Multi-dimensional carbon flux monitoring method

Abstract

The invention discloses a multi-dimensional carbon flux monitoring method, and particularly relates to the technical field of carbon flux monitoring, and the method comprises the following steps: collecting multi-source heterogeneous sensor data, and carrying out spatial resolution unification, time synchronization alignment and data format standardization preprocessing to ensure data consistency; executing fusion validity detection, and correcting time synchronization deviation and data redundancy conflicts; according to an abnormal detection result, dynamically adjusting a fusion weight or determining a weight based on historical clustering, and combining spatial adjacency interpolation and time interpolation to realize data fusion; and finally, inputting a carbon flux inversion model, and outputting a high-precision carbon flux monitoring value to realize dynamic monitoring and prediction. According to the method, a dynamic fusion weight adjustment mechanism is adopted, the fusion weight of each data source is dynamically optimized, adaptive fusion of multi-source observation data is realized, the problems of data discontinuity and insufficient data coverage in the prior art are solved, and a multi-dimensional, full-coverage and high-resolution carbon flux fusion data set is ensured to be formed.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon flux monitoring, and more specifically, the present invention relates to a multi-dimensional carbon flux monitoring method. Background Art

[0002] As an important parameter for measuring the transfer and exchange of greenhouse gases such as carbon dioxide between the earth's surface and the atmosphere, carbon flux is a key indicator in carbon cycle research and carbon emission supervision. Through high-precision and dynamic monitoring of carbon flux, it can provide a scientific basis for carbon emission accounting, carbon sink assessment, and carbon market trading. However, there are many deficiencies in existing carbon flux monitoring technologies, which are mainly reflected in the following aspects.

[0003] Most existing carbon flux monitoring methods rely on a single observation method or data source, usually using a single data type such as fixed ground observation stations, remote sensing image data, or unmanned aerial vehicle measurement data. This single data mode results in a limited spatial coverage range, inconsistent spatial and temporal resolutions of data, and it is difficult to meet the requirements of large-scale, refined, and continuous carbon flux monitoring. Secondly, there are significant differences in spatial resolution, time synchronization, data format, and accuracy standards among multi-source heterogeneous data. Due to different observation sampling frequencies, inconsistent time synchronization mechanisms, and diverse data acquisition means, there is a lack of a unified spatio-temporal benchmark and structural standard among the data, which directly affects the effectiveness of data fusion. Existing technologies often lack dynamic fusion algorithms and real-time weight adjustment mechanisms for multi-source data, and it is difficult to solve problems such as data redundancy conflicts, cumulative time deviation, and inconsistent observed values, further resulting in insufficient quality of the input data of the carbon flux inversion model and affecting the accuracy and reliability of the model prediction results.

[0004] Existing carbon flux inversion models usually adopt traditional statistical analysis methods or linear regression models, and cannot effectively mine the non-linear characteristics in complex observed data. Especially in the case of strong spatial heterogeneity, large fluctuations in data quality, and high observation anomaly rates, the model prediction accuracy is insufficient, and it is difficult to achieve precise monitoring and early warning of the dynamic changes of carbon emissions and carbon sinks.

[0005] In summary, existing carbon flux monitoring methods generally have technical bottlenecks such as insufficient multi-source data fusion ability, unstable observed data quality, and limited prediction accuracy of the inversion model. In order to overcome the above problems and meet the requirements of large-scale, high-precision, and high-timeliness carbon flux monitoring, there is an urgent need for a comprehensive monitoring method that can fuse multi-source heterogeneous observation data, dynamically adjust data weights, and improve the accuracy of the carbon flux inversion model. The present invention is precisely proposed under this background, aiming to solve the problems of ineffective multi-source data fusion and poor data consistency in existing technologies, resulting in insufficient accuracy of the carbon flux inversion model. Summary of the Invention

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-dimensional carbon flux monitoring method, comprising the following steps:

[0008] Collect multi-source heterogeneous sensor data and preprocess the multi-source heterogeneous sensor data, including: unified processing of spatial resolution, synchronous alignment of sampling time, and standardized conversion of data format, to ensure that data from different sources has a unified spatio-temporal basis and structural consistency;

[0009] Perform data fusion effectiveness detection, detect whether there are abnormal time synchronization deviations and data redundancy conflict abnormalities, and perform feedback correction;

[0010] When there is abnormal sensor data, dynamically adjust the fusion weight parameters of each data source. When there is no abnormal sensor data, obtain the fusion weight parameters of each data source this time based on the average level of the historical clustering similar groups, and through spatial adjacency interpolation and time interpolation, achieve the unified fusion of different data sources in terms of spatial resolution and time granularity, and obtain a multi-dimensional carbon flux fusion data set;

[0011] Input the fusion data set into a pre-trained carbon flux inversion model, and based on the prediction system of the carbon flux inversion model, output the carbon flux monitoring value.

[0012] In a preferred embodiment, the time synchronization deviation detection refers to: by comparing the difference between the acquisition timestamps of each data source and the reference benchmark time, if the difference is greater than the set time threshold, it is judged as an abnormal synchronization deviation;

[0013] The data redundancy conflict detection refers to: by comparing the difference degree of the observed data values at the same spatial position and time point of different data sources, if the difference degree exceeds the set difference threshold, it is judged as a redundant conflict abnormality.

[0014] In a preferred embodiment, in the time synchronization deviation detection step, the reference benchmark time adopts the synchronous time, which is synchronized by the built-in internal clock through the network time protocol, and the sampling time is marked in real time when receiving data, ensuring the accuracy of the time alignment processing of all data sources, dynamically monitoring the deviation change of the sampling time difference of all sensors, and when the sampling frequency fluctuates, automatically triggering the sampling time threshold correction mechanism to correct the set time threshold.

[0015] In a preferred embodiment, in the redundant conflict detection step, a dynamic spatial consistency evaluation mechanism is adopted. When evaluating the spatial consistency of the current data source, according to the dispersion of the data space distribution and the correlation between adjacent data nodes, the redundancy rate between different observation points is calculated using the spatial consistency function. By setting a dynamic difference threshold, early warning of redundant conflict anomalies for different time windows is realized, ensuring the data fusion quality in an environment of high concurrent sampling rate of heterogeneous data.

[0016] In a preferred embodiment, during feedback correction, invalid data is screened by the rule that the data credibility score is lower than the set credibility threshold. If the remaining valid data after screening does not meet the usage standard, supplementary sampling is performed.

[0017] In a preferred embodiment, when dynamically adjusting the fusion weight parameters of each data source, an initial weight value is obtained, and then the synchronization deviation influence factor of the current data source i is calculated : ; represents the deviation between the acquisition time of the current data source i and the reference time, is the maximum allowable time deviation threshold;

[0018] Calculate the redundant conflict influence factor : ; represents the average difference degree between the data value of the current data source i and the data of other data sources at the same spatial and temporal point, is the maximum allowable difference degree threshold;

[0019] The final fusion weight calculation formula is: ; represents the initial weight value of the current data source i, represents the fusion weight of the current data source i.

