A method for corroborating anomalies in carbon sink verification and monitoring data
By combining physical process model pre-simulation and spatial collaborative verification, the method of accurately identifying and correcting anomaly types in carbon sink verification and monitoring data has been achieved, solving the problem of insufficient anomaly type differentiation in existing technologies and improving the accuracy and responsiveness of detection.
Patent Information
- Application Number
- CN202611318988.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-28
AI Technical Summary
Existing carbon sequestration verification methods cannot effectively distinguish between different types of anomalies, resulting in insufficient accuracy and reliability of anomaly detection. They are prone to misjudging genuine anomalies as random noise or misjudging local pseudo-anomalies as genuine anomalies.
By combining hierarchical progressive anomaly identification with spatial collaborative verification, the anomaly types are determined by physical process model pre-simulation and spatial collaborative anomaly index (SCAI), including collaborative anomalies, local pseudo-anomalies and extreme anomalies, and targeted correction strategies are implemented for each.
It enables accurate classification and correction of carbon sink verification and monitoring data, improves the accuracy and reliability of anomaly detection, avoids misjudgment and inappropriate correction due to type confusion, has the ability to detect problems in advance, and improves the timeliness of response and the accuracy of decision-making.
Smart Images

Figure CN122818133B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of carbon sink data anomaly verification, specifically involving a method for verifying anomalies in carbon sink verification and monitoring data. Background Technology
[0002] Carbon sinks refer to the processes and capacity of terrestrial ecosystems such as forests and grasslands to absorb and store carbon dioxide from the atmosphere. Accurate monitoring and accounting of terrestrial ecosystem carbon sinks has become an important component of the national carbon accounting system. The core work of carbon sink verification involves periodically monitoring fixed sample plots in ecosystems such as forests and grasslands. This involves regularly collecting observational data on carbon storage, vegetation growth, and soil carbon content at pre-determined fixed sample plot locations, and comparing the temporal changes of the sample plot coordinates with the observed values to determine the consistency and reliability of the data.
[0003] With the diversification of monitoring methods such as IoT sensors, satellite remote sensing, and manual ground surveys, and the continuous improvement of data quality requirements in the carbon trading market, the carbon sink verification process needs to identify abnormal data. Early carbon sink verification mainly relied on manual ground surveys, obtaining carbon storage change data through periodic retesting of fixed sample sites. In this stage, data anomalies were primarily identified through the experience of verification personnel, lacking a systematic anomaly detection method.
[0004] With the development of spatial statistical methods, geostatistics has been introduced into the field of carbon sink data verification. Methods such as variograms and kriging interpolation are used to analyze the spatial variability and spatial autocorrelation of carbon sink indicators such as forest carbon density. The core idea of these methods is that spatially adjacent plots should have similar trends in their carbon sink observations; if the observations of a plot deviate significantly from the overall trend of its neighboring plots, then that plot may be an anomaly.
[0005] Meanwhile, data lineage management technology has also been introduced into the field of carbon sequestration data verification. Chinese patent CN108763304A discloses a method for verifying electricity ledger data based on lineage relationships. This method verifies the quality of ledger data during the transfer process by establishing an effective / ineffective status field for lineage nodes. The core logic of this method is that each step in the data transfer process from collection to processing constitutes a lineage path. By tracking the changes in the data's status at each stage, it is possible to determine whether any anomalies have occurred during the transfer process.
[0006] Despite the progress made in the aforementioned technologies, existing methods still have the following shortcomings. Current methods for detecting anomalies in carbon sequestration monitoring data typically label all observations that deviate from expected values as anomalies. Specifically, deviations in sample site monitoring data from expected values may stem from three entirely different causes: First, isolated anomalies (local pseudo-anomalies) may occur due to measurement errors, sensor malfunctions, or environmental interference. Second, physical processes such as landslides or pest infestations may occur in the area where the sample site is located, leading to structural changes in the carbon sequestration characteristics of the entire region, with the anomaly appearing spatially in conjunction with neighboring areas (coordinated anomalies). Third, deviations may occur in the data collection or transfer process due to human operational errors such as uncalibrated equipment or mislabeling (incorrect collection path). The corresponding processing strategies for these three causes are drastically different: local pseudo-anomalies should be corrected through spatial interpolation; coordinated anomalies should trigger regional verification and topology reconstruction; and incorrect collection paths should be traced back to equipment status and operational records.
[0007] Furthermore, traditional methods perform independent single-point verification on each sample plot, comparing the measured values of the plot with the baseline predicted values. If the deviation exceeds a threshold, it is judged as an anomaly. This method ignores the important information contained in the spatial autocorrelation structure. If the observed value of a sample plot deviates from the expectation, but its neighboring sample plots also deviate in the same direction and to the same extent, then the anomaly is likely a regional real change rather than a random error. Conversely, if a sample plot deviates from the expectation in isolation while its neighboring sample plots are normal, then the anomaly is more likely a measurement error or local disturbance. Existing techniques are prone to misclassifying true anomalies as random noise or misclassifying local spurious anomalies as true anomalies. Summary of the Invention
[0008] This invention provides a method for verifying anomalies in carbon sink verification and monitoring data. By combining hierarchical progressive anomaly identification with spatial collaborative verification, it achieves accurate classification of anomaly types in sample plots to enable the implementation of targeted correction strategies. This solves the problem that existing technologies rely solely on single-point deviation thresholds to identify anomalies and lack a neighborhood collaborative verification mechanism, making it difficult to effectively distinguish anomaly types with different causes and thus difficult to implement accurate corrections.
[0009] The technical solution adopted in this invention is as follows: A method for corroborating anomalies in carbon sink verification and monitoring data includes: Based on the acquired soil mechanical parameters, the physical process model is deduced in advance, and the safety factor and landslide probability of each grid cell are output. When the output landslide probability reaches the warning threshold, a physical warning event is generated. The method also includes, When the sample plot monitoring data arrives, the single-point deviation value of the sample plot is calculated based on the sample plot monitoring data. If the single-point deviation value does not exceed the corresponding first threshold, the data is determined to be normal; if any of the single-point deviation values exceeds the corresponding first threshold, it is marked as abnormal; wherein, the sample plot monitoring data includes at least coordinates and carbon reserves, and the single-point deviation value includes coordinate deviation value and / or carbon reserve fluctuation rate; In response to the sample plot being marked as anomalous, the spatial neighborhood set of the sample plot is queried. Based on the monitoring data and the baseline prediction data of the physical process model, the single-point residual of the anomalous sample plot and each sample plot in the spatial neighborhood set is obtained, and the spatial co-anomaly index of the anomalous sample plot is determined. Based on the value of the spatial coordinated anomaly index, the anomaly type of the sample plot is determined for correction purposes. The anomaly type includes at least coordinated anomalies, local pseudo-anomalies, and extreme anomalies. If the absolute value of the spatial coordination anomaly index is less than or equal to the second threshold, it is determined to be a coordination anomaly. If the absolute value of the spatial coordination anomaly index is greater than the second threshold and less than or equal to the third threshold, it is determined to be a local pseudo-anomaly. If the absolute value of the spatial coordination anomaly index is greater than the third threshold, it is determined to be an extreme anomaly; wherein the third threshold is greater than the second threshold.
[0010] Specifically, the single-point deviation value is: The single-point deviation value includes a coordinate deviation value, obtained by the difference between the current coordinate and the historical coordinate; and / or, The single-point deviation value includes the carbon storage volatility, which is obtained by the deviation rate between the current carbon storage and the baseline prediction value of the physical process model.
[0011] If the single-point deviation value does not exceed the corresponding first threshold, the data is determined to be normal; if any of the single-point deviation values exceeds the corresponding first threshold, it is marked as abnormal. Specifically: The single-point deviation value includes the coordinate deviation value. If the coordinate deviation value Δd ≤ 5m, the data is considered normal; if the coordinate deviation value Δd > 5m, it is marked as abnormal; or... The single-point deviation value includes the carbon storage volatility. If the carbon storage volatility Δc ≤ 10%, the data is considered normal; if the carbon storage volatility Δc > 10%, it is marked as abnormal; or... The single-point deviation value includes the coordinate deviation value and the carbon storage volatility. If the coordinate deviation value Δd ≤ 5m and the carbon storage volatility Δc ≤ 10%, the data is considered normal. If the coordinate deviation value Δd > 5m or the carbon storage volatility Δc > 10%, it is marked as abnormal.
[0012] The spatial coordination anomaly index is determined, specifically including: Calculate the single-point residuals of the anomaly plot i. ,in The sample monitoring data for sample plot i. Baseline prediction data for the physical process model; Computational spatial expected residual ,in Let i be the set of spatial neighborhoods of the sample plot i. Let i be the single-point residual of any plot j in the spatial neighborhood set of plot i. Let j be the kriging weight of plot j relative to plot i; Calculate the absolute value of the disconnection ; Divide the absolute value of the dissociation by the standard error of the Kriging theory. The spatial coordination anomaly index was obtained. The standard error of the Kriging theory mentioned therein It is determined by the spatial configuration of the spatial neighborhood set.
