Cloud physical parameter inversion method based on convolutional neural network

Abstract

The present invention provides a cloud physics parameter inversion method based on a convolutional neural network, comprising the following steps: S1, calculating the scattering characteristics of cloud particles in the terahertz frequency band, combining the particle spectrum distribution, simulating and calculating the functional relationship between the terahertz radar equivalent reflectivity factor and the cloud physics characteristic parameters, and establishing a forward physical model; S2, preprocessing the measured data and microphysical parameter products of the terahertz cloud radar to obtain a training database; S3, establishing the structure of the convolutional neural network, using the forward physical model as a constraint, and adaptively adjusting the learning rate based on the deviation between the predicted value and the actual value of the radar echo; S4, repeatedly training the convolutional neural network using the training database, and obtaining the cloud physics characteristic parameters by inverting the terahertz radar equivalent reflectivity factor. This method improves the inversion accuracy of cloud physics parameters, eliminates the reliance on empirical values, and is an algorithm with stronger generalization and adaptive learning.

Description

Technical Field

[0001] The present invention relates to the field of meteorological remote sensing technology, and in particular to a cloud physical parameter inversion method based on a convolutional neural network. Background Art

[0002] Clouds are crucial components of the Earth-atmosphere system, and their physical parameters play a crucial role in research on the Earth-atmosphere radiation budget, weather forecasting, climate prediction, and artificial rainmaking. Therefore, accurately extracting cloud physical parameters from terahertz cloud radar remote sensing data is of great significance. Compared to common millimeter-wave cloud radars, terahertz cloud radars have a wavelength closer to the scale of cloud particles, theoretically offering higher detection capabilities and sensitivity, making them a new approach to active cloud remote sensing.

[0003] The method of extracting cloud physical parameters using radar reflectivity factors obtained from radar observations is called an inversion method. Currently, the commonly used cloud parameter inversion method is based on the statistical characteristics of the data. It comprehensively analyzes millimeter-wave radar and measured cloud physical parameters to derive an empirical relationship between the radar reflectivity factor and cloud physical parameters. This method is computationally simple, but the algorithm has poor applicability and scalability. Domestic and foreign scholars have proposed an inversion method based on optimal estimation theory. This method obtains the optimal solution while minimizing the cost function, and iteratively obtains the cloud physical parameters. The algorithm's applicability and scalability have been significantly improved, but the setting of key parameters still depends on statistical results.

[0004] Therefore, in order to improve the inversion accuracy of cloud physical parameters, it is necessary to get rid of the dependence on empirical values ​​and choose algorithms with stronger generalization and adaptive learning capabilities. Summary of the Invention

[0005] The present invention provides a cloud physical parameter inversion method based on convolutional neural networks. This method improves the inversion accuracy of cloud physical parameters and gets rid of the dependence on empirical values. It is an algorithm with stronger generalization and adaptive learning capabilities.

[0006] To achieve the above and other related objectives, the present invention provides a cloud physical parameter inversion method based on a convolutional neural network, comprising the following steps:

[0007] S1. Calculate the scattering characteristics of cloud particles in the terahertz band, combine the particle spectrum distribution, simulate the functional relationship between the terahertz radar equivalent reflectivity factor and cloud physical characteristic parameters, and establish a forward physical model. The cloud physical characteristic parameters include cloud particle size, cloud particle number concentration and distribution parameters;

[0008] S2. Preprocessing the measured data and microphysical parameter products of the terahertz cloud radar to obtain a training database, wherein the input data of the training database is the measured terahertz radar equivalent reflectivity factor, and the output data is the cloud physical characteristic parameters;

[0009] S3. Establishing a convolutional neural network structure, using the forward physical model as a constraint, and adaptively adjusting the learning rate based on the deviation between the predicted value and the actual value of the radar echo to achieve accurate representation of the complex nonlinear function between cloud remote sensing characteristics and output cloud parameter results;

[0010] S4. Utilizing the training database, repeatedly training the convolutional neural network, and obtaining the cloud physical characteristic parameters by inverting and calculating the terahertz radar equivalent reflectivity factor.

[0011] Preferably, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer. The data of the input layer is the input data. The convolutional neural network extracts features of the input data through the convolutional layer and the pooling layer to obtain local features of the input layer, and then generates the cloud physical feature parameters through the fully connected layer and enters the output layer.

