A CO2 flux inversion method and device based on adaptive covariance and multi-model fusion

Abstract

This invention provides a method and apparatus for CO2 flux inversion using adaptive covariance and multi-model fusion, relating to the field of atmospheric environmental monitoring technology. The method includes: acquiring observational data and calculating an initial residual sequence based on an initial prior flux; introducing a variational cost function based on the initial residual sequence to dynamically solve for a spatiotemporal correlation scale matching the current atmospheric diffusion state, and constructing a dynamic observation error covariance matrix; acquiring high-frequency remote sensing data in real time, constructing a dynamic correction factor based on the high-frequency remote sensing data, and using the dynamic correction factor to scale the initial prior flux in real time to generate a dynamic prior flux; and running two atmospheric transport models in parallel to construct an augmented observation vector and an augmented error covariance matrix. This invention improves the physical consistency, accuracy, and robustness of the inversion results and enhances the ability to capture dynamic emission events.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment monitoring technology, and in particular to a method and apparatus for CO2 flux inversion using adaptive covariance and multi-model fusion. Background Technology

[0002] Against the backdrop of global greenhouse gas emission reduction, accurately quantifying the distribution of surface CO2 sources and sinks (i.e., CO2 flux) is a core scientific requirement. Current mainstream technologies rely on the inverse problem framework of atmospheric transport, that is, using observational information of atmospheric CO2 and combining it with atmospheric transport models to infer the flux intensity at the surface. These methods are usually based on a Bayesian inversion framework, but in practical applications, several structural assumptions limit their inversion accuracy.

[0003] First, existing technologies typically employ fixed spatiotemporal correlation structures when constructing error covariance. For example, spatial correlation is modeled as a pre-defined exponential decay function, while temporal correlation is confined to a fixed sliding window. However, the atmospheric diffusion capacity is not constant; it dynamically changes with meteorological conditions (such as wind speed and boundary layer height). Under stable weather conditions, local pollutant accumulation is strong, and the correlation should be stronger; while under windy conditions, turbulent mixing and advection transport are rapid, and the correlation should be weaker. Using a fixed correlation scale to fit all weather types leads to inversion results that are either overly smoothed, erasing strong local signals, or produce non-physical spatial structures, resulting in severe physical distortion.

[0004] Secondly, to improve the reliability of transport models, existing technologies often run two or more atmospheric transport models in parallel (such as FLEXPART and STILT), but the fusion of information from multiple models is rather crude. A common practice is to simply average or empirically weight the calculation results from different models. This "result-level fusion" ignores the systematic differences between different models in terms of physical parameterization, turbulence diffusion schemes, etc., which lead to different model error structures. Without optimal weighting at the error covariance level, the systematic bias of a single model will be directly inherited into the final flux field, failing to achieve effective error suppression.

[0005] Finally, the prior flux inventories upon which existing technologies rely suffer from spatiotemporal rigidity and slow event response. Commonly used fossil provenance inventories are mostly static data on a monthly or annual scale. When faced with short-term, high-frequency dynamic events such as wildfires, biomass burning, and surges in urban nighttime economic activity, these low-frequency priors cannot provide guiding information in sync with observations. This forces inversion algorithms to rely on observation residuals to correct fluxes, resulting in slow convergence and a significant lag in tracking abrupt events.

[0006] In summary, existing CO2 flux inversion techniques suffer from the following drawbacks: First, the fixed statistical structures (such as the covariance matrix) used are incompatible with the inherently time-varying characteristics of atmospheric diffusion processes; second, while multiple model information coexists, there is a lack of an optimal fusion architecture based on error covariance to eliminate systematic biases; and third, prior flux updates are lagging, failing to capture high-frequency dynamic emission events. Therefore, a method is urgently needed that can dynamically adjust spatiotemporal correlations, deeply fuse multi-model information, and correct prior fluxes in real time to overcome the accuracy bottleneck of existing technologies. Summary of the Invention

[0007] This invention provides a method and apparatus for CO2 flux inversion based on adaptive covariance and multi-model fusion, which addresses the shortcomings of existing CO2 flux inversion methods, such as rigid statistical structures, coarse multi-model fusion, and lagging updates of prior information, resulting in low inversion accuracy.

