Method for predicting high-resolution monthly-scale atmospheric CO2 concentration based on ERT model

Abstract

The invention provides a method for predicting high-resolution monthly-scale atmospheric CO2 concentration based on an ERT model, and the method comprises the steps: carrying out the space-time matching of CO2 satellite column concentration, CO2 foundation observation data, meteorological data, vegetation data, topographic data, night light data and other environment variable data, and forming a data set for the construction of the ERT model; taking the XCO2 as a target variable and the environment variable data as an explanatory variable to train a model, and performing parameter optimization on the ERT model; generating a high-resolution atmospheric CO2 concentration data set for predicting the high-resolution atmospheric CO2 concentration; and separating man-made carbon emissions by means of XCO2 anomalies on the basis of the data set. According to the method, the meteorological data, the vegetation data, the topographic data and the night light data are combined, so that the defect that the atmospheric CO2 concentration is predicted only by combining natural environment variable data in the past is overcome, and meanwhile, artificial CO2 emission can also be obtained.

Description

Technical Field

[0001] The present invention relates to the technical fields of remote sensing technology, atmospheric science, geographical science and other disciplines, and in particular to a method for predicting high-resolution monthly-scale atmospheric CO2 concentration based on an ERT model. Background Art

[0002] Monitoring atmospheric CO2 concentration is crucial to responding to global warming and reducing carbon dioxide emissions. It is also one of the hot issues in global change research and is of great significance.

[0003] Currently, there are two main methods for monitoring atmospheric CO2 concentrations: traditional ground-based observations and satellite remote sensing. Traditional ground-based observations, such as the Global Carbon Column Observing Network (TCCON), the Global Atmosphere Watch (GAW), and the World Data Centre for Greenhouse Gases (WDCGG), offer high accuracy. However, due to the small number and uneven distribution of ground-based monitoring sites, comprehensive monitoring is not possible (Lu Lijiang, 2019). In contrast, satellite remote sensing offers advantages such as wide coverage, short acquisition cycles, fast updates, and fewer restrictions. It can provide stable, long-term, and comprehensive coverage of atmospheric CO2 concentrations (Liu et al., 2021). Currently, widely used CO2 monitoring satellites include GOSAT / GOSAT-2 launched by Japan, the Orbiting Carbon Observatory (OCO-2 / OCO-3) launched by the United States, and the TanSat and GF-5 satellites launched by China (Liu et al., 2022).

[0004] Although satellite observations offer many advantages over ground-based observations, they still have some limitations. For example, the low spatial resolution and monitoring gaps in satellite CO2 data make it impossible to comprehensively and continuously monitor CO2 concentration changes. Currently, there are two main methods for filling the gaps in satellite inversion and reconstructing high-resolution XCO2 data: traditional spatial interpolation and machine learning algorithms based on regression models. Traditional spatial interpolation methods simply invert CO2 concentration values based on existing satellite observations. Although spatial interpolation can fill data gaps, due to data sparseness, the interpolation results often exhibit orbit-related striping characteristics and low resolution (Zhang et al., 2024). Therefore, this method has uncertainties in estimating atmospheric CO2 concentration.

[0005] In recent years, the complexity of CO2 transport across the ocean-land-air ecosystem has led to poor linear model results. However, machine learning, leveraging its powerful ability to fit nonlinear relationships, can combine natural and anthropogenic data to perform simulations and predictions, filling in the gaps in satellite observations (He et al., 2023). There are three main approaches to simultaneously measure anthropogenic carbon emissions: the first is to isolate anthropogenic CO2 emissions by calculating regional CO2 background values and further calculating CO2 anomalies; the second is to quantitatively estimate anthropogenic CO2 concentrations by calculating the difference between anthropogenic emission areas and clean background areas; and the third is to use machine learning to study the estimation of anthropogenic CO2 emissions.