[0013] After determining it to be a collaborative anomaly, attribution analysis is also included: Based on the spatial coordination anomaly index of the sample plot ,pass Obtain the spatial chain likelihood ; The independence score for the collection of the sample monitoring data from the sample plots was based on the data collected from the sample plots. ,pass Seemingly obtaining a bloodline chain ,in The neutral threshold is set to 0.5, and k is the steepness coefficient, set to 6.0. The physical deviation between the plot monitoring data and the predicted values of the physical process model is used as a basis. ,pass Obtain the physical chain likelihood ,in The standard deviation of the predictions for the physical process model; Based on the landslide probability, prior probabilities are set for the spatial chain likelihood, the lineage chain likelihood, and the physical chain likelihood, respectively. The posterior probabilities are then adjusted using Bayes' theorem to obtain the spatial chain likelihood, the lineage chain likelihood, and the physical chain likelihood. If the maximum posterior probability of multiple posterior probabilities is less than the fourth threshold, then output an uncertain state and end the process; If the maximum a posteriori probability is greater than or equal to the fourth threshold, then the hypothesis corresponding to the maximum a posteriori probability is adopted as the attribution result; Among them, the attribution result corresponding to the spatial chain likelihood is a local pseudo-anomaly; the attribution result corresponding to the physical chain likelihood is a physically driven true anomaly; and the attribution result corresponding to the lineage chain likelihood is an error in the acquisition path.
[0014] Based on the landslide probability, prior probabilities are set for the spatial chain likelihood, the kinship chain likelihood, and the physical chain likelihood, respectively, as follows: If the landslide probability reaches the warning threshold, the prior probability of the spatial chain likelihood is P(H1)=0.2, the prior probability of the physical chain likelihood is P(H2)=0.7, and the prior probability of the bloodline chain likelihood is P(H3)=0.1. Otherwise, based on the prior probability P(H1)=0.2 of the spatial chain likelihood, the prior probability P(H2)=0.7 of the physical chain likelihood, and the prior probability P(H3)=0.1 of the lineage chain likelihood, the prior probability of the physical chain likelihood is reduced, while the prior probabilities of the spatial chain likelihood and the lineage chain likelihood are increased.
[0015] If the attribution result is a physically driven true anomaly, then perform topology reconstruction based on the physical feasible region constraints: Read the physical boundary output by the physical process model, generate an allowable transition matrix based on the physical boundary, and migrate the sample plots located within the physical boundary to a new sub-network; After the reconstruction is completed, a freeze period is set, during which no new reconstruction requests will be accepted.
[0016] The topology reconstruction of the physically feasible region constraints specifically includes: Calculate the ratio η of the geometric intersection area of the polygon of the sample plot and the physical boundary to the total area of the sample plot itself. If η ≥ 50% and the safety factor of the center point of the sample plot is less than the first safety threshold, then migration is allowed. Otherwise, migration is prohibited.
[0017] If the attribution result is a local spurious anomaly or an incorrect acquisition path, then perform lineage inversion and correction: A causal graph is pre-constructed, which contains multiple nodes and causal edges connecting the nodes; the causal edges include four types: physical edges, spatial edges, lineage edges, and hybrid edges, and each causal edge carries an uncertainty weight; Tracing back from the node corresponding to the abnormal sample plot to the original acquisition node along the bloodline edge in the causal graph, the aggregation calculation parameters are adjusted or the acquisition device status is marked.
[0018] After the attribution result processing is completed, it also includes: A causal graph is pre-constructed, which contains multiple nodes and causal edges connecting the nodes; the causal edges include four types: physical edges, spatial edges, lineage edges, and hybrid edges, and each causal edge carries an uncertainty weight; The uncertainty value of the processing operation is calculated based on the uncertainty weight. If the uncertainty value exceeds the fifth threshold, automatic correction is rejected, the data flow status is marked as pending manual review, and the process is terminated. If the uncertainty value does not exceed the fifth threshold, then automatic correction is performed.
[0019] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are as follows: This invention establishes an anomaly discrimination architecture, classifying anomalies into three types: cooperative anomalies, local pseudo-anomalies, and extreme anomalies. When the deviation of a single point in a sample plot exceeds a first threshold, this invention does not directly determine the anomaly type. Instead, it further queries the spatial neighborhood set of the sample plot and calculates its Spatial Cooperative Anomaly Index (SCAI). This index quantifies the consistency between sample plot anomalies and spatial trends by comparing the degree of discrepancy between the residual of the target sample plot and the Kriging-weighted expected value of the residual of its neighboring sample plots. Based on this, the absolute value of SCAI is compared with a second and a third threshold: if |SCAI| ≤ the second threshold, it indicates that the plot anomaly is highly consistent with the neighborhood trend, and is judged as a coordinated anomaly. This type of anomaly reflects the real regional change and should trigger a regional verification; if the second threshold < |SCAI| ≤ the third threshold, it indicates that the plot anomaly is somewhat disconnected from the neighborhood trend but not completely independent, and is judged as a local pseudo-anomaly. This type of anomaly may originate from measurement error or local environmental interference and should be handled through spatial smoothing correction; if |SCAI| > the third threshold, it indicates that the plot anomaly is completely independent of the spatial structure and is judged as an extreme anomaly. This type of anomaly may originate from equipment failure or human error and should trigger an emergency on-site verification. Through the above three-level classification, this invention provides an accurate decision-making basis for subsequently implementing different correction strategies, avoiding misjudgments and inappropriate correction strategies caused by confusion of anomaly types.
[0020] By querying the spatial neighborhood set of the sample plots, calculating the spatial expected residual based on Kriging weights, and then constructing the Spatial Collaborative Anomaly Index (SCAI), neighborhood collaborative verification is achieved. This makes anomaly judgment no longer rely on single-point deviation values in isolation, but places the sample plots in the collaborative framework of their spatial neighborhoods for comprehensive judgment. This effectively avoids the problem of misjudging true anomalies as random noise or misjudging local pseudo-anomalies as true anomalies, and significantly improves the accuracy and reliability of anomaly judgment.
[0021] This invention, based on data-driven anomaly detection, introduces a physical process model deduction mechanism. By pre-deducing the physical process model based on acquired soil mechanical parameters, it outputs the safety factor and landslide probability for each grid cell, and generates a physical warning event when the landslide probability reaches a warning threshold. This mechanism enables the system to obtain warning information before physical processes (such as landslides) occur, providing important physical prior references for subsequent data anomaly verification. When sample plot monitoring data arrives, this invention combines the physical warning event with single-point deviation verification and spatial collaborative verification, forming cross-validation. This allows anomaly verification to no longer passively rely on post-data arrival verification, but rather possesses pre-emptive perception capabilities, further improving the system's response timeliness and decision-making accuracy. Attached Figure Description
[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the method for verifying anomalies in carbon sink verification and monitoring data according to one embodiment of the present invention. Detailed Implementation
[0023] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.
[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0025] like Figure 1 As shown, a method for corroborating anomalies in carbon sink verification and monitoring data includes: S1: Based on the acquired soil mechanical parameters, perform a physical process model deduction in advance, output the safety factor and landslide probability of each grid cell, and generate a physical early warning event when the output landslide probability reaches the early warning threshold.
[0026] The main purpose of step S1 is to establish a pre-emptive warning mechanism for physical processes. In carbon sequestration scenarios, physical processes such as landslides and pests can cause substantial changes in the spatial structure of sample plots, thereby affecting the validity of monitoring data. By performing a physical process model simulation before the data arrives, step S1 enables the system to obtain early warning information before the physical process occurs, providing a physical prior reference for subsequent data anomaly verification, and elevating anomaly detection from reactive post-event verification to proactive pre-event warning.
[0027] Specifically, the system first acquires the soil mechanical parameters of each grid cell in the target area, including but not limited to physical properties such as soil internal friction angle, cohesion, slope, and pore water pressure. These parameters are the core input variables driving the landslide stability mechanical model.
[0028] The system continuously performs physical process model simulations at preset time steps. In a preferred embodiment, this time step is on the order of one hour, meaning the system performs a model calculation every hour. In each simulation, the physical process model calculates the safety factor and landslide probability for each grid cell based on the soil mechanical parameters at the current moment. The safety factor reflects the stability of the grid cell under current conditions; a safety factor less than 1 indicates a risk of instability in the area. The landslide probability quantifies the likelihood of a landslide occurring in the grid cell within a future period.
[0029] Based on the model output, a warning threshold trigger is further configured. When the landslide probability of a certain grid cell or continuous area reaches the preset warning threshold, the system automatically generates a physical warning event. An optimal value for this warning threshold is 80%, meaning that a warning is triggered when the landslide probability exceeds 80%. The physical warning event includes information such as the coordinates of the affected area, the prediction time window, and the confidence level, and is injected into the global causal event queue for other modules to subscribe to.
