Greenhouse gas weak signal decoupling inversion method based on physical residual learning

Abstract

This invention discloses a method for decoupling and inverting weak greenhouse gas signals based on physical residual learning, belonging to the field of satellite remote sensing data processing and atmospheric parameter inversion technology. This method aims to solve the problem in infrared hyperspectral inversion where weak CO2 signals are masked by strong atmospheric temperature and humidity background signals, making it difficult for AI models to extract true CO2 change characteristics. Its core is to construct a three-stage processing architecture of "physical background stripping—residual feature mapping—incremental reconstruction." First, the physical background spectrum is simulated using a radiative transfer model, and the spectral residuals are calculated with the measured spectrum to strip away temperature and humidity interference. Then, a deep residual network is used to learn features from the spectral residuals to predict the increment of CO2 concentration relative to the baseline profile. Finally, the baseline profile is superimposed to obtain the final CO2 vertical profile. This invention combines the physical mechanism of atmospheric radiative transfer with the advantages of deep learning, improving the anti-interference ability, generalization, and physical interpretability of CO2 inversion.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing data processing and atmospheric parameter inversion technology. Specifically, it relates to a method for high-precision inversion of carbon dioxide (CO2) vertical profiles using infrared hyperspectral observation data by combining the physical mechanism of atmospheric radiative transfer with a deep residual network. Background Technology

[0002] As a major greenhouse gas, carbon dioxide (CO2) requires high-precision acquisition of its atmospheric vertical distribution profile for studying the global carbon cycle, climate change, and assessing greenhouse gas emission reduction. Infrared hyperspectral satellite detection (such as FY-3D / HIRAS), with its advantages of high spectral resolution and wide coverage, has become a core technology for atmospheric CO2 profile inversion.

[0003] Currently, CO2 inversion methods for infrared hyperspectral data are mainly divided into two categories: physical inversion methods and deep learning methods based on statistical regression. Physical inversion methods (such as the Optimal Estimation Method, OEM) construct a cost function based on the radiative transfer equation and find the optimal solution through multiple iterations. However, they suffer from extremely low computational efficiency, with single-point inversion taking several seconds. Furthermore, they are highly dependent on the initial guess (prior field). If the prior field has a large bias, the iterative process is prone to getting trapped in local minima, affecting the inversion accuracy.

[0004] Deep learning methods based on statistical regression construct deep neural networks and utilize a large number of spectral-parameter pairs for supervised training to establish a nonlinear statistical mapping relationship between satellite-observed radiance and atmospheric parameters, achieving end-to-end parameter inference. While this solves the problem of low efficiency in physical inversion methods, it still has significant drawbacks: First, the lack of explicit constraints on radiative transfer mechanisms makes it easy to learn spurious correlations in the data, incorrectly attributing temperature-induced spectral changes to CO2 concentration changes, resulting in inversion results containing temperature-related interference errors that are difficult to separate. Second, the extrapolation and generalization capabilities for "outside-domain" data are insufficient; essentially, it is a statistical interpolation based on the training dataset. When actual observation scenarios (such as extreme stratospheric warming or unknown aerosol types) exceed the statistical distribution range of the training dataset, it is prone to outputting erroneous profiles that violate atmospheric physics laws. Third, the interpretability of black-box models is poor; the end-to-end mapping masks the physical mechanisms of the sensitivity of spectral channels to different atmospheric stratifications, making it difficult to quantify the specific contributions of different channels to the inversion results and to conduct physical source analysis of inversion errors.

[0005] In summary, existing CO2 inversion methods cannot effectively solve the core problem of weak CO2 signals being masked by strong ambient temperature and humidity background signals in infrared hyperspectral inversion. This results in inversion accuracy, anti-interference ability, and generalization ability failing to meet the needs of practical applications. Therefore, there is an urgent need for a high-precision inversion method that can effectively decouple weak CO2 signals from ambient temperature and humidity background signals while taking into account both physical rationality and inversion efficiency. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for decoupling and inverting weak greenhouse gas signals based on physical residual learning. This method solves the problem that weak CO2 signals are masked by strong atmospheric temperature and humidity background signals in infrared hyperspectral inversion, making it difficult for AI models to extract the true CO2 change features, and achieves high-precision inversion of CO2 vertical profiles.

