An atmospheric radiation transfer model simulation method based on full connection and RNN neural network

Abstract

The application discloses a kind of atmospheric radiation transmission model simulation methods based on full connection and RNN neural network, comprising the following steps: the simulation object MODTRAN model is decomposed into atmospheric radiation transmission calculation and sensor spectral response calculation two links, and the parameters of these two processes are template parameter design;The template parameters of the two processes designed are input into MODTRAN model for simulation, respectively obtain atmospheric radiation transmission and sensor spectral response sample set;Establish forward simulation network, and train forward simulation network by atmospheric radiation transmission and sensor spectral response sample set;Based on forward simulation network, sensor entrance pupil radiance is simulated and calculated.The present application uses RNN neural network to simulate atmospheric radiation transmission model comprehensively, effectively avoids the great limitation of traditional method which only simulates a certain specific problem;And using RNN neural network to simulate atmospheric radiation transmission model, the speed and accuracy of calculation can be greatly improved.

Description

Technical Field

[0001] This invention belongs to the fields of atmospheric radiative transfer and deep learning technology, specifically relating to a method for simulating atmospheric radiative transfer models based on fully connected networks and RNN neural networks. Background Technology

[0002] Atmospheric radiative transfer models play a crucial role in the application of remote sensing imagery, but their enormous computational complexity and time-consuming nature have long been significant drawbacks. Researchers both domestically and internationally have actively explored ways to improve the efficiency of atmospheric radiative transfer models while maintaining their accuracy. Methodologically, these efforts can be divided into two categories: simplifying atmospheric radiative transfer model calculations using lookup tables and simulating atmospheric radiative transfer models based on deep learning.

[0003] The lookup table method primarily constructs a band-reflectivity-entry pupil radiance lookup table. The method involves: first, setting atmospheric condition parameters and geometric parameters for the simulation time and scene; then, under selected imaging conditions, writing specific reflectivity values ​​into the atmospheric radiative transfer model to obtain the entry pupil radiance at these reflectivity values, and further solving for the unknown parameters in the simplified equations; finally, discretely taking values ​​within the range [0,1] and writing them one by one into the known simplified model to obtain the entry pupil radiance at different wavelengths for each reflectivity, and then performing integration based on the sensor's spectral response function to obtain the entry pupil radiance values ​​for different bands, thus forming the band-reflectivity-entry pupil radiance lookup table. However, due to the inherent errors in the interpolation process of the lookup table method, and the fact that the size of the lookup table becomes enormous with changes in the satellite observation model and application scenario, the computational complexity of simplifying the atmospheric radiative transfer model using the lookup table method presents a significant challenge as the data volume grows rapidly.

[0004] Methods using deep learning to simulate atmospheric radiative transfer models primarily involve designing suitable deep learning networks for different application scenarios of the atmospheric radiative transfer model. Based on a deep learning-based atmospheric radiative transfer model framework, the selected model is simulated to achieve the corresponding input and output of surface reflectance versus sensor entrance pupil radiance. However, existing research on using deep learning to simulate atmospheric radiative transfer models focuses only on specific problems or sensors, rarely considering simulations based on the fundamental principles of atmospheric radiative transfer models. Therefore, it has significant limitations in application, failing to be usable in different application scenarios or for different sensors. Summary of the Invention

[0005] Purpose of the invention: In order to overcome the shortcomings of the existing technology, this invention provides a method for simulating atmospheric radiative transfer model based on fully connected networks and RNN neural networks, which can simulate the calculation process of atmospheric radiative transfer model and realize the corresponding input and output of surface reflectivity and sensor entrance pupil radiance.

[0006] Technical Solution: To achieve the above objectives, this invention provides a method for simulating atmospheric radiative transfer based on a fully connected and RNN neural network model, comprising the following steps:

[0007] S1: Decompose the simulated MODTRAN model into two parts: atmospheric radiative transfer calculation and sensor spectral response calculation, and design template parameters for the parameters of these two processes.

