Air-ground multi-source spatio-temporal information fusion perception method and system based on artificial intelligence

Abstract

The present application relates to the technical field of environmental monitoring, and particularly relates to an air-ground multi-source spatio-temporal information fusion perception method and system based on artificial intelligence, comprising: constructing a three-source data acquisition architecture, using an intelligent calibration model to dynamically calibrate the three-source data in space and time, and obtaining calibrated data; correlating and verifying the calibrated data and quantifying the weight of interference factors through a causal reasoning model, dynamically adjusting the field constraint parameters through a time sequence feature coding model, completing deep fusion through an intelligent fusion model, and outputting high-precision fusion results; constructing a field knowledge graph, mining potential business rules and identifying unknown abnormal patterns through rule mining and abnormality identification, and obtaining an iteratively updated knowledge graph dynamic; based on a collaborative learning framework and combined with privacy protection technology, realizing cross-region model collaborative evolution, constructing a scenario-based value conversion engine to generate a business structured report. The present application can improve the data fusion accuracy and the full-link traceability accuracy.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to an artificial intelligence-based method and system for fusion sensing of multi-source spatiotemporal information from air and ground. Background Technology

[0002] With the rapid development of surveying and mapping monitoring, water conservancy management, forestry protection, emergency rescue and other fields, the demand for refined, real-time and traceable spatiotemporal information perception is becoming increasingly prominent and has become a core requirement to support accurate business decision-making.

[0003] Traditional multi-source data sensing technologies and existing solutions generally suffer from pain points such as insufficient dynamic adaptation of data acquisition tasks, limited spatiotemporal calibration accuracy, low multimodal fusion efficiency, difficulty in balancing privacy and performance in cross-regional collaboration, and weak end-to-end traceability capabilities, which cannot meet the high-standard application requirements of spatiotemporal information sensing in various fields.

[0004] Therefore, it is necessary to construct a comprehensive, intelligent, highly accurate, highly reliable, and traceable multi-source spatiotemporal data fusion and sensing technology system that integrates air and ground data. Summary of the Invention

[0005] This invention takes causal-driven multimodal precision fusion as its core and constructs a closed-loop air-ground multi-source spatiotemporal information fusion perception system to improve the accuracy of data fusion and the accuracy of full-link traceability.

[0006] The technical solution proposed in this invention is: an artificial intelligence-based multi-source spatiotemporal information fusion sensing method for air and ground, the method comprising: A three-source data acquisition architecture is constructed, a unified spatiotemporal benchmark is preset, multiple types of modal data sources are integrated, collaborative acquisition tasks are generated based on dynamic game theory using a multi-objective optimization algorithm, nested acquisition logic is configured through adversarial training, and an intelligent calibration model is used to dynamically calibrate the three-source data in spatiotemporal manner to obtain calibration data. The calibration data is correlated and verified by a causal reasoning model and the weight of interference factors is quantified. The domain constraint parameters are dynamically adjusted by a time-series feature coding model. A cross-modal alignment algorithm is used to achieve semantic fusion of multiple modalities. Differentiated fusion weights are assigned in combination with data reliability assessment. Deep fusion is completed by an intelligent fusion model and high-precision fusion results are output. Based on the fusion results, a domain knowledge graph is constructed. Potential business rules are mined and unknown abnormal patterns are identified through rule mining and anomaly recognition. Knowledge to be verified is marked and pushed for manual review, and the dynamic knowledge graph is obtained through iterative updates. Based on a collaborative learning framework combined with privacy protection technology, cross-regional model collaborative evolution is achieved, generating structured business reports and constructing full-link traceability information by associating updated knowledge graphs.

[0007] Preferably, the specific process of generating collaborative acquisition tasks based on the multi-objective optimization algorithm through dynamic game theory is as follows: A multi-objective optimization function is defined, with the objectives of maximizing data integrity, minimizing acquisition cost, and optimizing task timeliness, and an objective function matrix is ​​constructed. Define the constraints of dynamic game, including equipment resource constraints, data quality constraints, and scenario adaptation constraints; The objective function matrix is ​​solved using a multi-objective optimization algorithm to generate multiple sets of candidate acquisition schemes; The candidate acquisition schemes are comprehensively evaluated, and the optimal candidate scheme is selected as the final collaborative acquisition task. The task execution parameters are then output.

[0008] Preferably, the specific process for performing dynamic spatiotemporal calibration on the three source data is as follows: Outlier removal and missing value filling are performed on the aggregated multi-modal data sources to obtain clean data; Based on a unified spatiotemporal benchmark, a dynamic time warping algorithm is used to align temporal data, and an image registration algorithm is used to align spatial data, achieving preliminary spatiotemporal alignment of the three-source data. The initially aligned data is input into a Bayesian neural network to model the probability distribution of satellite orbit deviation, UAV flight attitude error, and ground sensor drift. The calibrated data and the corresponding error confidence interval are output, and the error confidence interval is used as the data reliability assessment result.

[0009] Preferably, the specific process for obtaining the high-precision fusion result is as follows: The calibrated data is input into the causal inference model. The counterfactual intervention module simulates the correlation after removing confounding factors, quantifies the weight of interfering factors, removes false causal correlations, and obtains valid correlation data verified by causality. The effective associated data is input into the temporal feature encoding model to extract the temporal features of the scene. The loss function coefficients of the domain constraint parameters are dynamically adjusted based on the features to obtain associated data with dynamic constraint weights. The associated data with dynamic constraint weights is split into explicit modal data and implicit modal data. A multimodal contrastive learning algorithm is used to map the two types of data to the same semantic space and output cross-modal alignment features. Based on the data reliability assessment results, differentiated fusion weights are assigned to cross-modal aligned features. The weighted features are then input into the improved Transformer model for feature-level and decision-level fusion, outputting high-precision fusion results.

[0010] Preferably, the process of obtaining the iteratively updated knowledge graph is as follows: Based on the high-precision fusion results, core information is extracted to construct an initial domain knowledge graph; The initial domain knowledge graph and the high-precision fusion result are jointly input into the knowledge mining and anomaly identification module to mine potential business rules, identify unknown anomaly clusters that deviate from the normal pattern, and output the potential business rule set and unknown anomaly cluster data. Mark potential business rule sets and unknown anomaly cluster data as knowledge to be verified, sort them by priority and push them to manual review, and obtain the results of manual review; The approved potential business rules and unknown anomaly patterns are added to the initial domain knowledge graph, the rule base and anomaly pattern base of the graph are updated, and the iteratively updated domain knowledge graph is output.

[0011] Preferably, the specific implementation process of the cross-regional model co-evolution is as follows: Each region deploys a local model, loads the iteratively updated domain knowledge graph, combines local multi-source private data to fine-tune the local model, and outputs the model parameter update amount; Differential privacy technology is used to add Laplacian noise to the model parameter update, thus completing the privacy protection process and outputting the encrypted parameter update; Each region uploads the encrypted parameter update data to the central server, which then integrates the parameter update data from all regions using a parameter aggregation algorithm. A global cross-regional co-evolutionary model is generated based on the aggregated parameter update amount and distributed to each region. Each region updates its local model through the global model, thereby achieving cross-regional model co-evolution.

[0012] Preferably, the specific process for obtaining the structured business report is as follows: Pre-set templates to meet the differentiated needs of different assessment areas, including report indicators, data dimensions and format specifications specific to each area; The scenario-based value conversion engine reads the output of the cross-regional collaborative evolution model and matches the corresponding demand templates according to the target application area. Based on the matched demand template, core data is extracted from the output of the cross-regional co-evolutionary model. The core data is structured and organized, including data classification and statistics, trend analysis and visualization chart generation, and the data is presented according to template specifications to generate structured business reports for the corresponding fields.

