Method and system for limb atmospheric profile retrieval based on two-dimensional radiative transfer constraint
By employing a deep learning method based on two-dimensional radiative transfer constraints, and utilizing a temporal convolutional neural network and a differentiable radiative transfer module, the problems of insufficient inversion accuracy and low computational efficiency of limb detection in high gradient regions are solved, achieving efficient and accurate two-dimensional atmospheric profile inversion.
Patent Information
- Application Number
- CN202610223582.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-25
- Publication Date
- 2026-06-23
Smart Images
Figure CN122262995A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of atmospheric remote sensing inversion technology, specifically to a method and system for inverting lateral atmospheric profiles based on two-dimensional radiative transfer constraints. Background Technology
[0002] Limb sounding is a widely used atmospheric remote sensing technique that uses spaceborne or airborne instruments to observe the atmosphere "at the edge." Limb sounding observations cover the microwave, infrared, visible, and ultraviolet bands, observing both thermal radiation and scattered sunlight or airglow. Limb sounding offers several advantages in atmospheric temperature and composition retrieval: higher vertical resolution, the ability to perform retrievals at higher altitudes, and greater sensitivity to trace gases.
[0003] Traditional edge detection inversion methods generally employ the one-dimensional spherically symmetric atmospheric assumption, which assumes that atmospheric parameters only vary with altitude and are uniformly distributed in the horizontal direction. However, in actual edge observations, the detector's line-of-sight path traverses hundreds of kilometers of atmospheric layers, and horizontal gradients in temperature, pressure, gas concentration, and aerosol extinction inevitably exist along the path.
[0004] The systematic bias caused by this "spherical symmetry assumption" is particularly prominent in the following scenarios:
[0005] (1) Edge region of polar vortex: There are severe horizontal gradients of trace gases such as ozone and nitrogen oxides;
[0006] (2) After a volcanic eruption: Aerosols exhibit a highly uneven spatial distribution;
[0007] (3) Near the terminator: rapid changes in photochemical state lead to uneven concentration levels of components;
[0008] (4) Tropopause fold region: The stratosphere-tropospheric exchange region has a complex three-dimensional structure.
[0009] Existing research employs tomographic methods to address this problem by simultaneously inverting two-dimensional along-track profiles from multiple edge scans. Other studies utilize prior information about changes in the level of state variables during a single edge scan inversion. While these two-dimensional radiative transfer correction methods improve accuracy, they increase computational complexity by one to two orders of magnitude, making them unsuitable for operational use. Summary of the Invention
[0010] The technical problem to be solved by this invention is to provide a method and system for inverting lateral atmospheric profiles based on two-dimensional radiative transfer constraints, which automatically extracts the horizontal gradient features of atmospheric parameters from adjacent observation sequences along the track, effectively models the horizontal inhomogeneity on the lateral line-of-sight path while maintaining the high computational efficiency of deep learning methods, significantly improves the inversion accuracy of high gradient regions such as polar vortex edges and volcanic eruptions, and outputs diagnostic information such as the horizontal gradient field and two-dimensional correction contribution, thereby enhancing the interpretability of the inversion results.
[0011] To solve the above problems, the present invention adopts the following technical solution:
[0012] This invention proposes a method for inverting lateral atmospheric profiles based on two-dimensional radiative transfer constraints, comprising the following steps:
[0013] S1. Obtain the marginal radiation data of the target observation point. and its adjacent observation sequences along the track , where K is the number of adjacent points;
[0014] S2. Extract horizontal gradient feature vectors from adjacent observation sequences using a temporal convolutional neural network. ;
[0015] S3. Process target point observations using the first neural network. Predicting the initial profile under the assumption of one-dimensional spherical symmetry ;
[0016] S4, Fusion of horizontal gradient feature vectors Compared with the initial profile The second neural network is used to predict the two-dimensional correction amount. and horizontal gradient field The corrected profile was obtained ;
[0017] S5. Using a differentiable simplified two-dimensional radiative transfer module, based on the corrected profile... and horizontal gradient field Calculation of simulated radiation ;
[0018] S6. The neural network is trained end-to-end using a multi-task loss function that includes two-dimensional radiative transfer consistency constraints.
[0019] S7, Output the two-dimensional corrected atmospheric profile And related diagnostic information.
[0020] Preferably, in step S2, the temporal convolutional neural network adopts a causal dilated convolutional structure, including:
[0021] The convolutional blocks are multi-layered with causal dilation, and the dilation coefficient of each layer increases in power of 2, i.e., the dilation coefficient sequence is {1, 2, 4, 8, ...}.