[0030] Through step S1, this invention injects the ability to provide early warnings of physical processes into the data verification system. Through continuous deduction of the physical process model, the system obtains the evolution trend and early warning information of the physical process before the monitoring data arrives, thus possessing predictive tracing capabilities.
[0031] When monitoring data arrives, the system already possesses information such as whether the area is in a physical warning state and the level of confidence in the warning. This prior physical information serves as an important reference for subsequent anomaly determination, ensuring that anomaly evidence no longer relies solely on data-driven mechanisms, but rather achieves an organic synergy between physical and data-driven approaches. Moreover, the generation of physical warning events does not depend on the arrival of monitoring data; therefore, even during monitoring gaps, the system can continuously track the changing trends of physical processes, ensuring the timeliness and continuity of warnings.
[0032] The method also includes, S2: When the sample plot monitoring data arrives, calculate the single-point deviation value of the sample plot based on the sample plot monitoring data. If the single-point deviation value does not exceed the corresponding first threshold, the data is determined to be normal; if any of the single-point deviation values exceeds the corresponding first threshold, it is marked as abnormal; wherein, the sample plot monitoring data includes at least coordinates and carbon reserves, and the single-point deviation value includes coordinate deviation value and / or carbon reserve fluctuation rate.
[0033] The main purpose of step S2 is to quickly screen the arriving sample plot monitoring data. By calculating the single-point deviation value and comparing it with the threshold, a large amount of normal data is efficiently separated from a small amount of suspected abnormal data, providing a precise trigger entry point for subsequent in-depth analysis. The core design concept of this step is to quickly release normal data and accurately mark abnormal data, avoiding the need to perform complex spatial co-validation on all data, thereby significantly reducing the computational burden of the system.
[0034] Step S2 is initiated upon arrival of the plot monitoring data. The plot monitoring data includes at least the current coordinates of the plot and the observed carbon storage value. The coordinate data is obtained from field measurements using the Global Positioning System or the BeiDou Navigation Satellite System, while the carbon storage data is obtained from various monitoring methods such as IoT sensors, satellite remote sensing, or manual ground surveys.
[0035] The system first calculates the single-point deviation value for each sample plot based on the plot monitoring data. The single-point deviation value quantifies the degree of deviation between the current observed value and the expected value. After calculating the single-point deviation value, the system compares this value with a preset first threshold. The setting of the first threshold varies depending on the type of deviation value.
[0036] The single-point deviation value includes coordinate deviation value and / or carbon storage volatility. When the single-point deviation value only includes coordinate deviation value, if the coordinate deviation value does not exceed the corresponding first threshold, the data is judged to be normal; if the coordinate deviation value exceeds the corresponding first threshold, it is marked as abnormal.
[0037] When the single-point deviation value only includes carbon storage volatility, if the carbon storage volatility does not exceed the corresponding first threshold, the data is considered normal; if the carbon storage volatility exceeds the corresponding first threshold, it is marked as abnormal.
[0038] When a single-point deviation value includes both coordinate deviation and carbon storage volatility, if neither the coordinate deviation nor the carbon storage volatility exceeds its corresponding first threshold, the data is considered normal; if either the coordinate deviation or the carbon storage volatility exceeds its corresponding first threshold, the data is considered abnormal. In other words, the data is considered abnormal in all three cases: the coordinate deviation exceeds the corresponding first threshold, but the carbon storage volatility does not; the coordinate deviation does not exceed the corresponding first threshold, but the carbon storage volatility exceeds the corresponding first threshold; and both the coordinate deviation and the carbon storage volatility exceed the corresponding first threshold.
[0039] Through step S2, this invention achieves efficient and rapid release of normal data. The calculation of single-point deviation values involves only simple mathematical operations between the current observed value and the baseline predicted value, with extremely low computational complexity, which can be completed in milliseconds. For the normal sample plot data, which constitutes the vast majority in carbon sink verification, step S2 can quickly determine and release the data without triggering subsequent computationally intensive operations such as spatial collaborative verification, significantly reducing the overall computational resource consumption of the system.
[0040] By comparing the single-point deviation value with a first threshold, step S2 divides all sample plot data into two main categories: normal and abnormal. Sample plot data marked as abnormal will proceed to step S3 for deep spatial co-verification, while sample plot data determined to be normal will be directly released. This ensures that limited computing resources are concentrated on the in-depth analysis of suspected abnormal data, avoiding resource waste.
[0041] S3: In response to the sample plot being marked as abnormal, query the spatial neighborhood set of the sample plot, and obtain the single-point residual of the abnormal sample plot and each sample plot in the spatial neighborhood set based on the monitoring data and the baseline prediction data of the physical process model, and determine the spatial collaborative anomaly index of the abnormal sample plot.
[0042] The main purpose of step S3 is to perform spatial co-verification on the sample plots marked as anomalous in step S2. This involves introducing spatial neighborhood information to quantify the consistency between the anomaly and the spatial trend. The single-point deviation verification in step S2 only considers the deviation between the observed and expected values of the sample plot itself, but does not consider the spatial rationality of this deviation. A sample plot's deviation from the expected value may stem from measurement errors or reflect real regional changes; a single-point deviation value alone cannot distinguish between these two situations. Step S3 upgrades the single-point verification to a dual verification of single-point and neighborhood co-verification by querying the spatial neighborhood set of the sample plot and calculating the spatial co-anomaly index, providing a quantitative basis for the anomaly type classification in step S4.
[0043] Step S3 is triggered after a sample plot is marked as an anomaly in step S2. The system first queries the spatial neighborhood set of the anomaly sample plot. The spatial neighborhood set is pre-constructed based on the spatial autocorrelation structure of the sample plot network. Specifically, during the project initialization phase, the system calculates the empirical variogram based on historical monitoring data, fits the theoretical variogram model, determines the range, and defines all sample plots within the range as the spatial neighborhood of that sample plot.
[0044] After acquiring the spatial neighborhood set, the system calculates the single-point residual for each sample plot in the anomalous sample plot and its spatial neighborhood set based on the monitoring data and the baseline prediction data of the physical process model. The single-point residual is defined as the latest measured value of the sample plot minus the baseline prediction value. The baseline prediction value can be derived from historical time-series trends or extrapolated from the physical process model.
[0045] The system further determines the spatial collaborative anomaly index of anomalous plots based on the single-point residuals of all plots. Through step S3, this invention introduces spatial neighborhood collaborative information, upgrading anomaly determination from isolated judgment to collaborative judgment.
[0046] Understandably, the spatial co-anomaly index is centered at 0, and the larger the absolute value, the more significant the contradiction between plot anomalies and spatial trends. This quantitative expression makes the degree of anomalies between different plots and different regions comparable, providing a unified metric for the threshold classification in step S4.
[0047] S4: Based on the value of the spatial coordinated anomaly index, determine the anomaly type of the sample plot for correction, wherein the anomaly type includes at least coordinated anomaly, local pseudo-anomaly, and extreme anomaly; If the absolute value of the spatial coordination anomaly index is less than or equal to the second threshold, it is determined to be a coordination anomaly. If the absolute value of the spatial coordination anomaly index is greater than the second threshold and less than or equal to the third threshold, it is determined to be a local pseudo-anomaly. If the absolute value of the spatial coordination anomaly index is greater than the third threshold, it is determined to be an extreme anomaly; wherein the third threshold is greater than the second threshold.
[0048] The main purpose of step S4 is to classify anomalous plots into three types based on the spatial co-anomaly index calculated in step S3: co-anomalies, local pseudo-anomalies, and extreme anomalies. Different causes of anomalies require different handling strategies. Co-anomalies reflect real regional changes and should trigger regional verification; local pseudo-anomalies may originate from measurement errors or local interference and should be handled through spatial smoothing correction; extreme anomalies may originate from equipment failure or human error and should trigger emergency on-site verification. Step S4 automatically distinguishes between the three anomaly types using preset second and third thresholds, providing accurate type judgment criteria for subsequent targeted correction operations.
[0049] Step S4 is executed after the spatial coordination anomaly index is calculated in step S3. The system compares the absolute value of the spatial coordination anomaly index with preset second and third thresholds, where the third threshold is greater than the second threshold.
[0050] The specific rules for classification are as follows: When the absolute value of the spatial co-anomaly index is less than or equal to the second threshold, the system determines that the anomaly type of the sample plot is a co-anomaly. In this case, the degree of disconnect between the actual residual of the sample plot and the expected residual of its neighboring area is small, indicating that the anomaly state of the sample plot is highly consistent with the spatial trend. In other words, the sample plot does not deviate from the expectation in isolation; its neighboring sample plots have also deviated in the same direction and to the same extent. Therefore, the anomaly reflects a real regional change rather than random error.
[0051] When the absolute value of the spatial co-anomaly index is greater than the second threshold and less than or equal to the third threshold, the system determines the anomaly type of the sample plot as a local pseudo-anomaly. In this case, there is a certain degree of disconnect between the actual residual of the sample plot and the expected residual of the neighboring area, indicating that the anomalous part of the sample plot deviates from the spatial trend, but the degree of disconnection has not reached the level of complete independence. This type of anomaly may originate from measurement errors, sensor malfunctions, or local environmental disturbances, rather than regional changes in physical processes.
