason is that deep learning models are essentially highly complex function fitters, and the internal parameter optimization process aims to minimize the error between the predicted output and the true label, rather than building a causal model that conforms to physical or logical laws. This leads to the fact that even if the model can give accurate predictions, it cannot reveal which micro factors (such as irrigation behavior of a specific plot) drive the formation of macro situations (such as regional drought) through what path, thereby limiting the depth of understanding of complex system emergent phenomena. At the same time, these models usually output a single deterministic prediction value, and the design goal does not include the evaluation of the confidence of the prediction result itself. This paradigm does not systematically consider and propagate the risks derived from data noise and model structural uncertainty, so it cannot provide quantitative risk assessment for decision-making, limiting the reliability of the prediction result in critical applications.

[0006] In addition, the static nature of existing models makes it difficult to adapt to dynamically changing environments, leading to a decline in accuracy over time. This is because once the existing model is trained and deployed, its internal parameters and structure are usually fixed. However, geospatial systems are inherently non-stationary, and their inherent operating rules can change due to natural evolution or human intervention (such as new water conservancy projects, changes in agricultural policies). Existing models lack an internal, active feedback and correction mechanism, and cannot continuously compare their prediction results with new real-world observations to autonomously detect and diagnose the "cognitive dissonance" between the model and reality. This "open-loop" working mode makes the model unable to evolve and adapt itself when new data appears, and its representation of the real world will gradually become outdated over time, leading to a continuous decline in prediction performance. SUMMARY

[0007] To overcome the shortcomings of the prior art, the present application provides a multi-modal geospatial data large model situation analysis method and system, which solves the problem that existing geospatial analysis models cannot explain their prediction logic or adapt to environmental changes when running as static "black boxes".

[0008] To achieve the above purpose, the present application is implemented by the following technical solutions:

[0009] The present application provides a multi-modal geospatial data large model situation analysis method in the first aspect, which comprises the following steps:

[0010] One or more pre-trained geospatial visual base models are used to process multi-modal geospatial data to extract hierarchical causal variables containing micro-level variables and macro-level variables;

[0011] A causal discovery algorithm is applied to analyze the time series of the hierarchical causal variables to build a hierarchical causal world model containing causal relationships and causal confidence;

[0012] A state transition function is used to drive the hierarchical causal world model to perform dynamic process simulation, evolving the probability distribution of system state from the current time to the future time to generate a future state with prediction uncertainty;

[0013] The probability distance between the newly observed real state in the multi-modal geospatial data and the probability distribution of the future state is calculated to quantify the cognitive dissonance representing the difference between the two;

[0014] A preset threshold is designed, and when the cognitive dissonance exceeds the preset threshold, a large language model agent is activated to generate a set of revised hypotheses, the hierarchical causal world model is updated according to the revised hypotheses, and the dynamic process simulation is performed again by using the updated hierarchical causal world model to generate a revised future state as a new situation analysis result.

[0015] Preferably, the step of extracting hierarchical causal variables including micro-level variables and macro-level variables specifically comprises: identifying geographical entities and geographical entity states from the multi-modal geospatial data as the micro-level variables; generating a macro-level variable by aggregating the states of a set of micro-level variables within a specified spatio-temporal neighborhood through an aggregation function; and the hierarchical causal world model includes modeling of causal relationships between the micro-level variables, between the macro-level variables, and across hierarchical variables.

[0016] Preferably, in the step of constructing a hierarchical causal world model including causal relationships and causal confidence, the process of assigning the causal confidence to each causal relationship comprises:

[0017] The causal confidence is calculated according to the following formula

[0018]

[0019] In the formula, represents a causal relationship edge; is the causal confidence of the causal relationship edge; is a fusion function; is a statistical significance indicator of the causal relationship; is a goodness-of-fit score based on historical data; is a prior weight from a domain knowledge base.

[0020] Preferably, the step of performing dynamic process simulation specifically comprises:

[0021] The current state of the system is represented as a probability distribution including a mean vector and a covariance matrix;

[0022] The probability distribution of the next time state is predicted according to the probability distribution of the current state and the hierarchical causal world model through the state transition function;

[0023] In the prediction process, the model uncertainty represented by the causal confidence is propagated into the covariance matrix of the next time state.

