Carbon dioxide column concentration multi-source spatio-temporal data fusion method

Abstract

The invention relates to the technical field of atmospheric environment measurement and control, and particularly discloses a carbon dioxide column concentration multi-source spatio-temporal data fusion method, which comprises the following steps: by combining a hybrid model of HGT and Transformer, reanalyzing meteorological variables such as modeled XCO2 and wind speed and direction of ERA-5 in a product by utilizing XCO2 and CAMS-IO inverted by an orbital carbon observation satellite OCO-2 satellite, enhancing a CAMS-EGG4 data set, and obtaining a CAMS-EGG4 data set; according to the method, multi-source input features are fused in spatio-temporal joint modeling to realize high precision and strong generalization ability, a heterogeneous spatio-temporal diagram is constructed by combining a spatial proximity relationship with time sequence nodes, so that an internal spatio-temporal dependency structure in atmospheric CO2 observation is explicitly expressed, and the spatial coverage breadth advantage can be maintained while the spatial coverage breadth advantage is maintained. And the observation precision comparable with that of foundation observation is realized.

Description

Technical Field

[0001] This application relates to the field of atmospheric environment monitoring and control technology, and specifically discloses a method for fusing multi-source spatiotemporal data of carbon dioxide column concentration. Background Technology

[0002] Accurate monitoring, prediction, and assessment of the average mole fraction of carbon dioxide (XCO2) in dry air columns at global and regional scales are crucial for achieving carbon neutrality. Among various XCO2 monitoring methods, satellite observation and ground-based measurements are currently the two most widely used and reliable approaches. However, satellite observation is limited by its relatively low accuracy, while ground-based measurements are constrained by their limited spatial coverage.

[0003] While satellite datasets typically provide high-quality XCO2 estimates, their accuracy is often affected by atmospheric interference, and significant spatiotemporal gaps arise due to long revisit periods. Furthermore, XCO2 retrieval results from greenhouse gas (GHG) satellite missions often lack long-term continuity. In contrast, model-based data (such as CAMS) utilize data assimilation methods to combine multiple satellite and other observations with atmospheric chemistry models and emission priors, thereby obtaining physically consistent, best-in-class estimates of historical atmospheric states. This type of data offers advantages in terms of high temporal resolution and long-term continuity, but its consistency is inferior to that of satellite observations.

[0004] The development of big data technology has led to the increasing popularity of data-driven methods in addressing atmospheric issues. Traditional methods often rely on interpolation techniques to fuse satellite and model data to generate improved XCO2 products. However, traditional interpolation typically assumes spatial continuity and smoothness, thus failing to capture abrupt changes caused by atmospheric transport. In recent years, deep learning methods have been used to fuse satellite and model data to generate XCO2 datasets. By comprehensively modeling the spatiotemporal variations of XCO2 using convolutional neural networks (CNN), long short-term memory networks (LSTM), and spatial attention mechanisms, a data-driven deep learning method based on multi-source satellite observations has been established, integrating variables from land surface, anthropogenic, and meteorological sources.

[0005] Most current deep learning models are based on spatiotemporal sequence neural network architectures, which typically converge slowly, thus limiting their application to relatively small regions. Furthermore, whether it's interpolation or deep learning methods, their outputs are mostly difficult to form long-term continuous sequences; the temporal resolution is usually only once every 16 days or once per day, severely restricting the construction of long-term XCO2 datasets. Convolutional neural networks (CNNs) are suitable for regular grids but struggle to handle irregular spatial distributions and complex adjacency structures in real-world geographic spaces. Models based on recurrent neural networks (RNNs), such as gated recurrent units (GRUs) and long short-term memory networks (LSTMs), can capture temporal dependencies but often fail to effectively incorporate spatial relationships, resulting in insufficient characterization of spatial autocorrelation.

[0006] This invention provides a method for fusing multi-source spatiotemporal data of carbon dioxide column concentration to solve the above-mentioned problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method for fusing multi-source spatiotemporal data of carbon dioxide column concentration in spatiotemporal joint modeling, which can integrate multi-source input features, so as to achieve observation accuracy comparable to ground-based observation while maintaining the advantage of broad spatial coverage.

