Artificial Intelligence-Based Satellite Remote Sensing Big Data Processing and Analysis Methods and Systems

Abstract

This invention relates to the field of remote sensing data processing technology, and in particular to a satellite remote sensing big data processing and analysis method and system based on artificial intelligence. The method involves acquiring and preprocessing satellite big data to obtain a set of candidate variables, constructing a dynamic causal inference model to identify directed causal relationships between variables and generating a multidimensional causal association map. Based on this map, a counterfactual reasoning framework is constructed to simulate the causal effects of different band schemes on target identification accuracy. A decision support scheme is generated by combining resource constraints, and the model is corrected online using the deviation between the actual execution effect and the expected causal effect, thereby improving the accuracy and adaptability of remote sensing data analysis.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, and in particular to a method and system for satellite remote sensing big data processing and analysis based on artificial intelligence. Background Technology

[0002] Satellite remote sensing data processing and analysis technologies have been widely applied in fields such as agricultural monitoring, environmental protection, and urban planning. Current conventional practices typically involve preprocessing multi-source remote sensing images, such as radiometric and geometric corrections, then extracting spectral, textural, or spatial features, and finally using statistical models or machine learning algorithms (such as support vector machines, random forests, and convolutional neural networks) to classify ground features or retrieve parameters. These methods often rely on a large number of manually labeled samples, training the model by minimizing prediction errors, and ultimately outputting ground feature identification results or parameter estimates. Data fusion strategies are mostly based on empirical rules or correlation analysis, such as selecting specific band combinations to improve classification accuracy; however, this selection lacks a deep understanding of the intrinsic mechanisms between observed variables and target objects.

[0003] Existing conventional approaches suffer from two main drawbacks. First, most methods rely on statistical correlation rather than causal logic. Spurious correlations often exist between spectral reflectance and surface parameters; for example, cloud cover or atmospheric scattering can cause false correlations between observed values ​​and actual land cover attributes. Conventional models merely learn the data distribution patterns to fit these correlations, and their performance significantly deteriorates and their generalization ability is insufficient when the data distribution shifts due to seasonal, regional, or sensor variations. Second, traditional analytical frameworks lack the ability to assess the causal effects of decision-making options. In practical applications, it is necessary to select the optimal band combination or preprocessing procedure based on specific objectives (such as improving the accuracy of identifying a certain land cover). However, conventional methods can only provide correlation recommendations and cannot simulate the actual impact of changing an observation parameter (such as replacing a band) on the final identification result, let alone quantitatively compare multiple candidate options. This correlation-driven decision support often leads to wasted resources or poor results because the selected optimal solution may fail in new data environments. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for satellite remote sensing big data processing and analysis based on artificial intelligence, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for processing and analyzing satellite remote sensing big data based on artificial intelligence, comprising:

[0006] A set of big data to be analyzed is obtained from remote sensing satellites, and the set of big data to be analyzed is preprocessed to form a set of candidate variables;

[0007] A dynamic causal inference model is constructed. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, the directed causal relationship between remote sensing observation parameters, surface physical processes and target land cover characteristics is identified and the causal strength is quantified. At the same time, the directed causal relationship is traced back to the confounding factors and the causal validity is verified. False causal paths are pruned to obtain a multidimensional causal association map.

[0008] Based on the directed causal relationships in the multidimensional causal relationship map, a counterfactual reasoning framework is constructed with the target feature recognition accuracy as the decision target variable. The propagation path of different band combination schemes along the directed causal relationships under different weight configurations is simulated, and the expected causal effect of each band selection scheme on the target feature recognition accuracy and its confidence interval are calculated.

[0009] The expected causal effect is cross-validated with computational resource constraints to generate a decision support scheme set containing remote sensing data processing schemes. The actual execution effect of the remote sensing data processing scheme is collected, and the deviation between the actual execution effect and the expected causal effect is used to correct the causal strength weight and confusion factor verification threshold in the dynamic causal inference model online.

[0010] A dynamic causal inference model is constructed. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, the directed causal relationships between remote sensing observation parameters, surface physical processes, and target land cover characteristics are identified, and the causal strength is quantified.

[0011] For a pair of variables in the candidate variable set, the conditional mutual information of the pair of variables under a given set of conditional variables is calculated within a sliding time window. When the conditional mutual information converges to a stable value as the set of conditional variables expands, the set of conditional variables corresponding to the stable value is determined as the minimum separating set. Based on the minimum separating set, it is determined whether there is a direct causal path between the pair of variables.

[0012] Traverse all variable pairs in the candidate variable set, mark variable pairs with direct causal paths as candidate causal edges, and construct an initial undirected causal skeleton graph;

[0013] For each candidate causal edge in the initial undirected causal skeleton graph, the causal direction is oriented using time sequence constraints and non-cyclic constraints. By evaluating the variance of the prediction residual of the source variable to the target variable under different causal direction assumptions, the direction that minimizes the variance of the prediction residual is selected as the true causal direction, thus forming a directed causal graph.

[0014] For each directed edge in the directed causal graph, the contribution coefficient of the change in the source variable to the change in the target variable is calculated, and the contribution coefficient is used as the causal strength corresponding to the directed edge.

[0015] The causal direction is oriented using chronological order constraints and non-cyclic constraints. By evaluating the variance of the prediction residuals of the source variable on the target variable under different causal direction assumptions, the direction that minimizes the variance of the prediction residuals is selected as the true causal direction, forming a directed causal graph, including:

[0016] For each candidate causal edge in the initial undirected causal skeleton graph, the temporal priority is determined according to the difference in satellite overpass imaging timestamps of the connected variable pairs. The variables with the higher temporal priority are taken as causal candidates, and the variables with the lower temporal lag are taken as effect candidates. A positive causal hypothesis is established, and a negative hypothesis corresponding to the positive causal hypothesis is constructed as a control.

[0017] For the positive causal hypothesis and the negative hypothesis, it is examined whether their combination with the oriented edges in the initial undirected causal skeleton graph forms a directed loop. Hypotheses that violate the non-cyclic constraint are marked as invalid hypotheses, while valid hypotheses that satisfy the constraint are retained.

[0018] Using the historical observations of the causal variables in the effective hypothesis, the current observations of the effect variables are predicted. The residual variance between the actual observations and the predicted values ​​of the effect variables is calculated as the fitting error. The fitting errors of the positive causal hypothesis and the negative hypothesis are compared, and the hypothesis direction with the smallest fitting error is selected as the orientation result of the candidate causal edge.

[0019] The orientation results are applied to the initial undirected causal skeleton graph, the set of oriented edges is updated, and the constraint check and error comparison are repeated until all candidate causal edges are oriented, forming a directed causal graph.

[0020] The directed causal relationship is subjected to reverse tracing of confounding factors and verification of causal validity, and false causal paths are pruned to obtain a multidimensional causal relationship graph, including:

[0021] For each directed causal relationship in the directed causal graph, a common predecessor node that can reach the cause variable and effect variable of the directed causal relationship is searched in reverse in the directed causal graph. A causal propagation tree starting from the common predecessor node is constructed. The direct propagation branch through the cause variable and the bypass propagation branch that bypasses the cause variable in the causal propagation tree are decomposed. When the cumulative contribution of the bypass propagation branch to the effect variable accounts for more than the proportion of the observed causal strength, the common predecessor node is marked as a confounding factor.

[0022] Block the bypass propagation branch, calculate the net causal contribution of the cause variable to the effect variable, and when the decrease in the net causal contribution relative to the original observed causal strength exceeds the validity threshold, determine that the directed causal relationship is a false causal path and remove it; otherwise, retain it and update its causal strength with the net causal contribution.

[0023] The directed causal graph is traversed to complete the verification and pruning, and the retained directed causal relationships and their corrected causal strengths are constructed into a multidimensional causal association graph.

[0024] Based on the directed causal relationships in the multidimensional causal relationship map, a counterfactual reasoning framework is constructed with the target feature identification accuracy as the decision objective variable. This framework simulates the propagation path of different band combination schemes along the directed causal relationships under different weight configurations, calculating the expected causal effect of each band selection scheme on the target feature identification accuracy and its confidence interval, including:

[0025] In the multidimensional causal relationship map, the accuracy of target feature identification is used as the decision target variable. All directed causal relationship paths pointing to the decision target variable are traced backward, and the variables at the starting point of the directed causal relationship path are identified as the set of interventionable variables.