[0020] In a preferred embodiment, the initial weight setting process combines three indicators: the type of acquisition device, the data acquisition environment, and historical stability statistical data, constructs an initial weight multi-dimensional evaluation model, and generates a basic weight value through weighted average to provide an initial benchmark for dynamic weight adjustment.

[0021] In a preferred embodiment, the spatial adjacency interpolation algorithm is the inverse distance weighted interpolation or the Kriging interpolation algorithm. The spatial region is divided into multiple sub-regions. If the comprehensive weight mean value of the sub-region is higher than the set fusion threshold, then both the Kriging interpolation and the inverse distance weighted interpolation are jointly used within the sub-region. If the comprehensive weight mean value of the sub-region is not higher than the set fusion threshold, and the synchronization deviation influence factors of all data sources are greater than , and the redundant conflict influence factors of all data sources are greater than , the Kriging interpolation algorithm is adopted. If the average comprehensive weight of the sub-region is not higher than the set fusion threshold, and the synchronization deviation influence factor of the data source is less than or equal to or the redundancy conflict influence factor of the data source is less than or equal to , the inverse distance weighted interpolation algorithm is adopted.

[0022] In a preferred embodiment, during time interpolation, time alignment is performed, and the missing time points are predicted and supplemented through the constructed time series trend fitting model.

[0023] In a preferred embodiment, during supplementary sampling, the sampling density and the anomaly rate of the data source are obtained, and the adjustment coefficient is calculated:

[0024] ; the adjusted sampling frequency is , and the calculation formula is: ; is the preset basic sampling frequency.

[0025] The technical effects and advantages of the present invention:

[0026] By collecting and fusing multi-source heterogeneous carbon flux observation data, the present invention establishes a unified spatio-temporal reference and data structure standard, effectively solving the problem of difficult data fusion in the prior art due to diverse data sources, asynchronous time synchronization, and inconsistent spatial resolutions. By introducing preprocessing steps such as unified processing of spatial resolution, synchronous alignment of sampling time, and standardized conversion of data formats, the observation data from different sources can be seamlessly docked, ensuring the basic consistency of data fusion processing. During the data collection and processing stages, time synchronization deviation anomalies and data redundancy conflict anomalies are detected in real time, ensuring that the data entering the data fusion and carbon flux inversion models has high credibility and consistency, and significantly improving the standardization of carbon flux data collection and the reliability of data quality.

[0027] The present invention adopts a dynamic fusion weight adjustment mechanism. By real-time evaluating the time synchronization deviation influence factor and redundancy conflict influence factor of each data source, it dynamically optimizes the fusion weights of each data source to achieve the adaptive fusion of multi-source observation data. When there is no abnormality, based on the average level of the historical clustering similar groups and the distance between the current data state and the historical data state, the dynamic weighted calculation of the fusion weights is carried out to further enhance the rationality and intelligence level of weight allocation. Through the combined application of spatial adjacency interpolation and temporal interpolation, the missing points of carbon flux observation data in the spatial and temporal dimensions are fully supplemented, improving the spatial integrity and temporal continuity of carbon flux data, solving the problems of discontinuous data and insufficient data coverage in the prior art, and ensuring the formation of a multi-dimensional, full-coverage, and high-resolution carbon flux fusion dataset.

[0028] The present invention inputs the fused high-quality dataset into a pre-trained carbon flux inversion model. The model is designed with a convolutional neural network structure, which automatically extracts the spatial correlation and temporal dynamic features between different observation factors, effectively improving the prediction accuracy and adaptability of the carbon flux inversion model. This model can not only accurately predict the carbon emission and carbon sink information in each spatial region and different time periods, but also support the analysis and early warning of the spatio-temporal dynamic change trend of carbon flux. The finally output carbon flux monitoring value provides a scientific data basis for carbon emission verification, carbon sink assessment, and ecological environment management. The present invention significantly improves the intelligence and automation level of carbon flux monitoring, strengthens the dynamic monitoring ability of carbon emissions and carbon sinks, and helps to achieve national strategic goals such as carbon peak and carbon neutrality, having important economic and environmental benefits. Brief Description of the Drawings

[0029] For the convenience of those skilled in the art to understand, the present invention will be further described below in conjunction with the accompanying drawings;

[0030] Figure 1 It is a schematic diagram of a multi-dimensional carbon flux monitoring method in the present invention. Detailed Embodiments

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] Refer to Figure 1 The following embodiments are obtained:

[0033] Example 1: The present invention is proposed under the background of realizing the strategic goals of carbon peak and carbon neutrality and strengthening the management of ecological environment protection. At present, with the increasingly prominent problem of global climate change, the ecological environment problems caused by greenhouse gas emissions pose a severe challenge to the sustainable development of mankind. As an important parameter for measuring the intensity of the transfer and exchange of greenhouse gases such as carbon dioxide between the surface ecology and the atmosphere, carbon flux has become a core indicator in climate change research, carbon emission management, and ecological service function assessment. High-precision and dynamic carbon flux monitoring is an important basis for realizing scientific carbon management and promoting carbon emission reduction work. However, the existing carbon flux monitoring systems mostly rely on single data sources or static data processing methods, often suffering from problems such as insufficient data coverage, inconsistent time and space resolutions, and difficulty in fusing heterogeneous data, seriously restricting the accuracy and timeliness of carbon flux monitoring data and making it difficult to meet the needs of refined carbon emission accounting and scientific decision-making.

[0034] Under this background, the present invention proposes a multi-dimensional carbon flux monitoring method. By fusing multi-source heterogeneous data from remote sensing satellites, ground observation stations, unmanned aerial vehicle platforms, Internet of Things sensors, etc., it breaks through the limitations of single observation means and realizes the unification of the spatial scale and time granularity of data from different sources. By introducing a dynamic weight adjustment mechanism, according to the real-time credibility and data quality assessment results of each data source, the weight distribution of each data source in the data fusion process is dynamically optimized, effectively improving the robustness of data fusion and the accuracy of the results. At the same time, algorithms such as spatial interpolation and temporal interpolation are used to repair and supplement the spatial gaps and temporal breaks in the observed data, forming a multi-dimensional carbon flux fusion dataset with high resolution and full coverage. Furthermore, the fusion dataset is input into the carbon flux inversion model, and artificial intelligence technology is combined to improve the timeliness and accuracy of carbon flux prediction and estimation.

[0035] The present invention not only solves the problem of difficult effective fusion of multi-source heterogeneous carbon flux data, improves the spatial integrity and temporal continuity of carbon flux monitoring, but also enhances the scientific nature and real-time nature of dynamic analysis of carbon emissions and carbon sinks, and can provide scientific, accurate, and comprehensive data support for ecological environment management, carbon emission supervision, carbon market trading, and the realization of carbon peak and carbon neutrality goals, with significant social benefits and application prospects.

