Ice cloud scattering characteristics prediction model establishment method and prediction method based on deep reinforcement learning

Abstract

The present invention discloses a method for establishing a prediction model for ice cloud scattering characteristics based on deep reinforcement learning and a prediction method. The method belongs to the field of microwave remote sensing and detection and includes: establishing a first prediction model for predicting ice cloud scattering characteristic parameters based on ice cloud physical parameters (frequency, temperature, and ice cloud physical parameters); adding N network layers before the first initial prediction model to obtain a second prediction model for predicting ice cloud scattering characteristic parameters based on pressure, humidity, and ice cloud physical parameters, and embedding the model into a radiation transfer model to simulate brightness temperature; using the parameters of the second prediction model as the environmental state and minimizing the error between the simulated brightness temperature and the actual brightness temperature as the goal, using the corresponding data set to drive iterative training of deep reinforcement learning to update the parameters of the second prediction model. After the training is completed, the second prediction model in the radiation transfer model is output as the ice cloud scattering characteristic prediction model. The present invention can improve the prediction accuracy of ice cloud scattering characteristics.

Description

Technical Field

[0001] The present invention belongs to the field of microwave remote sensing and detection, and more specifically, to a method for establishing a prediction model for the scattering characteristics of ice clouds based on deep reinforcement learning and a prediction method. Background Art

[0002] Ice clouds have a significant impact on the Earth's energy cycle and water vapor cycle. Accurate and efficient calculation of the volume scattering characteristics of ice clouds is of great significance to research fields such as atmospheric radiation transfer models, climate and meteorological cloud models, and disaster weather forecasting.

[0003] Traditional methods for calculating the scattering properties of ice clouds first calculate the scattering properties of ice-phase particles in ice clouds based on physical parameters such as frequency, temperature, and ice cloud content, and then integrate them over an empirical distribution function to obtain the scattering properties. Ice-phase particles in ice clouds are primarily concentrated in the 20-600 micron range. Compared to microwaves, which can only detect larger particles, and infrared bands, which can only detect smaller particles, the terahertz band, whose wavelength range closely matches the size distribution of ice cloud particles, offers unique advantages for spaceborne terahertz radiometers using passive remote sensing. Ice clouds are composed of complex ice crystals with a wide variety of morphologies. Early radiation transfer calculations approximated ice cloud particles as spherical. With increasing research, several methods for calculating the scattering of non-spherical particles have emerged, such as the commonly used discrete dipole approximation, the finite time-domain difference method, and the T-matrix method.

[0004] Current radiative transfer models simplify complex and variable ice-phase particles into a single non-spherical shape and integrate the scattering characteristics of ice clouds over empirical distribution functions. This simplistic, one-size-fits-all approach can lead to simulation errors in radiative transfer over ice cloud regions. Furthermore, this method of first calculating the scattering characteristics of a single particle and then integrating them is computationally intensive and inefficient.

[0005] With the continuous development of deep learning technology, researchers have proposed a method for predicting the scattering parameters of ice clouds using deep learning models. This method eliminates the need for complex operations such as integration in actual predictions, greatly improving computational efficiency. However, this method requires training the model using a large number of ice cloud scattering parameters. Since ice cloud scattering parameters cannot be directly measured and can only be obtained indirectly through calculation, the ice cloud scattering parameters calculated using traditional calculation methods are inherently inaccurate, which limits the prediction accuracy of the trained model. Summary of the Invention

[0006] In response to the defects of the existing technology and the need for improvement, the present invention provides a method for establishing a prediction model for the scattering characteristics of ice clouds and a prediction method based on deep reinforcement learning, the purpose of which is to improve the prediction accuracy of the scattering characteristics of ice clouds.