[0007] Technical solution: The present invention provides a method for decoupling and inverting weak greenhouse gas signals based on physical residual learning, comprising the following steps:

[0008] S1. Data preparation and preprocessing: acquire satellite infrared hyperspectral observation data, meteorological background field data and CO2 true value data, filter the data, perform quality control and spatiotemporal matching interpolation, and construct an observation-background-true value aligned dataset;

[0009] S2. Simulation and residual construction of physical reference radiation field: Set CO2 reference profile, combine meteorological background field data with CO2 reference profile into mixed atmospheric state vector, input forward radiative transfer model to calculate physical simulation background spectrum, subtract satellite measured spectrum from physical simulation background spectrum channel by channel to obtain spectral residual vector;

[0010] S3. Construction and training of incremental residual mapping network: Select spectral channels sensitive to CO2 as input features, construct a deep residual network, take the spectral residual vector as input and the difference between the true value of CO2 and the baseline reference profile as label, train the network to realize the mapping of spectral residual to CO2 concentration increment.

[0011] S4. Online Inversion and Profile Reconstruction: Real-time satellite observation data and meteorological forecast field data are accessed, and real-time spectral residuals are calculated. The CO2 concentration increment is obtained by inputting the trained deep residual network. The increment is superimposed with the CO2 reference profile to obtain the final CO2 vertical profile.

[0012] Furthermore, in step S1, the satellite infrared hyperspectral observation data is the first-level radiation observation data of the HIRAS detector carried by the FY-3D satellite, including spectral radiance and satellite geometric observation parameters; the meteorological background field data is ERA5 reanalysis data, including meteorological parameters such as stratified atmospheric temperature profiles and stratified specific humidity profiles; the true CO2 data is profile data provided by the Global Carbon Assimilation System (such as CarbonTracker) or the Copernicus Atmospheric Monitoring Service (CAMS), or high-precision column concentration products from satellites such as OCO-2 (when using column concentration data, a pressure weighting function is introduced as an observation operator to convert the profile integral into column concentration) or ground TCCON station and flight (such as CARIBIC) observation data.

[0013] Furthermore, the quality control in step S1 includes spectral apodization and cloud detection screening. The spectral apodization uses a Hamming or Blackman-Harris window function, and the cloud detection screening uses a MERSI cloud mask product to remove pixels with more than 5% cloud cover in the field of view.

[0014] Furthermore, the CO2 reference profile mentioned in step S2 is a multi-year climatic average that varies with latitude and month, and the positive radiative transfer model is the RTTOV model.

[0015] Furthermore, the spectral channels selected in step S3 are 100-200 channels of the 15μm CO2 absorption band of the HIRAS detector, and the deep residual network is a ResNet-1D network containing 5-10 stacked residual blocks, with the residual blocks embedding an SE attention mechanism module.

[0016] Furthermore, in step S3, mean squared error is used as the loss function, and the Adam optimizer is used to train the network until the loss function converges.

[0017] Furthermore, in step S4, the meteorological forecast field data is NCEP GFS forecast data, and the real-time spectral residual is calculated using the real-time meteorological forecast field data and the CO2 reference profile.

[0018] The present invention also discloses a greenhouse gas weak signal decoupling inversion system based on physical residual learning, including a data preprocessing module, a physical background simulation and residual calculation module, a network training module and an online inversion module. The modules are connected in sequence to work together to complete the inversion of CO2 vertical profile.

[0019] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0020] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0021] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0022] 1. Strong anti-interference capability: Compared with the pure AI model, this solution explicitly removes more than 90% of atmospheric temperature and humidity background interference through the physical radiation transmission model. The AI ​​model only needs to deal with the "residual error", which greatly reduces the false inversion caused by temperature and humidity coupling, significantly improves the anti-interference capability of CO2 inversion, and solves the core problem of weak CO2 signal being masked by strong temperature and humidity background.