[0012] Preferably, the calculation of the convolutional neural network specifically includes the following steps:

[0013] S3.1, forward calculation: a single training sample is input into the convolutional neural network, and the value of each neuron is calculated step by step from front to back to obtain the prediction result x of the training sample out ;

[0014] S3.2. Reverse calculation: First, establish the prediction result x out The loss function E between the real inversion result x and the original inversion result x adopts square error. The loss function E of the output layer is calculated as follows:

[0015]

[0016] Where n is the total number of distance bins of radar detection profiles. Then, the derivative of the network parameters with respect to the loss function E is calculated layer by layer, i.e., the local gradient δ. The gradient descent method is used to optimize w and b of each layer, and the network parameters are iteratively updated. The iterative operation is performed until convergence. The calculation formula is as follows:

[0017] w l+1 =w l -Δw l

[0018]

[0019] b l+1 =b l -Δb l

[0020]

[0021] Where η represents the learning rate and l represents the number of cycles during training;

[0022] S3.3, using the adaptive learning rate scheme, calculate the inversion result x based on the actual prediction out , according to the forward physical model, calculate the predicted value Z of the radar echo sim , calculate the predicted value Z sim The deviation from the measured value Z is used to correct the learning rate η, ensuring that the learning rate η gradually decreases as the accuracy increases. The calculation formula is as follows:

[0023]

[0024] Among them, η0 represents the initial value of the learning rate η.

[0025] Preferably, when the convolutional neural network is calculated, the actions of each layer are as follows:

[0026] The input layer passes the input data to the convolutional layer;

[0027] The convolution layer sets three convolution kernels, calculates the pre-activation value, and obtains the first feature map through the activation function f. The element value C of the first feature map outk (k=1,2,…,n+2) is:

[0028]

[0029] Wherein, w and b are the multiplicative and additive biases corresponding to each first feature map, f represents the activation function, and y represents the input data;

[0030] The pooling layer retains useful information while reducing the size of the first feature map input by the convolution layer. The first feature map is downsampled by using the average value aggregation method of adjacent matrix regions to complete the screening and dimension reduction of feature parameters to obtain a second feature map. The element value P of the second feature map is outk (k=1,2,…,n+2) is:

[0031] P outk =down k (C outk );

[0032] In the fully connected layer, the obtained second feature maps are sequentially connected into vectors and input into the fully connected network. After weighted and biased operations, the activation function f is used to obtain the prediction results of the cloud physical feature parameters and send them to the output layer.

[0033] Preferably, the activation function f is a ReLU function.

[0034] Preferably, the predicted value Z of the radar echo sim The calculation of depends on the forward physical model. The specific steps are: calculating the complex refractive index of ice crystal particles in the terahertz band; using the discrete dipole approximation method to calculate the scattering characteristics of cloud particles in the terahertz band to obtain the backscattering cross section σ bk Table; through the backscattering cross section σ bk The table is combined with the particle spectrum distribution to obtain the functional relationship between the terahertz radar equivalent reflectivity factor and the cloud physical characteristic parameters.

[0035] Preferably, the pre-processing of the measured data and microphysical parameter products of the terahertz cloud radar includes: eliminating rainfall cases and removing data with a large impact of ground clutter.

[0036] In summary, the present invention provides a convolutional neural network-based cloud physics parameter inversion method. By using a designed training database and a constructed convolutional neural network structure, the convolutional neural network is repeatedly trained using the training database to achieve convolutional neural network-based cloud physics parameter inversion. Based on the radar reflectivity factor of the terahertz cloud radar, the network is fed with input to obtain particle size, particle number concentration, and distribution parameters—the characteristic parameters of the cloud particle spectrum distribution. This method improves the inversion accuracy of cloud physics parameters and eliminates reliance on empirical values, resulting in an algorithm with enhanced generalization and adaptive learning capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of a cloud physical parameter inversion method based on a convolutional neural network provided by one embodiment of the present invention;

[0038] Figure 2 A schematic diagram of a specific implementation flow of a cloud physical parameter inversion method based on a convolutional neural network provided in one embodiment of the present invention;

[0039] Figure 3 A schematic diagram of the convolutional neural network structure in the cloud physical parameter inversion method based on convolutional neural network provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following is combined with Figure 1-3The cloud physical parameter inversion method based on convolutional neural network of the present invention is further described in detail with specific implementation methods. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the drawings are in a very simplified form and are not in precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the implementation method of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the drawings. It should be noted that the structure, proportion, size, etc. illustrated in the drawings of this specification are only used to match the content disclosed in the specification, so that people familiar with this technology can understand and read them, and are not used to limit the implementation conditions of the present invention. Therefore, they have no technical significance. Any modification of the structure, change of the proportional relationship or adjustment of the size should still fall within the scope of the technical content disclosed by the present invention without affecting the efficacy and purpose that can be achieved by the present invention.