[0008] To achieve the above objectives, the basic solution of this invention provides a method for fusing multi-source spatiotemporal data of carbon dioxide column concentration, comprising the following steps: Step S1, Constructing a dynamic heterogeneous graph: Based on the existing dataset, dynamically construct a dynamic heterogeneous graph for multi-source XCO2 spatiotemporal modeling according to time steps; Step S2, Multi-source feature standardization: z-score standardization is performed on the original numerical features from the same source, and independent feature extraction and alignment strategies are adopted for different node types; Step S3: Construct a hybrid model combining HGT and Transformer: including a spatial modeling module and a temporal modeling module; In the spatial modeling module, the input features are used to model spatial autocorrelation through an adjacency matrix constructed based on geographic coordinates and wind field information, and specific types of node embeddings are generated. In the time modeling module, Transformer encoding is used to enhance the sensitivity to time order, and time-dependent features are extracted through structures such as multi-head self-attention, feedforward networks, and residual connections. Spatial and temporal features are spliced ​​together to form fused features, so as to obtain the predicted CO2 concentration using linear layer mapping; Step S4: Use a hybrid model to calculate and output the predicted XCO2 concentration.

[0009] Furthermore, in step S1, an independent graph snapshot is constructed at each time step, and the snapshots are arranged sequentially to form a time-series dynamic heterogeneous graph. The graph contains data nodes from different sources and at different resolutions, and multiple types of edges are automatically constructed based on spatial adjacency and cross-source semantic relationships.

[0010] Furthermore, in step S1, the edge type is defined based on the physical or statistical relationships between nodes, defining various meta-relationships: For grid points, construct the edges connecting the grid points based on the four-adjacency rule; For scattered points, a K-dimensional tree is used to find the neighbors of the observation points within a given radius r, and a local spatial association structure is constructed. Based on nearest neighbor or radius search, satellite observation nodes are connected to grid nodes at the same time to achieve cross-resolution alignment. Forward and backward time edges are constructed between adjacent time snapshots to achieve information continuity in the time series.

[0011] Furthermore, in step S2, a multi-source feature dimension alignment module using linear projection is used to project features of different node types onto a unified dimension.

[0012] Furthermore, in the spatial modeling module, a message passing mechanism is used to update the node representation, and an independent transformation matrix is ​​learned for each node type and edge type. Different attention weights are assigned to different types of neighbors to generate more discriminative node embedding representations. In the time modeling module, the Transformer takes the observations at each time step as input sequence elements and uses a multi-head self-attention mechanism to dynamically capture the correlation between different time points.

[0013] Furthermore, in step S3, the feature dimension output by the Transformer encoder is consistent with that of the spatial module.

[0014] Furthermore, in step S3, the fused features are input into a two-layer feedforward neural network for further fusion and regression.

[0015] Furthermore, in step S4, the calculation steps of the hybrid model are as follows: Step S41: Input the unified node features and their heterogeneous relationships into the HGT module and generate node-level spatial features; Step S42: Aggregate spatial features to each time step according to the participation relationship of time nodes; Step S43: Input the time series feature sequence formed by each time step into the time Transformer encoder to extract the time dependency structure across days or hours and obtain the time series representation; Step S44: Using a lightweight regression prediction head, the Transformer output vector is mapped to the final XCO2 concentration prediction value.

[0016] Furthermore, in step S41, the expression for generating node-level spatial features is as follows: ; in, This represents the node embedding representation obtained after processing by the HGT module in the t-th graph snapshot. This represents the initial feature set of all nodes input into the HGT at time slice t. This represents the set of all edge relationships in a dynamic heterogeneous graph during that time slice. In step S42, the expression for aggregating spatial features to each time step is as follows: ; Where t is the time node. This is the spatial representation of the aggregated time nodes, and serves as the input to the temporal Transformer. Aggregate(·) is a subset of all spatial nodes pointing to time node t. Aggregate(·) is an aggregation operation used to combine all spatial nodes belonging to the same time. In step S43, the timing representation expression is as follows: ; Where T is the length of the time-series feature sequence. is a temporal feature sequence, and μ is the output vector obtained after Transformer encoding; In step S43, the expression for the final predicted XCO2 concentration is as follows: ; Where y represents the XCO2 concentration value predicted by the model.

[0017] The principle and effect of this solution are as follows: 1. This invention combines a hybrid model of HGT and Transformer, utilizing XCO2 retrieved from the OCO-2 orbital carbon observation satellite, modeled XCO2 from the CAMS-IO reanalysis product, and meteorological variables such as wind speed and direction from ERA-5 to enhance the CAMS-EGG4 dataset. This allows for the fusion of multi-source input features in spatiotemporal joint modeling to achieve high accuracy and strong generalization ability. By combining spatial proximity relationships with time series nodes, a heterogeneous spatiotemporal map is constructed, which explicitly expresses the inherent spatiotemporal dependency structure in atmospheric CO2 observation. This achieves observation accuracy comparable to ground-based observations while maintaining the advantage of broad spatial coverage.