[0023] 2. Improved generalization and physical interpretability: The model is based on the physical equations of atmospheric radiative transfer, and the AI ​​model only plays the role of "repairing residuals". Even under extreme observation conditions (such as extreme stratospheric warming), the physical model can still ensure the rationality of the baseline CO2 profile, avoiding non-physical abrupt changes caused by pure AI models. At the same time, the physical meaning of the construction of spectral residuals is clear, which can quantify the contribution of different spectral channels to CO2 inversion, realize the physical source of inversion error, and improve the interpretability of the model.

[0024] 3. High data utilization and high training efficiency: This residual learning mode allows the network to learn the complex physical laws of atmospheric radiation transfer from scratch, only the weak changes in CO2 signal. This not only speeds up the network convergence speed, but also reduces the requirement for training sample size. Compared with end-to-end pure AI models, it can achieve higher inversion accuracy with the same sample size.

[0025] 4. High practicality and wide range of applications: This method is compatible with infrared hyperspectral satellite data such as FY-3D / HIRAS, and can achieve high-precision and rapid inversion of CO2 vertical profiles. It can be widely used in fields such as global carbon cycle research, climate change monitoring, and greenhouse gas emission reduction assessment, and has important practical application value. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention.

[0027] Figure 2 This is a detailed flowchart of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, the present invention provides a method for decoupling and inverting weak greenhouse gas signals based on physical residual learning, comprising the following steps:

[0030] Step (1): Data preparation and preprocessing

[0031] The core objective of this step is to construct an observation-background-truth dataset that is strictly aligned in time, space, and physical state, providing high-quality data support for subsequent physical simulations and network training.

[0032] (1.1) Acquisition of multi-source data:

[0033] Satellite observation data: Acquire Level 1 (L1) radiation observation data from the Infrared Hyperspectral Atmospheric Vertical Sounder (HIRAS) of Fengyun-3D satellite (FY-3D). The data includes the spectral radiance after Fourier transform of the interferogram, as well as geometric observation parameters such as satellite zenith angle and azimuth angle. These parameters are the basis for subsequent spectral processing and physical simulation.

[0034] Meteorological background field data: ERA5 reanalysis data released by the European Centre for Medium-Range Weather Forecasts (ECMWF) were obtained, and stratified atmospheric temperature profiles, stratified specific humidity profiles, stratified ozone profiles, surface temperature, 2-meter air temperature, surface air pressure, and 10-meter wind field data were extracted. These meteorological parameters are key factors affecting the atmospheric radiation transfer process and directly determine the simulation accuracy of the background spectrum.

[0035] True carbon dioxide data: As 'labels' for model training, high-precision carbon dioxide profile data provided by global carbon assimilation systems (such as CarbonTracker) or the Copernicus Atmospheric Monitoring Service (CAMS) are preferred. Alternatively, Fourier transform spectrometer observation data from ground-based TCCON stations or high-precision column concentration products from satellites such as OCO-2 can be used. When using column concentration data, a pressure weighting function is introduced as an observation operator at the network output to convert the predicted profile integral into column concentration for calculation, ensuring the accuracy and reliability of the labels and providing a true reference for network training.

[0036] (1.2) Data screening and quality control:

[0037] Spectral apodization: The raw HIRAS spectrum is apodized, preferably using the Hamming or Blackman-Harris window function. The purpose is to suppress the spectral sidelobe ringing effect caused by the truncation of the interferogram and reduce the interference of spectral noise on the subsequent inversion process.

[0038] Cloud detection and removal: Using the MERSI (Medium Resolution Spectral Imager) cloud mask product on the same platform as the FY-3D, pixels with cloud cover exceeding 5% within the field of view are removed, retaining only the clear sky field of view (Clear Sky FOV). Since the infrared band cannot penetrate clouds, the presence of clouds completely destroys the underlying CO2 information, rendering spectral observations worthless. Therefore, cloud detection and removal is a crucial step in ensuring data validity.