[0022] Each convolutional block contains, in sequence, a one-dimensional causal convolutional layer, a batch normalization layer, a ReLU activation function, and a residual connection;
[0023] The receptive field of the temporal convolutional neural network covers K observation points before and after the orbital direction, for a total of 2K+1 points, where K ranges from 3 to 7.
[0024] Preferably, in step S4, the horizontal gradient feature vectors are fused. Compared with the initial profile The fusion uses an attention mechanism to perform feature fusion, which combines the horizontal gradient feature vector with the initial profile feature in a weighted manner.
[0025] Preferably, in step S5, the differentiable simplified two-dimensional radiative transfer module adopts the linear horizontal gradient assumption, and the atmospheric parameters... Represented as:
[0026] ;
[0027] in, This is the vertical profile at the target point. The horizontal gradient along the track direction. The coordinates are along the track.
[0028] Preferably, in step S5, the differentiable simplified two-dimensional radiative transmission module discretizes the line-of-sight path as follows: The simulated radiation is calculated using the following steps at each sampling point:
[0029] S5.1: Calculate each sampling point Atmospheric parameters at the location: ;
[0030] S5.2: Calculate cumulative optical thickness: ;
[0031] S5.3: Calculate radiation using discrete integrals: ;
[0032] in For the source function, Extinction coefficient, The sampling interval is denoted as .
[0033] Preferably, in step S6, the multi-task loss function The calculation method is as follows:
[0034] ,
[0035] in:
[0036] The mean square error loss for profile prediction is calculated as follows: ;
[0037] The residual loss for two-dimensional radiative transmission is calculated as follows: ;
[0038] The horizontal gradient prediction loss is calculated as follows: ;
[0039] For vertical smoothness regularization loss of the profile;
[0040] , , These are the weighting coefficients.
[0041] Preferably, in step S7, the diagnostic information includes: the horizontal gradient field estimated in step S4. Two-dimensional correction contribution index Two-dimensional radiative transfer residuals .
[0042] Preferably, during the training phase of the neural network model, the training data used is constructed through the following steps:
[0043] Three-dimensional atmospheric field samples were extracted from the reanalysis data. These samples included horizontally non-uniform scenes at the edge of the polar vortex, after a volcanic eruption, near the terminator, and in the folded region of the troposphere.
[0044] A high-precision radiative transfer model was used to perform a two-dimensional radiative transfer forward modeling on the three-dimensional atmospheric field sample to simulate radiation under the near-edge observation geometry.
[0045] Construct training sample pairs, where the input is the simulated limb radiation sequence and observed geometric parameters, and the output label is the true atmospheric profile and horizontal gradient field of the target point.
[0046] Preferably, the differentiable simplified two-dimensional radiative transfer module employs the following computational acceleration strategy:
[0047] (1) Pre-calculate the lookup table for the extinction coefficient and source function;
[0048] (2) Implement the module using a deep learning framework that supports automatic differentiation;
[0049] (3) Process multiple observation points in batches and use GPUs for parallel computing.
[0050] Meanwhile, this invention proposes a near-edge atmospheric profile inversion system based on two-dimensional radiative transfer constraints, comprising:
[0051] The data preprocessing module is configured to perform the following process: acquire and preprocess the lateral radiation data of the target observation point and its adjacent observation sequences along the track;
[0052] The horizontal gradient feature extraction module is configured to perform the following process: extracting horizontal gradient feature vectors from the adjacent observation sequences using a temporal convolutional network;
[0053] The profile prediction module is configured to perform the following process: predict an initial profile based on target point observations, and fuse the horizontal gradient feature vector to predict a two-dimensional correction amount and a horizontal gradient field, thereby obtaining a corrected profile;
[0054] A differentiable simplified two-dimensional radiative transfer module is configured to perform the following process: calculate simulated radiative values based on the corrected profile and horizontal gradient field;
[0055] The post-processing output module is configured to perform the following process: output the final atmospheric profile inversion results and related diagnostic information.
[0056] The present invention adopts the above technical solution and has the following technical effects compared with the prior art:
[0057] (1) Horizontal non-uniformity modeling: This invention extracts the horizontal gradient information of adjacent observation sequences through a temporal convolutional network, effectively modeling the horizontal non-uniformity on the adjacent line-of-sight path, and improving the inversion accuracy by more than 30% in high gradient regions such as the edge of the polar vortex.
[0058] (2) High computational efficiency: Compared with the complete two-dimensional radiative transmission method, the present invention adopts a differentiable simplified two-dimensional radiative transmission module, which improves the computational speed by more than 50 times and meets the needs of business operation.