[0026] For the operable variables in the set of operable variables, the counterfactual values ​​of the operable variables under the band selection and adjustment state are set, and a change propagation operator of the operable variables from the current observation value to the counterfactual value is constructed. Along the directed causal relationship path from the operable variables to the decision target variable, the change propagation operator and the causal intensity of each segment on the directed causal relationship path are chained together to obtain the path causal effect.

[0027] For multiple parallel paths from the modifiable variable to the decision target variable, the path causal effects of the multiple parallel paths are aggregated to obtain the total causal effect of the counterfactual value on the decision target variable;

[0028] Based on the historical observation data of the interventionable variables, multiple sample sets are generated through perturbation resampling. For each sample set, path propagation and effect convergence are repeatedly performed to obtain the distribution characteristics of the total causal effect. The quantile intervals of the distribution characteristics are extracted as the expected causal effect and confidence interval of the intervention plan.

[0029] Cross-validating the expected causal effects with computational resource constraints generates a set of decision support solutions that includes remote sensing data processing schemes, including:

[0030] Obtain a set of computing resource constraints, which includes data transmission bandwidth constraints and recognition accuracy achievement constraints.

[0031] For each band selection scheme, the counterfactual values ​​of the modifiable variables in the band selection scheme are compared with the data transmission bandwidth constraints for feasibility verification, and the expected causal effect corresponding to the band selection scheme is compared with the recognition accuracy achievement constraints for conformity verification. When the counterfactual values ​​and the expected causal effect pass both the feasibility verification and the conformity verification, the band selection scheme is marked as a constraint-satisfied scheme.

[0032] For the constraint-satisfaction scheme, the expected causal effect confidence interval of the constraint-satisfaction scheme is extracted, the interval width of the expected causal effect confidence interval is calculated as the prediction uncertainty metric, and the distance between the lower bound of the expected causal effect confidence interval and the lower bound of the recognition accuracy achievement constraint is calculated as the target deviation metric.

[0033] Construct a risk assessment vector, and encapsulate the constraint-satisfying scheme, the expected causal effect, the confidence interval of the expected causal effect, and the risk assessment vector into a decision support scheme. Summarize all decision support schemes corresponding to the constraint-satisfying schemes to form a decision support scheme set.

[0034] The process of collecting the actual execution results of the remote sensing data processing scheme and using the deviation between the actual execution results and the expected causal effect to perform online correction of the causal strength weight and confounding factor verification threshold in the dynamic causal inference model includes:

[0035] After the implementation of the band selection scheme, the actual observed value of the target feature identification accuracy is collected. The difference between the actual observed value and the baseline observed value before the scheme adjustment is calculated as the actual execution effect. The difference between the actual execution effect and the expected causal effect is used as the effect deviation. The effect deviation is decomposed in reverse along the directed causal relationship path to obtain the contribution component of each segment of the directed causal relationship. The corresponding causal strength weight is corrected based on the contribution component.

[0036] The potential explanatory power of the bypass propagation branch for the effect bias is calculated based on the modified causal strength weight, and the confusion factor verification threshold is adjusted according to the comparison result of the potential explanatory power and the confusion reassessment triggering condition.

[0037] A second aspect of this invention provides an artificial intelligence-based satellite remote sensing big data processing and analysis system, comprising:

[0038] The data preparation unit is used to acquire a large dataset of data to be analyzed from remote sensing satellites, and to preprocess the dataset to form a set of candidate variables.

[0039] The causal inference unit is used to construct a dynamic causal inference model. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, it identifies the directed causal relationship between remote sensing observation parameters, surface physical processes and target land cover characteristics and quantifies the causal strength. At the same time, it performs reverse tracing of confounding factors and verification of causal validity on the directed causal relationship, prunes false causal paths, and obtains a multidimensional causal association map.

[0040] The causal effect unit is used to construct a counterfactual reasoning framework based on the directed causal relationship in the multidimensional causal relationship map, with the target feature identification accuracy as the decision target variable, to simulate the propagation path of different band combination schemes along the directed causal relationship under different weight configurations, and to calculate the expected causal effect of each band selection scheme on the target feature identification accuracy and its confidence interval.

[0041] The online correction unit is used to cross-validate the expected causal effect with the computational resource constraints, generate a decision support scheme set containing remote sensing data processing schemes, collect the actual execution effect of the remote sensing data processing schemes, and use the deviation between the actual execution effect and the expected causal effect to correct the causal strength weight and confusion factor verification threshold in the dynamic causal inference model online.

[0042] A third aspect of the present invention provides an electronic device, comprising:

[0043] processor;

[0044] Memory used to store processor-executable instructions;

[0045] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0046] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0047] By employing a dynamic causal inference model to perform temporal conditional independence tests and structure learning on the candidate variable set, physically meaningful directed causal relationships can be accurately extracted from high-dimensional, nonlinear remote sensing data. This reveals the intrinsic driving mechanism between remote sensing observation parameters and surface physical processes and target feature characteristics. The reverse tracing of confounding factors and the causal validity verification mechanism effectively eliminate spurious associations caused by environmental noise and observation errors. The pruned multidimensional causal association map significantly improves the robustness and interpretability of causal inference, providing a highly reliable causal knowledge framework for subsequent decision-making and avoiding misleading conclusions that are prone to occur in traditional correlation analysis.

[0048] A counterfactual reasoning framework based on a multidimensional causal correlation graph can simulate the propagation effect of various band combinations along directed causal paths to improve target feature identification accuracy. It accurately calculates the expected causal effects and their confidence intervals, enabling the evaluation of the potential effects of each scheme without relying on a large number of field observation samples. Cross-validation of the expected causal effects with computational resource constraints generates a decision support scheme set that balances improved identification accuracy with controlled resource consumption, effectively balancing causal gain and implementation cost. This significantly enhances the economy and relevance of remote sensing data processing schemes, improving the accuracy and stability of target feature identification.

[0049] The deviation between the actual implementation results and the expected causal effects is collected and fed back to the dynamic causal inference model. The causal strength weights and confounding factor verification thresholds are adjusted online, enabling the model to possess continuous adaptive capabilities. This allows it to sensitively respond to changes in data distribution, sensor degradation, and evolution of the land surface environment, maintaining high-precision causal inference and decision optimization performance over the long term. This closed-loop correction mechanism endows the system with strong robustness and sustainability in dynamic remote sensing application scenarios, ensuring the causal knowledge base is gradually iterated and improved, thereby continuously enhancing the comprehensive effectiveness of remote sensing big data analysis. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the satellite remote sensing big data processing and analysis method based on artificial intelligence, as described in an embodiment of the present invention.

[0051] Figure 2 This is a flowchart illustrating the generation process of remote sensing data decision support based on causal inference, as described in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the satellite remote sensing big data processing and analysis method based on artificial intelligence, as described in an embodiment of the present invention.

[0055] Artificial intelligence-based methods for satellite remote sensing big data processing and analysis include:

[0056] A set of big data to be analyzed is obtained from remote sensing satellites, and the set of big data to be analyzed is preprocessed to form a set of candidate variables;

[0057] A dynamic causal inference model is constructed. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, the directed causal relationship between remote sensing observation parameters, surface physical processes and target land cover characteristics is identified and the causal strength is quantified. At the same time, the directed causal relationship is traced back to the confounding factors and the causal validity is verified. False causal paths are pruned to obtain a multidimensional causal association map.

[0058] Based on the directed causal relationships in the multidimensional causal relationship map, a counterfactual reasoning framework is constructed with the target feature recognition accuracy as the decision target variable. The propagation path of different band combination schemes along the directed causal relationships under different weight configurations is simulated, and the expected causal effect of each band selection scheme on the target feature recognition accuracy and its confidence interval are calculated.

[0059] The expected causal effect is cross-validated with computational resource constraints to generate a decision support scheme set containing remote sensing data processing schemes. The actual execution effect of the remote sensing data processing scheme is collected, and the deviation between the actual execution effect and the expected causal effect is used to correct the causal strength weight and confusion factor verification threshold in the dynamic causal inference model online.

[0060] In one optional implementation, a dynamic causal inference model is constructed. This model identifies the directed causal relationships between remote sensing observation parameters, surface physical processes, and target land cover characteristics by performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, and quantifies the causal strength, including:

[0061] For a pair of variables in the candidate variable set, the conditional mutual information of the pair of variables under a given set of conditional variables is calculated within a sliding time window. When the conditional mutual information converges to a stable value as the set of conditional variables expands, the set of conditional variables corresponding to the stable value is determined as the minimum separating set. Based on the minimum separating set, it is determined whether there is a direct causal path between the pair of variables.

