Multi-source cross-scale satellite data adaptive greenhouse gas product fusion method

Abstract

The invention provides a multi-source cross-scale satellite data adaptive greenhouse gas product fusion method, which belongs to the technical field of atmospheric environment remote sensing application, and comprises the following steps: carrying out space-time unified matching, resampling and reprojection on observation data of different satellite platforms; performing error correction by using high-resolution OCO-2 data as hard data, performing system deviation correction on low-resolution GOSAT data based on an XGBoost algorithm, and constructing soft data; the hard data and the soft data are fused through a Bayesian maximum entropy framework, the fusion weight is optimized by utilizing a space-time covariance model and soft data constraints, and a posterior probability density mean value is obtained to serve as a final fusion result. And a high-precision CO2 concentration field in a global range is obtained by combining a space-time large-scale trend. According to the method, the contribution of low-uncertainty observation can be enhanced, high noise and geometric mismatch interference are effectively suppressed, a global greenhouse gas concentration fusion product with multi-source information and an uncertainty field is generated, and the precision and space continuity of greenhouse gas monitoring are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric environment remote sensing application technology, specifically relating to an adaptive greenhouse gas product fusion method based on multi-source, cross-scale satellite data. Background Technology

[0002] Traditional methods for monitoring atmospheric CO2 concentrations primarily rely on ground-based observations. While ground-based monitoring stations typically offer high measurement accuracy, their limited number and uneven distribution make global coverage difficult. Therefore, ground-based observations struggle to effectively reflect the spatiotemporal distribution characteristics of global atmospheric CO2, especially in remote areas and over oceans, where the lack of data further restricts their application. In contrast, satellite remote sensing technology provides broad spatial coverage, offering the advantage of real-time, global-scale observation. Through satellite remote sensing, the spatiotemporal variation characteristics of atmospheric CO2 can be acquired more comprehensively and systematically, providing crucial data support for global climate change research and carbon cycle models, and enhancing the ability to monitor CO2 concentrations.

[0003] However, satellite remote sensing technology also faces certain challenges. Different satellite sensors exhibit variations in measurement accuracy and detection capabilities, and factors such as clouds and aerosols in the atmosphere significantly impact the quality of satellite remote sensing data. This results in substantial uncertainty and error in the spatial distribution of observation data from a single satellite sensor globally. Furthermore, the spatiotemporal resolution of existing satellite observation data is limited, making it difficult to meet the urgent need for high-resolution, high-precision atmospheric CO2 data in carbon cycle, climate modeling, and regional pollution research. Therefore, effectively integrating data from different satellite sensors to improve the spatiotemporal resolution and accuracy of atmospheric CO2 concentration observations has become a crucial topic in current scientific research.

[0004] To address these challenges, data fusion technology based on multi-source satellite remote sensing data has gradually become an important research direction in recent years. By fusing data from different satellite sensors, the advantages of each can be fully utilized, compensating for the shortcomings of a single sensor, thereby improving the monitoring accuracy and spatiotemporal resolution of atmospheric CO2 concentration. Furthermore, with the continuous development of advanced technologies such as artificial intelligence and machine learning, multi-source data fusion algorithms based on these technologies have shown great potential in atmospheric CO2 concentration monitoring. By establishing more accurate fusion models, high-precision, high-spatiotemporal-resolution atmospheric CO2 concentration monitoring can be achieved globally, promoting the development of global carbon cycle and climate change research.

[0005] Therefore, this invention aims to develop a fusion algorithm based on multi-source satellite CO2 spatiotemporal data, in order to overcome the limitations of existing satellite remote sensing observation technology and achieve efficient monitoring and accurate estimation of atmospheric CO2 concentration using existing multi-source satellite CO2 concentration data. Summary of the Invention

[0006] This invention relates to an adaptive greenhouse gas product fusion method based on multi-source, multi-scale satellite data. It aims to improve the spatiotemporal resolution and accuracy of atmospheric CO2 concentration by comprehensively utilizing multi-source data from different satellite platforms, thereby better supporting research on global climate change, carbon cycle, and environmental protection. This invention combines the advantages of the Bayesian maximum entropy principle and the Extreme Gradient Boosting (XGBoost) algorithm, proposing an efficient and adaptive greenhouse gas product fusion framework that overcomes the limitations of traditional methods and has significant scientific research and practical application value.