[0059] (3) Strong physical consistency: This invention uses the two-dimensional radiative transfer equation as a soft constraint to embed the loss function, ensuring that the inversion results satisfy the physical laws of atmospheric radiative transfer;
[0060] (4) End-to-end learning: This invention adopts a multi-task joint training strategy to avoid error accumulation caused by step-by-step optimization;
[0061] (5) Strong diagnostic capability: This invention outputs a horizontal gradient field and a two-dimensional correction contribution, which enhances the interpretability and quality assessment capability of the inversion results. Attached Figure Description
[0062] Figure 1 The flowchart illustrates the deep learning-based atmospheric profile inversion method based on two-dimensional radiative transfer constraints involved in this embodiment.
[0063] Figure 2 This is a schematic diagram of the overall neural network architecture involved in the embodiment.
[0064] Figure 3 This is a schematic diagram of the differentiable two-dimensional radiative transmission module involved in the embodiment.
[0065] Figure 4 This is a geometric schematic diagram of the adjacent observation sequence along the track involved in the embodiment.
[0066] Figure 5 This is an example of a horizontal gradient scenario involved in the implementation plan, along with a comparison of the inversion results.
[0067] Figure 6 This is a schematic diagram of multi-head attention feature fusion involved in the embodiment.
[0068] Figure 7 This is a schematic diagram of the composite loss function structure involved in the embodiment.
[0069] Figure 8 This is an example of a horizontal gradient scenario involved in the implementation plan, along with a comparison of the inversion results. Detailed Implementation
[0070] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.
[0071] Example 1: This example proposes a deep learning-based method for inverting atmospheric profiles at the edge based on two-dimensional radiative transfer constraints. Figure 1 This paper demonstrates the complete workflow of a deep learning inversion method based on two-dimensional radiative transfer constraints, including an end-to-end training architecture for input data, neural network processing, physical constraints, and loss functions. The method includes the following steps:
[0072] S1. Obtain the marginal radiation data of the target observation point. and its adjacent observation sequences along the track , where K is the number of adjacent points; Figure 4 It illustrates the geometric configuration of the near-edge observation (satellite, line of sight, tangent point) and the spatial distribution relationship of the sequence of adjacent observation points along the track.
[0073] S2. Extract horizontal gradient feature vectors from adjacent observation sequences using a temporal convolutional neural network. The temporal convolutional neural network employs a causal dilated convolution structure, including:
[0074] The convolutional blocks are multi-layered with causal dilation, and the dilation coefficient of each layer increases in power of 2, i.e., the dilation coefficient sequence is {1, 2, 4, 8, ...}.
[0075] Each convolutional block contains, in sequence, a one-dimensional causal convolutional layer, a batch normalization layer, a ReLU activation function, and a residual connection;
[0076] The receptive field of the temporal convolutional neural network covers K observation points before and after the orbital direction (2K+1 in total), and the value of K ranges from 3 to 7.
[0077] Figure 3 The paper details the dilated causal convolution structure of the temporal convolutional network TCN, illustrating how different dilation rates can be used to expand the receptive field to capture the spatial correlation of observation sequences along the track.
[0078] S3. Process target point observations using the first neural network. Predicting the initial profile under the assumption of one-dimensional spherical symmetry ;
[0079] S4, Fusion of horizontal gradient feature vectors Compared with the initial profile The second neural network is used to predict the two-dimensional correction amount. and horizontal gradient field The corrected profile was obtained The fusion process employs an attention mechanism to perform feature fusion, which weights and combines the horizontal gradient feature vector with the initial profile features. Figure 2 The overall architecture of the neural network described in this embodiment is explained, from the data flow of input feature embedding through TCN temporal convolution, multi-head attention fusion to MLP decoding output atmospheric profile. Figure 6 The computational flow of the multi-head attention mechanism is described, including Q / K / V projection, multi-head parallel computation, feature stitching, and residual connection.
[0080] S5. Using a differentiable simplified two-dimensional radiative transfer module, based on the corrected profile... and horizontal gradient field Calculation of simulated radiation The differentiable simplified two-dimensional radiative transfer module adopts the linear horizontal gradient assumption and incorporates atmospheric parameters. Represented as:
[0081]
[0082] in This is the vertical profile at the target point. The horizontal gradient along the track direction. The coordinates are along the track.