[0008] S2: Input the template parameters of the two designed processes into the MODTRAN model for simulation to obtain the atmospheric radiative transfer and sensor spectral response sample sets, respectively;

[0009] S3: Establish a forward simulation network and train the forward simulation network using atmospheric radiative transfer and sensor spectral response sample sets;

[0010] S4: Simulate and calculate the entrance pupil radiance of the sensor based on a forward simulation network.

[0011] Furthermore, the template parameter design in step S1 includes setting invariant parameters and setting variable parameters. The invariant parameters include the simulated spectral band, absorption spectral line parameters, aerosol model, atmospheric model, model solution method and gas absorption calculation method. The variable parameters include the solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, boundary layer visibility, aerosol optical thickness and surface reflectivity.

[0012] Furthermore, the process of obtaining the atmospheric radiative transfer dataset in step S2 is as follows:

[0013] Based on the range and interval of surface reflectance values, the parameter combinations were sampled in 201 layers. Each layer randomly generated 480 parameter combinations, resulting in a total of 96,480 data combinations. Each data combination included solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, boundary layer visibility, aerosol optical thickness, surface reflectance, and MODIS bands 1 to 4, with atmospheric top radiance at 1 nm intervals.

[0014] Furthermore, the process of obtaining the sensor spectral response sample set in step S2 is as follows:

[0015] A1: Based on the atmospheric radiative transfer dataset, calculate the standard deviation of each parameter. The formula for standard deviation is:

[0016]

[0017] Where μ is the average number of times each parameter takes different values ​​within its range, and x i Let be the number of times the i-th value appears, and n be the number of possible values ​​for each parameter.

[0018] A2: Randomly select one set of data from the atmospheric radiative transfer dataset and obtain the radiance simulation results through the MODTRAN model;

[0019] A3: Within the solar radiation range given by MODTRAN, randomly select solar radiance values ​​for bands 1 to 4 of MODIS to generate 1,000,000 uniformly distributed radiance combinations, and convolve these combinations with the zero-padded spectral response function.

[0020] L k =∫L(λ)R k (λ)dλ

[0021] Among them, L k R represents the radiant energy value recorded by the sensor, where L(λ) is the radiant energy at wavelength λ at the entrance pupil. k (λ) represents the spectral responsivity of MODIS at wavelength λ in a certain band;

[0022] 1,000,000 sets of data were generated, each set consisting of the entrance pupil radiance of MODIS bands 1 to 4 and the radiance of the top of the atmosphere at 1 nm intervals.

[0023] Furthermore, in step S2, since the different parameters in the two datasets have different value ranges, in order to facilitate the training of the neural network model, each parameter in the data is standardized to between 0 and 1. The formula for standardization using the range method is as follows:

[0024]

[0025] Among them, X max X represents the maximum value of the feature. min It represents the minimum value of the feature.

[0026] Furthermore, the forward simulation network in step S3 includes an atmospheric radiative transfer calculation network and a MODIS spectral response calculation network.

[0027] The atmospheric radiative transfer calculation network consists of one input layer, four hidden layers, and one output layer. The loss function is MSE. The input layer has seven neurons, which are the solar zenith angle, the observed zenith angle, the relative azimuth angle, the altitude of the observation point, the boundary layer visibility, the aerosol optical thickness, and the surface reflectivity. The output layer has 125 neurons, which are the radiance of the top of the atmosphere at 1nm intervals in the first to fourth bands of MODIS.

[0028] The MODIS spectral response calculation network consists of three parts: data processing, network training, and network prediction. The data processing process includes data preprocessing and data normalization. The network training process optimizes the solution process by minimizing the loss function. The network prediction process predicts the sensor entrance pupil radiance time series based on the trained RNN algorithm.