[0008] In a first aspect, the present invention provides a CO2 flux inversion method based on adaptive covariance and multi-model fusion, comprising: Acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux; Based on the initial residual sequence, a variational cost function is introduced to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and to construct a dynamic observation error covariance matrix. High-frequency remote sensing data is acquired in real time, a dynamic correction factor is constructed based on the high-frequency remote sensing data, and the initial prior flux is scaled in real time using the dynamic correction factor to generate a dynamic prior flux. Two atmospheric transport models are run in parallel to construct augmented observation vectors and augmented error covariance matrices, and a comprehensive transport operator is derived based on the augmented observation vectors and augmented error covariance matrices. By combining the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux, the final posterior CO2 flux is solved through Bayesian inversion.

[0009] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, observation data is acquired, and an initial residual sequence is calculated by combining the initial prior flux, including: Acquire observational data for the target area; the observational data includes meteorological reanalysis data, ground / tower CO2 concentration monitoring data, and NPP-VIIRS satellite remote sensing-aided data. The meteorological reanalysis data were processed using a WRF model to generate a high-resolution meteorological field. Meteorological classification indices are defined based on boundary layer height and friction velocity; A quantile regression forest algorithm is introduced to fit the background value of CO2 and output a high signal-to-noise ratio CO2 elevation concentration sequence; Based on the prior flux, an initial residual sequence for the target region is generated.

[0010] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, a variational cost function is introduced based on the initial residual sequence to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and a dynamic observation error covariance matrix is ​​constructed, including: The initial residual sequence is input into the variational cost function, and the optimal spatial correlation scale and temporal correlation scale are solved in parallel by minimizing the objective function. Based on the spatial correlation scale and the temporal correlation scale, the observation error covariance matrix is ​​adaptively constructed.

[0011] According to the present invention, an adaptive covariance and multi-model fusion CO2 flux inversion method is provided, which acquires high-frequency remote sensing data in real time, constructs a dynamic correction factor based on the high-frequency remote sensing data, and uses the dynamic correction factor to scale the initial prior flux in real time to generate a dynamic prior flux, including: Read high-frequency remote sensing data and construct a dynamic correction factor based on the high-frequency remote sensing data; The initial prior flux is multiplied pixel by pixel by the dynamic correction factor to obtain the dynamic prior flux.

[0012] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, high-frequency remote sensing data is read, and a dynamic correction factor is constructed based on the high-frequency remote sensing data, including: Read the high-frequency remote sensing data; if a fire point is detected, extract the fire radiation power. If high-brightness nighttime lights are detected, the radiation value is extracted; The dynamic correction factor is calculated based on the fire radiation power, the radiation value, and the preset calibration coefficient.

[0013] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, the high-frequency remote sensing data includes nighttime light data and / or fire point radiative power data from the NPP-VIIRS satellite.

[0014] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, two atmospheric transport models are run in parallel to construct augmented observation vectors and augmented error covariance matrices, including: Two atmospheric transport models are run in parallel to generate their respective footprint fields; Calculate the prediction divergence of the atmospheric transport model under the same meteorological field, and adaptively adjust the observation error covariance matrix based on the prediction divergence. The atmospheric transport model fusion mechanism is implemented, and the transport matrices of each atmospheric transport model are obtained. If the errors of the two atmospheric transport models are uncorrelated, then an augmented observation vector and an augmented error covariance matrix are constructed.

[0015] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, the prediction divergence of the atmospheric transport model under the same meteorological field is calculated, and the observation error covariance matrix is ​​adaptively adjusted according to the prediction divergence, including: Calculate the prediction divergence of the atmospheric transport model under the same meteorological field; If the predicted divergence exceeds a preset threshold, it is determined that the meteorological field in the current period is complex and there is a risk of systematic bias between models. Increase the weight of the diagonal elements in the observation error covariance matrix corresponding to the time period, and reduce the confidence of the observation data until the prediction divergence falls back to within the preset threshold.

[0016] According to the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention, a comprehensive transport operator is derived based on the augmented observation vector and the augmented error covariance matrix, including: Based on the principle of generalized least squares, the comprehensive transmission operator is obtained by weighted fusion of the inverse of the augmented error covariance matrix.

[0017] Secondly, the present invention also provides a CO2 flux inversion device based on adaptive covariance and multi-model fusion, comprising: The acquisition module is used to acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux. The module is used to introduce a variational cost function based on the initial residual sequence, dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and construct a dynamic observation error covariance matrix. The correction module is used to acquire high-frequency remote sensing data in real time, construct a dynamic correction factor based on the high-frequency remote sensing data, and use the dynamic correction factor to scale the initial prior flux in real time to generate a dynamic prior flux. The computation module is used to run two atmospheric transport models in parallel, construct augmented observation vectors and augmented error covariance matrices, and derive the integrated transport operator based on the augmented observation vectors and augmented error covariance matrices. The inversion module is used to combine the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux to solve for the final posterior CO2 flux through Bayesian inversion.