[0007] To achieve the above objectives, according to one aspect of the present invention, a method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning is provided, comprising:

[0008] Step S1: constructing a neural network model and training the neural network model using a first data set consisting of ice cloud physical parameters and corresponding volume scattering parameters to obtain a first prediction model for predicting ice cloud volume scattering characteristic parameters based on the ice cloud physical parameters; the ice cloud physical parameters include frequency, temperature, and ice cloud content;

[0009] Step S2: Add N network layers before the input layer of the first prediction model to obtain a second prediction model, and embed the second prediction model into the radiation transfer model; the second prediction model uses pressure, humidity, and ice cloud physical parameters as input to predict ice cloud scattering characteristic parameters; N is a positive integer;

[0010] Step S3: After establishing a deep reinforcement learning training environment, using the parameters of the second prediction model in the radiative transfer model as the environmental state and minimizing the error between the simulated brightness temperature output by the radiative transfer model and the actual brightness temperature as the training objective, the second dataset is used to drive iterative deep reinforcement learning training to update the parameters of the second prediction model. After the training is completed, the second prediction model in the radiative transfer model is output as the ice cloud scattering characteristics prediction model.

[0011] Among them, the second data set consists of ice cloud physical parameters and corresponding pressure, humidity and real brightness temperature in the ice cloud sensitive frequency band under the target scenario; the greater the error between the simulated brightness temperature and the real brightness temperature, the smaller the environmental reward.

[0012] Further, in step S2, N=2;

[0013] Furthermore, in the two added network layers, the first network layer includes five nodes, one for inputting frequency, temperature, ice cloud content, humidity, and pressure, respectively. The output of the node for inputting frequency is directly connected to the node for inputting frequency in the first prediction model.

[0014] The second network layer includes three nodes, designated as the first internal node, the second internal node, and the third internal node. The outputs of these three nodes are connected to the three nodes in the first prediction model used for inputting frequency, temperature, and ice cloud content, respectively. The input of the first internal node is connected only to the output of the node used for inputting frequency in the first network layer. The input of the second internal node is connected to the output of all nodes in the first network layer. The input of the third internal node is connected to the output of all nodes in the first network layer. Furthermore, in step S3, the environmental state is the parameter of the network node newly added to the second prediction model relative to the first prediction model.

[0015] Furthermore, in step S1, the constructed neural network model is a BP neural network.

[0016] Furthermore, the target scene is a tropical cyclone scene.

[0017] Furthermore, the ice cloud sensitive frequency band is the 183GHz band.

[0018] According to another aspect of the present invention, a method for predicting scattering characteristics of an ice cloud is provided, comprising:

[0019] The physical parameters, humidity and pressure of the ice cloud are input into the ice cloud scattering characteristic prediction model established by the above-mentioned ice cloud scattering characteristic prediction model establishment method based on deep reinforcement learning provided by the present invention to obtain the ice cloud scattering characteristic parameters.

[0020] According to another aspect of the present invention, a computer program product is provided. When executed by a processor, the computer program product implements the above-mentioned method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided by the present invention, and / or the above-mentioned method for predicting ice cloud scattering properties provided by the present invention.

[0021] According to another aspect of the present invention, a computer-readable storage medium is provided, comprising a stored computer program. When the computer program is executed by a processor, the processor controls the device containing the computer-readable storage medium to execute the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided by the present invention, and / or the method for predicting ice cloud scattering properties provided by the present invention.

[0022] According to another aspect of the present invention, there is provided an electronic device, comprising:

[0023] a computer-readable storage medium for storing a computer program;

[0024] and a processor configured to read a computer program stored in a computer-readable storage medium and execute the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided by the present invention, and / or the method for predicting ice cloud scattering properties provided by the present invention.

[0025] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0026] (1) The traditional calculation method of ice cloud body scattering characteristic parameters believes that the ice cloud body scattering characteristics are only related to the three parameters of frequency, temperature and content. However, actual research shows that the formation of ice clouds is also related to pressure and humidity. The present invention first uses the traditional calculation method to obtain a data set of frequency, temperature and content and the corresponding ice cloud body scattering characteristic parameters, and uses this data set to train a first prediction model that can predict the ice cloud body scattering characteristics based on frequency, temperature and content. Then, the model structure is modified so that its input also includes pressure and humidity in addition to frequency, temperature and content. The improved model is embedded in the radiation transfer model, thereby converting the prediction of body scattering characteristics to the prediction of brightness temperature, solving the problem that the traditional method cannot use pressure and humidity to calculate body scattering characteristics. On this basis, the prediction model structure parameters are used as the environmental state and the brightness temperature prediction error is used as the environmental reward. The prediction model parameters are optimized through the deep reinforcement learning iterative training method. On the one hand, the model can simultaneously consider the influence of frequency, temperature, content, pressure and humidity on ice clouds. On the other hand, the brightness temperature prediction result is used to correct the prediction error of the body scattering characteristic parameters, effectively improving the prediction accuracy of the trained prediction model.