[0013] Preferably, the specific process of tracing the associated full-link traceability information is as follows: Collect process data from each stage, including data sources, processing algorithms, parameter configurations, and error records for each stage, to form a full-link process dataset; Establish a full-link data association index to bind the full-link process dataset with each indicator in the business structured report, forming a mapping relationship between indicators and process data; Construct a traceability query module that, based on the mapping relationship between indicators and process data, allows reverse querying of the corresponding original data, processing flow, and error source for any indicator in the business structured report; The traceability query module outputs full-link traceability information, including related indexes and query results.

[0014] The present invention also provides an artificial intelligence-based air-ground multi-source spatiotemporal information fusion sensing system, the system being used to execute the artificial intelligence-based air-ground multi-source spatiotemporal information fusion sensing method.

[0015] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the artificial intelligence-based multi-source spatiotemporal information fusion sensing method.

[0016] The beneficial effects of this invention are: 1. By employing multi-objective optimization-driven three-source active collaborative acquisition, extreme scenario antifragility prediction, and Bayesian neural network uncertainty calibration, the entire process of acquisition-calibration-preprocessing is integrated, completely eliminating the cumbersome manual configuration and link separation inherent in traditional technologies. This not only significantly reduces the complexity and labor costs of cross-scenario adaptation in multiple fields such as surveying, water conservancy, and forestry, but also significantly improves data integrity and real-time performance. Calibration errors are precisely quantified and controlled, providing a high-fidelity and highly reliable data foundation for subsequent deep integration. Simultaneously, it completely solves the problem of error-prone manual parameter configuration, laying a solid foundation for dynamic optimization across the entire process.

[0017] 2. By leveraging counterfactual causal reasoning, dynamic constraints based on scene temporal features, and multimodal cross-domain alignment technology, the core pain points of traditional fusion—namely, "false association misjudgment, rigid constraints, and single dimension"—have been successfully addressed. The accuracy of the fusion results is improved by more than 30% compared to existing technologies, and it can flexibly adapt to dynamic changes in scenarios such as flood / dry seasons in water conservancy and outbreaks of forestry pests and diseases. The deep alignment of implicit and explicit modalities enriches the fusion dimensions and significantly enhances the information perception capability in complex scenarios. The linkage between the error confidence level of the fusion process and the preceding data collection makes the fusion results more aligned with the actual business needs of the domain, providing accurate and reliable data support for subsequent value transformation.

[0018] 3. By leveraging a dual-driven knowledge graph for automatic updates, federated privacy-preserving cross-domain collaboration, and a scenario-based business value transformation engine, the system effectively addresses the challenges of traditional technologies, such as "human dependence on knowledge, data silos, and insufficient practicality." The model's cross-regional self-evolution capability reduces scenario expansion costs by 80%, while differential privacy technology ensures data privacy and security. The scenario-based adaptation engine directly transforms deep fusion results into structured business outputs such as surveying accuracy reports, water conservancy disaster early warnings, and forestry resource accounting, significantly improving the technology's practicality and decision-making efficiency. The combination of a full-link traceability mechanism and self-evolution capability ensures that the technology system can adapt to the dynamic needs of the field in the long term, providing core technical support for the stable operation of industries such as surveying, water conservancy, and forestry. Attached Figure Description

[0019] Figure 1 This is a flowchart of an AI-based multi-source spatiotemporal information fusion sensing method for air and ground. Figure 2 This is a flowchart of the perception process of an AI-based multi-source spatiotemporal information fusion perception method. Detailed Implementation

[0020] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0021] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0022] like Figure 1 and Figure 2As shown, a three-source data acquisition architecture is constructed, a unified spatiotemporal benchmark is preset, and multiple types of modal data sources are integrated. Collaborative acquisition tasks are generated through dynamic game theory based on a multi-objective optimization algorithm. Nested acquisition logic is configured through adversarial training. An intelligent calibration model is used to dynamically calibrate the three-source data in spatiotemporal space to obtain calibration data. A causal inference model is used to verify the correlation of the calibration data and quantify the weights of interference factors. A temporal feature encoding model is used to dynamically adjust domain constraint parameters. A cross-modal alignment algorithm is used to achieve semantic fusion of multiple modalities. Differentiated fusion weights are assigned based on data reliability assessment. A deep fusion model is used to complete deep fusion and output high-precision fusion results. A domain knowledge graph is constructed based on the fusion results. Potential business rules are mined and unknown abnormal patterns are identified through rule mining and anomaly recognition. Knowledge to be verified is marked and pushed for manual review, resulting in an iteratively updated dynamic knowledge graph. A collaborative learning framework combined with privacy protection technology is used to achieve cross-regional model collaborative evolution, generating structured business reports. The updated knowledge graph is then used to construct full-link traceability information.

[0023] Furthermore, the specific process of generating collaborative acquisition tasks based on the multi-objective optimization algorithm through dynamic game theory is as follows:

[0024] A multi-objective optimization function is defined, aiming to maximize data integrity, minimize acquisition cost, and optimize task timeliness, and an objective function matrix is ​​constructed. The constraints of the dynamic game are clarified, including equipment resource constraints, data quality constraints, and scenario adaptation constraints. The objective function matrix is ​​solved using a multi-objective optimization algorithm to generate multiple candidate acquisition schemes. The candidate acquisition schemes are comprehensively evaluated, and the optimal candidate scheme is selected as the final collaborative acquisition task, outputting the task execution parameters. A new scenario mutation adaptive mechanism is added to ensure the convergence of the algorithm and the quality of the solution under sudden scenarios.

[0025] Among them, the multi-objective optimization function uses the weighted summation method to construct the objective function matrix. The core is to achieve a balance between the three core objectives of maximizing data integrity, minimizing acquisition cost, and optimizing task timeliness. The specific calculation method is as follows: Minimize objective function value = data integrity weight × (1 - data integrity) + acquisition cost weight × acquisition cost + task timeliness weight × task timeliness.

[0026] In the formula, the parameter vector covers configurations related to collaborative data acquisition tasks, such as data source selection, acquisition frequency, and equipment scheduling scheme; the data integrity value ranges from 0 to 1, with higher values ​​representing more comprehensive data coverage, calculated as the ratio of the effective data acquisition volume to the theoretically required data acquisition volume; the acquisition cost unit is yuan, specifically including various related expenditures such as equipment energy consumption, drone flight costs, and manual assistance costs; the task timeliness unit is seconds, i.e., the total time consumed from task issuance to completion of data acquisition; the weights of data integrity, acquisition cost, and task timeliness are all determined based on the analytic hierarchy process, with values ​​of 0.4, 0.3, and 0.3 respectively, and the sum of the three weights is 1. The weight ratio of each objective can be dynamically adjusted according to the actual needs of specific application scenarios such as surveying and mapping monitoring and water conservancy management.

[0027] The objective function matrix is ​​organized in a standardized format of "objective type - function description - weight coefficient - constraint boundary" to ensure the relevance and standardization of the solution process.

