Method for constructing high-resolution atmospheric carbon dioxide concentration data set based on XGBoost-BO

Abstract

The invention relates to a method for constructing a high-resolution atmosphere carbon dioxide concentration data set based on XGBoost-BO, and belongs to the technical field of environment monitoring and artificial intelligence modeling. The method comprises the following steps: preprocessing OCO-2 satellite data and multi-source auxiliary data, and fusing the preprocessed OCO-2 satellite data and multi-source auxiliary data to obtain a new data set; a Bayesian optimization method is adopted to search for an optimal hyper-parameter, a target function is optimized through second-order Taylor expansion, a regular term is introduced to control the complexity of the model, and ten-fold cross validation is used to evaluate the performance of the model; quantizing the contribution degree of each feature to model prediction through a tree SHAP method, and analyzing global feature importance ranking and feature contribution distribution of individual samples; and performing model verification by using the test set and the site actual measurement data. According to the method, the problems that an existing model-based reconstruction method is insufficient in interpretation and prone to falling into local optimum are solved, and the temporal-spatial resolution of CO2 concentration monitoring can be improved.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and artificial intelligence modeling technology, and relates to a method for constructing a high-resolution atmospheric carbon dioxide concentration dataset based on XGBoost-BO. Background Technology

[0002] As a major greenhouse gas, CO2 is a primary driver of global warming. Therefore, accurate monitoring of its concentration changes is crucial for scientifically assessing climate change trends and formulating emission reduction policies. In China, with rapid economic development and increased energy consumption, CO2 emissions have continued to rise, leading to increasingly serious environmental pollution and greenhouse effect problems. High-precision CO2 monitoring can not only provide the government with scientific evidence to help formulate more reasonable environmental protection policies, but also provide data support for enterprises in carbon emission management and carbon footprint analysis. Furthermore, high spatiotemporal resolution CO2 monitoring technology can help relevant personnel analyze the distribution of carbon dioxide in different regions and at different times in greater detail, providing real-time, dynamic monitoring data, thereby providing more accurate decision support for addressing climate change. Therefore, developing and using CO2 monitoring products with high spatiotemporal resolution is of significant strategic importance for understanding and responding to the dynamic changes in carbon dioxide emissions and achieving emission reduction targets.

[0003] Currently, the main methods for acquiring CO2 data are ground monitoring stations and satellite remote sensing. Ground monitoring stations are the traditional method for acquiring CO2 data. They continuously measure the CO2 concentration in the air using sensors installed at specific locations. However, they can only provide data within a limited geographical area and cannot cover global or large-scale dynamic changes. Satellite remote sensing technology acquires CO2 concentration data in the Earth's atmosphere using sensors mounted on satellites. Modern satellites (such as NASA's OCO-2) can monitor the distribution of carbon dioxide in the atmosphere on a global scale and with high spatiotemporal resolution. They can observe the spatial distribution and changing trends of CO2, providing important data support, especially for some hard-to-monitor areas (such as oceans and remote regions). Although modern satellites can provide global CO2 monitoring data, the spatiotemporal resolution of existing satellites is still limited. The orbits and observation frequencies of most satellites are limited, preventing them from conducting high-frequency dynamic monitoring.

[0004] There are two main methods commonly used to construct a full-coverage XCO2 dataset: one is interpolation methods, such as Kriging interpolation, for example, using GOSAT and SCIAMACHY satellite data to draw a global land CO2 distribution map using Kriging interpolation. However, the Kriging method has high requirements for data distribution. If the data points are sparse or unevenly distributed, the interpolation results may be unreliable. Moreover, Kriging interpolation relies on spatial variability models (such as Gaussian models, exponential models, etc.). If the spatial variability of the actual data does not match the assumed model, it may lead to large interpolation errors. Therefore, constructing a regression model to predict missing XCO2 values ​​by establishing relationships between satellite observation data and auxiliary variables such as meteorology, vegetation, and topography has become a better approach. However, existing model-based reconstruction methods suffer from insufficient interpretability and are prone to getting trapped in local optima, making it difficult to guarantee that the obtained parameter combinations can maximize the model's ability to capture complex nonlinear relationships and feature interactions. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for constructing a high-resolution atmospheric carbon dioxide concentration dataset based on XGBoost-BO. This method combines the XGBoost model with the Bayesian optimization method and further introduces the SHAP method to construct the XGBoost-BO model. The XGBoost-BO model is used to generate a daily atmospheric CO2 concentration dataset with high spatial resolution, thereby improving the spatiotemporal resolution of CO2 concentration monitoring and providing high-quality data support for carbon cycle research and greenhouse gas monitoring.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for constructing a high-resolution atmospheric carbon dioxide concentration dataset based on XGBoost-BO, comprising:

[0008] The OCO-2 satellite data and multi-source auxiliary data are preprocessed and then fused to obtain a new dataset.

[0009] The optimal hyperparameters of XGBoost are searched using Bayesian optimization, the objective function is optimized by second-order Taylor expansion, a regularization term is introduced to control the model complexity, and ten-fold cross-validation is used to evaluate the model performance in order to construct the XGBoost-BO model.

[0010] The contribution of each feature to the model prediction is quantified by the tree SHAP method, and the global feature importance ranking and the feature contribution distribution of individual samples are analyzed.

[0011] Model validation was performed using the test set and real-world site data.

[0012] Furthermore, the preprocessing of OCO-2 satellite data and multi-source auxiliary data includes: filtering data and excluding data with a quality flag of 1 to perform quality control on OCO-2 satellite data; constructing a 0.1°×0.1° grid and retaining only the points falling within the grid; grouping the spatially connected data according to the longitude and latitude corresponding to the grid; calculating the average value of XCO2 for each group to obtain gridded OCO-2 data.

[0013] Multi-source auxiliary data is interpolated based on the longitude, latitude, and time information of the gridded OCO-2 data to generate new data points based on longitude, latitude, and time.

[0014] All datasets are merged to obtain new data. The missing parts in the new data are filled with interpolation methods to construct a new dataset.

[0015] Furthermore, a Bayesian optimization method is used to search for the optimal hyperparameters of XGBoost. First, Bayesian optimization is used to randomly and uniformly select points throughout the entire region. Each time a point is added, the loop begins. After finding N candidate solutions, a Gaussian regression model is established. The posterior probability of each point is calculated through the Gaussian regression model, and the maximum point is found as the next point. The maximum value among all points is found as the optimal solution, i.e., the optimal hyperparameters.

[0016] Furthermore, the XGBoost model is trained using hyperparameters, the objective function is expanded using a second-order Taylor series, and the second derivative of the function is used to train the tree model. Simultaneously, the complexity of the tree model is incorporated as a regularization term into the optimization objective to improve the efficiency of the learning model. The objective function is shown below:

[0017]

[0018] In the formula, the first term on the right-hand side represents the sample error, n is the sample size, l(·) is a function that measures sample error (such as RMSE), and y i Let be the true value of the i-th OCO-2XCO2 sample. Ω(f) is the predicted value of the i-th OCO-2XCO2 sample, which is the point-by-point summation of the outputs of all M trees; m ) represents the regularization term; f m For the m-th tree;

[0019] The final training output of the XGBoost model is the sum of m tree models:

[0020]

[0021] In the formula, CAMS is the predicted value of OCO-2 data at a specific coordinate on a specific day. ij GOSATij MODIS ij ERA5 ij ODIAC ij VIIRS ij GFED ij All of these are auxiliary variables for a specific coordinate on a specific day.

[0022] Furthermore, the XGBoost model output from the training was evaluated using 10-fold cross-validation. The original dataset was divided into 10 equally sized subsets. Each time, one subset was selected as the validation set, and the remaining 9 subsets were used as the training set. This process was repeated 10 times to ensure that each subset served as a validation set. After all 10 training and validation cycles were completed, the results of each validation cycle were summarized, and the average performance metric of the model was calculated.

[0023] Furthermore, the analysis of global feature importance ranking involves first measuring the contribution of each feature of an individual in the training dataset to the model output using a tree SHAP model; then, the feature contributions are ranked according to the average of all samples.