[0041] See Figure 1 The present invention provides a cloud physical parameter inversion method based on convolutional neural network, comprising the following steps:

[0042] S1. Calculate the scattering characteristics of cloud particles in the terahertz band, combine the particle spectrum distribution, simulate the functional relationship between the terahertz radar equivalent reflectivity factor and cloud physical characteristic parameters, and establish a forward physical model. The cloud physical characteristic parameters include cloud particle size, cloud particle number concentration and distribution parameters;

[0043] S2. Preprocess the measured data and microphysical parameter products of the terahertz cloud radar to obtain a training database. The input data of the training database is the measured terahertz radar equivalent reflectivity factor, which can be expressed as:

[0044] y=[Z(z1)…Z(z n )] T

[0045] Among them, z i (i=1,2,…,n) represents the height of range bin i, n is the number of range bins detected by the radar, z1 is the cloud base, z n For Genting.

[0046] The output data is cloud physical characteristic parameters, which can be expressed as:

[0047] x=[lgD g (z1)…lgD g (z n )lgN0(z1)…lgN0(z n )α(z1)…α(z n )] T

[0048] Here, lg represents the logarithm with base 10.

[0049] S3. Establishing a convolutional neural network structure, using the forward physical model as a constraint, and adaptively adjusting the learning rate based on the deviation between the predicted value and the actual value of the radar echo to achieve accurate representation of the complex nonlinear function between cloud remote sensing characteristics and output cloud parameter results;

[0050] S4. Utilizing the training database, repeatedly training the convolutional neural network, and obtaining the cloud physical characteristic parameters by inverting and calculating the terahertz radar equivalent reflectivity factor.

[0051] The explanation of convolutional neural network is as follows:

[0052] Convolutional neural networks (CNNs) are a type of feedforward neural network with a deep structure that incorporates convolutional computations. They are a representative algorithm for deep learning. Convolutional neural networks possess the ability to learn representations and perform translation-invariant classification of input information based on their hierarchical structure, hence the name "translation-invariant artificial neural network."

[0053] In this embodiment, the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The data of the input layer is the input data. The convolutional neural network extracts features from the input data through the convolutional layer and the pooling layer to obtain local features of the input layer, and then generates the cloud physical feature parameters through the fully connected layer and enters the output layer.

[0054] Among them, the input layer and output layer are used to input and output data and can process multi-dimensional data; the function of the convolution layer is to extract features of the input data. It contains multiple convolution kernels, and each element of the convolution kernel corresponds to a weight coefficient and a bias, which is similar to a neuron in a feedforward neural network; after feature extraction in the convolution layer, the output feature map will be passed to the pooling layer for feature selection and information filtering; the fully connected layer in the convolutional neural network is equivalent to the hidden layer in the traditional feedforward neural network. The fully connected layer is located at the last part of the convolutional neural network and only transmits signals to other fully connected layers. The feature map will lose its spatial topological structure in the fully connected layer, and will be expanded into a vector and passed through the activation function.

[0055] In this embodiment, see Figure 2 , the calculation of the convolutional neural network includes the following steps:

[0056] S3.1, forward calculation: a single training sample is input into the convolutional neural network, and the value of each neuron is calculated step by step from front to back to obtain the prediction result x of the training sample out ;

[0057] S3.2, reverse calculation: first establish the prediction result x out The loss function E between the real inversion result x and the original inversion result x adopts square error. The loss function E of the output layer is calculated as follows:

[0058]

[0059] Where n is the total number of distance bins of radar detection profiles. Then, the derivative of the network parameters with respect to the loss function E is calculated layer by layer, i.e., the local gradient δ. The gradient descent method is used to optimize w and b of each layer, and the network parameters are iteratively updated. The iterative operation is performed until convergence. The calculation formula is as follows:

[0060] w l+1 =w l -Δw l

[0061]

[0062] b l+1 =b l -Δb l

[0063]

[0064] Where η represents the learning rate and l represents the number of cycles during training;

[0065] S3.3, using the adaptive learning rate scheme, calculate the inversion result x based on the actual prediction out , according to the forward physical model, calculate the predicted value Z of the radar echo sim , calculate the predicted value Z sim The deviation from the measured value Z is used to correct the learning rate η, ensuring that the learning rate η gradually decreases as the accuracy increases. The calculation formula is as follows:

[0066]

[0067] Among them, η0 represents the initial value of the learning rate η.