[0027] (2) The present invention found that although frequency, temperature, content, pressure and humidity all affect the formation of ice clouds, the correlation between frequency and other parameters is not great. Based on this finding, in the preferred embodiment of the present invention, two network layers are specifically added before the input layer of the first model. The first network layer contains 5 nodes, which are used to input 5 parameters respectively. The second network layer includes 3 nodes (first internal node, second internal node and third internal node). The outputs of these 3 nodes are respectively connected to the three nodes for inputting frequency, temperature and ice cloud content in the first prediction model; the input of the first internal node is only connected to the output of the node for inputting frequency in the first network layer, the input of the second internal node is connected to the output of each node in the first network layer, and the input of the third internal node is connected to the output of each node in the first network layer. In this way, the network structure can be simplified to the greatest extent while fully considering the relationship between parameters, thereby improving the efficiency of subsequent deep reinforcement learning iterative training.

[0028] (3) In the preferred embodiment of the present invention, in deep reinforcement learning, only the newly added network node parameters are used as the environment state, so that only the newly added network node parameters need to be updated during the iterative training process, thereby reducing the computational complexity in the iterative training and improving the efficiency of the iterative training. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of a method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided by an embodiment of the present invention;

[0030] Figure 2A first prediction model established for an embodiment of the present invention;

[0031] Figure 3 A schematic diagram of a second prediction model established in an embodiment of the present invention;

[0032] Figure 4 are the scattering characteristic parameters of ice cloud calculated by traditional methods; where (a) is the absorption coefficient, (b) is the scattering coefficient, and (c) is the asymmetry factor;

[0033] Figure 5 ] are the scattering characteristic parameters of the ice cloud predicted by the BP neural network trained using an embodiment of the present invention; where (a) is the absorption coefficient, (b) is the scattering coefficient, and (c) is the asymmetry factor;

[0034] Figure 6 The relative errors between the BP neural network trained by the embodiment of the present invention and the volume scattering characteristics predicted by the traditional method; where (a) is the relative error of the absorption coefficient, (b) is the relative error of the scattering coefficient, and (c) is the relative error of the asymmetry factor;

[0035] Figure 7 The relative errors of the ice cloud scattering characteristic parameters predicted by the BP neural network trained by the embodiment of the present invention at the two frequencies of 176.31 GHz and 190.31 GHz;

[0036] Figure 8 Comparison of the simulated brightness temperatures using different methods under the target scenario; (a) is the actual brightness temperature, (b) is the simulated brightness temperature based on the ice cloud body scattering property prediction model, and (c) is the simulated brightness temperature based on the traditional body scattering calculation method.

[0037] Figure 9 Figure 2 shows the simulated brightness temperature residuals of different methods under the target scene. (a) is the simulated brightness temperature residual based on the ice cloud volume scattering property prediction model, and (b) is the simulated brightness temperature residual based on the traditional volume scattering calculation method.

[0038] Figure 10 is the root mean square error of different methods in different partitions under the target scenario;

[0039] Figure 11 Comparison of the brightness temperature simulations using different methods under the HAISHEN scenario; (a) is the actual brightness temperature, (b) is the simulated brightness temperature based on the ice cloud body scattering property prediction model, and (c) is the simulated brightness temperature based on the traditional body scattering calculation method.

[0040] Figure 12Figure 2 shows the simulated brightness temperature residuals of different methods under the HAISHEN scenario. (a) is the simulated brightness temperature residual based on the ice cloud volume scattering property prediction model, and (b) is the simulated brightness temperature residual based on the traditional volume scattering calculation method.