[0024] Furthermore, model validation using test set and site-tested data includes: using 20% ​​of the new dataset as the test set, and using goodness of fit, root mean square error, and mean absolute percentage error to evaluate model accuracy;

[0025] Obtain the measured XCO2 value at the TCCON site and compare it with the XCO2 value predicted by the model to verify the accuracy of the prediction results.

[0026] The beneficial effects of this invention are as follows:

[0027] (1) By integrating satellite data with meteorological and surface auxiliary variables, and combining 0.1°×0.1° high-precision gridding and spatiotemporal interpolation technology, this invention effectively fills the data gaps and generates a daily CO2 concentration dataset with wider coverage and stronger spatiotemporal continuity, thus solving the problems of insufficient ground station coverage and spatiotemporal fragmentation of satellite data.

[0028] (2) This invention employs Bayesian optimization (BO) to globally search for XGBoost hyperparameters, which avoids getting trapped in local optima. By optimizing the objective function through second-order Taylor expansion and introducing a regularization term, the model's ability to capture complex nonlinear relationships is enhanced, while suppressing overfitting. Furthermore, the use of a ten-fold cross-validation mechanism further ensures the stability and generalization of the model's performance.

[0029] (3) This invention introduces the tree SHAP method to quantify the contribution of each feature to the prediction results, realizing the global feature importance ranking and the contribution distribution analysis of individual samples. It reveals the influence mechanism of key environmental factors on CO2 concentration and provides a transparent explanation for the model decision-making process, overcoming the defects of traditional "black box" models.

[0030] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0032] Figure 1 A flowchart illustrating the method for constructing a high-resolution atmospheric carbon dioxide concentration dataset based on XGBoost-BO provided in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of data preprocessing;

[0034] Figure 3 A flowchart of the model training process;

[0035] Figure 4 Flowchart for model interpretation and verification;

[0036] Figure 5 This represents the average change in the model output when different datasets are removed from the model.

[0037] Figure 6 This relates the contribution of a single feature to the model's output to its own value. Detailed Implementation

[0038] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0039] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0040] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0041] This invention provides a method for constructing a high-resolution atmospheric carbon dioxide concentration dataset based on XGBoost-BO. Building upon the XGBoost model, this invention introduces a Bayesian optimization method to construct a new XGBoost-BO model. By performing global probabilistic modeling and iterative search on key hyperparameters of XGBoost, the prediction accuracy and generalization ability of the model are improved. Furthermore, to enhance the interpretability of the model, the SHAP method is introduced. By calculating the SHAP value of each auxiliary variable in each sample, the relative importance of each feature to the daily average XCO2 prediction can be quantified at the global level, and the impact of different feature values ​​on the model output can be revealed at the local level. Finally, the accuracy is verified by comparison with test set and TCCON observation station data.

[0042] like Figure 1 As shown, the method includes the following steps:

[0043] Step 1) Data Preprocessing: OCO-2 satellite data undergoes quality control and spatial gridding to construct a 0.1°×0.1° resolution grid. Multi-source auxiliary data, including CAMS-EGG4, GOSAT, MODIS, ERA5, ODIAC, VIIRS, and GFED datasets, are integrated. Data fusion is achieved through spatiotemporal interpolation. The datasets used are shown in the table below:

[0044] Table 1

[0045]

[0046]

[0047] Specifically, such as Figure 2 As shown, the OCO-2 data is first subjected to quality control, filtering the data and excluding data with a quality flag (XCO2_quality_flag) of 1 (which usually indicates low data quality or invalidity).

[0048] Then, a 0.1°×0.1° grid is constructed, retaining only the points that fall within the grid. The spatially connected data is grouped according to the longitude and latitude corresponding to the grid, and the average XCO2 value is calculated for each group to obtain the gridded OCO-2 data.

[0049] Next, other auxiliary data (CAMS-EGG4, GOSAT, MODIS, ERA5, ODIAC, VIIRS, and GFED) are interpolated in their corresponding datasets based on the longitude, latitude, and time information of the gridded OCO-2 data to generate new data points according to longitude, latitude, and time. These data maintain a consistent grid scale with the OCO-2 data.

