An atmospheric aerosol particle parameter inversion method and device and electronic equipment

Abstract

The application provides an inversion method and device for atmospheric aerosol particle parameters and an electronic device. The inversion method parameterizes non-spherical aerosol particles to obtain an aerosol parameter vector; a trained physical guided neural network model is constructed and obtained; the trained physical guided neural network model is used for forward simulation of a radiation transfer model and calculation of the derivative of the output with respect to the input aerosol parameters, and the Jacobian matrix of the output with respect to the aerosol parameters is determined through automatic differentiation; based on the Jacobian matrix and the top-layer atmospheric observation to be inverted, inversion iteration is performed to determine the target aerosol parameters corresponding to the top-layer atmospheric observation to be inverted, thereby solving the problem that the traditional scattering model is not derivable and the derivative is difficult to obtain, and improving the aerosol remote sensing inversion accuracy and efficiency.

Description

Technical Field

[0001] This application relates to the field of atmospheric optics and satellite remote sensing inversion, and more specifically, to a method, apparatus and electronic equipment for inverting atmospheric aerosol particle parameters. Background Technology

[0002] Atmospheric aerosols are crucial factors influencing climate, environment, and weather processes. Their microphysical properties (such as aerosol particle size distribution, complex refractive index, and shape) are key aspects of aerosol monitoring. Atmospheric aerosol particles are highly complex, especially non-spherical particles. Conventional Mie scattering theory only applies to ideal spherical particles, while real-world atmospheric aerosols are often non-spherical or irregularly shaped mixtures (e.g., dust, mineral particles, black carbon clusters). For non-spherical particles, current techniques primarily employ numerical scattering models such as the T-matrix method and the discrete dipole approximation (DDA). While these models can describe the scattering process of non-spherical particles, they are inherently highly nonlinear, lack differentiable analytical expressions, and cannot directly obtain explicit partial derivatives of the output with respect to input parameters (e.g., complex refractive index, particle size, or shape). This makes them difficult to directly embed into satellite inversion of aerosol parameters based on optimization estimation theory. Existing aerosol inversion methods often assume spherical or fixed particle shapes, which limits the accuracy of inversion for non-spherical aerosol mixtures. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, apparatus and electronic equipment for inverting atmospheric aerosol particle parameters, which solves the problems of non-differentiability and difficulty in obtaining derivatives in traditional scattering models, and improves the accuracy and efficiency of aerosol remote sensing inversion.

[0004] The method for inverting atmospheric aerosol particle parameters provided in this application includes:

[0005] S1: Parameterize the non-spherical aerosol particles and construct an aerosol parameter vector to characterize the microphysical properties of aerosols;

[0006] S2: Construct a physical-guided neural network model. The input of the physical-guided neural network model includes the aerosol parameter vector and satellite observation geometry. The output is the predicted atmospheric top-layer observation. The atmospheric top-layer observation is the atmospheric top-layer radiance or the Stokes parameter of the atmospheric top-layer.

[0007] S3: Train the physical guidance neural network model using a pre-prepared training dataset to obtain a trained physical guidance neural network model;

[0008] S4: Use the trained physical-guided neural network model for forward modeling of radiative transfer and calculate the derivative of the output with respect to the input aerosol parameters. Determine the Jacobian matrix of the output with respect to the aerosol parameters through automatic differentiation.

[0009] S5: Based on the Jacobi matrix and the atmospheric top-level observations to be inverted, perform inversion iteration to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted.

[0010] In some embodiments, in the method for inverting atmospheric aerosol particle parameters, the loss function of the physical-guided neural network model includes a data fitting term and a physical constraint term based on physical rules; the data fitting term characterizes the deviation between the predicted atmospheric top-level observations and the actual atmospheric top-level observations; the physical constraint term characterizes whether the output prediction result conforms to physical laws.

[0011] The physical constraints include at least one of the following: residuals of the radiative transfer equation, energy conservation conditions, and phase function normalization conditions.

[0012] In some embodiments, in the method for inverting atmospheric aerosol particle parameters, the loss function for:

[0013]

[0014] Among them, the Characterizing the data fitting term, the Characterizing the physical constraint term, the The weighting coefficients characterize the physical constraint terms.

[0015] In some embodiments, the method for inverting atmospheric aerosol particle parameters, wherein the parameterization description of non-spherical aerosol particles and the construction of an aerosol parameter vector for characterizing aerosol microphysical properties include:

[0016] For clustered light-absorbing particles, a parameterized description is performed based on a fractal aggregation model; the parameters of the clustered light-absorbing particles include: fractal dimension, fractal factor, rotation radius, number of basic spheres in the cluster, and radius of a single sphere;

[0017] For non-spherical mineral aerosols, a parametric description is performed based on an ellipsoidal model; the parameters of the non-spherical mineral aerosols include the equivalent sphere radius and aspect ratio.

[0018] For spherical aerosols, classical Mie theory is used to simulate scattering for parameterization; the parameters of the spherical aerosols include: effective radius, effective variance, and complex refractive index.

[0019] In some embodiments, the method for inverting atmospheric aerosol particle parameters involves using a trained physical-guided neural network model for forward modeling of a radiative transfer model and calculating the derivative of the output with respect to the input aerosol parameters. The method then determines the Jacobian matrix of the output relative to the aerosol parameters through automatic differentiation, including:

[0020] When calculating the Jacobian matrix, the atmospheric top-level observations output by the physical-guided neural network model are... aerosol parameters Taking the partial derivative, we obtain the Jacobian matrix. Jacobi matrix The first in ij The elements are:

[0021] ;

[0022] in, Jacobi matrix The first in ij The element represents the partial derivative of the i-th output with respect to the j-th input parameter; The i-th output of the physics-guided neural network model is represented. The j-th input parameter characterizes the physics-guided neural network model; The input value represents the input value of the p-th neuron in the L-th layer; This represents the input value of the r-th neuron in layer l.