[0041] Figure 13 is the root mean square error of different methods in different partitions under the HAISHEN scenario;

[0042] Figure 14 Comparison of the brightness temperature simulations using different methods under the MARIA scenario; (a) is the actual brightness temperature, (b) is the brightness temperature simulated based on the ice cloud body scattering property prediction model, and (c) is the brightness temperature simulated based on the traditional body scattering calculation method.

[0043] Figure 15 Figure 2 shows the simulated brightness temperature residuals of different methods under the MARIA scenario. (a) is the simulated brightness temperature residual based on the ice cloud volume scattering property prediction model, and (b) is the simulated brightness temperature residual based on the traditional volume scattering calculation method.

[0044] Figure 16 is the root mean square error of different methods in different partitions under the MARIA scenario;

[0045] Figure 17 Comparison of the simulated brightness temperatures using different methods under the Meranti scenario; (a) is the actual brightness temperature, (b) is the simulated brightness temperature based on the ice cloud body scattering property prediction model, and (c) is the simulated brightness temperature based on the traditional body scattering calculation method;

[0046] Figure 18 Figure 2 shows the simulated brightness temperature residuals of different methods under the Meranti scenario. (a) is the simulated brightness temperature residual based on the ice cloud volume scattering property prediction model, and (b) is the simulated brightness temperature residual based on the traditional volume scattering calculation method.

[0047] Figure 19 is the root mean square error of different methods in different partitions under the Meranti scenario;

[0048] Figure 20 The calculation time of different methods. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0050] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0051] In order to solve the technical problem of low calculation accuracy of existing methods for calculating the scattering characteristic parameters of ice clouds, the present invention provides a method for establishing an ice cloud body scattering characteristic prediction model and a prediction method based on deep reinforcement learning. The overall idea is to fully consider the influencing factors of ice cloud formation, construct a body scattering network to be optimized, and use the deep reinforcement learning DDPG algorithm to drive the radiation transfer model to calculate the simulated brightness temperature using the body scattering neural network to be optimized. The error between the simulated brightness temperature and the actual observed brightness temperature data is used as the training target of reinforcement learning, thereby obtaining an optimized ice cloud body scattering neural network.

[0052] The following are examples.

[0053] Example 1:

[0054] A method for establishing a prediction model for ice cloud scattering characteristics based on deep reinforcement learning, such as Figure 1 As shown, it includes: steps S1 to S3.

[0055] In this embodiment, step S1 includes: constructing a neural network model and training the neural network model using a first data set consisting of ice cloud physical parameters and corresponding volume scattering parameters to obtain a first prediction model for predicting ice cloud volume scattering characteristic parameters based on the ice cloud physical parameters; the ice cloud physical parameters include: frequency, temperature, and ice cloud content.

[0056] The first prediction model can be established based on any neural network. Optionally, in this embodiment, considering that the BP neural network has arbitrarily complex pattern classification capabilities and excellent multi-dimensional function mapping capabilities, the BP neural network is selected to establish the first prediction model.

[0057] In practical applications, the first data set can be established by using the traditional ice cloud scattering characteristic parameter calculation method. After using this data set to train the BP neural network, the first prediction model is obtained as follows: Figure 2 As shown in the figure, it includes an input layer, a hidden layer and an output layer. The input layer contains 3 nodes, which are used to input frequency, temperature and ice cloud content respectively. The hidden layer contains 128 nodes. The output layer includes 66 nodes, which are used to output volume scattering characteristic parameters such as absorption coefficient, scattering coefficient and scattering phase matrix.

[0058] It should be noted that the BP neural network is only an optional implementation of the present invention. In some other embodiments of the present invention, other neural networks such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM) may also be used.

[0059] In this embodiment, step S2 includes: adding N network layers before the input layer of the first prediction model to obtain a second prediction model, and embedding the second prediction model into the radiation transfer model; the second prediction model uses pressure, humidity, and ice cloud physical parameters as input to predict the scattering characteristic parameters of the ice cloud body; N is a positive integer.