[0052] When the absolute value of the spatial co-anomaly index exceeds the third threshold, the system classifies the anomaly type of the sample plot as an extreme anomaly. In this case, the actual residual of the sample plot deviates significantly from the expected residual of the neighboring area, indicating that the anomaly of the sample plot is completely independent of the spatial structure. This type of anomaly usually originates from errors in the data acquisition path, such as uncalibrated equipment, mislabeled tags, or GPS misalignment, requiring an emergency on-site verification.
[0053] In a preferred embodiment, the second threshold is set to 1.96, and the third threshold is set to 3.0. 1.96 corresponds to the critical value of the 95% confidence interval of the standard normal distribution, and 3.0 corresponds to the critical value of the 99.7% confidence interval. These two thresholds are obtained based on the standardization characteristics of the spatial co-anomaly index itself, requiring no additional calibration for different regions or monitoring methods, and thus possessing good universality.
[0054] Through step S4, the present invention achieves accurate three-level classification of anomaly types. Based on the comparison of spatial co-anomaly index and two-level thresholds, anomalies are refined into three types: co-anomalies, local pseudo-anomalies, and extreme anomalies, providing an accurate decision-making basis for subsequent targeted corrections.
[0055] Furthermore, collaborative anomalies and local pseudo-anomalies may appear similar in terms of single-point deviation values, both showing deviations from expected plot observations. However, their causes and processing strategies are completely different: collaborative anomalies require triggering regional verification, while local pseudo-anomalies only require spatial smoothing correction. Step S4, by introducing spatial neighborhood collaborative information, effectively distinguishes these two superficially similar but substantially different anomaly types, avoiding inappropriate correction strategies due to type confusion.
[0056] In a preferred embodiment of the present invention, the single-point deviation value is specifically: The single-point deviation value includes a coordinate deviation value, obtained by the difference between the current coordinate and the historical coordinate; and / or, The single-point deviation value includes the carbon storage volatility, which is obtained by the deviation rate between the current carbon storage and the baseline prediction value of the physical process model.
[0057] The main purpose of this implementation method is to provide clear operational guidance for the calculation of single-point deviation values by specifying single-point deviation values into two types: coordinate deviation values and carbon storage volatility.
[0058] The first type of single-point deviation is the coordinate deviation value. The coordinate deviation value is obtained by comparing the current coordinates with historical coordinates. In the actual operation of carbon sink verification, fixed sample plots are usually located with the assistance of the Global Positioning System (GPS) or the BeiDou Navigation Satellite System, and each sample plot has a unique historical coordinate record. When new monitoring data arrives, the system compares the currently measured sample plot coordinates with the historical coordinates stored in the database, calculates the spatial distance difference between the two, and this difference is the coordinate deviation value. The coordinate deviation value reflects whether the sample plot has experienced a positional shift outside the normal range.
[0059] The second type of single-point deviation is the carbon storage volatility. Carbon storage volatility is obtained by calculating the deviation rate between the current carbon storage and the baseline prediction of the physical process model. The physical process model has been pre-calculated based on soil mechanical parameters and can output the baseline prediction of carbon storage for each grid cell. When new monitoring data arrives, the system compares the currently measured carbon storage observation with the baseline prediction output by the physical process model and calculates the deviation rate, which is the carbon storage volatility. Carbon storage volatility reflects whether the observed carbon storage deviates from the prediction of the physical model.
[0060] It should be noted that the single-point deviation value can include only the coordinate deviation value, only the carbon storage volatility value, or both. For example, in scenarios where only the stability of the sample plot location is of concern, only the coordinate deviation value can be used as the single-point deviation value; in scenarios where only the magnitude of carbon storage changes is of concern, only the carbon storage volatility value can be used as the single-point deviation value; and in scenarios where a comprehensive assessment of the sample plot data quality is required, both can be used for a combined judgment.
[0061] Based on this, if the single-point deviation value does not exceed the corresponding first threshold, the data is determined to be normal; if any of the single-point deviation values exceeds the corresponding first threshold, it is marked as abnormal, specifically: The single-point deviation value includes the coordinate deviation value. If the coordinate deviation value Δd ≤ 5m, the data is considered normal; if the coordinate deviation value Δd > 5m, it is marked as abnormal; or... The single-point deviation value includes the carbon storage volatility. If the carbon storage volatility Δc ≤ 10%, the data is considered normal; if the carbon storage volatility Δc > 10%, it is marked as abnormal; or... The single-point deviation value includes the coordinate deviation value and the carbon storage volatility. If the coordinate deviation value Δd ≤ 5m and the carbon storage volatility Δc ≤ 10%, the data is considered normal. If the coordinate deviation value Δd > 5m or the carbon storage volatility Δc > 10%, it is marked as abnormal.
[0062] The main purpose of this embodiment is to clarify the specific conditions for determining whether data is normal and labeled as abnormal under different single-point deviation value types, and to realize the separate setting of the threshold for coordinate deviation value and the threshold for carbon storage volatility, as well as the separate provisions for single-condition judgment and combined-condition judgment.
[0063] The first approach is a single-condition judgment based on coordinate deviation. In this approach, the first threshold is specified as 5 meters. When the coordinate deviation is less than or equal to 5 meters, the system determines the data to be normal; when the coordinate deviation is greater than 5 meters, the system marks it as abnormal. The 5-meter threshold is set according to the industry standard for monitoring sample sites in carbon sink projects, which stipulates that the center point reset error of the sample site is usually required to be no more than ±5 meters.
[0064] The second approach is a single-condition judgment based on carbon storage volatility. In this approach, the first threshold is specified as 10%. When the carbon storage volatility is less than or equal to 10%, the system determines the data to be normal; when the carbon storage volatility is greater than 10%, the system marks it as abnormal. The 10% threshold setting corresponds to the common requirement in carbon sink measurement that the sampling accuracy is not less than 90%.
[0065] The third approach uses a combination of coordinate deviation and carbon storage volatility as the determining factor. In this approach, the system considers both deviations simultaneously. When the coordinate deviation is less than or equal to 5 meters and the carbon storage volatility is less than or equal to 10%, the system considers the data normal. When the coordinate deviation is greater than 5 meters or the carbon storage volatility is greater than 10%, the system marks it as abnormal. Data is only considered normal when both indicators do not exceed their respective thresholds; if either indicator exceeds its threshold, it is marked as abnormal.
[0066] The location stability of fixed sample plots is a fundamental prerequisite for the reliability of carbon sequestration monitoring data. When a sample plot deviates from its normal range, it may indicate that the plot markers have been damaged, the plot boundaries have undergone human or natural changes, or there is a significant error in the GPS measurement. By calculating the coordinate deviation value and comparing it with a first threshold in subsequent steps, the system can promptly identify location anomalies.
[0067] Setting a carbon storage volatility rate can effectively identify anomalies in carbon storage observations. Carbon storage is a core indicator for carbon sink verification, and its observed values should maintain reasonable consistency with the predictions of physical process models. When the carbon storage volatility rate exceeds the normal range, it may indicate sensor malfunction, human error in recording, or physical changes in the area that have not been captured by the model. By calculating the carbon storage volatility rate and comparing it with a first threshold in subsequent steps, the system can promptly identify carbon storage anomalies.
[0068] As a preferred embodiment of the present invention, determining the spatial coordination anomaly index specifically includes: Calculate the single-point residuals of the anomaly plot i. ,in The sample monitoring data for sample plot i. Baseline prediction data for the physical process model; Computational spatial expected residual ,in Let i be the set of spatial neighborhoods of the sample plot i. Let i be the single-point residual of any plot j in the spatial neighborhood set of plot i. Let j be the kriging weight of plot j relative to plot i; Calculate the absolute value of the disconnection ; Divide the absolute value of the dissociation by the standard error of the Kriging theory. The spatial coordination anomaly index was obtained. The standard error of the Kriging theory mentioned therein It is determined by the spatial configuration of the spatial neighborhood set.
[0069] The main objective of this implementation method is to transform the spatial co-anomaly index from raw monitoring data to a dimensionless standardized score by starting with the calculation of single-point residuals, proceeding through Kriging weighted estimation of spatial expected residuals, solving for the absolute value of the disjointness, and finally standardizing by dividing by the Kriging theoretical standard error. This index is used to quantify the degree of disjointness between the anomaly of the target sample plot and the expected state of its spatial neighborhood, providing quantitative input for anomaly type classification.
[0070] The system first calculates the single-point residuals of sample plot i marked as an anomaly, expressed as: ; in, The sample monitoring data for sample plot i is the latest measured value of that sample plot, which can be obtained from various monitoring methods such as IoT sensors, satellite remote sensing or manual ground collection. This represents the baseline prediction data for plot i at the current time using the physical process model. The sign of the single-point residual reflects the direction of deviation of the measured value from the baseline prediction; a positive value indicates that the measured value is higher than expected, and a negative value indicates that the measured value is lower than expected. The larger the absolute value, the more significant the anomaly.