[0050] Finally, all datasets are merged to obtain new data, and missing parts in the new data are filled using interpolation methods.

[0051] Step 2) Train the model using XGBoost-BO: Use Bayesian optimization to search for the optimal hyperparameters of the XGBoost model, optimize the objective function using second-order Taylor expansion, introduce a regularization term to control model complexity, and use 10-fold cross-validation to evaluate model performance. For example... Figure 3 As shown, the training process is as follows:

[0052] ① First, use Bayesian optimization to randomly and uniformly select points throughout the entire region. Repeat this process for each new point, until N candidate solutions are found. Then, establish a Gaussian regression model. Calculate the posterior probability of each point using the Gaussian regression model and identify the maximum point as the next point. The maximum value among all points is the optimal solution, i.e., the optimal hyperparameters.

[0053] Compared to conventional grid search and random search, utilizing information from points that have already been searched can improve the speed of the optimization process and the quality of the results.

[0054] ②Then, using the hyperparameters obtained in step ①, the XGBoost model is trained. The core principle of the XGBoost algorithm is to perform a second-order Taylor expansion of the objective function and then use the second derivative of the function to train the tree model. Furthermore, the complexity of the tree model is incorporated as a regularization term into the optimization objective to improve the efficiency of the learning model.

[0055] like Figure 3As shown, a model is first trained using a CART tree. Then, in each iteration, a second-order Taylor expansion is performed on the loss function between the current prediction and the true value. Based on the gradient and curvature information, a new CART tree is fitted. This process is repeated until the exit condition is met. The final XGBoost model is the sum of models 1 to n. During training, the hyperparameters remain unchanged.

[0056] Assume there are m tree models in total, and the output of each tree model is defined as f. i Therefore, the final output of the XGBoost model is as follows:

[0057]

[0058] In the formula, This represents the predicted value of OCO-2 data at a specific coordinate on a given day, CAMS. ij GOSAT ij MODIS ij ERA5 ij ODIAC ij VIIRS ij GFED ij Each of these represents an auxiliary variable corresponding to a specific coordinate on a specific day.

[0059] Once the tree structure is fixed, the node where each sample falls is determined. At this point, the following objective function is defined to measure the error of XGBoost on the overall sample size:

[0060]

[0061] In the formula, the first term on the right-hand side represents the sample error, n represents the sample size, l(·) is a function that measures sample error (such as RMSE), and y i Let be the true value of the i-th OCO-2XCO2 sample. The predicted value for the i-th OCO-2XCO2 sample is the pointwise summation of the outputs of all M trees; Ω(f) on the right side of the equation. m ) is a regularization term, the purpose of which is to reduce the risk of overfitting. m Let m be the m-th tree.

[0062] The second item is processed as follows:

[0063]

[0064] CART trees are built one by one. When optimizing the M-th tree, the first M-1 trees have already been computed, meaning their regularization values ​​are already determined, so they can be ignored during optimization. The M-th tree is defined as follows:

[0065]

[0066] In the formula, γ and λ are both penalty terms, and T is the total number of leaf nodes in the tree. This definition can be understood as follows: on the one hand, a large T value indicates that the tree is relatively deep, and the probability of overfitting will increase, so γ is used for penalty; on the other hand, a large w value indicates that the tree will occupy a large proportion in the entire model, that is, the prediction result mainly depends on the tree, and the risk of overfitting will also increase, so λ needs to be used for penalty.

[0067] ③ Use 10-fold cross-validation. 10-fold cross-validation is a model evaluation method primarily used to verify the model's generalization ability and performance stability. This method divides the original dataset into 10 equally sized subsets. Each time, one subset is selected as the validation set, and the remaining nine subsets are used as the training set. This process is repeated 10 times to ensure that each subset serves as a validation set. After all 10 training and validation iterations are completed, the results of each validation are summarized, and the model's average performance metric is calculated. This makes the model's evaluation results more stable and avoids random errors caused by uneven data partitioning.

