A deep learning-based method for calculating radiative transfer in three-dimensional clouds
Through deep learning-based methods, the cloud-containing scene attributes in the radiation transmission equation are combined, targeted convolution kernels and strategies are established, and the speed and accuracy in the calculation of three-dimensional cloud radiation transmission is solved using sub-pixel convolution module and logarithmic method. The problem of difficulty in unification has achieved efficient three-dimensional cloud monochromatic radiation calculation.
Patent Information
- Application Number
- CN202310493703.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-26
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2043-04-26
AI Technical Summary
The existing three-dimensional cloud radiation transmission calculation methods are difficult to meet the requirements of high accuracy and high speed at the same time, especially in complex scenarios containing cloud atmospheres, the calculation speed and accuracy are difficult to unify.
Using a deep learning-based method, targeted convolution kernels and strategies are established by combining the cloud-containing scene properties in the radiation transmission equation, sub-pixel convolution modules are introduced for pooling and upsampling operations, and logarithmic method is used to optimize. Loss function, calculates the luminance of the radiant layer of three-dimensional clouds.
It effectively improves the speed and accuracy of the monochrome radiation transmission calculation of three-dimensional clouds, and overcomes the difficulties in calculation speed and accuracy of conventional algorithms.
Smart Images
Figure CN116992751B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of three-dimensional cloud-containing remote sensing radiation calculation and machine learning, and specifically relates to a three-dimensional cloud layer radiation transmission calculation method based on deep learning. Background Art
[0002] With the development of atmospheric remote sensing science, the radiation transfer model of the cloudy atmosphere has become more important and is widely used in research fields such as climate change, environmental monitoring, and weather forecasting. Since clouds have strong non-uniformity in three-dimensional space and multiple scattering calculations are relatively complex, the accurate calculation of cloud radiation characteristics is very complicated. How to efficiently and accurately calculate the scattering in the three-dimensional cloud area is a difficult point in the radiation transfer calculation.
[0003] In recent years, machine learning has been applied to the field of physical simulation, greatly improving the speed of physical simulation. Machine learning can actively find the associations and features between data from a large amount of historical data or simulation data, and update the internal parameters of the machine learning algorithm through a large amount of training process to achieve machine self-learning.
[0004] The existing three-dimensional cloud radiation transfer methods are difficult to apply in the remote sensing field where high computational accuracy and speed are required. It is necessary to propose key technologies that are conducive to improving computational speed and accuracy based on the scene properties of the cloud-containing atmosphere. Summary of the invention
[0005] In order to overcome the problem that conventional three-dimensional cloud radiation transfer calculation methods are difficult to meet the requirements of both accuracy and speed, the present invention proposes a three-dimensional cloud radiation transfer calculation method based on deep learning.
[0006] The present invention proposes a three-dimensional cloud layer radiation transfer calculation method based on deep learning. The method systematically combines and arranges cloud-containing atmospheric scene attributes according to the radiation transfer equation, establishes targeted convolution kernels and strategies to represent the energy exchange between adjacent cloud units, introduces a sub-pixel convolution module for optimizing pooling and upsampling operations, and simultaneously creates a loss function combined with a logarithmic method to obtain an image of the logarithmic value of cloud layer radiation brightness. The specific process of the three-dimensional cloud layer radiation transfer calculation method is as follows:
[0007] Step 1: Combine solar radiation flux, path length, asymmetry factor, temperature, extinction coefficient and single scattering albedo into an array of cloud-containing atmospheric scene attributes, and arrange the cloud-containing atmospheric scene attributes according to the cloud unit position as the network input layer; Step 2: Establish targeted convolution kernels and convolution strategies, use three-dimensional convolution kernels to perform convolution operations on the cloud-containing atmospheric scene attributes in the input layer, calculate the radiation energy exchange between adjacent cloud units, and obtain new cloud scene attributes;
[0008] Step 3: Perform a pooling operation on the updated cloud scene attributes to obtain the cloud scene attributes after the cloud unit is merged, and use sub-pixel convolution to upsample the cloud scene attributes after the cloud unit is merged to obtain the cloud scene attributes after the cloud unit is re-split.
[0009] Step 4: Extract the radiation brightness information layer by layer from the cloud scene attributes after the cloud unit is re-divided, and finally obtain the logarithmic value image of the uplink radiation brightness of the three-dimensional cloud field.