[0028] The constraints of the dynamic game are constructed around three dimensions: equipment, data, and scenario. All constraints are quantitative and clearly defined in terms of priority (from high to low): data quality constraints > equipment resource constraints > scenario adaptation constraints, ensuring that core requirements are met first. At the same time, slack variables (within the range of 0-0.05) are introduced to handle constraint conflicts. When a high-priority constraint is satisfied, a small deviation (not exceeding 5% of the constraint value) is allowed for a low-priority constraint, so as to avoid the algorithm failing to converge due to constraint conflicts. Specific constraints are as follows: Equipment resource constraints: The satellite data receiving terminal has no more than 8 concurrent receiving channels (with dynamic expansion capability, which can be expanded to 16 channels through redundant equipment in high-density monitoring scenarios); the drone's endurance is no less than 2 hours (the duration of a single flight data collection is no more than 90 minutes, and the function of ground sensors and satellites coordinating to supplement data collection can be realized under extreme weather conditions); the number of concurrent data collection nodes in the ground sensor network is no more than 500; and the equipment operating parameters (voltage, current) are within the standard range (voltage 10-14V, current no more than 2A); Data quality constraints: The accuracy of collected data is no less than 98%, the completeness is no less than 99%, the proportion of outliers in a single data source is no more than 2%, and the proportion of missing values ​​is no more than 1%; Scenario adaptation constraints: The data collection accuracy in surveying and monitoring scenarios is no more than 0.1m for position and no more than 0.05m for elevation; the data collection frequency in water conservancy management scenarios is no less than once every 10 minutes; the infrared data resolution in forestry protection scenarios is no less than 640×480; and the timeliness of emergency rescue scenarios is no more than 300 seconds. The constraints are expressed using inequalities and embedded in the objective function solution process, serving as the core basis for candidate solution selection. At the same time, a constraint relaxation threshold (±10%) is set to flexibly adjust in response to unexpected scenarios.

[0029] The multi-objective optimization algorithm adopts the non-dominated sorting genetic algorithm No. 3, combined with dynamic game strategy and epsilon constraint processing technology to achieve multi-objective equilibrium solution. The algorithm parameters are set as follows: population size is set to 100, number of iterations is set to 200, crossover probability is set to 0.8, mutation probability is set to 0.05, and crowding threshold is set to 0.5.

[0030] The specific implementation of constraint handling is as follows: The epsilon constraint method is used to transform multi-objective optimization into sequential single-objective optimization. The main objective is to maximize data integrity, while the collection cost and task timeliness are used as constraint objectives. A first constraint threshold is set (upper limit of collection cost = optimal cost × (1 plus slack variable)) and a second constraint threshold is set (upper limit of task timeliness = optimal timeliness × (1 plus slack variable)). A penalty function is applied to the scheme that violates the constraints. The penalty function value is equal to the degree of constraint violation multiplied by 10 (degree of violation = (actual value minus constraint value) divided by constraint value). The penalty function value is included in the objective function value to reduce the ranking priority of the scheme that violates the constraints.

[0031] During the solution process, a dynamic game mechanism is used to simulate the checks and balances between the objectives, supplementing the static game model with complete information: data integrity, acquisition cost, and task timeliness are regarded as three players, and the payoff function of each player is the degree of achievement of the objective (payoff value = one minus (actual objective value divided by optimal objective value)). The game equilibrium condition is that the difference in payoff values ​​among all players does not exceed 0.1. When the performance improvement of one objective causes the payoff value of another objective to decrease by more than 0.1, the game adjustment is triggered. The textual description of the reference point adjustment method is: the new reference point is equal to the original reference point plus (the payoff value of the objective minus the average payoff value) multiplied by 0.3. The distribution of reference points is dynamically optimized to ensure that the three objectives are optimally coordinated. After the solution is completed, twenty to fifty candidate acquisition schemes are generated. Each scheme satisfies all constraints (including the allowable deviation of slack variables) and covers different equipment scheduling strategies and acquisition parameter configurations.

[0032] For sudden scenario switching (equipment failure, new task), it has a scenario mutation adaptive mechanism: Scenario mutation detection: Real-time monitoring of equipment operating status (voltage, signal strength) and task queue. When equipment failure (such as drone disconnection) or a new emergency task is detected, it is determined as a scenario mutation; Dynamic adjustment of algorithm parameters: Immediately reduce the population size to 50, the number of iterations to 100, and the crossover probability to 0.9 to quickly generate backup solutions; Online learning and updating: Uses a deep Q-network as an online learning algorithm, trains the decision model based on historical sudden scenario processing data (10,000+ sets), and adjusts the update frequency according to the scenario type (updates once every 5 triggers for equipment failure scenarios and once every 3 triggers for new task scenarios), ensuring that the algorithm converges within 100 iterations (convergence rate ≥ 95%), the solution generation time is ≤ 2 seconds, and the quality loss of the solution is ≤ 5%.

[0033] The comprehensive score of the candidate data collection scheme adopts a multi-dimensional quantitative scoring method, and the specific calculation method is as follows: Comprehensive score = Data integrity weight (value 0.4) × Data integrity value × 10 + Collection cost weight (value 0.3) × [1 - (Current scheme collection cost ÷ Maximum collection cost among all candidate schemes)] × 10 + Task timeliness weight (value 0.3) × [1 - (Current scheme task time ÷ Maximum task time among all candidate schemes)] × 10.

[0034] The comprehensive score ranges from 0 to 10 points. After ranking the scores from highest to lowest, the highest-scoring solution is selected as the final collaborative data collection task. If multiple solutions have a score difference of no more than 0.2, they are considered to have the same score. The priority rules for tie-breaking are as follows: the first priority is to select the solution with the higher data integrity value; if the data integrity values ​​are the same, the second priority is to select the solution with better task timeliness (shorter task duration); if the first two are the same, the final priority is to compare the collection cost and select the solution with the lower collection cost.

[0035] The final collaborative acquisition task outputs the following execution parameters: a list of data sources, acquisition parameters for each device (satellite resolution, UAV flight altitude, sensor acquisition frequency), device scheduling timetable, data transmission protocol, and task completion time limit. The parameters are stored in JSON format and can be sent to the fusion sensing terminal in real time.

[0036] Furthermore, the specific process of dynamic spatiotemporal calibration of the three-source data is as follows: Outlier removal and missing value filling are performed on the aggregated multi-modal data sources to obtain clean data. Based on a unified spatiotemporal benchmark, the time-series data is aligned using a dynamic time warping algorithm, and the spatial data is aligned using an image registration algorithm, achieving preliminary spatiotemporal alignment of the three-source data. The preliminarily aligned data is then input into a Bayesian neural network to model the probability distributions of satellite orbit deviation, UAV flight attitude error, and ground sensor drift. The calibrated data and corresponding error confidence intervals are output, with the error confidence intervals serving as the data reliability assessment results.

[0037] The aggregated multi-modal data sources encompass satellite remote sensing data (imagery, elevation, meteorology, etc.), UAV multi-payload data (LiDAR point clouds, infrared imagery, visible light imagery, etc.), ground sensor data (soil moisture, water level, air quality, etc.), and manually calibrated data. This data source supports multiple data types, including tagged image file formats, JSON format, and comma-separated value formats, and aggregates data to the data processing unit through multiple interfaces, with an aggregation latency of no more than 100ms.

[0038] The intelligent calibration model comprises five core processes: data preprocessing, spatiotemporal benchmark unification, temporal and spatial alignment, probabilistic error modeling, and calibration result output. Each step is deeply integrated, with precise implementation of technical details: First, outlier removal is achieved through the Shapiro-Wilk test and the Isolation Forest algorithm, followed by missing value filling using a differentiated strategy; then, a unified benchmark is established based on Coordinated Universal Time (UTC) 84; subsequently, preliminary temporal and spatial alignment is achieved through dynamic time warping and SIFT / ICP registration algorithms; next, a three-layer Bayesian neural network is used to quantify and model various error probability distributions, combined with Kalman filtering to optimize error propagation; finally, calibration data and a reliability assessment result with a 95% confidence interval are output, providing a high-fidelity data foundation for subsequent high-precision fusion. The specific steps are as follows: The Shapiro-Wilke test was used to test the data distribution. The selection criteria were: in small sample (n≤5000) and medium sample (5000<n≤10000) scenarios, this method showed significantly higher power than the Kolmogorov-Smirnov test for normality testing, and its sensitivity in identifying skewed distributions was improved by more than 30%. The judgment criteria for the test results were as follows: if the P-value was not lower than 0.05, the data was considered to conform to a normal distribution, and outliers exceeding the range of "mean - 3 standard deviations" to "mean + 3 standard deviations" were removed using the 3σ criterion; if the P-value was lower than 0.05, outliers exceeding the range of "first quartile - 1.5 interquartile range" to "third quartile + 1.5 interquartile range" were removed using the interquartile range. Meanwhile, outlier removal is combined with the isolated forest algorithm. Sensitivity analysis shows that when the threshold is 0.8, the recall rate is 98% and the precision is 92%; when the threshold is 0.85, the recall rate is 97% and the precision is 96%; and when the threshold is 0.9, the recall rate is 95% and the precision is 98%. After comprehensive consideration, the threshold of 0.85 is selected to ensure that the accuracy of anomaly identification is not less than 99%.