[0029] Furthermore, the training method for the forward simulation network in step S3 is as follows:

[0030] The atmospheric radiative transfer computational network was trained using an atmospheric radiative transfer sample dataset. Before training, 96,480 sample data were divided into training, testing, and validation sets in a 3:1:1 ratio.

[0031] The spectral response calculation network is trained using a sensor spectral response sample dataset. Before training, 1,000,000 sample data points are divided into training, test, and validation sets in a 3:1:1 ratio.

[0032] Further, step S4 specifically includes:

[0033] B1: Input the solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, aerosol optical thickness and surface reflectivity data into the atmospheric radiation transfer calculation network to calculate and obtain the radiance of the top of the atmosphere;

[0034] B2: Input the atmospheric top radiance time series into the MODIS spectral response calculation network to calculate and obtain the sensor entrance pupil radiance.

[0035] Beneficial effects: Compared with existing technologies, this invention differs from traditional deep learning-based atmospheric radiative transfer model simulation methods by focusing on simulating the principles of atmospheric radiative transfer models. It utilizes RNN neural networks to comprehensively simulate both atmospheric radiative transfer calculations and sensor spectral responses, effectively avoiding the significant limitations of traditional methods that only simulate specific problems. Furthermore, using RNN neural networks to simulate atmospheric radiative transfer models can significantly improve the speed and accuracy of calculations. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the acquisition of the sensor's entrance pupil radiance.

[0038] Figure 3 This is a structural diagram of the atmospheric radiative transfer calculation network;

[0039] Figure 4 This is a diagram of the network structure for calculating the spectral response.

[0040] Figure 5 This is a flowchart of the simulation calculation of the sensor's entrance pupil radiance. Detailed Implementation

[0041] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0042] This invention provides a method for simulating atmospheric radiative transfer models based on fully connected networks and RNN neural networks, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0043] S1: Decompose the simulated MODTRAN model into two parts: atmospheric radiative transfer calculation and sensor spectral response calculation, and design template parameters for the parameters of these two processes.

[0044] The specific process is as follows:

[0045] 1) The atmospheric radiative transfer model can be divided into two independent processes: atmospheric radiative transfer calculation and sensor spectral response, each with its own parameter inputs. The input parameters for the atmospheric radiative transfer calculation mainly include geometric parameters, aerosol parameters, atmospheric model parameters, etc.; the input parameter for the sensor spectral response process is the spectral response function.

[0046] 2) Set constant parameters. The specific settings for constant parameters are shown in Table 1. 3)

[0048] Table 1 - Design of Invariant Parameters for the MODTRAN Model

[0049]

[0050]

[0051] 3) Set the variable parameters. The specific settings for the variable parameters are shown in Table 2.

[0052] Table 2 - MODTRAN Model Variable Parameter Design Table

[0053]

[0054] 4) Design the sensor spectral response parameters and select the spectral response functions of the first to fourth bands of MODIS for simulation.

[0055] S2: Input the template parameters of the two designed processes into the MODTRAN model for simulation to obtain the atmospheric radiative transfer and sensor spectral response sample sets, respectively;

[0056] The specific process is as follows:

[0057] 1) Based on the range and interval of surface reflectance, the parameter combinations were sampled in 201 layers. 480 parameter combinations were randomly generated in each layer, for a total of 96,480 data combinations. Each data combination included solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, boundary layer visibility, aerosol optical thickness, surface reflectance, and atmospheric top radiance in MODIS bands 1 to 4 at 1 nm intervals.

[0058] 2) Based on the above steps, calculate the standard deviation of each parameter. The formula for standard deviation is:

[0059]

[0060] Where μ is the average number of times each parameter takes different values ​​within its range, and x i Let be the number of times the i-th value appears, and n be the number of possible values ​​for each parameter.

[0061] 3) Randomly select one set of data from the data combination obtained in step 1), and use the MODTRAN model to simulate the radiance.