[0010] Furthermore, the step 1 is specifically as follows:
[0011] According to the radiation transfer formulas (1) to (5), the parameters required for radiation transfer calculation include solar radiation flux, path length, asymmetry factor, temperature, extinction coefficient and single scattering albedo. The above parameters are arranged in a fixed way to construct an array H(x, y, z). The array H(x, y, z) is sorted according to the spatial position of the cloud unit to form a cloud scene attribute array, and the input layer that meets the network training requirements is obtained.
[0012]
[0013] Where: s is the position vector; Ω is the direction vector; k e (s) is the extinction coefficient; I(s,Ω) represents the radiation brightness and radiation source function along the Ω direction at position s, J E (s,Ω),J SS (s,Ω) and J MS (s,Ω) are the heat source function, single scattering source function and multiple scattering source function respectively.
[0014] The expression of the source function is as follows:
[0015] J E (s,Ω)=[1-ω(s)]B[T(s)] (2)
[0016]
[0017]
[0018] Where ω(s) is the single scattering albedo, F TOA is the solar radiation flux density reaching the top of the atmosphere, P(s; Ω, Ω′0) is the phase function of the radiation scattered from the Ω′0 direction to the Ω direction at s, I(s, Ω′) is the incident radiation intensity along the Ω′ direction at s, and B[T(s)] is the Planck function, which is a function of the atmospheric temperature T.
[0019] The expression of the phase function is as follows:
[0020]
[0021] Where g(s) is the asymmetry factor.
[0022] Furthermore, the step 2 is specifically as follows:
[0023] The size of the three-dimensional convolution kernel is set to three times the size of the cloud unit, so that the receptive field of the convolution kernel can sense the relevant information of the neighborhood. At the same time, the convolution step size corresponds to the size of the attribute array H(x, y, z), and the convolution operation is performed on the input cloud-containing atmospheric attributes. When the cloud scene represented by the current layer of the neural network has only one layer, only the radiation transmission process between horizontal domains needs to be considered, and the three-dimensional convolution kernel degenerates into a two-dimensional convolution kernel. Considering that the exchange between the cloud layer and the external radiation is different under different boundary conditions, when the boundary condition is a periodic boundary, cyclic filling is used to correspond to it, and when the boundary condition is an open boundary, zero filling is used to correspond to it.
[0024] Furthermore, the step 3 is specifically as follows:
[0025] The convolution operation with a stride of an integer multiple of the cloud unit size is used to achieve the pooling effect; sub-pixel convolution is used for upsampling. The sub-pixel convolution process is to first perform two convolutions with a stride of 1×1×1 on the attribute array of length×width×height c×h×d. Then perform a convolution with a stride of 1×1×1 to generate a channel number r 3 The intermediate calculation result is then length × width × height × number of channels is c × h × d × r 3 The intermediate calculation results are reorganized into up-sampled results of rc×rh×rd.
[0026] Furthermore, the step 4 is specifically as follows:
[0027] The radiation brightness information is extracted layer by layer from the cloud scene attributes after the cloud unit is re-split, and finally the calculation result of the upward radiation brightness of the three-dimensional cloud field is obtained. The radiation brightness of the three-dimensional cloud layer is scaled using the log method to reduce the relative error of the neural network in calculating the upward radiation brightness of the three-dimensional cloud layer. At the same time, regularization is added to the loss function to prevent overfitting problems. The loss function expression is as follows:
[0028]
[0029] In the formula, Loss represents the loss function, I j represents the true value of the radiation brightness of the jth cloud layer, I j ′ represents the cloud layer radiation brightness calculated by the jth network, λ is the regularization strength, and w i is the i-th model parameter, n is the number of model parameters, and N is the number of cloud layer radiation brightness calculation results.