[0036] A multi-dimensional carbon flux monitoring method, comprising the following steps:

[0037] Collect multi-source heterogeneous sensor data and preprocess the multi-source heterogeneous sensor data, including: unified processing of spatial resolution, synchronization and alignment of sampling time, and standardization conversion of data formats, to ensure that data from different sources have a unified spatio-temporal basis and structural consistency; collecting multi-source heterogeneous sensor data and preprocessing the multi-source heterogeneous sensor data, the significance of this step is to solve the problems of inconsistent spatial resolution, different time sampling granularities, and non-uniform data format standards caused by hardware differences and different acquisition mechanisms during the data acquisition stage of different types of sensor devices. Through the unified processing of spatial resolution, it can be ensured that the spatial ranges and granularities covered by different data sources are standardized, enabling data from different sources to be compared and analyzed at the same spatial scale. Sampling time synchronization and alignment are to solve the problem of inconsistent time axes caused by sensors collecting data at different time nodes, and to ensure that all data are compared and fused under the same time reference system through time alignment. The standardization conversion of data formats unifies the encoding methods, data structures, and data content description methods of various data files or data streams, providing a standardized and consistent data basis for subsequent data fusion and model training. This series of preprocessing operations ensure that data from different sources have a unified spatio-temporal basis and structural consistency, laying a solid foundation for subsequent data fusion and analysis.

[0038] Perform data fusion validity detection to detect whether there are abnormal time synchronization offsets and data redundancy conflict anomalies, and provide feedback for correction; the significance of performing data fusion validity detection is to ensure the quality and reliability of the input data during the data fusion process. By detecting whether there are abnormal time synchronization offsets, it is possible to timely discover data time deviations caused by device time drift, clock asynchronization, etc., thereby avoiding interference of data in the time dimension on the fusion result. Detecting data redundancy conflict anomalies is to identify the problem of observation value conflicts existing at the same spatial location and time point of different data sources. By analyzing the differences in observation values between data sources, abnormal deviations caused by redundant data can be timely discovered. This detection mechanism effectively avoids errors caused by inconsistencies between data sources and abnormal data by timely identifying the abnormal state of data and providing feedback for correction, ensuring the accuracy and stability of the subsequent data fusion processing results.

[0039] When there is abnormal sensor data, the fusion weight parameters of each data source are dynamically adjusted. When there is no abnormal sensor data, the fusion weight parameters of each data source for this time are obtained based on the average level of the historical clustering similar groups. Through spatial adjacency interpolation and temporal interpolation, the unified fusion of different data sources in terms of spatial resolution and temporal granularity is achieved, and a multi-dimensional carbon flux fusion dataset is obtained. When abnormal sensor data is detected, it is of great significance to dynamically adjust the fusion weight parameters of each data source. By dynamically adjusting the fusion weights, the credibility of the data can be dynamically and optimally allocated according to the current state of the data source. For data sources with large synchronous deviations or serious redundancy conflicts, their weights in the fusion process are reduced to minimize the impact of abnormal data on the overall fusion result. In the case of no abnormalities, the weights are allocated for this data fusion based on the average level of the similar groups analyzed by historical clustering, making the fusion result more robust and reasonable. Through the dynamic adjustment mechanism, the fairness and reliability of different data sources participating in data fusion can be maximally improved, ensuring that the fusion result objectively reflects the true state of the spatio-temporal distribution of carbon fluxes.

[0040] Achieving the unified fusion of different data sources in terms of spatial resolution and temporal granularity through spatial adjacency interpolation and temporal interpolation is a key link in the data processing flow of the present invention. The significance of spatial adjacency interpolation lies in solving the problem of missing spatial data caused by uneven distribution of data sampling points or limited observation range. By means of interpolation technology, the data in the blank areas are supplemented to improve the integrity of spatial data. The significance of temporal interpolation lies in solving problems such as different sensor sampling frequencies, discontinuous time series, or data loss. Through temporal completion means, the continuity and smoothness of time data are ensured. The combined effect of these two interpolation methods realizes the seamless fusion of data from different sources and in different formats under the unified spatial and temporal dimensions, generating a multi-dimensional carbon flux fusion dataset, providing a comprehensive, fine-grained, and high-precision data basis for subsequent modeling and analysis.

[0041] Input the fused dataset into the pre-trained carbon flux inversion model, and based on the prediction system of the carbon flux inversion model, output the carbon flux monitoring values. Inputting the fused dataset into the pre-trained carbon flux inversion model and outputting the carbon flux monitoring values based on the prediction system of the carbon flux inversion model is significant for realizing the automatic inversion and prediction of the spatio-temporal dynamic information of carbon flux. By inputting the high-quality dataset that has been fused and processed, the inversion model can learn and extract the spatio-temporal evolution characteristics in the data, and identify the variation laws of carbon flux in different spatial regions and time stages based on the deep learning model. The prediction system combines historical data with real-time data through the computing power of the model to dynamically generate the carbon flux monitoring values. These monitoring values not only provide detailed and refined carbon flux distribution information, but also provide a scientific basis for carbon emission accounting, carbon sink assessment, carbon trading market analysis, and the formulation of the carbon peak and carbon neutrality strategies, improving the intelligent, precise, and efficient level of environmental governance.

[0042] When there is no abnormal sensor data, the fusion weight parameters of each data source for this time are obtained based on the average level of the historical clustering similar groups, specifically:

[0043] When there is no abnormal sensor data, the fusion weight parameters of each data source for this time are determined based on the average level of the historical clustering similar groups. The core purpose is to extract the performance laws of the data sources in different operating states through the analysis and classification of historical observation data, and then guide the dynamic assignment of weights in the current data fusion process to improve the reliability and stability of the fused dataset.

[0044] Specifically, first, use the long-term accumulated historical observation data to conduct clustering analysis on all data sources. The basis for clustering includes but is not limited to key indicators such as the type of data acquisition equipment, historical sampling stability, sampling environmental conditions, time synchronization accuracy of historical data, redundancy conflict frequency, and data anomaly rate. Adopt a multi-dimensional clustering algorithm to divide the data sources with similar characteristics and operating states into different similar groups. The data sources in each group have similar data quality characteristics and operating performances in the previous observation periods, and can represent the conventional behavior patterns of a certain type of data source under specific environmental or application conditions.

[0045] After clustering is completed, statistical analysis is performed on the historical weight allocation within each similar group. The specific approach is to perform weighted aggregation and average calculation on the fusion weights of all data sources within the group over different past observation periods to obtain the average weight level of the similar group. When the current data source enters the fusion weight calculation, the state characteristics of the current sampled data are first extracted to form the current feature vector. Subsequently, by calculating the similarity distance between the current feature vector and all historical state samples within the similar group, the similarity between the current data source and the historical samples is measured. The distance here is usually based on the weighted Euclidean distance of each dimension in the multi-dimensional feature space or other distance metrics such as the Euclidean distance that can be selected in the present invention. The smaller the distance value, the closer the state of the current data source is to a certain historical sample within the similar group, and the more representative it is of the typical performance of the group.

[0046] Based on the magnitude of the distance value, different weighting coefficients are assigned to each historical sample within the similar group. The closer the sample is, the larger its weighting coefficient, indicating that its influence on the current weight allocation is stronger; the farther the sample is, the smaller its weighting coefficient, and the corresponding influence on the current weight is weakened. For example, by dividing the number one by the sum of the distance value and a constant, the weighting coefficient is obtained. This constant is usually taken as a minimum value to avoid the situation of a zero denominator. By performing weighted average calculation on the fusion weight values of all samples within the similar group, the fusion weight parameter of the current data source is finally obtained. This process not only considers the overall level of the group's historical weights but also dynamically reflects the similarity between the current data source state and the historical samples, making the weight allocation more refined and dynamic.