[0018] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the adaptive covariance and multi-model fusion CO2 flux inversion method as described in the first aspect above.

[0019] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CO2 flux inversion method of adaptive covariance and multi-model fusion as described in the first aspect above.

[0020] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the adaptive covariance and multi-model fusion CO2 flux inversion method as described in the first aspect above.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. Improved physical consistency of inversion results: This invention introduces a variational cost function to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state using the initial residual sequence, and constructs a dynamic observation error covariance matrix based on this. This approach allows the covariance structure to be adjusted in real time according to real meteorological conditions such as boundary layer height and stability. Under stable weather conditions, it automatically enhances spatial correlation to capture local accumulation, and under windy weather conditions, it automatically shrinks correlation to avoid non-physical diffusion. This makes the inversion results more consistent with the physical facts of atmospheric diffusion and avoids physical distortion caused by using a fixed correlation scale.

[0022] 2. Enhanced Inversion Accuracy and Robustness: This invention constructs augmented observation vectors and augmented error covariance matrices, and derives a comprehensive transport operator based on the generalized least squares principle, achieving optimal fusion of two parallel atmospheric transport models at the error structure level. This approach differs from simple result averaging; it adaptively adjusts weights according to the model's prediction divergence, effectively suppressing systematic biases in physical parameterization of a single model, mathematically approximating the minimum mean square error, and significantly improving the stability and reliability of the final posterior CO2 flux.

[0023] 3. Enhanced Capability for Detecting Dynamic Emission Events: This invention acquires high-frequency remote sensing data such as NPP-VIIRS in real time and constructs a dynamic correction factor to scale the static prior flux pixel by pixel in real time. This method solves the problems of lagging updates and inability to respond to abrupt events in traditional prior inventories. When facing high-frequency dynamic events such as wildfire outbreaks or peak nighttime emissions in cities, dynamically updated prior fluxes can provide a more accurate initial field for inversion, thereby accelerating the algorithm's convergence speed and significantly improving the ability to detect sudden emission sources and the accuracy of quantification. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the CO2 flux inversion method based on adaptive covariance and multi-model fusion provided by the present invention; Figure 2 This is a schematic diagram of the CO2 flux inversion method framework in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-model fusion structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] Please see Figure 1 This application provides a CO2 flux inversion method based on adaptive covariance and multi-model fusion, aiming to solve the problems in existing technologies where the statistical structure (such as spatiotemporal covariance) of the inversion model is fixed and cannot adapt to the dynamic changes of the real atmospheric diffusion process, and the information fusion method of multiple atmospheric transport models is coarse, resulting in insufficient inversion accuracy and physical realism. This method achieves high-precision and high-timeliness inversion of CO2 flux through a series of steps. The method specifically includes the following steps: Step S101: Obtain the observation data and calculate the initial residual sequence by combining it with the initial prior flux; Step S102: Based on the initial residual sequence, a variational cost function is introduced to dynamically solve the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and a dynamic observation error covariance matrix is ​​constructed. Step S103: Acquire high-frequency remote sensing data in real time, construct a dynamic correction factor based on the high-frequency remote sensing data, and use the dynamic correction factor to scale the initial prior flux in real time to generate a dynamic prior flux. Step S104: Run two atmospheric transport models in parallel to construct augmented observation vectors and augmented error covariance matrices, and derive the integrated transport operator based on the augmented observation vectors and augmented error covariance matrices. In step S105, the final posterior CO2 flux is solved by Bayesian inversion by combining the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux.

[0028] like Figure 2 As shown, Figure 2 This is a schematic diagram of the CO2 flux inversion method framework in an embodiment of the present invention. In one specific implementation, the method first acquires observational data and calculates an initial residual sequence based on the initial prior flux. This step is the starting point of the entire dynamic inversion framework. By calculating the difference between the preliminary simulation results of the model and the actual observations, it provides a basis for subsequent dynamic adjustments. Its design aims to quantify the sources of uncertainty under the current model settings, providing a driving force for the optimization of subsequent steps. By acquiring this initial residual sequence, the problem of lacking effective feedback signals to adjust the internal parameters of the model in traditional methods is solved, laying the foundation for achieving adaptive adjustment.

