Atmospheric flux estimation method based on multi-scale dynamic correlation

Abstract

The present application relates to the technical field of atmospheric flux calculation, in particular to an atmospheric flux estimation method based on multi-scale dynamic correlation. In the existing multi-scale atmospheric flux calculation technology, multi-scale data integration relies on manual interpolation or step-by-step calculation, which cannot effectively capture the spatio-temporal correlation of different scale atmospheric processes, and is prone to significant scale connection errors; and traditional iterative calculation needs to repeatedly check parameters, which is difficult to meet the time requirement in the scene of atmospheric pollution emergency monitoring, real-time carbon flux evaluation, etc. The present application firstly aligns the multi-source multi-scale observation data in time and space and suppresses noise, and then adopts an AI model to construct a cross-scale dynamic correlation mechanism to replace the traditional manual experience formula. This method can effectively reduce the scale connection error, solve the problem that the traditional technology cannot balance the precision and speed, and is suitable for scenes such as atmospheric pollution emergency response and regional carbon cycle dynamic monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of atmospheric flux calculation, more particularly, it relates to an atmospheric flux estimation method based on multi-scale dynamic correlation. BACKGROUND

[0002] Atmospheric flux estimation, as a core supporting technology in the field of atmospheric flux calculation, is directly related to the decision-making scientificity of scenarios such as atmospheric pollution emergency response and regional carbon cycle dynamic monitoring. Its estimation accuracy and response timeliness are key indicators for measuring the practicality of the technology.

[0003] There are two major core pain points in current multi-scale atmospheric flux calculation technology: first, multi-scale data integration relies on manual interpolation or step-by-step calculation, which cannot effectively capture the spatio-temporal correlation of different scale atmospheric processes, resulting in significant scale connection errors. For example, satellite remote sensing, ground observation and numerical simulation data are prone to data conflicts due to rough spatio-temporal matching; second, traditional iterative calculation needs to repeatedly check parameters, which is difficult to meet the rapid response demand in time-sensitive scenarios such as atmospheric pollution emergency monitoring and real-time carbon flux evaluation. In addition, the existing data fusion lacks a physical-driven two-way verification mechanism, which cannot dynamically quantify the representative error and quality of the data source; the construction of cross-scale correlation relies on manual experience formula, which lacks dynamic optimization structure support, further limiting the precision improvement.

[0004] Therefore, the present application proposes an atmospheric flux estimation method based on multi-scale dynamic correlation to solve the technical bottlenecks of the prior art. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide an atmospheric flux estimation method based on multi-scale dynamic correlation.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] The atmospheric flux estimation method based on multi-scale dynamic correlation specifically comprises the following steps:

[0008] Step 1, spatio-temporal alignment and noise suppression of multi-scale observation data.

[0009] S11, data acquisition: collecting three types of multi-scale data sources, namely satellite remote sensing data, ground observation data and numerical simulation data;

[0010] S12, spatio-temporal alignment: taking the ground flux tower observation timestamp as the reference, performing time alignment and space alignment on the satellite remote sensing data and numerical simulation data;

[0011] S13, noise suppression: according to the different noise characteristics of satellite remote sensing data, ground observation data and numerical simulation data, differential denoising strategies are adopted;

[0012] S14, physically driven multi-source data fusion: based on bidirectional tracing, multi-source data adaptive confidence weighted fusion is performed for each grid point after spatio-temporal alignment and denoising, the confidence weights of satellite remote sensing data, ground observation data and numerical simulation data are dynamically calculated, and physically driven data fusion is realized;

[0013] Step two, construction of a cross-scale dynamic correlation mechanism based on an AI model.

[0014] S21, AI model selection and training: a dynamic graph neural network is selected as the correlation model, a graph node and edge weight function are defined, a message passing mechanism is designed, a training set is constructed and the model is trained;

[0015] S22, dynamic correlation calculation for pollutant flux: the reasonableness of the initial flux field is verified based on the mass balance equation, the edge contribution is calculated by the integral gradient method and the graph structure is optimized until the convergence condition is met, and the final pollutant flux distribution field is output.