[0007] The present invention proposes an adaptive greenhouse gas product fusion method based on multi-source, multi-scale satellite data, which mainly includes the following steps:

[0008] S1. Acquisition and Resampling of Multi-Source Multi-Scale Satellite Observation Data

[0009] Acquire synchronous remote sensing observation datasets of greenhouse gases from different platforms such as OCO-2 and GOSAT, and perform unified spatiotemporal benchmark matching and resampling;

[0010] First, synchronous greenhouse gas remote sensing datasets are acquired from multiple satellite platforms (such as OCO-2 and GOSAT). Data preprocessing is performed on all satellite data, including removing observations contaminated by clouds / strong aerosols. Due to differences in spatial resolution and observation time among different satellite sensors, spatiotemporal matching, resampling, and reprojection of these data are necessary to ensure that multi-source data can be fused under the same spatiotemporal reference. Greenhouse gas data from both OCO-2 and GOSAT platforms need to be resampled into a regular grid format. The specific method is as follows:

[0011]

[0012] in For the original satellite pixels, This represents the area of ​​the intersection between the original pixel and the target regular grid. To account for the impact of missing satellite data values ​​on the area weighting calculation, The observation value of each satellite's original pixel, The area weighting value for the target grid.

[0013] A unified spatiotemporal reference is applied to all preprocessed satellite data, unifying multi-source data to the unified spatiotemporal reference of GOSAT satellite data.

[0014] S2, Building Hard and Soft Data

[0015] In the data fusion process, the construction of hard and soft data is crucial. Hard data is typically provided by high-resolution satellite observation data (such as OCO-2 data), which has high reliability in quality and is used for error correction and basic data support. First, spatiotemporal trend removal processing is performed on the hard data, and a spatiotemporal covariance model is established. Based on considerations of spatial and temporal neighborhoods, empirical model parameters are fitted and used as auxiliary information in subsequent data fusion. Specifically, this includes:

[0016] S2.1, Hard Data Construction:

[0017] Spatiotemporal trend removal processing was performed on the OCO-2 data, and a spatiotemporal covariance model in the form of a nested exponential function was established to fit the parameters of the empirical model. The spatial and temporal neighborhoods were comprehensively considered and used as auxiliary information for fusion.

[0018] S2.1.1: First, a spatiotemporal window needs to be determined to remove the average values ​​in both the spatial and temporal domains. The size of this window in the spatial domain is determined by the ratio 'a' of GOSAT to the spatial resolution (area scale) of OCO-2. Regarding the size of the temporal window, considering the OCO-2 revisit period of 16 days, this study ultimately determined the maximum temporal distance window to be 33 days. A spatiotemporal filter window of size a×b×c is defined.

[0019] S2.1.2: Perform spatiotemporal trend removal processing on the OCO-2 data, and then fit the empirical model parameters based on the spatiotemporal covariance; the formula is as follows:

[0020]

[0021] Where (x, y, t) are the corresponding pixel coordinates, where x, y, and t represent longitude, latitude, and time coordinates, respectively. a, b, and c represent the upper and lower limits of the summation window in the spatial dimensions, determining the size and position of the window.

[0022] S2.2, Soft Data Construction:

[0023] The XGBoost machine learning method of extreme gradient boosting is used to correct systematic bias. Historical data of simultaneous observation of OCO-2 and GOSAT are selected. Based on the data of the overlapping period of OCO-2 and GOSAT, the XGBoost algorithm is applied to remove systematic bias and obtain corrected GOSAT data. Spatiotemporal trend removal processing is performed on the corrected GOSAT-2 data, and the processed GOSAT data is used as soft data.

[0024] Specifically, it includes:

[0025] S2.2.1: Based on the overlap period data of OCO-2 and GOSAT, extract image data synchronized with OCO-2 observations from daily GOSAT images, and resample to the same spatial resolution as OCO-2 using the nearest neighbor method.

[0026] S2.2.2: Based on the resampled GOSAT data obtained in step one, the XGBoost algorithm is applied to remove the systematic bias of the GOSAT data, as shown in the following formula:

[0027]

[0028]

[0029] in, This represents the gradient of the left branch. This represents the second derivative (Hessian) of the left-hand branch. This represents the gradient of the right-hand branch. This represents the second derivative (Hessian) of the right-hand branch. This represents the regularization term, used to prevent overfitting. This represents the reduction term, used to control the gain (usually adjusted during tree splitting). The total number of samples, This is the actual value (real label). This is the predicted value (model output). The mean squared error (MSE) is used for calculation, which measures the model's prediction error.