[0039] Missing values ​​are filled using a differentiated strategy: linear interpolation is used for continuous time-series data, K-nearest neighbor (K=5) interpolation is used for discrete categorical data, and the mean of neighboring pixels is used for satellite image data. The error of the filled data does not exceed 5%, ensuring the integrity and accuracy of clean data.

[0040] The unified spatiotemporal reference adopts Coordinated Universal Time (UTC) 84 (spatial reference) and Coordinated Universal Time (UTC) (time reference): the timestamps of the three sources of data are unified through the time synchronization module of the Global Positioning System, and the time synchronization accuracy does not exceed 1 millisecond; the spatial coordinates are uniformly converted to Coordinated Universal Time 84 through the coordinate transformation algorithm, and the transformation error does not exceed 0.01 meters, ensuring the consistency of the spatiotemporal reference.

[0041] The dynamic time warping algorithm is used to align three-source time series data. It adopts an adaptive window adjustment mechanism, and the specific calculation method is: window size = basic window coefficient × (1 + noise correction coefficient × data standard deviation).

[0042] The base window coefficient is set based on the data frequency, with a value of 30 for high-frequency data and 60 for low-frequency data. The noise correction coefficient is fixed at 0.2, and the data standard deviation is used to quantify the degree of data fluctuation. The actual adjustment rules are as follows: if the data frequency is high (e.g., sensor data acquisition once per minute) and the noise is low (standard deviation ≤ 0.01), the window size is set to 30-40; if the data frequency is low (e.g., satellite data acquisition once per day) and the noise is high (standard deviation > 0.01), the window size is set to 60-80. Euclidean distance is used as the distance metric, and the time axis of the time series data is adjusted through dynamic programming to ensure time series alignment accuracy.

[0043] Image registration algorithms are used to align spatial data: Scale-Invariant Feature Transform (SIFT) feature matching is used between sky-source remote sensing imagery and air-source UAV imagery, with a matching threshold of 0.7 and a matching accuracy of no less than 95%. Iterative Closest Point (ICP) registration is used between air-source lidar point cloud data and ground-source sensor location data, with a registration error not exceeding 0.05 meters. These algorithms achieve precise spatial alignment of the three data sources, with a preliminary spatiotemporal deviation of no more than 0.1 meters.

[0044] The Bayesian neural network employs a three-layer structure: input layer, hidden layer, and output layer. The input layer consists of pre-aligned three-source data (256 dimensions). The hidden layer comprises two layers (the first layer has 128 neurons with a linear rectified function as the activation function; the second layer has 64 neurons with a linear rectified function with leakage as the activation function). The output layer contains the probability distributions of satellite orbital deviation, UAV flight attitude error, and ground sensor drift (using a Gaussian mixture model to adapt to non-Gaussian noise). Network training utilizes a variational inference algorithm with a learning rate of 0.001 and 1000 iterations. The training set contains over 100,000 sets of three-source data with error labels (covering both Gaussian and non-Gaussian noise scenarios). Training continues until the validation set loss converges (the loss value does not exceed 0.008). During the modeling process, Monte Carlo sampling (sampling times = 100) was used to estimate the posterior probability of each error, quantifying the probability distribution range of satellite orbit deviation (not exceeding 0.03m), UAV flight attitude error (pitch angle not exceeding 0.5°, roll angle not exceeding 0.5°), and ground sensor drift (not exceeding 0.02m). Kalman filtering was introduced to correct error propagation, with the filter parameters set as follows: the process noise covariance matrix is ​​a diagonal matrix with all diagonal elements of 1e-6, the observation noise covariance matrix is ​​a diagonal matrix with all diagonal elements of 1e-5, and the initial state covariance matrix is ​​a diagonal matrix with all diagonal elements of 1e-4. The error calibration accuracy was improved by more than 25% in non-Gaussian noise scenarios.

[0045] The calibrated data is obtained through an error compensation algorithm, which subtracts the expected value of the corresponding error from the initially aligned data to ensure that the calibration error does not exceed 0.05m. The error confidence interval is calculated based on the posterior probability distribution, with a confidence level set at 95%, meaning that 95% of the error falls within this interval. The smaller the interval width, the higher the data reliability. The error confidence interval serves as the data reliability assessment result and is stored in a format categorized as data type-confidence interval width-reliability level. The reliability level is divided into three levels (Excellent: confidence interval width not exceeding 0.02m; Good: confidence interval width greater than 0.02m and not exceeding 0.05m; Poor: confidence interval width greater than 0.05m), which is used as the basis for subsequent fusion weight allocation. The calibrated data is stored synchronously with the reliability assessment results, providing support for subsequent link calls and verification.

[0046] Furthermore, the specific process for obtaining the high-precision fusion result is as follows: The calibrated data is input into a causal inference model. A counterfactual intervention module simulates the relationships after removing confounding factors, quantifies the weights of interfering factors, eliminates spurious causal associations, and obtains valid association data verified by causality. This valid association data is then input into a temporal feature encoding model to extract scene temporal features. Based on these features, the loss function coefficients of the domain constraint parameters are dynamically adjusted to obtain association data with dynamic constraint weights. This association data with dynamic constraint weights is then split into explicit and implicit modal data. A multimodal contrastive learning algorithm is used to map the two types of data to the same semantic space, outputting cross-modal alignment features. Based on the data reliability assessment results, differentiated fusion weights are assigned to the cross-modal alignment features. The weighted features are then input into an improved Transformer model for feature-level and decision-level dual-level fusion, outputting a high-precision fusion result.

[0047] The causal inference model consists of three parts: a causal graph, a structural equation model, and a counterfactual intervention module. Based on the causal inference framework, the specific implementation details are as follows: Causal graph construction: A hybrid approach combining domain knowledge and data-driven methods is adopted. An initial causal graph (containing 20+ core nodes and 30+ causal edges) is first constructed based on the experience of experts in fields such as water conservancy and surveying. Then, the existence of edges is optimized based on data feedback using the Peter-Clarke algorithm, resulting in a final causal graph accuracy of ≥92%. Structural equation model: A linear regression equation is used (with a determination coefficient of not less than 0.85). Nonlinear relationships are modeled by mapping to a high-dimensional space using kernel functions. Counterfactual intervention module: The module simulates the correlation after removing confounding factors (such as weather interference and temporary equipment errors) through interference quantum simulation. The weights of each interference factor are quantified using Shapley's additive interpretation (interference factor weights less than 0.1 are considered invalid interferences and are removed). False causal associations (such as feature pairs that appear related but have no actual causal relationship) are removed to ensure the reliability of the effective correlation data. The false association removal rate for the effective correlation data after causal verification is not less than 98%.