[0006] Therefore, it is of great significance to develop a method based on the ERT model to predict high-resolution monthly atmospheric CO2 concentration. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model, which can predict atmospheric CO2 concentration from a fine temporal and spatial scale, and separate anthropogenic CO2 concentration based on the atmospheric CO2 concentration to achieve carbon emission reduction.

[0008] To achieve the above object, the present invention provides a method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model, the method comprising:

[0009] Obtaining CO2 concentration observation data within the study area, wherein the CO2 concentration observation data includes OCO-2XCO2 data and TCCON data, and the OCO-2XCO2 data is used as dependent variable data input into the ERT model;

[0010] Obtain several environmental variables in the study area as independent variable data input into the ERT model;

[0011] All data are preprocessed. After preprocessing, the independent variable data and the dependent variable data are temporally and spatially matched to the target grid scale to obtain temporally and spatially matched data, and the temporally and spatially matched data are used as the XCO2 training data set;

[0012] Constructing an ERT model, expressing the independent variable data and the dependent variable data in the XCO2 training data set in the form of equations and inputting them into the ERT model for training to obtain a trained ERT model, and outputting an XCO2 predicted value from the trained ERT model;

[0013] According to the TCCON data, the XCO2 prediction value is evaluated based on three common indicators, namely, the coefficient of determination R 2, root mean square error RMSE and mean absolute error MAE. If the XCO2 predicted value passes the evaluation, the XCO2 abnormal concentration is calculated according to the XCO2 predicted value, and the artificial CO2 concentration is separated according to the calculated XCO2 abnormal concentration.

[0014] Furthermore, the several environmental variables in the study area are specifically: meteorological data, vegetation data, night light data and elevation data.

[0015] Furthermore, the specific process of preprocessing all the data is: using bilinear interpolation to uniformly resample all the data to the same spatial resolution of the target grid scale, and uniformly resample the data to the same time resolution according to the time extraction method.

[0016] Furthermore, the spatiotemporal matching of the independent variable data and the dependent variable data specifically includes: spatiotemporal scale conversion; wherein the spatiotemporal scale conversion includes spatiotemporal scale conversion of OCO-2XCO2 data, spatiotemporal scale conversion of vegetation data, spatial scale conversion of night light data, and spatial scale conversion of elevation data;

[0017] Furthermore, the spatiotemporal scale conversion of the OCO-2XCO2 data is specifically as follows: the acquired OCO-2XCO2 data is converted from daily scale data to monthly scale data according to the time extraction method, and then the average value of the XCO2 value within the target grid scale is calculated; the scale conversion of the vegetation data is specifically as follows: the acquired vegetation data is converted from daily scale data to monthly scale data according to the maximum synthesis method, and then the average value of the NDVI value and EVI value within the target grid scale is calculated; the spatial scale conversion of the night light data and elevation data is specifically as follows: the acquired night light data and elevation data are both converted to the target grid scale according to the bilinear interpolation method.

[0018] Furthermore, the independent variable data and the dependent variable data are expressed in the form of equations as follows:

[0019] XCO2=f(t2m,d2m,sp,speed,dir,NDVI,EVI,DN,DEM);

[0020] Among them, XCO2 represents the XCO2 prediction value output by the ERT model; t2m, d2m, sp, speed, dir, NDVI, EVI, DN, and DEM represent 2-meter temperature, 2-meter dew point temperature, surface pressure, wind speed, wind direction, normalized difference vegetation index, enhanced vegetation index, night light, and elevation, respectively.

[0021] Furthermore, the ERT model is composed of four layers of spatiotemporal features; the first layer is the input layer, which is used to represent the independent variables and dependent variables according to a specific equation and serve as the input of the ERT model; the second layer is the feature layer, which is used to randomly select eigenvalues from the dependent variable, and the eigenvalues include factor features and spatiotemporal features; the third layer is the output layer, which is used to output the results of the second layer; and the fourth layer is the result layer, which is used to integrate the output results of the third layer into XCO2 prediction values according to spatiotemporal features.