[0039] (1.3) Spatiotemporal matching interpolation:

[0040] Considering that the satellite observation points and the grid points of the ERA5 reanalysis data do not overlap, bilinear interpolation is used to interpolate the ERA5 meteorological field data to the center latitude and longitude of each HIRAS observation cell to achieve spatial alignment. In the time dimension, the two ERA5 time intervals closest to the HIRAS observation time (usually with an interval of 1 hour) are selected for linear interpolation to ensure that the meteorological background field can accurately reflect the atmospheric state at the instant of observation and reduce the error caused by spatiotemporal misalignment.

[0041] Step (2): Simulation and residual construction of the physical reference radiation field

[0042] This step is the core innovation of the present invention. By explicitly simulating atmospheric background radiation through a physical model, the CO2 signal is initially decoupled from the temperature and humidity background signal, most of the background interference is removed, and the weak changes in the CO2 signal are highlighted.

[0043] (2.1) Define the reference state:

[0044] A baseline carbon dioxide profile vector, denoted as XCO2_ref, is constructed. In this embodiment, XCO2_ref is set to a multi-year climatic mean that varies with latitude and month. For example, for an observation point at 30 degrees North latitude in January, the values ​​of this vector at all pressure layers are fixed prior values. This baseline profile is known and deterministic at the input end, and does not include the actual daily CO2 anomaly fluctuations. Its function is to provide a stable background benchmark for separating anomalous changes in CO2.

[0045] (2.2) Running the forward radiative transfer model:

[0046] RTTOV (Radiative Transfer for TOVS) was selected as the forward radiative transfer model. This model has the advantages of high computational efficiency and high simulation accuracy, and is suitable for radiative transfer simulation of infrared hyperspectral data. The real meteorological background field (temperature, water vapor, and surface parameters) obtained in step (1) was combined with the baseline CO2 profile (XCO2_ref) in step (2.1) to construct a "mixed atmospheric state vector". This vector contains complete information on atmospheric temperature and humidity background, surface condition, and baseline CO2 concentration. The mixed atmospheric state vector was input into the RTTOV model to simulate and calculate the theoretical top spectral radiance under this mixed state, which is denoted as the physical simulation background spectrum Yphy. This spectrum mainly reflects the radiation characteristics under the combined effect of atmospheric temperature and humidity background, surface condition, and baseline CO2 concentration, that is, most of the radiation signals caused by non-CO2 anomaly changes.

[0047] (2.3) Calculate the spectral residuals:

[0048] The spectral residual vector ΔY is obtained by subtracting the measured HIRAS spectrum Yobs from the physically simulated spectrum Yphy channel by channel. The formula for calculating ΔY is as follows:

[0049] ΔY = Yobs - Yphy

[0050] In this spectral residual vector, most of the radiation signals caused by atmospheric temperature and humidity background, surface conditions and reference CO2 concentration have been stripped away. It mainly contains the radiation signals caused by real CO2 anomalies, as well as a small amount of background noise that has not been completely stripped away. This achieves the initial decoupling of weak CO2 signals from strong temperature and humidity background signals, laying the foundation for subsequent AI models to extract real CO2 change features.

[0051] Step (3): Construction and training of incremental residual mapping network

[0052] This step involves constructing a deep residual network to accurately learn the weak CO2 signal in the spectral residual vector, thereby mapping the spectral residual to the CO2 concentration increment and further improving the inversion accuracy.

[0053] (3.1) Channel feature selection:

[0054] The HIRAS detector has 2275 channels, but not all channels are sensitive to CO2. Some channels are significantly affected by other gases such as water vapor and ozone. If all of these channels were used for network training, redundant information and interference would be introduced, reducing the network training efficiency and inversion accuracy. Therefore, in this embodiment, approximately 100-200 channels located in the 15μm CO2 absorption band are selected as input features. In this scheme, the Jacobian sensitivity matrix of the HIRAS full spectrum to CO2 concentration is calculated using the forward radiative transfer model. A sensitivity threshold is set, and channels with high cross-sensitivity to water vapor and ozone are eliminated. Finally, approximately 100-200 channels located in the 15μm CO2 absorption band and with obvious peak values ​​of the weight function in the lower troposphere are selected as input features. These channels contain the main absorption characteristics of CO2 in the lower troposphere, are sensitive to changes in CO2 concentration, and are less affected by other gases, which can effectively improve the learning efficiency and inversion accuracy of the network. (3.2) Network structure design:

[0055] A deep residual network (ResNet-1D) is constructed to adapt for feature learning of one-dimensional spectral residual data. The network mainly consists of an input layer, a feature extraction layer, and an output layer.