[0048] The temporal feature encoding model focuses on extracting multi-scale scene temporal features and dynamically adapting to domain constraints. It consists of a temporal convolutional network (TCN) and a channel attention network, which work together to accurately extract temporal features and enhance key information, providing scenario-based adaptation support for subsequent cross-modal fusion. The specific implementation is as follows: The temporal convolutional network (TCNN) employs an architecture combining TCNN and an attention mechanism. The input is valid associated data (temporal length = 100, feature dimension = 64). The TCNN contains four convolutional layers (kernel size = 3, stride = 1, padding value = 1), expanding the receptive field through dilated convolutions to extract multi-scale temporal features. The attention mechanism uses a channel attention network to strengthen the weights of key temporal features and suppress interference from irrelevant temporal features. After feature extraction, based on the dynamic changes in temporal features (such as vegetation growth timeline and water level change timeline), an adaptive adjustment algorithm dynamically modifies the loss function coefficients of the domain constraint parameters. For example, in water conservancy management scenarios, when water level changes drastically, the loss coefficient of the water level feature constraint is increased (from 0.6 to 0.8) to ensure the targeted nature of feature extraction. The associated data with dynamically constrained weights have weights that are positively correlated with the temporal feature change trend, with weight values ​​ranging from 0.3 to 0.9.

[0049] Explicit modal data refers to data that can be directly observed and whose features are easily extracted (such as satellite visible light imagery, UAV lidar point clouds, and real-time values ​​from ground sensors). Implicit modal data refers to feature data that is difficult to observe directly and requires inference to obtain (such as vegetation growth trends, soil moisture gradients, and water level change rates). The splitting is performed using a feature observability quantification algorithm, with the core indicators defined as mutual information and Pearson correlation coefficient: observability = 0.6 × mutual information + 0.4 × Pearson correlation coefficient. The mutual information calculation adopts deep mutual information maximization (deep learning method), which maximizes the mutual information between input features and global features by training a dual-branch network to ensure the accuracy of mutual information calculation under nonlinear relationships. Data with an observability of not less than 0.7 is classified as explicit modal data, and data with an observability of less than 0.7 is classified as implicit modal data. The multimodal contrastive learning algorithm adopts a simple contrastive learning framework. The loss function is explicitly defined as the information-noise contrastive estimation loss function. Positive and negative sample pairs are generated through data augmentation (rotation and pruning augmentation for explicit modality data, and noise addition and translation augmentation for implicit modality data). The learning rate is 0.0001 and the number of iterations is 800. The two types of data are mapped to the same semantic space (128 dimensions). The cross-modal alignment error does not exceed 0.03. The output cross-modal alignment features contain the core information of both explicit and implicit features.

[0050] Differentiated fusion weights are assigned based on data reliability assessment results, using a linear allocation algorithm: weight equals the reliability level coefficient multiplied by the feature importance coefficient. The reliability level coefficient is (Excellent: 1.0; Good: 0.7; Poor: 0.3), and the feature importance coefficient is determined based on domain knowledge and feature contribution (range 0.2-0.8). The sum of all feature weights is 1. The improved transformer model adds a feature fusion layer to the traditional transformer model. The input is a weighted cross-modal aligned feature (128 dimensions). The encoder has 6 layers (8 multi-head attention heads, 256 hidden layer dimensions), and the decoder has 3 layers. Feature-level fusion extracts the correlation information between features through the encoder to generate a fused feature vector. Decision-level fusion uses the decoder to vote on the feature-level fusion results, combining domain rules to correct fusion bias. This two-level fusion ensures a fusion accuracy of no less than 99%. The high-precision fusion result includes spatial location information, temporal variation information, and environmental parameter information. It can output in both JSON and tagged image file formats and can be directly used for subsequent knowledge graph construction and business applications.

[0051] Furthermore, the process of obtaining the iteratively updated knowledge graph is as follows: Based on the high-precision fusion results, core information is extracted to construct an initial domain knowledge graph. The initial domain knowledge graph and the high-precision fusion results are then jointly input into the knowledge mining and anomaly identification module to mine potential business rules, identify unknown anomaly clusters that deviate from normal patterns, and output a set of potential business rules and data on unknown anomaly clusters. The set of potential business rules and data on unknown anomaly clusters are marked as knowledge to be verified, sorted by priority, and pushed for manual review to obtain the results. The verified potential business rules and unknown anomaly patterns are added to the initial domain knowledge graph, the rule base and anomaly pattern base of the graph are updated, and the iteratively updated domain knowledge graph is output.

[0052] The core information extraction employs a strategy combining keyword extraction and entity recognition. Based on a bidirectional encoder-representation converter model, it extracts entities (such as terrain entities, environmental parameter entities, and equipment entities), relationships (such as terrain and elevation, sensor and monitoring values, anomalies and causes), and attributes (such as elevation values, monitoring time, and anomaly levels) from the high-precision fusion results. The entity recognition accuracy is no less than 97%, and the relationship extraction accuracy is no less than 95%. The normal mode refers to the stable distribution state of the entity-relationship-attribute triple based on domain-based business rules and historical compliant data. Specifically, it is characterized by entity attribute values ​​being within a preset reasonable range (e.g., water level 0.5-3 meters in water conservancy scenarios, vegetation coverage ≥60% in forestry scenarios), entity relationships conforming to business logic (e.g., stable "sensor-monitoring value" correspondence without logical contradictions), and smooth temporal change trends without abrupt changes (e.g., quarterly surface elevation change ≤0.05 meters). The initial domain knowledge graph is stored using a graph database (Neo4j). The graph structure contains entity-relationship-attribute triples, with nodes representing entities (no fewer than 1000), edges representing relationships between entities (no fewer than 2000), and attributes representing the specific parameters of entities and relationships. The initial graph also includes a basic rule base (covering 50+ basic domain business rules) and an anomaly pattern base (covering 30+ known anomaly patterns), providing rule query and anomaly matching capabilities.

[0053] The knowledge mining and anomaly detection module integrates association rule mining and anomaly detection algorithms. The association rule mining employs a frequent pattern growth algorithm, implemented using the Spark distributed computing framework with a minimum support of 0.2 and a minimum confidence of 0.8. Efficiency comparison experiments show that when processing 1 million data points, the prior algorithm takes 120 seconds, while the frequent pattern growth algorithm takes 72 seconds, representing a 40% efficiency improvement. When processing 5 million data points, the prior algorithm takes 680 seconds, while the frequent pattern growth algorithm takes 390 seconds, representing a 42.6% efficiency improvement. This demonstrates its ability to handle large-scale data processing. Based on the rapid processing function; anomaly identification combines isolated forest and density clustering algorithms. The method for determining the cluster radius of 0.5 for density clustering algorithm is as follows: the elbow rule is used to analyze the sum of squared errors within the cluster under different radii (0.3-0.7). A clear inflection point appears at the radius of 0.5, and the silhouette coefficient reaches the maximum value of 0.78, ensuring the optimal clustering effect; potential business rules are mined from the initial knowledge graph and high-precision fusion results (such as "if the soil moisture is less than 15% and the vegetation coverage is less than 30%, the vegetation is abnormally short of water"). The potential business rule set (no less than 20 rules) is output.

[0054] The prioritization of knowledge to be verified employs a combined risk level and confidence level ranking method. The priority score is equal to 0.6 multiplied by the risk level coefficient plus 0.4 multiplied by the confidence level. The risk level coefficient is (high risk: 1.0; medium risk: 0.7; low risk: 0.3), and the confidence level represents the reliability of the potential rule / anomaly cluster (range 0-1). Knowledge is ranked from highest to lowest score, with higher-priority knowledge (such as high-risk unknown anomaly clusters and high-confidence potential rules) being pushed to the manual review terminal first. Manual review uses a combination of single-person and dual-person review. Review indicators include rule rationality, anomaly accuracy, and business relevance. Review results are categorized into three types: passed, passed after modification, and rejected. The pass rate is no less than 85%. Knowledge that passes after modification is included in the update scope after a second review confirmation.