[0030] S2.2.3: Based on the corrected GOSAT data obtained in step two, spatiotemporal trend removal processing is performed on the corrected GOSAT data, and the cumulative probability density function of the two datasets is calculated pixel by pixel. The processed GOSAT data is used as soft data, as shown in the following formula:

[0031]

[0032] in, The probability density function expression for soft data. This is background information reflected in low spatial resolution GOSAT imagery; It is the variance of the error calculated using the XGBoost algorithm after resampling the low-resolution GOSAT image.

[0033] S3, BME Spatiotemporal Fusion

[0034] BME spatiotemporal fusion is one of the core technologies of this invention. The specific process is as follows:

[0035] S3.1, Prior Stage: Utilizing the maximum entropy principle, a joint probability distribution is constructed through a spatiotemporal covariance model to obtain the prior probability density function under Lagrange constraints. This stage provides prior knowledge for subsequent fusion. The specific method is as follows:

[0036]

[0037]

[0038] The Lagrange constraint is shown in the following formula:

[0039]

[0040] in, This represents all data values, including hard data, soft data, and data points to be estimated; It is the prior probability density function (PDF) of generalized knowledge / data for the entire study area; It is a Lagrange operator. It is an information constraint, represented by the mean and covariance functions; The number of vectors representing the constraints; This represents the regularization parameter for each category; Indicates category α in position Expected value at that location; Indicates the L function pair The loss; Indicates the location of each category Characteristic function at; Indicates the location of each category The expected value at that location.

[0041] S3.2, Posterior Stage: Using OCO-2 data as hard data and corrected GOSAT data as soft data, a Bayesian maximum entropy update is performed to obtain the posterior probability distribution. This process combines hard data, soft data, and prior knowledge to achieve effective information fusion among multi-source data. The specific method is as follows:

[0042] The posterior mean is taken as the final fusion estimate and superimposed with the spatiotemporal large-scale trend to obtain the complete XCO2 field;

[0043]

[0044]

[0045] in, Indicates to The weighted mean; Representation function exist The value at; Represents the weighted variance; Represents the coordinates of the sample points; express The mean.

[0046] S3.3, Estimation and Reconstruction: By taking the posterior mean as the final fusion estimate and superimposing it with large-scale spatiotemporal trends, a complete XCO2 field is obtained. This fusion result provides a more accurate spatiotemporal distribution of CO2 concentration and has global-scale adaptability.

[0047] S4. Evaluation of Fusion Results

[0048] To ensure the effectiveness of the fusion method, this invention proposes a multi-dimensional evaluation scheme:

[0049] S4.1 Accuracy Verification: By comparing data from ground observation stations such as TCCON station by station, the correlation coefficient (R) and root mean square error (RMSE) are calculated to verify the accuracy of the fusion results. Specifically, this includes:

[0050] The fusion results were cross-validated with data from ground-based observation stations such as TCCON to demonstrate the reliability and applicability of the fusion method. The results are as follows:

[0051]

[0052] The CO2 value predicted by the representative model Compared with actual observed values The correlation is between 0 and 1, with a higher value indicating a stronger correlation.

[0053]

[0054] The CO2 value predicted by the representative model Compared with actual observed values The smaller the arithmetic square root of the mean of the sum of squared differences, the smaller the model error.

[0055] S4.2 Integrity Assessment: Statistically assess the effective coverage rate and compare the differences between multi-source data fusion and single satellite products to evaluate the integrity of the fusion results.

[0056] S4.3 Consistency Comparison: In areas where single satellite data is missing or multi-source satellite data overlap, the fusion result is compared with the original product to verify the robustness of the fusion result.

[0057] The technical effects of this invention are as follows:

[0058] (1) High precision and high resolution: By combining hard data (high-resolution data) and soft data (low-resolution data after correction), this invention solves the problem of missing data on a global scale from a single satellite observation while maintaining data precision.

[0059] (2) Strong adaptability: The XGBoost algorithm is used for system bias correction, which can flexibly adapt to the data characteristics of different satellite sensors and automatically optimize the fusion process.

[0060] (3) Global applicability: This method is not only applicable to CO2 concentration monitoring, but can also be extended to the fusion of other greenhouse gas or environmental monitoring data, and has broad application prospects.

[0061] (4) Efficient fusion process: By introducing the Bayesian maximum entropy (BME) framework, multi-source data is effectively fused, the computational efficiency is optimized, and the needs of large-scale data processing are met.

[0062] The technical solution of this invention can be widely applied in fields such as atmospheric environment monitoring, global climate change research, carbon cycle model analysis, and greenhouse gas emission reduction policy evaluation. By providing high-precision, high spatiotemporal resolution atmospheric CO2 concentration data, it can provide important data support for climate change prediction, environmental protection, and policy decision-making.