[0071] The system further calculates the spatial expected residual, which is expressed as follows: ; in, Let i be the set of spatial neighborhoods of sample plot i. This set is determined by the spatial coordination module based on the range of the variogram function during the system initialization phase. That is, all sample plots within the range of the variogram constitute the spatial neighborhood of the sample plot. Let be the single-point residual of any location j in the spatial neighborhood set; Let be the kriging weight of neighboring plot j relative to target plot i.
[0072] Kriging weights are obtained by solving the ordinary Kriging equations and satisfy the constraint that the sum of the weights of all neighboring plots is 1. Their magnitude depends on the distance and spatial orientation of the neighboring plots relative to the target plot. The closer the neighboring plots are and the stronger their spatial correlation, the greater the weight they receive.
[0073] Spatial expected residuals indicate that, if the spatial autocorrelation structure holds, the residual of the target plot i should be close to the weighted average of the residuals of its neighboring plots. This weighted average is the residual level that the spatial structure believes the target plot should have.
[0074] The system calculates the absolute value of the difference between the actual single-point residual of the target sample plot and the spatially expected residual, i.e. The absolute value of this disjointness reflects the gap between the actual residuals of the target plot and the expected residuals of its neighbors. The larger the difference, the more significant the contradiction between the plot's anomalous state and its spatial trend.
[0075] It is understandable that, because the spatial prediction accuracy varies in different regions, the prediction error is inherently larger in sparse areas, and the same absolute value of the disconnection has different statistical significance in sparse and dense areas.
[0076] The system divides the absolute value of the dissociation by the standard error of the Kriging theory. The spatial coordination anomaly index is obtained, and its expression is: ; Among them, the standard error of Kriging theory It is determined by the spatial configuration of the spatial neighborhood set. Specifically, It is the theoretical standard error of ordinary kriging interpolation at location i of the target plot, i.e., the kriging variance. The square root of.
[0077] The standard error is determined by two factors. First, the parameters of the variogram model, including the sill value, nugget value, and range, are obtained by fitting historical monitoring data during the initialization of the spatial coordination module. Second, the spatial configuration of the neighboring plots, i.e., the distance and orientation distribution of each neighboring plot relative to the target plot.
[0078] In mathematics, the square of the standard error of Kriging's theory, i.e., the Kriging variance, can be expressed as: ; in, The prior variance of the target sample plot is equal to the sill value of the variogram. The covariance vector between the target plot and each neighboring plot is calculated by the variogram model based on spatial distance. The covariance matrix between all pairs of neighboring sample plots is also calculated using the variogram model. By dividing by the Kriging theory standard error, the absolute value of the disjointness is transformed into a dimensionless standardized fraction, making the degree of disjointness comparable across different regions and sampling densities.
[0079] This invention achieves a quantitative and standardized expression of the degree of anomaly disconnect. The Spatial Co-anomaly Index expresses the degree of disconnect between the anomaly of the target plot and the spatial trend as a dimensionless standardized score centered at 0. The larger the absolute value, the more significant the contradiction between the plot anomaly and the spatial trend. This quantitative expression provides a unified measurement scale for the degree of anomaly between different plots and different regions, providing a quantifiable input basis for classification and discrimination.
[0080] The introduction of Kriging weights ensures optimal unbiasedness in the estimation of spatial expected residuals. The ordinary Kriging method solves for the optimal weights of each neighboring plot by minimizing the variance of the estimation error, while simultaneously satisfying the constraint of unbiased estimation. Compared to simple inverse distance weighting or arithmetic mean, Kriging weights fully consider the spatial autocorrelation structure among plots, making the estimation of spatial expected residuals statistically optimal. This characteristic ensures that the spatial co-anomaly index can truly reflect the degree of deviation between plot anomalies and spatial trends, rather than being biased due to improper weight selection.
[0081] Furthermore, the introduction of the Kriging theory standard error enables the standardization process to be adaptive. Different regions have different sample plot densities, resulting in varying spatial prediction accuracies; sparsely populated areas inherently have larger prediction errors. By dividing the absolute value of the discrepancy by the Kriging theory standard error, spatial prediction accuracy is incorporated into the standardization process. The same absolute value of discrepancy will be assigned different standardization scores in sparsely populated and densely populated areas because the Kriging theory standard error is larger in sparse areas and smaller in dense areas. This adaptive standardization avoids misjudgments caused by differences in spatial sampling density, ensuring that the spatial co-anomaly index reliably reflects the degree of spatial discrepancy of anomalies under different monitoring conditions.
[0082] The calculation of the Spatial Co-anomaly Index (SCAI) ensures reliable quantitative input for classification. A smaller SCAI level indicates that the plot anomaly is consistent with the spatial trend and should be classified as a co-anomaly; a SCAI level in the middle range indicates that the plot anomaly is partially disconnected from the spatial trend and should be classified as a local pseudo-anomaly; a larger SCAI level indicates that the plot anomaly is completely independent of the spatial structure and should be classified as an extreme anomaly.
[0083] In a preferred embodiment of the present invention, after determining that a cooperative anomaly has occurred, attribution analysis is also included: Based on the spatial coordination anomaly index of the sample plot ,pass Obtain the spatial chain likelihood ; The independence score for the collection of the sample monitoring data from the sample plots was based on the data collected from the sample plots. ,pass Seemingly obtaining a bloodline chain ,in The neutral threshold is set to 0.5, and k is the steepness coefficient, set to 6.0. The physical deviation between the plot monitoring data and the predicted values of the physical process model is used as a basis. ,pass Obtain the physical chain likelihood ,in The standard deviation of the predictions for the physical process model; Based on the landslide probability, prior probabilities are set for the spatial chain likelihood, the lineage chain likelihood, and the physical chain likelihood, respectively. The posterior probabilities are then adjusted using Bayes' theorem to obtain the spatial chain likelihood, the lineage chain likelihood, and the physical chain likelihood. If the maximum posterior probability of multiple posterior probabilities is less than the fourth threshold, then output an uncertain state and end the process; If the maximum a posteriori probability is greater than or equal to the fourth threshold, then the hypothesis corresponding to the maximum a posteriori probability is adopted as the attribution result; Among them, the attribution result corresponding to the spatial chain likelihood is a local pseudo-anomaly; the attribution result corresponding to the physical chain likelihood is a physically driven true anomaly; and the attribution result corresponding to the lineage chain likelihood is an error in the acquisition path.
[0084] The main objective of this implementation is to provide an attribution analysis method that is performed when a plot anomaly is determined to be a collaborative anomaly.
[0085] Although anomalies have been categorized into three types—coordinated anomalies, local pseudo-anomalies, and extreme anomalies—coordinated anomalies can still originate from two different fundamental causes: first, physical processes driving true anomalies, where physical processes such as landslides and pests cause substantial changes in the spatial structure of the sample plot; and second, errors in the data collection path, where data deviations occur during the collection or transfer process due to human intervention such as uncalibrated equipment or mislabeled labels. The corresponding handling strategies for these two causes are drastically different. Physically driven true anomalies should trigger topology reconstruction and regional verification, while errors in the data collection path should be addressed by tracing equipment status and operational records.
[0086] By introducing a Bayesian causal arbitration mechanism, the inferences of spatial chains, physical chains, and kinship chains are incorporated into a unified probabilistic framework for quantitative comparison, providing an interpretable probabilistic decision-making basis for the attribution analysis of collaborative anomalies.
[0087] The system calculates the spatial chain likelihood based on the Spatial Co-occurrence Anomaly Index (SCAI) of the sample plots, and its expression is as follows: ; This formula substitutes the SCAI value into the probability density function of the standard normal distribution. The density value is highest when SCAI=0, indicating that the plot anomaly is completely consistent with the spatial trend; the larger |SCAI| is, the lower the density value, indicating a more significant contradiction between the plot anomaly and the spatial trend. Spatial chain likelihood reflects the conditional probability density of observing the current SCAI value under the assumption of a local spurious anomaly dominated by the spatial chain.
[0088] The system calculates the kinship likelihood based on the sampling independence score I of the sample plot monitoring data, and its expression is: ; in The neutral threshold is set to 0.5; k is the steepness coefficient, set to 6.0. The acquisition independence score I is weighted and synthesized by the lineage state machine based on three sub-dimensions: device consistency, operational consistency, and link uniqueness, with a value ranging from 0 to 1. A value closer to 1 indicates a more independent acquisition path and a greater likelihood of unique errors; a value closer to 0 indicates a highly shared acquisition path. When I = 0.8, the lineage chain likelihood is approximately 0.86; when I = 0.2, the lineage chain likelihood is approximately 0.14, showing a smooth mapping and good discriminative power. The lineage chain likelihood reflects the conditional probability density of the current independence score observed under the assumption of lineage-dominated acquisition path errors.