[0055] The knowledge graph is updated incrementally without rebuilding the entire graph: approved potential business rules are added to the rule base, updating their confidence levels and applicable scenarios; unknown anomaly patterns (anomaly cluster features, judgment criteria) are added to the anomaly pattern library, establishing the association between anomaly patterns and business rules; new entities, relationships, and attributes are added, and existing entity attribute values ​​and relationships are updated to ensure the timeliness of the graph data. The iteratively updated domain knowledge graph has at least a 20% increase in the number of entities compared to the initial graph, and at least a 30% increase in the number of rules and anomaly pattern libraries. It possesses dynamic knowledge updates and fast query capabilities (query latency not exceeding 5ms), and can be used for subsequent cross-regional model collaborative evolution and business report generation.

[0056] Furthermore, the specific implementation process of cross-regional model co-evolution is as follows: Each region deploys a local model, loads the iteratively updated domain knowledge graph, and performs local fine-tuning of the local model using local multi-source private data, outputting the model parameter update amount. Differential privacy technology is used to add Laplace noise to the model parameter update amount to complete privacy protection processing, outputting the encrypted parameter update amount. Each region uploads the encrypted parameter update amount to the central server, and the central server integrates the parameter update amounts of all regions through a parameter aggregation algorithm. Based on the aggregated parameter update amount, a global cross-regional co-evolutionary model is generated and distributed to each region. Each region updates its local model through the global model, realizing cross-regional model co-evolution.

[0057] The local models in each region are built on an improved transformer model, maintaining core consistency with the model architecture of the high-precision fusion module. At the same time, targeted improvements have been made to address local small sample fine-tuning, knowledge graph constraints, and low latency requirements. The overall parameter dimension is approximately 1 million, which is suitable for edge computing node deployment. It adopts a five-segment architecture of "input preprocessing layer → knowledge embedding layer → encoder module → decoder module → output layer". The core improvements are concentrated in the design of the knowledge embedding layer and the top adjustable layer, which retain the feature extraction capabilities of the high-precision fusion module while adapting to local data characteristics and knowledge constraints.

[0058] The input preprocessing layer receives 128-dimensional feature vectors from local multi-source private data. A new local data adaptation sublayer eliminates dimensional differences through layer normalization and linear projection, maps to 256 dimensions, and adds data type identifiers. The preprocessing time for a single data entry is less than 1 millisecond. The knowledge embedding layer, as a core improvement module, encodes over 50 business rules into 128-dimensional vectors through a rule embedding unit. These vectors are then fused with the input features via a gating mechanism. The entity-relationship embedding unit converts the "entity-relationship-attribute" triples into 64-dimensional relation vectors. Semantic alignment is enhanced using dot product attention, and the output is a 256-dimensional feature vector incorporating knowledge constraints. The input process takes no more than 2 milliseconds; the encoder module has a 6-layer structure, with each layer containing a multi-head self-attention sublayer and a feedforward neural network sublayer. The bottom 4 layers are completely frozen, and the top 2 layers have 30% adjustable parameters, adapting to the preservation of general features and local data adaptation; the decoder module has a 3-layer structure, with the addition of a local data adaptation sublayer and a decision calibration sublayer, dynamically adjusting feature weights and calibrating decision bias. The top 1 layer is fully adjustable, the middle 1 layer is partially adjustable, and the bottom 1 layer is frozen; the output layer has two branches: feature output and parameter update output, which output a 128-dimensional feature vector and a globally compatible parameter update vector, respectively. The end-to-end latency of the entire process does not exceed 200 milliseconds.

[0059] The model's structural adaptability fully supports the requirements of few-sample fine-tuning and knowledge constraints: the knowledge embedding layer dynamically integrates business rules through a gating mechanism, serving both as a loss function constraint to avoid overfitting and directly participating in feature processing, ensuring that local fine-tuning does not deviate from the domain's business logic; the top adjustable layer has approximately 300,000 adjustable parameters, accounting for 30% of the total parameters. Combined with the few-sample fine-tuning algorithm, only 10,000 sets of local data are needed for effective adaptation, while freezing 700,000 dimensions of the underlying core parameters to ensure the model's generalization ability; each layer of the model adopts a lightweight design, with the encoder feedforward neural network hidden layer dimension reduced to 512, and the single model inference memory usage not exceeding 2GB, fully adapting to the hardware resource limitations of edge computing nodes (clock speed ≥ 1.2GHz, memory ≥ 8GB), meeting the application requirements of low latency and high adaptability locally.

[0060] Differential privacy technology employs a Laplace mechanism to protect the privacy information in local model parameter updates (avoiding the inference of local private data from parameters). The noise addition is described as follows: the encrypted parameter update equals the original parameter update plus Laplace-based noise, calculated by dividing the sensitivity of the parameter update by the privacy budget. The sensitivity of the parameter update is dynamically adjusted based on data sensitivity: sensitivity for sensitive data (such as data from core emergency rescue areas) is set to 0.05, and sensitivity for non-sensitive data (such as general terrain data) is set to 0.2. The privacy budget is based on a privacy-utility curve balance: for high privacy requirements (such as monitoring classified areas), the privacy budget is set to 0.5, where the privacy protection strength (based on the ε-differential privacy definition) satisfies a privacy leakage probability ≤ 0.01 and a model performance loss ≤ 8%; for ordinary privacy requirements, it is set to 1.5, where the privacy protection strength results in a privacy leakage probability ≤ 0.03 and a model performance loss ≤ 3%, achieving an optimal balance between privacy and performance. The encrypted parameter update is further encrypted using a 256-bit encryption algorithm of the Advanced Encryption Standard to ensure security during transmission and prevent parameter tampering or leakage.

[0061] Each region and the central server use a dedicated 5G mobile communication network for transmission, with a transmission rate of no less than 100Mbps and a transmission latency of no more than 50ms. Communication compression technology (parameter sparsity + quantization encoding) is employed to compress parameter updates to 30% of their original size, with an information loss rate of ≤3%, significantly reducing communication overhead. Each region uploads encrypted parameter updates at fixed intervals (e.g., every 24 hours), enabling emergency updates (e.g., immediately uploading parameter updates when new anomalies occur locally). The central server deploys a parameter aggregation module, using a federated averaging algorithm to integrate parameter updates from all regions. The aggregation is expressed as follows: the globally aggregated parameter equals the sum of the weighted parameter updates from all regions divided by the sum of the weights of each region. The weight of each region is calculated as: weight = data volume weight × model performance weight. Data volume weight = data volume of that region / total data volume of all regions. Model performance weight = model accuracy of that region / average model accuracy of all regions, balancing the fairness of data volume and model performance. During the aggregation process, a Byzantine fault tolerance mechanism is introduced. A threshold voting method is used (if more than 60% of the regional parameters are consistent, they are retained) to remove the abnormal parameter update of malicious nodes. Experiments have verified that this mechanism can tolerate up to 33% of malicious nodes (that is, when the proportion of malicious nodes is ≤33%, the accuracy of aggregation parameters is ≥95%), and the abnormal parameter removal rate does not exceed 3%, ensuring the rationality of aggregation parameters.

[0062] The global cross-regional co-evolutionary model is generated based on aggregated parameters. After generation, performance testing is conducted (the test set uses shared public test data across regions, excluding private data). Test metrics include a fusion accuracy of no less than 99%, anomaly detection accuracy of no less than 98%, and processing latency of no more than 300ms. Once the test is passed, the model is distributed to each region. Each region receives the global model and uses a model fusion algorithm to merge the global model parameters with its local model parameters, updating the local model while preserving its adaptability to the local scenario. Simultaneously, it absorbs excellent parameters from other regions, achieving cross-regional model co-evolution. The co-evolution cycle can be dynamically adjusted. For regions with large data volumes and complex scenarios, the co-evolution cycle can be shortened (e.g., every 12 hours) to ensure continuous optimization of model performance.