[0018] 2. The HGT module effectively models the spatial autocorrelation between observation stations, enabling the network to represent complex geographic and atmospheric connectivity relationships that are difficult to capture by traditional convolutional or statistical models. By utilizing the graph structure, it maintains the spatial continuity of CO2 changes and enhances regional consistency under different environments.

[0019] 3. The Transformer module excels in handling temporal dependencies. Through its self-attention mechanism, it can adaptively learn long-range temporal interactions and seasonal variations, demonstrating superior performance in representing interannual and sub-seasonal variability compared to recurrence-based models (such as LSTM). This ensures a more stable and interpretable understanding of the temporal evolution of XCO2; the integration of spatial and temporal advantages significantly improves the accuracy and robustness of XCO2 simulations.

[0020] 4. As can be seen, the hybrid model of this invention can effectively improve the fusion capability of multi-source spatiotemporal data on atmospheric XCO2 concentration, demonstrating a strong ability to capture the spatial autocorrelation and temporal dependence of multi-source datasets. This method provides a new approach for generating high-resolution CO2 concentration fields, which can support climate monitoring, carbon budget assessment, and policy decisions for achieving carbon neutrality goals. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a multi-source spatiotemporal data fusion method for carbon dioxide column concentration proposed in an embodiment of this application is shown; Figure 2 This paper presents an example diagram illustrating the construction of a dynamic heterogeneous graph in a multi-source spatiotemporal data fusion method for carbon dioxide column concentration proposed in an embodiment of this application. Figure 3 This paper illustrates an example diagram of the projection of the multi-source feature dimension alignment module in a multi-source spatiotemporal data fusion method for carbon dioxide column concentration proposed in an embodiment of this application. Figure 4 This paper illustrates the structure of a hybrid model in a multi-source spatiotemporal data fusion method for carbon dioxide column concentration proposed in an embodiment of this application. Figure 5 An example image of global OCO-2 satellite XCO2 observations in the dataset of this application embodiment is shown; Figure 6 This paper shows an example of the spatial distribution of XCO2 in the global CAMS-EGG4 reanalysis dataset of an embodiment of this application; Figure 7 This paper illustrates an example of the spatial distribution of the global CAMS-I0 reanalysis XOO2 dataset in an embodiment of this application. Figure 8 An example diagram of the TCCON Hefei station in the dataset of this application embodiment is shown; Figure 9 An example diagram of TCCON Xianghe Station in the dataset of this application embodiment is shown; Figure 10 This application illustrates an example of a global 10-meter height wind field based on ERA-5 in the dataset of an embodiment. Figure 11 A comparison chart of XCO2 values ​​between CAMS and TCCON sites in an embodiment of this application is shown; Figure 12 This paper presents a comparison chart of the predicted values ​​and the XCO2 values ​​at the TCCON site in an embodiment of this application. Figure 13 The diagram shows a comparison between the predicted XCO2 value and the original CAMS-XCO2 data for 2017 in an embodiment of this application, where (a) is the predicted XCO2 value on January 1, 2017, (b) is the original CAMS data on January 1, 2017, (c) is the predicted XCO2 value on April 1, 2017, (d) is the original CAMS data on April 1, 2017, (e) is the predicted XCO2 value on July 1, 2017, (f) is the original CAMS data on July 1, 2017, (g) is the predicted XCO2 value on October 1, 2017, and (h) is the original CAMS data on October 1, 2017. Figure 14 The diagram shows a comparison between the predicted XCO2 value and the original CAMS-XCO2 data for 2020 in the implementation of this application, where (a) is the predicted XCO2 value on January 1, 2020, (b) is the original CAMS data on January 1, 2020, (c) is the predicted XCO2 value on April 1, 2020, (d) is the original CAMS data on April 1, 2020, (e) is the predicted XCO2 value on July 1, 2020, (f) is the original CAMS data on July 1, 2020, (g) is the predicted XCO2 value on October 1, 2020, and (h) is the original CAMS data on October 1, 2020. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0024] A method for fusing multi-source spatiotemporal data of carbon dioxide column concentration, implementing, for example... Figure 1 As shown, it includes the following steps: Step S1: Construct a dynamic heterogeneous graph; To address the spatial heterogeneity of multi-source XCO2 data, this embodiment employs a time-step dynamic heterogeneous graph modeling method based on existing datasets. During the modeling process, an independent graph snapshot is constructed at each time step, and these snapshots are sequentially arranged to form a dynamic heterogeneous time-series graph. The graph contains data nodes from different sources and at different resolutions, and automatically constructs multiple types of edges based on spatial adjacency and cross-source semantic relationships, thereby achieving a unified representation of multi-source data.