[0060] This embodiment improves the structure of the first prediction model so that the model can simultaneously consider the effects of pressure, humidity, frequency, temperature and ice cloud content on ice cloud formation. Considering that the correlation between frequency and other parameters is not significant, in order to simplify the network structure, Figure 3 As shown, in this embodiment, specifically Figure 2 Two network layers are added before the input layer of the first prediction model shown, i.e., N = 2. Of the two added network layers, the first network layer includes five nodes for inputting frequency, temperature, ice cloud content, humidity, and pressure, respectively. The output of the node for inputting frequency is directly connected to the node for inputting frequency in the first prediction model.

[0061] The second network layer includes three nodes, which are respectively recorded as the first internal node, the second internal node, and the third internal node. The outputs of these three nodes are connected to the three nodes used for inputting frequency, temperature, and ice cloud content in the first prediction model, respectively; the input of the first internal node is only connected to the output of the node used for inputting frequency in the first network layer, the input of the second internal node is connected to the output of each node in the first network layer, and the input of the third internal node is connected to the output of each node in the first network layer.

[0062] Based on the second prediction model obtained by the above improvements, the network structure is simplified to the greatest extent while fully considering the parameter relationship, thereby improving the efficiency of subsequent deep reinforcement learning iterative training.

[0063] Due to limitations of traditional calculation methods, it's impossible to calculate ice cloud scattering parameters based on pressure, humidity, frequency, temperature, and ice cloud content, making it impossible to directly construct a corresponding dataset to optimize and update the improved model. This embodiment embeds the model to be optimized, namely the second prediction model, into the radiation transfer model. The prediction results of the second prediction model can be embedded into the brightness temperature simulation process. This allows the prediction accuracy of the second model to be evaluated by measuring the error between the simulated brightness temperature and the actual brightness temperature observed by a spaceborne terahertz radiometer.

[0064] The radiative transfer model, as an observation operator, links atmospheric physical parameters with satellite-observed emissivity data. The radiative transfer model includes a scattering module and an absorption module, respectively used to account for the multiple scattering of electromagnetic waves by suspended particles as they propagate through the air, and the absorption of electromagnetic waves by factors such as water vapor and oxygen. It's easy to understand that embedding the second prediction model within the radiative transfer model specifically refers to using the second prediction model as the scattering module within the radiative transfer model.

[0065] The DOTLRT (Discrete Ordinate Tangent Linear Radiative Transfer) radiation transfer model uses the discrete ordinate method and the cascade method to achieve a fast analytical solution to the microwave radiation transfer equation and the calculation of the Jacobian matrix. It is a commonly used radiation transfer model. Optionally, in step (S2) of this embodiment, the second prediction model is specifically embedded in the DOTLRT radiation transfer model. It should also be noted that the DOTLRT radiation transfer model is only an optional embodiment of the present invention. In some other embodiments of the present invention, other radiation transfer models such as Atmospheric Radiative Transfer Simulator (ARTS), Community Radiative Transfer Model (CRTM), and Radiative Transfer for TOVS (RTTOV) may also be used.

[0066] In this embodiment, step S3 includes: using the parameters of the second prediction model in the radiative transfer model as the environmental state and minimizing the error between the simulated brightness temperature output by the radiative transfer model and the actual brightness temperature as the training objective, establishing a deep reinforcement learning training environment, using the second dataset to drive iterative training of the deep reinforcement learning to update the parameters of the second prediction model. After the training is completed, the second prediction model in the radiative transfer model is output as the ice cloud scattering property prediction model;

[0067] The second data set consists of the physical parameters of ice clouds and the corresponding pressure, humidity, and true brightness temperature within the ice cloud-sensitive frequency band under the target scenario. Optionally, in this embodiment, the target scenario is a tropical cyclone scenario, in which the ice cloud content is relatively high, and more data can be observed for updating the model parameters. The selected ice cloud-sensitive frequency band is specifically 183 GHz.

[0068] In order to further improve the efficiency of deep reinforcement learning iterative training, in step S3 of this embodiment, the environment state is the parameters of the network nodes newly added to the second prediction model relative to the first prediction model, specifically including the weights and biases of the newly added network nodes.