[0016] Further, in step S14, the physically driven multi-source data fusion includes three sub-steps:

[0017] S141, uplink tracing: a Lagrangian random particle diffusion model is used to quantify the spatial representative error of ground observation data and numerical simulation data;

[0018] S142, downlink evaluation: based on real-time weather condition data quality diagnosis, the data quality is quantified by real-time meteorological elements, and the dynamic quality factors of the three types of multi-scale data sources are calculated;

[0019] S143, dynamic weight synthesis and multi-source data fusion: based on the quantitative indicators of uplink tracing and downlink evaluation, the final fusion weight is calculated, and the fusion value of the target point data is calculated.

[0020] Further, the uplink tracing has the following specific process:

[0021] Calculation of the representativeness of the flux tower data: for the flux tower site, the footprint function is used to calculate the underlying surface representativeness score of the target grid point;

[0022] Calculation of the representativeness of numerical simulation data: for the model grid, the representativeness error is measured by the underlying surface heterogeneity; the model grid is the basic calculation unit for spatial discretization in numerical simulation.

[0023] Further, the downlink evaluation has the following specific process:

[0024] The satellite data quality factor is calculated in combination with the cloud cover ratio and the ground visibility; the numerical simulation data stability factor is calculated based on the Richardson number; and the ground station flow field correlation factor is calculated in combination with the angle between the flux tower site and the prevailing wind direction and the horizontal attenuation.

[0025] Further, the dynamic weight synthesis is combined with multi-source data fusion, and the specific process is as follows:

[0026] First, a satellite remote sensing data confidence factor is calculated based on the satellite data quality factor, a ground observation data confidence factor is calculated based on the flux tower data representativeness and ground site flow field correlation factor, and a numerical simulation data confidence factor is calculated based on the numerical simulation data representativeness and numerical simulation data stability factor; second, the optimal weight of the target point data is calculated based on three types of multi-scale data sources; and finally, the final fusion value of the target point data is calculated based on the initial fusion weight and the optimal weight.

[0027] Further, the initial flux field rationality is verified based on the mass balance equation, and the specific process is as follows:

[0028] The physical residual of the initial flux value and the actual observation value is calculated by discretizing the mass balance equation, and the average physical residual in the region is calculated.

[0029] Further, the integral gradient method is used to calculate the edge contribution and optimize the graph structure, and the specific process is as follows:

[0030] Locate the high residual node: identify the high residual node through the average physical residual in the region;

[0031] Calculate the edge contribution: calculate the contribution of the high residual node and its incoming edge by the integral gradient method;

[0032] Dynamic optimization of graph structure: normalize the contribution by the Sigmoid function, and adjust the edge weight by the following formula:

[0033] ;

[0034] wherein, is the learning rate, is the current edge weight, is the Sigmoid function, is the edge contribution of the high residual node and its incoming edge.

[0035] Further, in the step S22, the convergence condition is as follows:

[0036] ;

[0037] wherein, is the average physical residual in the region calculated for the time, is the convergence threshold, is the iteration number.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] 1. Bidirectional tracing improves data fusion quality: The bidirectional tracing mechanism of uplink tracing and downlink evaluation is adopted. The uplink quantifies the spatial representative error of ground observation and numerical simulation data through the Lagrangian random particle diffusion model. The downlink combines real-time meteorological elements (cloud cover, Richardson number, etc.) to calculate the dynamic quality factor of each data source. Finally, the confidence weight is dynamically generated to realize multi-source fusion. This method replaces the traditional manual interpolation, effectively captures the physical correlation of data at different scales, significantly reduces the scale connection error, and provides a high-quality data basis for flux estimation.