[0025] like Figure 2 As shown, based on the physical meaning of different data sources, there are multiple node types. Satellite observation nodes represent single-point spectral inversion XCO2 observations, including features such as latitude and longitude, timestamp, XCO2 value, and quality identifier; reanalysis grid nodes represent CO2 or meteorological parameters of grid points, such as temperature, wind speed, and humidity; time nodes are used to aggregate nodes within the same time snapshot, serving as the connection point for the time Transformer input.

[0026] The edge type is constructed based on the physical or statistical relationships between nodes, defining various meta-relationships. For grid points, edges connecting grids are constructed based on the four-adjacency rule. For scattered points, K-dimensional trees (KDTree) are used to find the neighbors of observation points within a given radius r, constructing a local spatial association structure. Finally, based on nearest neighbor (KNN) or radius search, satellite observation nodes are connected to grid nodes at the same time to achieve cross-resolution alignment. Forward and backward time edges are constructed between adjacent time snapshots to achieve information continuity in the time series.

[0027] This graphing approach can achieve natural connections between cross-source data while ensuring the expressive power of spatial structure, enabling subsequent heterogeneous graph neural networks to simultaneously capture spatial local relationships and cross-source global correlations.

[0028] Step S2, multi-source feature standardization; The data processed in this embodiment comes from multiple heterogeneous sources, and these sources differ in sampling location, physical meaning, resolution, time frequency, and data dimensions. To enable these data to be processed by the same model structure, this embodiment employs the following standardized data preprocessing method to ensure consistency and fusion of multi-source data within the graph structure, as detailed below: Step S21 involves performing z-score standardization on the original numerical features from the same source, including mean centering and scale standardization, to eliminate differences in the numerical range of different physical quantities and ensure that the model is not biased during training due to inconsistent feature scales. The z-score standardization formula is as follows: ; in, These are the standardized eigenvalues. These are the original eigenvalues. The mean, The standard deviation is denoted as .

[0029] Step S22: Adopt independent feature extraction and alignment strategies for different node types, namely: map multi-source observation data into a dynamic heterogeneous graph structure, and map the original feature vectors of different node types to a unified dimension through a linear projection layer specific to the node type. This step ensures that all nodes can be processed by the same spatial encoder, thereby achieving the alignment of cross-source heterogeneous features.

[0030] Because the feature dimensions of different types of nodes are not consistent—for example, ERA-5 nodes contain 3 feature dimensions, while CAMS mesh nodes contain only 1 feature dimension—in order to enable HGT to perform cross-type message passing, such as... Figure 3 As shown, this embodiment uses a multi-source feature dimension alignment module with linear projection to project features of different node types onto a unified 128 dimensions, as shown in the following formula: ; in, For node type, Embed the projected nodes. It is a linear projection matrix. This is the original feature vector for this node type. This is a bias term.

[0031] After performing independent linear transformations on each node type using a linear projection module, all nodes are represented as feature vectors of the same dimension, thus enabling consistent message passing and computation in subsequent heterogeneous graph networks.

[0032] In terms of spatial information processing, this embodiment retains the geographical location information of each type of node, including latitude and longitude, serial number coordinates or grid positioning number, so that the subsequent heterogeneous graph can identify its spatial adjacency relationship and enable the model to better capture spatial distribution patterns.

[0033] Regarding time information, this embodiment ensures that each time step snapshot contains a dedicated type of "time node." The time node records the time characteristics within the step, such as date, time sequence number, and period interval, for subsequent time series modeling. In the integration phase, this embodiment processes data generated by various data sources at the same time and combines them into a single graph snapshot. All graph snapshots are organized in chronological order to obtain a dynamic heterogeneous graph sequence for training. This approach not only solves the problem of inconsistent dimensions between data from different sources but also provides a standardized data structure foundation for subsequent heterogeneous graph learning and Transformer time series modeling.