[0024] Preferably, the step of quantifying the cognitive dissonance representing the difference between the two specifically comprises:

[0025] ​​The Mahalanobis distance between the two is calculated to quantify the cognitive dissonance :

[0026] ;

[0027] wherein, is the quantified value of the cognitive dissonance; is the newly observed true state vector; is the mean vector of the probability distribution of the future state; is the covariance matrix of the probability distribution of the future state; superscript is the matrix transpose operation; superscript is the matrix inverse operation.

[0028] Preferably, the step of activating a large language model agent to generate a set of revised hypotheses further comprises:

[0029] analyzing, by the large language model agent, causal paths leading to the prediction bias, and generating the set of revised hypotheses;

[0030] designing and dispatching, by the large language model agent, one or more active probing tasks autonomously for verifying the set of revised hypotheses.

[0031] Preferably, the step of updating the hierarchical causal world model according to the revised hypotheses specifically comprises:

[0032] obtaining new data returned by the active probing tasks;

[0033] updating the posterior belief on the set of revised hypotheses according to the new data and Bayesian inference;

[0034] selecting the revised hypothesis with the highest posterior belief, and updating the causal relationship or causal confidence of the hierarchical causal world model accordingly.

[0035] Preferably, the method further comprises:

[0036] continuously monitoring the simulation prediction results of the macro-level variables;

[0037] and identifying an emergent effect and generating an early warning when it is predicted that the value of a certain macro-level variable will exceed a predefined critical threshold in the future.

[0038] Preferably, the method further comprises the step of conducting counterfactual reasoning:

[0039] receiving a user input query and parsing the user input query into an external intervention action vector;

[0040] running a new dynamic process simulation in the hierarchical causal world model with the external intervention action vector as input;

[0041] outputting the simulation result containing prediction uncertainty as a response to the query.

[0042] Correspondingly, the present application also provides a multi-modal geospatial data large model situation analysis system, which is used to perform any one of the above-mentioned multi-modal geospatial data large model situation analysis methods, and the system comprises:

[0043] a variable extraction module configured with one or more pre-trained geospatial visual base models for processing multi-modal geospatial data and extracting hierarchical causal variables containing micro-level variables and macro-level variables;

[0044] a world model construction module configured with a causal discovery algorithm for analyzing time series of the hierarchical causal variables and constructing a hierarchical causal world model containing causal relationships and causal confidence;

[0045] a dynamic simulation module configured with a state transition function for driving the hierarchical causal world model to perform dynamic process simulation, evolving the probability distribution of system state from the current time to the future time, and generating a future state with prediction uncertainty;

[0046] a cognitive dissonance detection module for calculating the probability distance between the newly observed real state in the multi-modal geospatial data and the probability distribution of the future state, quantifying the cognitive dissonance representing the difference between the two;

[0047] a model correction module configured with a large language model agent for generating a set of correction hypotheses when the cognitive dissonance exceeds a preset threshold, updating the hierarchical causal world model according to the correction hypotheses, and providing the updated hierarchical causal world model to the dynamic simulation module for subsequent situation analysis.

[0048] The present application provides a multi-modal geospatial data large model situation analysis method and system. The present application has the following beneficial effects:

[0049] 1、The present application constructs a hierarchical causal world model containing micro-level variables and macro-level variables, and explicitly mines and quantifies the causal relationships between variables, rather than only performing correlation statistics. This enables the system not only to predict changes in macro-situation, but also to trace the macro-emergence effect to its fundamental micro-level driving factors, providing a clear and logical causal chain for the analysis results, thereby significantly enhancing the depth of understanding and interpretation ability of the dynamic evolution process of complex geospatial systems.

[0050] 2、The application uses the constructed hierarchical causal world model to perform dynamic process simulation, and can prospectively deduce the future evolution path of the geospatial situation. More importantly, in the simulation process, the system quantitatively propagates the risks derived from data noise and uncertainty of causal relationships through a covariance matrix. This enables each prediction result to have an explicit confidence evaluation, so that the system can actively warn of emerging effects that may break through the critical threshold based on probability, providing more reliable and measurable decision-making basis than traditional methods.