[0047] This average level reflects the comprehensive weight obtained by the data sources of this group based on the evaluation of actual data quality and stability during the historical fusion process. It is a weight value verified by historical practice and has a certain degree of stability and representativeness.

[0048] In the current data fusion task, if it is detected that no synchronization deviation anomalies and redundant conflict anomalies occur in all data sources, it indicates that the current sampling state is normal and the data quality is stable. At this time, the average weight level of the corresponding similar group will be directly called as the basic fusion weight parameter for each data source in this instance. Each data source automatically inherits the average weight of the group it belongs to according to its historical clustering group, avoiding the problem of unreasonable weight allocation caused by insufficient characteristics of single-sampling data or real-time data analysis delay.

[0049] By determining the fusion weight parameter based on the average level of the historical clustering similar group, not only can the complexity of real-time calculation be reduced, but also the historical experience data can be fully utilized to improve the scientificity and rationality of weight allocation, ensuring the robustness of the multi-source data fusion process and the credibility of the carbon flux monitoring results.

[0050] Time synchronization deviation detection refers to: by comparing the difference between the acquisition timestamps of each data source and the reference benchmark time, if the average time difference is greater than the set time threshold, it is judged that the synchronization deviation is abnormal;

[0051] Data redundancy conflict detection refers to: by comparing the degree of difference in the observed data values at the same spatial position and time point of different data sources, if the degree of difference exceeds the set difference threshold, it is judged that the redundancy conflict is abnormal.

[0052] In the time synchronization deviation detection step, the reference benchmark time uses the synchronization time, which is synchronized by the built-in internal clock through the Network Time Protocol. When receiving data, the sampling time is marked in real time to ensure the accuracy of time alignment processing for all data sources. Dynamically monitor the deviation change of the sampling time differences of all sensors. When the sampling frequency fluctuates, automatically trigger the sampling time threshold correction mechanism to correct the set time threshold. The specific steps are as follows:

[0053] Step 1: Establishment of reference time synchronization. Each acquisition end node (data source) i integrates a high-precision internal clock, and the clock signal source is the local crystal oscillator. Perform standard time synchronization through the Network Time Protocol NTP. The global reference benchmark time is denoted as , and at fixed synchronization intervals , each device calibrates the clock deviation, and the clock deviation is: ; Calibration formula (device local time correction): ; is the time after calibration of data source i, and before calibration is , ensuring the accuracy of basic time synchronization.

[0054] Step 2: Record the sampling time of the sensor. At the moment when each sensor collects data, based on the local time , record the single sampling time : ; i represents the i-th sensor, i.e., the data source, and n represents the n-th sampling, forming a sampling time series: ; represents the total amount of data in the current sampling period of the i-th data source.

[0055] Step 3: Calculate the sampling time deviation. Calculate the deviation amount between each sampling moment and the reference benchmark time : ; represents the time of the global benchmark time at the n-th sampling moment. Calculate the average time deviation (mean value within the sliding window): ; W is the length of the sliding window, defining the time range for sampling detection.

[0056] Step 4: Sampling frequency volatility detection, calculate the time interval between every two sampling points n and n-1 in real time : ; The instantaneous sampling frequency is: ; The sliding window average sampling frequency is: ; The volatility is: ; The volatility represents the relative fluctuation degree of the sampling frequency, and

[0057] is the standard deviation of the sampling frequency.Step 5: Dynamic correction of the time synchronization threshold, the default initial time threshold is : ; is the preset sensitivity coefficient, which adjusts the influence weight of frequency fluctuation on threshold adjustment. The finally dynamically adjusted time threshold is: ; is the automatically adjusted allowable time synchronization deviation range, that is, the time threshold. If ; then trigger the time synchronization anomaly alarm and start the automatic compensation mechanism at the same time.

[0058] Considering the correction of the internal clock frequency deviation of the device, a compensation method with a clock drift rate is adopted:

[0059] The device clock deviation is modeled as: ; t is the time variable, which is the length of time elapsed since the last time synchronization or the reference time setting. is the reading of the local clock of device i at time t, is the time of the reference reference time at time t, is the initial time deviation (absolute error) of device i, is the clock deviation rate of device i, with the unit of ppm (one in a million). The clock deviation rate is estimated in real time as : ; is the sliding window time span, and the corrected local time is: .

[0060] In the redundant conflict detection step, a dynamic spatial consistency evaluation mechanism is adopted. When evaluating the spatial consistency of the current data source, according to the dispersion of the data space distribution and the correlation between adjacent data nodes, a spatial consistency function is used to calculate the redundancy rate between different observation points. By setting a dynamic difference threshold, early warning of redundant conflict anomalies for different time windows is realized, ensuring the data fusion quality in an environment of high sampling rate of heterogeneous data concurrency.

[0061] Redundant conflict detection aims to discover whether the observed values of multi-source heterogeneous data in the same spatial cell (or geographical location) and at the same time point conflict. Specifically, through a dynamic spatial consistency evaluation mechanism, the consistency of the observed results of different data sources is measured, specifically as follows:

[0062] Step 1: Spatial unit division. The monitoring area is divided into several spatial cells G(x, y). Each cell is a rectangular grid or an irregular polygon, numbered , and there are multiple observed values of data source i at time t in each cell, which are , where i is the data source number, k is the cell number, and t is the sampling time point.

[0063] Step 2: Calculation of spatial observation value difference degree. For each cell , the mean value and the spatial observation value difference degree (dispersion), that is, the standard deviation , are calculated respectively according to the set of observed values, which reflects the spatial consistency of the observed values of different data sources. The greater the dispersion, the more serious the inconsistency.

[0064] Step 3: Calculation of spatial correlation coefficient. Calculate the spatial correlation between cell and its adjacent cell : ; is the covariance, is the standard deviation of data source i in cell , is the standard deviation of data source j in the adjacent cell .

[0065] Step 4: Construction of spatial consistency function. Define the spatial consistency coefficient of cell :

[0066] ; M is the number of adjacent cells, and the adjacent range is defined by distance or topological relationship. This coefficient reflects the similarity of the spatial observed values between cell and its adjacent cells. Tending to 1 indicates strong correlation (good spatial consistency), and tending to 0 indicates weak correlation (poor spatial consistency).

[0067] Step Five: Redundancy Rate Calculation and Dynamic Difference Threshold Judgment, Redundancy Rate Calculation: ; Basic Threshold Setting: ; is the mean value of historical redundancy rate, is the standard deviation of historical redundancy rate, is the preset sensitivity adjustment parameter, obtain the concurrent sampling rate of the current data source , and then calculate: ; μ is the preset concurrent sensitivity coefficient, is the average sampling frequency, is the dynamic sensitivity adjustment parameter, the final dynamic difference threshold is: If ; then the cell is determined as a redundant conflict abnormal area at time t.