[0068] ④ Repeat steps ① to ③ above, and select the set of hyperparameters with the best evaluation results to train the XGBoost model to obtain the optimal model.

[0069] Step 3), as Figure 4 As shown, the SHAP model is used to interpret the model. The contribution of each feature to the XGBoost model prediction is quantified by the tree SHAP method, and the global feature importance ranking and the feature contribution distribution of individual samples are analyzed.

[0070] Global Feature Contribution: Tree SHAP first measures the contribution of each feature of an individual in the training dataset to the model output; then, the feature contributions are ranked according to the average of all samples. XGBoost, as a tree ensemble model, explicitly shows the segmentation contribution of each feature for each decision path, making SHAP computation more intuitive and efficient. See also Figure 5 The x-axis essentially represents the average change in the model's output when a feature is removed from the model. Features are ordered according to the absolute total value of their impact on the model.

[0071] Individual Feature Contribution: Since the tree SHAP value is derived through an individualized model interpretation method, an individualized interpretation can be obtained for each sample from the XGBoost model. The process is as follows: Take a sample, use the baseline as the "initial output", for each feature, simulate the output difference "if added vs. if not added" around that feature, and average this difference over all possible subsets to obtain its feature contribution value. Feature contribution value + baseline = actual predicted value of the sample. Figure 6 This demonstrates how the contribution of a single feature to the model output is affected by its value. Figure 6 In the diagram, the x-position of a point represents the influence of that feature on the result, and the color of the point represents the value of that feature.

[0072] Step 4) Validate the model using the test set and TCCON observation station data: using R... 2 The accuracy of the model was evaluated using three indicators: RMSE and MAPE, and the accuracy of the model's prediction results was verified using measured data from two TCCON sites in Hefei and Xianghe, China.

[0073] The evaluation is conducted using three commonly used metrics—goodness of fit (Rfit). 2 ), root mean square error (RMSE) and mean absolute percentage error (MAPE).

[0074] Goodness of fit (R) 2 R0 is used to evaluate the goodness of fit of a regression model, representing the degree to which the independent variable explains the variation in the dependent variable. Its value ranges from 0 to 1 (sometimes negative), with values ​​closer to 1 indicating a better model fit. Its characteristics are as follows: R0 2 The larger the R value, the better the model fit. When the model is perfectly accurate, R... 2 =1. If the model performs poorly, even worse than the average prediction, R = 1. 2 It may be a negative value. R 2 The calculation formula is as follows:

[0075]

[0076] The root mean square error (RMSE) measures the deviation between predicted and actual values, reflecting the accuracy of the model's predictions. It ranges from 0 to +∞; a smaller RMSE value indicates a more accurate prediction. Its characteristics include a larger penalty for errors and greater susceptibility to outliers. It is commonly used to evaluate the accuracy of regression model predictions. The RMSE value has the same units as the actual value, facilitating practical interpretation. The formula for calculating RMSE is as follows:

[0077]

[0078] Mean Absolute Percentage Error (MAPE) is a commonly used metric for evaluating regression models, representing the relative magnitude of the error between predicted and actual values. It calculates the relative error of each predicted value and averages the results to reflect the percentage of the model's prediction error. A value of 0% indicates a perfect prediction (i.e., the predicted value equals the actual value), while a value greater than 0% indicates the presence of error. A higher MAPE value indicates a larger model error. MAPE is asymmetric; underestimating the error (predicted value less than the actual value) and overestimating the error (predicted value greater than the actual value) have different effects. The formula for calculating MAPE is as follows:

[0079]

[0080] Using 20% ​​of the dataset as the test set, we comprehensively tested R... 2 The accuracy of the model is evaluated using three metrics: RMSE, MAPE, and RMSE.