[0040] 2. Dynamic graph structure optimization balances accuracy and timeliness: After constructing the cross-scale correlation model with dynamic graph neural network, the physical residual is calculated by the discretization of the mass balance equation to identify high residual nodes that meet the preset conditions. The edge contribution degree (quantifying the influence degree and positive and negative effects of the edge on the high residual node flux) is calculated for the high residual node and its incoming edge by the integral gradient method. Then, the contribution degree is normalized by the Sigmoid function to dynamically adjust the edge weight, and the local forward reasoning is performed to correct the flux of the high residual area and its neighborhood nodes. Finally, the regional average physical residual is verified for convergence to output the global flux field. This process replaces the redundant operations of traditional manual parameter verification and manual correction of abnormal flux values. It not only improves the cross-scale correlation accuracy through precise optimization of high residual areas, but also reduces invalid iteration steps, solving the problem of balancing accuracy and speed in traditional technology, and meeting the needs of time-sensitive scenarios such as atmospheric pollution emergency monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 Flowchart of the atmospheric flux estimation method based on multi-scale dynamic correlation;

[0042] Figure 2 Schematic diagram of the physical driven multi-source data fusion of the present application;

[0043] Figure 3 Flowchart of the dynamic correlation calculation for pollutant flux of the present application. DETAILED DESCRIPTION

[0044] Embodiment, refer to Figure 1 The atmospheric flux estimation method based on multi-scale dynamic correlation of the present embodiment specifically includes the following steps:

[0045] Step one, time and space alignment and noise suppression of multi-scale observation data.

[0046] S11, data acquisition: collect three types of multi-scale data sources;

[0047] Satellite remote sensing data: For the scenario of atmospheric pollutant flux estimation, the pollutant column concentration data of Sentinel-5 Precursor (Sentinel-5P) such as VOCs column concentration (Volatile Organic Compounds) are adopted, with a spatial resolution of 3.5 km and daily update;

[0048] Ground observation data: Based on the FLUXNET flux tower network, pollutant flux data with a spatial resolution of 50-200 m and a time resolution of 30 minutes are collected; through the ground automatic monitoring site, pollutant concentration data with a spatial resolution of 1-5 km and a time resolution of 1 hour are obtained;

[0049] Numerical simulation data: The Weather Research and Forecasting Model (WRF) is used to generate meteorological field data with a spatial resolution of 10-50 km and a time resolution of 1 hour, including wind speed, temperature field, etc.;

[0050] S12, spatiotemporal alignment: taking the ground flux tower observation timestamp as the benchmark, satellite data and numerical simulation data are matched in time and space;

[0051] Temporal alignment: the linear interpolation method is used to unify the time resolution of satellite / simulation data to 30 minutes, and the missing time node data is supplemented by trend extrapolation of adjacent time data; wherein the extrapolation supplement process specifically includes: using the least square method to linearly regress the atmospheric pollutant flux data of continuous multiple adjacent time points to obtain a trend equation, and predicting the atmospheric pollutant flux of the next time point according to the trend equation obtained by fitting; if multiple time data need to be supplemented, iterative calculation is required;

[0052] Spatial alignment: a unified geographical grid (such as 500m x 500m in the embodiment) is constructed, and the adaptive weight interpolation method is used to realize data fusion, specifically as follows:

[0053] The interpolation weight W is dynamically adjusted according to the spatial resolution R of the data source, and the calculation formula is:

[0054] ;

[0055] Wherein, is the resolution of the i th data source, is the total number of data sources;

[0056] ​Taking the unified geographic grid as the reference, each grid unit is traversed, and for different scale data source data points falling into the unit, a weighted average is performed according to the calculated weight, and the weighted result is taken as the final data value of the grid unit, so that different scale data is accurately mapped to the same grid, and spatial distortion caused by traditional fixed interpolation is avoided;

[0057] S13, noise suppression: for the noise characteristics of different data sources, a differentiated denoising strategy is adopted;

[0058] Satellite remote sensing data: using wavelet threshold denoising method, selecting Db4 wavelet basis, and calculating threshold value adaptively through data signal-to-noise ratio, to eliminate abnormal values caused by cloud coverage and atmospheric scattering;

[0059] Ground observation data: using Kalman filter method, according to statistical characteristics such as atmospheric pollutant flux fluctuation variance, mean deviation, and abnormal value distribution frequency of historical observation data within 30 days, to filter noise caused by instrument drift and transient interference;

[0060] Numerical simulation data: first, using numerical models (such as WRF-Noah-MP, EC-Earth, etc. Atmospheric-land coupled model) based on meteorological reanalysis data, land use data and topographic parameters, atmospheric pollutant fluxes at different height layers are calculated by solving atmospheric dynamics and thermodynamics equations, to obtain simulation values; then using residual correction method, the residual between simulation value and contemporaneous ground observation value is used as a correction term to dynamically adjust numerical simulation data and reduce system bias;