[0051] 3、The application introduces a self-correction mechanism based on cognitive dissonance. By continuously comparing model predictions with new real observations, the system can autonomously identify the deviation between its internal world model and the real world. Once the deviation exceeds a threshold, the system can correct and improve its causal model through a series of operations such as generating correction hypotheses and dispatching active detection tasks. This closed-loop learning and iteration mechanism enables the model to continuously learn from new data, adapt to dynamic changes in the environment, and thus ensure its long-term application accuracy and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a schematic diagram of the system structure of the application;

[0053] Figure 2 is a schematic diagram of the method flow of the application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the application will be described below in conjunction with the drawings in the specification of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0055] As shown in Figure 1 and Figure 2 , the application provides a multi-modal geospatial data large model situation analysis method, which comprises the following steps:

[0056] Step 1: Multi-modal perception and hierarchical causal variable extraction.

[0057] This step aims to convert multi-source, heterogeneous raw geospatial data streams into structured, multi-scale, and time series variables that can be used for causal modeling.

[0058] Firstly, the system receives multi-modal geospatial data, including but not limited to time-series satellite remote sensing imagery, UAV video stream, GIS vector data, ground IoT sensor data, and web text information. One or more pre-trained geospatial vision base models and natural language processing models process the received data to identify geospatial entities and their state attributes at different time instants.

[0059] Then, the system generates hierarchical causal variables from the identified entity state attributes according to a pre-defined geospatial ontology and scale rules. The hierarchical causal variables include:

[0060] Micro-level variables : directly correspond to the quantified state of the th geospatial entity at time instant , such as the soil moisture of a specific plot or the pollutant emission rate of a factory.

[0061] Macro-level variables : represent the aggregated state of the th region at time instant , generated by a set of spatially correlated micro-level variables through an aggregation function . For example, the total pollutant load of a region is the weighted sum of the pollutant emission rates of all factories within the region. This process can be represented as:

[0062] ;

[0063] where is the set of micro-entity indices that define the macro-region ; is the th macro-level variable of the region at time instant ; is the aggregation function; is the th micro-level variable of the geospatial entity at time instant .

[0064] Finally, the state of the system at each time instant is composed of all micro-level variables and macro-level variables into a comprehensive state vector .

[0065] Step two: Construction of the hierarchical causal spatio-temporal world model.

[0066] The core of this step is to learn and construct a hierarchical causal world model that can represent the dynamic mechanisms of the real world based on historical state vector time series data.

[0067] First, within the micro and macro layers, the system applies a causal discovery algorithm (e.g., constraint-based PC algorithm or score-based GES algorithm) to process the time series of historical state vectors, mine the causal relationships among variables, and form a directed acyclic graph as the causal skeleton.

[0068] Then, to quantify the uncertainty of the model, the system assigns a dynamic causal confidence to each causal relationship edge in the causal graph The causal confidence is computed by a fusion function

[0069]

[0070] where represents a causal relationship edge; is the causal confidence of the causal relationship edge; is a fusion function; is a statistical significance indicator of the causal relationship; is a goodness-of-fit score based on historical data; is a prior weight from a domain knowledge base.

[0071] In addition, the system also learns and establishes cross-scale causal connectors between variables at different levels through models such as graph neural networks, to represent how micro events aggregate to cause macro state changes.

[0072] Step three: dynamic process simulation based on uncertainty propagation.

[0073] This step uses the constructed hierarchical causal world model to perform forward-looking simulation and deduction of the system's future state.

[0074] The system models the state at any time as a multi-dimensional Gaussian probability distribution, defined by the mean vector and the covariance matrix . Through a state transition function , according to the current state probability distribution and the hierarchical causal world model , the state probability distribution at the next time is predicted:

[0075]

[0076] where is the mean vector of the state probability distribution of the system at future time ; and is the covariance matrix of the state probability distribution of the system at future time .​​​​ is a state transition function; is the mean vector of the state probability distribution of the system at the current time instant ; is the covariance matrix of the state probability distribution of the system at the current time instant ; is an external intervention action vector applied at time instant ; is the hierarchical causal world model at time instant .

[0077] During the simulation process, the uncertainty is explicitly treated and propagated.

[0078] In particular, the calculation of the covariance matrix of the next time instant depends not only on the at the current time instant, but also on the confidence of each causal edge in the hierarchical causal world model. A causal path with lower confidence will result in a larger variance in the prediction result, so that each prediction result carries a quantitative uncertainty evaluation.

[0079] Step four: self-correction based on cognitive dissonance.

[0080] This step enables the method to learn from prediction errors and self-improve, which is the core of adaptive optimization of the situation analysis method.

