Earth surface CO2 concentration space-time prediction method fused with satellite-earth observation data

Abstract

The invention discloses an earth surface CO2 concentration space-time prediction method fused with satellite-earth observation data, and is suitable for the field of atmosphere and space data monitoring. Preprocessing the acquired XCO2 observation data of the multi-source satellite, the CO2 observation data of the ground station and the auxiliary variable data, and fusing the XCO2 data observed by the multi-source satellite based on an uncertainty weighting method; quantifying the correlation between each auxiliary variable data and the CO2 concentration of the ground station based on a Pearson's correlation coefficient, carrying out significance test, and screening auxiliary variable data with significant correlation; a CNN-GRU-Attention deep fusion model is constructed, spatial-temporal correlation modeling is carried out on the CO2 concentration of a ground station, multi-source satellite observation XCO2 fusion data around the ground station and remarkably related auxiliary variables, and a spatial-temporal characteristic relation is mined, so that spatial-temporal prediction is carried out on the CO2 concentration of the earth surface. The method is strong in generalization ability and high in prediction precision, and can be suitable for space-time prediction of the surface CO2 concentration in different regions.

Description

Technical Field

[0001] The present invention relates to a spatiotemporal prediction method for surface CO2 concentration integrating satellite and ground observation data, which is applicable to the field of atmosphere and space data monitoring. Background Art

[0002] At present, the temporal and spatial characteristics and changing laws of global CO2 concentration are quantitatively studied. Among them, the ground station observation data has the characteristics of high precision and long time series, but its spatial distribution is sparse and the monitoring range is limited, which makes it difficult to achieve large-scale monitoring in spatiotemporal continuity, making it difficult to directly obtain the spatiotemporal continuous high-precision spatial distribution of CO2 on the surface. In contrast, although the accuracy of satellite observations is not as high as that of ground stations, its wide spatial coverage capability provides the possibility to make up for the shortcomings of ground monitoring. The fusion of satellite and ground observation data can effectively compensate for the sparseness of the spatial distribution of ground stations, thereby effectively improving the CO2 monitoring capability and providing a solution for obtaining spatiotemporal continuous surface CO2 data, thereby providing solid data support for the formulation of scientific and reasonable emission reduction strategies.

[0003] Traditional satellite-to-ground observation data fusion methods often focus more on processing the direct nonlinear relationship between satellite observation data and ground station observation data, so as to establish a mapping relationship between satellite-to-ground observation data. This method often focuses on the geographical location of the station to reveal the intrinsic characteristic association between satellite-to-ground observation data. However, this method has certain limitations, ignoring the spatiotemporal influence of the surrounding environment on the CO2 concentration of the ground station, and failing to fully consider the interaction between the ground station and its surrounding environment in the spatiotemporal dimension. In fact, the CO2 concentration of the ground station is not only directly affected by the satellite XCO2 observation data and related auxiliary variables at the location of the station, but also has a certain spatial relationship with the satellite XCO2 observation data and auxiliary variable observation data in the surrounding area. Similarly, the observation data at the previous M moments will also have a certain impact on the CO2 concentration at the current moment. Therefore, in order to overcome the shortcomings of traditional methods, the present invention comprehensively considers the spatiotemporal relationship between satellite-to-ground observation data, and pays attention to the mining of spatiotemporal features to achieve spatiotemporal continuous surface CO2 concentration prediction.

[0004] The prior art with patent number CN202110722157.5 discloses a multi-satellite multi-source XCO2 data fusion method and system. The method is based on the HASM method, and the atmospheric XCO2 spatiotemporal continuous distribution simulation data is fused with the XCO2 observation data obtained by TanSat satellite, GOSAT satellite and OCO-2 satellite observations, so as to obtain a multi-satellite fused XCO2 data set. The above method only uses the XCO2 observation data of TanSat satellite, GOSAT satellite and OCO-2 satellite for fusion, and has not yet considered GOSAT-2 satellite and OCO-3 satellite. The improvement of the spatial coverage of satellite XCO2 data is limited. Secondly, it is only the fusion of satellite XCO2 observation data, and it is impossible to obtain the spatiotemporal continuous surface CO2 concentration.

[0005] The prior art with patent number CN202210765379.X discloses a method for estimating surface NO2 concentration based on satellite remote sensing data. This method matches satellite observation data with ground observation data in time and space, and estimates near-surface NO2 concentration by combining RF and XGBOOST models. However, this method only models the nonlinear relationship between satellite data and ground station observation data, ignoring the temporal and spatial characteristics between observation data. Summary of the invention

[0006] Technical problem: In view of the shortcomings of the existing technology, a method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and ground observation data is provided. The spatial continuity advantage of satellite observation is utilized to compensate for the problem that the sparse spatial distribution of ground stations makes it difficult to obtain continuous surface CO2 concentration data. The XCO2 data of multi-source satellite observations and significantly correlated meteorological auxiliary data are integrated, while taking into account the spatiotemporal correlation characteristics between multi-source observation data. The method can predict the surface CO2 concentration with high accuracy and high spatiotemporal continuity.