[0023] In some embodiments, the method for inverting atmospheric aerosol particle parameters, wherein the inversion iteration based on the Jacobian matrix and the atmospheric top-level observations to be inverted to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted includes:

[0024] Based on the Jacobi matrix, the parameter correction amount for aerosol parameters is determined during the inversion iteration of the atmospheric top-level observations to be inverted.

[0025] The aerosol parameters are updated in the next iteration based on the parameter correction, and the aerosol parameters are finally inverted to determine the target aerosol parameters corresponding to the top-level atmospheric observations to be inverted.

[0026] In some embodiments, the method for inverting atmospheric aerosol particle parameters, based on the Jacobian matrix, determines the parameter correction amount for aerosol parameters during the inversion iteration of the atmospheric top-level observations to be inverted, including:

[0027] In each inversion iteration, the difference between the predicted atmospheric top-layer observations and the atmospheric top-layer observations to be inverted corresponding to the aerosol parameters in a single iteration is calculated.

[0028] Based on the Jacobian matrix, a linear approximation relationship is constructed between the difference and the aerosol parameters in this iteration;

[0029] Solving the linear approximation relationship determines the parameter correction amount for the aerosol parameter in a single iteration.

[0030] In some embodiments, the training dataset in the atmospheric aerosol particle parameter inversion method is constructed based on the following steps:

[0031] Obtain sample aerosol parameter vectors and sample satellite observation geometric conditions from multiple sets of sample data;

[0032] Based on the physical scattering model, the scattering output corresponding to the sample aerosol parameter vector and the sample satellite observation geometry in each set of sample data is calculated, and the scattering output is coupled into the radiative transfer model for forward modeling to obtain the top atmospheric observation of each set of sample data.

[0033] In some embodiments, an inversion device for atmospheric aerosol particle parameters is also provided, the inversion device comprising:

[0034] The first construction module is used to parameterize the non-spherical aerosol particles and construct an aerosol parameter vector to characterize the microphysical properties of aerosols.

[0035] The second construction module is used to construct a physics-guided neural network model. The input of the physics-guided neural network model includes the aerosol parameter vector and satellite observation geometry, and the output is the predicted atmospheric top-layer observation. The atmospheric top-layer observation is the atmospheric top-layer radiance or the Stokes parameter of the atmospheric top-layer.

[0036] The training module is used to train the physical guided neural network model using a pre-prepared training dataset to obtain a trained physical guided neural network model.

[0037] The first determining module is used to apply the trained physical-guided neural network model to the forward modeling of the radiative transfer model and calculate the derivative of the output with respect to the input aerosol parameters, and determine the Jacobian matrix of the output relative to the aerosol parameters through automatic differentiation.

[0038] The second determining module is used to perform inversion iteration based on the Jacobi matrix and the atmospheric top-level observations to be inverted, and to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted.

[0039] In some embodiments, an electronic device is also provided, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps of the atmospheric aerosol particle parameter inversion method are performed.

[0040] This application provides a method, apparatus, and electronic device for inverting atmospheric aerosol particle parameters. First, the geometric and compositional characteristics of complex non-spherical aerosol particles are described in a parameterized manner. Then, a trainable deep neural network is constructed, incorporating physical scattering laws into the loss function to ensure that the network output is consistent with the actual physical scattering model. The sensitivity of the optical output to microphysical parameters (Jacobi matrix) is obtained through automatic differentiation for use by the inversion algorithm, thereby solving the problems of non-differentiability and difficulty in obtaining derivatives in traditional scattering models, and improving the accuracy and efficiency of aerosol remote sensing inversion. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of the method for inverting atmospheric aerosol particle parameters according to an embodiment of this application is shown;

[0043] Figure 2 A flowchart illustrating the method for constructing a training dataset according to an embodiment of this application is shown;

[0044] Figure 3 A flowchart of the method for determining the target aerosol parameters corresponding to the top-level atmospheric observations to be inverted, as described in an embodiment of this application, is shown.

[0045] Figure 4 A flowchart of the method for determining the parameter correction amount for aerosol parameters during the inversion iteration of atmospheric top-level observations to be inverted, as described in an embodiment of this application, is shown.

[0046] Figure 5 This paper shows a schematic diagram of the structure of the atmospheric aerosol particle parameter inversion device described in an embodiment of this application;

[0047] Figure 6 A schematic diagram of the structure of the electronic device described in an embodiment of this application is shown. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0049] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0050] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0051] Atmospheric aerosols are crucial factors influencing climate, environment, and weather processes. Their microphysical properties (such as aerosol particle size distribution, complex refractive index, and shape) are key aspects of aerosol monitoring. Atmospheric aerosol particles are highly complex, especially non-spherical particles. Conventional Mie scattering theory only applies to ideal spherical particles, while real-world atmospheric aerosols are often non-spherical or irregularly shaped mixtures (e.g., dust, mineral particles, black carbon clusters). For non-spherical particles, current techniques primarily employ numerical scattering models such as the T-matrix method and the discrete dipole approximation (DDA). While these models can describe the scattering process of non-spherical particles, they are inherently highly nonlinear, lack differentiable analytical expressions, and cannot directly obtain explicit partial derivatives of the output with respect to input parameters (e.g., complex refractive index, particle size, or shape). This makes them difficult to directly embed into satellite inversion of aerosol parameters based on optimization estimation theory. Existing aerosol inversion methods often assume spherical or fixed particle shapes, which limits the accuracy of inversion for non-spherical aerosol mixtures.