[0081] First, the system compares the newly observed real-world state vector in step one with the mean vector of the predicted state at the corresponding time instant in step three. The cognitive dissonance is quantified by calculating the Mahalanobis distance between the two:

[0082] ;

[0083] where is the quantification value of the cognitive dissonance; is the newly observed real state vector; is the mean vector of the probability distribution of the future state; is the covariance matrix of the probability distribution of the future state; the superscript is the matrix transpose operation; the superscript is the matrix inverse operation.

[0084] When the value of cognitive dissonance exceeds a preset statistical threshold When the cognitive dissonance is detected, a revision process is initiated, which includes a large language model agent. In the process, the large language model agent performs the following operations:

[0085] Backtracking the causal path leading to the prediction bias, and making causal attribution;

[0086] Generating a set of revision hypotheses about the cause of the dissonance , such as parameter drift or the existence of unobserved confounding variables;

[0087] To verify the revision hypotheses, one or more active probing tasks are autonomously designed and dispatched to request new observation data .

[0088] After obtaining new data, the system applies Bayesian inference to update the posterior belief of each hypothesis, and selects the most credible hypothesis to guide the update of the structure or causal confidence of the hierarchical causal world model.

[0089] Step five: generate the situation analysis result.

[0090] This step is the final output link of the method. After updating the model in step four, the method uses the updated hierarchical causal world model to return to step three to re-perform dynamic process simulation, generate a revised future state that is closer to the evolution trend of the real world, and output this revised future state as a new situation analysis result.

[0091] In addition, the method can also continuously monitor the simulation prediction results of macro-level variables. When it is predicted that the value of a macro variable will break through the pre-defined critical threshold, it is identified as an emergent effect and a warning report is generated.

[0092] Please refer to the attached Figure 1 , Figure 1 is a system structure schematic diagram of a multi-modal geospatial data large model situation analysis system according to an embodiment of the present application. The embodiment of the present application provides a multi-modal geospatial data large model situation analysis system, which is implemented on a computing device. The computing device includes a processor, a memory, and a network interface for communicating with other devices. The memory stores a computer program, and the processor executes the computer program to implement the various functional modules and methods of the system described below.

[0093] The system specifically includes a variable extraction module 10, a world model construction module 20, a dynamic simulation module 30, a cognitive dissonance detection module 40, and a model revision module 50.

[0094] The variable extraction module 10 is configured to receive external multi-modal geospatial data through a network interface, including time-series satellite remote sensing images, geographic information system vector data, and ground sensor data. The variable extraction module 10 is internally configured with one or more pre-trained geospatial visual base models to process the multi-modal geospatial data and extract structured hierarchical causal variables containing micro-level variables and macro-level variables therefrom. After processing, the variable extraction module 10 outputs the generated time-series data of the hierarchical causal variables.

[0095] The world model construction module 20 is communicatively connected to the variable extraction module 10 to receive the time-series data of the hierarchical causal variables output by the variable extraction module 10. The world model construction module 20 is internally configured with a causal discovery algorithm library and a confidence calculation unit. The world model construction module 20 executes the causal discovery algorithm to analyze the time-series data to establish a causal relationship graph between variables, and calls the confidence calculation unit to assign a causal confidence to each causal relationship, thereby constructing and maintaining a hierarchical causal world model. The constructed hierarchical causal world model is output to the dynamic simulation module 30.

[0096] The dynamic simulation module 30 is communicatively connected to the world model construction module 20 to receive the hierarchical causal world model. The dynamic simulation module 30 is internally configured with a state transition function solver and an uncertainty propagation unit. The dynamic simulation module 30 drives the hierarchical causal world model to perform dynamic process simulation according to a probability distribution of a current state, evolves the system state from the current time to the future time, and uses the uncertainty propagation unit to convert the causal confidence in the model into the covariance of the prediction result, finally generates a future state with prediction uncertainty. The future state is output to the cognitive dissonance detection module 40.

[0097] The cognitive dissonance detection module 40 is communicatively connected to the variable extraction module 10 and the dynamic simulation module 30. The cognitive dissonance detection module 40 obtains a real state newly observed from the variable extraction module 10, and obtains a predicted future state from the dynamic simulation module 30. The cognitive dissonance detection module 40 is internally configured with a probability distance calculation unit and a threshold comparator, which are used to calculate the probability distance between the probability distributions of the real state and the future state, quantify the cognitive dissonance representing the difference between the two, and compare the quantified cognitive dissonance value with a preset threshold. When the cognitive dissonance value exceeds the preset threshold, the cognitive dissonance detection module 40 outputs a trigger signal.