[0063] The proposed improved and optimized Bayesian Maximum Entropy (XGBoost-BME) fusion method offers a novel solution to the problems of low spatiotemporal resolution, missing data, and uncertainty in global-scale greenhouse gas observations. By comprehensively utilizing multi-source satellite data and advanced machine learning algorithms, this invention can provide more accurate and comprehensive atmospheric CO2 concentration monitoring results, offering important technical support for climate change research and environmental protection. Attached Figure Description

[0064] Figure 1 This is a flowchart of the method for adaptive CO2 product fusion of multi-source cross-scale satellite data according to the present invention;

[0065] Figure 2 This is the original data image from which the spatiotemporal trend has been removed in this invention;

[0066] Figure 3 This is the detrended image after removing the spatiotemporal trend according to the present invention;

[0067] Figure 4 This is the present invention. Figure 3 Spatiotemporal covariance model diagram. Detailed Implementation

[0068] The specific embodiments of the present invention include the following steps:

[0069] like Figure 1 As shown, an adaptive greenhouse gas product fusion method based on multi-source, multi-scale satellite data mainly includes the following steps:

[0070] S1. Acquisition and Resampling of Multi-Source Multi-Scale Satellite Observation Data

[0071] Acquire synchronous remote sensing observation datasets of greenhouse gases from different platforms such as OCO-2 and GOSAT, and perform unified spatiotemporal benchmark matching and resampling;

[0072] First, synchronous greenhouse gas remote sensing datasets are acquired from multiple satellite platforms (such as OCO-2 and GOSAT). Data preprocessing is performed on all satellite data, including removing observations contaminated by clouds / strong aerosols. Due to differences in spatial resolution and observation time among different satellite sensors, spatiotemporal matching, resampling, and reprojection of these data are necessary to ensure that multi-source data can be fused under the same spatiotemporal reference. Greenhouse gas data from both OCO-2 and GOSAT platforms need to be resampled into a regular grid format. The specific method is as follows:

[0073]

[0074] in For the original satellite pixels, This represents the area of ​​the intersection between the original pixel and the target regular grid. To account for the impact of missing satellite data values ​​on the area weighting calculation, The observation value of each satellite's original pixel, The area weighting value for the target grid.

[0075] A unified spatiotemporal reference is applied to all preprocessed satellite data, unifying multi-source data to the unified spatiotemporal reference of GOSAT satellite data.

[0076] S2, Building Hard and Soft Data

[0077] In the data fusion process, the construction of hard and soft data is crucial. Hard data is typically provided by high-resolution satellite observation data (such as OCO-2 data), which has high reliability in quality and is used for error correction and basic data support. First, spatiotemporal trend removal processing is performed on the hard data, and a spatiotemporal covariance model is established. Based on considerations of spatial and temporal neighborhoods, empirical model parameters are fitted and used as auxiliary information in subsequent data fusion. Specifically, this includes:

[0078] S2.1, Hard Data Construction:

[0079] Spatiotemporal trend removal processing was performed on the OCO-2 data, and a spatiotemporal covariance model in the form of a nested exponential function was established to fit the parameters of the empirical model. The spatial and temporal neighborhoods were comprehensively considered and used as auxiliary information for fusion.

[0080] S2.1.1: First, a spatiotemporal window needs to be determined to remove the average values ​​in both the spatial and temporal domains. The size of this window in the spatial domain is determined by the ratio 'a' of GOSAT to the spatial resolution (area scale) of OCO-2. Regarding the size of the temporal window, considering the OCO-2 revisit period of 16 days, this study ultimately determined the maximum temporal distance window to be 33 days. A spatiotemporal filter window of size a×b×c is defined.

[0081] S2.1.2: Perform spatiotemporal trend removal processing on the OCO-2 data to satisfy the second-order stationarity assumption required for subsequent modeling (mean and variance remain constant in the spatiotemporal domain); the formula is as follows:

[0082]

[0083] Where (x, y, t) are the corresponding pixel coordinates, where x, y, and t represent longitude, latitude, and time coordinates, respectively. a, b, and c represent the upper and lower limits of the summation window in the spatial dimensions, determining the size and position of the window.

[0084] S2.2, Soft Data Construction:

[0085] The XGBoost machine learning method of extreme gradient boosting is used to correct systematic bias. Historical data of simultaneous observation of OCO-2 and GOSAT are selected. Based on the data of the overlapping period of OCO-2 and GOSAT, the XGBoost algorithm is applied to remove systematic bias and obtain corrected GOSAT data. Spatiotemporal trend removal processing is performed on the corrected GOSAT-2 data, and the processed GOSAT data is used as soft data.