[0062] Traverse all variable pairs in the candidate variable set, mark variable pairs with direct causal paths as candidate causal edges, and construct an initial undirected causal skeleton graph;

[0063] For each candidate causal edge in the initial undirected causal skeleton graph, the causal direction is oriented using time sequence constraints and non-cyclic constraints. By evaluating the variance of the prediction residual of the source variable to the target variable under different causal direction assumptions, the direction that minimizes the variance of the prediction residual is selected as the true causal direction, thus forming a directed causal graph.

[0064] For each directed edge in the directed causal graph, the contribution coefficient of the change in the source variable to the change in the target variable is calculated, and the contribution coefficient is used as the causal strength corresponding to the directed edge.

[0065] In constructing a dynamic causal inference model, for each pair of variables in the candidate variable set, their conditional independence relationship needs to be evaluated within a sliding time window. The sliding time window design takes into account both the temporal continuity of remote sensing data and the dynamic characteristics of surface physical processes; the window length can be adaptively adjusted according to the seasonal variation cycle or observation frequency of the target ground features. For variable pairs within the window, the statistical dependence between the two variables is measured by calculating the conditional mutual information under a given set of conditional variables. Let the variables... With variables The variable pairs to be tested are, and the set of condition variables is... Then conditional mutual information is defined as Its measurement is known In the case of and The strength of the residual statistical association between them.

[0066] With the set of condition variables The gradual expansion of conditional mutual information It will gradually decrease as more common influencing factors are brought under control. Follow When the expansion no longer changes significantly and converges to a certain stable value, the corresponding set of condition variables is determined as the minimal separating set. The physical meaning of the minimal separating set is that it contains exactly all sets that can block... and Intermediate or codependent variables that are statistically correlated with each other make it possible to control for... back, and The conditional mutual information approaches zero. If in the minimum separating set... Down If the value is significantly greater than zero (statistical significance is determined by permutation test or kernel density estimation), then it is considered... and If a direct causal path exists between them, otherwise determine that the two are in a given... The post-conditions are independent, and there is no direct causal connection. This test mechanism effectively distinguishes between the true physical causal relationship between remote sensing observation parameters and spurious correlations caused by common driving factors (such as solar radiation intensity, atmospheric water vapor content, etc.).

[0067] After traversing all variable pairs in the candidate variable set and performing conditional independence tests, all variable pairs determined to have a direct causal path are marked as candidate causal edges. Using each variable in the candidate variable set as a node and the candidate causal edges as connections, an initial undirected causal skeleton graph is constructed. At this stage, the skeleton graph only records the connection topology between variables; the direction of each edge has not yet been determined. The sparsity of the skeleton graph directly reflects the density of effective causal associations in the candidate variable set. An overly dense skeleton graph often indicates that the significance threshold for the conditional independence test is set too leniently and needs to be appropriately tightened to reduce the introduction of noisy edges.

[0068] After constructing the skeleton graph, each candidate causal edge is oriented. This orientation process utilizes two types of constraints: temporal sequence constraints and non-cyclic constraints. Temporal sequence constraints are based on the fundamental principle of causality, meaning the cause precedes the effect in time. By examining the temporal lag cross-correlation characteristics of two variables, the direction of the temporally preceding variable is prioritized as the source variable. Non-cyclic constraints require that the final directed graph does not contain directed loops to ensure the logical consistency of the causal structure. For candidate causal edges whose direction cannot be uniquely determined by temporal sequence constraints alone, the prediction residual variance under different causal direction assumptions is further evaluated to aid in the judgment: [The text then describes different assumptions and their implications.] and Two directions are considered. Under each assumption, regression prediction of the target variable is performed using the source variables, and the variance of the prediction residuals is calculated. Let the directional assumptions be... The variance of the predicted residuals is directional assumption The variance of the predicted residuals is The direction that minimizes the variance of the predicted residuals is selected as the true causal direction. The theoretical basis for this criterion is that the function mapping relationship under the true causal direction is usually simpler and smoother, with a relatively smaller residual variance, while the reverse mapping often requires the introduction of more complex nonlinear structures to fit the data, resulting in a relatively larger residual variance. After the above orientation steps, the initial undirected causal skeleton graph is transformed into a directed causal graph.

[0069] After the directed causal graph is constructed, the causal strength of each directed edge in the graph is quantified. Let the source variable of a certain directed edge be... The target variable is The quantification of causal strength is achieved through calculation. Change in amount The contribution coefficient of the change is used to achieve this. Specifically, in the control directed causal graph, other pairs... Given a parent node variable with a direct causal influence, for Apply a unit perturbation and observe. The magnitude of the response change is defined as the ratio of this magnitude of the response change to the unit disturbance, which is then defined as the contribution coefficient. ,Right now Contribution coefficient The absolute value reflects the strength of the causal influence, while the sign reflects the direction of the influence (positive promotion or negative inhibition). In specific applications of remote sensing data, such as the causal strength of near-infrared reflectance (as a source variable) on vegetation index (as a target variable), it can be precisely quantified by assessing the change in vegetation index caused by a unit change in near-infrared reflectance, while controlling other parent nodes such as red band reflectance. The contribution coefficients corresponding to all directed edges are then used. Assigning values ​​to the edges of the directed causal graph creates a weighted multidimensional causal relationship graph, providing a quantitative basis for the causal propagation path for the subsequent construction of a counterfactual reasoning framework.

[0070] Throughout the temporal conditional independence test and structure learning process, the step size of the sliding time window needs to be matched with the revisit cycle of the remote sensing satellites to ensure that each time window contains a sufficient number of valid observation samples, avoiding excessive bias in conditional mutual information estimation due to insufficient sample size. For high-dimensional candidate variable sets, the search space of the conditional variable set grows exponentially with the number of variables. A heuristic search algorithm based on the maximum-minimum parent node priority strategy can be used to accelerate the identification process of the minimum separating set, effectively reducing computational complexity while ensuring test accuracy. In addition, to address the common problem of missing cloud cover values ​​in remote sensing data, interpolation and completion of the time series data or robust estimation methods capable of handling missing data need to be performed before calculating conditional mutual information to ensure the reliability of the causal structure learning results.

[0071] In one optional implementation, the causal direction is oriented using chronological order constraints and non-cyclic constraints. By evaluating the variance of the prediction residuals of the source variable on the target variable under different causal direction assumptions, the direction that minimizes the variance of the prediction residuals is selected as the true causal direction, forming a directed causal graph, including:

[0072] For each candidate causal edge in the initial undirected causal skeleton graph, the temporal priority is determined according to the difference in satellite overpass imaging timestamps of the connected variable pairs. The variables with the higher temporal priority are taken as causal candidates, and the variables with the lower temporal lag are taken as effect candidates. A positive causal hypothesis is established, and a negative hypothesis corresponding to the positive causal hypothesis is constructed as a control.

[0073] For the positive causal hypothesis and the negative hypothesis, it is examined whether their combination with the oriented edges in the initial undirected causal skeleton graph forms a directed loop. Hypotheses that violate the non-cyclic constraint are marked as invalid hypotheses, while valid hypotheses that satisfy the constraint are retained.

[0074] Using the historical observations of the causal variables in the effective hypothesis, the current observations of the effect variables are predicted. The residual variance between the actual observations and the predicted values ​​of the effect variables is calculated as the fitting error. The fitting errors of the positive causal hypothesis and the negative hypothesis are compared, and the hypothesis direction with the smallest fitting error is selected as the orientation result of the candidate causal edge.

[0075] The orientation results are applied to the initial undirected causal skeleton graph, the set of oriented edges is updated, and the constraint check and error comparison are repeated until all candidate causal edges are oriented, forming a directed causal graph.

[0076] After obtaining the initial undirected causal skeleton graph, each candidate causal edge in the skeleton graph needs to be oriented to ultimately form a directed causal graph that reflects the true causal propagation path. The orientation process relies on two core principles: one is the temporal sequence information naturally carried by satellite remote sensing data, and the other is statistical testing criteria based on the comparison of prediction residual variances. These two principles work together to ensure the physical rationality and statistical consistency of the orientation results.