[0022] Furthermore, the specific process of evaluating the XCO2 predicted value is as follows: evaluating the XCO2 predicted value according to three indicators, wherein the three indicators are the coefficient of determination R 2 , root mean square error RMSE and mean absolute error MAE, specifically expressed as:

[0023]

[0024] Where n represents the sample size in the OCO-2XCO2 data, y i represents the satellite observation value of sample i in the OCO-2XCO2 data, represents the predicted XCO2 value of sample i, represents the average satellite observation of sample i in the OCO-2XCO2 data.

[0025] Furthermore, the specific process of separating the anthropogenic CO2 concentration based on the calculated XCO2 abnormal concentration is: calculating the mean of the XCO2 predicted value output by the trained ERT model, and using the mean as the background atmospheric CO2 concentration value of the study area, subtracting the background atmospheric CO2 concentration value from the XCO2 predicted value to obtain the separated anthropogenic CO2 concentration.

[0026] The beneficial effect of the present invention is that the present invention uses the ERT machine learning algorithm in combination with satellite observation CO2 data and other auxiliary data for assimilation to generate a high-temporal-resolution atmospheric CO2 concentration spatiotemporal dataset for predicting high-resolution atmospheric CO2 concentration. This method improves the defects of using geostatistical interpolation, and adds machine learning models and environmental auxiliary variable data on the basis of interpolation, providing a new technical solution for accurately estimating the atmospheric CO2 concentration in the study area; the present invention uses satellite observation CO2 data for inversion estimation, and uses ground station observation CO2 data for verification and evaluation, thereby improving the accuracy and spatiotemporal resolution of simulated atmospheric CO2 concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The above is an overview of the technical solution of the present invention. To more clearly illustrate the technical solution used in the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0028] Figure 1 This is a schematic diagram of a method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model of the present invention;

[0029] Figure 2 This is a schematic diagram of the principle of preprocessing all data in the present invention;

[0030] Figure 3 This is a result diagram of the atmospheric CO2 concentration prediction value output by the trained ERT model in the present invention based on sample cross-validation and grid cross-validation;

[0031] Figure 4 Schematic diagram of the accuracy of the trained ERT model in the present invention in reconstructing the XCO2 foundation of China's atmosphere on a monthly time scale. DETAILED DESCRIPTION

[0032] The invention will be further described below with reference to the accompanying drawings and in combination with specific implementations, so that those skilled in the art can implement the invention with reference to the description. The protection scope of the invention is not limited to the specific implementations.

[0033] The present invention relates to a method for predicting high-resolution monthly atmospheric CO2 concentration based on an ERT model, the method comprising:

[0034] Obtain CO2 concentration observation data in the study area, such as Figure 1 As shown, the CO2 concentration observation data includes OCO-2XCO2 data and TCCON data, and the OCO-2XCO2 data is used as the dependent variable data input into the ERT model;

[0035] In a specific embodiment, the CO2 concentration observation data comes from the L2 data product of the Orbiting Carbon Observatory-2 (OCO-2) satellite, the first carbon observation satellite successfully launched by the United States. Quality control is performed on the data product, and the data with xco2_quality_flag=0 in the OCO-2XCO2 data is selected as the OCO-2XCO2 data described in the present invention, and the data with an error of less than 0.5ppm in the TCCON data is selected as the TCCON data described in the present invention.

[0036] Obtain several environmental variables in the study area as independent variable data input into the ERT model;

[0037] In a specific embodiment, several environmental variables in the study area are: meteorological data, vegetation data, night light data and elevation data; the meteorological data selects five variables, namely 2-meter temperature, 2-meter dew point temperature, 10-meter U-shaped wind, 10-meter V-shaped wind and surface pressure, and their units are converted; the vegetation data is selected from the normalized vegetation index NDVI and enhanced vegetation index EVI in the MOD13A2 data set; the night light data comes from the NPP / VIIRS data set, and the elevation data comes from the STRM3_V4.1 data set, all of which are used as input data for the ERT model.

