A Bayesian Deep Learning-Based Method and System for Inverting Limb Atmospheric Profiles

Abstract

This invention discloses a method and system for retrieving limb atmospheric profiles based on Bayesian deep learning, belonging to the field of atmospheric remote sensing inversion technology. The invention includes: constructing a Bayesian neural network comprising a shared feature extraction backbone and multiple task branches; the backbone network models the weight distribution based on variational Bayesian inference; the average kernel prediction branch outputs a low-rank representation of the average kernel matrix; designing a multi-task loss function that integrates negative log-likelihood, KL divergence, average kernel consistency, and physical constraints; training multiple models using deep ensemble learning; quantifying randomness and cognitive uncertainty through Monte Carlo sampling; and finally outputting a complete inversion product containing the profile, hierarchical uncertainty, and average kernel matrix. This invention achieves, for the first time, end-to-end joint learning of average kernel information and deep learning inversion, improving inference speed by approximately 100 times while maintaining the same accuracy as traditional methods, meeting the needs of near real-time operational applications.

Description

Technical Field

[0001] This invention relates to the field of atmospheric remote sensing inversion technology, specifically to a method and system for inverting limb atmospheric profiles based on Bayesian deep learning. Background Technology

[0002] Traditional atmospheric profile inversion primarily employs the Optimal Estimation (OE) method. While this method provides uncertainty information such as the average kernel matrix and the posterior error covariance matrix, it suffers from the following problems:

[0003] (1) Low computational efficiency: The optimal estimation method requires iteratively running the forward model to simulate the measured radiation values ​​corresponding to the inverted parameters, which makes the inversion process computationally expensive because the forward model involves radiative transfer simulation calculations. This characteristic makes it difficult to meet the requirements of near real-time applications. In addition, traditional inversion algorithms rely heavily on iterative calculations involving radiative transfer models, which are both computationally and time-consuming, limiting their application in real-time or near real-time scenarios.

[0004] (2) Lack of Uncertainty in Neural Network Methods: In recent years, neural network methods have made significant progress in atmospheric inversion. Neural networks offer a promising alternative because, once trained, they can generate inversion results in a fraction of the time required by traditional methods. However, existing neural network inversion methods have a key drawback: obtaining theoretically rigorous uncertainty quantification using artificial neural networks remains a challenge. Traditional neural networks only output point estimates and cannot provide uncertainty quantification information comparable to optimal estimation methods.

[0005] (3) Importance of Averaging Kernel Information: The AK (Averaging Kernel) is a key indicator characterizing the vertical resolution and sensitivity of an inversion system. The AK characterizes the sensitivity of the inversion to the actual atmospheric conditions and describes the vertical resolution of the measurement. Averaging kernel information is indispensable in data assimilation and product validation: it is essential for comparison with other data products and for data assimilation. Existing neural network methods generally lack effective fusion and output of AK information.

[0006] Existing research indicates that Bayesian neural networks can provide a fundamental method for uncertainty quantification, laying the theoretical foundation for probabilistic modeling in deep learning models. However, this technology has not yet been effectively combined with average kernel information, and current applications are mainly concentrated in nadir observation scenarios, with relatively little research on its application in edge observation. Monte Carlo Dropout (MCD), as a practical approximation method for Bayesian inference, has been widely used for uncertainty estimation in neural networks. However, this method suffers from high computational overhead, requiring multiple forward propagation samplings during the inference phase, and its output lacks physical interpretability, making it difficult to integrate with the traditional inversion theory framework. Existing research proposes using efficient perturbation methods to approximate the average kernel matrix of neural networks. However, these methods only treat the average kernel calculation as an independent post-processing step, failing to embed the average kernel prediction into the end-to-end training process of the neural network, thus failing to achieve joint optimization of the average kernel and the inversion model. Existing research has achieved near real-time aerosol inversion based on neural networks using OMPS-LP (Ozone Mapping and Profiler Suite - Edge Profiler), achieving approximately 60 times the computational speedup compared to traditional methods. However, this method only outputs point estimates of the profile and fails to provide complete information on uncertainty quantification, including key products such as posterior error covariance and mean kernel matrix. Summary of the Invention

[0007] The technical problem to be solved by this invention is to address the shortcomings of traditional edge atmospheric profile inversion methods, specifically including the low computational efficiency of optimal estimation methods; although neural network methods are fast, they lack theoretically rigorous uncertainty quantification capabilities and cannot output average kernel information compatible with traditional methods, thus limiting the application of inversion products in data assimilation and product verification.