[0034] Step S3: Construct a hybrid model combining HGT and Transformer; By constructing a hybrid model combining HGT and Transformer (HGT-Transformer), the quality of the generated XCO2 estimates is improved. Specifically, each time step corresponds to an independent graph snapshot, and all time steps together form a graph sequence. Within each snapshot, multiple types of nodes are defined based on different data sources (e.g., OCO-2 satellite inversion, CAMS reanalysis products, and ERA-5 meteorological fields) to reflect the heterogeneity of the input data. Each type of node contains multidimensional features, such as XCO2 concentration, wind field components, and temperature.

[0035] Spatial edges connect geographically adjacent nodes, constructed based on physical distance or correlation; temporal edges connect the representations of the same node at adjacent time steps, thus maintaining temporal continuity. Furthermore, cross-data source interaction edges are constructed to model cross-domain relationships between different datasets. This results in a heterogeneous graph at each time step t. This dynamic heterogeneous graph enables the model to simultaneously learn spatial, temporal, and cross-domain dependency features within a unified representation space.

[0036] In spatial modeling, HGT updates node representations based on type-aware multi-head attention and meta-relation message passing mechanisms. HGT learns independent transformation matrices for each node and edge type, thus preserving semantic differences and physical properties during neighborhood information aggregation. By assigning different attention weights to different types of neighbors, HGT can effectively model cross-domain interactions between OCO-2, CAMS, and ERA-5 nodes, generating more discriminative node embedding representations.

[0037] In terms of temporal modeling, Transformer takes observations at each time step (e.g., XCO2, CAMS-IO, wind speed) as input sequence elements. Utilizing a multi-head self-attention mechanism, the model can dynamically capture correlations between different time points, thereby extracting potential temporal patterns and modeling long-term dependencies.

[0038] The overall HGT–Transformer hybrid model structure is as follows: Figure 4 As shown, the system consists of a spatial modeling module and a temporal modeling module, used to learn the spatial autocorrelation and temporal dependence of satellite observation data, respectively. In the spatial modeling module, inputs include CAMS-EGG4, CAMS-IO, and ERA-5 wind fields, and spatial autocorrelation is modeled using an adjacency matrix constructed based on geographic coordinates and wind field information. After processing by the HGT module, the input features generate specific types of node embeddings, effectively capturing spatial autocorrelation across regions and data sources, providing rich spatial representations.

[0039] The temporal modeling module employs a standard Transformer encoder. Inputs are time-series images from CAMS-EGG4 and CAMS-IO. The model first uses positional encoding to enhance its sensitivity to temporal order, then extracts temporally dependent features through structures such as multi-head self-attention, feedforward networks, and residual connections. The feature dimensions output by the Transformer encoder are consistent with those of the spatial module, ensuring dimensionality matching during stitching and promoting the effective fusion of spatial and temporal features.

[0040] After concatenating spatial and temporal features along the channel dimension, the fused features are input into a two-layer feedforward neural network for further fusion and regression. The final linear layer maps the high-dimensional features to the predicted CO2 concentration. To prevent overfitting, dropout is introduced at key locations in the fusion model, and the Adam optimizer is used for backpropagation training. The mean squared error (MSE) loss function is used to measure the deviation between the predicted and observed values.

[0041] Step S4: Calculate and output the predicted XCO2 concentration using a hybrid model; The calculation steps for the hybrid model are as follows: Step S41: Input the unified node features and their heterogeneous relationships into the HGT module. HGT performs relationship-aware message passing on information between different types of nodes according to the dynamic relationship constructed in this invention, generating a node-level spatial representation as a spatial feature, as shown below: ; in, This represents the node embedding representation obtained after processing by the HGT module in the t-th graph snapshot. This represents the initial feature set of all nodes in the r-th time slice before they are input into the HGT. This represents the set of all edge relationships in a dynamic heterogeneous graph during that time slice.

[0042] Step S42: After obtaining the spatial features of various nodes, this embodiment further aggregates the spatial features to each time step according to the participation relationship of the time nodes: ; Where t is the time node. This is the spatial representation of the aggregated time nodes, and serves as the input to the temporal Transformer. For all spatial nodes pointing to time node t, Aggregate(·) is an aggregation operation used to combine all spatial nodes belonging to the same time.