[0061] S14, physically driven multi-source data fusion: as shown in Figure 2 , based on bidirectional source tracing, the confidence weight of satellite remote sensing data, ground observation data and numerical simulation data is dynamically calculated for each grid point after time and space alignment and denoising, to realize physically driven intelligent fusion and improve data quality;

[0062] S141, uplink source tracing: based on representative evaluation of Lagrangian footprint model, Lagrangian stochastic particle diffusion model (LSM) is used to quantify the spatial representative error of ground observation data and numerical simulation data;

[0063] Flux tower data representative: for flux tower site s, the footprint function is used to calculate the underlying surface representative score of target grid point i , the formula is as follows:

[0064] ;

[0065] Wherein, is the effective source area of site s; is the ground point the underlying surface type of the target grid point, the corresponding underlying surface type, the corresponding underlying surface type, is an indicator function, when the condition in the bracket is true, the indicator function value is 1, when the condition in the bracket is not true, the indicator function value is 0; , the higher the value, the better the representation; the underlying surface type includes farmland, forest, city, etc.

[0066] Numerical simulation data representativeness: for the model grid m, the representativeness error of underlying surface heterogeneity is represented by , the calculation formula is as follows:

[0067] ;

[0068] wherein, is the total number of high-resolution grids in the model grid , represents the underlying surface type of the th high-resolution grid in the model grid m, , the higher the value, the worse the representation; the model grid is the basic calculation unit used for spatial discretization in numerical simulation;

[0069] S142, downward evaluation: data quality diagnosis based on real-time meteorological conditions, quantifying data quality through real-time meteorological elements, calculating dynamic quality factors;

[0070] Satellite data quality factor , the calculation formula is as follows:

[0071] ;

[0072] wherein, is the cloud cover ratio, i.e. the estimated proportion of cloud coverage of the sky, the cloud cover ratio is calculated by combining satellite sensors and algorithms, is the ground visibility, unit: km, is the reference threshold value;

[0073] Numerical simulation data stability factor , the calculation formula is as follows:

[0074] ;

[0075] wherein, is the Richardson number, a is an adjustment parameter, which is determined by domain experience, in the embodiment , , the higher the value, the more unstable the atmosphere, the higher the simulation confidence;

[0076] Ground station flow field correlation factor , the calculation formula is as follows:

[0077] ;

[0078] wherein, is the angle between the line connecting the flux tower site and the target point and the prevailing wind direction, is the horizontal distance, the attenuation scale is an empirical value; the prevailing wind direction refers to the highest frequency of occurrence of the wind direction in a certain region within a certain period of time obtained by statistical method, reflecting the dominant trend of atmospheric movement in the region;

[0079] S143, dynamic weight synthesis and multi-source data fusion: comprehensive quantitative index calculation of final fusion weight;

[0080] Dynamic confidence factor calculation:

[0081] Satellite remote sensing data confidence factor: ;

[0082] Ground observation data confidence factor: ; is the number of sites within the influence range;

[0083] Numerical simulation data confidence factor: ;

[0084] The maximum weight and the fusion value:

[0085] The maximum weight of the data source src (including satellite remote sensing data sat, ground observation data grd and numerical simulation data mod) at the target point i is determined by the static weight and the dynamic confidence factor of the target point i , the calculation formula is as follows:

[0086] ;

[0087] The final fusion value of the target point i is:

[0088] ;

[0089] wherein, is the observation value or simulation value of the data source at the point ;

[0090] The static weight is the initial fusion weight determined only according to the historical average accuracy of the data source without introducing the bidirectional tracing mechanism, and the specific process is as follows:

[0091] The historical average accuracy of satellite remote sensing data sat, ground observation data grd and numerical simulation data mod in the target area is counted respectively (value range , the higher the value, the better the historical estimated accuracy, determined by the degree of coincidence between long-term observed / simulated values and true values); Calculate the static weight by normalization:

[0092] ;

[0093] Step two, construction of cross-scale dynamic correlation mechanism based on AI model.