[0089] The system calculates the physical chain likelihood based on the physical deviation Δ between the sample plot monitoring data and the predicted values from the physical process model. The expression for this likelihood is: ; in Δ represents the standard deviation of the physical process model's predictions, calculated from the historical backsubstitution error during the model calibration phase. Δ is the absolute value of the difference between the measured values and the physical model's predicted values. The closer Δ is to 0, the higher the density value, indicating a high degree of agreement between the measured data and the physical calculations, strongly supporting a physical-driven attribution; Δ exceeding 2 to 3 times indicates a higher density. The extremely low temporal density indicates that the physical model cannot explain the anomaly. The physical chain likelihood reflects the conditional probability density of observing the current physical deviation value under the assumption of a physically driven true anomaly dominated by the physical chain.
[0090] The system sets prior probabilities for spatial chain likelihood, lineage chain likelihood, and physical chain likelihood based on the landslide probability. The prior probabilities are dynamically injected by the physical model; when the landslide probability is high, the prior probability of the physical-driven true anomaly hypothesis is increased.
[0091] Specifically, based on the landslide probability, prior probabilities are set for the spatial chain likelihood, the kinship chain likelihood, and the physical chain likelihood, respectively, as follows: If the landslide probability reaches the warning threshold, the prior probability of the spatial chain likelihood is P(H1)=0.2, the prior probability of the physical chain likelihood is P(H2)=0.7, and the prior probability of the bloodline chain likelihood is P(H3)=0.1. Otherwise, based on the prior probability P(H1)=0.2 of the spatial chain likelihood, the prior probability P(H2)=0.7 of the physical chain likelihood, and the prior probability P(H3)=0.1 of the lineage chain likelihood, the prior probability of the physical chain likelihood is reduced, while the prior probabilities of the spatial chain likelihood and the lineage chain likelihood are increased.
[0092] The setting of prior probabilities directly affects the calculation results of Bayesian posterior probabilities, and thus the final arbitration decision. This embodiment provides quantifiable and reproducible prior setting rules for Bayesian arbitration by specifying the prior probability values for two scenarios: when the landslide probability reaches the warning threshold and when it does not.
[0093] The first scenario is when the landslide probability reaches the warning threshold. In this case, the system sets the prior probability P(H2) of the physical chain likelihood to 0.7, the prior probability P(H1) of the spatial chain likelihood to 0.2, and the prior probability P(H3) of the lineage chain likelihood to 0.1. When the physical model predicts that the landslide probability reaches the warning threshold, the physical-driven true anomaly hypothesis obtains the highest prior probability.
[0094] The second scenario is when the landslide probability does not reach the warning threshold. In this case, based on the prior probabilities of the first scenario, the system reduces the prior probability of the physical chain likelihood while increasing the prior probabilities of the spatial chain likelihood and the lineage chain likelihood. This adjustment reflects that when physical warning support is lacking, the system's confidence in the physical driving hypothesis should decrease accordingly, while allocating more confidence to the hypotheses corresponding to the spatial chain and the lineage chain.
[0095] For example, P(H2) = 0.5, P(H1) = 0.3, P(H3) = 0.2; or P(H1) = 0.4, P(H2) = 0.3, P(H3) = 0.3, and the present invention does not impose specific limitations on these values.
[0096] When the landslide probability reaches the warning threshold, the physical chain hypothesis receives the highest prior; when the landslide probability does not reach the warning threshold, the physical chain prior is reduced. This dynamic adjustment mechanism allows Bayesian arbitration to fully utilize ex-ante warning information about physical processes. A landslide probability reaching the warning threshold means that the physical process model has identified a risk of instability in the region; under this condition, the physical-driven true anomaly hypothesis receives the highest prior.
[0097] Based on this, the system adjusts the three likelihoods separately using Bayes' theorem to obtain their respective posterior probabilities. The magnitudes of the multiple posterior probabilities are then compared.
[0098] Specifically: ; in, It is the product of the three-chain likelihoods, that is k takes any value from 1 to 3; j takes any value from 1 to 3.
[0099] Obtain the posterior probability of the spatial chain likelihood. Posterior probability of kinship likelihood ; Posterior probability of physical chain likelihood .
[0100] If the maximum a posteriori probability is less than the fourth threshold, the system outputs an uncertain state and terminates the process without forcing a decision. If the maximum a posteriori probability is greater than or equal to the fourth threshold, the system adopts the hypothesis corresponding to the maximum a posteriori probability as the attribution result. The attribution result corresponding to the spatial chain likelihood is a local pseudo-anomaly, the attribution result corresponding to the physical chain likelihood is a physically driven true anomaly, and the attribution result corresponding to the lineage chain likelihood is an error in the acquisition path. Specifically, the fourth threshold is set to 60%.
[0101] This invention achieves unified probabilistic arbitration for three-chain conflicts. By mapping the three inferences to a unified hypothesis space and using Bayes' theorem to calculate posterior probabilities for decision-making, the resolution of multi-source conflicts is upgraded from simple voting or priority ranking to interpretable probabilistic comparison.
[0102] Precise quantification of likelihood was achieved through continuous probability density mapping. All three likelihoods were calculated using continuous probability density functions: the spatial chain used SCAI as the standard normal input to obtain its density value; the physical chain used the physical deviation value substituted into the Gaussian density function; and the kinship chain used the independence score transformed by the Logistic function. All three are dimensionless relative probability density values and do not require normalization. The denominator of the Bayesian formula automatically performed weighted normalization, ensuring the statistical rigor of the arbitration results.
[0103] When the maximum posterior probability is less than the fourth threshold, the system does not force a decision but outputs an uncertain state and terminates the process. This truncation mechanism allows the system to proactively acknowledge uncertainty when evidence is insufficient, avoiding erroneous attributions that might result from blind decision-making, demonstrating the system's self-awareness.
[0104] The prior probability is not fixed, but dynamically adjusted based on the landslide probability output by the physical model. When the landslide probability is high, the prior probability of the physical-driven true anomaly hypothesis is increased accordingly, enabling the arbitration result to fully utilize the pre-warning information of the physical process, thus achieving a deep integration of physical-driven and data-driven approaches.
[0105] In one embodiment, if the attribution result is a physically driven true anomaly, then a topology reconstruction of the physical feasible region constraints is performed: Read the physical boundary output by the physical process model, generate an allowable transition matrix based on the physical boundary, and migrate the sample plots located within the physical boundary to a new sub-network; After the reconstruction is completed, a freeze period is set, during which no new reconstruction requests will be accepted.
[0106] The main objective of this embodiment is to provide a method for topology reconstruction based on physical feasible domain constraints when the attribution result is a physically driven true anomaly. When Bayesian arbitration confirms that a plot anomaly originates from physical processes such as landslides, the spatial topology of the plot has undergone substantial changes, and plots that originally belonged to the same spatial network have become decoupled due to the physical process. If the original spatial neighborhood relationships are continued for subsequent verification, it will lead to erroneous judgments based on failed spatial relationships.
[0107] By introducing the physical boundary output by the physical model as a hard constraint for topology reconstruction, the search space for topology reconstruction is restricted to the physically feasible domain, and a freeze period is set after reconstruction is completed, which fundamentally eliminates the cascading effects that may be caused by the free editing mode.
[0108] The system first reads the physical boundary output by the physical process model. This physical boundary is output by the landslide stability mechanics model or the pest and disease diffusion model, such as the landslide body boundary polygon output by the landslide model. This boundary represents the spatial extent of the influence of the physical process and is the basis for determining which sample plots should be relocated.
[0109] The system generates an allowable transition matrix based on physical boundaries. The allowable transition matrix is a binary matrix where each element corresponds to a sample plot. A value of 1 indicates that the sample plot is allowed to be migrated to a new subnetwork, and a value of 0 indicates that the sample plot is not allowed to migrate. The condition for allowing migration is that the sample plot is located inside the physical boundary.
[0110] The system migrates sample plots located within the physical boundaries to a new subnetwork. The migration operation includes disconnecting the spatial edges between the migrated sample plots and the original spatial network, and establishing a hybrid edge between the migrated sample plots and the new subnetwork. This hybrid edge carries a physical attribution identifier and a physical model version number, recording the causal basis for this topology change.
[0111] After the refactoring is completed, a freeze period is set. During the freeze period, the system will not accept new refactoring requests. Even if a new physical warning event arrives, the system will only record the event log and will not perform the refactoring operation.
[0112] This embodiment uses the physical model output as a hard constraint for topology reconstruction. By using the physical boundary as the basis for generating the allowed transition matrix, the infinite free editing problem is reduced to finite editing within the physically feasible domain, eliminating the risks of cascading effects and physically infeasible reconstruction.
[0113] The freeze period effectively prevents system oscillations caused by frequent reconstructions. After a physical process occurs, the system needs time to confirm the stability of the new spatial structure. By setting a freeze period after reconstruction, it ensures that no new reconstruction requests are accepted for at least one monitoring cycle, avoiding repeated reconstructions triggered by minor fluctuations in the physical model output, thus enhancing system stability.