[0056] Input layer: Receives the standardized spectral residual ΔY. Standardization can eliminate the influence of differences in spectral amplitudes between different channels and accelerate network convergence.

[0057] Feature extraction layer: Contains 5-10 stacked residual blocks. Each residual block consists of two one-dimensional convolutional layers (Conv1D), a batch normalization layer, and a ReLU activation function. The kernel size is preferably 3 or 5, which can effectively capture the correlation features between adjacent spectral channels and extract the hidden CO2 signal variation patterns in the spectral residuals. The batch normalization layer can alleviate the gradient vanishing problem during network training and improve training stability. The ReLU activation function can introduce nonlinear mapping and enhance the network's ability to fit complex features.

[0058] Attention mechanism (optional, preferred): An SE (Squeeze-and-Excitation) module is embedded in the residual block. This module can adaptively weight different spectral channels, automatically suppress channels with poor signal-to-noise ratio, enhance attention to the core channel of CO2 absorption line, and further improve the network's ability to extract weak CO2 signals.

[0059] Output layer: A fully connected layer is used, with the output dimension corresponding to the number of vertical atmospheric layers (e.g., 20 layers), to output the CO2 concentration increment of each atmospheric layer.

[0060] (3.3) Definition of prediction target:

[0061] Unlike existing technologies that directly predict the absolute CO2 concentration, the output of the network in this invention is not the absolute CO2 concentration, but rather a concentration increment vector ΔX. The training label is defined as: the true CO2 profile (from the ground truth data in step (1)) minus the baseline reference profile XCO2_ref, calculated as: ΔXlabel = Xtrue - XCO2_ref. This incremental learning strategy reduces the dynamic range of the neural network's fitting function, making it easier for the network to capture minute changes in CO2 concentration and improving the accuracy of inversion for weak signals.

[0062] (3.4) Model training:

[0063] The network update is guided by a joint loss function (Loss), which includes not only the mean squared error term but also a vertical smoothing regularization term to eliminate non-physical inter-layer oscillations. The formula for calculating the loss function is as follows:

[0064] In the formula, the first term is the mean square error between the predicted increment and the actual increment; the second term is the second-order difference smoothing constraint term, Δx. iLet γ represent the concentration increment of the i-th layer, and γ be the smoothing weight hyperparameter. The Adam optimizer is used for gradient descent iterations, with appropriate learning rates and iteration counts, until the loss function converges, resulting in a trained deep residual network model. This model can achieve accurate mapping from spectral residuals to CO2 concentration increments.

[0065] Step (4): Online inversion and profile reconstruction

[0066] This step describes the operational process during actual business operation after the model training is completed, realizing real-time, high-precision inversion of the CO2 vertical profile.

[0067] (4.1) Real-time data stream access: Receive real-time HIRAS observation data streams, including real-time spectral radiance and geometric observation parameters; at the same time, receive real-time updated numerical weather prediction fields (such as NCEP GFS forecast data) for real-time background spectrum simulation.

[0068] (4.2) Forward physical noise reduction: Based on the temperature and humidity data of the real-time numerical weather forecast field and the built-in fixed reference CO2 profile (XCO2_ref), the RTM model (RTTOV) is called to quickly calculate the current background spectrum Yphy; the real-time observed spectrum Yobs is subtracted from the background spectrum Yphy channel by channel to calculate the real-time spectral residual ΔY, thus completing the removal of real-time background interference.

[0069] (4.3) Network inference: After standardizing the real-time calculated spectral residual ΔY, it is input into the trained deep residual network, and the network quickly outputs the predicted CO2 concentration increment vector ΔXpred.

[0070] (4.4) Final Result Reconstruction: The predicted CO2 concentration increment vector ΔXpred is superimposed with the baseline reference profile XCO2_ref to obtain the final inverted vertical carbon dioxide distribution profile, which is calculated as follows:

[0071] Xfinal_result = XCO2_ref + ΔXpred

[0072] The output Xfinal_result is a CO2 vertical profile product that meets the needs of practical applications, and can be further used to generate global or regional CO2 concentration distribution maps.