[0083] The differentiable simplified two-dimensional radiative transmission module discretizes the line-of-sight path as follows: The simulated radiation is calculated using the following steps at each sampling point:
[0084] S5.1: Calculate each sampling point Atmospheric parameters at the location ;
[0085] S5.2: Calculate cumulative optical thickness ;
[0086] S5.3: Calculation of radiation using discrete integrals ;
[0087] in For the source function, Extinction coefficient, The sampling interval is defined as follows. Furthermore, the differentiable simplified two-dimensional radiative transfer module employs the following computational acceleration strategy:
[0088] A lookup table for pre-calculated extinction coefficients and source functions;
[0089] Gradient propagation is implemented using a deep learning framework that supports automatic differentiation;
[0090] Batch processing of multiple observation points utilizes GPUs for parallel computation.
[0091] Figure 5 The working principle of the differentiable two-dimensional radiative transfer module is explained, and it is demonstrated how to extend the inversion profile into a two-dimensional atmospheric field and generate simulated radiation through path integral.
[0092] S6. The neural network is trained end-to-end using a multi-task loss function that includes two-dimensional radiative transfer consistency constraints; the multi-task loss function include:
[0093]
[0094] in:
[0095] Loss due to mean square error in profile prediction: ;
[0096] For the two-dimensional radiative transfer residual loss: ;
[0097] For horizontal gradient prediction loss: ;
[0098] For vertical smoothness regularization loss of the profile;
[0099] , , These are the weighting coefficients. Figure 7The structure of the composite loss function is shown, which includes four components: MSE supervised loss, physical constraints of the method of this invention, gradient loss, and smoothing regularization, along with their weight configurations.
[0100] S7, Output the two-dimensional corrected atmospheric profile and related diagnostic information. The diagnostic information includes: the estimated horizontal gradient field. Two-dimensional correction contribution index Two-dimensional radiative transfer residuals .
[0101] The steps for constructing training data for the neural network model described in this method are as follows:
[0102] Three-dimensional atmospheric field samples were extracted from the reanalysis data. These samples included horizontally non-uniform scenes at the edge of the polar vortex, after a volcanic eruption, near the terminator, and in the folded region of the troposphere.
[0103] A high-precision radiative transfer model was used to perform a two-dimensional radiative transfer forward modeling on the three-dimensional atmospheric field sample to simulate radiation under the near-edge observation geometry.
[0104] Construct training sample pairs, with the input being the simulated radiation sequence and observed geometric parameters, and the labels being the true profile of the target point and the horizontal gradient.
[0105] Finally, the inversion results of the traditional 1D method and the method described in this embodiment are compared in a horizontal gradient scene. The significant improvement of the method is demonstrated through profile curves and error histograms, such as... Figure 8 As shown, the improved neural network algorithm was used to invert atmospheric temperature and humidity profiles from microwave and infrared hyperspectral observations. Verification results show that the algorithm significantly improves both inversion error and vertical resolution. The species profile inversion error and vertical resolution obtained by the neural network technique are comparable to those of the optimal estimation method. By including different learning terms during training, it exhibits strong robustness to uncertainties in prior information and is faster than the optimal estimation method when nonlinear inversion is required.
[0106] Example 2: This example proposes a limb atmospheric profile inversion system based on two-dimensional radiative transfer constraints, including:
[0107] The data preprocessing module is configured to perform the following process: acquire and preprocess the lateral radiation data of the target observation point and its adjacent observation sequences along the track;
[0108] The horizontal gradient feature extraction module is configured to perform the following process: extracting horizontal gradient feature vectors from the adjacent observation sequences using a temporal convolutional network;
[0109] The profile prediction module is configured to perform the following process: predict an initial profile based on target point observations, and fuse the horizontal gradient feature vector to predict a two-dimensional correction amount and a horizontal gradient field, thereby obtaining a corrected profile;
[0110] A differentiable simplified two-dimensional radiative transfer module is configured to perform the following process: calculate simulated radiative values based on the corrected profile and horizontal gradient field;
[0111] The post-processing output module is configured to perform the following process: output the final atmospheric profile inversion results and related diagnostic information.
[0112] It should be noted that the processing flow of Embodiment 2 corresponds to the specific steps of the method provided in Embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in Embodiment 1 of the present invention.
[0113] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0114] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.
