Atmospheric ozone vertical profile retrieval method based on differentiable radiative transfer physical constraint

Abstract

The application discloses a method for atmospheric ozone vertical profile inversion based on a differentiable radiative transfer physical constraint, and comprises the following steps: collecting atmospheric ozone data and auxiliary data to construct a data set; constructing a logarithmic residual inversion network model based on a logarithmic space residual learning mechanism; embedding a differentiable forward simulation operator at the end of the logarithmic residual inversion network model; introducing a physical constraint loss function; training the model through a preheating-fine-tuning two-stage training method based on the data set, and realizing the inversion of the atmospheric ozone vertical profile based on the trained model. The application effectively solves the problem of neural network gradient disappearance or weight imbalance caused by the large concentration span (from ppb level to ppm level) of ozone in the vertical direction.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing atmospheric detection technology, and particularly relates to a method for inverting the vertical profile of atmospheric ozone based on the physical constraints of microradiative transfer. Background Technology

[0002] Among existing meteorological inversion techniques, the Physical Iteration Method (OEM) is the traditional best estimation method, but its drawbacks are: extremely time-consuming computation, making it difficult to handle massive amounts of satellite data; and ozone inversion not only relies on the infrared band but also has a strong dependence on prior information. If the prior profile is inaccurate, it can easily lead to large errors in the inversion results in the troposphere. Pure data-driven deep learning directly establishes a mapping from brightness temperature to ozone concentration, but its drawbacks are: the vertical distribution of ozone in the atmosphere is extremely uneven (mostly concentrated in the stratosphere, with very little near the ground), and pure neural networks are prone to learning an "average state," making it difficult to capture the fine structure of the stratospheric-tropospheric exchange (STE) process. Furthermore, the integrated output profile often does not conserve the total column ozone observed by physical means.

[0003] Therefore, there is an urgent need in this field for a method to invert the vertical profile of atmospheric ozone, which can solve the technical problems of existing deep learning methods lacking physical mechanism constraints when inverting ozone profiles, resulting in stratospheric peak position shifts, inaccurate tropospheric ozone capture, and non-closed radiative transfer. Summary of the Invention

[0004] Purpose of the Invention: The purpose of this invention is to provide a method for inverting the vertical profile of atmospheric ozone based on the physical constraints of minute radiative transfer. This effectively solves the problem of gradient vanishing or weight imbalance in neural networks caused by the large concentration range of ozone in the vertical direction (from ppb to ppm).

[0005] Technical solution: The present invention provides a method for inverting the vertical profile of atmospheric ozone based on the physical constraint of minute radiative transfer, comprising the following steps:

[0006] Step 1: Collect atmospheric ozone data and auxiliary data to construct a dataset;

[0007] Step 2: Construct a logarithmic residual inversion network model based on the logarithmic space residual learning mechanism;

[0008] Step 3: Embed a differentiable positive simulation operator at the end of the numerical residual inversion network model;

[0009] Step 4: Introduce the physical constraint loss function;

[0010] Step 5: Based on the dataset, train the model using a two-stage training method of preheating and fine-tuning, and then use the trained model to invert the vertical profile of atmospheric ozone.

[0011] Further, step 1 specifically involves: selecting a hyperspectral infrared detector mounted on a polar-orbiting meteorological satellite for data acquisition, and using Jacobi matrix analysis within the wavenumber range of 980 cm⁻¹. -1 Up to 1080 cm -1 Channels highly sensitive to ozone changes were selected as input features; auxiliary data included satellite zenith angle (SatZA), surface temperature (ST), and surface pressure (SP).

[0012] Training tag: Using ECMWF ERA5 reanalysis data, extract the corresponding spatiotemporal ozone mass mixing ratio vertical profile and interpolate it to a fixed 101-layer pressure layer;

[0013] Verification label: Data from global ozone radiosondes is used to verify the precision of the vertical structure;

[0014] Prior background field construction: Statistical analysis of the global ozone average state over the past 5-10 years to construct a climatological average profile based on latitude and month.

[0015] Furthermore, step 2 specifically involves: to address the weight imbalance caused by the large concentration range of ozone in the troposphere (ppb level) and stratosphere (ppm level), a logarithmic space residual learning mechanism is designed:

[0016] The network input is a standardized brightness temperature vector. ; where y obs For input data, μ y δ is the mean brightness temperature of all samples. y represents the standard deviation of the brightness temperature for all samples.

[0017] The backbone network adopts a fully connected residual network, which contains 4 residual blocks, each containing a fully connected layer - BatchNorm - Swish activation function;

[0018] Output definition: The network does not directly output ozone concentration x, but instead outputs the logarithmic concentration correction Δξ relative to the prior background field;

[0019]

[0020] Where f NN For a neural network, θ represents the network parameters;

[0021] Profile reconstruction equation: x ap For the prior background field;

[0022] By predicting ln(x / x) apThis ensures that the output ozone concentration is always positive, and the network focuses on learning outliers that deviate from the average state of the current atmosphere.