[0114] The migrated sample plots are associated with physical attribution identifiers through mixed edges, so that each topological change has a clear physical causal basis, providing a complete causal chain for subsequent historical tracing and auditing.
[0115] Specifically, the topology reconstruction of the physically feasible region constraints includes: Calculate the ratio η of the geometric intersection area of the polygon of the sample plot and the physical boundary to the total area of the sample plot itself. If η ≥ 50% and the safety factor of the center point of the sample plot is less than the first safety threshold, then migration is allowed. Otherwise, migration is prohibited.
[0116] The system first calculates the ratio η of the geometric intersection area between the plot polygon and the physical boundary to the total area of the plot itself. The plot polygon is a polygonal region formed in space according to the fixed boundary of the plot, and the physical boundary is the boundary of the influence area output by the physical process model. The geometric intersection area reflects the degree to which the plot is affected by the physical process; the larger the intersection area, the deeper the plot is located within the physical influence area.
[0117] The system further determines whether two conditions are met simultaneously. The first condition is whether the geometric intersection area ratio η is greater than or equal to 50%. The second condition is whether the safety factor of the sample plot center point is less than the first safety threshold. The safety factor is a mechanical stability index output by the physical process model; when the safety factor is less than 1, it indicates that the area has a risk of instability. Only when both conditions are met simultaneously will the system allow the sample plot to migrate to the new subnetwork; otherwise, the system prohibits the migration of the sample plot.
[0118] The dual-condition setting enhances the robustness of the decision. Judging solely based on the percentage of geometric intersection area may lead to misjudgments; a plot may have most of its area within the boundary, but its center point might still be in a stable region. Migration is only permitted when the plot is both spatially covered by the physical boundary and mechanically unstable, thus avoiding erroneous migrations that might result from relying solely on geometric determination.
[0119] Furthermore, when more than half of the area of a sample plot is located inside the physical boundary, the sample plot has been substantially affected by physical processes, and it is physically reasonable to relocate it to a new subnetwork; when only a small part of the sample plot is located inside the boundary, the main part of the sample plot is still in the original spatial structure, and it is more reasonable to retain it in the original network.
[0120] In another embodiment of this implementation, if the attribution result is a local spurious anomaly or an incorrect acquisition path, then lineage inversion and correction are performed: A causal graph is pre-constructed, which contains multiple nodes and causal edges connecting the nodes; the causal edges include four types: physical edges, spatial edges, lineage edges, and hybrid edges, and each causal edge carries an uncertainty weight; Tracing back from the node corresponding to the abnormal sample plot to the original acquisition node along the bloodline edge in the causal graph, the aggregation calculation parameters are adjusted or the acquisition device status is marked.
[0121] The main objective of this embodiment is to provide a lineage inversion and correction method when the attribution result is a local pseudo-anomaly or an incorrect data acquisition path. When Bayesian arbitration confirms that the anomaly in the sample plot originates from an incorrect data acquisition path or a local pseudo-anomaly, it is not necessary to trigger topology reconstruction. Instead, the source of the anomaly should be located and the corresponding correction should be performed by tracing the data flow link. By pre-constructing a causal graph containing various types of causal edges and tracing back from the anomaly node to the original data acquisition node along the lineage edges, accurate location and targeted correction of the anomaly source are achieved.
[0122] The system pre-constructs a causal graph. This graph contains multiple nodes and causal edges connecting them. Nodes represent entities involved in various stages of data collection, processing, aggregation, and verification, while causal edges represent the causal relationships between nodes. There are four types of causal edges: physical edges representing the influence of physical processes on sample plots, spatial edges representing spatial statistical dependencies between sample plots, lineage edges representing dependencies between steps in the data flow process, and hybrid edges representing composite causal relationships between physical processes, spatial structures, and data flow. Each causal edge carries an uncertainty weight for subsequent uncertainty propagation calculations.
[0123] The system traces back along the lineage edges in the causal graph from the node corresponding to the anomalous sample plot to the original data collection node. The lineage edges record the entire lifecycle path of data from collection to verification, including the original data collection node, standardization node, aggregation node, and verification node. Reverse tracing starts from the current anomalous node and traverses upwards along the reverse direction of the lineage edges until the original data collection node is reached.
[0124] The system adjusts aggregation calculation parameters or marks the status of acquisition devices based on the tracing results. When the attribution result indicates an incorrect acquisition path, the system traces back to the specific acquisition device, marks its fault status, and generates a calibration notification; when the attribution result indicates a local pseudo-anomaly, the system adjusts the relevant parameters in the aggregation calculation.
[0125] This invention enables precise location of the anomaly source. By tracing back from the anomaly node to the original acquisition node along the bloodline edge, the system can accurately pinpoint which stage of the data flow the anomaly occurred at—whether it was the sensor acquisition stage, the data transmission stage, or the data processing stage. This precise location capability allows corrective measures to directly target the anomaly source, rather than merely superficially correcting the final conclusion.
[0126] The introduction of uncertainty weights provides the foundation for subsequent uncertainty propagation and truncation. The uncertainty weights carried by each causal edge enable the system to quantify the propagation path and cumulative effect of uncertainty in the causal graph, providing input for uncertainty truncation checks.
[0127] In a preferred embodiment of the present invention, after the attribution result processing is completed, the method further includes: A causal graph is pre-constructed, which contains multiple nodes and causal edges connecting the nodes; the causal edges include four types: physical edges, spatial edges, lineage edges, and hybrid edges, and each causal edge carries an uncertainty weight; The uncertainty value of the processing operation is calculated based on the uncertainty weight. If the uncertainty value exceeds the fifth threshold, automatic correction is rejected, the data flow status is marked as pending manual review, and the process is terminated. If the uncertainty value does not exceed the fifth threshold, then automatic correction is performed.
[0128] The main objective of this implementation is to provide an uncertainty truncation check mechanism that is executed after the attribution result processing is complete. The system has completed a full processing flow from single-point deviation screening, spatial co-verification, Bayesian causal arbitration to topological reconstruction or lineage inversion correction. However, each of the above steps inevitably introduces uncertainty, which may be amplified cascaded after propagating along the path in the causal graph. If automatic correction is rashly executed under conditions of excessive uncertainty, the corrected conclusion may deviate from the true situation, or even be worse than not correcting at all.
[0129] By setting an uncertainty cutoff before correction is implemented, the accumulated uncertainty is quantitatively assessed. Automatic correction is only allowed when the uncertainty is within a controllable range. When the uncertainty exceeds the threshold, the data flow status is proactively marked as pending manual review. This avoids the risk of the system forcibly outputting unreliable conclusions when its own decision-making reliability is insufficient.
[0130] The system calculates the uncertainty value of the current correction operation based on the uncertainty weights carried by each edge in the causal graph. The propagation of uncertainty in the causal graph follows path-dependent propagation rules. Specifically, for serial paths, uncertainty propagates according to the variance accumulation rule, meaning the total uncertainty of a serial path equals the square root of the sum of the variances of the uncertainties of each path segment. For parallel paths, uncertainty propagates according to the precision-weighted fusion rule, meaning that when multiple pieces of parallel evidence simultaneously support the same conclusion, the fused uncertainty is lower than the uncertainty of a single path. Starting from the current decision node, the system traverses all causal paths related to the correction operation along the causal graph, propagating and accumulating uncertainty level by level according to the above rules, ultimately obtaining the comprehensive uncertainty value of the correction operation.
[0131] The system compares the calculated overall uncertainty value with a preset fifth threshold. A preferred value for the fifth threshold is 20%, which is set based on the acceptable range of uncertainty in carbon sink accounting.
[0132] If the uncertainty value exceeds the fifth threshold, the system will refuse automatic correction, mark the data flow status as pending manual review, and terminate the process. In this case, the system will not force a correction conclusion but will instead delegate the decision-making power to humans. Simultaneously, the system will generate a pending work order containing an uncertainty propagation path diagram and the original evidence package of the three chains, and push it to the manual review queue. This data will not be included in the current period's aggregated carbon emissions calculation.
[0133] If the uncertainty value does not exceed the fifth threshold, the system performs automatic correction. The specific content of the automatic correction depends on the type of attribution result. For physically driven true anomalies, the system adjusts the aggregation calculation parameters according to the loss rate predicted by the physical model; for local pseudo-anomalies, the system replaces the original anomaly value with the Kriging interpolation result; for acquisition path errors, the system marks the status of the acquisition device and generates a calibration notification.
[0134] This invention achieves quantitative self-awareness of uncertainty. The system can accurately track the source, propagation path, and cumulative effect of uncertainty along the causal graph, and has a quantitative cognitive ability to recognize the reliability of its own decisions, rather than blindly performing corrective operations.