[0007] To achieve the above technical objectives, the present invention discloses a spatiotemporal prediction method for surface CO2 concentration by integrating satellite and ground observation data, taking into account the high accuracy of CO2 observation data from ground stations, the low accuracy of satellite observation XCO2 observation but the large coverage, and the strong correlation between the atmospheric XCO2 data observed by satellite and the CO2 concentration data observed by ground stations. Therefore, the multi-source satellite observation XCO2 fusion data, ground station CO2 concentration data and auxiliary variable data are combined, so that the satellite observation data can be used as a supplement to the ground observation, thereby improving the accuracy and reliability of the prediction of surface CO2 concentration.

[0008] The specific steps are as follows:

[0009] S1. Obtain multi-source satellite XCO2 observation data, ground station CO2 observation data, and auxiliary variable data in the study area. The ground station can obtain high-precision and long-time series CO2 concentration data, but the distribution is sparse. The accuracy of multi-source satellite observation XCO2 fusion data is not as good as the ground station CO2 concentration data, but the coverage is wide; multi-source satellite observation data include XCO2 observation data from GOSAT satellite, GOSAT-2 satellite, OCO-2 satellite and OCO-3 satellite. The CO2 concentration data of the ground station comes from the CO2 observation data of the Global Atmospheric Observation Program GAW station. The auxiliary variable data comes from the global climate reanalysis ERA5 data set of the European Center for Medium-Range Weather Forecasts, mainly including 2m temperature, dew point, 10m wind speed, 10m wind direction, surface air pressure and total precipitation;

[0010] S2. Perform quality control, anomaly elimination, and reprojection and resampling on multi-source satellite XCO2 observation data, ground station CO2 observation data, and auxiliary variable data to obtain various observation data at a unified spatiotemporal scale, that is, to unify them to the same coordinate system: WGS84 coordinates, the same spatial resolution: 0.1°, and the same temporal resolution: 1 day;

[0011] S3. Create a 0.1°×0.1° grid cell with the same scope as the study area. Due to the differences in observation accuracy of different satellites, based on the XCO2 observation values ​​of GOSAT satellite, GOSAT-2 satellite, OCO-2 satellite, and OCO-3 satellite in each grid cell and their corresponding observation accuracy, use the uncertainty weighting method to perform weighted fusion on the XCO2 observation values ​​of each satellite in each grid cell. The observation values ​​of different satellites falling in the same grid cell at the same time are fused according to their respective observation accuracy, so that the satellite data with higher observation accuracy is assigned a higher weight during fusion, which is beneficial to improve the overall accuracy of the fused data and increase the observation range of the satellite in the region, so as to obtain the multi-source satellite XCO2 fusion data XCO2_satellites with high coverage of the study area.

[0012] S4. Analyze the correlation between each auxiliary variable and the CO2 concentration of the ground station in the study area, use the Pearson correlation coefficient to quantify the influence of each auxiliary variable data on the CO2 concentration of the ground station, and the correlation between the auxiliary variable data and the CO2 concentration of the ground station, and obtain the auxiliary variables that are significantly correlated with the CO2 concentration of the ground station through significance test screening;

[0013] S5. With the grid corresponding to the CO2 concentration observation value of each ground station as the center, set a spatial window of size N×N, and perform spatial correlation modeling on the CO2 concentration of the grid where the ground station is located, the XCO2 observation value after multi-source satellite fusion of the surrounding N×N grids, and the significantly correlated auxiliary variable observation value; set the time window length to M, and perform temporal correlation modeling on the XCO2 observation value after multi-source satellite fusion at the previous M moments, the significantly correlated auxiliary variable observation value, the ground station CO2 concentration observation value, and the ground station CO2 concentration observation value at the current moment;

[0014] S6. Based on the above-mentioned multi-source satellite XCO2 fusion data XCO2_satellites, CO2 observation data of ground stations, and significantly related auxiliary variables, the three models of convolutional network CNN, gated recurrent unit GRU and attention network Attention are coupled to construct a CNN-GRU-Attention deep fusion model. The CNN-GRU-Attention deep fusion model can fully mine the spatiotemporal characteristics between the CO2 concentration of the ground station and its surrounding multi-source satellite XCO2 fusion data XCO2_satellites and significantly related auxiliary variables; the trained CNN-GRU-Attention deep fusion model is used to perform spatiotemporal prediction of the surface CO2 concentration, so as to obtain spatiotemporally continuous surface CO2 concentration information. The training process and data set of the CNN-GRU-Attention deep fusion model are prior arts in this field.