[0077] For each candidate causal edge in the skeleton graph, the satellite transit imaging timestamps connecting the variables at both ends of the edge are first extracted. Remote sensing data is typically collected in units of orbital periods, and the observation times of different variables naturally differ. Let a candidate edge connect the variables... With variables ,like The corresponding imaging timestamps are generally earlier than The corresponding imaging timestamp will be It was identified as a time-priority variable and used as a candidate cause in the positive causal hypothesis. As a candidate for outcome, a positive causal hypothesis is established. Simultaneously, construct corresponding reverse hypotheses. As a control group, in practice, the determination of time series differences is based on the average time difference over multiple observation periods to avoid misjudgment caused by a single orbital disturbance. When the timestamp differences between the two variables are not significant (i.e., the time difference is lower than the set minimum resolution threshold), both the positive and negative hypotheses are retained, and the final determination is made by the subsequent residual variance comparison step.

[0078] After establishing the positive and negative hypotheses, acyclic constraint checks need to be performed on each candidate edge for both hypotheses. Directed causal graphs require a directed acyclic graph structure, meaning no directed cycles are allowed. During the process of orienting each edge of the skeleton graph, as the set of oriented edges expands, newly added directed edges and existing oriented edges together form cycles. Regarding the positive hypothesis... Temporarily add it to the currently directed edge set, and perform a depth-first search or topological sorting check to determine if there exists a path from it. Depart via existing path and return A directed path exists; if it exists, the positive hypothesis violates the acyclicity constraint and is marked as a invalid hypothesis. For the negative hypothesis... Perform the same testing procedure. After constraint testing, if both the positive and negative hypotheses are valid, proceed to the residual variance comparison stage; if only one is valid, directly adopt the valid hypothesis as the orientation result; if both are invalid, temporarily set aside the side and re-evaluate after the other sides have been oriented.

[0079] For valid hypotheses that pass the non-cyclic constraint test, the variance of the predicted residuals is used as the orientation indicator. This is based on the positive hypothesis. For example, using variables From a series of observations at several historical moments, methods such as nonlinear regression or Gaussian process regression are used to construct a regression model for the variables. The prediction model for the current observation value is calculated. The residual sequence between the actual observed values ​​and the predicted values ​​is used to obtain the residual variance under the positive assumption. Similarly, regarding the converse hypothesis... Perform the same process, to Historical observations prediction The residual variance under the reverse assumption is calculated from the current observations. In remote sensing scenarios, the input features of the prediction model can also incorporate auxiliary variables such as land cover type and seasonal cycle coding to improve prediction accuracy and reduce the interference of confusion effects on residual variance estimation.

[0080] Compare and The hypothesis direction with smaller residual variance is selected as the orientation result for the candidate causal edge. Smaller residual variance means that when predicting the effect variable using historical information of the cause variable, the model fitting error is lower, indicating that the causal propagation relationship under this direction has stronger statistical support. When the difference between residual variances is small (i.e., the difference between the two is below the preset significance threshold), the judgment result of time priority is given priority, and the time priority variable is used as the cause to avoid instability in orientation due to statistical fluctuations. For satellite remote sensing data, the causal propagation between physical quantities such as surface reflectance, vegetation index, and surface temperature often has clear physical mechanism support. Based on the comparison of residual variances, domain prior constraints can be introduced to assign higher confidence weights to specific physical paths, further improving the physical rationality of the orientation results.

[0081] The orientation results of each candidate edge are sequentially written into the set of oriented edges, updating the current structure of the directed causal graph. During the update process, the acyclic constraint check is retried after each new directed edge is added to ensure that the entire graph structure always meets the requirements of a directed acyclic graph. For candidate edges that were previously shelved because both hypotheses were invalid, constraint checks and residual variance comparisons are performed again after other edges are oriented and the set of oriented edges is updated, until all candidate edges are oriented or are identified as unorientable edges (i.e., violating the acyclic constraint in any direction under the current graph structure; such edges usually correspond to synchronous response relationships between variables and require special handling in the subsequent causal strength quantification stage).

[0082] After all candidate causal edges are oriented, a complete directed causal graph is obtained. This graph uses remote sensing observation parameters as source nodes, surface physical process parameters as intermediate nodes, and target feature features as sink nodes, clearly depicting the causal propagation chain between remote sensing data acquisition, surface physical response, and feature identification results. The subsequent counterfactual reasoning framework will use this directed causal graph as a basis to simulate the causal effects of different band combinations on the accuracy of target feature identification along the directed causal path, thus providing interpretable quantitative evidence for optimizing remote sensing data processing schemes.

[0083] In one optional implementation, the directed causal relationship is subjected to reverse tracing of confounding factors and verification of causal validity, and false causal paths are pruned to obtain a multidimensional causal association graph, including:

[0084] For each directed causal relationship in the directed causal graph, a common predecessor node that can reach the cause variable and effect variable of the directed causal relationship is searched in reverse in the directed causal graph. A causal propagation tree starting from the common predecessor node is constructed. The direct propagation branch through the cause variable and the bypass propagation branch that bypasses the cause variable in the causal propagation tree are decomposed. When the cumulative contribution of the bypass propagation branch to the effect variable accounts for more than the proportion of the observed causal strength, the common predecessor node is marked as a confounding factor.

[0085] Block the bypass propagation branch, calculate the net causal contribution of the cause variable to the effect variable, and when the decrease in the net causal contribution relative to the original observed causal strength exceeds the validity threshold, determine that the directed causal relationship is a false causal path and remove it; otherwise, retain it and update its causal strength with the net causal contribution.

[0086] The directed causal graph is traversed to complete the verification and pruning, and the retained directed causal relationships and their corrected causal strengths are constructed into a multidimensional causal association graph.

[0087] After identifying directed causal relationships and quantifying causal strength, it is necessary to systematically trace back the confounding factors and verify the causal validity of the obtained directed causal graph in order to eliminate false association paths driven by common causes and ensure that each directed edge retained in the final multidimensional causal association graph has real causal explanatory power.

[0088] For each directed causal relationship in a directed causal graph, the causal variable is denoted as... The result-end variable is denoted as Perform a reverse graph search within the directed causal graph. Specifically, from... and Each person starts from the beginning and traverses backwards along the directed edges to collect all items that can be reached simultaneously. and All ancestor nodes are used to form a common predecessor node candidate set. In remote sensing scenarios, common precursor nodes often correspond to physical quantities such as solar altitude angle, atmospheric optical thickness, and terrain slope that simultaneously affect multiple remote sensing observation parameters. If these variables are not included in the condition set, they are highly susceptible to... and This creates a false statistical correlation between them.

[0089] for Each common predecessor node ,by Construct a causal propagation tree for the root node, enumerating from... Departure can reach All propagation paths. These paths are decomposed into two categories: those via dependent variables. The direct propagation branch (denoted as) ) and bypass Direct arrival The bypass propagation branch (denoted as) Bypass propagation branch representative right The independent influence, if strong enough, will cause Non-causal components are mixed into the observed correlations.

[0090] To quantify the side propagation branch on the resultant variable The cumulative contribution is obtained by summing the chained products of the causality strength coefficients on each bypass path. .Will Causal strength with original observations Compare and calculate the percentage of bypass contribution. .when Exceeding the preset confusion recognition threshold At that time, Marked for causal relationship The system identifies and records the set of side propagation paths of confusing factors. The confusion identification threshold is then defined. It can be adaptively adjusted according to the data noise level and sample size. In satellite remote sensing big data scenarios, it is usually set in the range of 0.15 to 0.30 to balance detection sensitivity and false alarm rate.

[0091] After marking the confounding factors, blocking operations are performed on all marked bypass propagation branches. The blocking method adopts the backdoor path closure criterion. right When calculating the net causal contribution, all confounding factors are considered. The fixed value of is equivalent to setting in the statistical model. Conditional processing is performed. After blocking... right Net causal contribution Only retained direct drive The part and the contribution of the legitimate mediation path via the intermediate variable remove the confusing components introduced by the common predecessor node.

[0092] Net causal contribution After the calculation is completed, a causal validity check needs to be performed, calculating the attenuation of the net causal contribution relative to the original observed causal strength. Its expression is .when Exceeding the validity threshold When, it indicates the original observation The association is largely driven by confounding factors rather than a genuine causal mechanism; therefore, this directed causal relationship is classified as a spurious causal path and removed. Validity threshold. The setting needs to be combined with domain knowledge. In remote sensing ground object recognition tasks, if the net causal contribution of a causal path decreases by more than 50% after confusion correction, it is generally considered that the path does not have sufficient causal explanatory power and should be deleted from the directed causal graph. Not exceeding Then determine For a true causal relationship, Replace the original causal strength to complete the correction of this directed edge.