[0063] Furthermore, the specific process for obtaining the structured business report is as follows: The system pre-sets differentiated requirement templates for different assessment areas, each containing specific report metrics, data dimensions, and format specifications. A scenario-based value conversion engine reads the output of the cross-regional collaborative evolution model and matches the corresponding requirement template to the target application area. Based on the matched requirement template, it extracts core data from the output of the cross-regional collaborative evolution model. The core data is then structured, including data classification and statistics, trend analysis, and visualization chart generation. The data is presented according to the template specifications, generating a structured business report for the corresponding area.

[0064] Among them, the differentiated demand templates cover four core areas: surveying and mapping monitoring, water conservancy management and control, forestry protection, and emergency rescue. Each area has 3-5 sub-scenario templates (such as the surveying and mapping monitoring field, which includes templates for topographic measurement and surface change monitoring). It can achieve compatibility with Extensible Markup Language / Word Processing Software formats and can be customized according to user needs. The core content of the templates for each field is designed differently: Surveying and Mapping Monitoring: Reporting indicators include topographic measurement accuracy, surface change area, and rate of change; data dimensions cover spatial location, elevation, and temporal changes; format specifications include measurement result tables, change comparison maps, and accuracy acceptance reports. Water Conservancy Management: Reporting indicators include water level, water quantity, water quality, and flood warning levels; data dimensions cover real-time monitoring values, historical comparison values, and warning thresholds; format specifications include monitoring time-series maps, warning reports, and disposal recommendations. Forestry Protection: Reporting indicators include vegetation coverage, pest and disease incidence, and growth level; data dimensions cover vegetation parameters, pest and disease characteristics, and spatial distribution; format specifications include vegetation growth maps, pest and disease distribution maps, and prevention and control recommendations. Emergency Rescue: Reporting indicators include disaster type, disaster scope, severity of damage, and rescue priority; data dimensions cover disaster characteristics, spatial location, and temporal changes; format specifications include disaster situation maps, rescue plans, and disposal progress.

[0065] The scenario-based value conversion engine integrates a domain identification module and a template matching module. The domain identification module uses a classification algorithm (with an accuracy rate of no less than 99%) to determine the target application domain based on scenario features (such as data types and parameter indicators) in the output of the cross-regional co-evolutionary model. The template matching module matches the optimal required template from the template library based on the target domain and sub-scenario, with a matching priority of: precise matching of sub-scenario ≥ general matching of domain ≥ custom template matching, and a matching latency of no more than 100ms. The output of the cross-regional co-evolutionary model includes high-precision fusion results, anomaly identification results, and knowledge graph association results, all of which are structured data, enabling rapid extraction of core data.

[0066] Core data extraction employs a template-driven extraction algorithm. Based on the report indicators and data dimensions in the matched template, corresponding data is extracted from the model output, achieving an accuracy of no less than 98% and an extraction efficiency of no less than 1000 records / second. The extracted core data includes real-time data, historical comparison data, anomaly data, and knowledge-related data, which are categorized and organized according to template requirements. Data classification and statistics utilize statistical analysis algorithms (such as mean, variance, and percentage statistics) to generate statistical reports. Trend analysis employs time-series trend fitting algorithms (such as linear fitting and exponential fitting) to analyze data trends (rising, falling, stable) and generate trend analysis conclusions. Visual chart generation utilizes an optimized chart visualization tool. For large data volume scenarios (data points > 10,000), the maximum triangle three-barrel downsampling algorithm is used. Accuracy comparison before and after downsampling: When data points reach 100,000… Downsampling to 10,000 points results in an error rate of ≤2%; when the data points are 1 million, downsampling to 10,000 points results in an error rate of ≤3%, ensuring that the core features of the chart are not distorted; the graphics processor acceleration is implemented based on the unified computing device architecture 11.0, supporting parallel rendering computation, combined with a hierarchical rendering strategy (prioritizing the rendering of core indicators and delaying the loading of secondary indicators), generating line charts (time series changes), bar charts (classification comparison), heat maps (spatial distribution), pie charts (percentage statistics), etc., with chart accuracy of no less than 99%, interactive viewing capabilities (such as zooming in, zooming out, and detailed query), and rendering latency of no more than 300ms under large data volumes.

[0067] The structured business report is automatically formatted according to a template, including modules such as report title, report summary, core data statistics, trend analysis, visualization charts, conclusions and recommendations, and report signature. It is compatible with word processing software and portable document formats, allowing for direct download, printing, and export. Report generation has a delay of no more than 500ms. After generation, it automatically performs format and content validation. Format validation ensures standardized layout (consistent font, line spacing, and page numbers), while content validation ensures accurate data, clear logic, and reasonable conclusions. If validation fails, it is automatically regenerated (with a maximum of two retries). The generated structured business report can be simultaneously pushed to relevant business terminals and supports manual editing (such as adding manual data and adjusting conclusions and recommendations) to meet the needs of different business scenarios.

[0068] Furthermore, the specific process of tracing the source of information across the entire supply chain is as follows: Collect process data from each stage, including data sources, processing algorithms, parameter configurations, and error records, to form a full-link process dataset. Establish a full-link data association index to bind the full-link process dataset to each indicator in the business structured report, forming a mapping relationship between indicators and process data. Construct a traceability query module to query the corresponding original data, processing flow, and error source for any indicator in the business structured report based on the mapping relationship between indicators and process data. Output full-link traceability information containing the association index and query results through the traceability query module.

[0069] The end-to-end process data covers data from all stages of the entire technical system, collected categorized by stage. The specific content collected at each stage includes: Data Acquisition Stage: Data source name, acquisition device number, acquisition time, acquisition parameters, raw data; Collaborative Acquisition Task Stage: Task parameters, device scheduling records, task execution logs; Dynamic Spatiotemporal Calibration Stage: Clean data, alignment algorithm parameters, calibration error, reliability assessment results; High-Precision Fusion Stage: Causal verification results, time-series feature data, fusion weights, fusion algorithm parameters; Knowledge Graph Construction Stage: Core information extraction records, latent rule mining logs, manual review results, graph update records; Cross-Regional Collaborative Evolution Stage: Local model fine-tuning parameters, encrypted parameter update volume, aggregation parameters, global model parameters; Report Generation Stage: Core data extraction records, statistical analysis results, visualization chart parameters. The process data is collected in real-time and stored in its entirety, with a storage period of no less than one year. It utilizes a combined relational database and file system for storage (structured process data is stored in the relational database, and unstructured data such as raw images is stored in the file system), enabling rapid data retrieval and reading.

[0070] The end-to-end data association index adopts a distributed index architecture, built on an elastic search cluster (3 master nodes + 6 data nodes). The hardware configuration is: each node has a CPU with ≥16 cores, ≥32GB of memory, and ≥2TB of solid-state drive, with horizontal scalability (the number of nodes can be dynamically increased to 12). Index fields include "Indicator ID, Stage Name, Data ID, Timestamp, Association Parameters, and Error Records," employing a sharding strategy (12 shards per index, 3 replicas), enabling multi-condition combined queries with a query latency of no more than 10ms. The association index creation process is as follows: First, a unique ID is assigned to each indicator in the business structured report, and a unique data ID is assigned to each piece of data in the end-to-end process data. Then, based on data flow and causal relationships, the indicator ID is bound to the corresponding process data ID, recording the binding relationship (e.g., the water level monitoring value indicator in the report is bound to the raw water level data from the collection stage, the water level calibration data from the calibration stage, and the water level fusion data from the fusion stage). Finally, a mapping relationship table between indicators and process data is generated and stored in the index database, enabling dynamic updates of the mapping relationship (e.g., automatically updating the mapping relationship when new process data is added).