[0068] During feedback correction, invalid data is screened through the rule that the data credibility score is lower than the set credibility threshold. If the remaining valid data after screening does not meet the usage standard, supplementary sampling is performed. Through time synchronization deviation detection, the sampling time of each data source is calibrated. According to the deviation degree between the data source collection time and the reference benchmark time, the stability of time synchronization is evaluated. The smaller the deviation, the more accurate and reliable the data source sampling time is, and the higher the time synchronization deviation impact factor score; the larger the deviation, the more likely there are clock drift or sampling delay problems in the data source, and the corresponding time synchronization deviation impact factor score is reduced.

[0069] Through redundant conflict detection, the observed values from different data sources are compared and analyzed. If the observed value of a certain data source is highly consistent with the observed values of other data sources at the same spatial position and the same time point, it indicates that the data validity of this data source is strong, and the redundant conflict impact factor score is high. On the contrary, if there are significant differences in the observed values, exceeding the preset reasonable threshold range, it is considered that there are abnormalities or conflict risks in the observed data of this data source, and the redundant conflict impact factor score is reduced accordingly. After calculating the time synchronization deviation impact factor and the redundant conflict impact factor respectively, the final data credibility score of this data source is comprehensively formed by means of weighted average or weighted product. The range of the score value is usually set between zero and one. The higher the score, the better the data quality of this data source and the more trustworthy it is; the lower the score, the more potential abnormalities or errors exist in this data source, the priority of participating in data fusion is reduced, and the credibility threshold is a preset judgment standard for screening valid data and invalid data.

[0070] The threshold value is determined based on the severity of the data reliability requirements in different application scenarios and the analysis results of historical data accumulated during long-term operation. Usually, the credibility threshold is expressed in percentage or as a decimal between zero and one, and it is a dynamically adjustable parameter. When the credibility score of a certain data source is higher than or equal to the credibility threshold, the data is determined to be valid data and is allowed to participate in subsequent data fusion and the training or prediction of the carbon flux inversion model; while when the credibility score of a certain data source is lower than the threshold, it is automatically identified as invalid data or low-quality data and excluded from the data fusion calculation to ensure the overall data quality and the accuracy of the final carbon flux monitoring results.

[0071] The determination method of the credibility threshold can be analyzed based on the statistical mean and standard deviation of the historical data distribution and dynamically adjusted in combination with the current operating state. Specifically, it can be determined by the expert assignment method. For example, when the abnormal rate of the sampling data source increases or in the high-concurrency operating state, the credibility threshold can be appropriately increased to ensure that the final data quality meets the standard; while in the case of a good data acquisition environment and high consistency of the sampled data, the credibility threshold can be appropriately reduced according to the requirements to accommodate the observation data of more data sources to participate in the fusion and improve the overall data coverage and the comprehensiveness of carbon flux monitoring. Supplementary sampling refers to the situation where the relevant indicators of the valid data remaining after screening, such as the coverage rate and the total data volume, do not meet the usage standards, and supplementary sampling is carried out.

[0072] When dynamically adjusting the fusion weight parameters of each data source, obtain the initial weight value, and then calculate the synchronization deviation influence factor of the current data source i : ; represents the deviation between the acquisition time of the current data source i and the reference time, is the maximum allowable time deviation threshold;

[0073] Calculate the redundant conflict influence factor : ; represents the average difference degree between the data value of the current data source i and the data of other data sources at the same spatial and temporal point, is the maximum allowable difference degree threshold;

[0074] The final fusion weight calculation formula is: ; represents the initial weight value of the current data source i, represents the fusion weight of the current data source i.

[0075] When dynamically adjusting the fusion weight parameters of each data source, it is first necessary to determine an initial weight value for each data source. On this basis, the time synchronization deviation impact factor of this data source is further calculated. The time synchronization deviation impact factor is used to measure the deviation degree between the current sampling time of this data source and the global reference time. If the deviation between the sampling time and the reference time is small, it indicates that the time synchronization performance of this data source is good, and its time synchronization deviation impact factor score is high. Conversely, if the deviation approaches or exceeds the set maximum allowable time deviation threshold, the score of this impact factor will be correspondingly reduced. This factor directly reflects the impact weight of the data source time consistency on the fusion process.

[0076] At the same time, the redundancy conflict impact factor of this data source is calculated. This impact factor is used to measure the difference degree between the observed data value of this data source and the observed values of other data sources at the same spatial position and the same time point. If the consistency between the data source observation value and the observation values of other data sources is good and the difference is small, it indicates that the data validity of this data source is high, and the redundancy conflict impact factor score is high. Conversely, if the observation value difference is significant, approaching or exceeding the set maximum allowable difference threshold, the score of this impact factor will be reduced. This factor reflects the credible impact of the data source data content on the fusion result in terms of spatial consistency.

[0077] Finally, the initial weight value of this data source is comprehensively calculated with its corresponding time synchronization deviation impact factor and redundancy conflict impact factor to obtain the dynamic fusion weight value of this data source at the current moment. The comprehensive calculation method can adopt the weighted product or weighted average method, taking the initial weight value as the basis, and dynamically adjusting the participation weight of this data source in the multi-source heterogeneous data fusion process based on the impact factors in terms of time synchronization and data consistency. The higher the weight value, the higher the data quality of this data source and the stronger the fusion credibility, and it occupies a higher weight proportion in the final multi-dimensional carbon flux fusion dataset; the lower the weight value, the corresponding reduction in the impact of the data source data on the fusion result, and it may even be identified as invalid data and excluded. Through this dynamic weight adjustment mechanism, the unified fusion of different data sources in terms of spatial resolution and time granularity is achieved, improving the overall quality and fusion accuracy of carbon flux data.

[0078] The initial weight setting process combines three indicators: the type of acquisition device, the data acquisition environment, and historical stability statistical data, constructs an initial weight multi-dimensional evaluation model, and generates a basic weight value through the weighted average method, providing an initial benchmark for dynamic weight adjustment.

[0079] In the process of setting the initial weights of each data source, an evaluation model based on multi-dimensional indicators is adopted to comprehensively analyze and score each data source to generate its basic weight value. This model mainly combines three core indicators: the type of acquisition device, the data acquisition environment, and historical stability statistical data, ensuring the rationality and scientific nature of the initial weights and providing a reliable benchmark basis for subsequent dynamic weight adjustment.

[0080] As one of the important indicators for initial weight evaluation, the type of acquisition device refers to the accuracy level, stability, and applicability of different sensors or data source hardware itself. According to the accuracy level, technical maturity, manufacturing standards, and actual deployment of the sensors, the device types are classified and basic scores are assigned. Generally, sensors with high precision, advanced technology, and designed specifically for carbon flux monitoring will have a higher weight score than general-purpose, non-professional, or devices with long-term maintenance problems.

[0081] As the second evaluation indicator, the data acquisition environment mainly evaluates the geographical area, climate conditions, and environmental interference factors where the acquisition device is located. Acquisition devices with stable environments, less interference, and excellent sampling conditions will score higher on this indicator; while devices located in complex terrains, highly polluted, or extreme climate conditions will be given a relatively lower score because the data is greatly affected by the environment, resulting in a decrease in the stability and reliability of the observed data.