[0062] 4) Within the solar radiation range given by MODTRAN, randomly select solar radiance values ​​from MODIS values ​​1 to 4 to generate 1,000,000 uniformly distributed radiance combinations, and convolve these combinations with the zero-padded spectral response function.

[0063] L k =∫L(λ)R k (λ)dλ

[0064] Among them, L k R represents the radiant energy value recorded by the sensor, where L(λ) is the radiant energy at wavelength λ at the entrance pupil. k (λ) represents the spectral responsivity of the first band of MODIS at wavelength λ.

[0065] 1,000,000 sets of data were generated, each set consisting of the entrance pupil radiance of MODIS bands 1 to 4 and the radiance of the top of the atmosphere at 1 nm intervals.

[0066] 5) Since the different parameters in the two datasets obtained in steps (1) and (4) have different value ranges, in order to facilitate the training of the neural network model, it is necessary to standardize each parameter in the data to a value between 0 and 1. The formula for standardization using the range method is as follows:

[0067]

[0068] Among them, X max X represents the maximum value of the feature. min It represents the minimum value of the feature.

[0069] S3: Establish a forward simulation network, which includes an atmospheric radiative transfer calculation network and a MODIS spectral response calculation network;

[0070] like Figure 3 As shown, the atmospheric radiative transfer calculation network consists of one input layer, four hidden layers, and one output layer, with the loss function being MSE. The input layer has seven neurons, which are the solar zenith angle, the observed zenith angle, the relative azimuth angle, the altitude of the observation point, the boundary layer visibility, the aerosol optical thickness, and the surface reflectivity. The output layer has 125 neurons, which are the atmospheric top radiance at 1nm intervals in the first to fourth bands of MODIS.

[0071] like Figure 4 As shown, the MODIS spectral response calculation network consists of three parts: data processing, network training, and network prediction. The data processing process includes data preprocessing and data normalization. The network training process optimizes the solution process by minimizing the loss function. The network prediction process predicts the sensor entrance pupil radiance time series based on the trained RNN algorithm.

[0072] The atmospheric radiative transfer computational network was trained using an atmospheric radiative transfer sample dataset. Before training, 96,480 sample data were divided into training, testing, and validation sets in a 3:1:1 ratio.

[0073] The spectral response calculation network is trained using a sensor spectral response sample dataset. Before training, 1,000,000 sample data points are divided into training, test, and validation sets in a 3:1:1 ratio.

[0074] S4: Simulation calculation of sensor entrance pupil radiance based on a forward simulation network, such as... Figure 5 As shown, the specific calculation process includes the following steps:

[0075] 1) Input the solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, aerosol optical thickness and surface reflectivity data into the atmospheric radiation transfer calculation network to calculate and obtain the radiance of the top of the atmosphere;

[0076] 2) Input the atmospheric top radiance time series into the MODIS spectral response calculation network to calculate and obtain the sensor entrance pupil radiance.

Claims

1. A method for simulating atmospheric radiative transfer based on a fully connected and RNN neural network, characterized in that, The steps include the following: S1: Decompose the simulated MODTRAN model into two parts: atmospheric radiative transfer calculation and sensor spectral response calculation, and design template parameters for the parameters of these two processes. S2: Input the template parameters of the two designed processes into the MODTRAN model for simulation to obtain the atmospheric radiative transfer and sensor spectral response sample sets, respectively; S3: Establish a forward simulation network and train the forward simulation network using atmospheric radiative transfer and sensor spectral response sample sets; S4: Simulation calculation of sensor entrance pupil radiance based on forward simulation network; The template parameter design in step S1 includes setting invariant parameters and setting variable parameters. Invariant parameters include simulation spectral band, absorption spectral line parameters, aerosol model, atmospheric model, model solution method and gas absorption calculation method. Variable parameters include solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, boundary layer visibility, aerosol optical thickness and surface reflectivity. The forward simulation network in step S3 includes an atmospheric radiative transfer calculation network and a MODIS spectral response calculation network. The atmospheric radiative transfer calculation network consists of one input layer, four hidden layers, and one output layer. The loss function is MSE. The input layer has seven neurons, which are the solar zenith angle, the observed zenith angle, the relative azimuth angle, the altitude of the observation point, the boundary layer visibility, the aerosol optical thickness, and the surface reflectivity. The output layer has 125 neurons, which are the radiance of the top of the atmosphere at 1nm intervals in the first to fourth bands of MODIS. The MODIS spectral response calculation network consists of three parts: data processing, network training, and network prediction. The data processing process includes data preprocessing and data normalization. The network training process optimizes the solution process by minimizing the loss function. The network prediction process predicts the sensor entrance pupil radiance time series based on the trained RNN algorithm.