[0043] Step S43: Each time step constitutes a time-series feature sequence of length T. The sequence is input to a time-varying Transformer encoder to extract the time dependency structure across days or hours, resulting in the following time series representation: ; Where μ is the output vector obtained after Transformer encoding.

[0044] Step S44: The system uses a lightweight regression prediction head to map the Transformer output vector to the final predicted XCO2 concentration value. ; Where y represents the XCO2 concentration value predicted by the model.

[0045] In this embodiment, the dataset source is shown in Table 1.

[0046] Table 1. Datasets used in this invention

[0047] Dataset Time range Time resolution Spatial resolution variable Download link CAMS-EGG4 2017-2020 3h 0.75° <![CDATA[XCO2]]> https: / / ads.atmosphere.copernicus.eu CAMS-IO 2017-2020 3h 1.4°×0.7° <![CDATA[XCO2]]> https: / / ads.atmosphere.copernicus.eu OCO-2 2017-2020 16d 2.25km × 1.29km <![CDATA[XCO2]]> https: / / search.earthdata.nasa.gov ERA-5 2017-2020 1h 0.25° wind speed https: / / cds.climate.copernicus.eu TCCON - - - <![CDATA[XCO2]]> https: / / tccondata.org The datasets are as follows: OCO-2: The Orbiting Carbon Observatory-2 (OCO-2) satellite was launched by NASA in July 2014 to monitor atmospheric CO2. The satellite operates in a sun-synchronous polar orbit at an altitude of 705 km, with an equatorial transit time of approximately 13:30 local time. Its onboard spectrometer measures atmospheric reflectance spectra in three spectral channels with center wavelengths of 0.76, 1.6, and 2.1 µm. Physical inversion algorithms are used to synthesize these spectral observations to obtain XCO2 measurements. OCO-2 has a spatial resolution of 2.25 km × 1.29 km and a temporal resolution of 16 days. This embodiment uses the latest version of ACOS / OCO-2 Level 2 XCO2 Lite product, such as... Figure 5 As shown.

[0048] CAMS-EGG4: The fourth version of the Copernicus Atmospheric Monitoring Service's global gridded product (CAMS-EGG4) reanalysis XCO2 dataset is generated based on multi-satellite observations and atmospheric simulations, providing global XCO2 measurements with a spatial resolution of 0.75° and a temporal resolution of 3 hours. Figure 6As shown in the figure, the validation report reveals a trend in the bias over time. For example, before 2014, the model tended to overestimate XCO2, followed by a period of underestimation, before overestimating again in 2020. This type of bias may be related to updates to the model's input data or changes in atmospheric conditions. This embodiment uses the latest version of the dataset, which provides global XCO2 estimates for 2020.

[0049] CAMS-IO: The Copernicus Atmospheric Monitoring Service Interpolated Output (CAMS-IO) greenhouse gas product is generated from multiple types of input data, including fossil fuel emissions, cement production emissions, fire emission fluxes, ocean fluxes, ground observations, and satellite measurements, and processed using the PYVAR integrated inversion system. The CAMS-IO product provides two independent datasets based on satellite observations and ground air sampling. This example uses CAMS-IO average CO2 data obtained from ground air sampling inversion, such as... Figure 7 As shown. In comparison, satellite-based products may exhibit higher consistency. This dataset has a spatial resolution of 1.4° × 0.7° and a temporal resolution of 3 hours.

[0050] TCCON: The Total Column Concentration Observatory Network (TCCON) provides reliable ground-based measurements of various greenhouse gases. Its CO2 measurements are widely used for satellite observations and model output validation. This embodiment uses TCCON XCO2 observation data generated based on the GGG2020 algorithm, focusing on observation data from the Hefei and Xianghe stations in China, such as... Figure 8 As shown Figure 9 As shown. A total of 149,290 observation records from 2018 to 2024 were ultimately used as example data.

[0051] ERA-5: In addition to the XCO2 dataset mentioned above, atmospheric transport processes are also a key factor controlling the spatial distribution of atmospheric CO2. Therefore, this embodiment incorporates wind field components and wind speed as additional input features into the model. The wind field data comes from the fifth-generation global atmospheric reanalysis (ERA-5) dataset, such as... Figure 10 As shown in the figure. This dataset covers the period from 2017 to 2020, providing wind field data with a spatial resolution of 0.25 and a temporal resolution of 1 hour.