[0015] Furthermore, in S1, the spatial resolution of the XCO2 observation data of the GOSAT satellite is 10.5km×10.5km, and the revisit period is 3 days; the spatial resolution of the XCO2 observation data of the GOSAT-2 satellite is 9.7km×9.7km, and the revisit period is 3 days; the spatial resolution of the XCO2 observation data of the OCO-2 satellite is 2.25km×1.25km, and the revisit period is 16 days; the spatial resolution of the XCO2 observation data of the OCO-3 satellite is 2.25km×1.29km; the temporal resolution of the GAW ground station observation data is hourly; the spatial resolution of the ERA5 meteorological auxiliary data is 0.25°×0.25°, and the temporal resolution is hourly.

[0016] Furthermore, in S2, the satellite observation data itself is affected by aerosols, orbital spacing, cloud cover and inversion algorithms, so it is necessary to perform quality control on it. According to the instructions in the user manuals of each satellite product, the XCO2_quality_flag observation value is used as the quality judgment standard for satellite XCO2 observation products. The XCO2_quality_flag value of 0 indicates high-quality observation data. Therefore, it is necessary to use this as a basis to perform quality control on the XCO2 observation data of each satellite and preliminarily screen high-quality satellite data. Furthermore, in order to eliminate the outliers of the XCO2 observation data within the spatial range, the 4-times standard deviation method is used to perform spatial filtering on the observation data in the study area, that is, the data with XCO2 observation values ​​within the range of (μ-4σ, μ+4σ) are valid observation values, μ is the mean within the spatial range, and σ is the standard deviation within the spatial range.

[0017] Furthermore, in S3, for the 0.1°×0.1° grid within the study area, each grid cell has XCO2 observations from different satellites and the accuracy information corresponding to the satellite observations. Based on the observation accuracy between different satellites, the uncertainty weighted method is used to perform weighted fusion processing on the XCO2 observations from different satellites in each grid cell to generate an XCO2 fusion data XCO2_satellites with significantly improved spatial coverage compared to single satellite observations. The weighted fusion method based on uncertainty fully considers the observation accuracy differences of each satellite to improve the overall fusion accuracy. The calculation method is as follows:

[0018]

[0019] Among them, XCO'2 represents the fused XCO2 observation value, XCO 2(i) represents the XCO2 observation value of the i-th satellite, w i represents the weight of the i-th satellite observation, ω i It represents the observation accuracy of the i-th satellite, that is, the a posteriori error of the satellite inversion algorithm, which is proportional to the inverse of the weight of the satellite observation. That is, the smaller the a posteriori error, the higher the observation accuracy and the larger the weight. n represents the number of satellites.

[0020] Furthermore, in S4, the correlation between each auxiliary variable data and the CO2 concentration observed at the ground station is quantified by the Pearson correlation coefficient, which is calculated as follows:

[0021]

[0022] Among them, r represents the Pearson correlation coefficient, and its value range is [-1, 1]; X i and Respectively represent the index value and average value of auxiliary variable data X; Y i and Respectively represent the observed value and average value of CO2 concentration at the ground station; r conforms to the t distribution with t-2 degrees of freedom; the significance between the auxiliary variable data factor and the CO2 concentration observed at the ground station is further judged by table lookup test, and the auxiliary variable data that passes the significance test is screened.

[0023] Furthermore, in S5, the CO2 concentration observation value of the ground station not only has a complex mapping relationship with the XCO2_satellites satellite observation data and the significantly related auxiliary variable observation data at its location, but the level of CO2 concentration will also be affected by the XCO2_satellites satellite observation data and the significantly related auxiliary variable data around its location. Therefore, with the location of the ground station as the center point of the grid, an N×N spatial window is set, and a convolutional neural network is used to capture the spatial characteristics of the CO2 concentration observation value of the ground station, the XCO2_satellites satellite observation data, and the significantly related auxiliary variable data within the N×N spatial window; since the CO2 concentration observation value of the current ground station is also affected by the CO2 observation value of the ground station, the XCO2_satellites satellite observation data, and the significantly related auxiliary variable data at the previous M moments, the time window length is set to M, and the gated cyclic unit is used to capture the temporal characteristics of the CO2 concentration observation value of the ground station, the XCO2_satellites satellite observation data, and the significantly related auxiliary variable data;

[0024] The spatiotemporal mapping relationship between satellite and ground observation data is expressed as:

[0025] CO2_Site=f(XCO2_satellites,Auxiliary,Spatial_feature,Temporal_feature)

[0026] Among them, CO2_Site represents the CO2 concentration data observed by the ground site, XCO2_satellites represents the XCO2 observation data after multi-source satellite fusion, Auxiliary represents the significantly correlated auxiliary variable data, Spatial_feature represents the spatial relationship characteristics between the ground site observation data and the XCO2_satellites satellite observation data and the significantly correlated auxiliary variable data, and Temporal_feature represents the temporal relationship characteristics between the ground site observation data and the XCO2_satellites satellite observation data and the significantly correlated auxiliary variable data.