[0081] Subsequently, the accuracy of the prediction results was verified by comparing the XCO2 values ​​from two TCCON stations in China, Hefei and Xianghe, with the XCO2 values ​​predicted by the model. Furthermore, the CAMS dataset was compared with the values ​​from the two TCCON stations to contrast with the model. The CAMS dataset is a dataset of atmospheric carbon dioxide concentration at 0.75°C; comparing the CAMS dataset with the station values ​​verified the accuracy of the dataset resolution obtained during model training.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a high-resolution atmospheric carbon dioxide concentration dataset based on XGBoost-BO, characterized in that, The method includes: The OCO-2 satellite data and multi-source auxiliary data are preprocessed and then fused to obtain a new dataset. The optimal hyperparameters of XGBoost are searched using Bayesian optimization, the objective function is optimized by second-order Taylor expansion, a regularization term is introduced to control the model complexity, and ten-fold cross-validation is used to evaluate the model performance in order to construct the XGBoost-BO model. The contribution of each feature to the model prediction is quantified by the tree SHAP method, and the global feature importance ranking and the feature contribution distribution of individual samples are analyzed. Model validation was performed using the test set and real-world site data.

2. The method according to claim 1, characterized in that, Preprocessing of OCO-2 satellite data and multi-source auxiliary data includes: filtering data and excluding data with a quality flag of 1 to perform quality control on OCO-2 satellite data; constructing a 0.1°×0.1° grid and retaining only the points falling within the grid; grouping the spatially connected data according to the longitude and latitude corresponding to the grid; calculating the average value of XCO2 for each group to obtain gridded OCO-2 data. Multi-source auxiliary data is interpolated based on the longitude, latitude, and time information of the gridded OCO-2 data to generate new data points based on longitude, latitude, and time. All datasets are merged to obtain new data. The missing parts in the new data are filled with interpolation methods to construct a new dataset.

3. The method according to claim 2, characterized in that, The optimal hyperparameters of XGBoost are searched using the Bayesian optimization method. First, Bayesian optimization is used to randomly and uniformly select points in the entire region. Each time a point is added, the loop is started. After finding N candidate solutions, a Gaussian regression model is established. The posterior probability of each point is calculated through the Gaussian regression model, and the maximum point is found as the next point. The maximum value among all points is found as the optimal solution, i.e., the optimal hyperparameters.

4. The method according to claim 3, characterized in that, The XGBoost model is trained using hyperparameters. A second-order Taylor expansion of the objective function is performed, and the second derivative of the function is then used to train the tree model. Simultaneously, the complexity of the tree model is incorporated as a regularization term into the optimization objective to improve the efficiency of the learning model. The objective function is shown below: In the formula, the first term on the right-hand side represents the sample error, n is the sample size, l(·) is a function that measures sample error (such as RMSE), and y i Let be the true value of the i-th OCO-2XCO2 sample. The predicted value for the i-th OCO-2XCO2 sample is the point-by-point summation of the outputs of all M trees; Ω(f m ) represents the regularization term; f m For the m-th tree; The final training output of the XGBoost model is the sum of m tree models: In the formula, CAMS is the predicted value of OCO-2 data at a specific coordinate on a specific day. ij GOSAT ij MODIS ij ERA5 ij ODIAC ij VIIRS ij GFED ij All of these are auxiliary variables for a specific coordinate on a specific day.

5. The method according to claim 4, characterized in that, The XGBoost model was evaluated using 10-fold cross-validation. The original dataset was divided into 10 equally sized subsets. One subset was selected as the validation set each time, and the remaining 9 subsets were used as the training set. This process was repeated 10 times to ensure that each subset served as a validation set. After all 10 training and validation cycles were completed, the results of each validation cycle were summarized, and the average performance metric of the model was calculated.

6. The method according to claim 1, characterized in that, The analysis of global feature importance ranking includes first measuring the contribution of each feature of an individual in the training dataset to the model output using a tree SHAP model; Then, the feature contributions are sorted according to the average of all samples.

7. The method according to claim 1, characterized in that, Model validation using test set and site-tested data includes: using 20% ​​of the new dataset as the test set and evaluating model accuracy using goodness of fit, root mean square error, and mean absolute percentage error. Obtain the measured XCO2 value at the TCCON site and compare it with the XCO2 value predicted by the model to verify the accuracy of the prediction results.

Images