[0052] The entire dataset was divided into training and test sets, with 80% of the data (2017–2019) used for training and the remaining 20% ​​(2020) used for performance evaluation. Given the significant time-series characteristics of the data, the data split strictly followed chronological order to avoid future information leakage. Specifically, the first 72% of the samples in the time series were used for the training set, the next 8% for the validation set, and the last 20% for the test set. No random shuffling was performed during the split to ensure that the training and evaluation processes followed the realistic prediction scenario of "predicting the future from the past." This strategy improved the reliability and generalization ability of the experimental results.

[0053] The experiments in this embodiment used a fixed set of hyperparameters, listed in Table 2. The model was trained using a fixed learning rate of 0.001, and an early stopping mechanism was triggered when the loss function showed no improvement in 50 consecutive iterations. Training ran for a maximum of 1,000 epochs, with mean squared error (MSE) as the loss function.

[0054] Table 2. Model Hyperparameter Settings

[0055] category name Numerical / Description data Number of input features 5 Data partitioning ratio Training set: 80%, Test set: 20% batch 2 HGT HGT number of layers 2 HGT attention head count 4 Hidden Dimensions 128 Transformer Transformer layer number 2 Number of attention heads 4 Feedforward layer dimension 128 Dropout Ratio 0.1 Sequence length 3 train Optimizer Adam (learning rate 0.0001) loss function Masked MSE Number of rounds Maximum 100 rounds Early stop mechanism No improvement after 20 rounds

[0056] To evaluate the effectiveness of each module in the HGT-Transformer model, an ablation experiment was designed in this embodiment. The model was compared and evaluated using three variants: retaining the HGT module for spatial feature extraction while removing the Transformer temporal modeling module and using a simple fully connected layer for output prediction; retaining the Transformer module for temporal feature extraction while flattening the input data into a sequence and removing the HGT spatial modeling module; and the proposed method, which retains both the HGT spatial modeling module and the Transformer temporal modeling module. The ablation experiment results are shown in Table 3.

[0057] Table 3. Ablation Experiment Results

[0058] Model RMSE MAE <![CDATA[R 2 ]]> HGT alone 1.35 1.01 0.896 standalone Transformer 2.12 1.44 0.82 HGT-Transformer 0.99 0.79 0.969

[0059] The results show that, compared with the fusion model, the model using only HGT performs better in RMSE and MAE, indicating that simple spatial modeling is insufficient to fully utilize temporal dependencies. The model using only Transformer also has lower accuracy than the fusion model, suggesting that ignoring spatial neighborhood information leads to insufficient capture of spatial patterns in CO2 column concentration inversion. The hybrid model combining HGT and Transformer performs best in RMSE, MAE, and R² indices, demonstrating that integrating spatial and temporal modeling is both necessary and effective in improving inversion accuracy.

[0060] This embodiment improves the quality of the CAMS-EGG4XCO2 dataset. To evaluate the improvement, the original data and the improved results were validated against XCO2 observations from the Henan and Xianghe TCCON ground stations. The matching criterion was to average the CAMS data and the improved results within a spatial window of ±1° and a time window of ±1 hour around each TCCON station. The validation results are as follows: Figure 11 and Figure 12 As shown, Figure 11 For raw CAMS data, Figure 12 The training prediction results are shown. The improved XCO2 estimate is approximately 0.38 higher on R² than the original data, significantly improving the accuracy of the original dataset.

[0061] This embodiment also compares the spatial distribution of the improved dataset with that of the original CAMS-EGG4 dataset. Both datasets exhibit similar XCO2 spatial patterns in terms of signal range and annual average, but show significant differences in XCO2 concentration. Figure 13 and Figure 14 The spatial distribution of mean XCO2 in the improved CAMS-EGG4 dataset for 2017 and 2020 is presented separately. The results show that the improved dataset exhibits higher XCO2 concentrations in 2017 and lower concentrations in 2020 compared to the original CAMS-EGG4 measurements. This difference is attributed to the underestimation in 2017 and the overestimation in 2020 by the CAMS-EGG4 data.