[0008] To solve the above problems, the present invention adopts the following technical solution:

[0009] First, this invention proposes a method for inverting atmospheric profiles at proximal edges based on Bayesian deep learning, comprising the following steps:

[0010] S1. Obtain edge observation data and construct a multi-branch Bayesian neural network model. The model includes a shared feature extraction backbone network, a profile estimation branch, a heteroscedasticity uncertainty estimation branch, and an average kernel prediction branch. The shared feature extraction backbone network adopts a variational Bayesian inference framework, and models the network weight parameters as a probability distribution. The average kernel prediction branch outputs a low-rank decomposition representation of the average kernel matrix.

[0011] S2. Design a multi-task joint loss function, which includes negative log-likelihood loss, KL divergence regularization term, average kernel consistency loss and physical constraint regularization term;

[0012] S3. Multiple independent Bayesian neural network models are trained using a deep ensemble learning strategy, and the decomposition and quantification of random uncertainty and cognitive uncertainty are achieved through Monte Carlo sampling and ensemble statistics.

[0013] S4. Input the edge observation data into the trained Bayesian neural network model, perform forward inference, and output a complete inversion product containing atmospheric composition profile estimates, stratification uncertainties, and average kernel matrix.

[0014] Preferably, in step S1, the low-rank decomposition of the average kernel matrix is ​​expressed as:

[0015] ;

[0016] in, It is a low-rank factor. For the preset rank parameter and , The number of vertical layers. This is a diagonal correction term.

[0017] Preferably, in step S2, the average kernel consistency loss is defined as:

[0018] ;

[0019] in, The average kernel matrix of the network predictions. The reference average kernel matrix is ​​calculated using the optimal estimation method. It is the Frobenius norm. The trace of the matrix, These are the trace constraint weight coefficients.

[0020] Preferably, in step S3, the method for decomposing and quantifying the random uncertainty and cognitive uncertainty is as follows:

[0021] for Each integrated member and each member After Monte Carlo sampling, the total variance is decomposed into:

[0022] ;

[0023] Among them, random uncertainty , The random uncertainty (variance) of the output of the m-th ensemble member at the t-th sampling; cognitive uncertainty. , This is a single predicted value. It is the average predicted value.

[0024] Preferably, in step S2, the physical constraint regularization term includes:

[0025] Non-negativity constraint terms: ;

[0026] Vertical smoothing constraint term: ,in This represents the concentration value of the atmospheric component in the i-th layer predicted by the neural network;

[0027] Chapman layer shape prior constraint: For a target gas with a characteristic peak distribution, constrain the deviation of the peak position and shape parameters of the inverted profile from the climatological Chapman layer parameters.

[0028] Preferably, the edge observation data comes from satellite edge detection instruments, including the OMPS edge profiler, SCIAMACHY, MLS, or the OMS-L payload of the Fengyun series satellites in China; the satellite edge detection instruments measure multi-wavelength radiance spectra, observation geometric parameters, prior atmospheric profiles, and auxiliary meteorological data at different tangent heights; normalization and feature construction are performed on the input data to generate network input feature vectors.

[0029] The boundary condition at the top of the atmosphere is: the incident diffuse radiation is zero.

[0030] The surface boundary conditions are: calculate the surface reflected radiation based on the Lambert reflection model or the two-way reflection distribution function;

[0031] The boundary condition loss term is: ,

[0032] in, The number of boundary points. The number of boundary points (sum upper limit). The neural network predicted the value. These are the boundary condition constraint values.

[0033] Preferably, the Bayesian neural network model trained by the method can be applied to new target instruments through cross-instrument transfer learning; the transfer learning step includes:

[0034] The input feature distribution is adjusted using domain adaptive techniques, while keeping the structure of the average kernel prediction branch unchanged, and only the profile estimation branch and the uncertainty estimation branch are fine-tuned.