[0069] The ice cloud scattering characteristic prediction model finally established in this embodiment has a simple structure. In practical applications, the prediction results of the ice cloud scattering characteristic parameters can be obtained by inputting frequency, temperature, ice cloud content, pressure and humidity, and the prediction is efficient.

[0070] In general, the ice cloud scattering characteristic parameter prediction network constructed by the present invention has the ability to accurately and efficiently predict the scattering characteristics of the ice cloud body. It can directly predict the scattering characteristics of the ice cloud body from the physical parameters of the ice cloud, and then apply it to the radiation transfer model for calculation.

[0071] Example 2:

[0072] A method for predicting scattering characteristics of ice clouds, comprising:

[0073] The physical parameters, humidity, and pressure of the ice cloud are input into the ice cloud scattering characteristic prediction model established by the ice cloud scattering characteristic prediction model establishment method based on deep reinforcement learning provided in the above embodiment 1 to obtain the ice cloud scattering characteristic parameters.

[0074] Example 3:

[0075] A computer program product, when executed by a processor, implements the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided in the above-mentioned embodiment 1, and / or the method for predicting ice cloud scattering properties provided in the above-mentioned embodiment 2.

[0076] Example 4:

[0077] A computer-readable storage medium includes a stored computer program. When the computer program is executed by a processor, the processor controls the device where the computer-readable storage medium is located to execute the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided in the above-mentioned embodiment 1, and / or the method for predicting ice cloud scattering properties provided in the above-mentioned embodiment 2.

[0078] Example 5:

[0079] An electronic device, comprising:

[0080] a computer-readable storage medium for storing a computer program;

[0081] and a processor configured to read a computer program stored in a computer-readable storage medium and execute the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning provided in the above-mentioned embodiment 1, and / or the method for predicting ice cloud scattering properties provided in the above-mentioned embodiment 2.

[0082] The following is a further analysis of the beneficial effects achieved by the present invention using specific prediction examples. A dataset was constructed using Typhoon Haishen in 2020, Typhoon Maria in 2018, and Typhoon Meranti in 2016. The ice cloud scattering property prediction model was implemented in Python, and the deep reinforcement learning algorithm environment was built in MATLAB.

[0083] The dataset used to train the BP neural network was constructed as follows: a range of ice cloud physical parameters was set, and a traditional method for calculating the scattering properties of non-spherical particles was used to generate the ice cloud scattering property dataset. The frequencies were 183.31-7 GHz and 183.31+7 GHz, the GWC range was 1e-6 to 0.5 g / m³, and the temperature range was 190-270 K. The ice cloud scattering properties were obtained by integrating the Block Column shape from the ARTS single-particle scattering database over the Field07 particle size distribution function.

[0084] The traditional calculation method was used as a comparison to predict the scattering characteristic parameters of ice cloud bodies using the traditional calculation method and the trained BP neural network. Figure 4 (a), (b), and (c) show the absorption coefficient, scattering coefficient, and asymmetric factor calculated by the traditional calculation method, respectively. The asymmetric factor is used to measure the prediction accuracy of the scattering phase matrix. Figure 5 (a), (b), and (c) are the absorption coefficient, scattering coefficient, and asymmetric factor predicted by BP neural network, respectively. Figure 6 (a), (b), and (c) show the relative errors of the two methods for the absorption coefficient, scattering coefficient, and asymmetry factor, respectively. Figure 4 、 Figure 5 and Figure 6 In the figure, the horizontal axis GWC is the ice cloud content, the vertical axis Temperature is the temperature, the vertical axis absorption coefficient is the absorption coefficient, the scatter coefficient is the scattering coefficient, and the Asymmetry factor is the asymmetry factor. Figure 7 The relative error of the absorption coefficient (relative error of abs), the relative error of the scattering coefficient (relative error of sca) and the average relative error of the asymmetry factor (relative error of pha) predicted by the BP neural network at the two frequencies of 176.31GHz and 190.31GHz are listed.