[0062] As can be seen, this embodiment constructs a heterogeneous spatiotemporal map by combining spatial proximity relationships with time series nodes, thereby explicitly expressing the inherent spatiotemporal dependency structure in atmospheric CO2 observations. At the same time, it adopts an HGT-Transformer fusion model, using HGT to capture irregular spatial structures and using the Transformer architecture to model long-term time dependencies. In spatiotemporal joint modeling, it integrates multi-source input features to achieve high accuracy and strong generalization ability. While maintaining the advantage of broad spatial coverage, it achieves observation accuracy comparable to ground-based observations.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any indirect modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for fusing multi-source spatiotemporal data of carbon dioxide column concentration, characterized in that, Includes the following steps: Step S1, Constructing a dynamic heterogeneous graph: Based on the existing dataset, dynamically construct a dynamic heterogeneous graph for multi-source XCO2 spatiotemporal modeling according to time steps; Step S2, Multi-source feature standardization: z-score standardization is performed on the original numerical features from the same source, and independent feature extraction and alignment strategies are adopted for different node types; Step S3: Construct a hybrid model combining HGT and Transformer: including a spatial modeling module and a temporal modeling module; In the spatial modeling module, the input features are used to model spatial autocorrelation through an adjacency matrix constructed based on geographic coordinates and wind field information, and specific types of node embeddings are generated. In the time modeling module, Transformer encoding is used to enhance the sensitivity to time order, and time-dependent features are extracted through structures such as multi-head self-attention, feedforward networks, and residual connections. Spatial and temporal features are spliced ​​together to form fused features, so as to obtain the predicted CO2 concentration using linear layer mapping; Step S4: Use a hybrid model to calculate and output the predicted XCO2 concentration.

2. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 1, characterized in that, In step S1, an independent graph snapshot is constructed at each time step, and the snapshots are arranged in sequence to form a time-series dynamic heterogeneous graph. The graph contains data nodes from different sources and at different resolutions, and multiple types of edges are automatically constructed based on spatial adjacency and cross-source semantic relationships.

3. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 2, characterized in that, In step S1, the edge type is defined based on the physical or statistical relationships between nodes, defining various meta-relationships: For grid points, construct the edges connecting the grid points based on the four-adjacency rule; For scattered points, a K-dimensional tree is used to find the neighbors of the observation points within a given radius r, and a local spatial association structure is constructed. Based on nearest neighbor or radius search, satellite observation nodes are connected to grid nodes at the same time to achieve cross-resolution alignment. Forward and backward time edges are constructed between adjacent time snapshots to achieve information continuity in the time series.

4. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 1, characterized in that, In step S2, a multi-source feature dimension alignment module using linear projection is used to project features of different node types onto a unified dimension.

5. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 1, characterized in that, In the spatial modeling module, a message passing mechanism is used to update node representations, and independent transformation matrices are learned for each node type and edge type. Different attention weights are assigned to different types of neighbors to generate more discriminative node embedding representations. In the time modeling module, the Transformer takes the observations at each time step as input sequence elements and uses a multi-head self-attention mechanism to dynamically capture the correlation between different time points.

6. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 5, characterized in that, In step S3, the feature dimension output by the Transformer encoder is consistent with that of the spatial module.

7. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 6, characterized in that, In step S3, the fused features are input into a two-layer feedforward neural network for further fusion and regression.

8. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 1, characterized in that, In step S4, the calculation steps of the hybrid model are as follows: Step S41: Input the unified node features and their heterogeneous relationships into the HGT module and generate node-level spatial features; Step S42: Aggregate spatial features to each time step according to the participation relationship of time nodes; Step S43: Input the time series feature sequence formed by each time step into the time Transformer encoder to extract the time dependency structure across days or hours and obtain the time series representation; Step S44: Using a lightweight regression prediction head, the Transformer output vector is mapped to the final XCO2 concentration prediction value.

9. The method for fusing multi-source spatiotemporal data of carbon dioxide column concentration according to claim 1, characterized in that, In step S41, the expression for generating node-level spatial features is as follows: ； in, This represents the node embedding representation obtained after processing by the HGT module in the t-th graph snapshot. This represents the initial feature set of all nodes input into the HGT at time slice t. This represents the set of all edge relationships in a dynamic heterogeneous graph during that time slice. In step S42, the expression for aggregating spatial features to each time step is as follows: ； Where t is the time node. This is the spatial representation of the aggregated time nodes, and serves as the input to the temporal Transformer. Aggregate(·) is a subset of all spatial nodes pointing to time node t. Aggregate(·) is an aggregation operation used to combine all spatial nodes belonging to the same time. In step S43, the timing representation expression is as follows: ； Where T is the length of the time-series feature sequence. is a temporal feature sequence, and μ is the output vector obtained after Transformer encoding; In step S43, the expression for the final predicted XCO2 concentration is as follows: ； Where y represents the XCO2 concentration value predicted by the model.

Images