[0023] Furthermore, step 3 specifically includes the following steps:

[0024] Step 3.1, Radiative Transfer Equation:

[0025] Under clear sky conditions, the radiation intensity R(v) observed by the satellite can be approximately expressed as:

[0026]

[0027] Where: v is the wavenumber; ε is the surface emissivity; B is the Planck function; τ is the atmospheric transmittance, a function of ozone concentration; and T... s P represents the surface temperature. s Let τ be the surface pressure, T be the vertical profile of atmospheric temperature, T(P) represent the relationship between atmospheric temperature T and atmospheric pressure altitude P, and τ be the atmospheric transmittance. It is the partial derivative of transmittance with respect to air pressure;

[0028] Step 3.2, Differentiable layer implementation:

[0029] Construct a PyTorch differentiable layer using the tangent linear mode K-matrix model of RTTOV:

[0030]

[0031] in, Represents the simulated brightness temperature vector, RTM diff It is a differentiable radiative transport operator. The predicted ozone profile is given, and T represents the vertical atmospheric temperature profile. This is a water vapor profile.

[0032] Furthermore, step 4 specifically involves:

[0033] The training process uses a multi-task loss function, as shown in the following formula:

[0034]

[0035] Where, λ 1， λ 2，λ3 is a preset hyperparameter weighting coefficient used to balance the contribution of each loss term in the joint inversion. Since the brightness temperature error (unit: K) and concentration error (unit: ppmv) are of completely different magnitudes, λ1 plays a role in dimensional balance to ensure that radiation error can be perceived by the network; λ2 is the total amount conservation weight used to constrain the total ozone column content after vertical integration and reduce the ill-posedness of the inversion; λ3 adjusts the smoothness of the profile. If the inversion results are found to oscillate violently, λ3 is increased; if the results are found to be too rigid and cannot capture tropospheric abrupt changes, λ3 is decreased. For profile accuracy error:

[0036]

[0037] Using the log-mean squared error, where w i For the height layer weights, x true，i The ozone reference value for the i-th layer (e.g., radiosonde data) is given, and M represents the total number of vertical atmospheric stratifications (101 layers in this implementation). The ozone concentration value predicted by the neural network at the i-th layer is shown.

[0038] This represents the radiation closure error; the difference between simulated brightness temperature and satellite-observed brightness temperature, where N represents the total number of selected satellite infrared hyperspectral feature channels. Indicates the simulated brightness temperature value;

[0039]

[0040] To constrain the total column volume, a vertical integral is performed on the predicted profile.

[0041]

[0042] Among them, TCO ref For reference total ozone column concentration, P s ρ represents the surface pressure, p represents the atmospheric pressure, and g represents the acceleration due to gravity (usually taken as 9.8 m / s²). 2 ), used to convert the concentration integral in barometric coordinates into mass per unit area or Dobson units (DU). This represents the predicted ozone concentration as a function of atmospheric pressure p.

[0043] Smoothing constraint loss L smooth :

[0044]

[0045] By using second-order difference constraints to ensure the smoothness of the vertical profile, non-physical sawtooth oscillations can be eliminated.

[0046] Furthermore, step 5 specifically involves: employing a two-stage training method of preheating and fine-tuning; stage one, preheating: setting λ1 = 0, and using only the profile accuracy error. The network is trained to quickly approximate the statistical regularity of ERA5; Phase 2: Physical fine-tuning: λ1 and λ2 are enabled, and radiative transfer constraints are introduced; at this time, the network uses physical equations to fine-tune the weights, so that the inversion results are not only statistically accurate, but also have closed radiative characteristics; inference application: real-time received satellite L1 level data is input into the trained network, and after one forward propagation, a high-precision ozone vertical profile can be output in milliseconds.

[0047] 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.

[0048] 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.

[0049] The present invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method of the present invention.

[0050] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0051] Dual physical constraint mechanism: Not only is the radiative transfer equation introduced as a radiative domain constraint, but the total ozone amount conservation is also introduced as an integral domain constraint, which solves the ill-posed problem of ozone vertical distribution inversion.

[0052] Log-residual learning strategy: A method for predicting ozone concentration corrections in logarithmic space is requested for protection. This method effectively solves the problem of vanishing gradients or weight imbalances in neural networks caused by the large vertical concentration range of ozone (from ppb to ppm).

[0053] Channel sensitivity weighting: At the input end, strong absorption channels in the 9.6µm band are dynamically selected based on the Jacobian matrix, and channels that are severely affected by water vapor interference are removed.

[0054] More accurate vertical structure: Compared with pure AI models, this method uses physical equations to enforce constraints, which can more accurately locate the height of the ozone maximum layer and reduce the vertical shift of the stratospheric peak.