[0027] Furthermore, in S6, the temporal and spatial feature relationships among multi-source satellite XCO2 fusion data, ground observation site CO2 concentration data, and significantly related auxiliary variable data are integrated. Based on convolutional neural networks and gated recurrent units, the attention mechanism is introduced to construct a CNN-GRU-Attention deep fusion model to achieve spatiotemporal prediction of surface CO2 concentration. It mainly includes three parts: data preprocessing, model structure design, and model accuracy verification. The specific steps are as follows:

[0028] a. Data preprocessing: Obtain multi-source satellite XCO2 fusion data XCO2_satellites, ground station CO2 concentration observation data and significantly related auxiliary variable data in the study area, perform spatiotemporal matching of the ground station CO2 concentration data with the multi-source satellite XCO2 fusion data XCO2_satellites and significantly related auxiliary variable data to construct a spatiotemporal unified sample set, divide the data set into the first 70% of samples for training and the last 30% of samples as a test set to verify the model effect;

[0029] b. Network structure design: Construct a CNN-GRU-Attention spatiotemporal fusion model, including a sequentially connected input layer, a convolutional neural network, a gated recurrent unit network, an attention layer, and an output layer. The convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer, and the gated recurrent unit network includes multiple recurrent layers.

[0030] The input layer receives a sample data set consisting of ground station CO2 concentration observation data, multi-source satellite XCO2 fusion data XCO2_satellites, and significantly related auxiliary variable data; the spatial feature information is extracted through the convolutional layer of the convolutional neural network, and then the feature dimension is reduced through the pooling layer to increase the nonlinear expression ability of the model; the output of the convolutional layer is then flattened and input into the gated recurrent unit network, and multiple gated recurrent units are used to extract temporal feature information; the attention layer is then used to dynamically adjust and enhance the ability to extract spatiotemporal information; finally, the extracted spatiotemporal features are mapped to the final prediction results through the fully connected layer, and the final surface CO2 concentration prediction value is generated through the output layer; the loss function of the spatiotemporal fusion model is the mean square error, in order to minimize the difference between the predicted value and the actual observed value, and the Adam algorithm is used as the optimizer of the spatiotemporal fusion model to more efficiently update the network parameters;

[0031] c. Performance evaluation of the CNN-GRU-Attention spatiotemporal fusion model: The root mean square error (RMSE) and mean absolute error (MAE) were used to evaluate the prediction performance of the CNN-GRU-Attention spatiotemporal fusion model.

[0032] The root mean square error represents the deviation between the predicted value and the true value, and the formula is as follows:

[0033]

[0034] Among them, x i Represents the prediction result of the spatiotemporal prediction model on the surface CO2 concentration, y i represents the actual CO2 concentration observation value of the ground station, and n represents the total amount of data;

[0035] The mean absolute error refers to the average error between the predicted value and the true value, reflecting the actual error. The formula is as follows:

[0036]

[0037] Among them, x i Represents the prediction result of the spatiotemporal prediction model on the surface CO2 concentration, y i represents the actual CO2 concentration observation value of the ground station, and n represents the total amount of data.

[0038] A computer device comprises a processor and a memory, wherein the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used to execute a method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data.

[0039] Beneficial Effects

[0040] This method realizes the spatiotemporal prediction of surface CO2 concentration by fusing multi-source satellite and ground observation data, which utilizes XCO2 observation data from multiple satellites: GOSAT, GOSAT-2, OCO-2, and OCO-3. The fusion of multi-source satellite observation data can effectively reduce the impact of missing single satellite observation data and improve the overall spatial coverage of satellite observation data, thereby obtaining more comprehensive and continuous XCO2 observation data. At the same time, the observation accuracy of different satellites varies, so the fusion method based on uncertainty weighting can assign higher weights to satellite data with higher observation accuracy during fusion, which is conducive to improving the overall accuracy of the fused data. Compared with ground station observations, the wide spatial coverage of satellite observations can effectively supplement the missing ground station observation data. Therefore, the idea of ​​fusing satellite and ground observation data is adopted to realize the prediction of surface CO2 concentration. In addition, auxiliary variables such as temperature and dew point will have more or less influence on CO2 concentration. Auxiliary variable data with high spatiotemporal resolution can help more accurately simulate and predict the surface CO2 concentration.

[0041] The constructed CNN-GRU-Attention deep fusion model aims to capture the interaction between CO2 concentration and the surrounding environment to improve the accuracy and reliability of the prediction results. In the model structure design, different convolution kernel sizes are set in the CNN part to capture spatial features and seek the optimal spatial window size. Similarly, the GRU part also sets time windows of different lengths to compare the effects of different window sizes on the accuracy of the prediction results. The deep fusion model uses convolutional neural networks to mine the spatial features of multi-source observation data, and uses gated recurrent units to mine the temporal features of multi-source observation data. The attention mechanism is introduced to enhance the ability to extract spatiotemporal features. The spatiotemporal correlation between ground station observation data, satellite observation data, and meteorological auxiliary variable data is fully and effectively considered to achieve high-precision spatiotemporal prediction of surface CO2 concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the main flow chart of the spatiotemporal prediction method of the surface CO2 concentration by integrating satellite and earth observation data of the present invention.