[0052] Based on this, this application provides a method, apparatus, and electronic device for inverting atmospheric aerosol particle parameters. First, the geometric and compositional characteristics of complex non-spherical aerosol particles are described in a parameterized manner. Then, a trainable deep neural network is constructed, incorporating physical scattering laws into the loss function to ensure that the network output is consistent with the actual physical scattering model. The sensitivity of the optical output to microphysical parameters (Jacobi matrix) is obtained through automatic differentiation for use by the inversion algorithm, thereby solving the problems of non-differentiability and difficulty in obtaining derivatives in traditional scattering models, and improving the accuracy and efficiency of aerosol remote sensing inversion.

[0053] Please refer to Figure 1 , Figure 1 A flowchart of the method for inverting atmospheric aerosol particle parameters according to an embodiment of this application is shown; as follows: Figure 1 As shown, the method for inverting atmospheric aerosol particle parameters includes the following steps S1-S5:

[0054] S1: Parameterize the non-spherical aerosol particles and construct an aerosol parameter vector to characterize the microphysical properties of aerosols;

[0055] S2: Construct a physical-guided neural network model. The input of the physical-guided neural network model includes the aerosol parameter vector and satellite observation geometry. The output is the predicted atmospheric top-layer observation. The atmospheric top-layer observation is the atmospheric top-layer radiance or the Stokes parameter of the atmospheric top-layer.

[0056] S3: The physical guided neural network model is trained using a pre-prepared training dataset to obtain a trained physical guided neural network model; the training dataset is constructed based on real atmospheric top-level observations;

[0057] S4: Use the trained physical-guided neural network model for forward modeling of radiative transfer and calculate the derivative of the output with respect to the input aerosol parameters. Determine the Jacobian matrix of the output with respect to the aerosol parameters through automatic differentiation.

[0058] S5: Based on the Jacobi matrix and the atmospheric top-level observations to be inverted, perform inversion iteration to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted.

[0059] In step S1, non-spherical aerosol particles are parameterized to construct an aerosol parameter vector for characterizing the microphysical properties of aerosols.

[0060] In the real atmosphere, aerosol particles have complex and diverse forms. For example, black carbon clusters are often formed by the aggregation of irregular spheres, while mineral dust is in the form of polyhedral or flaky particles, and there are even mixed phases and cluster structures.

[0061] In some embodiments, the parameterization of non-spherical aerosol particles to construct an aerosol parameter vector for characterizing the microphysical properties of aerosols includes:

[0062] For clustered light-absorbing particles, a parameterized description is performed based on a fractal aggregation model; the parameters of the clustered light-absorbing particles include: fractal dimension. fractal factor Radius of rotation Number of basic spheres in a cluster radius of a single sphere ;

[0063] For non-spherical mineral aerosols, a parameterized description is performed based on an ellipsoidal model; the parameters of the non-spherical mineral aerosols include the equivalent sphere radius. Aspect Ratio ;

[0064] For spherical aerosols, classical Mie theory is used to simulate scattering for parameterization; the parameters of the spherical aerosol include: effective radius. Effective variance and complex refractive index Complex refractive index middle, Characterizing the complex refractive index, Characterizing the real part of the complex refractive index; Represents the imaginary unit; Characterizes the imaginary part of the complex refractive index.

[0065] In some embodiments, a parameterized mixed-state model is employed for complex aerosols.

[0066] Specifically, clustered light-absorbing particles are represented by black carbon, non-spherical mineral aerosols are represented by sand and dust, and spherical aerosols are more common, such as sulfates, nitrates, organic carbon, and sea salt.

[0067] It should be noted that the above parameters are the main parameters. In some embodiments, other parameters may also be included in the parameterization description of non-spherical aerosol particles.

[0068] For example, using fractal dimension fractal factor Radius of rotation The structure of black carbon clusters is described by parameters, and their relationship can be expressed by a fractal aggregation formula:

[0069] (1)

[0070] in, The number of basic spheres in the cluster. The radius of a single sphere, fractal dimension Let be the radius of rotation. The fractal factor is used; for non-spherical dust, ellipsoidal or polyhedral parameters can be used to describe it, such as including the equivalent sphere radius. Shape parameters (such as aspect ratio) For mixed-state particles, volume ratio or mass fraction is used. Indicates the proportions of different components.

[0071] In summary, we define the aerosol parameter vector. Fully describe the microphysical characteristics of complex aerosols.

[0072] The aerosol parameter vector That is, the parameter vector to be inverted, which is the parameter that needs to be obtained by inverting the observations of the top of the atmosphere.

[0073] In step S2, a physical-guided neural network model is constructed. The input of the physical-guided neural network model includes the aerosol parameter vector and satellite observation geometry, and the output is the predicted atmospheric top-layer observation. The atmospheric top-layer observation is the atmospheric top-layer radiance or the Stokes parameter of the atmospheric top-layer.

[0074] In some embodiments, the loss function of the physics-guided neural network model includes a data fitting term and a physics constraint term based on physics rules; the data fitting term characterizes the deviation between the predicted atmospheric top-level observations and the actual atmospheric top-level observations; the physics constraint term characterizes whether the output prediction results conform to physical laws.

[0075] The physical constraints include at least one of the following: residuals of the radiative transfer equation, energy conservation conditions, and phase function normalization conditions.

[0076] The loss function for:

[0077]

[0078] Among them, the Characterizing the data fitting term, the Characterizing the physical constraint term, the The weighting coefficients characterize the physical constraint terms.

[0079] To train the physical guidance neural network model, a training dataset needs to be constructed; please refer to... Figure 2 The training dataset is constructed based on the following steps S201-S202:

[0080] S201. Obtain the sample aerosol parameter vector and sample satellite observation geometry from multiple sets of sample data;

[0081] S202. Based on the physical scattering model, calculate the scattering output corresponding to the sample aerosol parameter vector and sample satellite observation geometry in each set of sample data, and couple the scattering output into the radiative transfer model for forward modeling to obtain the sample atmospheric top-level observation in each set of sample data.