[0085] Download the ERA5 historical reanalysis data, select the time and location of tropical cyclones Haishen, Maria, and Meranti, and input them into the WRF numerical forecast model to calculate the atmospheric state parameters for the scenario area. In this example, the regional grid size is set to 96×180 and the vertical layer of the atmospheric profile is set to 59.

[0086] Two network layers are introduced before the BP neural network to construct a second prediction model, which is then embedded into the DOTLRT radiation transfer model to drive the iterative training of the deep reinforcement learning algorithm. After the training, a prediction model for the scattering characteristics of ice clouds is obtained.

[0087] From the above three scenes, some networks are selected to establish the target scene. The actual brightness temperature in this scene is as follows: Figure 8 As shown in (a) of the figure, the brightness temperature of the same scene is simulated based on the ice cloud body scattering characteristics prediction model and the traditional body scattering calculation method. The results are as follows: Figure 8 Based on the brightness temperature simulation results, the brightness temperature residual maps corresponding to the ice cloud body scattering characteristics prediction model and the traditional body scattering calculation method in this scenario are calculated, as shown in (b) and (c) respectively. Figure 9 As shown in (a) and (b) in Figure 2, the root mean square error (RMSE) of the simulated brightness temperature corresponding to the ice cloud body scattering characteristics prediction model and the traditional body scattering calculation method in the cloud area (CLOUDY), convection area (CONVECTION) and no partition (ALL) is further calculated as follows: Figure 10 As shown in the figure, RMSE_DDPG represents the root mean square error of the ice cloud scattering characteristics prediction model, and RMSE_SSD represents the root mean square error of the traditional volume scattering calculation method. Figure 9 and Figure 10 The results show that the ice cloud scattering characteristics prediction model established by the present invention has a smaller error.

[0088] Furthermore, we selected a grid near the typhoon eye from the three tropical cyclone scenarios mentioned above as a test scenario, and input the ice cloud scattering characteristic parameters output by the ice cloud scattering characteristic prediction model into the DOTLRT radiation transfer model to calculate the simulated brightness temperature. The simulated brightness temperature was compared with the actual observed brightness temperature, and the BIAS and RMSE of the two were calculated to indirectly verify the prediction performance of the model. The selected scenarios specifically include the HAISHEN scenario, the MARIA scenario, and the Meranti scenario. The results under the HAISHEN scenario are as follows: Figures 11 to 13 As shown in the figure, the calculation results under the MARIA scenario are as follows Figures 14 to 16 As shown in the figure, the calculation results in the Meranti scenario are as follows Figures 17 to 19 shown.

[0089] Figure 11 、 Figure 14 and Figure 17 In the figure, (a), (b) and (c) are the actual observed brightness temperature, the brightness temperature simulated using the volume scattering network embedded in DOTLRT simulation and the brightness temperature simulated using the traditional volume scattering calculation method, respectively. The horizontal axis "longitude" represents longitude, and the vertical axis represents latitude. Figure 12 、 Figure 15 and Figure 18 In the figure, (a) and (b) are the residual diagrams of the brightness temperature simulated by using the volume scattering network embedded in DOTLRT and the brightness temperature simulated by using the traditional volume scattering calculation method, respectively. Figure 13 、 Figure 16 and Figure 19 Figure 2 shows the RMSE of the simulated brightness temperatures obtained by the two methods for clear sky (CLEANSKY), cloud (CLOUDY), convective (CONVECTION), and all regions (ALL). RMSE_DDPG represents the root mean square error (RMSE) of the ice cloud scattering prediction model, while RMSE_SSD represents the root mean square error (RMSE) of the traditional volume scattering calculation method. Comparison of the brightness temperature residuals and RMSE for each scenario shows that the ice cloud scattering prediction model developed by the present invention achieves significantly lower error in all scenarios.