[0030] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:
[0031] The present invention can be applied to the calculation of monochromatic radiation transmission of three-dimensional clouds, which overcomes the problem that conventional algorithms are difficult to unify in terms of calculation speed and accuracy, and effectively improves the calculation speed of monochromatic radiation transmission of three-dimensional clouds. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a schematic diagram of the overall process of the three-dimensional cloud layer radiation transmission calculation method of the present invention;
[0033] Figure 2 It is a schematic diagram of the input layer of the three-dimensional cloud layer radiation transmission calculation method of the present invention;
[0034] Figure 3 It is a schematic diagram of a three-dimensional convolution kernel of a three-dimensional cloud layer radiation transmission calculation method of the present invention;
[0035] Figure 4 It is a schematic diagram of the sub-pixel convolution process of the three-dimensional cloud layer radiation transmission calculation method of the present invention;
[0036] Figure 5 This is an example diagram of the calculation results of the three-dimensional cloud layer radiation transmission calculation method of the present invention;
[0037] Figure 6 It is a radiation brightness distribution histogram of the calculation results of the three-dimensional cloud layer radiation transmission calculation method of the present invention. DETAILED DESCRIPTION
[0038] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings:
[0039] The present invention proposes a method for calculating the radiation transfer of three-dimensional clouds based on deep learning. The method systematically combines and arranges the attributes of cloud-containing atmospheric scenes according to the radiation transfer equation, establishes targeted convolution kernels and strategies to represent the energy exchange between adjacent cloud units, introduces a sub-pixel convolution module for the optimization of pooling and upsampling operations, and simultaneously creates a loss function combined with a logarithmic method to obtain an image of the logarithmic value of the cloud radiation brightness. The present invention can quickly and accurately calculate the monochromatic radiation of three-dimensional clouds, and effectively improve the calculation accuracy of the monochromatic radiation of three-dimensional clouds while meeting the accuracy requirements.
[0040] The three-dimensional cloud layer radiation transfer calculation method based on deep learning includes the following steps:
[0041] Step 1: According to the radiation transfer formulas (1) to (5), the parameters required for radiation transfer calculation include solar radiation flux, path length, asymmetry factor, temperature, extinction coefficient and single scattering albedo. The above parameters are arranged in a fixed manner to construct an array H(x, y, z). The array H(x, y, z) is sorted according to the spatial position of the cloud unit to form a cloud scene attribute array, thus obtaining an input layer that meets the network training requirements.
[0042]
[0043] Where: s is the position vector; Ω is the direction vector; k e (s) is the extinction coefficient; I(s,Ω) represents the radiation brightness and radiation source function along the Ω direction at position s, J E (s,Ω),J SS (s,Ω) and J MS (s,Ω) are the heat source function, single scattering source function and multiple scattering source function respectively.
[0044] The expression of the source function is as follows:
[0045] J E (s,Ω)=[1-ω(s)]B[T(s)] (2)
[0046]
[0047]
[0048] Where ω(s) is the single scattering albedo, F TOA is the solar radiation flux density reaching the top of the atmosphere, P(s; Ω, Ω′0) is the phase function of the radiation scattered from the Ω′0 direction to the Ω direction at s, I(s, Ω′) is the incident radiation intensity along the Ω′ direction at s, and B[T(s)] is the Planck function, which is a function of the atmospheric temperature T.
[0049] The expression of the phase function is as follows:
[0050]
[0051] Where g(s) is the asymmetry factor.
[0052] Step 2: Set the size of the three-dimensional convolution kernel to three times the size of the cloud unit, so that the receptive field of the convolution kernel can sense the relevant information of the neighborhood. At the same time, the convolution step size corresponds to the size of the attribute array H(x, y, z), and the input cloud-containing atmospheric attributes are convolved. When the cloud scene represented by the current layer of the neural network has only one layer, only the radiation transmission process between horizontal domains needs to be considered, and the three-dimensional convolution kernel degenerates into a two-dimensional convolution kernel. Considering that the exchange between the cloud layer and the external radiation is different under different boundary conditions, when the boundary condition is a periodic boundary, cyclic filling is used to correspond to it, and when the boundary condition is an open boundary, zero filling is used to correspond to it.
[0053] Step 3: Use convolution operation with a stride of integer multiples of the cloud unit size to achieve pooling effect; use sub-pixel convolution for upsampling operation. The process of sub-pixel convolution is to first perform two convolutions with a stride of 1×1×1 on the attribute array of length×width×height c×h×d. Then perform a convolution with a stride of 1×1×1 to generate a channel number r 3 The intermediate calculation result is then length × width × height × number of channels is c × h × d × r 3 The intermediate calculation results are reorganized into up-sampled results of rc×rh×rd.
[0054] Step 4: Extract the radiation brightness information layer by layer from the cloud scene attributes after the cloud unit is re-split, and finally obtain the calculation result of the upward radiation brightness of the three-dimensional cloud field. Use the log method to scale the radiation brightness of the three-dimensional cloud layer to reduce the relative error of the neural network in calculating the upward radiation brightness of the three-dimensional cloud layer. At the same time, add regularization to the loss function to prevent overfitting problems. The loss function expression is as follows:
[0055]
[0056] In the formula, Loss represents the loss function, I j represents the true value of the radiation brightness of the jth cloud layer, I j ′ represents the cloud layer radiation brightness calculated by the jth network, λ is the regularization strength, and w i is the i-th model parameter, n is the number of model parameters, and N is the number of cloud layer radiation brightness calculation results.