[0082] The final training dataset includes a large number of sample pairs, each of which includes:

[0083] Input features: Comprising two parts: a) Aerosol microphysical parameter vector β: such as effective particle radius, complex refractive index, fractal dimension, composition ratio, etc.; b) Satellite observation geometry: such as solar zenith angle, observation zenith angle, relative azimuth angle. Output features: Atmospheric top-level observations corresponding to the input features, such as radiance or Stokes parameters.

[0084] The physical scattering model can be the T-matrix method, DDA, or multi-sphere Mie method, etc.

[0085] Based on this, the corresponding scattering output is calculated using methods such as the T-matrix method, DDA, or multi-sphere Mie, and then coupled into the radiative transfer model for forward modeling. The physical model is used to generate training data samples, and the neural network is trained to achieve high fitting accuracy to the training data while satisfying physical constraints.

[0086] Here, classical scattering theory is introduced as the physical basis, for a given aerosol parameter vector The scattering characteristics of particles can be calculated using methods such as the T matrix, multi-sphere Mie, or DDA, to determine their scattering phase function. Scattering angle, single scattering albedo and extinction efficiency factor Equal optical quantities; during radiative transmission, assuming the atmosphere has a parallel layer structure, then the photon in the optical thickness coordinate... The propagation on the surface satisfies the radiative transfer equation (vector form) as follows: (2)

[0087]

[0088] in, For Stokes vectors, Let cosine be the zenith angle of the direction of emission. Let cosine be the zenith angle of the incident direction. The azimuth angle of the launch direction. The azimuth angle is the direction of incidence. Sun zenith cosine The azimuth of the sun. The vertical optical thickness of the entire atmosphere; The Mueller matrix is ​​obtained by rotating the scattering phase function through a reference plane. This represents the solar radiation flux density. This is an aerosol parameter vector. Albedo is the single-scattering albedo. The specific scattering phase function elements are calculated from the physical model. For the top-level radiance observed by satellite... For example, for Stokes vectors The first item.

[0089] Construct a physics-guided neural network model; specifically, construct a deep neural network (PINN) for approximate mapping. or approximate mapping .

[0090] Deep neural networks (PINNs) output scattering-related quantities to embed physical constraints. Network inputs include aerosol parameters. Including observation geometry (such as solar zenith angle, azimuth angle, satellite zenith angle, etc.); the output is the corresponding radiance. and Consider the Stokes vector for polarization. The network structure can be a multi-layer fully connected network, including an input layer, multiple hidden layers, and an output layer. The input layer nodes represent parameters and geometric quantities, the multiple hidden layers use nonlinear activation functions (such as tanh or ReLU), and the output layer provides the predicted optical output, which is a linear function.

[0091] During training, the network loss function consists of both a data fitting term and a physical constraint term:

[0092] Data fitting term: Training samples generated using physical scattering models (T matrix, multi-sphere Mie, DDA, etc.). The prediction results of the network output Calculate the mean square error ;in, middle, The aerosol parameter vector representing the i-th sample. Characterize the atmospheric top-level observations of the i-th sample.

[0093] Physical loss term: The residual constructed based on the radiative transfer equation (2), for example, by forcibly satisfying the equation at several scattering angles and depth nodes. Or apply the laws of conservation of energy and optical theorems. Given the known relationships, add the loss; where Characterizes extinction efficiency; The imaginary part of the forward scattering amplitude S(0) of the scattering matrix; S(0) The value of the scattering amplitude function in the forward direction (i.e., when the scattering angle is 0°). The total loss can be written as...

[0094] (3)

[0095] in For the network weight parameter set, The weighting coefficients for the physical loss term, the Characterizing the data fitting term, the Characterizes physical constraints.

[0096] In some embodiments, the weighting coefficient It can be set up through methods such as cross-validation.

[0097] Through this physical loss term, the network is guided to learn a mapping that satisfies the physical laws of real scattering, thereby alleviating the overfitting problem of purely data-driven models.

[0098] The input layer of the physics-guided neural network model receives aerosol parameter vectors. Based on satellite observation geometric information, the intermediate hidden layer performs nonlinear feature extraction, and the output layer provides predicted values ​​of radiance or Stokes vector. Through automatic differentiation (Autodiff), the partial derivatives of the output with respect to each input parameter can also be explicitly calculated, thus directly obtaining the Jacobian matrix.

[0099] In step S3, the physical guidance neural network model is trained using a pre-prepared training dataset to obtain a trained physical guidance neural network model.

[0100] The Adam or gradient descent optimization algorithm is used during the training of the physical guided neural network model.

[0101] A batch of input features from the training dataset is fed into the PINN model. The data is passed layer by layer in the network and processed by linear transformations and nonlinear activation functions (such as ReLU and tanh). Finally, the network outputs the predicted atmospheric top-level observations corresponding to this batch of inputs.

[0102] The total loss for this training is calculated based on the loss function. Then, using a gradient descent optimization algorithm (such as Adam), all weights and bias parameters in the network are updated in reverse according to the calculated gradient direction to reduce the total loss.

[0103] The network parameters are iteratively optimized by traversing multiple batches in the training dataset. After each batch of training is completed, the loss is calculated using the validation set to monitor model performance and prevent overfitting. Training stops when the validation loss no longer decreases significantly or when the preset number of iterations is reached, resulting in a well-trained and converged PINN model.