[0029] Subsequently, based on the acquired initial residual sequence, this method introduces a variational cost function to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and constructs a dynamic observation error covariance matrix. This step is one of the core innovations of this method, as it abandons the traditional approach of using fixed, artificially set correlation scales. The principle is that the statistical characteristics of the observation residuals (such as their spatiotemporal correlation) directly reflect the physical process of atmospheric diffusion. By minimizing a cost function that includes the covariance matrix, the "optimal" spatiotemporal correlation scale that best explains the current residual structure can be inferred. In this way, the covariance matrix is ​​no longer a static parameter, but can dynamically evolve with weather conditions (such as stable weather or strong winds), thus resolving the fundamental contradiction between the static covariance assumption and atmospheric physical reality, and significantly improving the physical consistency of the inversion results.

[0030] Next, the method acquires high-frequency remote sensing data in real time, constructs a dynamic correction factor based on this data, and uses this dynamic correction factor to scale the initial prior flux in real time to generate a dynamic prior flux. This step aims to address the problems of low update frequency and spatiotemporal rigidity of traditional prior flux inventories (such as ODIAC), which are unable to capture sudden emission events. The principle is to use high-frequency fire point and nighttime light data provided by satellites such as NPP-VIIRS as real-time indicators of human activity and biomass burning to construct a correction factor related to actual emission intensity. Applying this factor to a static prior inventory "activates" it, reflecting the dynamic changes in surface emissions in real time. By introducing this dynamic prior, the dependence of the inversion algorithm on observational data is greatly reduced, the convergence speed is accelerated, and the ability to capture sudden events such as wildfires and urban nighttime emission peaks is significantly improved.

[0031] Then, this method runs two atmospheric transport models in parallel, constructs augmented observation vectors and augmented error covariance matrices, and derives a comprehensive transport operator based on these. This step aims to address the problems of insufficient utilization of multi-model information and coarse fusion methods in existing technologies. Its principle is based on optimal estimation theory, treating the outputs of two independent models (such as FLEXPART and STILT) as an augmented observation system, and assigning a weight determined by the covariance of each model's error to its output. The comprehensive transport operator derived through generalized least squares is the optimal fusion result in the sense of minimum mean square error. In this way, the systematic bias of a single model can be effectively suppressed, and the advantages of different models can be fully utilized, thereby obtaining a more accurate and robust transport simulation than any single model or simple averaging, solving the problem of limited inversion accuracy caused by the lack of theoretical support for model fusion strategies.

[0032] Finally, this method combines the dynamic observation error covariance matrix generated in the preceding steps, the optimal integrated transport operator, and the dynamic prior flux to solve for the final posterior CO2 flux using a Bayesian inversion framework. This step is the endpoint of the entire process, integrating all the results of dynamic optimization and information fusion into a unified mathematical framework. By using a dynamic covariance matrix, dynamic prior information, and the optimal transport operator, the final Bayesian inversion can be solved under physically more consistent, statistically better, and informationally more complete conditions. This ensures that the final output posterior CO2 flux is not only more accurate but also has a more reliable uncertainty assessment because each of its components has undergone dynamic, data-driven optimization.

[0033] Furthermore, in a preferred embodiment, the steps of acquiring observational data and calculating the initial residual sequence are refined. Specifically, firstly, observational data for the target area are acquired. This data is multi-source, including meteorological reanalysis data used to drive the atmospheric model, ground / tower CO2 concentration monitoring data as inversion constraints, and NPP-VIIRS satellite remote sensing data used for auxiliary analysis. Subsequently, the original meteorological reanalysis data is downscaled using the WRF (Weather Research and Forecasting) model to generate a high-resolution meteorological field, providing a more refined driver for subsequent atmospheric transport simulations. Next, to more scientifically handle background concentrations, a meteorological classification index based on boundary layer height and friction velocity is defined to guide the selection of subsequent data processing windows. Then, the quantile regression forest algorithm is introduced. Compared to traditional methods such as moving averages, this algorithm can more robustly fit the background CO2 values, thereby outputting a CO2 elevation concentration sequence with a higher signal-to-noise ratio. Finally, this elevation concentration sequence is combined with prior flux to generate the initial residual sequence used to drive subsequent optimizations. This series of meticulous preprocessing steps aims to improve the quality of input data from the source, providing a solid foundation for the accuracy and stability of the entire inversion system.

[0034] For example, meteorological reanalysis data (e.g., ERA5), surface / tower CO2 concentration monitoring data, and NPP-VIIRS satellite remote sensing-aided data for the target area are acquired. The meteorological reanalysis data are processed using a WRF model to generate a high-resolution meteorological field (e.g., 3km resolution). The fixed interquartile range (IQR) window is abandoned in favor of an adaptive windowing strategy based on meteorological classification. A meteorological classification index is defined based on boundary layer height and friction velocity.