[0098] The model correction module 50 is communicatively connected with the cognitive imbalance detection module 40 to receive the trigger signal. The model correction module 50 is internally configured with a large language model agent, an active probing task generator, and a Bayesian inference unit. Upon receiving the trigger signal, the large language model agent analyzes the causal path leading to the prediction bias and generates a set of correction hypotheses;

[0099] The active probing task generator generates specific data collection instructions based on the correction hypotheses; upon obtaining new data, the Bayesian inference unit verifies the hypotheses and selects the optimal hypothesis to generate an update instruction for the hierarchical causal world model.

[0100] The model correction module 50 is further communicatively connected with the world model construction module 20 and the dynamic simulation module 30. The model correction module 50 sends the generated update instruction to the world model construction module 20 to update the hierarchical causal world model. At the same time, the model correction module 50 prompts the dynamic simulation module 30 to call the updated hierarchical causal world model to re-perform dynamic process simulation and generate a corrected future state as a new situation analysis result output.

[0101] Please refer to the accompanying Figure 2 , Figure 2 is a method flowchart of a multi-modal geospatial data large model situation analysis method according to an embodiment of the present application. In a specific embodiment, the first stage of the method flow of the present application is multi-modal perception and hierarchical causal variable extraction, which is performed by the variable extraction module 10 in the system.

[0102] First, the variable extraction module 10 receives geospatial data of two different modalities:

[0103] One is time-series multi-spectral remote sensing images from Landsat 8;

[0104] The other is time-series soil moisture data collected by ground Internet of Things sensors deployed in a specific geographic area.

[0105] After receiving the data, the variable extraction module 10 pre-processes the data of different modalities. For remote sensing images, the module performs radiometric calibration and atmospheric correction to convert the original digital quantization values (DN) to surface reflectance, eliminating sensor errors and atmospheric effects. For ground sensor data, the module time-re-samples the collected hourly soil moisture readings by calculating daily averages to align their time resolution with the revisit period of the remote sensing images.

[0106] After preprocessing, the variable extraction module 10 begins generating hierarchical causal variables. Micro-level variables are defined as the quantized state directly corresponding to a single geographic entity. In this embodiment, for each 30-meter resolution image pixel... The Normalized Difference Vegetation Index (NDVI) was calculated using its surface reflectance data at time [time value missing]. The NDVI value of that pixel constitutes a micro-level variable. .

[0107] Macro-level variables are defined as the aggregated state of a set of micro-level variables within a specified spatiotemporal neighborhood. In this embodiment, the system loads a geographic information system vector layer that defines administrative divisions (e.g., townships). For each township area... Through aggregation functions Calculate the aggregated values ​​of the microscopic layer variables of all pixels within the region to generate the region at time [time value missing]. Macro-level variables The process is represented as:

[0108] ;

[0109] In the formula, To define the macro region A collection of micro-entity indexes; For the first Each region at time Macro-level variables; It is an aggregate function; For the first A geographical entity at time Micro-level variables.

[0110] Finally, the variable extraction module 10 integrates and aligns the time series data of all micro-level variables (NDVI values ​​of each pixel) and macro-level variables (average NDVI values ​​of each township) to form a unified, multi-scale time series dataset, and outputs it to the world model construction module 20.

[0111] In the flowchart of the method of the present invention, the next stage following variable extraction is the construction of a hierarchical causal world model. This stage is performed by the world model construction module 20 in the system, which receives a time series dataset containing multi-scale variables from the output of the variable extraction module 10.

[0112] First, the world model building module 20 calls its internal causal discovery algorithm library and selects the constraint-based PC algorithm to process the received time series dataset to discover the causal structure between variables. The PC algorithm performs a series of conditional independence tests to gradually eliminate indirect associations between variable pairs, ultimately generating a directed acyclic graph as the skeleton of the hierarchical causal world model. This graph contains causal relationships between variables at the micro level, between variables at the macro level, and between variables spanning both micro and macro levels.