[0038] All data are preprocessed. After preprocessing, the independent variable data and the dependent variable data are temporally and spatially matched to the target grid scale to obtain temporally and spatially matched data, and the temporally and spatially matched data are used as the XCO2 training data set;

[0039] In a specific embodiment, the specific process of preprocessing all data is as follows: all data are uniformly resampled to the same spatial resolution of the target grid scale using bilinear interpolation, and the data are uniformly resampled to the same temporal resolution according to the time extraction method; that is, all variables are uniformly resampled to the same spatiotemporal resolution (target grid scale, month). In the OCO-2XCO2 data, data with xco2_quality_flag=0 are regarded as high-quality data, and low-quality data with xco2_quality_flag=1 are filtered out; only data with an error less than 0.5ppm are retained in the XCO2 data of the TCCON data; the 2m temperature and 2m dew point temperature in the meteorological data are converted from Fahrenheit to Celsius; the u component and v component of the 10m wind field are calculated according to the following formulas (1) and (2) to obtain wind speed and wind direction; the surface pressure is converted from Pa to hPa; the normalized vegetation index and enhanced vegetation index in the vegetation data are removed from outliers and the value range is controlled between 0 and 1; and outliers are also removed from the night light data.

[0040] speed=sqrt(u 2 +v 2 ) (1)

[0041] dir=mod(270-arctan2(u,v),360) (2)

[0042] Where speed and dir represent wind speed and wind direction, respectively, and u and v represent the u and v components of the 10 m wind field, respectively.

[0043] In a specific embodiment, the spatiotemporal matching of the independent variable data and the dependent variable data specifically includes: spatiotemporal scale conversion; wherein the spatiotemporal scale conversion includes the spatiotemporal scale conversion of satellite CO2 observation data, the spatiotemporal scale conversion of vegetation data, and the spatial scale conversion of nighttime light and elevation data;

[0044] Among them, the acquired OCO-2XCO2 data is converted from daily scale data to monthly scale data according to the time extraction method, and then the average value of the XCO2 value within the target grid scale is calculated; the scale conversion of the vegetation data is specifically to convert the acquired vegetation data from daily scale data to monthly scale data according to the maximum synthesis method, and then the average value of the NDVI and EVI values within the target grid scale is calculated; the spatial scale conversion of the night light data and elevation data is specifically to convert the acquired night light data and elevation data (500m resolution of NPP / VIIRS and 90m resolution of SRTM3) to the target grid scale according to the bilinear interpolation method.

[0045] Constructing an ERT model, expressing the independent variable data and the dependent variable data in the XCO2 training data set in the form of an equation and inputting them into the ERT model for training to obtain a trained ERT model, and outputting an XCO2 predicted value from the trained ERT model, which is the predicted atmospheric CO2 concentration value;

[0046] In a specific embodiment, the ERT model is composed of four layers of spatiotemporal features. The first layer is the input layer, which is used to represent the independent variables and dependent variables according to a specific equation and serve as the input of the ERT model. The second layer is the feature layer, which is used to randomly select feature values from the dependent variable. The feature values include factor features and spatiotemporal features. The third layer is the output layer, which is used to output the results of the second layer. The fourth layer is the result layer, which is used to integrate the output results of the third layer into XCO2 prediction values according to the spatiotemporal features.

[0047] In a specific embodiment, the independent variable data and the dependent variable data are expressed in the form of equations as follows:

[0048] XCO2=f(t2m,d2m,sp,speed,dir,NDVI,EVI,DN,DEM);

[0049] Where XCO2 represents the atmospheric CO2 concentration predicted by the ERT model, t2m, d2m, sp, speed, dir, NDVI, EVI, DN, and DEM represent 2-meter temperature, 2-meter dew point temperature, surface pressure, wind speed, wind direction, normalized difference vegetation index, enhanced vegetation index, night light, and elevation, respectively; S2: The independent variable data is used as the model characteristic parameter and the R 2, RMSE and MAE values are used as the objectives of model optimization to generate a spatiotemporal dataset of predicted atmospheric CO2 concentration for predicting high-resolution atmospheric CO2 concentration.