[0035] Meanwhile, this invention proposes a near-edge atmospheric profile inversion system based on Bayesian deep learning, comprising:

[0036] The data preprocessing module is configured to perform the following processes: receive radiance spectra and auxiliary data from limb observations, and perform standardization and feature construction;

[0037] The Bayesian neural network inference module is configured to perform the following processes: loading a pre-trained deep ensemble model and performing multiple Monte Carlo forward propagations;

[0038] The uncertainty quantification module is configured to perform the following processes: statistically integrating forecast results, calculating random uncertainty and cognitive uncertainty;

[0039] The average kernel reconstruction module is configured to perform the following process: reconstruct the complete average kernel matrix from the low-rank factors output by the network;

[0040] The product generation module is configured to perform the following process: output inversion results, uncertainties, and average kernel products that conform to the standard format of satellite data products.

[0041] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described in the present invention.

[0042] Finally, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the computer program is executed, it implements the steps of the method described in the present invention.

[0043] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0044] (1) This invention replaces the traditional iterative physical inversion with Bayesian deep learning, which greatly improves the inference speed and meets the near real-time business requirements; at the same time, the inversion accuracy is comparable to that of the traditional method through multi-task loss and physical constraints.

[0045] (2) The present invention adds an average kernel prediction branch to the neural network architecture and designs an average kernel consistency loss function, so that the model can directly learn and output an average kernel matrix that is compatible with traditional physical inversion methods, so that deep learning inversion products can be directly used for data assimilation and multi-source product verification.

[0046] (3) This invention embeds physical regularization terms such as non-negativity constraints, vertical smoothness constraints, Chapman layer priors and boundary condition constraints into the multi-task loss function, deeply integrating prior knowledge into the model learning process, ensuring the rationality of the inversion profile in terms of shape and magnitude, and improving the physical consistency and business acceptance of the product. Attached Figure Description

[0047] Figure 1 This invention relates to a flowchart of a method for inverting atmospheric profiles at the edge based on Bayesian deep learning.

[0048] Figure 2This is a schematic diagram of the Bayesian deep learning network structure involved in the embodiment.

[0049] Figure 3 This is a schematic diagram illustrating the uncertainty quantification and average kernel output involved in the embodiment. Detailed Implementation

[0050] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Example 1: This example is a specific implementation of a method for retrieving atmospheric profiles at proximal edges based on Bayesian deep learning. (Refer to...) Figure 1 The specific process is as follows:

[0052] Step S1: Acquire edge observation data. The edge observation data comes from satellite edge detection instruments, including the OMPS edge profiler, SCIAMACHY, MLS, or the OMS-L payload of the Fengyun series satellites. The satellite edge detection instruments measure multi-wavelength radiance spectra, observation geometric parameters, prior atmospheric profiles, and auxiliary meteorological data at different tangent heights. Normalize and construct features on the input data to generate network input feature vectors.

[0053] Constructing a Bayesian neural network architecture based on average kernel perception:

[0054] Construct a multi-branch Bayesian neural network architecture, referencing Figure 2 ,include:

[0055] S1.1 Input Feature Construction Module: The radiance spectrum of the near edge observation, the observation geometric parameters (including the tangent height sequence, solar zenith angle, relative azimuth angle, and scattering angle), and the auxiliary meteorological data (including temperature profile and air pressure profile) are standardized and then concatenated into an input feature vector.

[0056] S1.2 Shared Feature Extraction Backbone Network: A variational Bayesian inference framework is adopted, and the network weight parameters are modeled as probability distributions rather than deterministic values.

[0057] Specifically, for the network... Layer weight matrix Assume it follows a Gaussian distribution:

[0058] ,

[0059] in and For learnable parameters, Indicates the first The mean parameter of the weight distribution, For the first Standard deviation parameter of the layer distribution.

[0060] Backpropagation of gradients is achieved using reparameterization techniques:

[0061] ,

[0062] in, This represents element-wise multiplication. This represents a standard normally distributed random noise vector. Represents a multi-standard normal distribution. Represents a vector with zero mean. This represents the unit covariance matrix.

[0063] S1.3 Profile Estimation Branch: Outputs estimated concentrations of target atmospheric components (such as ozone and aerosol extinction coefficient) at various altitude levels. ,in This represents the number of vertical layers.