[0104] A well-trained physical-guided neural network model, given any set of input parameters and geometric conditions, can instantly simulate the corresponding top-level atmospheric observations; it can also efficiently and analytically calculate the Jacobian matrix K of the observations with respect to the input parameters through automatic differentiation.

[0105] In step S4, the trained physical-guided neural network model is used for forward modeling of the radiative transfer model and the derivative of the output with respect to the input aerosol parameters is calculated. The Jacobian matrix of the output relative to the aerosol parameters is determined by automatic differentiation.

[0106] After training, the PINN model establishes a differentiable functional mapping from particle parameters to observed radiance. Since the network consists of differentiable modules (linear transformations + differentiable activation functions), the network output... The partial derivatives with respect to the input parameters can be obtained by back-calculation between network layers using the chain rule.

[0107] The trained physical-guided neural network model is used for forward modeling of the radiative transfer model and the derivative of the output with respect to the input aerosol parameters is calculated. The Jacobian matrix of the output relative to the aerosol parameters is determined by automatic differentiation, including:

[0108] When calculating the Jacobian matrix, the atmospheric top-level observations output by the physical-guided neural network model are... aerosol parameters Taking the partial derivative, we obtain the Jacobian matrix. Jacobi matrix The first in ij The elements are:

[0109] (4);

[0110] in, Jacobi matrix The first in ij The element represents the partial derivative of the i-th output with respect to the j-th input parameter; The i-th output of the physics-guided neural network model is represented. The j-th input parameter characterizes the physics-guided neural network model; The input value represents the input value of the p-th neuron in the L-th layer; This represents the input value of the r-th neuron in layer l.

[0111] Here, the Jacobi matrix The final result is accumulated layer by layer through the chain rule.

[0112] The calculation of the Jacobian matrix K is achieved using automatic differentiation technology, which avoids traditional numerical perturbation calculations and improves computational efficiency and accuracy.

[0113] In step S5, an inversion iteration is performed based on the Jacobi matrix and the atmospheric top-level observations to be inverted to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted.

[0114] Here, the calculated Jacobi matrix is ​​used to execute the Levenberg-Marquardt or variational algorithm within the optimization inversion framework to estimate aerosol morphology parameters, achieving fast convergence and high-precision inversion.

[0115] For example, in the actual inversion process, the Jacobi matrix can be used to map the observed radiance deviation to the particle parameter correction amount, guiding the gradient descent optimization.

[0116] Based on this, please refer to Figure 3 The inversion iteration based on the Jacobi matrix and the top-level atmospheric observations to be inverted, to determine the target aerosol parameters corresponding to the top-level atmospheric observations to be inverted, includes the following steps S301-S302:

[0117] S301. Based on the Jacobi matrix, determine the parameter correction amount for aerosol parameters during the inversion iteration of the atmospheric top-level observations to be inverted.

[0118] S302. Update the aerosol parameters for the next iteration based on the parameter correction amount, and finally realize the inversion of the aerosol parameters to determine the target aerosol parameters corresponding to the top-level atmospheric observations to be inverted.

[0119] Please refer to Figure 4 Based on the Jacobian matrix, the parameter correction amounts for aerosol parameters are determined during the inversion iteration of the atmospheric top-level observations to be inverted, including the following steps S401-S403:

[0120] S401. In each inversion iteration, calculate the difference between the predicted atmospheric top layer observations and the atmospheric top layer observations to be inverted corresponding to the aerosol parameters in a single iteration.

[0121] S402. Based on the Jacobian matrix, construct a linear approximation relationship between the difference and the aerosol parameters in this iteration;

[0122] S403. Solve the linear approximation relationship to determine the parameter correction amount for the aerosol parameter in a single iteration.

[0123] Before calculating the difference between the predicted atmospheric top-layer observations and the atmospheric top-layer observations to be inverted corresponding to the aerosol parameters in a single iteration, the method further includes:

[0124] The trained physical-guided neural network model processes aerosol parameters in a single iteration and outputs the predicted top-level atmospheric observations corresponding to the aerosol parameters in a single iteration.

[0125] In the actual inversion process, the obtained Jacobian matrix can be used to map the observed radiance deviation or Stokes parameter deviation to the particle parameter correction, guiding gradient descent optimization; in each iteration, the difference between the observed vector and the forward simulation result is linearized:

[0126] (5)

[0127] And update the parameters according to the least squares criterion. ;in, Characterizing the bias in observations at the top of the atmosphere, Characterizing the aerosol parameter vector deviation, The representation obtained is the Jacobian matrix.

[0128] In this way, compared with traditional numerical difference or adjoint methods, the analytical differentiability of neural networks avoids a large number of repeated calls to the scattering model, thereby significantly improving the efficiency and stability of inversion calculations.

[0129] It should be noted that performing inversion iterations on the atmospheric top-level observations to be inverted and determining the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted is a relatively complex process and is existing technology. The above steps only illustrate the role of the Jacobian matrix obtained in the embodiments of this application in the inversion process, and are not the complete inversion process. Other processes will not be described in detail.

[0130] In some embodiments, the process of implementing the atmospheric aerosol particle parameter inversion method is as follows:

[0131] (1) Training data generation: Select the aerosol parameter distribution under typical pollution scenarios, and use methods such as T matrix method, DDA or multi-sphere Mie to calculate the corresponding scattering output (such as scattering phase function, single scattering albedo and extinction efficiency factor). This data is then coupled into an atmospheric radiative transfer model for forward simulation, forming a training dataset.

[0132] (2) Network design and training: Design the number of layers and nodes per layer of PINN according to the parameter dimension and complexity, and select an appropriate activation function; add a constraint physical term to the loss function, and train the network using the backpropagation algorithm until the accuracy requirements are met.