[0035] in, Indicates the boundary layer height. Indicates friction speed, This indicates a meteorological classification index. If... =1, use a 72-hour long window to smooth the background; if it is 0, use a 24-hour short window.

[0036] Subsequently, a quantile regression forest algorithm was introduced to fit the background values ​​of CO2 concentration. The quantiles were then set. Please provide a solution:

[0037] Wherein, represents the background value of CO2 concentration. express Quantiles This represents the observed concentration. Compared to a simple moving average, this method effectively preserves weak source signals and outputs a high signal-to-noise ratio CO2 elevation concentration sequence. .

[0038] This concentration elevation sequence contains information not captured by model errors and real physical processes, and serves as the basis for driving adjustments to the covariance structure. Let... H is the observation vector, and H is the transfer matrix (footprint field). Let be the prior flux vector. Calculate the initial residual sequence:

[0039] Where r represents the initial residual sequence, H represents the observation vector, and H represents the transfer matrix. This represents the prior flux vector.

[0040] In another preferred embodiment, the steps for constructing the dynamic observation error covariance matrix are refined. Specifically, the initial residual sequence generated in the previous step is input into a preset variational cost function. This cost function is designed such that its value is correlated with the difference between the theoretical covariance matrix and the actual residual sample covariance matrix. By minimizing this objective function, the optimal spatial correlation scale and temporal correlation scale can be solved in parallel and automatically. These two scale parameters directly reflect the spatial and temporal correlation range of CO2 concentration disturbance information under the current atmospheric conditions. Finally, based on these two dynamically solved optimal scale parameters, the observation error covariance matrix is ​​adaptively constructed. This method enables the covariance structure to "sense" and respond to changes in atmospheric diffusion conditions, realizing the dynamic coupling of the statistical model and the physical process.

[0041] For example, the variational cost function is constructed as follows. The relevant scale is used as an optimization variable:

[0042] in, Indicates spatially relevant scales. Indicates a time-related scale. The observation error covariance matrix is ​​represented by its off-diagonal elements, which are defined by an exponential function:

[0043] in, For the site i and j Spatial distance between For the observation time difference, This represents the standard deviation of the observation error.

[0044] By minimizing the variational cost functionJ Dynamically solve for the optimal spatial correlation scale Time-related scales :

[0045] In calm weather conditions, the optimizer will tend to increase... and This allows information to spread over a larger spatial and temporal range; however, under strong winds and convective weather, the relevant scale is reduced, limiting the non-physical diffusion of information. The resulting dynamic covariance matrix can accurately match the real-time atmospheric diffusion capacity.

[0046] Furthermore, to overcome the problem of lagging updates in static inventories (such as ODIAC), the steps for generating dynamic prior fluxes were refined, and NPP-VIIRS nighttime light and fire point data were introduced for real-time correction. First, the system reads high-frequency remote sensing data in real time. Then, a dynamic correction factor is constructed based on this data. Finally, the initial, static prior flux map is multiplied pixel-by-pixel with this dynamic correction factor to obtain a dynamic prior flux that reflects real-time emission changes. This approach injects the "dynamic" information from high-frequency remote sensing observations into the low-frequency updated "static" prior inventory, achieving real-time updates of prior information.

[0047] In one alternative implementation, the process of constructing the dynamic correction factor is described in more detail. When reading high-frequency remote sensing data, the system performs feature detection. If a fire point is detected, the fire radiative power (FRP) of that fire point is extracted, a value directly related to the intensity of biomass combustion. If high-intensity nighttime lights (NTLs) are detected in the nighttime data, their radiative values ​​are extracted, a value related to the intensity of human activity in urban and industrial areas. Finally, based on the extracted fire radiative power, nighttime light radiative values, and a set of pre-defined calibration coefficients, the final dynamic correction factor is calculated using a weighted formula. This process transforms remotely sensed physical quantities into direct adjustment coefficients for CO2 emission fluxes.

[0048] For example, the time of reading t NPP-VIIRS data. If the grid ( i , j If a fire point is detected, its fire radiation power is extracted; if a high-brightness nighttime light is detected, its radiation value is extracted. A dynamic correction factor is constructed.

[0049] in, Indicates the dynamic correction factor. Indicates fire radiation power. Indicates the radiation value. This represents the calibration coefficient.