[0043] Figure 2 It is a structural block diagram of the CNN-GRU-Attention spatiotemporal depth fusion prediction model of the present invention. DETAILED DESCRIPTION

[0044] The method of the present invention is further described below with reference to the accompanying drawings and examples:

[0045] like Figure 1 As shown, the present invention discloses a method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and ground observation data, the steps of which are: firstly, obtaining multi-source satellite XCO2 observation data (GOSAT, GOSAT-2, OCO-2, OCO-3), ground station CO2 data, and auxiliary variable data; then, fusing the multi-source satellite XCO2 observation data based on an uncertainty weighted method to improve the spatial coverage of satellite observation data; then, using the Pearson correlation coefficient to analyze the correlation between the auxiliary variable data and the CO2 concentration data of the ground station, and obtaining significantly correlated auxiliary variables; finally, based on a convolutional neural network, a gated recurrent unit, and an attention mechanism, a deep fusion model is constructed to mine the spatiotemporal correlation features between satellite and ground observation data, so as to realize spatiotemporal prediction of surface CO2 concentration.

[0046] The specific steps are:

[0047] S1. Obtain multi-source satellite XCO2 observation data, ground station CO2 observation data, and auxiliary variable data in the study area. The multi-source satellite observation data include XCO2 observation data from GOSAT satellite, GOSAT-2 satellite, OCO-2 satellite, and OCO-3 satellite. The CO2 concentration data of the ground station comes from the CO2 observation data of the Global Atmospheric Observation Program GAW station. The auxiliary variable data comes from the global climate reanalysis ERA5 data set of the European Center for Medium-Range Weather Forecasts, mainly including 2m temperature, dew point, 10m wind speed, 10m wind direction, surface air pressure, and total precipitation.

[0048] S2. Perform quality control, anomaly removal, and reprojection and resampling on multi-source satellite XCO2 observation data, ground station CO2 observation data, and auxiliary variable data to obtain various observation data at a unified spatiotemporal scale (i.e., the same coordinate system: WGS84 coordinates, the same spatial resolution: 0.1°, and the same temporal resolution: 1 day);

[0049] S3. Create a 0.1°×0.1° grid cell with the same scope as the study area. Based on the XCO2 observation values ​​of GOSAT satellite, GOSAT-2 satellite, OCO-2 satellite, and OCO-3 satellite in each grid cell and their corresponding observation accuracy, use the uncertainty weighting method to perform weighted fusion on the XCO2 observation values ​​of each satellite in each grid cell to obtain XCO2 fusion data XCO2_satellites with higher spatial coverage than single satellite observations in the study area.

[0050] S4. Analyze the correlation between each auxiliary variable and the CO2 concentration of the ground station in the study area, use the Pearson correlation coefficient to quantify the influence of each auxiliary variable on the CO2 concentration of the ground station, and screen out the auxiliary variables that are significantly correlated with the CO2 concentration of the ground station;

[0051] S5. With the grid corresponding to the CO2 concentration observation value of each ground station as the center, set a spatial window of size N×N, and perform spatial correlation modeling on the CO2 concentration of the grid where the ground station is located, the XCO2 observation value after multi-source satellite fusion of the surrounding N×N grids, and the significantly correlated auxiliary variable observation value; set the time window length to M, and perform temporal correlation modeling on the XCO2 observation value after multi-source satellite fusion at the previous M moments, the significantly correlated auxiliary variable observation value, the ground station CO2 concentration observation value, and the ground station CO2 concentration observation value at the current moment;

[0052] S6. Based on the above-mentioned multi-source satellite XCO2 fusion data XCO2_satellites, CO2 observation data of ground stations, and significantly related auxiliary variables, a CNN-GRU-Attention (convolutional neural network-gated recurrent unit-attention mechanism) deep fusion model is constructed and trained to mine the spatiotemporal characteristics between the CO2 concentration of ground stations and their surrounding XCO2_satellites and significantly related auxiliary variables; the trained CNN-GRU-Attention deep fusion model is used to perform spatiotemporal prediction of surface CO2 concentration.

[0053] like Figure 2 As shown in the figure, the spatiotemporal feature relationship between multi-source satellite XCO2 fusion data, ground observation site CO2 concentration data, and significantly related auxiliary variable data is integrated. Based on convolutional neural network and gated recurrent unit, the attention mechanism is introduced to construct a CNN-GRU-Attention deep fusion model to achieve spatiotemporal prediction of surface CO2 concentration. It mainly includes three parts: data preprocessing, model structure design, and model accuracy verification. The specific description is as follows:

[0054] 1. Data preprocessing

[0055] The multi-source satellite XCO2 fusion data XCO2_satellites, ground station CO2 concentration observation data and significantly correlated auxiliary variable data in the study area were obtained, and the CO2 concentration data of the ground stations were matched with the multi-source satellite XCO2 fusion data XCO2_satellites and significantly correlated auxiliary variable data in time and space to construct a unified sample set in time and space. The data set was divided into the first 70% of the samples for training and the last 30% of the samples as a test set to verify the model effect.