[0093] For scenarios involving multiple confounding factors, all labeled confounding factors should be blocked simultaneously, rather than processed individually. This is because the bypass paths of different confounding factors overlap; blocking them individually would lead to repeated deduction of bypass contributions, thus underestimating the net causal contribution. In implementation, a joint conditionalization model encompassing all confounding factors is constructed to calculate the causal contribution in one step, under the condition that all confounding factor values ​​are fixed. right The net effect ensures the accuracy of bypass contribution deduction.

[0094] Traverse all directed edges in the directed causal graph, repeating the above-described process of obfuscation tracing, bypass decomposition, blocking calculation, and validity determination for each edge. During the traversal, it is crucial to handle loop structures within the directed causal graph. For causal graphs constructed from time-series remote sensing data, the graph structure is typically required to satisfy the directed acyclic graph constraint. If a loop is encountered while searching for common predecessor nodes in reverse, the dynamic causal graph must first be transformed into a directed acyclic form using time expansion before performing obfuscation tracing.

[0095] After completing the full graph traversal and pruning, all retained directed causal relationships and their corrected causal strengths are analyzed. A multidimensional causal relationship graph is constructed by summarizing the findings. Each directed edge in this graph has been verified against confounding factors, and its causal strength reflects the net causal effect after removing confounding influences. This allows for a more accurate characterization of the true driving relationship between remote sensing parameters, surface physical processes, and target land cover features. The multidimensional causal relationship graph also stores a list of confounding factors associated with each edge and their side-channel contribution ratios, providing complete prior information on the causal structure for subsequent counterfactual reasoning frameworks. It also provides an explainable path for tracing the root causes of prediction biases during the online correction phase.

[0096] In one optional implementation, based on the directed causal relationships in the multidimensional causal relationship map, a counterfactual reasoning framework is constructed with the target feature identification accuracy as the decision target variable. This framework simulates the propagation path of different band combination schemes along the directed causal relationships under different weight configurations, and calculates the expected causal effect of each band selection scheme on the target feature identification accuracy and its confidence interval, including:

[0097] In the multidimensional causal relationship map, the accuracy of target feature identification is used as the decision target variable. All directed causal relationship paths pointing to the decision target variable are traced backward, and the variables at the starting point of the directed causal relationship path are identified as the set of interventionable variables.

[0098] For the operable variables in the set of operable variables, the counterfactual values ​​of the operable variables under the band selection and adjustment state are set, and a change propagation operator of the operable variables from the current observation value to the counterfactual value is constructed. Along the directed causal relationship path from the operable variables to the decision target variable, the change propagation operator and the causal intensity of each segment on the directed causal relationship path are chained together to obtain the path causal effect.

[0099] For multiple parallel paths from the modifiable variable to the decision target variable, the path causal effects of the multiple parallel paths are aggregated to obtain the total causal effect of the counterfactual value on the decision target variable;

[0100] Based on the historical observation data of the interventionable variables, multiple sample sets are generated through perturbation resampling. For each sample set, path propagation and effect convergence are repeatedly performed to obtain the distribution characteristics of the total causal effect. The quantile intervals of the distribution characteristics are extracted as the expected causal effect and confidence interval of the intervention plan.

[0101] After the multidimensional causal relationship map is constructed, the target feature identification accuracy is used as the decision target variable. A reverse tracing is performed on all directed causal paths in the map pointing to this variable. Specifically, starting from the decision target variable node, the map is traversed layer by layer upwards in the reverse direction of the directed edges, recording each complete directed path from the map's starting node to the decision target variable. The starting node of the path is defined as the root node with an in-degree of zero, i.e., a variable that has no parent node in the multidimensional causal relationship map. These variables physically correspond to directly configurable or adjustable original input parameters in remote sensing observations, such as reflectance observations of each band, band selection flags, or band weight coefficients. All root nodes satisfying the above conditions are aggregated to form a set of operable variables. Each operable variable in this set has a clear physical operational meaning and its value can be changed by artificially setting band combination schemes, thereby triggering a chain propagation effect along the causal path and ultimately affecting the target feature identification accuracy.

[0102] For each operable variable in the set of operable variables, a counterfactual value is assigned under a specific band selection and adjustment scheme. The counterfactual value means that the variable is assumed to be forcibly intervened to a specified value, rather than its actual value under its natural observation state. Let the current observed value of a certain operable variable be... Its counterfactual value under a certain band combination scheme is Then the change in the value of this variable is .by As input, a change propagation operator is constructed, which describes the propagation of change when an actionable variable occurs. When a disturbance occurs, how does that disturbance propagate downstream along the directed causal path? For each directed edge on the path, its propagation contribution is weighted by the corresponding causal strength coefficient, and the causal strength coefficient between adjacent nodes is denoted as... ,in Indicates the upstream node, Indicates a downstream node. Along a path from... The path formed by directed edges of segments is subjected to a chained composition operation on the change propagation operator and the causal strength coefficients of each segment, that is... By multiplying the causality strength coefficients of each segment along the path, the path causality effect of the path is obtained. The calculation method is as follows ,in For path number The causality strength coefficient corresponding to the directed edge of the segment. This represents the total number of path segments. The physical meaning of this chain-like compound operation is: the net effect contribution generated when the disturbance of the influencing variable finally reaches the decision target variable after being propagated through multiple intermediate physical process nodes.

[0103] When multiple parallel directed causal paths exist between an interventionist variable and the target variable, the path causal effects of each path need to be converged to obtain the total causal effect of the intervention on the target variable. Suppose that multiple directed causal paths exist between a certain interventionist variable and the target variable... Parallel path, number The path causality of the path is Then the modifiable variable takes a counterfactual value. The total causal effect on the decision objective variable for When multiple modifiable variables in the set of modifiable variables are simultaneously assigned counterfactual values, the total causal effects of each modifiable variable can be linearly superimposed under the assumption of path independence, yielding the comprehensive expected causal effect of the entire band combination scheme on the decision target variable. For cases where there are shared intermediate nodes between paths, the effect contributions of shared nodes need to be deduplicated during the convergence phase to avoid overestimation due to repeated accumulation. Deduplication is achieved by marking already included nodes in the spectrum, retaining only one valid accumulation for multiple propagation contributions from the same node.

[0104] To quantify the uncertainty of the total causal effect and generate confidence intervals, multiple sample sets were generated based on historical observation data of the interventionable variables using a perturbation resampling method. Specifically, several samples were randomly drawn with replacement from the historical observation data to form a resampling sample set; this process was repeated to generate multiple resampling sample sets. Group sample set, in which The typical value range is 500 to 2000, which can be flexibly adjusted according to computational resource constraints. For each set of resampled samples, the causal strength coefficient of each directed edge is re-estimated. And the observed mean of the modifiable variables in this sample set is used as... The estimate maintains the counterfactual value. Without changing the process, the complete calculation procedure for path propagation and effect convergence is executed again to obtain the estimated total causal effect value for this sample set. ,in .go through After repeated sampling and calculation, the empirical distribution of the total causal effect is obtained. This distribution reflects the combined impact of causal strength estimation error, historical data fluctuations, and path structure uncertainty on the final effect estimation.

[0105] Quantile intervals are extracted from the above empirical distributions to serve as a quantitative representation of the expected causal effect and confidence interval of the intervention plan. One total causal effect estimate Sort in ascending order, and take the [number]th [item]. The nth value is used as the lower bound of the confidence interval, and the nth value is taken as the lower bound. The values ​​are used as the upper bound of the confidence interval, where and These are the preset lower and upper quantiles, respectively, with a typical configuration as follows: , The 95% confidence interval is used. The median or mean of the empirical distribution is used as a point estimate of the expected causal effect. For different band combinations, the above complete process is executed to generate the expected causal effect point estimate and corresponding confidence interval for each scheme, which is used for subsequent selection and ranking of the decision support scheme set. The width of the confidence interval itself also serves as a measure of scheme uncertainty. Schemes with narrower confidence intervals indicate robust effect estimates, while schemes with wider confidence intervals suggest insufficient historical data coverage or significant structural uncertainty in the causal path, requiring additional risk weighting during decision-making.