[0133] (3) Model application and Jacobi calculation: The trained network is used as a substitute for the forward simulation of the radiative transfer model, given any input. The predicted TOA radiance or Stokes parameter is obtained through forward propagation, and the Jacobian matrix is ​​calculated using automatic differentiation. .

[0134] (4) Inversion integration: In the satellite observation inversion algorithm, this PINN replaces the traditional scattering module, and the obtained Jacobi matrix is ​​used to perform iterative optimization and uncertainty analysis.

[0135] The following is an example of the atmospheric aerosol particle parameter inversion method described in this application in a specific application scenario.

[0136] First, set the parameters to be inverted: the parameters of the black carbon clusters include the radius of a single sphere. a、 fractal dimension fractal factor and the volume ratio of black carbon (BC) to the total aerosol volume. The parameters of dust particles include the equivalent sphere radius. Aspect Ratio and dust composition ratio Parameters for particles such as organic carbon, sulfates, and nitrates include effective radius. Effective variance and complex refractive index The total optical thickness, including the vertical profile of the aerosol, is also included. .

[0137] Then, sample data were generated using the T-matrix method, DDA, or multi-sphere Mie method. Specifically, the Stokes four-component scattering matrix, single-scattering albedo, and extinction cross section were calculated for different aerosol particles in the visible light band (e.g., 550 nm). The influence of the organic carbon coating layer on particle scattering during internal mixing was considered in the calculation. The resulting sample set contained approximately several thousand instances, which served as training data for PINN.

[0138] Constructing PINN: The input layer contains parameters

[0139] The network employs four fully connected hidden layers, with 256, 256, 128, and 64 neurons per layer, respectively, and uses tanh as the activation function. The output layer predicts TOA radiance (depending on whether polarization is considered). The loss function is a Stokes vector and includes: (1) the mean square error of the network output compared with the sample data; (2) the residual constraints of the forward modeling of the corresponding radiative transfer model; (3) physical constraints such as optical conservation and phase function normalization. The Adam optimizer is used during training until the verification error converges to an acceptable range.

[0140] After training, a combination of input parameters from a real-world observation scenario is selected. The simulated radiance is obtained using PINN forward prediction, and the Jacobian matrix is ​​obtained through an automatic differentiation module. Assume that the multi-angle radiance deviation observed in a certain iteration is... Then the linear equation can be solved. Obtain parameter correction amount This is used to update the next iteration. After multiple iterations, the goal is to achieve... Estimation of parameters such as...

[0141] Based on the same inventive concept, this application also provides an inversion device for atmospheric aerosol particle parameters corresponding to the inversion method for atmospheric aerosol particle parameters. Since the principle of the device in this application is similar to the above-mentioned inversion method for atmospheric aerosol particle parameters in this application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0142] Please refer to Figure 5 , Figure 5 This illustration shows a schematic diagram of the structure of the atmospheric aerosol particle parameter inversion device according to an embodiment of this application. The inversion device includes:

[0143] The first construction module 501 is used to parameterize the non-spherical aerosol particles and construct an aerosol parameter vector to characterize the microphysical properties of aerosols.

[0144] The second construction module 502 is used to construct a physical-guided neural network model. The input of the physical-guided neural network model includes the aerosol parameter vector and satellite observation geometry, and the output is the predicted atmospheric top-layer observation. The atmospheric top-layer observation is the atmospheric top-layer radiance or the Stokes parameter of the atmospheric top-layer.

[0145] Training module 503 is used to train the physical guided neural network model using a pre-prepared training dataset to obtain a trained physical guided neural network model.

[0146] The first determining module 504 is used to use the trained physical guidance neural network model for forward modeling of radiative transfer and to calculate the derivative of the output with respect to the input aerosol parameters, and to determine the Jacobian matrix of the output relative to the aerosol parameters through automatic differentiation.

[0147] The second determining module 505 is used to perform inversion iteration based on the Jacobi matrix and the atmospheric top-level observations to be inverted, and to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted.

[0148] In some embodiments, in the atmospheric aerosol particle parameter inversion device, the loss function of the physical-guided neural network model includes a data fitting term and a physical constraint term based on physical rules; the data fitting term characterizes the deviation between the predicted atmospheric top-level observations and the actual atmospheric top-level observations; the physical constraint term characterizes whether the output prediction result conforms to physical laws.

[0149] The physical constraints include at least one of the following: residuals of the radiative transfer equation, energy conservation conditions, and phase function normalization conditions.

[0150] In some embodiments, in the atmospheric aerosol particle parameter inversion device, the loss function for:

[0151]

[0152] Among them, the Characterizing the data fitting term, the Characterizing the physical constraint term, the The weighting coefficients characterize the physical constraint terms.

[0153] In some embodiments, in the atmospheric aerosol particle parameter inversion device, the first construction module, when parameterizing non-spherical aerosol particles and constructing an aerosol parameter vector to characterize the microphysical properties of aerosols, is specifically used for:

[0154] For clustered light-absorbing particles, a parameterized description is performed based on a fractal aggregation model; the parameters of the clustered light-absorbing particles include: fractal dimension, fractal factor, rotation radius, number of basic spheres in the cluster, and radius of a single sphere;

[0155] For non-spherical mineral aerosols, a parametric description is performed based on an ellipsoidal model; the parameters of the non-spherical mineral aerosols include the equivalent sphere radius and aspect ratio.

[0156] For spherical aerosols, classical Mie theory is used to simulate scattering for parameterization; the parameters of the spherical aerosols include: effective radius, effective variance, and complex refractive index.