[0071] The source tracing and query module is deployed on the core server, enabling access from multiple terminals including web pages and mobile devices. It offers various query methods (index name query, index ID query, time range query, and spatial range query), and incorporates a REDIS caching mechanism to cache high-frequency query results (cache validity of 1 hour, cache hit rate ≥80%). A query distribution strategy (based on consistent hashing algorithm) evenly distributes concurrent query requests across nodes in the elastic search cluster, supporting massive concurrent queries (no less than 1000 concurrent queries per second). Query operations are simple and convenient (enter query conditions and click query to obtain results). During the query process, based on the mapping relationship between indicators and process data, it traces back the entire process data corresponding to the indicator: raw data (initial data in the collection stage), processing flow (processing steps, algorithms, and parameters used in each stage), and error sources (error records, error confidence intervals, and error cause analysis in each stage). This achieves full-link reverse tracing from "indicator → fusion result → calibration data → raw data → collection device," with a tracing accuracy of no less than 99%.

[0072] The end-to-end traceability information is output in a structured format, including associated indexes (indicator ID, data ID, and relationships) and query results (raw data details, processing logs, error record details, and equipment information). It supports JSON and tabular output and can be directly downloaded and exported. The traceability information also includes a visual traceability chain diagram, intuitively displaying the end-to-end data flow and processing of the corresponding indicator. Users can click on a node to view its detailed data. Error source analysis is also provided, clarifying the impact of errors at each stage on the final indicator (e.g., calibration errors account for 30% of water level monitoring values), supporting data reliability assessment and equipment fault diagnosis. Furthermore, the traceability query module has log recording capabilities, recording all query operations (queryer, query time, query conditions, and query results) for easy subsequent auditing and traceability.

[0073] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include computer program products comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0074] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0075] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. An artificial intelligence-based multi-source spatiotemporal information fusion sensing method, characterized in that, The method includes: A three-source data acquisition architecture is constructed, a unified spatiotemporal benchmark is preset, multiple types of modal data sources are integrated, collaborative acquisition tasks are generated based on dynamic game theory using a multi-objective optimization algorithm, nested acquisition logic is configured through adversarial training, and an intelligent calibration model is used to dynamically calibrate the three-source data in spatiotemporal manner to obtain calibration data. The calibration data is correlated and verified by a causal reasoning model and the weight of interference factors is quantified. The domain constraint parameters are dynamically adjusted by a time-series feature coding model. A cross-modal alignment algorithm is used to achieve semantic fusion of multiple modalities. Differentiated fusion weights are assigned in combination with data reliability assessment. Deep fusion is completed by an intelligent fusion model and high-precision fusion results are output. Based on the fusion results, a domain knowledge graph is constructed. Potential business rules are mined and unknown abnormal patterns are identified through rule mining and anomaly recognition. Knowledge to be verified is marked and pushed for manual review, and the dynamic knowledge graph is obtained through iterative updates. Based on a collaborative learning framework combined with privacy protection technology, cross-regional model collaborative evolution is achieved, generating structured business reports and constructing full-link traceability information by associating updated knowledge graphs.

2. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 1, characterized in that, The specific process of generating collaborative data acquisition tasks based on the dynamic game theory using the multi-objective optimization algorithm is as follows: A multi-objective optimization function is defined, with the objectives of maximizing data integrity, minimizing acquisition cost, and optimizing task timeliness, and an objective function matrix is ​​constructed. Define the constraints of dynamic game, including equipment resource constraints, data quality constraints, and scenario adaptation constraints; The objective function matrix is ​​solved using a multi-objective optimization algorithm to generate multiple sets of candidate acquisition schemes; The candidate acquisition schemes are comprehensively evaluated, and the optimal candidate scheme is selected as the final collaborative acquisition task. The task execution parameters are then output.

3. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 2, characterized in that, The specific process of performing dynamic spatiotemporal calibration on the three-source data is as follows: Outlier removal and missing value filling are performed on the aggregated multi-modal data sources to obtain clean data; Based on a unified spatiotemporal benchmark, a dynamic time warping algorithm is used to align temporal data, and an image registration algorithm is used to align spatial data, achieving preliminary spatiotemporal alignment of the three-source data. The initially aligned data is input into a Bayesian neural network to model the probability distribution of satellite orbit deviation, UAV flight attitude error, and ground sensor drift. The calibrated data and the corresponding error confidence interval are output, and the error confidence interval is used as the data reliability assessment result.

4. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 3, characterized in that, The specific process for obtaining the high-precision fusion result is as follows: The calibrated data is input into the causal inference model. The counterfactual intervention module simulates the correlation after removing confounding factors, quantifies the weight of interfering factors, removes false causal correlations, and obtains valid correlation data verified by causality. The effective associated data is input into the temporal feature encoding model to extract the temporal features of the scene. The loss function coefficients of the domain constraint parameters are dynamically adjusted based on the features to obtain associated data with dynamic constraint weights. The associated data with dynamic constraint weights is split into explicit modal data and implicit modal data. A multimodal contrastive learning algorithm is used to map the two types of data to the same semantic space and output cross-modal alignment features. Based on the data reliability assessment results, differentiated fusion weights are assigned to cross-modal aligned features. The weighted features are then input into the improved Transformer model for feature-level and decision-level fusion, outputting high-precision fusion results.

5. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 4, characterized in that, The process of obtaining the iteratively updated knowledge graph is as follows: Based on the high-precision fusion results, core information is extracted to construct an initial domain knowledge graph; The initial domain knowledge graph and the high-precision fusion result are jointly input into the knowledge mining and anomaly identification module to mine potential business rules, identify unknown anomaly clusters that deviate from the normal pattern, and output the potential business rule set and unknown anomaly cluster data. Mark potential business rule sets and unknown anomaly cluster data as knowledge to be verified, sort them by priority and push them to manual review, and obtain the results of manual review; The approved potential business rules and unknown anomaly patterns are added to the initial domain knowledge graph, the rule base and anomaly pattern base of the graph are updated, and the iteratively updated domain knowledge graph is output.

6. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 5, characterized in that, The specific implementation process of the cross-regional model co-evolution is as follows: Each region deploys a local model, loads the iteratively updated domain knowledge graph, combines local multi-source private data to fine-tune the local model, and outputs the model parameter update amount; Differential privacy technology is used to add Laplacian noise to the model parameter update, thus completing the privacy protection process and outputting the encrypted parameter update; Each region uploads the encrypted parameter update data to the central server, which then integrates the parameter update data from all regions using a parameter aggregation algorithm. A global cross-regional co-evolutionary model is generated based on the aggregated parameter update amount and distributed to each region. Each region updates its local model through the global model, thereby achieving cross-regional model co-evolution.

7. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 6, characterized in that, The specific process for obtaining the structured business report is as follows: Pre-set templates to meet the differentiated needs of different assessment areas, including report indicators, data dimensions and format specifications specific to each area; The scenario-based value conversion engine reads the output of the cross-regional collaborative evolution model and matches the corresponding demand templates according to the target application area. Based on the matched demand template, core data is extracted from the output of the cross-regional co-evolutionary model. The core data is structured and organized, including data classification and statistics, trend analysis and visualization chart generation, and the data is presented according to template specifications to generate structured business reports for the corresponding fields.

8. The artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to claim 7, characterized in that, The specific process of tracing the associated full-link source information is as follows: Collect process data from each stage, including data sources, processing algorithms, parameter configurations, and error records for each stage, to form a full-link process dataset; Establish a full-link data association index to bind the full-link process dataset with each indicator in the business structured report, forming a mapping relationship between indicators and process data; Construct a traceability query module that, based on the mapping relationship between indicators and process data, allows reverse querying of the corresponding original data, processing flow, and error source for any indicator in the business structured report; The traceability query module outputs full-link traceability information, including related indexes and query results.

9. An artificial intelligence-based multi-source spatiotemporal information fusion sensing system, characterized in that, The system is used to execute the AI-based multi-source spatiotemporal information fusion sensing method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is executed by a processor to implement the artificial intelligence-based multi-source spatiotemporal information fusion sensing method according to any one of claims 1-8.

Images