[0106] In practical remote sensing big data processing scenarios, the variables in the set of operable variables typically correspond to observation channels sensitive to specific land cover types, such as near-infrared bands, shortwave infrared bands, and red-edge bands. By systematically enumerating the counterfactual values ​​of these bands—for example, adjusting the weight of a certain band from its current configuration to a higher or lower level—the potential for different band combination strategies to improve the accuracy of target land cover identification can be pre-assessed without actually performing data collection. This counterfactual simulation method based on causal inference, compared to traditional exhaustive experiments or empirical parameter tuning methods, can more efficiently locate the optimal band configuration range under conditions of limited computational resources and provide a quantifiable effect benchmark for subsequent online correction.

[0107] In one optional implementation, cross-validating the expected causal effect with computational resource constraints to generate a decision support scheme set including remote sensing data processing schemes includes:

[0108] Obtain a set of computing resource constraints, which includes data transmission bandwidth constraints and recognition accuracy achievement constraints.

[0109] For each band selection scheme, the counterfactual values ​​of the modifiable variables in the band selection scheme are compared with the data transmission bandwidth constraints for feasibility verification, and the expected causal effect corresponding to the band selection scheme is compared with the recognition accuracy achievement constraints for conformity verification. When the counterfactual values ​​and the expected causal effect pass both the feasibility verification and the conformity verification, the band selection scheme is marked as a constraint-satisfied scheme.

[0110] For the constraint-satisfaction scheme, the expected causal effect confidence interval of the constraint-satisfaction scheme is extracted, the interval width of the expected causal effect confidence interval is calculated as the prediction uncertainty metric, and the distance between the lower bound of the expected causal effect confidence interval and the lower bound of the recognition accuracy achievement constraint is calculated as the target deviation metric.

[0111] Construct a risk assessment vector, and encapsulate the constraint-satisfying scheme, the expected causal effect, the confidence interval of the expected causal effect, and the risk assessment vector into a decision support scheme. Summarize all decision support schemes corresponding to the constraint-satisfying schemes to form a decision support scheme set.

[0112] like Figure 2 As shown, the method includes:

[0113] When obtaining the set of computing resource constraints, two types of key constraint information need to be collected from the actual deployment environment: data transmission bandwidth constraints and recognition accuracy achievement constraints. Data transmission bandwidth constraints describe the maximum data throughput available between the satellite ground station and the data processing center, typically expressed as the amount of data that can be transmitted per unit time (e.g., MB / s), and are directly related to the number of selected bands and the sampling resolution of each band. Recognition accuracy achievement constraints specify the minimum threshold range that the target feature recognition accuracy must meet, including two endpoints: a lower bound and an upper bound. These are used in subsequent conformity checks to determine whether the expected causal effect of a particular band selection scheme falls within an acceptable range. These two types of constraints together constitute the hard boundary conditions for screening remote sensing data processing schemes.

[0114] For each band selection scheme, the core logic of feasibility verification is to compare the actual amount of band data corresponding to the counterfactual values ​​of the modifiable variables in that scheme with the data transmission bandwidth constraints. Specifically, for a given band selection scheme, the counterfactual values ​​of its modifiable variables... This determines the combination of bands activated by the scheme, and thus allows for the estimation of the amount of raw data to be transmitted under that combination. Let the total number of bands activated by a certain scheme be... The amount of single-scene data for each band within the target observation area is (Unit: MB), then the total data transfer requirement of this scheme is .like Exceeding the maximum allowable transmission amount by bandwidth constraints If the feasibility check fails, the proposed solution is deemed infeasible and directly removed from the candidate set, without proceeding to the next stage. Only when... Only then did the plan pass the feasibility verification and enter the compliance verification stage.

[0115] The conformity check will determine the expected causal effect of the band selection scheme. Lower bound of accuracy that is constrained by recognition accuracy A comparison needs to be made. Since the expected causal effect reflects the net improvement in target feature identification accuracy of the scheme, it is necessary to compare it with the current baseline identification accuracy. Based on this, the expected causal effects are superimposed to obtain the prediction and identification accuracy. .like If the scheme meets the recognition accuracy requirement, it is considered to pass the compliance check. When a scheme for selecting a certain band passes both the feasibility check and the compliance check, it is marked as a constraint-satisfied scheme and included in the subsequent risk assessment process.

[0116] For all constraint-satisfied schemes, it is necessary to further extract the confidence intervals of the expected causal effects for each scheme, and then calculate two quantitative indicators based on these: Prediction Uncertainty Measures. Defined as the difference between the upper and lower bounds of the confidence interval, i.e. ,in and These are the causal effect estimates corresponding to the upper and lower quantiles of the confidence interval, respectively. The larger the value, the more statistically unstable the expected effect of the plan is, and the higher the risk of implementation. The smaller the value, the more concentrated the causal effect estimation of the scheme is under different resampling conditions, and the stronger the prediction reliability.

[0117] Target deviation metric Defined as the absolute distance between the lower bound of the confidence interval for the expected causal effect and the lower bound of the constraint on the accuracy of prediction and recognition. When the prediction accuracy corresponding to the lower bound of the confidence interval is exactly equal to the lower bound of the precision, Approaching zero means that the scheme barely meets the accuracy requirements even under the most pessimistic causal effect estimation, belonging to the high-risk boundary scheme; when When the prediction accuracy corresponding to a larger confidence interval lower bound is higher than the accuracy lower bound, it indicates that the scheme has sufficient accuracy safety margin and can still reliably achieve the goal in the worst case.

[0118] When constructing the risk assessment vector, the predicted uncertainty measure will be used. Deviation from target measurement Combined into a two-dimensional vector This serves as a quantitative representation of the risk of the scheme that satisfies the constraint. The physical meanings of the two components are complementary: It measures the degree of fluctuation in the internal causal effect estimates of the scheme, reflecting the stability of the forecast itself; Measuring the distance between the proposed solution and the accuracy target under the most pessimistic scenario reflects the size of the solution's margin relative to the constraints. In practical decision support scenarios, decision-makers can adjust the approach based on the tolerance of the two types of risks for the specific task. and Different priority weights are assigned to allow for personalized sorting and selection within the decision support solution set.

[0119] The band combination configuration information and expected causal effects of the constraint-satisfying scheme will be included. Confidence interval for expected causal effect Prediction uncertainty measure Target deviation metric and risk assessment vector Each decision support solution is encapsulated as a complete record. Each record also includes the data transmission requirements for that solution. Prediction and recognition accuracy and benchmark recognition accuracy This allows decision-makers to intuitively obtain information on resource consumption and expected effects when comparing different options. After iterating through all options that satisfy the constraints and completing the encapsulation operations described above, all decision support options are recorded and summarized to form the final decision support option set. This option set can be further processed according to… Descending order Ascending or The system outputs data using various sorting strategies, including descending order, for use in subsequent remote sensing data processing workflow scheduling and execution.

[0120] In practical applications, the observation window for satellite remote sensing missions is limited, and computational resource constraints dynamically change with factors such as orbital coverage time periods and ground station receiving capabilities. Therefore, the aforementioned cross-validation process is designed as a reentrant execution mode: when constraints are updated, feasibility and compliance checks can be retried for the candidate band selection scheme set without rebuilding the dynamic causal inference model. Only the screening and encapsulation process needs to be re-executed based on the existing expected causal effects and confidence interval results. This significantly reduces computational overhead while ensuring decision quality, meeting the dual requirements of real-time performance and adaptability in remote sensing big data processing scenarios.

[0121] In one optional implementation, collecting the actual execution effect of the remote sensing data processing scheme and using the deviation between the actual execution effect and the expected causal effect to perform online correction of the causal strength weight and confounding factor verification threshold in the dynamic causal inference model includes:

[0122] After the implementation of the band selection scheme, the actual observed value of the target feature identification accuracy is collected. The difference between the actual observed value and the baseline observed value before the scheme adjustment is calculated as the actual execution effect. The difference between the actual execution effect and the expected causal effect is used as the effect deviation. The effect deviation is decomposed in reverse along the directed causal relationship path to obtain the contribution component of each segment of the directed causal relationship. The corresponding causal strength weight is corrected based on the contribution component.

[0123] The potential explanatory power of the bypass propagation branch for the effect bias is calculated based on the modified causal strength weight, and the confusion factor verification threshold is adjusted according to the comparison result of the potential explanatory power and the confusion reassessment triggering condition.