[0056] 2. Network structure design: The main body is the CNN-GRU-Attention spatiotemporal fusion model, which includes a sequentially connected input layer, a convolutional neural network, a gated recurrent unit network, an attention layer, and an output layer. The convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer, and the gated recurrent unit network includes multiple recurrent layers.

[0057] The input layer receives a sample data set consisting of ground station CO2 concentration observation data, multi-source satellite XCO2 fusion data XCO2_satellites, and significantly related auxiliary variable data; the convolutional layer of the convolutional neural network is used to extract spatial feature information, and then the pooling layer is used to reduce the feature dimension to increase the nonlinear expression ability of the model; the output of the convolutional layer is then flattened and input into the gated recurrent unit network, and multiple gated recurrent units are used to extract temporal feature information; the attention layer is then used to dynamically adjust and enhance the ability to extract spatiotemporal information; finally, the fully connected layer maps the extracted spatiotemporal features to the final prediction result, and the output layer generates the final surface CO2 concentration prediction value. The mean squared error (MSE) is used as the loss function of the model to minimize the difference between the predicted value and the actual observed value. The model optimizer uses the Adam (Adaptive Moment Estimation) algorithm to more efficiently update the network parameters.

[0058] 3. Performance evaluation of CNN-GRU-Attention spatiotemporal fusion model

[0059] The prediction performance of the CNN-GRU-Attention spatiotemporal fusion model was evaluated using the root mean square error (RMSE) and mean absolute error (MAE).

[0060] The root mean square error represents the deviation between the predicted value and the true value, and is calculated as follows:

[0061]

[0062] Among them, x i Represents the prediction result of the spatiotemporal prediction model on the surface CO2 concentration, y i represents the actual CO2 concentration observation value of the ground station, and n represents the total amount of data;

[0063] The mean absolute error refers to the average error between the predicted value and the true value, which reflects the actual error. The calculation formula is as follows:

[0064]

[0065] Among them, x i Represents the prediction result of the spatiotemporal prediction model on the surface CO2 concentration, y i represents the actual CO2 concentration observation value of the ground station, and n represents the total amount of data.

Claims

1. A method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data, characterized by: In order to achieve the above technical purpose, the present invention discloses a spatiotemporal prediction method for surface CO2 concentration by integrating satellite and ground observation data, taking into account the high accuracy of CO2 observation data from ground stations, the low accuracy of satellite observation XCO2 observation but the large coverage, and the strong correlation between the atmospheric XCO2 data observed by satellite and the CO2 concentration data observed by ground stations, thus combining the multi-source satellite observation XCO2 fusion data, the ground station CO2 concentration data and the auxiliary variable data, so that the satellite observation data can be used as a supplement to the ground observation, thereby improving the accuracy and reliability of the prediction of surface CO2 concentration; The specific steps are as follows: S1. Obtain multi-source satellite XCO2 observation data, ground station CO2 observation data, and auxiliary variable data in the study area. The ground station can obtain high-precision and long-time series CO2 concentration data, but the distribution is sparse. The accuracy of multi-source satellite observation XCO2 fusion data is not as good as the ground station CO2 concentration data, but the coverage is wide; multi-source satellite observation data include XCO2 observation data from GOSAT satellite, GOSAT-2 satellite, OCO-2 satellite and OCO-3 satellite. The CO2 concentration data of the ground station comes from the CO2 observation data of the Global Atmospheric Observation Program GAW station. The auxiliary variable data comes from the global climate reanalysis ERA5 data set of the European Center for Medium-Range Weather Forecasts, mainly including 2m temperature, dew point, 10m wind speed, 10m wind direction, surface air pressure and total precipitation; S2. Perform quality control, anomaly elimination, and reprojection and resampling on multi-source satellite XCO2 observation data, ground station CO2 observation data, and auxiliary variable data to obtain various observation data at a unified spatiotemporal scale, that is, to unify them to the same coordinate system: WGS84 coordinates, the same spatial resolution: 0.1°, and the same temporal resolution: 1 day; S3. Create a 0.1°×0.1° grid cell with the same scope as the study area. Due to the differences in observation accuracy of different satellites, based on the XCO2 observation values ​​of GOSAT satellite, GOSAT-2 satellite, OCO-2 satellite, and OCO-3 satellite in each grid cell and their corresponding observation accuracy, use the uncertainty weighting method to perform weighted fusion on the XCO2 observation values ​​of each satellite in each grid cell. The observation values ​​of different satellites falling in the same grid cell at the same time are fused according to their respective observation accuracy, so that the satellite data with higher observation accuracy is assigned a higher weight during fusion, which is beneficial to improve the overall accuracy of the fused data and increase the observation range of the satellite in the region, so as to obtain the multi-source satellite XCO2 fusion data XCO2_satellites with high coverage of the study area. S4. Analyze the correlation between each auxiliary variable and the CO2 concentration of the ground station in the study area, use the Pearson correlation coefficient to quantify the influence of each auxiliary variable data on the CO2 concentration of the ground station, and the correlation between the auxiliary variable data and the CO2 concentration of the ground station, and obtain the auxiliary variables that are significantly correlated with the CO2 concentration of the ground station through significance test screening; S5. With the grid corresponding to the CO2 concentration observation value of each ground station as the center, set a spatial window of size N×N, and perform spatial correlation modeling on the CO2 concentration of the grid where the ground station is located, the XCO2 observation value after multi-source satellite fusion of the surrounding N×N grids, and the significantly correlated auxiliary variable observation value; set the time window length to M, and perform temporal correlation modeling on the XCO2 observation value after multi-source satellite fusion at the previous M moments, the significantly correlated auxiliary variable observation value, the ground station CO2 concentration observation value, and the ground station CO2 concentration observation value at the current moment; S6. Based on the above multi-source satellite XCO2 fusion data XCO2_satellites, CO2 observation data of ground stations, and significantly related auxiliary variables, a CNN-GRU-Attention deep fusion model is constructed. The CNN-GRU-Attention deep fusion model can fully mine the spatiotemporal characteristics between the CO2 concentration of ground stations and their surrounding multi-source satellite XCO2 fusion data XCO2_satellites and significantly related auxiliary variables; the trained CNN-GRU-Attention deep fusion model is used to perform spatiotemporal prediction of the surface CO2 concentration, thereby obtaining spatiotemporally continuous surface CO2 concentration information.