[0082] Historical stability statistical data is the most dynamic and objective evaluation indicator in the initial weight setting. Through long-term statistical analysis of the historical observation records of the data source, multiple indicators such as sampling integrity, data anomaly rate, time synchronization deviation, and fluctuation range of observed values within the historical period are evaluated to reflect the long-term stability and consistency of the data source. The higher the quality of historical data, the better the data continuity, and the lower the anomaly rate, the higher the score for this item.

[0083] After comprehensively considering the above three-dimensional indicators, the initial basic weight value of each data source is calculated using the weighted average method. During the weighted average process, different weight coefficients are set for each evaluation indicator according to its importance to the overall weight. Generally, historical stability statistical data accounts for a relatively large proportion because it directly reflects the long-term operation performance of the data source; followed by the type of acquisition device, and then the data acquisition environment. The proportion of the weight coefficients of each indicator can be adjusted according to actual application requirements and business logic.

[0084] In the specific weighting process, first, the scoring values of each indicator are normalized to ensure comparability between different indicators. After normalization, the score values of their respective indicators are multiplied by the preset weight coefficients, and finally, the results after weighting all indicators are summed to obtain the initial basic weight value of the data source. It should be noted that: this weight value is not directly calculated based on the total amount of valid data of the data source, but is based on a comprehensive evaluation of the device performance, environmental adaptability, and historical data quality of the data source, reflecting the credibility and importance of the data source in the initial stage of data fusion. The finally obtained basic weight value will be used as one of the input parameters of the dynamic weight adjustment algorithm to participate in the real-time calculation and optimization of the dynamic fusion weight, ensuring the scientificity and accuracy of multi-source data fusion.

[0085] During time interpolation, time alignment is performed, and through the constructed time series trend fitting model, missing time points are predicted and supplemented. Time series trend fitting model: ; These model parameters are obtained by fitting the least squares method based on historical data. m represents the total number of fitting coefficients, t represents the time point, and t minus one is equal to the value of m. Substituting the time point t corresponding to the missing data into the time series trend fitting model to obtain the data value corresponding to the time point t , thus completing the supplementation of the time data series.

[0086] The spatial adjacency interpolation algorithm is the inverse distance weighted interpolation or the Kriging interpolation algorithm. The spatial region is divided into multiple sub-regions. If the average value of the comprehensive weights of the sub-regions is higher than the set fusion threshold, then the Kriging interpolation and the inverse distance weighted interpolation are jointly used within the sub-region. If the average value of the comprehensive weights of the sub-regions is not higher than the set fusion threshold, and the synchronous deviation influence factors of all data sources are greater than , the redundant conflict influence factors of all data sources are greater than , then the Kriging interpolation algorithm is used. If the average value of the comprehensive weights of the sub-regions is not higher than the set fusion threshold, and the synchronous deviation influence factor of the data source is less than or equal to or the redundant conflict influence factor of the data source is less than or equal to , then the inverse distance weighted interpolation algorithm is used.

[0087] By dynamically selecting the interpolation algorithm in different spatial regions, the problems of inconsistent accuracy and fluctuating data reliability in the spatial fusion process of multi-source heterogeneous data are solved. In carbon flux monitoring, the observed data from different data sources often have synchronous errors, different spatial resolutions, and differences in data acquisition accuracy. If a single spatial interpolation algorithm is simply used, it will cause the interpolation results in some regions to be distorted, reducing the overall data fusion quality.

[0088] First, divide the entire spatial region, decomposing the large region into multiple small regions or sub-regions. For each sub-region, dynamically evaluate the comprehensive data quality situation in this region based on indicators such as the fusion weights of various data sources, the influence factors of time synchronization deviation, and the influence factors of redundancy conflict. If the average value of the comprehensive weights of the data within a region is higher than the set fusion threshold, it indicates that the reliability of the data sources within this region is better, and it is inclined to use an interpolation method with stronger spatial trend modeling ability to maintain high-precision prediction. On the contrary, if the comprehensive weights of the data within this region are low, it indicates that there are data anomalies or unstable factors. To reduce the impact of local abnormal data on the overall result, it is more suitable to use an algorithm with higher interpolation robustness to avoid error diffusion.

[0089] According to the average indicators of the data sources in terms of time synchronization and redundancy conflict, further refine the selection of the interpolation algorithm. By introducing dynamic judgment conditions, ensure that the optimal interpolation strategy is adopted in data regions with different quality levels, improving the overall spatial consistency and prediction accuracy of the fused data set.

[0090] Inverse distance weighted interpolation method: The inverse distance weighted interpolation method is an interpolation method for numerical prediction based on the principle of the inverse ratio of spatial distances. Its core idea is that the closer the known observation points are to the interpolation position in space, the greater the influence of their observed values on the predicted value of the target position, and the higher the weight; the farther the observation points are, the smaller the influence and the lower the weight.

[0091] The specific approach is to calculate the spatial distances between the target interpolation point and the surrounding known observation points. Then, according to the reciprocals of these distances, calculate the weights of each observation point. Usually, the weight values are in a certain power relationship with the inverse ratio of the distances. Next, through the method of weighted average, use the observed values of each observation point and their corresponding weights to calculate the predicted value of the target interpolation point.

[0092] The inverse distance weighted interpolation method has the advantages of simple implementation and high calculation efficiency, and is suitable for interpolation tasks when the data points are relatively densely distributed, the data quality is balanced, and there is no obvious spatial heterogeneity. Since this method mainly relies on spatial distances for weighted calculation, it can effectively reduce the influence of local abnormal observed values on the interpolation result and has strong robustness. However, this method has limited ability to describe spatial trend changes and cannot handle well scenarios with complex spatial structures or obvious heterogeneity.

[0093] Kriging interpolation method: The Kriging interpolation method is an advanced interpolation algorithm based on spatial statistics theory, usually used to handle interpolation problems with complex spatial variability and strong data heterogeneity. Its core idea is to analyze the variation characteristics of spatial data, establish a spatial autocorrelation model, and then calculate the predicted value of the target position according to the spatial covariance relationship between different positions.

[0094] The specific process includes, first, conducting a spatial variability analysis on the known observation point data to construct a variogram model, which is used to describe the law of the correlation change between data at different spatial positions with distance. Then, based on this model, calculate the spatial covariance between the target interpolation point and all known observation points, and further solve the weight values of each observation point through the principle of optimal unbiased estimation. These weight values consider not only the spatial distance factor but also the spatial correlation and variability of the data, ensuring that the interpolation result conforms to both the characteristics of local observations and the overall spatial distribution trend.

[0095] The Kriging interpolation method can effectively mine and utilize the spatial structure information in the data, and is suitable for scenarios with high-quality observation data, obvious spatial distribution patterns, or strong spatial heterogeneity within the region. This method can maximize the accuracy and rationality of the interpolation result, but its algorithm complexity is high, the calculation amount is large, and it has high requirements for the quality of observation data and the number of sample points.