[0124] After the band selection scheme is actually deployed and implemented, the actual observed values ​​of the target feature identification accuracy are collected. The difference between these actual observed values ​​and the baseline observed values ​​before the scheme adjustment is the actual implementation effect of the scheme. This actual implementation effect is then subtracted from the expected causal effect calculated using the counterfactual reasoning framework to obtain the effect bias. The effect bias reflects the systematic deviation between the model prediction and the real-world response and is the core signal driving the online correction of the dynamic causal inference model.

[0125] Let the difference between the actual observed value and the reference observed value be... The expected causal effect is Then the effect bias is .when When the value is positive, it indicates that the model underestimates the actual effect of the solution; when... A negative value indicates that the model overestimates the actual effect of the proposed solution. In either case, it is necessary to follow the directed causal path to... Perform reverse decomposition to pinpoint the contribution of each segment's directed causal relationship to the total bias.

[0126] The specific process of reverse decomposition is as follows: Following the directed causal path from the modifiable variable to the decision target variable, trace back segment by segment from the target node to the source node. For each segment of the directed causal relationship on the path, based on the proportion of the causal strength coefficient of that segment in the total path propagation gain, the total effect bias is calculated. Distribute proportionally to this segment. Suppose a path contains a total of... The first segment has a directional causal relationship. The causal strength coefficient of the segment is The contribution weight of this segment to the total propagation gain is then determined. Satisfying the constraint that the sum of the contribution weights of all segments is normalized, i.e. ,in The causality strength coefficient of this segment is determined by the ratio of the causality strength coefficients of this segment to the sum of the causality strength coefficients of all segments along the path, and is expressed as follows: . No. The contribution component assigned to the directed causal relationship is .

[0127] When multiple parallel paths exist, the effect bias components of each path must first be allocated according to the path weights, and then decomposed segment by segment within each path. The path weight is determined by the proportion of the path's causal effect in the total effect of all parallel paths, thereby ensuring the consistency and traceability of the decomposition results in the multi-path scenario.

[0128] Obtain the contribution components of each segment of the directed causal relationship. Then, the corresponding causal strength weights are adjusted. Let the first... The current causal strength coefficient of the directed causal relationship is Corrected causality strength coefficient By contributing components Transform into This is achieved through incremental changes, specifically... ,in The learning rate hyperparameter, This represents the actual change at the source node of the directed causal relationship. The physical meaning of this correction method is: if the contribution of a certain causal relationship to the result in actual execution is higher than expected, then its causal strength coefficient is increased; if it is lower than expected, then its causal strength coefficient is decreased, thereby making the model more accurate in depicting the true causal propagation law.

[0129] To prevent model oscillations caused by excessively large single corrections, upper and lower bound constraints are set for each correction amount, i.e. ,in This is the preset maximum single correction range. When the calculated correction exceeds this range, it is truncated to the boundary value. Furthermore, after accumulating multiple actual execution data, a sliding window averaging method can be used to smooth the contribution components of each segment, reducing the interference of single observation noise on the causal strength weight correction results.

[0130] After adjusting for causal strength weights, the potential explanatory power of bypass propagation branches on effect bias is further evaluated to determine whether the confounding factor check threshold needs adjustment. A bypass propagation branch is a propagation path that bypasses the causal variable and directly reaches the effect variable; its existence indicates that the observed causal effect includes a contribution from confounding factors. Potential explanatory power is calculated by using the proportion of effect bias explained by the bypass propagation branch under the adjusted causal strength weights as a metric. Let the explanatory power of the bypass propagation branch on effect bias be... Then the potential explanatory power . The product of the corrected causal strength coefficients of each segment on the bypass path and the sum of the products are obtained by multiplying the product of the actual changes at the source node. This reflects the degree of fit of bypass propagation to the bias of the total effect under the current model parameters.

[0131] Potential Explanatory Power Compared to the obfuscation re-evaluation trigger condition, the obfuscation re-evaluation trigger condition is determined by a preset trigger threshold. Confirmed. When This indicates that the bypass propagation branch has a strong explanatory power for effect bias, and the current confusion factor verification threshold is too lenient, resulting in some confusion paths not being effectively identified and eliminated. In this case, the confusion factor verification threshold should be tightened, that is, the confusion identification threshold should be increased. By proportionally reducing the threshold, the identification of subsequent confounding factors becomes more rigorous, thereby eliminating more potential false causal paths. This indicates that the explanatory power of the bypass propagation branch for effect bias is limited. The current threshold setting for checking confounding factors is reasonable and does not need to be adjusted. Alternatively, it may be appropriately relaxed to avoid excessive pruning that could lead to the loss of the true causal path.

[0132] Adjustment amount of the confounding factor test threshold and its potential explanatory power The degree of exceedance is directly proportional; let the current confusion recognition threshold be... The adjusted threshold is ,but ,in To adjust the step size factor, this value is in Effective immediately. To ensure the threshold does not fall below the minimum allowable value. The adjustment result is truncated, that is... This prevents the threshold from being too low, which would cause all paths to be judged as confusing and thus degrade the causal graph.

[0133] The aforementioned causal strength weight correction and confounding factor verification threshold adjustment are both performed online, without requiring a complete re-execution of the structure learning process. Only local path parameters affected by effect bias are updated, thus ensuring the computational efficiency and adaptability of the dynamic causal inference model during continuous operation. After each round of online correction, the corrected causal strength weights and confounding factor verification thresholds are written back to the multidimensional causal association graph for use in subsequent counterfactual inference and decision support scheme generation, forming a closed-loop continuous optimization mechanism.

[0134] A second aspect of this invention provides an artificial intelligence-based satellite remote sensing big data processing and analysis system, comprising:

[0135] The data preparation unit is used to acquire a large dataset of data to be analyzed from remote sensing satellites, and to preprocess the dataset to form a set of candidate variables.

[0136] The causal inference unit is used to construct a dynamic causal inference model. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, it identifies the directed causal relationship between remote sensing observation parameters, surface physical processes and target land cover characteristics and quantifies the causal strength. At the same time, it performs reverse tracing of confounding factors and verification of causal validity on the directed causal relationship, prunes false causal paths, and obtains a multidimensional causal association map.

[0137] The causal effect unit is used to construct a counterfactual reasoning framework based on the directed causal relationship in the multidimensional causal relationship map, with the target feature identification accuracy as the decision target variable, to simulate the propagation path of different band combination schemes along the directed causal relationship under different weight configurations, and to calculate the expected causal effect of each band selection scheme on the target feature identification accuracy and its confidence interval.

[0138] The online correction unit is used to cross-validate the expected causal effect with the computational resource constraints, generate a decision support scheme set containing remote sensing data processing schemes, collect the actual execution effect of the remote sensing data processing schemes, and use the deviation between the actual execution effect and the expected causal effect to correct the causal strength weight and confusion factor verification threshold in the dynamic causal inference model online.

[0139] A third aspect of the present invention provides an electronic device, comprising:

[0140] processor;

[0141] Memory used to store processor-executable instructions;

[0142] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0143] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0144] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A satellite remote sensing big data processing and analysis method based on artificial intelligence, characterized in that, include: A set of big data to be analyzed is obtained from remote sensing satellites, and the set of big data to be analyzed is preprocessed to form a set of candidate variables; A dynamic causal inference model is constructed. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, the directed causal relationship between remote sensing observation parameters, surface physical processes and target land cover characteristics is identified and the causal strength is quantified. At the same time, the directed causal relationship is traced back to the confounding factors and the causal validity is verified. False causal paths are pruned to obtain a multidimensional causal association map. Based on the directed causal relationships in the multidimensional causal relationship map, a counterfactual reasoning framework is constructed with the target feature recognition accuracy as the decision target variable. The propagation path of different band combination schemes along the directed causal relationships under different weight configurations is simulated, and the expected causal effect of each band selection scheme on the target feature recognition accuracy and its confidence interval are calculated. The expected causal effect is cross-validated with computational resource constraints to generate a decision support scheme set containing remote sensing data processing schemes. The actual execution effect of the remote sensing data processing scheme is collected, and the deviation between the actual execution effect and the expected causal effect is used to correct the causal strength weight and confusion factor verification threshold in the dynamic causal inference model online.