[0055] Strong generalization ability with small samples: Due to the introduction of the physical laws of radiative transfer, the model no longer relies entirely on massive amounts of labeled data. High-precision tropospheric ozone inversion capability can be obtained by fine-tuning with a small amount of ozone radiosonde data.

[0056] The result is physically explainable: the output profile naturally satisfies the conservation of radiative energy, and is not only numerically accurate, but also physically consistent. Attached Figure Description

[0057] Figure 1 This is a flowchart of the present invention.

[0058] Figure 2 This is a flowchart illustrating the specific process of the present invention. Detailed Implementation

[0059] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0060] like Figure 1 , 2 As shown, the present invention provides a method for inverting the vertical profile of atmospheric ozone based on the physical constraint of differential radiative transfer, comprising the following steps:

[0061] Step 1: Collect atmospheric ozone data and auxiliary data to construct a dataset;

[0062] Step 2: Construct a logarithmic residual inversion network model based on the logarithmic space residual learning mechanism;

[0063] Step 3: Embed a differentiable positive simulation operator at the end of the numerical residual inversion network model;

[0064] Step 4: Introduce the physical constraint loss function;

[0065] Step 5: Based on the dataset, train the model using a two-stage training method of preheating and fine-tuning, and then use the trained model to invert the vertical profile of atmospheric ozone.

[0066] Example:

[0067] (1) Data

[0068] Dataset Construction: Global FY3D / HIRAS Level 1 brightness temperature (L1C) data were collected from 2022 to 2023. The ozone-sensitive 9.6μm band was selected, and cloud pollution points were removed by cloud detection to obtain clear sky brightness temperature data.

[0069] The corresponding spatiotemporal ERA5 ozone profile data was matched and unified to layer 101 (1100 hPa to 0.005 hPa) through interpolation. The corresponding spatiotemporal WOUDC global ozone radiosonde data was matched as verification data.

[0070] (2) Model

[0071] Network structure: Input dimension is 314 (number of channels) + 4 (auxiliary factor). The number of neurons in each of the 5 residual blocks is [1024, 512, 512, 256, 101]. The Swish activation function is used.

[0072] Phase 1 (Warm-up): Physical constraints are set to λ1=0, and the learning rate is 10. -3 The batch size is 512. Iterate for 100 epochs on a GPU (NVIDIA A100) processor. At this point, L... profile The model rapidly declined and learned the global statistical mean distribution of ERA5.

[0073] Phase Two (Fine-tuning): Settings: Enable the physical layer, set λ1=0.5, λ2=0.2. Reduce the learning rate to 10. -4 During the training loop, each batch of ERA5 data is mixed with 10 sets of WOUDC sounding data to force the network to fit a high-precision vertical gradient.

[0074] (3) Verification of experimental results

[0075] Verification and Application: The ozone profile derived from the inversion was compared with that of independent radiosonde stations, with a focus on calculating the relative errors in the troposphere (0-10km) and the lower stratosphere (10-25km).

[0076]

[0077] The above embodiments are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and equivalent substitutions without departing from the principle of the present invention. All such improvements and equivalent substitutions to the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A method for inverting the vertical profile of atmospheric ozone based on the physical constraint of differential radiative transfer, characterized in that, Includes the following steps: Step 1: Collect atmospheric ozone data and auxiliary data to construct a dataset; Step 2: Construct a logarithmic residual inversion network model based on the logarithmic space residual learning mechanism; Step 3: Embed a differentiable positive simulation operator at the end of the numerical residual inversion network model; Step 4: Introduce the physical constraint loss function; Step 5: Based on the dataset, train the model using a two-stage training method of preheating and fine-tuning, and then use the trained model to invert the vertical profile of atmospheric ozone.

2. The method for inverting the vertical profile of atmospheric ozone based on the physical constraint of negligible radiative transfer as described in claim 1, characterized in that, Step 1 specifically involves: selecting a hyperspectral infrared detector mounted on a polar-orbiting meteorological satellite for data acquisition, and using Jacobi matrix analysis within the wavenumber range of 980 cm⁻¹. -1 Up to 1080 cm -1 Channels highly sensitive to ozone changes were selected as input features; Supporting data include satellite zenith angle (SatZA), surface temperature (ST), and surface air pressure (P). s ; Training tags: Using ECMWF ERA5 reanalysis data, extract the corresponding spatiotemporal ozone mass mixing ratio vertical profile, interpolate to a fixed 101-layer pressure layer, and adopt a multi-source data gradient collaborative training strategy; Since ERA5 data has the characteristic of global coverage, it is first used to build a large-scale sample library to complete the baseline parameter training of the model, so that the model has the ability to generalize and invert the vertical structure of atmospheric ozone on a global scale. Verification label: Data from global ozone radiosondes is used to verify the precision of the vertical structure; Prior background field construction: Statistical analysis of the global ozone average state over the past 5-10 years to construct a climatological average profile based on latitude and month.