[0113] After obtaining the causal structure skeleton, the confidence calculation unit of the world model building module 20 is each causal relationship edge in the graph. Calculate a causal confidence score This calculation is performed through a fusion function. Complete. In this embodiment, the fusion function is a weighted average function. The specific calculation formula is as follows:

[0114] ;

[0115] In the formula, Represents a causal relationship edge; The causal confidence level of the edges of a causal relationship; For a fusion function; This serves as a statistical significance indicator for causal relationships. The goodness-of-fit score is based on historical data; These are the prior weights derived from the domain knowledge base.

[0116] The specific methods for obtaining the parameters in the formula are as follows: The final conditional independence test performed by the PC algorithm when determining this causal edge. Value export, specifically . By establishing relationships It is obtained by fitting a linear regression model and calculating its coefficient of determination (R-squared value). It is obtained by querying a pre-built domain knowledge base; for example, if the knowledge base records variables. The regions and variables represented If the region it represents has known physical connections (such as irrigation canals), then the causal edge is assigned a higher prior weight value.

[0117] Finally, the world model building module 20 encapsulates the causal relationship graph with causal confidence weights into a complete hierarchical causal world model. It is then output to the dynamic simulation module 30.

[0118] In the flowchart of the method of this invention, the next stage following the world model construction is dynamic process simulation. This stage is executed by the dynamic simulation module 30 in the system, which receives the hierarchical causal world model output from the world model construction module 20. .

[0119] The dynamic simulation module 30 will simulate the system at the current moment. The state is represented by a multidimensional Gaussian probability distribution, which consists of a mean vector. and a covariance matrix Fully defined. Among them, This represents the most likely estimate of each variable in the system state. This indicates the uncertainty of these estimates and the correlation between variables.

[0120] The module uses a state transition function Based on the current state probability distribution and the received hierarchical causal world model Calculate and predict the system's performance at the next moment. State probability distribution In this embodiment, the state transition function This is a Vector Autoregressive (VAR) model. The entire prediction process is represented as:

[0121] ;

[0122] In the formula, For the system at future moments The mean vector of the state probability distribution; For the system at future moments The covariance matrix of the state probability distribution; It is a state transition function; For the system at the current moment The mean vector of the state probability distribution; For the system at the current moment The covariance matrix of the state probability distribution; For a moment The vector of external intervention actions applied; For at any time A hierarchical causal world model.

[0123] In this simulation, uncertainty is explicitly propagated. Specifically, the covariance matrix at the next time step... The calculation not only inherits the uncertainty of the current moment. Furthermore, process noise caused by the model's inherent imperfections was also included. The covariance matrix of the process noise... It is based on the hierarchical causal world model The causal confidence level is dynamically constructed. For any causal relationship edge in the model... Its confidence level The lower the value, the lower the process noise covariance matrix. The higher the corresponding element value, the better. This results in a larger variance in the prediction of variables driven by low-confidence causal relationships, thus mathematically achieving the technical effect of propagating model uncertainty into the prediction results.

[0124] Finally, the dynamic simulation module 30 will calculate the mean vector. Covariance Matrix The probability distribution of the future state is used as a whole to predict the result and output to the cognitive dissonance detection module 40.

[0125] In the flowchart of the method of this invention, the next stage following the dynamic process simulation is cognitive dissonance detection and self-correction. This stage is executed collaboratively by the cognitive dissonance detection module 40 and the model correction module 50 in the system.

[0126] The cognitive dissonance detection module 40 receives the newly observed true state vector from the variable extraction module 10. and receive the mean vector from the dynamic simulation module 30. Covariance Matrix The module predicts future states. The probability distance calculation unit within the module calculates the Mahalanobis distance between the actual state vector and the probability distribution of the predicted future states based on the received input, in order to quantify cognitive dissonance. .

[0127] Cognitive dissonance is quantified by calculating the Mahalanobis distance between the two. The specific calculation formula is as follows:

[0128] ;

[0129] In the formula, This is a quantitative value for cognitive dissonance; This represents the newly observed true state vector; Let be the mean vector of the probability distribution of future states; The covariance matrix of the probability distribution of future states; superscript For matrix transpose operation; superscript Invert a matrix.

[0130] After the calculation is completed, the threshold comparator within the module will calculate the cognitive dissonance value. With a preset statistical threshold A comparison is made. In this embodiment, the threshold... The critical value is set based on the chi-square distribution. Because... Under the null hypothesis (i.e., the observations do indeed come from the predicted distribution), it follows a distribution with degrees of freedom of The chi-square distribution of (state vector dimension) allows for the determination of a high confidence threshold based on a preset significance level (e.g., 0.05).