[0068] In this embodiment, see Figure 3 , when the convolutional neural network is calculated, the actions of each layer are as follows:

[0069] The input layer passes the input data to the convolutional layer;

[0070] The convolution layer sets three convolution kernels, calculates the pre-activation value, and obtains the first feature map through the activation function f. The element value C of the first feature map outk (k=1,2,…,n+2) is:

[0071]

[0072] Wherein, w and b are the multiplicative and additive biases corresponding to each first feature map, f represents the activation function, and y represents the input data;

[0073] The pooling layer retains useful information while reducing the size of the first feature map input by the convolution layer. The first feature map is downsampled by using the average value aggregation method of adjacent matrix regions to complete the screening and dimension reduction of feature parameters to obtain a second feature map. The element value P of the second feature map is outk (k=1,2,…,n+2) is:

[0074] P outk =down k (C outk );

[0075] In the fully connected layer, the obtained second feature maps are sequentially connected into vectors and input into the fully connected network. After weighted and biased operations, the activation function f is used to obtain the prediction results of the cloud physical feature parameters and send them to the output layer.

[0076] In this embodiment, the activation function f is a ReLU function.

[0077] Preferably, the predicted value Z of the radar echo sim The calculation of depends on the forward physical model. The specific steps are: calculating the complex refractive index of ice crystal particles in the terahertz band; using the discrete dipole approximation method to calculate the scattering characteristics of cloud particles in the terahertz band to obtain the backscattering cross section σ bk Table; through the backscattering cross section σ bk The table is combined with the particle spectrum distribution to obtain the functional relationship between the terahertz radar equivalent reflectivity factor and the cloud physical characteristic parameters.

[0078] In this embodiment, the activation function f is a ReLU function.

[0079] In this embodiment, the calculation method of the terahertz radar equivalent reflectivity factor is as described in S1. The specific steps are as follows:

[0080] Based on the survey results of cloud droplet spectrum characteristics, the size range of cloud particles was set to 2-2000μm, the complex refractive index of typical ice crystal particles in the terahertz band was calculated, and the discrete dipole approximation method was used to calculate the scattering characteristics of cloud particles in the terahertz band to obtain a backscattering cross-section table; based on the backscattering cross-section table and combined with the particle spectrum distribution, the functional relationship between the terahertz radar equivalent reflectivity factor and the cloud physical characteristic parameters was obtained.

[0081] The cloud particle number concentration is the number of cloud droplets per unit volume. Its variation with scale is called the cloud particle spectrum. Assuming that the cloud particle spectrum satisfies the gamma distribution, the formula is as follows:

[0082] N(D)=N0e α (D / D g ) α exp[-αD / D g ]

[0083] Where: N0 is the cloud particle number concentration; D is the cloud particle diameter; D g is the median length; α is the gamma distribution parameter. The definition of the terahertz cloud radar reflectivity factor is as follows:

[0084]

[0085] Where: σ bk The backscattering cross section of a spherical particle with a particle diameter D is calculated from the negative refractive index m of the particle, K = (m 2 -1) / (m 2 +2).

[0086] In this way, the inversion of cloud physical parameters based on convolutional neural networks can be realized. According to the radar reflectivity factor of the terahertz cloud radar, the particle size, particle number concentration and distribution parameters, that is, the cloud particle spectrum distribution characteristic parameters, can be obtained after inputting the network.

[0087] The advantages of the present invention are that it improves the inversion accuracy of cloud physical parameters and gets rid of the dependence on empirical values. It is an algorithm with stronger generalization and adaptive learning capabilities.