[0094] S21, AI model selection and training: Dynamic Graph Neural Network (DyGNN) is selected as the correlation model to replace traditional artificial experience formula, adopting a three-layer message passing structure, the specific design is as follows:

[0095] Node definition: the multi-scale data grid unit processed in step one is taken as a graph node, each node contains data features such as normalized vegetation index (NDVI), land surface temperature, wind speed; physical properties such as underlying surface type, elevation;

[0096] Dynamic calculation of edge weight: based on atmospheric physical processes (such as turbulence intensity, water vapor gradient), the edge weight function is designed, the formula is as follows:

[0097] ;

[0098] wherein, is the edge weight of node p and node q; is the temperature difference between the two nodes; is the spatial distance; is the average wind speed, which is obtained through meteorological station monitoring equipment, meteorological satellite data or numerical weather prediction model, and the data covers wind speed information at different height layers, which is processed by time average to obtain the required average wind speed; is the roughness length, which can be obtained based on field measurement or geographic information system data, or calculated by using laser radar and other equipment to measure terrain undulation; is the dynamic adjustment coefficient, which is optimized by gradient descent method according to real-time weather conditions;

[0099] Message passing mechanism: the first layer node embedding update formula is:

[0100] ;

[0101] wherein, LeakyReLU activation function, neighbor feature mean aggregation function, and the first layer trainable parameter matrix, the neighbor node set of node In this embodiment, the neighbor node set is the node within a spatial neighborhood radius of 5 km, and it is generally considered that the meteorological elements within a radius of 5 km have strong spatial correlation; denotes the edge weight between node u and node v; denotes the feature embedding of the l-th layer node u; denotes the feature embedding of the l-th layer node v; denotes the updated feature embedding of the l+I-th layer node v;

[0102] Model training: The FLUXNET 2025 global flux observation data and the corresponding multi-scale auxiliary data are used to construct a training set, and the ground flux tower measured value is used as a label. The model prediction error (root mean square error RMSE) is minimized by using the Adam optimizer. The training set and the validation set are divided according to the ratio of 8:2. Batch processing (batchsize=128) is used for training, and 100 iterations are performed. After each round, the RMSE is evaluated on the validation set. When the RMSE of the validation set does not decrease for 10 consecutive rounds, the early stopping mechanism is triggered, and the optimal model parameters are saved;

[0103] Model output: The estimated value of the atmospheric pollutant flux of each grid, with a time resolution of 30 minutes;

[0104] S22, dynamic correlation calculation for pollutant flux: This step realizes dynamic correlation under physical constraints by simulating-evaluating-optimizing online dynamic cycle, so that the graph structure responds to the real-time atmospheric transmission state; as shown in Figure 3 The specific steps are as follows:

[0105] S221, initial correlation graph construction and reasoning:

[0106] The preprocessed and fused multi-scale pollutant concentration and meteorological field data are input into the trained DyGNN model, and the global graph structure is initialized wherein, is a node, is an edge, is an initial edge weight matrix;

[0107] The multi-scale pollutant concentration and meteorological field data are organized in the form of a structured multi-dimensional tensor, with the dimension defined as [time dimension x spatial dimension x variable dimension]. If the research area covers a time length of hours, after preprocessing, there are uniform spatial grids, including one pollutant and one meteorological variable, the data tensor shape is ;

[0108] where, the time dimension (T): corresponds to the minute number (or hour number), embodies the time scale; ;

[0109] the space dimension (S): corresponds to the total number of uniform resampled grid numbers, embodies the fusion of multiple spatial scales; ;

[0110] the variable dimension (V): corresponds to the total number of pollutant concentration variable numbers and meteorological field variable numbers; for example, if there are 5 pollutants and 6 meteorological parameters, then ;

[0111] Run forward inference to get the initial flux field where, the flux of node i is ,

[0112] where, is the node feature matrix;

[0113] S222, physical process feedback verification based on mass balance:

[0114] Verify the rationality of the initial flux field by discretizing the mass balance equation:

[0115] ;