[0073] Example:

[0074] CO2 inversion based on FY-3D HIRAS data

[0075] This embodiment uses data from the FY-3D satellite's HIRAS detector to achieve high-precision inversion of the CO2 vertical profile. The specific steps are as follows:

[0076] (1) Data preparation:

[0077] FY-3D HIRAS L1 observation data from 2020 to 2023 were selected, covering major global regions, with a temporal resolution of 1 day and a spatial resolution of 13km × 13km. The corresponding ERA5 reanalysis data were selected as meteorological background data, with a temporal resolution of 1 hour and a spatial resolution of 0.25° × 0.25°, and parameters such as stratified atmospheric temperature, specific humidity, and surface temperature were extracted. Ground TCCON station data (such as Beijing Shangdianzi Station and Qinghai Waliguan Station) and high-precision column concentration products from GOSAT / OCO-2 satellites were selected as CO2 label ground truth values ​​to ensure that the label data is spatiotemporally aligned with the HIRAS observation data.

[0078] Channel selection: From the 2275 channels of HIRAS, 150 strong absorption channels located in the 15μm CO2 absorption band (wavenumber 650-760 cm⁻¹) were selected using the information entropy method. Channels that were severely affected by O3 and water vapor were removed to ensure that the selected channels were sensitive to changes in CO2 concentration.

[0079] (2) Data preprocessing:

[0080] Spectral apodization: Hamming window function is used to apodize the raw HIRAS spectrum to suppress spectral sidelobe ringing effect; Cloud detection and removal: MERSI cloud mask product is used to remove pixels with more than 5% cloud cover in the field of view, retaining only clear-sky pixels; Spatiotemporal matching interpolation: Bilinear interpolation is used to interpolate ERA5 meteorological data to the latitude and longitude of the HIRAS observation pixel center. In the time dimension, the two ERA5 time intervals closest to the observation time are selected for linear interpolation to construct an "observation-background-true value" aligned dataset.

[0081] (3) Physical background simulation and residual construction:

[0082] Constructing the CO2 baseline reference profile XCO2_ref: Using global CO2 multi-year climate average data, the baseline CO2 profiles for different regions and months are determined by latitude (each 10° interval) and month (12 months), resulting in a total of 36 sets of baseline profiles to ensure that the baseline profiles can reflect the climate characteristics of different regions.

[0083] The RTTOV-13 fast radiative transfer mode was integrated, and a physical simulator (RTM Interface) was constructed. The input interface was set to receive parameters such as temperature, humidity, and surface temperature from ERA5, and the CO2 profile was fixed as the aforementioned reference profile. The preprocessed meteorological background data and the reference CO2 profile were combined into a mixed atmospheric state vector, which was then input into the RTTOV-13 model to simulate and calculate the physical background spectrum Yphy. The measured HIRAS spectrum Yobs was subtracted from Yphy channel by channel to obtain the spectral residual vector ΔY.

[0084] (4) Network construction and training:

[0085] A ResNet-1D deep residual network is constructed, which contains 7 stacked residual blocks. Each residual block consists of two Conv1D layers (with a kernel size of 3), a Batch Normalization layer, and a ReLU activation function. An SE attention mechanism module is embedded in each residual block. The input layer has a dimension of 150 (corresponding to the residual values ​​of 150 optimized channels), and the output layer has a dimension of 20 (corresponding to the CO2 concentration increment of 20 atmospheric layers).

[0086] The spectral residual vector ΔY is standardized and used as the network input. The difference between the TCCON / GOSAT / OCO-2 ground truth data and the baseline reference profile is used as the training label. The joint loss function (Loss) is used as the loss function, and the Adam optimizer (learning rate set to 0.001, number of iterations set to 100) is used for training until the loss function converges, thus obtaining the trained network model.