[0086] Specifically, it includes:

[0087] S2.2.1: Based on the overlap period data of OCO-2 and GOSAT, extract image data synchronized with OCO-2 observations from daily GOSAT images, and resample to the same spatial resolution as OCO-2 using the nearest neighbor method.

[0088] S2.2.2: Based on the resampled GOSAT data obtained in step one, the XGBoost algorithm is applied to remove the systematic bias of the GOSAT data, as shown in the following formula:

[0089]

[0090]

[0091] in, This represents the gradient of the left branch. This represents the second derivative (Hessian) of the left-hand branch. This represents the gradient of the right-hand branch. This represents the second derivative (Hessian) of the right-hand branch. This represents the regularization term, used to prevent overfitting. This represents the reduction term, used to control the gain (usually adjusted during tree splitting). The total number of samples, This is the actual value (real label). This is the predicted value (model output). The mean squared error (MSE) is used for calculation, which measures the model's prediction error.

[0092] S2.2.3: This invention uses explicit quantification of observational uncertainties to participate in fusion estimation, constructing Gaussian probabilistic soft data for GOSAT data. Specifically, it utilizes the accumulated observational uncertainties during resampling and systematic bias correction processes to characterize satellite inversion errors.

[0093]

[0094] in, The probability density function expression for soft data. It is the background mean information reflected in low spatial resolution GOSAT imagery; It is the variance of the error calculated using the XGBoost algorithm after resampling the low-resolution GOSAT image.

[0095] S2.3, Removal of Spatiotemporal Trends:

[0096] S2.3.1: Based on the hard and soft data obtained from S2.2 and S2.3, spatiotemporal trend removal processing is performed. A spatiotemporal covariance model in the form of a nested exponential function is established to fit the parameters of the empirical model. Spatial and temporal neighborhoods are comprehensively considered and used as auxiliary information for fusion. The specific operation is as follows:

[0097] S2.3.3.1: First, a spatiotemporal window needs to be determined to remove the average values ​​in both the spatial and temporal domains. The size of this window in the spatial domain is determined by the ratio 'a' of GOSAT to OCO-2 spatial resolution (area scale). Regarding the temporal window size, considering the OCO-2 revisit period of 16 days, this study ultimately determined the maximum temporal distance window to be 33 days. A spatiotemporal filter window of size a × b × c is defined.

[0098] S2.3.3.2: Perform spatiotemporal trend removal processing on both hard and soft data to satisfy the second-order stationarity assumption required for subsequent modeling (mean and variance remain constant in the spatiotemporal domain); the formula is as follows:

[0099]

[0100] Where (x, y, t) are the corresponding pixel coordinates, where x, y, and t represent longitude, latitude, and time coordinates, respectively. a, b, and c represent the upper and lower limits of the summation window in the spatial dimensions, determining the size and position of the window.

[0101] This invention uses a spatiotemporal window to estimate the trend term in CO2 data and then removes it, as shown below. Figure 2 and Figure 3 As shown: the original CO2 frequency distribution is mainly concentrated around 420–426, and the distribution pattern is obviously asymmetric / skewed and affected by large-scale spatiotemporal background changes. Figure 2 After de-temporal trend processing, the CO2 residual distribution changed from its original skewed characteristics to an approximately symmetrical, approximately normal shape centered around 0. Figure 3 Meanwhile, the systematic mean structure in the original data is effectively removed, and the residual fluctuations are more representative of local random disturbances. This result shows that the detrending step significantly weakens spatiotemporal nonstationarity, providing a more reasonable second-order stationarity premise for subsequent spatiotemporal correlation structure characterization and statistical modeling in the residual domain (such as covariance / variogram estimation and fusion inversion).

[0102] S3, Spatiotemporal Covariance Modeling

[0103] After removing the spatiotemporal trend and obtaining stationary residuals, this invention characterizes the spatial and temporal correlation structure of AOD by constructing a spatiotemporal covariance function.