[0116] where, is the physical residual, representing the difference between the initial flux value and the actual observation value, is the future time pollutant observation concentration, is the current pollutant concentration, represents the time interval, represents the pollutant flux estimated by the initial flux field, is the wind speed vector, is the chemical degradation coefficient, is the advection term;

[0117] Calculate the area-averaged physical residual:

[0118] ;

[0119] where, is the total number of nodes in the verification area;

[0120] S223, edge contribution degree calculation and graph optimization based on integral gradient: optimize the graph structure by quantifying the contribution of edges to high residual nodes;

[0121] Locate high residual nodes: identify nodes that satisfy ​Node combination of the application Wherein, is a high residual coefficient, which is determined according to the optimization fineness of the graph structure under the actual scene in the embodiment ;

[0122] Calculate the edge contribution: for the high residual node and the incoming edge , the contribution is calculated by integral gradient method :

[0123] ;

[0124] Wherein, is the current edge weight, is the baseline weight (usually 0), is the flux output along the weight path, is the partial derivative of the output to the edge weight;

[0125] The scalar calculated by the formula is directly proportional to the contribution of the edge to the node i pollutant flux, and the sign represents promotion or inhibition;

[0126] Dynamic optimization of graph structure: normalize the contribution by Sigmoid function, adjust the edge weight:

[0127] ;

[0128] Wherein, is the learning rate, usually 0.1~0.3, in the embodiment , by comparing the goodness of fit of pollutant flux under different , physical residual convergence speed and other indicators, select the which makes the validation set accuracy optimal, the goodness of fit is determined by the coefficient of determination as the evaluation index, the convergence speed takes the iteration number as the evaluation index, is the Sigmoid function;

[0129] Local forward inference is performed on and its neighborhood nodes to obtain the corrected flux ;

[0130] S224, output the pollutant flux distribution field: backfill the corrected flux globally to obtain , repeat S222~S223 until the convergence condition is met:

[0131] ;

[0132] Wherein, a convergence threshold, a number of iterations;

[0133] Finally, output the pollutant flux distribution field meeting physical consistency ;

[0134] Through the detailed introduction of the above-mentioned embodiments, the atmospheric flux estimation method based on multi-scale dynamic correlation of the application realizes efficient estimation through the progressive process of data preprocessing-physical fusion-AI correlation optimization: firstly, the three types of data of satellite remote sensing, ground observation and numerical simulation are time and space aligned and differentiated noise is suppressed, and the data basis deviation is eliminated; secondly, relying on the two-way tracing, a physical driving fusion mechanism is constructed, the data source quality is dynamically quantified and weighted fusion is performed, and the data basis is consolidated; finally, a cross-scale correlation is constructed by a dynamic graph neural network, the graph structure is optimized by combining quality balance verification and integral gradient method, and the precise flux distribution field is output after convergence judgment. Through the two core designs of two-way tracing and graph structure optimization, the method effectively solves the problems of large scale connection error and time deficiency of the prior art, and balances the accuracy and speed, providing reliable technical support for atmospheric pollution emergency response and regional carbon cycle monitoring.

[0135] The above formulas are all dimensionless numerical calculations, and the preset parameters in the formulas are set by a person skilled in the art according to the actual situation.

[0136] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired or wireless (such as infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center and the like containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD) or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0137] It should be understood that the magnitude of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0138] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0139] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0140] In several embodiments provided by the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or in other forms.