[0157] In some embodiments, in the atmospheric aerosol particle parameter inversion device, the first determining module, when using the trained physical-guided neural network model for forward modeling of the radiative transfer model and calculating the derivative of the output with respect to the input aerosol parameters, and determining the Jacobian matrix of the output relative to the aerosol parameters through automatic differentiation, is specifically used for:

[0158] When calculating the Jacobian matrix, the atmospheric top-level observations output by the physical-guided neural network model are... aerosol parameters Taking the partial derivative, we obtain the Jacobian matrix. Jacobi matrix The first in ij The elements are:

[0159] ;

[0160] in, Jacobi matrix The first in ij The element represents the partial derivative of the i-th output with respect to the j-th input parameter; The i-th output of the physics-guided neural network model is represented. The j-th input parameter characterizes the physics-guided neural network model; The input value represents the input value of the p-th neuron in the L-th layer; This represents the input value of the r-th neuron in layer l.

[0161] In some embodiments, in the atmospheric aerosol particle parameter inversion device, the second determining module, when performing inversion iterations based on the Jacobian matrix and the atmospheric top-level observations to be inverted to determine the target aerosol parameters corresponding to the atmospheric top-level observations to be inverted, is specifically used for:

[0162] Based on the Jacobi matrix, the parameter correction amount for aerosol parameters is determined during the inversion iteration of the atmospheric top-level observations to be inverted.

[0163] The aerosol parameters are updated in the next iteration based on the parameter correction, and the aerosol parameters are finally inverted to determine the target aerosol parameters corresponding to the top-level atmospheric observations to be inverted.

[0164] In some embodiments, in the atmospheric aerosol particle parameter inversion device, the second determining module, based on the Jacobian matrix, determines the parameter correction amount for aerosol parameters during the inversion iteration of the atmospheric top-level observations to be inverted, including:

[0165] In each inversion iteration, the difference between the predicted atmospheric top-layer observations and the atmospheric top-layer observations to be inverted corresponding to the aerosol parameters in a single iteration is calculated.

[0166] Based on the Jacobian matrix, a linear approximation relationship is constructed between the difference and the aerosol parameters in this iteration;

[0167] Solving the linear approximation relationship determines the parameter correction amount for the aerosol parameter in a single iteration.

[0168] In some embodiments, the atmospheric aerosol particle parameter inversion device further includes a third construction module for constructing the training dataset based on the following steps:

[0169] Obtain sample aerosol parameter vectors and sample satellite observation geometric conditions from multiple sets of sample data;

[0170] Based on the physical scattering model, the scattering output corresponding to the sample aerosol parameter vector and the sample satellite observation geometry in each set of sample data is calculated, and the scattering output is coupled into the radiative transfer model for forward modeling to obtain the top atmospheric observation of each set of sample data.

[0171] Based on the same inventive concept, this application also provides an electronic device corresponding to the method for inverting atmospheric aerosol particle parameters. Since the principle of solving the problem by the electronic device in this application is similar to the method for inverting atmospheric aerosol particle parameters described above in this application, the implementation of the electronic device can refer to the implementation of the method, and the repeated parts will not be described again.

[0172] Please refer to Figure 6 , Figure 6 A schematic diagram of the structure of the electronic device according to an embodiment of this application is shown. The electronic device 600 includes: a processor 602, a memory 601, and a bus. The memory 601 stores machine-readable instructions executable by the processor 602. When the electronic device 600 is running, the processor 602 communicates with the memory 601 through the bus. When the machine-readable instructions are executed by the processor 602, the steps of the atmospheric aerosol particle parameter inversion method are performed.

[0173] Based on the same inventive concept, this application also provides a computer-readable storage medium corresponding to the method for inverting atmospheric aerosol particle parameters. Since the principle of the computer-readable storage medium in this application is similar to the method for inverting atmospheric aerosol particle parameters described above, the implementation of the computer-readable storage medium can refer to the implementation of the method, and the repeated parts will not be described again.

[0174] A computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the method for inverting atmospheric aerosol particle parameters.

[0175] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0176] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0177] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0178] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0179] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of inversion of parameters of atmospheric aerosol particles, characterized in that, The inversion method comprises: S1: parameterizing the non-spherical aerosol particles to construct an aerosol parameter vector for representing aerosol microphysical characteristics; S2: constructing a physics-guided neural network model, the input of the physics-guided neural network model comprising the aerosol parameter vector and satellite observation geometric conditions, and the output being a predicted atmospheric top layer observation; the atmospheric top layer observation being an atmospheric top layer radiance or an atmospheric top layer Stokes parameter; S3: training the physics-guided neural network model using a pre-prepared training data set to obtain a trained physics-guided neural network model; S4: using the trained physics-guided neural network model for radiative transfer model forward simulation and calculating the derivative of the output with respect to the input aerosol parameters, and determining the Jacobian matrix of the output with respect to the aerosol parameters through automatic differentiation; S5: performing inversion iteration based on the Jacobian matrix and the atmospheric top layer observation to be inverted to determine the target aerosol parameters corresponding to the atmospheric top layer observation to be inverted. The parameterization of the non-spherical aerosol particles to construct an aerosol parameter vector for representing aerosol microphysical characteristics comprises: For cluster-shaped light-absorbing particles, parameterization is performed based on a fractal aggregation model; the parameters of the cluster-shaped light-absorbing particles include fractal dimension, fractal factor, rotation radius, number of basic spheres in the cluster, and single sphere radius; for non-spherical mineral aerosols, parameterization is performed based on an ellipsoid model; the parameters of the non-spherical mineral aerosols include equivalent sphere radius and aspect ratio; for spherical aerosols, parameterization is performed by simulating scattering using classical Mie theory; the parameters of the spherical aerosols include effective radius, effective variance, and complex refractive index. using the trained physics-guided neural network model for forward modeling of a radiative transfer model and calculating the derivative of an output with respect to the aerosol parameters, determining a Jacobian matrix of the output with respect to the aerosol parameters by automatic differentiation, comprising: when calculating the Jacobian matrix, the atmospheric top-layer observation output by the physics-guided neural network model with respect to the aerosol parameters The loss function of the physics-guided neural network model comprises a data fitting term and a physical constraint term based on physical rules; the data fitting term represents the deviation between the predicted atmospheric top layer observation and the actual atmospheric top layer observation; the physical constraint term represents whether the output predicted result conforms to the physical law; the i-th element in the Jacobian matrix is​​​ ； wherein, is the Jacobian matrix is the element in the Jacobian matrix The physical constraint term comprises at least one of the following: radiative transfer equation residual, energy conservation condition, and phase function normalization condition. characterizing the partial derivative of the i-th output with respect to the j-th input parameter; characterizing the i-th output of the physics-guided neural network model; characterizing the j-th input parameter of the physics-guided neural network model; characterizing the input value of the p-th neuron in the L-th layer; characterizing the input value of the r-th neuron in the l-th layer.