[0064] S1.4 Heteroscedasticity Uncertainty Estimation Branch: Outputs the log-standard deviation of each altitude layer. It is used to characterize aleatoric uncertainty caused by observation noise.

[0065] S1.5 Average Kernel Prediction Branch: Outputs a compact representation of the average kernel matrix. This takes into account the full average kernel matrix. Due to its high dimensionality, a low-rank decomposition is used to represent it.

[0066] ,

[0067] in It is a low-rank factor. It is a rank parameter. This is a diagonal correction term. (Network output) , and .

[0068] Step S2: Design the joint loss function for multiple tasks:

[0069] Design a composite loss function consisting of four components:

[0070] S2.1 Negative Log-Likelihood Loss: Assuming the inversion error follows a heteroscedastic Gaussian distribution, the negative log-likelihood loss is:

[0071] ,

[0072] in For the first The truth value of the layer, and These are the estimated value and standard deviation of the network output, respectively.

[0073] S2.2 KL divergence regularization term: Constrains the posterior distribution of the weights to approximate the prior distribution, preventing overfitting.

[0074] ,

[0075] in, Indicates the divergence sign. KL divergence loss term, KL divergence (relative entropy) measures the difference between two probability distributions, prior to the distribution. Set to a standard normal distribution or a Gaussian mixture distribution. It is a variational posterior distribution.

[0076] S2.3 Average Kernel Consistency Loss: Ensures that the average kernel matrix predicted by the network is consistent with the reference average kernel matrix calculated by traditional methods.

[0077] ,

[0078] in, The average kernel matrix of the network predictions. The reference average kernel matrix is ​​calculated using the optimal estimation method. It is the Frobenius norm. The trace of the matrix, These are the trace constraint weight coefficients.

[0079] S2.4 Physical constraint regularization term: embedding prior physical knowledge of atmospheric profile:

[0080] ,

[0081] in:

[0082] Non-negativity constraint;

[0083] For smoothing constraints;

[0084] Prior constraints on the shape of the Chapman layer (for gases with characteristic peak distributions, such as ozone).

[0085] S2.5 Total Loss Function:

[0086] ,

[0087] in , , These are adjustable weight hyperparameters.

[0088] Step S3: Uncertainty quantification with ensemble learning enhancement, refer to Figure 3 The specific steps include:

[0089] S3.1 Deep Ensemble Training: Training with different random initializations and data subsets Individual Bayesian neural network models are integrated into a deep ensemble.

[0090] S3.2 Two-level uncertainty decomposition: for input observations Integrated predictions were obtained through Monte Carlo sampling:

[0091] For the Each integrated member executes Each forward propagation (implemented by sampling different weights each time) yields the prediction set. .

[0092] The overall forecast mean and variance are decomposed as follows:

[0093] ,

[0094] ,

[0095] S3.3 Average Kernel Integration: The integration estimate of the average kernel matrix is:

[0096] ,

[0097] Simultaneously, the uncertainty of the average kernel is calculated to characterize the differences in vertical sensitivity estimation among different models.

[0098] Step S4: Generate the complete inversion product

[0099] S4.1 Profile Product: Outputs estimated atmospheric composition concentrations at each altitude level and their posterior standard deviations.

[0100] S4.2 Uncertainty Products: Outputs random uncertainty (from observation noise) and cognitive uncertainty (from model uncertainty) respectively, allowing users to choose to use them according to the application scenario.

[0101] S4.3 Average Kernel Product: Outputs a complete average kernel matrix to characterize the vertical resolution and sensitivity of the inversion, supporting interoperability and data assimilation applications with traditional inversion products.

[0102] S4.4 Quality Marker Products: Set up tiered quality marks based on uncertainty level and average cross-diagonal elements.

[0103] This embodiment uses observation data from the OMPS Limper Profiler (OMPS-LP) carried by the US Suomi-NPP satellite as an example to illustrate the specific implementation process of the method of the present invention.

[0104] 1.1 Data Acquisition and Preparation

[0105] (1) Acquisition of observation data

[0106] Download the OMPS-LP Level 1B data product, version V2.6, from the NASA GES DISC data center. The observation data includes:

[0107] Ultraviolet (UV) band: 290-380nm, spectral resolution 1nm;

[0108] Visible light band (VIS): 380-1000nm, spectral resolution 1nm;

[0109] Observation height range: 0-80km, vertical sampling interval approximately 1.6km;

[0110] Time range: April 2012 to December 2023.