[0050] Static prior flux Multiplying the dynamic prior flux by the dynamic correction factor pixel by pixel yields the dynamic prior flux:

[0051] in, This represents dynamic prior flux. As a new prior input to the Bayesian inversion framework, this closed-loop correction mechanism significantly reduces the inversion's dependence on observational data and improves its ability to capture sudden emission events such as wildfires and urban nighttime activities. In each inversion iteration, high-frequency remote sensing information is used to perform pixel-level scaling of the prior flux, forming a closed-loop correction link of "remote sensing indication - prior scaling - inversion optimization," ensuring that the prior field can accurately reflect the actual emission fluctuations on the land surface in real time.

[0052] In another preferred embodiment, the steps for deriving the integrated transport operator are refined. First, two different atmospheric transport models (e.g., FLEXPART and STILT) are run in parallel, driven by the same high-resolution meteorological field, each generating its own footprint field (i.e., transport matrix). Next, the prediction divergence of the two models under the same conditions is calculated, i.e., the difference in their concentration contributions to the same flux field. This divergence is used as a diagnostic indicator, adaptively adjusting the observation error covariance matrix according to its magnitude; a large model divergence indicates high model uncertainty, and the weight of the observation data should be appropriately reduced. Then, a state-space model fusion mechanism is executed, and the transport matrix of each atmospheric transport model is obtained. Assuming that the errors of the two models are independent, an augmented observation vector (stacked from the outputs of the two models) and an augmented error covariance matrix (a block diagonal matrix, where the diagonal blocks are the error covariances of each model) are constructed.

[0053] Furthermore, the adjustment steps in the multi-model fusion process described above are explained in detail. After calculating the prediction divergence of the two atmospheric transport models under the same meteorological field, the system compares it with a preset threshold. If the divergence exceeds the threshold, it is determined that the meteorological field for the current period is relatively complex, and there may be a significant risk of systematic bias between the models. At this time, the system automatically increases the weight of the diagonal elements in the observation error covariance matrix corresponding to the current period. The physical meaning of this operation is to proactively reduce the confidence level of the observation data for that period to avoid unreliable model transport results excessively affecting the inversion. This weight will continue to increase until the prediction divergence falls back below the preset threshold, thereby achieving a dynamic risk control of model uncertainty.

[0054] For example, the high-resolution meteorological field output by WRF is used to drive the FLEXPART model and the STILT model respectively, to obtain two independent sensitivity footprint fields. , To quantify the uncertainty of the model structure, the prediction divergence between the two models under the same meteorological field was calculated. To quantify the uncertainty of the model structure, the specific formula is as follows:

[0055] in, Indicates the degree of divergence in predictions. and Indicates a footprint field. This represents the prior flux vector. If If the divergence exceeds a preset threshold (e.g., twice the standard deviation of historical divergence), the meteorological field for that period is deemed complex, indicating a significant risk of systematic bias between models. In this case, the weights of the diagonal elements corresponding to that period in the observation error covariance matrix are automatically increased to reduce the confidence level of the observation data until the divergence falls back to a reasonable range. Simultaneously, a dual-model fusion mechanism is triggered; otherwise, the process can be simplified to outputting the mean of a single model.

[0056] Assuming the errors of the two models are uncorrelated, construct the augmented observation vector and the block diagonal augmented error covariance matrix:

[0057]

[0058] Where Y represents the augmented observation vector, R aug This represents the block diagonal augmented error covariance matrix. and These are the concentration vectors simulated by the two models, respectively; and These are the observation error covariance matrices of the two models, and their weights are determined by the historical prediction error statistics or uncertainty quantification results of their respective models (i.e., the larger the error, the smaller the weight).

[0059] In one alternative implementation, the specific derivation of the synthesized transport operator is explained. It is based on the generalized least squares principle and is obtained by weighted fusion using the inverse of the augmented error covariance matrix. Specifically, the weight of each model's transport matrix is ​​proportional to the inverse of its error covariance matrix. This fusion method, where "the smaller the error, the larger the weight," ensures that the final synthesized transport operator is mathematically optimal and can minimize the uncertainty caused by model transport errors.

[0060] For example, such as Figure 3 As shown, Figure 3This is a schematic diagram of the multi-model fusion structure in an embodiment of the present invention. Based on the generalized least squares principle, a comprehensive transport operator is constructed:

[0061] Among them, H optimal Let H represent the integrated transmission operator, and let H represent the transmission matrix. R aug This represents the block diagonal augmented error covariance matrix. The optimal synthesized transport operator is calculated using a Bayesian solver (such as MCMC or variational inference), and the posterior flux is calculated. x The solution and the output of uncertainty quantification.