2. The method of retrieval of parameters of atmospheric aerosol particles according to claim 1, characterized in that, The inversion iteration based on the Jacobian matrix and the atmospheric top layer observation to be inverted to determine the target aerosol parameters corresponding to the atmospheric top layer observation to be inverted comprises: Based on the Jacobian matrix, determining the parameter correction amount for the aerosol parameters in the process of inversion iteration of the atmospheric top layer observation to be inverted; 3. The method of retrieval of parameters of atmospheric aerosol particles according to claim 2, characterized in that, The loss function is: wherein the characterizing a data fitting term, the characterizing a physical constraint term, the characterizing a weight coefficient of the physical constraint term.

4. The method of claim 1, wherein, Based on the parameter correction amount, updating the aerosol parameters for the next iteration inversion, and finally realizing the inversion of the aerosol parameters to determine the target aerosol parameters corresponding to the atmospheric top layer observation to be inverted. Based on the Jacobian matrix, determining the parameter correction amount for the aerosol parameters in the process of inversion iteration of the atmospheric top layer observation to be inverted comprises: In each inversion iteration, calculating the difference between the predicted atmospheric top layer observation corresponding to the aerosol parameters in a single iteration and the atmospheric top layer observation to be inverted.

5. The method of retrieval of parameters of atmospheric aerosol particles according to claim 4, characterized in that, ​ ​ constructing a linear approximation relationship between the difference and the aerosol parameter in the current iteration based on the Jacobian matrix; solving the linear approximation relationship to determine a parameter correction amount for the aerosol parameter in a single iteration.

6. The method of claim 1, wherein The training data set is constructed based on the following steps: obtaining sample aerosol parameter vectors and sample satellite observation geometric conditions in multiple groups of sample data; calculating scattering outputs corresponding to the sample aerosol parameter vectors and sample satellite observation geometric conditions in each group of sample data based on a physical scattering model, and coupling the scattering outputs into a radiative transfer model for forward modeling to obtain sample top-of-atmosphere observations in each group of sample data.

7. An atmospheric aerosol particle parameter inversion device, characterized by, The inversion device comprises: a first construction module configured to parameterize non-spherical aerosol particles to construct an aerosol parameter vector representing aerosol microphysical properties; a second construction module configured to construct a physically guided neural network model, wherein the input of the physically guided neural network model comprises the aerosol parameter vector and satellite observation geometric conditions, and the output is predicted top-of-atmosphere observations; the top-of-atmosphere observations are top-of-atmosphere radiance or Stokes parameters at the top of the atmosphere; a training module configured to train the physically guided neural network model using a pre-prepared training data set to obtain a trained physically guided neural network model; a first determination module configured to use the trained physically guided neural network model for radiative transfer model forward modeling and to calculate the derivative of the output with respect to the input aerosol parameter, and to determine the Jacobian matrix of the output with respect to the aerosol parameter through automatic differentiation; a second determination module configured to perform inversion iteration based on the Jacobian matrix and the top-of-atmosphere observations to be inverted to determine target aerosol parameters corresponding to the top-of-atmosphere observations to be inverted; The first construction module, when parameterizing non-spherical aerosol particles to construct an aerosol parameter vector representing aerosol microphysical properties, is specifically configured to: parameterize cluster-shaped light-absorbing particles based on a fractal aggregation model; the parameters of the cluster-shaped light-absorbing particles include fractal dimension, fractal factor, rotation radius, number of basic spheres in the cluster, and single sphere radius; parameterize non-spherical mineral aerosols based on an ellipsoid model; the parameters of the non-spherical mineral aerosols include equivalent sphere radius and aspect ratio; and parameterize spherical aerosols by simulating scattering based on classical Mie theory; the parameters of the spherical aerosols include effective radius, effective variance, and complex refractive index. The first determining module is specifically configured to: when the Jacobian matrix is calculated, the atmospheric top layer observation quantity output by the physical guided neural network model to the aerosol parameters , the partial derivative is obtained to obtain the Jacobian matrix ; and the first element in the Jacobian matrix is The method comprises the following steps: ​ ； wherein, is the Jacobian matrix is the element in the Jacobian matrix a processor, a memory, and a bus, wherein the memory stores machine readable instructions executable by the processor, and the processor communicates with the memory through the bus when the electronic device is running, and the machine readable instructions are executed by the processor to perform the steps of the atmospheric aerosol particle parameter inversion method according to any one of claims 1 to 6. characterizing the partial derivative of the i-th output with respect to the j-th input parameter; characterizing the i-th output of the physics-guided neural network model; characterizing the j-th input parameter of the physics-guided neural network model; characterizing the input value of the p-th neuron in the L-th layer; characterizing the input value of the r-th neuron in the l-th layer.

8. An electronic device, comprising: ​ ​

Images