[0111] In this embodiment, normalized radiation at a wavelength of 675 nm is selected as the main inversion input. This wavelength is sensitive to ozone absorption and relatively less affected by aerosols. Each profile contains normalized radiation values ​​at 50 tangent heights, i.e., input dimension Ny=50.

[0112] (2) Acquisition of prior data

[0113] Ozone prior profile data are obtained from the following sources:

[0114] MERRA-2 reanalysis data: provides ozone volume mixing ratios over an altitude range of 10-80 km;

[0115] McPeters / Labow Ozone Climate State: As Supplementary Prior Information;

[0116] Prior covariance matrix: constructed based on historical observation statistics, taking into account latitude and seasonal dependence;

[0117] The a priori profile is interpolated to a uniform vertical grid with 40 grid layers (Nz=40), a height range of 10-70km, and a vertical resolution of approximately 1.5km.

[0118] (3) Radiative transfer simulation

[0119] A training dataset was constructed using the SASKTRAN spherical atmospheric radiative transfer model for forward simulation.

[0120] Atmospheric profile sampling: 100,000 ozone profiles were randomly selected from the MERRA-2 data;

[0121] Observational geometry sampling: solar zenith angle 20°-85°, relative azimuth angle 0°-180°, tangent height 10-70km;

[0122] Perturbation strategy: Apply ±30% random perturbation to the true profile to enhance data diversity;

[0123] Noise model: Gaussian noise was added based on the signal-to-noise ratio characteristics of the OMPS-LP instrument;

[0124] In the end, 200,000 training samples, 30,000 validation samples, and 20,000 test samples were generated.

[0125] 1.2 Data Preprocessing and Feature Extraction

[0126] (1) Normalization of radiation data

[0127] The raw radiation data is normalized:

[0128]

[0129] in To measure radiation, For reference altitude (60km) radiation, and These are the mean and standard deviation of the training set statistics, respectively.

[0130] (2) Auxiliary feature construction

[0131] In addition to normalized radiation, auxiliary input features are constructed as shown in Table 1:

[0132] Table 1

[0133]

[0134] The total input feature dimension is Ny=145.

[0135] (3) Output target processing

[0136] The target value for the ozone profile is obtained by logarithmic transformation:

[0137]

[0138] in This represents the ozone volume mixing ratio. Logarithmic transformation can make the numerical distribution closer to a Gaussian distribution, which is beneficial for network training.

[0139] 1.3 Network Structure Configuration

[0140] The specific configuration of the Bayesian deep learning network used in this implementation case is shown in Table 2:

[0141] Table 2

[0142]

[0143] The total number of network parameters is approximately 1.2M, of which the Bayesian layer parameters account for 60%.

[0144] 1.4 Loss Function and Training Configuration

[0145] (1) Weight setting of loss function

[0146] The weights of each term in the multi-task joint loss function are determined through grid search:

[0147]

[0148] The selection criteria for the loss function weights are shown in Table 3.

[0149] Table 3

[0150]

[0151] (2) The training hyperparameter configuration is shown in Table 4.

[0152] Table 4

[0153]

[0154] (3) Monitoring of the training process

[0155] The typical evolution of each loss term during training is shown in Table 5:

[0156] Table 5

[0157]

[0158] 1.5 Reasoning and Quantification of Uncertainty

[0159] (1) Monte Carlo sampling configuration

[0160] The inference phase employs a dual Monte Carlo sampling strategy:

[0161] Number of weighted sampling times: M=20;

[0162] Dropout sampling times: T=5;

[0163] Total number of forward propagations: M×T=100 times.

[0164] (2) Comparison of inference time (single profile, GPU: NVIDIA V100) is shown in Table 6.

[0165] Table 6

[0166]

[0167] (3) Uncertainty decomposition results

[0168] Uncertainty decomposition was performed on the test set, and the uncertainty decomposition statistics (unit: ppmv) are shown in Table 7:

[0169] Table 7

[0170]

[0171] The results show that the uncertainty in the lower atmosphere mainly comes from measurement noise (random uncertainty), while the uncertainty in the upper atmosphere mainly comes from insufficient training data (cognitive uncertainty).