2. The method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data according to claim 1 is characterized in that: In S1, the spatial resolution of XCO2 observation data from GOSAT satellite is 10.5km×10.5km, and the revisit period is 3 days; the spatial resolution of XCO2 observation data from GOSAT-2 satellite is 9.7km×9.7km, and the revisit period is 3 days; the spatial resolution of XCO2 observation data from OCO-2 satellite is 2.25km×1.25km, and the revisit period is 16 days; the spatial resolution of XCO2 observation data from OCO-3 satellite is 2.25km×1.29km; The temporal resolution of GAW ground station observation data is hourly; the spatial resolution of ERA5 meteorological auxiliary data is 0.25°×0.25°, and the temporal resolution is hourly.

3. The method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data according to claim 1, characterized in that: In S2, the satellite observation data itself is affected by aerosols, orbital spacing, cloud cover and inversion algorithms, so it is necessary to perform quality control on it. According to the instructions in the user manuals of each satellite product, the XCO2_quality_flag observation value is used as the quality judgment standard for satellite XCO2 observation products. The XCO2_quality_flag value of 0 indicates high-quality observation data. Therefore, it is necessary to use this as a basis to perform quality control on the XCO2 observation data of each satellite and preliminarily screen high-quality satellite data. Furthermore, in order to eliminate the outliers of the XCO2 observation data within the spatial range, the 4-times standard deviation method is used to perform spatial filtering on the observation data in the study area, that is, the data with XCO2 observation values ​​within the range of (μ-4σ, μ+4σ) are valid observation values, μ is the mean within the spatial range, and σ is the standard deviation within the spatial range.

4. The method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data according to claim 1, characterized in that: In S3, for the 0.1°×0.1° grid within the study area, each grid cell has XCO2 observations from different satellites and the accuracy information corresponding to the satellite observations. Based on the observation accuracy between different satellites, the uncertainty weighted method is used to perform weighted fusion processing on the XCO2 observations from different satellites in each grid cell to generate an XCO2 fusion data XCO2_satellites with significantly improved spatial coverage compared to single satellite observations. The weighted fusion method based on uncertainty fully considers the differences in observation accuracy of each satellite to improve the overall fusion accuracy. The calculation method is as follows: Among them, XCO'2 represents the fused XCO2 observation value, XCO 2(i) represents the XCO2 observation value of the i-th satellite, w i represents the weight of the i-th satellite observation, ω i It represents the observation accuracy of the i-th satellite, that is, the a posteriori error of the satellite inversion algorithm, which is proportional to the inverse of the weight of the satellite observation. That is, the smaller the a posteriori error, the higher the observation accuracy and the larger the weight. n represents the number of satellites.

5. The method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data according to claim 1 is characterized in that In S4, the correlation between each auxiliary variable data and the CO2 concentration observed at the ground station is quantified by the Pearson correlation coefficient, which is calculated as follows: Among them, r represents the Pearson correlation coefficient, and its value range is [-1, 1]; X i and Respectively represent the index value and average value of auxiliary variable data X; Y i and Respectively represent the observed value and average value of CO2 concentration at the ground station; r conforms to the t distribution with t-2 degrees of freedom; the significance between the auxiliary variable data factor and the CO2 concentration observed at the ground station is further judged by table lookup test, and the auxiliary variable data that passes the significance test is screened.