[0050] According to the TCCON data, the XCO2 prediction value is evaluated based on three common indicators, namely, the coefficient of determination R 2 , root mean square error RMSE and mean absolute error MAE. If the XCO2 predicted value passes the evaluation, the XCO2 abnormal concentration is calculated according to the XCO2 predicted value, and the artificial CO2 concentration is separated according to the calculated XCO2 abnormal concentration.

[0051] In a specific embodiment, the accuracy of the obtained XCO2 prediction value is evaluated using XCO2 data observed by TCCON ground stations and a ten-fold cross-validation method; the TCCON data includes two stations, Hefei and Xianghe; the ten-fold cross-validation method includes ten-fold sample cross-validation and ten-fold grid cross-validation. The ten-fold sample cross-validation method selects 90% of the samples in the XCO2 prediction value for training, and the remaining 10% for validation, and repeats 10 times. The ten-fold grid cross-validation method also selects 90% of the samples for training, and the remaining 10% for validation, and repeats 10 times.

[0052] Among them, the coefficient of determination R 2 , root mean square error RMSE and mean absolute error MAE are specifically expressed as:

[0053]

[0054] Where n represents the sample size in the OCO-2XCO2 data, y i represents the satellite observation value of sample i in the OCO-2XCO2 data, represents the predicted XCO2 value of sample i, represents the average satellite observation of sample i in the OCO-2XCO2 data.

[0055] The specific process of separating the anthropogenic CO2 concentration based on the calculated XCO2 abnormal concentration is as follows: calculating the mean of the XCO2 predicted values output by the trained ERT model, using the mean as the background atmospheric CO2 concentration value of the study area, and subtracting the background atmospheric CO2 concentration value from the XCO2 predicted value to obtain the separated anthropogenic CO2 concentration.

[0056] like Figure 3 As shown, Figure 3The results of the atmospheric CO2 concentration prediction values output by the trained ERT model in the present invention are based on sample cross-validation and grid cross-validation. The monthly XCO2 values estimated by the method proposed in the present invention are very consistent with the original OCO-2XCO2 results. The results obtained based on sample cross-validation are: R 2 =0.988, RMSE=0.732, MAE=0.501; the results obtained based on grid cross validation are: R 2 =0.947, RMSE=1.523, MAE=1.124. The validation results show that the prediction accuracy of both cross-validation methods is high. The ERT model prediction results are close to the atmospheric CO2 concentration monitored by the OCO-2 satellite, and have strong performance in predicting monthly CO2 concentration. Figure 4 The accuracy of the trained ERT model in reconstructing ground-based XCO2 in China's atmosphere on a monthly time scale is demonstrated. Figure 4 In the data, it can be seen that when the CO2 concentration data at Hefei and Xianghe stations are compared with the ERT model predictions, both have high correlations of 0.951 and 0.865, respectively. The linear regressions are also closely aligned with the 1:1 line, with RMSEs ranging from 1.339 to 1.840, indicating that the model has high accuracy at these stations. In summary, ERT_XCO2 has high consistency with satellite observations and TCCON data, demonstrating the robustness of the proposed method and the accuracy of XCO2 estimation.

Claims

1. A method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model, characterized by: The method comprises: Obtaining CO2 concentration observation data within the study area, wherein the CO2 concentration observation data includes OCO-2XCO2 data and TCCON data, and the OCO-2XCO2 data is used as dependent variable data input into the ERT model; Obtain several environmental variables in the study area as independent variable data input into the ERT model; All data are preprocessed. After preprocessing, the independent variable data and the dependent variable data are temporally and spatially matched to the target grid scale to obtain temporally and spatially matched data, and the temporally and spatially matched data are used as the XCO2 training data set; Constructing an ERT model, expressing the independent variable data and the dependent variable data in the XCO2 training data set in the form of equations and inputting them into the ERT model for training to obtain a trained ERT model, and outputting an XCO2 predicted value from the trained ERT model; According to the TCCON data, the XCO2 prediction value is evaluated based on three common indicators, namely, the coefficient of determination R 2 , root mean square error RMSE and mean absolute error MAE. If the XCO2 predicted value passes the evaluation, the XCO2 abnormal concentration is calculated according to the XCO2 predicted value, and the artificial CO2 concentration is separated according to the calculated XCO2 abnormal concentration.

2. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 1, characterized in that: The several environmental variables in the study area are specifically: meteorological data, vegetation data, night light data and elevation data.

3. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 2, characterized in that: The specific process of preprocessing all the data is: using bilinear interpolation to uniformly resample all the data to the same spatial resolution of the target grid scale, and uniformly resample the data to the same temporal resolution according to the time extraction method.

4. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 3, characterized in that: The spatiotemporal matching of the independent variable data and the dependent variable data specifically includes: spatiotemporal scale conversion; wherein the spatiotemporal scale conversion includes the spatiotemporal scale conversion of OCO-2XCO2 data, the spatiotemporal scale conversion of vegetation data, the spatial scale conversion of night light data, and the spatial scale conversion of elevation data.

5. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 4, characterized in that: The spatiotemporal scale conversion of the OCO-2XCO2 data is specifically as follows: the acquired OCO-2XCO2 data is converted from daily scale data to monthly scale data according to the time extraction method, and then the average value of the XCO2 value within the target grid scale is calculated; the scale conversion of the vegetation data is specifically as follows: the acquired vegetation data is converted from daily scale data to monthly scale data according to the maximum synthesis method, and then the average value of the NDVI value and EVI value within the target grid scale is calculated; the spatial scale conversion of the night light data and elevation data is specifically as follows: the acquired night light data and elevation data are converted to the target grid scale according to the bilinear interpolation method.

6. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 5, characterized in that: The independent variable data and dependent variable data are expressed in the form of equations as follows: XCO2=f(t2m,d2m,sp,speed,dir,NDVI,EVI,DN,DEM); Among them, XCO2 represents the XCO2 prediction value output by the ERT model; t2m, d2m, sp, speed, dir, NDVI, EVI, DN, and DEM represent 2-meter temperature, 2-meter dew point temperature, surface pressure, wind speed, wind direction, normalized difference vegetation index, enhanced vegetation index, night light, and elevation, respectively.

7. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 5, characterized in that: The ERT model consists of four layers of spatiotemporal features. The first layer is the input layer, which is used to represent the independent and dependent variables according to a specific equation and serve as the input of the ERT model. The second layer is the feature layer, which is used to randomly select eigenvalues from the dependent variable. The eigenvalues include factor features and spatiotemporal features. The third layer is the output layer, which is used to output the results of the second layer. The fourth layer is the result layer, which is used to integrate the output results of the third layer into XCO2 prediction values according to spatiotemporal features.

8. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 7, characterized in that: The specific process of evaluating the XCO2 predicted value is as follows: evaluating the XCO2 predicted value according to three indicators, wherein the three indicators are the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE, specifically expressed as: Where n represents the sample size in the OCO-2XCO2 data, y i represents the satellite observation value of sample i in the OCO-2XCO2 data, represents the predicted XCO2 value of sample i, represents the average satellite observation of sample i in the OCO-2XCO2 data.

9. The method for predicting high-resolution monthly atmospheric CO2 concentration based on the ERT model according to claim 8, characterized in that: The specific process of separating the anthropogenic CO2 concentration based on the calculated XCO2 abnormal concentration is as follows: according to the XCO2 predicted value output by the trained ERT model, its mean is calculated, and the mean is used as the background atmospheric CO2 concentration value of the study area, and the background atmospheric CO2 concentration value is subtracted from the XCO2 predicted value to obtain the separated anthropogenic CO2 concentration.

Images