[0087] (5) Online inversion:

[0088] It receives real-time HIRAS L1 observation data and NCEP GFS forecast field data (time resolution of 6 hours); using a physics simulator, it quickly calculates the background spectrum Yphy using temperature and humidity data from the GFS forecast field and a built-in baseline CO2 profile, and calculates the residual ΔY between the real-time observed spectrum and the background spectrum; after standardizing ΔY, it inputs it into the trained network model to obtain the CO2 concentration increment ΔXpred; and then overlays ΔXpred with the baseline reference profile to obtain the final CO2 vertical profile product, generating a global CO2 concentration distribution map (spatial resolution 13km×13km, time resolution 1 day) for practical operational applications.

[0089] The inversion results of this embodiment show that the CO2 profile inversion accuracy of this method is more than 15% higher than that of the pure AI model, and the efficiency is more than 100 times higher than that of the traditional physical inversion method. It can effectively remove the background interference of temperature and humidity, accurately extract weak CO2 signals, and meet the actual needs of global carbon cycle research and climate change monitoring.

[0090] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A greenhouse gas weak signal decoupling inversion method based on physical residual learning, characterized by, Includes the following steps: S1. Data preparation and preprocessing: acquire satellite infrared hyperspectral observation data, meteorological background field data and CO2 true value data, filter the data, perform quality control and spatiotemporal matching interpolation, and construct an observation-background-true value aligned dataset; S2. Simulation and residual construction of physical reference radiation field: Set CO2 reference profile, combine meteorological background field data with CO2 reference profile into mixed atmospheric state vector, input forward radiative transfer model to calculate physical simulation background spectrum, subtract satellite measured spectrum from physical simulation background spectrum channel by channel to obtain spectral residual vector; S3. Construction and training of incremental residual mapping network: Select spectral channels sensitive to CO2 as input features, construct a deep residual network, take the spectral residual vector as input and the difference between the true value of CO2 and the baseline reference profile as label, train the network to realize the mapping of spectral residual to CO2 concentration increment. S4. Online Inversion and Profile Reconstruction: Real-time satellite observation data and meteorological forecast field data are accessed, and real-time spectral residuals are calculated. The CO2 concentration increment is obtained by inputting the trained deep residual network. The increment is superimposed with the CO2 reference profile to obtain the final CO2 vertical profile.

2. The method of claim 1, wherein, In step S1, the satellite infrared hyperspectral observation data is the first-level radiation observation data of the HIRAS detector carried by the FY-3D satellite, including spectral radiance and satellite geometric observation parameters; the meteorological background field data is ERA5 reanalysis data, including meteorological parameters such as stratified atmospheric temperature profiles and stratified specific humidity profiles; the CO2 true value data is OCO-2, GOSAT satellite high-precision column concentration products or ground TCCON station and flight observation data.

3. The method of claim 1, wherein, The quality control in step S1 includes spectral apodization and cloud detection screening. The spectral apodization uses a Hamming or Blackman-Harris window function, and the cloud detection screening uses a MERSI cloud mask product to remove pixels with more than 5% cloud cover in the field of view.

4. The method of claim 1, wherein, The CO2 reference profile mentioned in step S2 is a multi-year climatic average that varies with latitude and month, and the positive radiative transfer model is the RTTOV model.

5. The method of claim 1, wherein, In step S3, the selected spectral channels are 100-200 channels of the 15μm CO2 absorption band of the HIRAS detector. The deep residual network is a ResNet-1D network containing 5-10 stacked residual blocks, and the residual blocks embed an SE attention mechanism module.

6. The method of claim 1, wherein, In step S3, mean squared error is used as the loss function, and the Adam optimizer is used to train the network until the loss function converges.

7. The method of claim 1, wherein, In step S4, the meteorological forecast field data is NCEPGFS forecast data, and the real-time spectral residual is calculated using the real-time meteorological forecast field data and the CO2 baseline reference profile.

8. A system for implementing the method of any one of claims 1 to 7, characterized in that, It includes a data preprocessing module, a physical background simulation and residual calculation module, a network training module, and an online inversion module. These modules are connected in sequence and work together to complete the inversion of the CO2 vertical profile.

9. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 8. The processor executes the computer program to implement the steps of the method of claim 1.

10. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images