[0141] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the technical solutions that essentially contribute to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0142] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in 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 method for atmospheric flux estimation based on multiscale dynamic correlation, characterized in that, The method flow is as follows: Step one, spatiotemporal alignment and noise suppression of multi-scale observation data: S11, data collection: collect three types of multi-scale data sources, namely satellite remote sensing data, ground observation data and numerical simulation data; S12, spatiotemporal alignment: taking the ground flux tower observation timestamp as the benchmark, the satellite remote sensing data, numerical simulation data are time-aligned and space-aligned; S13, noise suppression: for different noise characteristics of satellite remote sensing data, ground observation data and numerical simulation data, differential denoising strategies are adopted; S14, physical driving multi-source data fusion: based on the bidirectional tracing of multi-source data adaptive confidence weighted fusion, for each grid point that has completed spatiotemporal alignment and denoising, the confidence weight of satellite remote sensing data, ground observation data and numerical simulation data is dynamically calculated to realize physical driving data fusion; specifically including three sub-steps: S141, uplink tracing: using the Lagrangian random particle diffusion model to quantify the spatial representative error of ground observation data and numerical simulation data; S142, downlink evaluation: based on real-time meteorological condition data quality diagnosis, the data quality is quantified through real-time meteorological elements, and the dynamic quality factor of the three types of multi-scale data sources is calculated; S143, dynamic weight synthesis and multi-source data fusion: based on the quantitative indexes of uplink tracing and downlink evaluation, the final fusion weight is calculated, and the fusion value of the target point data is calculated; Step two, construction of cross-scale dynamic correlation mechanism based on AI model: S21, AI model selection and training: selecting dynamic graph neural network as the correlation model, defining the graph node and edge weight function, taking the multi-scale data grid unit processed in step one as the graph node, designing the edge weight function based on atmospheric physical process, designing the message passing mechanism, constructing the training set and training the model; S22, dynamic correlation calculation for pollutant flux: based on the mass balance equation to verify the rationality of the initial flux field, the edge contribution degree is calculated by integral gradient method and the graph structure is optimized until the convergence condition is met, and the final pollutant flux distribution field is output.

2. The method of estimating atmospheric fluxes based on multiscale dynamic correlation according to claim 1, characterized in that, The specific process of the uplink tracing is as follows: Calculation of flux tower data representativeness: for the flux tower site, the under-surface representativeness score of the target grid point is calculated through the footprint function; Calculation of numerical simulation data representativeness: for the model grid, the representativeness error is measured by the under-surface heterogeneity; the model grid is the basic calculation unit for spatial discretization in numerical simulation.

3. The method for atmospheric flux estimation based on multiscale dynamic correlation according to claim 1, characterized in that, The specific process of the downlink evaluation is as follows: The satellite data quality factor is calculated in combination with the cloudiness ratio and the ground visibility; the numerical simulation data stability factor is calculated based on the Richardson number; the ground station flow field correlation factor is calculated in combination with the angle between the flux tower site and the prevailing wind direction and the horizontal attenuation.

4. The method of atmospheric flux estimation based on multiscale dynamic correlation according to claim 3, characterized in that, The specific process of the dynamic weight synthesis and multi-source data fusion is as follows: Firstly, the satellite remote sensing data confidence factor is calculated based on the satellite data quality factor, the ground observation data confidence factor is calculated based on the flux tower data representativeness and the ground station flow field correlation factor, and the numerical simulation data confidence factor is calculated based on the numerical simulation data representativeness and the numerical simulation data stability factor; Secondly, the minimum weight of the target point data is calculated based on three types of multi-scale data sources; Finally, the final fusion value of the target point data is calculated based on the initial fusion weight and the minimum weight.

5. The method for atmospheric flux estimation based on multi-scale dynamic correlation according to claim 1, characterized in that, The initial flux field is verified for reasonableness based on the mass balance equation, and the specific process is as follows: The physical residual of the initial flux value and the actual observation value is calculated by discretizing the mass balance equation, and the average physical residual in the region is calculated.

6. The method of atmospheric flux estimation based on multiscale dynamic correlation according to claim 5, characterized in that, The integral gradient method is used to calculate the edge contribution and optimize the graph structure, and the specific process is as follows: Locate high residual nodes: identify high residual nodes through the average physical residual in the region; Calculate the edge contribution: for high residual nodes and their incoming edges, calculate the contribution through the integral gradient method; Dynamic optimization of graph structure: normalize the contribution through the Sigmoid function, and adjust the edge weight through the following formula: ； wherein, is a learning rate, is a current edge weight, is a sigmoid function, is a high residual node and its edge contribution degree of the incoming edge.

7. The method for atmospheric flux estimation based on multi-scale dynamic correlation according to claim 1, characterized in that, In step S22, the convergence condition is as follows: ； wherein, is the first computed area average physical residual, is the second computed area average physical residual, is a convergence threshold, is the number of iterations.

Images