[0088] Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description is not intended to limit the present invention. After reading the above description, various modifications and substitutions of the present invention will become apparent to those skilled in the art. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A cloud physical parameter inversion method based on convolutional neural network, characterized in that: The steps include: S1. Calculate the scattering characteristics of cloud particles in the terahertz band, combine the particle spectrum distribution, simulate the functional relationship between the terahertz radar equivalent reflectivity factor and cloud physical characteristic parameters, and establish a forward physical model. The cloud physical characteristic parameters include cloud particle size, cloud particle number concentration and distribution parameters; S2. Preprocessing the measured data and microphysical parameter products of the terahertz cloud radar to obtain a training database, wherein the input data of the training database is the measured terahertz radar equivalent reflectivity factor, and the output data is the cloud physical characteristic parameters; S3. Establishing the structure of a convolutional neural network, using the forward physical model as a constraint condition, and adaptively adjusting the learning rate based on the deviation between the predicted value and the actual value of the radar echo to achieve accurate representation of the complex nonlinear function between cloud remote sensing characteristics and output cloud parameter results; the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The data of the input layer is the vertical profile of the terahertz cloud radar reflectivity factor. The convolutional neural network extracts features from the input data through the convolutional layer and the pooling layer to obtain local features of the input layer, and then generates the cloud physical feature parameters through the fully connected layer, which enter the output layer; S4. Utilizing the training database, repeatedly training the convolutional neural network, and obtaining the cloud physical characteristic parameters by inverting and calculating the terahertz radar equivalent reflectivity factor.

2. The cloud physical parameter inversion method based on convolutional neural network according to claim 1, characterized in that: The calculation of the convolutional neural network specifically includes the following steps: S3.1, forward calculation: a single training sample is input into the convolutional neural network, and the value of each neuron is calculated step by step from front to back to obtain the prediction result x of the training sample out ; S3.2, reverse calculation: first establish the prediction result x out The loss function E between the real inversion result x and the original inversion result x adopts square error. The loss function E of the output layer is calculated as follows: Where n is the total number of distance bins of radar detection profiles. Then, the derivative of the network parameters with respect to the loss function E is calculated layer by layer, i.e., the local gradient δ. The gradient descent method is used to optimize w and b of each layer, and the network parameters are iteratively updated. The iterative operation is performed until convergence. The calculation formula is as follows: In l+1 =in l -Δw l b l+1 =b l -Δb l Where η represents the learning rate and l represents the number of cycles during training; S3.3, using the adaptive learning rate scheme, calculate the inversion result x based on the actual prediction out , according to the forward physical model, calculate the predicted value Z of the radar echo sim , calculate the predicted value Z sim The deviation from the measured value Z is used to correct the learning rate η, ensuring that the learning rate η gradually decreases as the accuracy increases. The calculation formula is as follows: Among them, η0 represents the initial value of the learning rate η.

3. The cloud physical parameter inversion method based on convolutional neural network according to claim 2, characterized in that: When the convolutional neural network is calculated, the actions of each layer are as follows: The input layer passes the input data to the convolutional layer; The convolution layer sets three convolution kernels, calculates the pre-activation value, and obtains the first feature map through the activation function f. The element value C of the first feature map outk (k=1,2,…,n+2) is: Wherein, w and b are the multiplicative and additive biases corresponding to each first feature map, f represents the activation function, and y represents the input data; The pooling layer retains useful information while reducing the size of the first feature map input by the convolution layer. The first feature map is downsampled by using the average value aggregation method of adjacent matrix regions to complete the screening and dimension reduction of feature parameters to obtain a second feature map. The element value P of the second feature map is outk (k=1,2,…,n+2) is: P outk =down k (C outk ); In the fully connected layer, the obtained second feature maps are sequentially connected into vectors and input into the fully connected network. After weighted and biased operations, the activation function f is used to obtain the prediction results of the cloud physical feature parameters and send them to the output layer.

4. The cloud physical parameter inversion method based on convolutional neural network according to claim 3, characterized in that: The activation function f is the ReLU function.

5. The cloud physical parameter inversion method based on convolutional neural network according to claim 3, characterized in that: The predicted value Z of the radar echo sim The calculation of depends on the forward physical model. The specific steps are: calculating the complex refractive index of ice crystal particles in the terahertz band; using the discrete dipole approximation method to calculate the scattering characteristics of cloud particles in the terahertz band to obtain the backscattering cross section σ bk Table; through the backscattering cross section σ bk The table is combined with the particle spectrum distribution to obtain the functional relationship between the terahertz radar equivalent reflectivity factor and the cloud physical characteristic parameters.

6. The cloud physical parameter inversion method based on convolutional neural network according to claim 1, characterized in that: The preprocessing of the measured data and microphysical parameter products of the terahertz cloud radar includes: Exclude rainfall cases and remove data with significant influence from ground clutter.

Images