2. The atmospheric radiative transfer model simulation method based on fully connected networks and RNN neural networks according to claim 1, characterized in that, The process of obtaining the atmospheric radiative transfer dataset in step S2 is as follows: Based on the range and interval of surface reflectance values, the parameter combinations were sampled in 201 layers. Each layer randomly generated 480 parameter combinations, resulting in a total of 96,480 data combinations. Each data combination included solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, boundary layer visibility, aerosol optical thickness, surface reflectance, and MODIS bands 1 to 4, with atmospheric top radiance at 1 nm intervals.

3. The atmospheric radiative transfer model simulation method based on fully connected networks and RNN neural networks according to claim 2, characterized in that, The process of obtaining the sensor spectral response sample set in step S2 is as follows: A1: Based on the atmospheric radiative transfer dataset, calculate the standard deviation of each parameter. The formula for standard deviation is: ； in, This is the average number of times each parameter takes different values ​​within its range. For the first The number of times each value appears. The number of possible values ​​for each parameter; A2: Randomly select one set of data from the atmospheric radiative transfer dataset and obtain the radiance simulation results through the MODTRAN model; A3: Within the solar radiation range given by MODTRAN, randomly select solar radiance values ​​for bands 1 to 4 of MODIS to generate 1,000,000 uniformly distributed radiance combinations, and convolve these combinations with the zero-padded spectral response function. ； in, The value of radiation energy recorded by the sensor. The wavelength at the entrance pupil is Radiant energy, For a certain band of MODIS at wavelength Spectral response at; 1,000,000 sets of data were generated, each set consisting of the entrance pupil radiance of MODIS bands 1 to 4 and the radiance of the top of the atmosphere at 1 nm intervals.

4. The atmospheric radiative transfer model simulation method based on fully connected networks and RNN neural networks according to claim 3, characterized in that, In step S2, since the two sets of datasets have different parameter ranges, each parameter in the data is standardized to between 0 and 1 to facilitate the training of the neural network model. The formula for standardization using the range method is as follows: ； in, The maximum value of the feature. It represents the minimum value of the feature.

5. The atmospheric radiative transfer model simulation method based on fully connected networks and RNN neural networks according to claim 1, characterized in that, The training method for the forward simulation network in step S3 is as follows: The atmospheric radiative transfer computational network was trained using an atmospheric radiative transfer sample dataset. Before training, 96,480 sample data were divided into training, testing, and validation sets in a 3:1:1 ratio. The spectral response calculation network is trained using a sensor spectral response sample dataset. Before training, 1,000,000 sample data points are divided into training, test, and validation sets in a 3:1:1 ratio.

6. The atmospheric radiative transfer model simulation method based on fully connected networks and RNN neural networks according to claim 1, characterized in that, Step S4 specifically involves: B1: Input the solar zenith angle, observation zenith angle, relative azimuth angle, observation point altitude, aerosol optical thickness, and surface reflectivity into the atmospheric radiative transfer calculation network to obtain the atmospheric top radiance of MODIS bands 1 to 4. B2: Input the atmospheric top radiance obtained in the previous step into the spectral response calculation network to obtain the sensor entrance pupil radiance.

Images