[0057] A specific example of three-dimensional cloud radiation transfer calculation is shown below:
[0058] Application step 1: According to the radiation transfer formulas (1) to (5), the parameters required for radiation transfer calculation include solar radiation flux, path length, asymmetry factor, temperature, extinction coefficient and single scattering albedo. The above parameters are arranged in a fixed manner to construct an array H(x, y, z). The array H(x, y, z) is sorted according to the spatial position of the cloud unit to form a cloud scene attribute array, and the input layer that meets the network training requirements is obtained.
[0059]
[0060] Where: s is the position vector; Ω is the direction vector; k e (s) is the extinction coefficient; I(s,Ω) represents the radiation brightness and radiation source function along the Ω direction at position s, J E (s,Ω),J SS (s,Ω) and J MS (s,Ω) are the heat source function, single scattering source function and multiple scattering source function respectively.
[0061] The expression of the source function is as follows:
[0062] J E (s,Ω)=[1-ω(s)]B[T(s)] (2)
[0063]
[0064]
[0065] Where ω(s) is the single scattering albedo, F TOA is the solar radiation flux density reaching the top of the atmosphere, P(s; Ω, Ω′0) is the phase function of the radiation scattered from the Ω′0 direction to the Ω direction at s, I(s, Ω′) is the incident radiation intensity along the Ω′ direction at s, and B[T(s)] is the Planck function, which is a function of the atmospheric temperature T.
[0066] The expression of the phase function is as follows:
[0067]
[0068] Where g(s) is the asymmetry factor.
[0069] Application step 2: Set the size of the three-dimensional convolution kernel to three times the size of the cloud unit, so that the receptive field of the convolution kernel can sense the relevant information of the neighborhood. At the same time, the convolution step size corresponds to the size of the attribute array H(x, y, z), and the input cloud-containing atmospheric attributes are convolved. When the cloud scene represented by the current layer of the neural network has only one layer, only the radiation transmission process between horizontal domains needs to be considered, and the three-dimensional convolution kernel degenerates into a two-dimensional convolution kernel. Considering that the exchange between the cloud layer and the external radiation is different under different boundary conditions, when the boundary condition is a periodic boundary, cyclic filling is used to correspond to it, and when the boundary condition is an open boundary, zero filling is used to correspond to it.
[0070] Application step 3: Use convolution operation with a stride of integer multiples of the cloud unit size to achieve pooling effect; use sub-pixel convolution for upsampling operation. The sub-pixel convolution process is to first perform two convolutions with a stride of 1×1×1 on the attribute array of length×width×height c×h×d. Then perform a convolution with a stride of 1×1×1 to generate a channel number r 3 The intermediate calculation result is then length × width × height × number of channels is c × h × d × r 3 The intermediate calculation results are reorganized into up-sampled results of rc×rh×rd.
[0071] Application step 4: Extract the radiation brightness information layer by layer from the cloud scene attributes after the cloud unit is re-split, and finally obtain the calculation result of the upward radiation brightness of the three-dimensional cloud field. Use the log method to scale the radiation brightness of the three-dimensional cloud layer to reduce the relative error of the neural network in calculating the upward radiation brightness of the three-dimensional cloud layer. At the same time, add regularization to the loss function to prevent overfitting problems. The loss function expression is as follows:
[0072]
[0073] In the formula, Loss represents the loss function, I j represents the true value of the radiation brightness of the jth cloud layer, I j ′ represents the cloud layer radiation brightness calculated by the jth network, λ is the regularization strength, and w i is the i-th model parameter, n is the number of model parameters, and N is the number of cloud layer radiation brightness calculation results.