[0096] When the mean value of the comprehensive weights of the data within a certain sub-region is higher than the set fusion threshold, it indicates that the overall quality of the observation data of each data source within this region is high, with good time synchronization and spatial consistency. In order to make full use of the respective advantages of different interpolation algorithms, the Kriging interpolation method and the inverse distance weighted interpolation method are jointly used for interpolation prediction within this sub-region.

[0097] The specific approach is to first use the Kriging interpolation method and the inverse distance weighted interpolation method respectively, and calculate the predicted values of the target interpolation position based on the same set of known observation point data. The Kriging interpolation method can fully reflect the spatial structure and correlation of the data, providing a high-precision prediction result based on spatial variability analysis, while the inverse distance weighted interpolation method enhances the robustness and anti-interference ability of local data through distance weight allocation.

[0098] Subsequently, according to the preset interpolation weight allocation strategy, the predicted values obtained by the two interpolation methods are weighted and fused. The weight ratio of the weighted fusion can find the weight ratio with the best data quality after weighted fusion of all data sources through the genetic algorithm, so as to achieve the optimal combination of the two interpolation results. The finally generated interpolation predicted value not only inherits the accurate characterization ability of the Kriging interpolation method for the spatial structure but also has the adaptability and robustness of the inverse distance weighted interpolation method to local data changes, effectively improving the spatial distribution rationality and prediction accuracy of the interpolation result.

[0099] During supplementary sampling, obtain the sampling density of the data source and the anomaly rate , calculate the adjustment coefficient :

[0100] ; The adjusted sampling frequency is , the calculation formula is: ; is the preset basic sampling frequency.

[0101] During supplementary sampling, the sampling frequency is dynamically adjusted according to the sampling density and anomaly rate of the data source to ensure the balance of data quality and coverage. The specific approach is as follows: First, obtain the sampling density of the current data source, which is the number of observed data points of the data source per unit space or unit time. Then obtain the anomaly rate of the data source, which is the proportion of data marked as invalid or of low quality in its historical data. Based on these two factors, calculate an adjustment coefficient, the magnitude of which is determined by the ratio of the sampling density to the anomaly rate. If the sampling density of the data source is high and the anomaly rate is low, it indicates that the data quality of the data source is stable, the value of the adjustment coefficient approaches the original level, and the change in the supplementary sampling frequency is small. If the anomaly rate of the data source is high, it indicates that there are many errors or uncertainties in the data. To make up for data missing and improve the reliability of the overall data, the adjustment coefficient will be reduced, thereby reducing the dependence on the data source and the increase in the sampling frequency. Finally, through the calculated adjustment coefficient, the basic sampling frequency is adjusted so that the supplementary sampling frequency can dynamically adapt to the actual situation of the data source, while ensuring data coverage, reducing the impact of abnormal data on the final data fusion result.

[0102] Embodiment 2: This embodiment addresses the problem in the prior art that multi-source heterogeneous carbon flux observation data is asynchronous in the time dimension and the cumulative time stamp error leads to data fusion failure, and proposes a method based on dynamic detection and compensation of time synchronization deviation. First, through the built-in time synchronization module, the internal clocks of all acquisition devices are uniformly calibrated by the Network Time Protocol to ensure that the acquisition times of different data sources have a unified reference benchmark. Then, during the data reception process, the sampling time of each data source is recorded in real time, and the deviation between the sampling time and the reference benchmark time is compared. Through sliding window detection and drift trend analysis of the time deviation, if it is found that the sampling time of a certain data source has a persistent deviation or sudden anomaly, an immediate time synchronization anomaly warning is triggered, and the automatic compensation mechanism is started. The compensation mechanism dynamically corrects the sampling time according to the average rate and drift trend of the time deviation, so that the sampled data can be accurately aligned to the standard time axis, solving the problems of time distortion and fusion failure of carbon flux observation data caused by time asynchronization in the prior art.

[0103] Example 3: Aiming at the problems in the prior art that the multi-source heterogeneous carbon flux observation data have low spatial interpolation accuracy and poor data fusion consistency due to different spatial resolutions, sparse spatial data distribution or conflicting observation values, a method for jointly applying a dynamic interpolation algorithm is proposed. First, according to the spatial division results of the carbon flux observation area, the data quality of each sub-area is evaluated, and the mean value of the data source fusion weight, the time synchronization deviation and the redundancy conflict index of each area are calculated. If the data quality in a certain sub-area is high and the observation values are in good agreement, the Kriging interpolation method and the inverse distance weighted interpolation method are jointly used for spatial interpolation. Specifically, the interpolation values are calculated based on the Kriging interpolation method and the inverse distance weighted interpolation method respectively, and then according to the regional spatial heterogeneity index and the time non-stationarity index, the weighted ratio of the two interpolation methods is dynamically allocated to achieve an optimized combination of the interpolation results. Through this implementation method, the interpolation accuracy of the carbon flux observation data in the process of unifying the spatial resolution is effectively improved, the continuity and rationality of the spatial data are ensured, and the problem that the spatial interpolation method in the prior art is single and it is difficult to take into account both the spatial trend and the local fine features is solved.

[0104] Example 4: Aiming at the problems in the prior art that the sampling density of the carbon flux observation data is uneven and the observation abnormality rate is high, resulting in serious data loss and discontinuous time series data, a supplementary sampling method based on dynamic adjustment of the sampling frequency of the data source is proposed. First, the real-time data sampling quality of each data source is detected to obtain the sampling density and the observation abnormality rate information of the current data source. According to the sampling density and the abnormality rate, the adjustment coefficient is calculated to dynamically adjust the sampling frequency. If it is detected that the data credibility score of a certain data source is lower than the preset threshold and the effective data in the spatial or time unit is not enough to meet the fusion calculation requirements, the supplementary sampling mechanism is immediately triggered to increase the data sampling frequency or enable the standby observation node for sampling. At the same time, through the time series trend fitting model, the historical data trend is modeled and analyzed, and the values are supplemented based on trend prediction and spatial adjacency interpolation at the missing time points. Through this implementation method, the problems that the sampling frequency of the carbon flux data in the prior art is inflexible, the proportion of abnormal data is high, and the missing data is difficult to be automatically repaired are solved, and the integrity of the carbon flux data time series and the continuity of the input data of the inversion model are significantly improved.

[0105] In the present invention, after steps such as preprocessing, dynamic weight adjustment, spatial interpolation, and temporal interpolation of the fused dataset, a multi-dimensional carbon flux fused dataset with consistent spatial resolution, unified temporal granularity, and excellent data quality is formed. This dataset has a complete time series structure and spatial grid distribution information, and covers information from multiple heterogeneous data sources such as remote sensing image data, atmospheric observation data, ground sensor data, and meteorological environment data. This dataset not only reflects the spatio-temporal dynamic characteristics of carbon fluxes in different regions and different time periods, but also eliminates the redundant conflicts and temporal synchronization biases between data sources through data fusion processing, improving the overall reliability and effectiveness of the data.