3. The method for inverting the vertical profile of atmospheric ozone based on the physical constraint of negligible radiative transfer as described in claim 1, characterized in that, Step 2 specifically involves: To address the weight imbalance caused by the large concentration range of ozone in the troposphere (ppb level) and stratosphere (ppm level), a logarithmic space residual learning mechanism is designed: The network input is a standardized brightness temperature vector. ; where y obs For input data, μ y δ is the mean brightness temperature of all samples. y The standard deviation of the brightness temperature for all samples; The backbone network adopts a fully connected residual network, which contains 4 residual blocks, each containing a fully connected layer - BatchNorm - Swish activation function; Output definition: The network does not directly output ozone concentration x, but instead outputs the logarithmic concentration correction Δξ relative to the prior background field; ； Where f NN For a neural network, θ represents the network parameters; Profile reconstruction equation: x ap For the prior background field; By predicting ln(x / x) ap This ensures that the output ozone concentration is always positive, and the network focuses on learning outliers that deviate from the average state of the current atmosphere.

4. The method for inverting the vertical profile of atmospheric ozone based on the physical constraint of differential radiative transfer as described in claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1, Radiative Transfer Equation: Under clear sky conditions, the radiation intensity R(v) observed by the satellite can be approximately expressed as: ； Where: v is the wavenumber; ε is the surface emissivity; B is the Planck function; T s P represents the surface temperature. s Let τ be the surface pressure, T be the vertical profile of atmospheric temperature, T(P) represent the relationship between atmospheric temperature T and atmospheric pressure altitude P, and τ be the atmospheric transmittance. It is the partial derivative of transmittance with respect to air pressure; Step 3.2, Differentiable layer implementation: Construct a PyTorch differentiable layer using the tangent linear mode K-matrix model of RTTOV: ； in, Represents the simulated brightness temperature vector, RTM diff It is a differentiable radiative transport operator. The predicted ozone profile is given, and T represents the vertical atmospheric temperature profile. This is a water vapor profile.

5. The method for inverting the vertical profile of atmospheric ozone based on the physical constraint of negligible radiative transfer according to claim 3, characterized in that, Step 4 is as follows: The training process uses a multi-task loss function, as shown in the following formula: ； Where, λ 1， λ 2， λ3 is a preset hyperparameter weighting coefficient used to balance the contribution of each loss term in the joint inversion. Since the brightness temperature error and concentration error are of completely different magnitudes, λ1 plays a role in dimensional balance to ensure that radiation error can be perceived by the network; λ2 is the total amount conservation weight used to constrain the total ozone column content after vertical integration and reduce the ill-posedness of the inversion; λ3 adjusts the smoothness of the profile. If the inversion result is found to oscillate violently, λ3 is increased; if the result is found to fail to capture tropospheric abrupt changes, λ3 is decreased. For profile accuracy error: ； Using the log-mean squared error, where w i For the height layer weights, x true，i Let be the true reference value for ozone corresponding to the i-th layer, and represent the total number of vertical atmospheric layers. The ozone concentration value predicted by the neural network at the i-th layer is shown. To simulate the difference between the brightness temperature and the satellite observation brightness temperature for radiation closure error, N represents the total number of selected satellite infrared hyperspectral feature channels. Indicates the simulated brightness temperature value; ； To constrain the total column volume, perform vertical integration on the predicted profile; ； Among them, TCO ref For reference total ozone column, P s ρ represents surface pressure, p represents atmospheric pressure, and g represents gravitational acceleration. This is used to convert the concentration integral in a pressure coordinate system to mass per unit area or Dobson's unit DU. This represents the predicted ozone concentration as a function of atmospheric pressure p. Smoothing constraint loss L smooth : ； By using second-order difference constraints to ensure the smoothness of the vertical profile, non-physical sawtooth oscillations can be eliminated.

6. The method for inverting the vertical profile of atmospheric ozone based on the physical constraint of differential radiative transfer according to claim 5, characterized in that, Step 5 specifically involves: employing a two-stage training method of preheating and fine-tuning; Stage 1, preheating: setting λ1 = 0, using only the profile accuracy error. The network is trained to quickly approximate the statistical regularity of ERA5; Phase 2: Physical fine-tuning: λ1 and λ2 are enabled, and radiative transfer constraints are introduced; at this time, the network uses physical equations to fine-tune the weights, so that the inversion results are not only statistically accurate, but also have closed radiative characteristics; inference application: real-time received satellite L1 level data is input into the trained network, and after one forward propagation, a high-precision ozone vertical profile can be output in milliseconds.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

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

9. A computer program product comprising a computer program / 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