[0074] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A three-dimensional cloud radiation transfer calculation method based on deep learning, characterized in that: The method systematically combines and arranges cloud-containing atmospheric scene attributes according to the radiation transfer equation, establishes targeted convolution kernels and strategies to represent the energy exchange between adjacent cloud units, introduces a sub-pixel convolution module for the optimization of pooling and upsampling operations, creates a loss function based on the logarithmic method, and obtains an image of the logarithmic value of cloud layer radiation brightness; The specific process of the three-dimensional cloud layer radiation transfer calculation method is as follows: Step 1: Create a cloud atmosphere scene attribute array and arrange the cloud atmosphere scene attributes according to the cloud unit position as the network input layer; Step 2: Establish targeted convolution kernels and convolution strategies, use three-dimensional convolution kernels to perform convolution operations on the cloud-containing atmospheric scene attributes, calculate the radiation energy exchange between adjacent cloud units, and obtain updated cloud scene attributes; Step 3: Perform a pooling operation on the updated cloud scene attributes to obtain the cloud scene attributes after the cloud unit is merged, and use sub-pixel convolution to upsample the cloud scene attributes after the cloud unit is merged to obtain the cloud scene attributes after the cloud unit is split; Step 4: Extract the radiation brightness information layer by layer from the cloud scene attributes that have been re-divided by cloud units, and finally obtain a three-dimensional cloud field uplink radiation brightness logarithmic value image; The step 1 is specifically as follows: The parameters required for the radiation transfer calculation are set, including solar radiation flux, path length, asymmetry factor, temperature, extinction coefficient and single scattering albedo. According to the radiation transfer formulas (1) to (5), the above parameters are arranged in a fixed manner to construct an array H(x, y, z). The array H(x, y, z) is sorted according to the spatial position of the cloud unit to form a cloud scene attribute array, and an input layer that meets the network training requirements is obtained. Where: s is the position vector; Ω is the direction vector; k e (s) is the extinction coefficient; I(s,Ω) represents the radiation brightness and radiation source function along the Ω direction at position s, J E (s,Ω),J SS (s,Ω) and J MS (s,Ω) are heat source function, single scattering source function and multiple scattering source function respectively; The expressions of heat source function, single scattering source function and multiple scattering source function are as follows: J E (s,Ω)=[1-ω(s)]B[T(s)] (2) Where ω(s) is the single scattering albedo, F TOA is the solar radiation flux density reaching the top of the atmosphere, P(s;Ω,Ω′0) is the phase function of scattering from the Ω′0 direction to the Ω direction at s, I(s,Ω′) is the incident radiation intensity along the Ω′ direction at s, and B[T(s)] is the Planck function, which is a function of the atmospheric temperature T; The expression of the phase function is as follows: Where g(s) is the asymmetry factor.
2. A three-dimensional cloud layer radiation transfer calculation method based on deep learning according to claim 1, characterized in that: The step 2 is specifically as follows: The size of the three-dimensional convolution kernel is set to three times the size of the cloud unit, so that the receptive field of the convolution kernel can sense the relevant information of the neighborhood. At the same time, the convolution step size corresponds to the size of the attribute array H(x, y, z), and the convolution operation is performed on the input cloud-containing atmospheric attributes. When the cloud scene represented by the current layer of the neural network has only one layer, only the radiation transmission process between horizontal domains needs to be considered, and the three-dimensional convolution kernel degenerates into a two-dimensional convolution kernel. Considering that the exchange between the cloud layer and the external radiation is different under different boundary conditions, when the boundary condition is a periodic boundary, cyclic filling is used to correspond to it, and when the boundary condition is an open boundary, zero filling is used to correspond to it.
3. The method for calculating three-dimensional cloud layer radiation transfer based on deep learning according to claim 1, characterized in that: The step 3 is specifically as follows: The convolution operation with a stride of an integer multiple of the cloud unit size is used for pooling effect; sub-pixel convolution is used for upsampling operation. The process of sub-pixel convolution is to first perform two convolutions with a stride of 1×1×1 on the attribute array of length×width×height c×h×d; then perform a convolution with a stride of 1×1×1 to generate a channel number r 3 The intermediate calculation result is then length × width × height × number of channels is c × h × d × r 3 The intermediate calculation results are reorganized into up-sampled results of rc×rh×rd.
4. The method for calculating three-dimensional cloud radiation transfer based on deep learning according to claim 1, characterized in that: The step 4 is specifically as follows: The radiation brightness information is extracted layer by layer from the cloud scene attributes after the cloud unit is re-split, and finally the calculation result of the upward radiation brightness of the three-dimensional cloud field is obtained. The radiation brightness of the three-dimensional cloud layer is scaled using the log method to reduce the relative error of the network calculation of the upward radiation brightness of the three-dimensional cloud layer. Regularization is added to the loss function to prevent overfitting problems. The loss function expression is as follows: In the formula, Loss represents the loss function, I j represents the true value of the radiation brightness of the jth cloud layer, I j ′ represents the cloud layer radiation brightness calculated by the jth network, λ is the regularization strength, and w i is the i-th model parameter, n is the number of model parameters, and N is the number of cloud layer radiation brightness calculation results.