[0172] 1.6 Verification of Inversion Accuracy

[0173] (1) Comparison with the actual profile:

[0174] The inversion accuracy was evaluated using an independent test set, and the results are shown in Table 8.

[0175] Table 8

[0176]

[0177] (2) Comparison with traditional methods

[0178] A comparison with NASA's official OMPS-LP V2.6 product (based on the OE method) is shown in Table 9:

[0179] Table 9

[0180]

[0181] (3) The average kernel matrix prediction accuracy is shown in Table 10.

[0182] Table 10

[0183]

[0184] Example 2: This example refers to the method proposed in Example 1, and uses the edge imaging spectrometer (OMS-L) carried by my country's Fengyun-3E / F satellite as a platform to illustrate the application of the method of the present invention on domestic satellite data.

[0185] 2.1 Instrument Characteristics and Data Description

[0186] The main parameters of the OMS-L instrument are shown in Table 11.

[0187] Table 11

[0188]

[0189] 2.2 Multi-component joint inversion strategy

[0190] This embodiment achieves the joint inversion of ozone and nitrogen dioxide, and the network structure is adjusted accordingly:

[0191] The input feature design is shown in Table 12.

[0192] Table 12

[0193]

[0194] The multi-task output configuration is shown in Table 13.

[0195] Table 13

[0196]

[0197] 2.3 Training Data Construction

[0198] Since the actual observation data of OMS-L is limited, a transfer learning strategy is adopted:

[0199] (1) Pre-training phase

[0200] • Data source: Synthetic data based on OMS-L instrument response function simulation;

[0201] • Atmospheric state sampling: MERRA-2 (O3) + GEOS-Chem (NO2);

[0202] • Sample size: 500,000 records;

[0203] • Number of pre-training epochs: 100 epochs.

[0204] (2) Fine-tuning stage

[0205] • Data source: OMS-L L1B measured data and TROPOMI / OMI cross-validation data;

[0206] • Sample size: 50,000 records;

[0207] • Fine-tuning rounds: 50 epochs;

[0208] • Learning rate: 1 / 10 of the initial learning rate.

[0209] 2.4 Verification of Inversion Results

[0210] The accuracy of OMS-L ozone profile inversion (compared with ozone sounding) is shown in Table 14.

[0211] Table 14

[0212]

[0213] The accuracy of OMS-L nitrogen dioxide profile inversion (compared with ACE-FTS) is shown in Table 15.

[0214] Table 15

[0215]

[0216] Example 3: This example proposes a near-edge atmospheric profile inversion system based on Bayesian deep learning, including:

[0217] The data preprocessing module is configured to perform the following processes: receive radiance spectra and auxiliary data from limb observations, and perform standardization and feature construction;

[0218] The Bayesian neural network inference module is configured to perform the following processes: loading a pre-trained deep ensemble model and performing multiple Monte Carlo forward propagations;

[0219] The uncertainty quantification module is configured to perform the following processes: statistically integrating forecast results, calculating random uncertainty and cognitive uncertainty;

[0220] The average kernel reconstruction module is configured to perform the following process: reconstruct the complete average kernel matrix from the low-rank factors output by the network;

[0221] The product generation module is configured to perform the following process: output inversion results, uncertainties, and average kernel products that conform to the standard format of satellite data products.

[0222] The integrated deployment scheme of this embodiment in the satellite ground application system is as follows:

[0223] 3.1 System Architecture Design

[0224] The hardware configuration of the business system is shown in Table 16.

[0225] Table 16

[0226]

[0227] 3.2 Processing efficiency indicators

[0228] The assessment of business processing capabilities is shown in Table 17.

[0229] Table 17

[0230]

[0231] 3.3 Product Output Specifications

[0232] The output product specifications are shown in Table 18.

[0233] Table 18

[0234]

[0235] 3.4 Quality Control Process

[0236] The automatic quality control rules are shown in Table 19. A comprehensive comparison between the method of this invention and existing methods is shown in Table 20.

[0237] Table 19

[0238]

[0239] Table 20

[0240]

[0241] Example 4: This example proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the control method as described in this invention.