[0090] Furthermore, the computation time of different methods is compared, such as Figure 20 As shown in the figure, "Mie" represents the time for a single calculation using the Mie method (spherical calculation theory), "SSD" represents the time for a single calculation using the non-spherical scattering model, and "NN" represents the time for a single calculation using the ice cloud scattering characteristics prediction model established by the present invention. In order to reduce the impact of the computer system scheduling mechanism on the single radiation transfer calculation, the calculation was repeated 50 times. The calculation time of each method above was averaged. The calculation time using Mie theory took 2.27 seconds, the calculation time using the non-spherical scattering model took 0.45 seconds, and the calculation time using the ice cloud scattering characteristics prediction model took 0.12 seconds. It can be found that the calculation speed of the ice cloud scattering characteristics prediction model is significantly improved compared to the volume scattering model, with an increase of about 73%.

[0091] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements 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 method for establishing an ice cloud scattering characteristics prediction model based on deep reinforcement learning, characterized in that: include: Step S1: constructing a neural network model and training the neural network model using a first data set consisting of ice cloud physical parameters and corresponding volume scattering parameters to obtain a first prediction model for predicting ice cloud volume scattering characteristic parameters based on the ice cloud physical parameters; the ice cloud physical parameters include frequency, temperature, and ice cloud content; Step S2: Add N network layers before the input layer of the first prediction model to obtain a second prediction model, and embed the second prediction model into the radiation transfer model; the second prediction model uses pressure, humidity, and ice cloud physical parameters as input to predict ice cloud scattering characteristic parameters; N is a positive integer; Step S3: After establishing a deep reinforcement learning training environment, using the parameters of the second prediction model in the radiative transfer model as the environmental state and minimizing the error between the simulated brightness temperature output by the radiative transfer model and the actual brightness temperature as the training objective, the second dataset is used to drive iterative training of the deep reinforcement learning to update the parameters of the second prediction model. After the training is completed, the second prediction model in the radiative transfer model is output as the ice cloud scattering characteristics prediction model. The second data set consists of ice cloud physical parameters and corresponding pressure, humidity and true brightness temperature within the ice cloud sensitive frequency band under the target scene.

2. The method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to claim 1, characterized in that: In step S2, N=2; Moreover, among the two added network layers, the first network layer includes five nodes, which are used to input frequency, temperature, ice cloud content, humidity and pressure respectively; The second network layer includes three nodes, which are respectively recorded as the first internal node, the second internal node, and the third internal node. The outputs of these three nodes are connected to the three nodes used for inputting frequency, temperature, and ice cloud content in the first prediction model, respectively; the input of the first internal node is only connected to the output of the node used for inputting frequency in the first network layer, the input of the second internal node is connected to the output of each node in the first network layer, and the input of the third internal node is connected to the output of each node in the first network layer.

3. The method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to claim 1 or 2, wherein: In step S3, the environmental state is the parameter of the network node newly added by the second prediction model relative to the first prediction model.

4. The method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to claim 3, wherein: In step S1, the constructed neural network model is a BP neural network.

5. The method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to claim 1 or 2, wherein: The target scene is a tropical cyclone scene.

6. The method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to claim 5, characterized in that: The ice cloud sensitive frequency band is the 183 GHz band.

7. A method for predicting the scattering characteristics of ice clouds, characterized in that: include: The physical parameters, humidity and pressure of the ice cloud are input into the ice cloud scattering characteristic prediction model established by the method for establishing an ice cloud scattering characteristic prediction model based on deep reinforcement learning according to any one of claims 1 to 6 to obtain the ice cloud scattering characteristic parameters.

8. A computer program product, characterized in that When the computer program product is executed by a processor, it implements the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to any one of claims 1 to 6, and / or the method for predicting ice cloud scattering properties according to claim 7.

9. A computer-readable storage medium, characterized in that The invention comprises a stored computer program, which, when executed by a processor, controls the device where the computer-readable storage medium is located to execute the method for establishing an ice cloud body scattering characteristic prediction model based on deep reinforcement learning according to any one of claims 1 to 6, and / or the ice cloud body scattering characteristic prediction method according to claim 7.

10. An electronic device, characterized in that: include: a computer-readable storage medium for storing a computer program; and a processor configured to read the computer program stored in the computer-readable storage medium and execute the method for establishing an ice cloud scattering property prediction model based on deep reinforcement learning according to any one of claims 1 to 6, and / or the method for predicting ice cloud scattering properties according to claim 7.

Images