2. The method according to claim 1, characterized in that, A dynamic causal inference model is constructed. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, the directed causal relationships between remote sensing observation parameters, surface physical processes, and target land cover characteristics are identified, and the causal strength is quantified. For a pair of variables in the candidate variable set, the conditional mutual information of the pair of variables under a given set of conditional variables is calculated within a sliding time window. When the conditional mutual information converges to a stable value as the set of conditional variables expands, the set of conditional variables corresponding to the stable value is determined as the minimum separating set. Based on the minimum separating set, it is determined whether there is a direct causal path between the pair of variables. Traverse all variable pairs in the candidate variable set, mark variable pairs with direct causal paths as candidate causal edges, and construct an initial undirected causal skeleton graph; For each candidate causal edge in the initial undirected causal skeleton graph, the causal direction is oriented using time sequence constraints and non-cyclic constraints. By evaluating the variance of the prediction residual of the source variable to the target variable under different causal direction assumptions, the direction that minimizes the variance of the prediction residual is selected as the true causal direction, thus forming a directed causal graph. For each directed edge in the directed causal graph, the contribution coefficient of the change in the source variable to the change in the target variable is calculated, and the contribution coefficient is used as the causal strength corresponding to the directed edge.

3. The method according to claim 2, characterized in that, The causal direction is oriented using chronological order constraints and non-cyclic constraints. By evaluating the variance of the prediction residuals of the source variable on the target variable under different causal direction assumptions, the direction that minimizes the variance of the prediction residuals is selected as the true causal direction, forming a directed causal graph, including: For each candidate causal edge in the initial undirected causal skeleton graph, the temporal priority is determined according to the difference in satellite overpass imaging timestamps of the connected variable pairs. The variables with the higher temporal priority are taken as causal candidates, and the variables with the lower temporal lag are taken as effect candidates. A positive causal hypothesis is established, and a negative hypothesis corresponding to the positive causal hypothesis is constructed as a control. For the positive causal hypothesis and the negative hypothesis, it is examined whether their combination with the oriented edges in the initial undirected causal skeleton graph forms a directed loop. Hypotheses that violate the non-cyclic constraint are marked as invalid hypotheses, while valid hypotheses that satisfy the constraint are retained. Using the historical observations of the causal variables in the effective hypothesis, the current observations of the effect variables are predicted. The residual variance between the actual observations and the predicted values ​​of the effect variables is calculated as the fitting error. The fitting errors of the positive causal hypothesis and the negative hypothesis are compared, and the hypothesis direction with the smallest fitting error is selected as the orientation result of the candidate causal edge. The orientation results are applied to the initial undirected causal skeleton graph, the set of oriented edges is updated, and the constraint check and error comparison are repeated until all candidate causal edges are oriented, forming a directed causal graph.

4. The method according to claim 1, characterized in that, The directed causal relationship is subjected to reverse tracing of confounding factors and verification of causal validity, and false causal paths are pruned to obtain a multidimensional causal relationship graph, including: For each directed causal relationship in the directed causal graph, a common predecessor node that can reach the cause variable and effect variable of the directed causal relationship is searched in reverse in the directed causal graph. A causal propagation tree starting from the common predecessor node is constructed. The direct propagation branch through the cause variable and the bypass propagation branch that bypasses the cause variable in the causal propagation tree are decomposed. When the cumulative contribution of the bypass propagation branch to the effect variable accounts for more than the proportion of the observed causal strength, the common predecessor node is marked as a confounding factor. Block the bypass propagation branch, calculate the net causal contribution of the cause variable to the effect variable, and when the decrease in the net causal contribution relative to the original observed causal strength exceeds the validity threshold, determine that the directed causal relationship is a false causal path and remove it; otherwise, retain it and update its causal strength with the net causal contribution. The directed causal graph is traversed to complete the verification and pruning, and the retained directed causal relationships and their corrected causal strengths are constructed into a multidimensional causal association graph.

5. The method according to claim 1, characterized in that, Based on the directed causal relationships in the multidimensional causal relationship map, a counterfactual reasoning framework is constructed with the target feature identification accuracy as the decision objective variable. This framework simulates the propagation path of different band combination schemes along the directed causal relationships under different weight configurations, calculating the expected causal effect of each band selection scheme on the target feature identification accuracy and its confidence interval, including: In the multidimensional causal relationship map, the accuracy of target feature identification is used as the decision target variable. All directed causal relationship paths pointing to the decision target variable are traced backward, and the variables at the starting point of the directed causal relationship path are identified as the set of interventionable variables. For the operable variables in the set of operable variables, the counterfactual values ​​of the operable variables under the band selection and adjustment state are set, and a change propagation operator of the operable variables from the current observation value to the counterfactual value is constructed. Along the directed causal relationship path from the operable variables to the decision target variable, the change propagation operator and the causal intensity of each segment on the directed causal relationship path are chained together to obtain the path causal effect. For multiple parallel paths from the modifiable variable to the decision target variable, the path causal effects of the multiple parallel paths are aggregated to obtain the total causal effect of the counterfactual value on the decision target variable; Based on the historical observation data of the interventionable variables, multiple sample sets are generated through perturbation resampling. For each sample set, path propagation and effect convergence are repeatedly performed to obtain the distribution characteristics of the total causal effect. The quantile intervals of the distribution characteristics are extracted as the expected causal effect and confidence interval of the intervention plan.

6. The method according to claim 1, characterized in that, Cross-validating the expected causal effects with computational resource constraints generates a set of decision support solutions that includes remote sensing data processing schemes, including: Obtain a set of computing resource constraints, which includes data transmission bandwidth constraints and recognition accuracy achievement constraints. For each band selection scheme, the counterfactual values ​​of the modifiable variables in the band selection scheme are compared with the data transmission bandwidth constraints for feasibility verification, and the expected causal effect corresponding to the band selection scheme is compared with the recognition accuracy achievement constraints for conformity verification. When the counterfactual values ​​and the expected causal effect pass both the feasibility verification and the conformity verification, the band selection scheme is marked as a constraint-satisfied scheme. For the constraint-satisfaction scheme, the expected causal effect confidence interval of the constraint-satisfaction scheme is extracted, the interval width of the expected causal effect confidence interval is calculated as the prediction uncertainty metric, and the distance between the lower bound of the expected causal effect confidence interval and the lower bound of the recognition accuracy achievement constraint is calculated as the target deviation metric. Construct a risk assessment vector, and encapsulate the constraint-satisfying scheme, the expected causal effect, the confidence interval of the expected causal effect, and the risk assessment vector into a decision support scheme. Summarize all decision support schemes corresponding to the constraint-satisfying schemes to form a decision support scheme set.

7. The method according to claim 1, characterized in that, The process of collecting the actual execution results of the remote sensing data processing scheme and using the deviation between the actual execution results and the expected causal effect to perform online correction of the causal strength weight and confounding factor verification threshold in the dynamic causal inference model includes: After the implementation of the band selection scheme, the actual observed value of the target feature identification accuracy is collected. The difference between the actual observed value and the baseline observed value before the scheme adjustment is calculated as the actual execution effect. The difference between the actual execution effect and the expected causal effect is used as the effect deviation. The effect deviation is decomposed in reverse along the directed causal relationship path to obtain the contribution component of each segment of the directed causal relationship. The corresponding causal strength weight is corrected based on the contribution component. The potential explanatory power of the bypass propagation branch for the effect bias is calculated based on the modified causal strength weight, and the confusion factor verification threshold is adjusted according to the comparison result of the potential explanatory power and the confusion reassessment triggering condition.

8. An artificial intelligence-based satellite remote sensing big data processing and analysis system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data preparation unit is used to acquire a large dataset of data to be analyzed from remote sensing satellites, and to preprocess the dataset to form a set of candidate variables. The causal inference unit is used to construct a dynamic causal inference model. By performing temporal conditional independence tests and structure learning on the variables in the candidate variable set, it identifies the directed causal relationship between remote sensing observation parameters, surface physical processes and target land cover characteristics and quantifies the causal strength. At the same time, it performs reverse tracing of confounding factors and verification of causal validity on the directed causal relationship, prunes false causal paths, and obtains a multidimensional causal association map. The causal effect unit is used to construct a counterfactual reasoning framework based on the directed causal relationship in the multidimensional causal relationship map, with the target feature identification accuracy as the decision target variable, to simulate the propagation path of different band combination schemes along the directed causal relationship under different weight configurations, and to calculate the expected causal effect of each band selection scheme on the target feature identification accuracy and its confidence interval. The online correction unit is used to cross-validate the expected causal effect with the computational resource constraints, generate a decision support scheme set containing remote sensing data processing schemes, collect the actual execution effect of the remote sensing data processing schemes, and use the deviation between the actual execution effect and the expected causal effect to correct the causal strength weight and confusion factor verification threshold in the dynamic causal inference model online.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images