[0242] Example 5: This example proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of the method described in this invention.

[0243] It should be noted that the processing flow of embodiments 2-5 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.

[0244] 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.

[0245] 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 inverting atmospheric profiles at proximal edges based on Bayesian deep learning, characterized in that, Includes the following steps: To acquire edge observation data, a multi-branch Bayesian neural network model is constructed. The model includes a shared feature extraction backbone network, a profile estimation branch, a heteroscedasticity uncertainty estimation branch, and an average kernel prediction branch. The shared feature extraction backbone network adopts a variational Bayesian inference framework, modeling the network weight parameters as a probability distribution. The average kernel prediction branch outputs a low-rank decomposition representation of the average kernel matrix. Design a multi-task joint loss function, which includes negative log-likelihood loss, KL divergence regularization term, average kernel consistency loss and physical constraint regularization term; Multiple independent Bayesian neural network models are trained using a deep ensemble learning strategy, and the decomposition and quantification of random uncertainty and cognitive uncertainty are achieved through Monte Carlo sampling and ensemble statistics. The near-edge observation data is input into the trained Bayesian neural network model, forward inference is performed, and a complete inversion product containing atmospheric composition profile estimates, stratification uncertainties, and average kernel matrix is ​​output.

2. The method according to claim 1, characterized in that, The low-rank decomposition of the average kernel matrix is ​​expressed as: ； in, It is a low-rank factor. For the preset rank parameter and , The number of vertical layers. This is a diagonal correction term.

3. The method according to claim 1, characterized in that, The average kernel consistency loss is defined as: ； in, The average kernel matrix of the network predictions. The reference average kernel matrix is ​​calculated using the optimal estimation method. It is the Frobenius norm. The trace of the matrix, These are the trace constraint weight coefficients.

4. The method according to claim 1, characterized in that, The decomposition and quantification method for the aforementioned random uncertainty and cognitive uncertainty is as follows: for Each integrated member and each member After Monte Carlo sampling, the total variance is decomposed into: ； Among them, random uncertainty , The random uncertainty of the output of the m-th ensemble member at the t-th sampling; cognitive uncertainty. , This is a single predicted value. It is the average predicted value.

5. The method according to claim 1, characterized in that, The physical constraint regularization term includes: Non-negativity constraint terms: ; Vertical smoothing constraint term: ,in This represents the concentration value of the atmospheric component in the i-th layer predicted by the neural network; Chapman layer shape prior constraint: For a target gas with a characteristic peak distribution, constrain the deviation of the peak position and shape parameters of the inverted profile from the climatological Chapman layer parameters.

6. The method according to claim 1, characterized in that, The edge observation data comes from satellite edge detection instruments, including the OMPS edge profiler, SCIAMACHY, MLS, or the OMS-L payload of the Fengyun series satellites in China; the satellite edge detection instruments measure multi-wavelength radiance spectra, observation geometric parameters, prior atmospheric profiles, and auxiliary meteorological data at different tangent heights; normalization and feature construction are performed on the input data to generate network input feature vectors.

7. The method according to claim 1, characterized in that, The Bayesian neural network model trained by the method is applied to a new target instrument through cross-instrument transfer. The transfer learning steps include: The input feature distribution is adjusted using domain adaptive techniques, while keeping the structure of the average kernel prediction branch unchanged, and only the profile estimation branch and the uncertainty estimation branch are fine-tuned.

8. A system for inverting atmospheric profiles at limb edges based on Bayesian deep learning, characterized in that, include: The data preprocessing module is configured to perform the following processes: receive radiance spectra and auxiliary data from limb observations, and perform standardization and feature construction; The Bayesian neural network inference module is configured to perform the following processes: loading a pre-trained deep ensemble model and performing multiple Monte Carlo forward propagations; The uncertainty quantification module is configured to perform the following processes: statistically integrating forecast results, calculating random uncertainty and cognitive uncertainty; The average kernel reconstruction module is configured to perform the following process: reconstruct the complete average kernel matrix from the low-rank factors output by the network; The product generation module is configured to perform the following process: output inversion results, uncertainties, and average kernel products that conform to the standard format of satellite data products.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 7.

Images