[0062] In summary, this method, through the synergy of adaptive covariance and optimal fusion of two models, achieves an upgrade inversion kernel from "rigid statistical assumptions" to "physical dynamic following." Furthermore, by combining high-frequency remote sensing dynamic correction with static inventory, this method enables a leap from "monthly dead maps" to "event-level live maps" guided by prior knowledge. This method provides a complete solution for highly dynamic, high-resolution carbon flux inversion needs such as regional carbon neutrality monitoring and urban carbon emission MRV verification, offering physical consistency, statistical optimization, and excellent timeliness, significantly improving inversion accuracy and the ability to capture abrupt events.

[0063] This invention also provides an adaptive covariance and multi-model fusion CO2 flux inversion device. The following describes the adaptive covariance and multi-model fusion CO2 flux inversion device provided by this invention. The adaptive covariance and multi-model fusion CO2 flux inversion device described below can be referred to in correspondence with the adaptive covariance and multi-model fusion CO2 flux inversion method described above. The device includes: The acquisition module is used to acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux. The module is designed to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state based on the initial residual sequence and a variational cost function, and to construct a dynamic observation error covariance matrix. The correction module is used to acquire high-frequency remote sensing data in real time, construct dynamic correction factors based on the high-frequency remote sensing data, and use the dynamic correction factors to scale the initial prior flux in real time to generate dynamic prior flux. The computation module is used to run two atmospheric transport models in parallel, construct augmented observation vectors and augmented error covariance matrices, and derive the integrated transport operator based on the augmented observation vectors and augmented error covariance matrices. The inversion module is used to combine the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux to solve for the final posterior CO2 flux through Bayesian inversion.

[0064] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other via the communication bus 404. The processor 401 can call logical instructions in the memory 403 to execute an adaptive covariance and multi-model fusion CO2 flux inversion method, which includes: Acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux; Based on the initial residual sequence, a variational cost function is introduced to dynamically solve the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and a dynamic observation error covariance matrix is ​​constructed. High-frequency remote sensing data is acquired in real time. A dynamic correction factor is constructed based on the high-frequency remote sensing data. The initial prior flux is scaled in real time using the dynamic correction factor to generate a dynamic prior flux. Two atmospheric transport models were run in parallel to construct augmented observation vectors and augmented error covariance matrices, and a comprehensive transport operator was derived based on the augmented observation vectors and augmented error covariance matrices. By combining the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux, the final posterior CO2 flux is solved through Bayesian inversion.

[0065] Furthermore, the logical instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the adaptive covariance and multi-model fusion CO2 flux inversion method provided by the above methods, the method comprising: Acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux; Based on the initial residual sequence, a variational cost function is introduced to dynamically solve the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and a dynamic observation error covariance matrix is ​​constructed. High-frequency remote sensing data is acquired in real time. A dynamic correction factor is constructed based on the high-frequency remote sensing data. The initial prior flux is scaled in real time using the dynamic correction factor to generate a dynamic prior flux. Two atmospheric transport models were run in parallel to construct augmented observation vectors and augmented error covariance matrices, and a comprehensive transport operator was derived based on the augmented observation vectors and augmented error covariance matrices. By combining the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux, the final posterior CO2 flux is solved through Bayesian inversion.

[0067] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the adaptive covariance and multi-model fusion CO2 flux inversion method provided by the methods described above, the method comprising: Acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux; Based on the initial residual sequence, a variational cost function is introduced to dynamically solve the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and a dynamic observation error covariance matrix is ​​constructed. High-frequency remote sensing data is acquired in real time. A dynamic correction factor is constructed based on the high-frequency remote sensing data. The initial prior flux is scaled in real time using the dynamic correction factor to generate a dynamic prior flux. Two atmospheric transport models were run in parallel to construct augmented observation vectors and augmented error covariance matrices, and a comprehensive transport operator was derived based on the augmented observation vectors and augmented error covariance matrices. By combining the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux, the final posterior CO2 flux is solved through Bayesian inversion.

[0068] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for CO2 flux inversion using adaptive covariance and multi-model fusion, characterized in that, include: Acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux; Based on the initial residual sequence, a variational cost function is introduced to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and to construct a dynamic observation error covariance matrix. High-frequency remote sensing data is acquired in real time, a dynamic correction factor is constructed based on the high-frequency remote sensing data, and the initial prior flux is scaled in real time using the dynamic correction factor to generate a dynamic prior flux. Two atmospheric transport models are run in parallel to construct augmented observation vectors and augmented error covariance matrices, and a comprehensive transport operator is derived based on the augmented observation vectors and augmented error covariance matrices. By combining the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux, the final posterior CO2 flux is solved through Bayesian inversion.

2. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 1, characterized in that, Acquire observational data and calculate the initial residual sequence using the initial prior flux, including: Acquire observational data for the target area; the observational data includes meteorological reanalysis data, ground / tower CO2 concentration monitoring data, and NPP-VIIRS satellite remote sensing-aided data. The meteorological reanalysis data were processed using a WRF model to generate a high-resolution meteorological field. Meteorological classification indices are defined based on boundary layer height and friction velocity; A quantile regression forest algorithm is introduced to fit the background value of CO2 and output a high signal-to-noise ratio CO2 elevation concentration sequence; Based on the prior flux, an initial residual sequence for the target region is generated.

3. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 1, characterized in that, Based on the initial residual sequence, a variational cost function is introduced to dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and a dynamic observation error covariance matrix is ​​constructed, including: The initial residual sequence is input into the variational cost function, and the optimal spatial correlation scale and temporal correlation scale are solved in parallel by minimizing the objective function. Based on the spatial correlation scale and the temporal correlation scale, the observation error covariance matrix is ​​adaptively constructed.

4. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 1, characterized in that, Real-time acquisition of high-frequency remote sensing data; construction of a dynamic correction factor based on the high-frequency remote sensing data; and real-time scaling of the initial prior flux using the dynamic correction factor to generate a dynamic prior flux, including: Read high-frequency remote sensing data and construct a dynamic correction factor based on the high-frequency remote sensing data; The initial prior flux is multiplied pixel by pixel by the dynamic correction factor to obtain the dynamic prior flux.

5. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 4, characterized in that, Reading high-frequency remote sensing data and constructing a dynamic correction factor based on the high-frequency remote sensing data includes: Read the high-frequency remote sensing data; if a fire point is detected, extract the fire radiation power. If high-brightness nighttime lights are detected, the radiation value is extracted; The dynamic correction factor is calculated based on the fire radiation power, the radiation value, and the preset calibration coefficient.

6. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 5, characterized in that, The high-frequency remote sensing data includes nighttime light data and / or fire point radiant power data from the NPP-VIIRS satellite.

7. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 1, characterized in that, Two atmospheric transport models are run in parallel to construct augmented observation vectors and augmented error covariance matrices, including: Two atmospheric transport models are run in parallel to generate their respective footprint fields; Calculate the prediction divergence of the atmospheric transport model under the same meteorological field, and adaptively adjust the observation error covariance matrix based on the prediction divergence. The atmospheric transport model fusion mechanism is implemented, and the transport matrices of each atmospheric transport model are obtained. If the errors of the two atmospheric transport models are uncorrelated, then an augmented observation vector and an augmented error covariance matrix are constructed.

8. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 7, characterized in that, Calculating the prediction divergence of the atmospheric transport model under the same meteorological field, and adaptively adjusting the observation error covariance matrix based on the prediction divergence, includes: Calculate the prediction divergence of the atmospheric transport model under the same meteorological field; If the predicted divergence exceeds a preset threshold, it is determined that the meteorological field in the current period is complex and there is a risk of systematic bias between models. Increase the weight of the diagonal elements in the observation error covariance matrix corresponding to the time period, and reduce the confidence of the observation data until the prediction divergence falls back to within the preset threshold.

9. The CO2 flux inversion method based on adaptive covariance and multi-model fusion according to claim 1, characterized in that, Based on the augmented observation vector and the augmented error covariance matrix, a comprehensive transport operator is derived, including: Based on the principle of generalized least squares, the comprehensive transmission operator is obtained by weighted fusion of the inverse of the augmented error covariance matrix.

10. A CO2 flux inversion device with adaptive covariance and multi-model fusion, characterized in that, include: The acquisition module is used to acquire observation data and calculate the initial residual sequence by combining it with the initial prior flux. The module is used to introduce a variational cost function based on the initial residual sequence, dynamically solve for the spatiotemporal correlation scale that matches the current atmospheric diffusion state, and construct a dynamic observation error covariance matrix. The correction module is used to acquire high-frequency remote sensing data in real time, construct a dynamic correction factor based on the high-frequency remote sensing data, and use the dynamic correction factor to scale the initial prior flux in real time to generate a dynamic prior flux. The computation module is used to run two atmospheric transport models in parallel, construct augmented observation vectors and augmented error covariance matrices, and derive the integrated transport operator based on the augmented observation vectors and augmented error covariance matrices. The inversion module is used to combine the observation error covariance matrix, the integrated transport operator, and the dynamic prior flux to solve for the final posterior CO2 flux through Bayesian inversion.

Images