[0106] The above-mentioned fused dataset is input into a pre-trained carbon flux inversion model. This carbon flux inversion model is designed with a convolutional neural network structure. Through multi-layer convolutional operations and feature extraction modules, it automatically analyzes and learns the complex non-linear relationships between different observation factors in the fused dataset. At the input layer of the model, the multi-dimensional observation data in the fused dataset is input into the model in a unified format, forming a multi-dimensional input tensor containing spatial distribution, time series, and observation features. The convolutional neural network extracts the spatial feature information in the input data through layer-by-layer convolutional operations and pooling operations, identifies the spatial change trends of carbon fluxes in different regions, and learns the dynamic evolution law of carbon fluxes in the time dimension through feature stacking or sliding window mechanisms in the time dimension.

[0107] The model adopts a feature fusion mechanism internally, multi-level integrates the high-dimensional features obtained from the convolutional operations of data of different observation factors, and combines activation functions to improve the model's fitting ability for complex non-linear changes. The fully connected layer part of the model forms the prediction result of the regional carbon flux estimation value by weighted combination of the comprehensive feature vectors of different spatial grid cells. The output layer outputs the carbon flux inversion values at different time points and different spatial units as carbon flux monitoring values in a standard format based on the prediction system.

[0108] The finally obtained carbon flux monitoring values are a set of spatio-temporal prediction data covering all spatial grids and time periods of the monitoring area. Each spatial unit and time point corresponds to a carbon flux monitoring value, reflecting the carbon emission or carbon sink intensity in this area at a specific time. These carbon flux monitoring values can not only be used for carbon emission accounting, carbon sink assessment, and ecological environment supervision, but also provide data support for carbon market trading and carbon peak and carbon neutrality strategic decision-making. This carbon flux inversion model effectively improves the prediction accuracy and timeliness of carbon fluxes after multi-source heterogeneous data fusion processing through the convolutional neural network structure, and solves the problems of low accuracy of carbon flux inversion models and inability to handle complex spatial heterogeneity and temporal dynamic characteristics in the existing technologies.

[0109] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0110] It should be understood that in various embodiments of the present application, the magnitudes of the serial numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0111] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present application.

[0112] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described devices and units can refer to the corresponding processes in the foregoing method embodiments and will not be described herein again.

[0113] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present application, and all should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-dimensional carbon flux monitoring method, characterized in that, It includes the following steps: Collect multi-source heterogeneous sensor data and preprocess the multi-source heterogeneous sensor data, including: unified processing of spatial resolution, synchronous alignment of sampling time, and standardized conversion of data format, to ensure that data from different sources have a unified spatio-temporal basis and structural consistency; Execute data fusion effectiveness detection, detect whether there are abnormal time synchronization deviations and data redundancy conflict anomalies, and perform feedback correction; When there are abnormal sensor data, dynamically adjust the fusion weight parameters of each data source. When there is no abnormal sensor data, obtain the fusion weight parameters of each data source this time based on the average level of the historical clustering similar groups, and through spatial adjacency interpolation and time interpolation, achieve the unified fusion of different data sources in terms of spatial resolution and time granularity, and obtain a multi-dimensional carbon flux fusion dataset; Input the fusion dataset into a pre-trained carbon flux inversion model, and based on the prediction system of the carbon flux inversion model, output the carbon flux monitoring value.

2. The multi-dimensional carbon flux monitoring method according to claim 1, wherein Time synchronization deviation detection refers to: by comparing the difference between the acquisition timestamps of each data source and the reference benchmark time, if the difference is greater than the set time threshold, it is judged as an abnormal synchronization deviation; Data redundancy conflict detection refers to: by comparing the difference degree of the observed data values at the same spatial position and time point of different data sources, if the difference degree exceeds the set difference threshold, it is judged as a redundant conflict anomaly.

3. The multi-dimensional carbon flux monitoring method according to claim 2, wherein In the time synchronization deviation detection step, the reference benchmark time adopts the synchronous time, which is synchronized by the built-in internal clock through the Network Time Protocol. The sampling time is marked in real time when receiving data to ensure the accuracy of time alignment processing for all data sources. Dynamically monitor the deviation change of the sampling time difference of all sensors. When the sampling frequency fluctuates, automatically trigger the sampling time threshold correction mechanism to correct the set time threshold.

4. A multi-dimensional carbon flux monitoring method according to claim 3, characterized in that In the redundant conflict detection step, a dynamic spatial consistency evaluation mechanism is adopted. When evaluating the spatial consistency of the current data source, based on the dispersion of the data spatial distribution and the correlation between adjacent data nodes, use the spatial consistency function to calculate the redundancy rate between different observation points, and through setting a dynamic difference threshold, realize the early warning of redundant conflict anomalies in different time windows, and ensure the data fusion quality in the heterogeneous data concurrent high sampling rate environment.

5. A multi-dimensional carbon flux monitoring method according to claim 4, characterized in that, When dynamically adjusting the fusion weight parameters of each data source, obtain the initial weight value, and then calculate the synchronization deviation impact factor of the current data source i : ; represents the deviation between the acquisition time of the current data source i and the reference time, is the maximum allowable time deviation threshold; Calculate the impact factor of redundant conflict : ; represents the average difference degree between the data value of the current data source i and the data of other data sources at the same spatio-temporal point, is the maximum allowable difference degree threshold; The formula for calculating the final fusion weight is as follows: ; represents the initial weight value of the current data source i, represents the fusion weight of the current data source i.

6. The multi-dimensional carbon flux monitoring method according to claim 5, characterized in that During feedback correction, screen out invalid data through the rule that the data credibility score is lower than the set credibility threshold. If the remaining valid data after screening does not meet the usage standard, supplementary sampling is performed.

7. The multi-dimensional carbon flux monitoring method according to claim 6, characterized in that In the initial weight setting process, combine three indicators of the acquisition device type, data acquisition environment, and historical stability statistical data to construct an initial weight multi-dimensional evaluation model, and generate the basic weight value through weighted average to provide an initial benchmark for dynamic weight adjustment.

8. A multi-dimensional carbon flux monitoring method according to claim 7, characterized in that The spatial adjacency interpolation algorithm is the inverse distance weighted interpolation or the Kriging interpolation algorithm. The spatial region is divided into multiple sub-regions. If the average value of the comprehensive weights of the sub-region is higher than the set fusion threshold, then the Kriging interpolation and the inverse distance weighted interpolation are jointly used within the sub-region. If the average value of the comprehensive weights of the sub-region is not higher than the set fusion threshold, and the synchronous deviation influence factors of all data sources are greater than , and the redundant conflict influence factors of all data sources are greater than , then the Kriging interpolation algorithm is used. If the average value of the comprehensive weights of the sub-region is not higher than the set fusion threshold, and the synchronous deviation influence factor of the data source is less than or equal to or the redundant conflict influence factor of the data source is less than or equal to , then the inverse distance weighted interpolation algorithm is used.

9. A multi-dimensional carbon flux monitoring method according to claim 8, characterized in that, During time interpolation, perform time alignment, and through the constructed time series trend fitting model, predict and supplement the missing time points.

10. A multi-dimensional carbon flux monitoring method according to claim 9, characterized in that During supplementary sampling, obtain the sampling density of the data source and the abnormality rate , and calculate the adjustment coefficient : ; The adjusted sampling frequency is , and the calculation formula is: ; is the preset base sampling frequency.

Images