[0135] The path propagation rule, which combines series variance accumulation with parallel precision weighting, ensures the scientific rigor of uncertainty assessment. The uncertainty in the series path accumulates and amplifies at each level, reflecting the real physical law of error propagation along the causal chain; while the precision weighting fusion in the parallel path embodies the statistical principle of increased confidence when multiple sources of evidence corroborate each other. These two rules correspond to the uncertainty propagation characteristics under different topological structures.
[0136] The fifth threshold setting allows the system to proactively refuse automatic correction when uncertainty is too high. This truncation mechanism prevents the system from forcibly outputting unreliable conclusions when its own decision-making reliability is insufficient. Marking the data flow status as pending manual review instead of discarding it directly ensures that there is still a safety net of manual verification when automated processing capabilities are insufficient, achieving an organic connection between automation and manual review.
[0137] For any parts not mentioned in this invention, existing technologies can be used or referenced.
[0138] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0139] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.
Claims
1. A method for corroborating anomalies in carbon sink verification and monitoring data, characterized in that, include: Based on the acquired soil mechanical parameters, the physical process model is deduced in advance, and the safety factor and landslide probability of each grid cell are output. When the output landslide probability reaches the warning threshold, a physical warning event is generated. The method also includes, When the sample plot monitoring data arrives, the single-point deviation value of the sample plot is calculated based on the sample plot monitoring data. If the single-point deviation value does not exceed the corresponding first threshold, the data is determined to be normal. If any of the single-point deviation values exceeds the corresponding first threshold, it is marked as abnormal. The sample plot monitoring data includes at least coordinates and carbon storage, and the single-point deviation value includes coordinate deviation value and / or carbon storage volatility. In response to the sample plot being marked as anomaly, the spatial neighborhood set of the sample plot is queried. Based on the monitoring data and the baseline prediction data of the physical process model, the single-point residual of the anomalous sample plot and each sample plot in the spatial neighborhood set is obtained. The spatial co-anomaly index of the anomalous sample plot is determined, and the consistency between the sample plot anomaly and the spatial trend is quantified. Based on the value of the spatial co-anomaly index, the anomaly type of the sample plot is determined for correction purposes, wherein the anomaly type includes at least co-anomaly, local pseudo-anomaly, and extreme anomaly; If the absolute value of the spatial coordination anomaly index is less than or equal to the second threshold, it is determined to be a coordination anomaly, indicating that the sample plot anomaly is consistent with the spatial trend, triggering regional verification. If the absolute value of the spatial co-anomaly index is greater than the second threshold and less than or equal to the third threshold, it is determined to be a local pseudo-anomaly, indicating that the plot anomaly is out of sync with the neighborhood trend, and is processed by spatial smoothing correction. If the absolute value of the spatial coordination anomaly index is greater than the third threshold, it is determined to be an extreme anomaly; wherein, the third threshold is greater than the second threshold, indicating that the sample plot anomaly is independent of the spatial structure, triggering an emergency on-site verification.
2. The method of claim 1, wherein, The single-point deviation value is specifically: The single-point deviation value includes a coordinate deviation value, obtained by the difference between the current coordinate and the historical coordinate; and / or, The single-point deviation value includes the carbon storage volatility, which is obtained by the deviation rate between the current carbon storage and the baseline prediction value of the physical process model.
3. The method of claim 2, wherein, If the single-point deviation value does not exceed the corresponding first threshold, the data is determined to be normal; if any of the single-point deviation values exceeds the corresponding first threshold, it is marked as abnormal. Specifically: The single-point deviation value includes the coordinate deviation value. If the coordinate deviation value Δd ≤ 5m, the data is considered normal; if the coordinate deviation value Δd > 5m, it is marked as abnormal; or... The single-point deviation value includes the carbon storage volatility. If the carbon storage volatility Δc ≤ 10%, the data is considered normal; if the carbon storage volatility Δc > 10%, it is marked as abnormal; or... The single-point deviation value includes the coordinate deviation value and the carbon storage volatility. If the coordinate deviation value Δd ≤ 5m and the carbon storage volatility Δc ≤ 10%, the data is considered normal. If the coordinate deviation value Δd > 5m or the carbon storage volatility Δc > 10%, it is marked as abnormal.
4. The method of claim 1, wherein, The spatial coordination anomaly index is determined, specifically including: computing the single-point residuals of the plot i that are anomalous wherein is plot monitoring data for the plot i, is baseline prediction data for the physical process model; Computing spatial expected residuals wherein is a set of spatial neighbors of the plot i, is a single-point residual of any plot j in the set of spatial neighbors of the plot i, is a kriging weight of plot j on the plot i. Computing the absolute value of the decoupling ; dividing the dislocation absolute value by a kriging theoretical standard error to obtain a spatial synergy anomaly index ; wherein the kriging theoretical standard error is determined by a spatial configuration of a spatial neighborhood set.
5. The method of claim 1, wherein, After determining it to be a collaborative anomaly, attribution analysis is also included: According to the spatially coordinated anomaly index of the sample plot , by obtaining a spatial chain likelihood ; a plot monitoring data collection independence score for the plot , by obtaining the likelihood of the bloodline chain wherein is a neutral threshold value, having a value of 0.5, and k is a steepness coefficient, having a value of 6.0; a physical bias value between plot monitoring data of the plot and a predicted value of the physical process model , by obtaining a physical chain likelihood wherein is a predicted standard deviation of the physical process model; Based on the landslide probability, prior probabilities are set for the spatial chain likelihood, the lineage chain likelihood, and the physical chain likelihood, respectively. The posterior probabilities are then adjusted using Bayes' theorem to obtain the spatial chain likelihood, the lineage chain likelihood, and the physical chain likelihood. If the maximum posterior probability of multiple posterior probabilities is less than the fourth threshold, then output an uncertain state and end the process; If the maximum a posteriori probability is greater than or equal to the fourth threshold, then the hypothesis corresponding to the maximum a posteriori probability is adopted as the attribution result; Among them, the attribution result corresponding to the spatial chain likelihood is a local pseudo-anomaly; the attribution result corresponding to the physical chain likelihood is a physically driven true anomaly; and the attribution result corresponding to the lineage chain likelihood is an error in the acquisition path.
6. The method of claim 5, wherein, Based on the landslide probability, prior probabilities are set for the spatial chain likelihood, the kinship chain likelihood, and the physical chain likelihood, respectively, as follows: If the landslide probability reaches the warning threshold, the prior probability of the spatial chain likelihood is P(H1)=0.2, the prior probability of the physical chain likelihood is P(H2)=0.7, and the prior probability of the bloodline chain likelihood is P(H3)=0.
1. Otherwise, based on the prior probability P(H1)=0.2 of the spatial chain likelihood, the prior probability P(H2)=0.7 of the physical chain likelihood, and the prior probability P(H3)=0.1 of the lineage chain likelihood, the prior probability of the physical chain likelihood is reduced, while the prior probabilities of the spatial chain likelihood and the lineage chain likelihood are increased.
7. The method of claim 5, wherein, If the attribution result is a physically driven true anomaly, then perform topology reconstruction based on the physical feasible region constraints: Read the physical boundary output by the physical process model, generate an allowable transition matrix based on the physical boundary, and migrate the sample plots located within the physical boundary to a new sub-network; After the reconstruction is completed, a freeze period is set, during which no new reconstruction requests will be accepted.
8. The method according to claim 7, characterized in that, The topology reconstruction of the physically feasible region constraints specifically includes: Calculate the ratio η of the geometric intersection area of the polygon of the sample plot and the physical boundary to the total area of the sample plot itself. If η ≥ 50% and the safety factor of the center point of the sample plot is less than the first safety threshold, then migration is allowed. Otherwise, migration is prohibited.
9. The method according to claim 5, characterized in that, If the attribution result is a local spurious anomaly or an incorrect acquisition path, then perform lineage inversion and correction: A causal graph is pre-constructed, which contains multiple nodes and causal edges connecting the nodes; the causal edges include four types: physical edges, spatial edges, lineage edges, and hybrid edges, and each causal edge carries an uncertainty weight; Tracing back from the node corresponding to the abnormal sample plot to the original acquisition node along the bloodline edge in the causal graph, the aggregation calculation parameters are adjusted or the acquisition device status is marked.
10. The method according to any one of claims 7 to 9, characterized in that, After the attribution result processing is completed, it also includes: A causal graph is pre-constructed, which contains multiple nodes and causal edges connecting the nodes; the causal edges include four types: physical edges, spatial edges, lineage edges, and hybrid edges, and each causal edge carries an uncertainty weight; The uncertainty value of the processing operation is calculated based on the uncertainty weight. If the uncertainty value exceeds the fifth threshold, automatic correction is rejected, the data flow status is marked as pending manual review, and the process is terminated. If the uncertainty value does not exceed the fifth threshold, then automatic correction is performed.
Citation Information
Patent Citations
Electric power standing book data checking method and device based on blood relationship
CN108763304A
Multi-dimensional carbon flux monitoring method
CN120352580A
Foundation pit safety early-warning method based on multi-source monitoring data fusion
WO2026044862A1