[0104] S3.1 Calculate the spatiotemporal covariance of the experiment based on the detrended joint AOD data, and analyze its decay characteristics as spatial distance and time interval increase;

[0105] S3.2. A nested exponential spatiotemporal covariance model is used to fit the experimental covariance to characterize multi-scale spatiotemporal variability. The least squares method is used to estimate the model parameters, and the search radius of the spatial and temporal neighborhoods in the BME fusion process is determined accordingly, thus providing a key prior statistical structure for subsequent Bayesian maximum entropy inversion. The specific formula is as follows:

[0106]

[0107] Spatiotemporal covariance function, representing the distance of CO2 in a given space. With time interval The covariance value under the given conditions. Spatial lag distance represents the Euclidean distance between two pixels in two-dimensional spatial coordinates. Time lag: The time difference (in days) between two observation times. (x, y, t): Spatial and temporal coordinates of the pixel to be estimated. (x′, y′, t′): Spatial and temporal coordinates of the observed pixels in the neighborhood. , The partial sill values ​​corresponding to the two exponential covariance models reflect the intensity of variation at different scales. , The spatial correlation scale (spatial range) of the two exponential models represents the feature distance where the spatial correlation decays to near zero. , The time correlation scale (time range) of the two index models represents the characteristic time length where time correlation is significant.

[0108] Figure 4 This paper presents the experimental covariance of detrended CO2 data and the relationship between its fitted covariance model and lag distance. The results show that CO2 exhibits significant spatiotemporal correlation within small lag scales, and the covariance rapidly decays with increasing lag distance, eventually approaching zero, indicating a limited effective correlation range for CO2. The fitted covariance model effectively characterizes the overall decay trend of the experimental covariance, demonstrating its ability to reflect the main spatiotemporal correlation structure of CO2. Although some dispersion exists in the experimental covariance at extremely small lag scales, it does not affect the judgment of the overall correlation characteristics. This result provides a reasonable prior covariance basis for subsequent BME spatiotemporal fusion of CO2.

[0109] S4, BME Spatiotemporal Fusion

[0110] BME spatiotemporal fusion is one of the core technologies of this invention. The specific process is as follows:

[0111] S4.1, Prior Stage: Utilizing the maximum entropy principle, a joint probability distribution is constructed through a spatiotemporal covariance model to obtain the prior probability density function under Lagrange constraints. This stage provides prior knowledge for subsequent fusion. The specific method is as follows:

[0112]

[0113]

[0114] The Lagrange constraint is shown in the following formula:

[0115]

[0116] in, This represents all data values, including hard data, soft data, and data points to be estimated; It is the prior probability density function (PDF) of generalized knowledge / data for the entire study area; It is a Lagrange operator. It is an information constraint, represented by the mean and covariance functions; The number of vectors representing the constraints; This represents the regularization parameter for each category; Indicates category α in position Expected value at that location; Indicates the L function pair The loss; Indicates the location of each category Characteristic function at; Indicates the location of each category The expected value at that location.

[0117] S4.2, Posterior Stage: Using OCO-2 data as hard data and corrected GOSAT data as soft data, a Bayesian maximum entropy update is performed to obtain the posterior probability distribution. This process combines hard data, soft data, and prior knowledge to achieve effective information fusion among multi-source data. The specific method is as follows:

[0118] The posterior mean is taken as the final fusion estimate and superimposed with the spatiotemporal large-scale trend to obtain the complete XCO2 field;

[0119]

[0120]

[0121] in, Indicates to The weighted mean; Representation function exist The value at; Represents the weighted variance; Represents the coordinates of the sample points; express The mean.

[0122] S4.3, Estimation and Reconstruction: By taking the posterior mean as the final fusion estimate and superimposing it with large-scale spatiotemporal trends, a complete XCO2 field is obtained. This fusion result provides a more accurate spatiotemporal distribution of CO2 concentration and has global-scale adaptability.

[0123] The fusion results inherit the high-precision characteristics of OCO-2 observations in terms of overall spatial pattern, while significantly improving the discontinuities in spatial coverage caused by single-satellite inversion. XCO2 exhibits a reasonable spatial gradient distribution within the study area, with smooth transitions between high and low value regions, and no obvious striping effects or abrupt boundary changes, indicating that the fusion process effectively suppressed the influence of single-source observation noise. Compared with the original sparse distribution of GOSAT and the orbital-type observations of OCO-2, the fused XCO2 mean field is more spatially continuous, fully utilizing the soft constraint information provided by GOSAT in areas lacking OCO-2 data, thereby supplementing the background variation characteristics at the regional scale. This demonstrates that the BME method successfully introduces statistical information from soft data while maintaining the dominant constraints of hard data, achieving a balance between physical rationality and spatial integrity in the fusion results.

[0124] In regions with dense OCO-2 orbital coverage, the overall uncertainty is low, reflecting the strong constraint of hard data on the fusion results; while in regions where only GOSAT provides information or observations are sparse, the uncertainty is relatively high, but still maintains a continuous and smooth spatial distribution.