[0131] When the comparison result is At this time, the cognitive dissonance detection module 40 outputs a trigger signal to the model correction module 50. Upon receiving this signal, the large-scale language model agent configured within the model correction module 50 is invoked to execute the correction process. The agent first modifies the hierarchical causal world model... Perform a retrospective analysis to pinpoint the causal path that contributed most to the prediction bias.

[0132] Based on the backtracking analysis results, the large language model agent generates a set of structured revised hypotheses. Each hypothesis is a verifiable explanation for the cause of model mismatch; for example, "hypothesis..." "There is an unmodeled causal relationship between variable A and variable B" or "hypothesis" "The causal influence of variable C on variable D has undergone parameter drift."

[0133] Subsequently, the active probe task generator within the module designs one or more specific active probe tasks for each revised hypothesis. These tasks aim to acquire new data that can verify or falsify the corresponding hypothesis. For example, to verify a hypothesis... The task instruction is "Request higher spatial resolution remote sensing imagery of the region where variable A is located"; to verify the hypothesis. The task instruction is to "call an external economic database and query the latest policy documents related to variable C".

[0134] Active detection mission executes and returns new data. Then, the Bayesian inference unit within the module uses the new data... And Bayes' theorem, update the posterior belief for each hypothesis. .

[0135] Finally, the model correction module 50 selects the correction hypothesis with the highest posterior belief and generates a hierarchical causal world model based on the content of this hypothesis. The specific modification instructions. For example, if If selected, then in Add a causal edge from A to B to the graph structure; if If selected, update the edge. The causal confidence level was determined. The modified model will be used for subsequent simulations and analyses.

[0136] In the flowchart of the method of this invention, the final stage is the generation and interaction of situation analysis results. This stage is executed after the model correction module 50 completes the update of the hierarchical causal world model, or when an external query from a user is received.

[0137] In one embodiment, after the model correction process is executed, the system flow automatically returns to the dynamic simulation module 30. The dynamic simulation module 30 calls the hierarchical causal world model updated by the model correction module 50, and uses the previous state as the initial condition to re-execute the dynamic process simulation. This simulation generates a corrected probability distribution of future states, which is considered to more accurately reflect the evolutionary trend of the real world, and is output as the final situation analysis result.

[0138] The results are presented in a visual format, including a situation evolution curve with a range of predictive uncertainty, and an emergent effect warning list, which clearly lists macroscopic variables that are expected to break through predefined critical thresholds in the future, the breakthrough time, and the corresponding probability.

[0139] Furthermore, the method in this embodiment of the invention also includes a counterfactual reasoning step. A user inputs a hypothetical intervention query through an interactive interface; upon receiving the query, the system parses and quantifies it into an external intervention action vector. This vector is a vector with the same dimensions as the system state vector, which has non-zero values ​​in the dimensions corresponding to the variables being intervened (for example, the dimension corresponding to irrigation water volume is 1.2 times the original value), while it is zero in other dimensions.

[0140] Subsequently, the dynamic simulation module 30 uses the system's current state, the existing hierarchical causal world model, and this non-zero external intervention action vector. As input, a new dynamic process simulation is run. The result of this simulation—a future state simulation with predictive uncertainty that includes the effects of interventions—is output as a response to the query.

[0141] In one embodiment, the response is presented as a comparison graph, showing the future evolution path of the target variable and its respective confidence interval under both no-intervention and intervention scenarios, thereby providing a quantitative basis for decision-making.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for situational analysis of large-scale multimodal geospatial data, characterized in that, The method includes the following steps: Multimodal geospatial data is processed using one or more pre-trained geospatial vision basic models to extract hierarchical causal variables that include micro-level and macro-level variables. The time series of the hierarchical causal variables are analyzed using causal discovery algorithms to construct a hierarchical causal world model that includes causal relationships and causal confidence. The hierarchical causal world model is driven by a state transition function to perform dynamic process simulation, which evolves the probability distribution of the system state from the current moment to the future moment, generating a future state with predictive uncertainty. Calculate the probability distance between the probability distributions of newly observed real states and future states in the multimodal geospatial data, and quantify the cognitive dissonance of the difference between the two. A preset threshold is designed. When the cognitive dissonance exceeds the preset threshold, a large language model agent is activated to generate a set of corrective hypotheses. The hierarchical causal world model is updated according to the corrective hypotheses, and the dynamic process simulation is re-performed using the updated hierarchical causal world model to generate a corrected future state as a new situational analysis result. The step of extracting hierarchical causal variables that include both micro-level and macro-level variables specifically includes: Geographic entities and their states are identified from the multimodal geospatial data and used as micro-level variables. By using an aggregation function, the states of a set of micro-level variables within a specified spatiotemporal neighborhood are aggregated to generate a macro-level variable; The hierarchical causal world model includes modeling causal relationships between variables at the micro level, between variables at the macro level, and between variables across levels.

2. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, In the step of constructing a hierarchical causal world model that includes causal relationships and causal confidence levels, the process of assigning the causal confidence level to each causal relationship includes: Calculate the causal confidence level using the following formula. : ； In the formula, Represents a causal relationship edge; The causal confidence level of the causal relationship edge; For a fusion function; The statistical significance index of the causal relationship is used. The goodness-of-fit score is based on historical data; These are the prior weights derived from the domain knowledge base.

3. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, The specific steps for performing dynamic process simulation include: The current state of the system is represented as a probability distribution containing a mean vector and a covariance matrix; Using the state transition function, the probability distribution of the state at the next moment is predicted based on the probability distribution of the current state and the hierarchical causal world model. During the prediction process, the model uncertainty represented by the causal confidence level is propagated to the covariance matrix of the state at the next time step.

4. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, The steps for quantifying the cognitive dissonance between the two specifically include: The Mahalanobis distance between the two is calculated using the following formula to quantify cognitive dissonance. : ； In the formula, This is the quantitative value of the cognitive dissonance. This represents the newly observed true state vector; Let be the mean vector of the probability distribution of the future state; Let be the covariance matrix of the probability distribution of the future state; superscript For matrix transpose operation; superscript Invert a matrix.

5. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, The step of activating a large language model agent to generate a set of revised hypotheses further includes: The large language model agent analyzes the causal paths that lead to prediction bias and generates the set of corrective hypotheses. The large language model agent autonomously designs and dispatches one or more active probing tasks to verify the set of revised hypotheses.

6. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, The step of updating the hierarchical causal world model according to the revised hypothesis specifically includes: Acquire new data returned by the active probing mission; Update the posterior beliefs of the set of revised hypotheses based on the new data and Bayesian inference; Select the revised hypothesis with the highest posterior belief, and update the causal relationship or causal confidence of the hierarchical causal world model accordingly.

7. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, The method further includes: Continuously monitor the simulation prediction results for the macroscopic layer variables; And when it is predicted that the value of a certain macro-level variable will exceed a predefined critical threshold in the future, an emergent effect can be identified and an early warning can be generated.

8. The method for situational analysis of large-scale multimodal geospatial data according to claim 1, characterized in that, The method also includes a counterfactual deduction step: Receive a query input from a user and parse the query into an external intervention action vector; Using the external intervention action vector as input, a new dynamic process simulation is run in the hierarchical causal world model; The simulation results, which include predictive uncertainty, are output as a response to the query.

9. A multimodal geospatial data large model situation analysis system, characterized in that, The system is used to perform a multimodal geospatial data large model situational analysis method according to any one of claims 1-8, the system comprising: The variable extraction module is configured with one or more pre-trained geospatial vision basic models to process multimodal geospatial data and extract hierarchical causal variables containing micro-level and macro-level variables. The world model construction module is equipped with a causal discovery algorithm, which is used to analyze the time series of the hierarchical causal variables and construct a hierarchical causal world model that includes causal relationships and causal confidence. The dynamic simulation module is configured with a state transition function to drive the hierarchical causal world model to perform dynamic process simulation, which evolves the probability distribution of the system state from the current moment to the future moment, generating a future state with predictive uncertainty. The cognitive dissonance detection module is used to calculate the probability distance between the probability distributions of newly observed real states and future states in the multimodal geospatial data, and to quantify the cognitive dissonance of the difference between the two. The model correction module is equipped with a large language model agent, which is used to generate a set of correction hypotheses when the cognitive dissonance exceeds a preset threshold, update the hierarchical causal world model according to the correction hypotheses, and provide the updated hierarchical causal world model to the dynamic simulation module for subsequent situation analysis.

Images