6. The method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data according to claim 1, characterized in that: In S5, the CO2 concentration observation value of the ground station not only has a complex mapping relationship with the XCO2_satellites satellite observation data and the significantly related auxiliary variable observation data at its location, but the level of CO2 concentration will also be affected by the XCO2_satellites satellite observation data and the significantly related auxiliary variable data around its location. Therefore, the location of the ground station is taken as the center point of the grid, an N×N spatial window is set, and a convolutional neural network is used to capture the spatial characteristics of the CO2 concentration observation value of the ground station, the XCO2_satellites satellite observation data, and the significantly related auxiliary variable data within the N×N spatial window; Since the CO2 concentration observation value of the current ground station is also affected by the CO2 observation value of the ground station, the XCO2_satellites satellite observation data, and the significantly related auxiliary variable data at the previous M moments, the time window length is set to M, and the gated cyclic unit is used to capture the temporal characteristics of the CO2 concentration observation value of the ground station, the XCO2_satellites satellite observation data, and the significantly related auxiliary variable data; The spatiotemporal mapping relationship between satellite and ground observation data is expressed as: CO2_Site=f(XCO2_satellites,Auxiliary,Spatial_feature,Temporal_feature) Among them, CO2_Site represents the CO2 concentration data observed by ground stations, XCO2_sotellites represents the XCO2 observation data after multi-source satellite fusion, Auxiliary represents significantly correlated auxiliary variable data, and Spatial_feature represents the ground station Temporal_feature represents the spatial relationship characteristics between point observation data and XCO2_satellites satellite observation data and significantly related auxiliary variable data. Temporal_feature represents the temporal relationship characteristics between ground station observation data and XCO2_satellites satellite observation data and significantly related auxiliary variable data.

7. The method for spatiotemporal prediction of surface CO2 concentration by integrating satellite and earth observation data according to claim 1, characterized in that: In S6, the temporal and spatial feature relationships among multi-source satellite XCO2 fusion data, ground observation site CO2 concentration data, and significantly related auxiliary variable data are integrated. Based on convolutional neural networks and gated recurrent units, the attention mechanism is introduced to construct a CNN-GRU-Attention deep fusion model to achieve spatiotemporal prediction of surface CO2 concentration. It mainly includes three parts: data preprocessing, model structure design, and model accuracy verification. The specific steps are as follows: a. Data preprocessing: Obtain multi-source satellite XCO2 fusion data XCO2_satellites, ground station CO2 concentration observation data and significantly related auxiliary variable data in the study area, perform spatiotemporal matching of the ground station CO2 concentration data with the multi-source satellite XCO2 fusion data XCO2_satellites and significantly related auxiliary variable data to construct a spatiotemporal unified sample set, divide the data set into the first 70% of samples for training and the last 30% of samples as a test set to verify the model effect; b. Network structure design: Construct a CNN-GRU-Attention spatiotemporal fusion model, including a sequentially connected input layer, a convolutional neural network, a gated recurrent unit network, an attention layer, and an output layer. The convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer, and the gated recurrent unit network includes multiple recurrent layers. The input layer receives a sample data set consisting of ground station CO2 concentration observation data, multi-source satellite XCO2 fusion data XCO2_satellites, and significantly related auxiliary variable data; the spatial feature information is extracted through the convolutional layer of the convolutional neural network, and then the feature dimension is reduced through the pooling layer to increase the nonlinear expression ability of the model; the output of the convolutional layer is then flattened and input into the gated recurrent unit network, and multiple gated recurrent units are used to extract temporal feature information; the attention layer is then used to dynamically adjust and enhance the ability to extract spatiotemporal information; finally, the extracted spatiotemporal features are mapped to the final prediction results through the fully connected layer, and the final surface CO2 concentration prediction value is generated through the output layer; the loss function of the spatiotemporal fusion model is the mean square error, in order to minimize the difference between the predicted value and the actual observed value, and the Adam algorithm is used as the optimizer of the spatiotemporal fusion model to more efficiently update the network parameters; c. Performance evaluation of the CNN-GRU-Attention spatiotemporal fusion model: The root mean square error (RMSE) and mean absolute error (MAE) were used to evaluate the prediction performance of the CNN-GRU-Attention spatiotemporal fusion model. The root mean square error represents the deviation between the predicted value and the true value, and the formula is as follows: Among them, x i Represents the prediction result of the spatiotemporal prediction model on the surface CO2 concentration, y i represents the actual CO2 concentration observation value of the ground station, and n represents the total amount of data; The mean absolute error refers to the average error between the predicted value and the true value, reflecting the actual error. The formula is as follows: Among them, x i Represents the prediction result of the spatiotemporal prediction model on the surface CO2 concentration, y i represents the actual CO2 concentration observation value of the ground station, and n represents the total amount of data.

8. A computer device, characterized in that: It includes a processor and a memory, the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used to execute the spatiotemporal prediction method of surface CO2 concentration by integrating satellite and earth observation data as described in any one of claims 1-7.

Images