[0125] The uncertainty distribution indicates that the BME fusion method can adaptively adjust the estimation confidence level based on the relative information content of soft and hard data: the fusion result is more reliable in regions dominated by high-quality hard data; in regions with insufficient observations, the uncertainty is reasonably increased, thus avoiding overly optimistic estimations of the result accuracy. Overall, the uncertainty field and the mean field have good spatial consistency and statistical self-consistency, further verifying the reliability of the fusion result.

[0126] S4. Evaluation of Fusion Results

[0127] To ensure the effectiveness of the fusion method, this invention proposes a multi-dimensional evaluation scheme:

[0128] S4.1 Accuracy Verification: By comparing data from ground observation stations such as TCCON station by station, the correlation coefficient (R) and root mean square error (RMSE) are calculated to verify the accuracy of the fusion results. Specifically, this includes:

[0129] The fusion results were cross-validated with data from ground-based observation stations such as TCCON to demonstrate the reliability and applicability of the fusion method. The results are as follows:

[0130]

[0131] The CO2 value predicted by the representative model Compared with actual observed values The correlation is between 0 and 1, with a higher value indicating a stronger correlation.

[0132]

[0133] The CO2 value predicted by the representative model Compared with actual observed values The smaller the arithmetic square root of the mean of the sum of squared differences, the smaller the model error.

[0134] S4.2 Integrity Assessment: Statistically assess the effective coverage rate and compare the differences between multi-source data fusion and single satellite products to evaluate the integrity of the fusion results.

[0135] S4.3 Consistency Comparison: In areas where single satellite data is missing or multi-source satellite data overlap, the fusion result is compared with the original product to verify the robustness of the fusion result.

Claims

1. A method for adaptive greenhouse gas product fusion from multi-source, multi-scale satellite data, characterized in that, Includes the following steps: Acquisition and resampling of multi-source, multi-scale satellite observation data; construction of hard / soft data; spatiotemporal fusion of BME; specifically: S1. Acquisition and resampling of multi-source, multi-scale satellite observation data; Acquire synchronous remote sensing observation datasets of greenhouse gases from different platforms such as OCO-2 and GOSAT, and perform unified spatiotemporal benchmark matching and resampling; S2, Constructing hard data and soft data; S2.1, Constructing hard data; Spatiotemporal trend removal processing was performed on the OCO-2 data, and a spatiotemporal covariance model in the form of a nested exponential function was established to fit the parameters of the empirical model. The spatial and temporal neighborhoods were comprehensively considered and used as auxiliary information for fusion. S2.2, Building Soft Data This paper utilizes the XGBoost machine learning method to correct systematic bias. Historical observation data from simultaneous observations of OCO-2 and GOSAT are selected. Based on the overlap period data of OCO-2 and GOSAT, the XGBoost algorithm is applied to remove systematic bias, resulting in corrected GOSAT data. The data undergoes spatiotemporal trend removal processing, and the processed GOSAT data is used as soft data. S3 and BME spatiotemporal fusion specifically include: S3.1, Prior Stage: Using the maximum entropy principle, the prior probability density function under the Lagrange constraint is obtained; the joint probability distribution is established from the spatiotemporal covariance; S3.2, Posterior Stage: The OCO-2 data is used as hard data, and combined with the soft data constructed in S2.2, Bayesian maximum entropy update is performed to obtain the posterior probability distribution; S3.3 Estimation and Reconstruction: The posterior mean is taken as the final fusion estimate and superimposed with the spatiotemporal large-scale trend to obtain the complete XCO2 field; S4. Evaluation of fusion results, specifically including: S4.1 Accuracy Verification: Compare station by station with TCCON, calculate correlation coefficient R, root mean square error, and system bias; S4.2 Integrity Assessment: Statistically assess effective coverage, and compare the differences between multi-satellite fusion and single-satellite products; S4.3 Consistency Comparison: In areas where single-star data is missing or multiple-star data overlap, the fusion results are compared with the original product to verify its robustness.

2. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 1, characterized in that, The acquisition and resampling of multi-source, multi-scale satellite observation data in S1 specifically includes: S1.1: Acquisition of Multi-Source Satellite Data Obtain synchronous remote sensing observation datasets of greenhouse gases from different platforms such as OCO-2 and GOSAT through official platforms; S1.2: Data Preprocessing All satellite data were preprocessed, including removing observations contaminated by clouds / strong aerosols; S1.3: Data Resampling Greenhouse gas data from both OCO-2 and GOSAT platforms need to be resampled into a regular grid format; the specific method is as follows: ； in For the original satellite pixels, This represents the area of ​​the intersection between the original pixel and the target regular grid. To account for the impact of missing satellite data values ​​on the area weighting calculation, The observation value of each satellite's original pixel, Weight the area of ​​the target grid; S1.4: Data Spatiotemporal Registration A unified spatiotemporal reference is applied to all preprocessed satellite data, unifying multi-source data to the unified spatiotemporal reference of GOSAT satellite data.

3. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 1, characterized in that, The hard data construction in S2.1 is specifically implemented as follows: S2.1.1: First, a spatiotemporal window is determined to remove the average values ​​in the spatial and temporal domains. The size of this window in the spatial domain is determined by the ratio 'a' of the spatial resolution of GOSAT to OCO-2. Regarding the size of the temporal window, considering that the OCO-2 revisit period is 16 days, the maximum time distance window is finally determined to be 33 days. A spatiotemporal filter window of size a×b×c is defined. S2.1.2: Perform spatiotemporal trend removal processing on the OCO-2 data, and then fit the empirical model parameters based on the spatiotemporal covariance; the formula is as follows: ； Where (x, y, t) are the corresponding pixel coordinates, x, y and t represent longitude, latitude and time coordinates respectively, and a, b and c represent the upper and lower limits of the summation window in the spatial dimensions, which determine the size and position of the window.

4. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 1, characterized in that, The soft data construction in S2.2 is specifically implemented as follows: S2.2.1: Based on the overlap period data of OCO-2 and GOSAT, extract image data that is synchronized with the OCO-2 observation from the GOSAT image and resample it to the same spatial resolution as OCO-2; S2.2.2: Based on the resampled GOSAT data, the XGBoost algorithm is applied to remove the systematic bias of the GOSAT data, as shown in the following formula: ； ； in, This represents the gradient of the left branch. This represents the second derivative of the left-hand branch. This represents the gradient of the right-hand branch. This represents the second derivative of the right-hand branch. This represents the regularization term, used to prevent overfitting. This indicates a reduction term, used to control gain. The total number of samples, This is the actual value. For predicted values, The mean squared error is used for calculation, which measures the prediction error of the model. S2.2.3: Based on the corrected GOSAT data, spatiotemporal trend removal processing is performed on the corrected GOSAT data, and the cumulative probability density function of the two datasets is calculated pixel by pixel. The processed GOSAT data is used as soft data, as shown in the following formula: ； in, The probability density function expression for soft data. This is background information reflected in low spatial resolution GOSAT imagery; It is the variance of the error calculated using the XGBoost algorithm after resampling the low-resolution GOSAT image.

5. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 1, characterized in that: S3.

1. In the prior stage, the principle of maximum entropy is used to obtain the prior probability density function under the joint probability distribution established by spatiotemporal covariance and Lagrange constraints. The specific method is as follows: ； ； The Lagrange constraint is shown in the following formula: ； in, This represents all data values, including hard data, soft data, and data points to be estimated; It is the prior probability density function PDF of generalized knowledge / data for the entire study area; It is a Lagrange operator. It is an information constraint, represented by the mean and covariance functions; The number of vectors representing the constraints; This represents the regularization parameter for each category; Indicates category α in position Expected value at that location; Indicates the L function pair The loss; Indicates the location of each category Characteristic function at; Indicates the location of each category The expected value at that location.

6. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 5, characterized in that: S3.2, Posterior Stage: The soft and hard data, along with the prior probability density function, are imported into the XGBoost-BME model to obtain the posterior probability distribution. The specific method is as follows: ； in, Indicates to The weighted mean; Representation function exist The value at; Represents the weighted variance; Represents the coordinates of the sample points; express The mean.

7. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 1, characterized in that: S3.3 Estimation and Reconstruction: The posterior mean is taken as the final fusion estimate and superimposed with the spatiotemporal large-scale trend to obtain the complete XCO2 field.

8. The method for adaptive greenhouse gas product fusion of multi-source, multi-scale satellite data according to claim 1, characterized in that: S4. Evaluation of fusion results, specifically including: The fusion results were cross-validated with TCCON site data to demonstrate the reliability and applicability of the fusion method; the formula is as follows: ； The CO2 value predicted by the representative model Compared with actual observed values The correlation between the two values ​​ranges from 0 to 1; the higher the value, the stronger the correlation. ； The CO2 value predicted by the representative model Compared with actual observed values The smaller the arithmetic square root of the mean of the sum of squared differences, the smaller the model error.

Images