Claims
1. A method for limb atmospheric profile retrieval based on two-dimensional radiative transfer constraints, characterized in that, Includes the following steps: S1, acquiring the edge radiation data of the target observation point and the observation sequence along the adjacent track wherein K is the number of adjacent points; S2, extracting a horizontal gradient feature vector from the adjacent observation sequence using a time series convolutional neural network ; S3, processing the target point observations with the first neural network , predicting an initial profile under the one-dimensional spherically symmetric assumption ; S4, fusing level gradient feature vectors with the initial profile , predicting two-dimensional correction amounts using the second neural network and the horizontal gradient field , obtaining the corrected profile ; S5, using a differentiable simplified 2D radiance transport module, based on the corrected profile and horizontal gradient fields computational simulation of radiation ; S6. The neural network is trained end-to-end using a multi-task loss function that includes two-dimensional radiative transfer consistency constraints. S7, output two-dimensional corrected atmospheric profile and related diagnostic information.
2. The method of claim 1, wherein, In step S2, the temporal convolutional neural network adopts a causal dilated convolutional structure, including: The convolutional blocks are multi-layered with causal dilation, and the dilation coefficient of each layer increases in power of 2, i.e., the dilation coefficient sequence is {1,2,4,8,...}. Each convolutional block contains, in sequence, a one-dimensional causal convolutional layer, a batch normalization layer, a ReLU activation function, and a residual connection; The receptive field of the temporal convolutional neural network covers K observation points before and after the orbital direction, for a total of 2K+1 points, where K ranges from 3 to 7.
3. The method according to claim 1, characterized in that, In step S4, the horizontal gradient feature vectors are fused. Compared with the initial profile The fusion uses an attention mechanism to perform feature fusion, which combines the horizontal gradient feature vector with the initial profile feature in a weighted manner.
4. The method according to claim 1, characterized in that, In step S5, the differentiable simplified two-dimensional radiative transfer module adopts the linear horizontal gradient assumption and incorporates atmospheric parameters. Represented as: ； in, This is the vertical profile at the target point. The horizontal gradient along the track direction. The coordinates are along the track.
5. The method according to claim 1, characterized in that, In step S5, the differentiable simplified two-dimensional radiative transmission module discretizes the line-of-sight path into... The simulated radiation is calculated using the following steps at each sampling point: Calculate each sampling point Atmospheric parameters at the location: ; Calculate the cumulative optical thickness: ; Radiation is calculated using discrete integration: ; in For the source function, Extinction coefficient, The sampling interval is denoted as .
6. The method according to claim 1, characterized in that, In step S6, the multi-task loss function The calculation method is as follows: ， in: The mean square error loss for profile prediction is calculated as follows: ; The residual loss for two-dimensional radiative transmission is calculated as follows: ; The horizontal gradient prediction loss is calculated as follows: ; For vertical smoothness regularization loss of the profile; , , These are the weighting coefficients.
7. The method according to claim 1, characterized in that, In step S7, the diagnostic information includes: the horizontal gradient field estimated in step S4. Two-dimensional correction contribution index Two-dimensional radiative transfer residuals .
8. The method according to claim 1, characterized in that, During the training phase of the neural network model, the training data used is constructed through the following steps: Three-dimensional atmospheric field samples were extracted from the reanalysis data. These samples included horizontally non-uniform scenes at the edge of the polar vortex, after a volcanic eruption, near the terminator, and in the folded region of the troposphere. A high-precision radiative transfer model was used to perform a two-dimensional radiative transfer forward modeling on the three-dimensional atmospheric field sample to simulate radiation under the near-edge observation geometry. Construct training sample pairs, where the input is the simulated limb radiation sequence and observed geometric parameters, and the output label is the true atmospheric profile and horizontal gradient field of the target point.
9. The method according to claim 5, characterized in that, The differentiable simplified two-dimensional radiative transfer module employs the following computational acceleration strategy: (1) Pre-calculate the lookup table for the extinction coefficient and source function; (2) Implement the module using a deep learning framework that supports automatic differentiation; (3) Process multiple observation points in batches and use GPUs for parallel computing.
10. A system for inverting limb atmospheric profiles based on two-dimensional radiative transfer constraints, characterized in that, include: The data preprocessing module is configured to perform the following process: acquire and preprocess the lateral radiation data of the target observation point and its adjacent observation sequences along the track; The horizontal gradient feature extraction module is configured to perform the following process: extracting horizontal gradient feature vectors from the adjacent observation sequences using a temporal convolutional network; The profile prediction module is configured to perform the following process: predict an initial profile based on target point observations, and fuse the horizontal gradient feature vector to predict a two-dimensional correction amount and a horizontal gradient field, thereby obtaining a corrected profile; A differentiable simplified two-dimensional radiative transfer module is configured to perform the following process: calculate simulated radiative values based on the corrected profile and horizontal gradient field; The post-processing output module is configured to perform the following process: output the final atmospheric profile inversion results and related diagnostic information.