Satellite-borne microwave radiometer on-orbit joint calibration method based on fusion of physical model and neural network

Abstract

The invention provides a satellite-borne microwave radiometer in-orbit joint calibration method based on fusion of a physical model and a neural network. The method comprises the following steps: modeling a microwave radiometer aperture surface brightness temperature by using the physical model; modeling the brightness temperature observed by the microwave radiometer by adopting a neural network; and alternately optimizing the parameters of the oral surface brightness temperature model and the observation brightness temperature model until a set optimization condition is met. The method has the advantages that the calibration precision and robustness of the model are improved; the precision and the efficiency are remarkably improved while the global optimization capability is maintained by the combined optimization strategy; according to the alternative optimization strategy, a complex joint optimization problem is decomposed, the optimization difficulty is reduced, the calculation efficiency is improved, and convergence can be more stable.

Description

Technical Field

[0001] This application belongs to the technical field of satellite remote sensing data processing, and specifically relates to an on-orbit joint calibration method for spaceborne microwave radiometers based on the fusion of physical models and neural networks. Background Art

[0002] The fixed-beam-pointing microwave radiometer is a non-scanning passive remote sensing device, and the pointing of its observation antenna is in a fixed direction. There are two main application scenarios for this type of radiometer: 1) observing targets in a specific direction; 2) performing co-path observation with a radar altimeter. By receiving the microwave radiation from the surface, the atmosphere, and the cold sky background, the wet path delay caused by water vapor and liquid water in clouds when the radar altimeter signal passes through the Earth's atmosphere can be corrected in real time. The prerequisite for the microwave radiometer (for simplicity of expression, the microwave radiometer mentioned in this application hereinafter refers to the fixed-beam-pointing microwave radiometer) to achieve the above applications is to accurately calibrate it. Currently, the calibration of the microwave radiometer can be divided into two stages: the calibration of the antenna aperture brightness temperature T A and the calibration of the observed brightness temperature T B Among them, the calibration of the antenna aperture brightness temperature T A refers to converting the observed voltage value measured by the microwave radiometer into the aperture brightness temperature of the antenna; the calibration of the observed brightness temperature T B refers to converting the antenna aperture brightness temperature T A into the brightness temperature observed within the main lobe of the antenna, that is, the brightness temperature of the target.

[0003] The observation antenna of the microwave radiometer usually consists of a parabolic reflector and a feed source. Among them, the reflector is mainly used to reflect the incident microwave signal into the feed source. Since the reflector of the microwave radiometer receives the microwave radiation of all targets within the antenna illumination range, and the contribution of the microwave radiation of each target to the aperture brightness temperature is related to the antenna pattern. Therefore, the microwave radiometer usually performs on-orbit calibration by analyzing the antenna pattern. The antenna pattern correction algorithm refers to expressing the observed brightness temperature of the antenna as a linear weighted sum of the antenna efficiency and the effective brightness temperature of the corresponding part.

[0004] In the prior art, the microwave radiometer usually performs on-orbit calibration in a serial manner, that is, first performing the calibration of the antenna aperture brightness temperature and then the calibration of the observed brightness temperature. Specifically, by substituting the calibration parameters obtained in the ground thermal vacuum test into the on-orbit observation data, the aperture brightness temperature of the antenna is obtained, and then the observed brightness temperature is obtained through the antenna pattern correction algorithm. However, this serial method fails to fully utilize the complementarity between the two. In addition, due to the multi-solution of the aperture calibration equation, the accuracy of the antenna pattern correction algorithm modeling, the change of the dynamic environment during the on-orbit operation of the microwave radiometer, and the lack of an effective global optimization strategy during the optimization process, there is still a large room for improvement in the on-orbit calibration accuracy of the microwave radiometer. Summary of the Invention

[0005] The purpose of this application is to address the above-mentioned deficiencies and propose an on-orbit joint calibration method for spaceborne microwave radiometers based on the fusion of physical models and neural networks, including: Step 1: Model the aperture brightness temperature of the microwave radiometer using a physical model; Step 2: Model the observed brightness temperature of the microwave radiometer using a neural network; Step 3: Alternately optimize the parameters of the observed aperture model and the observed brightness temperature model until the set optimization conditions are met.

[0006] As an improvement to the above method, Step 1 includes: Establish an aperture brightness temperature model for the microwave radiometer:

[0007]

[0008] wherein, and respectively represent the predicted aperture brightness temperature values of the observation branch and the cryogenic calibration branch; a 1 、 a 2 、 a 3 、 c 1 、 c 2 、 c 3 and u are physical model parameters to be optimized; T aw and T cw respectively represent the physical temperatures of the waveguides in the observation branch and the cryogenic calibration branch; T as and T cs respectively represent the physical temperatures of the switches in the observation branch and the cryogenic calibration branch; Q A represents the non-linear error of the aperture brightness temperature; V A 、 V H and V C respectively represent the output voltages of the receivers on the observation branch and the high- and low-temperature calibration branches; represents the equivalent input brightness temperature of the high-temperature calibration reference source; represents the equivalent input brightness temperature of the low-temperature calibration reference source; represents the equivalent input brightness temperature of the observation branch; Q C Indicates the nonlinear error of the brightness temperature of the aperture in the low temperature calibration branch; A ground thermal vacuum test was conducted, and multiple working temperature test points of microwave radiometers were set. At each working temperature test point, when the low-temperature calibration variable temperature source was stable at the set temperature, the observation variable temperature source was heated or cooled with a step length of S1 in the range of T1 K to T2 K, and the data set groundDataTa was obtained; when the observation variable temperature source was stable to the set temperature, the low-temperature calibration variable temperature source was heated or cooled with a step length of S2 in the range of T1 K to T2 K, and the data set groundDataTc was obtained; at the same time, the temperature measurement data of the observation variable temperature source and the high and low temperature variable temperature sources, the corresponding output voltages, and each attenuation component on each transmission branch were collected; when V A =V H 、V A =V C and V H =V C The corresponding data sets groundDataVAVH, groundDataVAVC and groundDataVHVC are obtained respectively.

[0009] As an improvement of the above method, step 2 comprises: The on-orbit observation data of the microwave radiometer is used as the original data set, and the original data set is quality controlled and normalized to obtain the test data; Divide the experimental data set into training set and test set; Choose a feedforward neural network, a convolutional neural network, or a recurrent neural network and their variants to model the observed brightness temperature.

[0010] As an improvement of the above method, step 3 comprises: Step 3-1: Initialize the parameters of the aperture brightness temperature model and the observation brightness temperature model and set the upper and lower limits of each parameter; Step 3-2: Alternately optimize the parameters of the surface brightness temperature model and the observation brightness temperature model, including: Step 3-2-1: Fix the parameters of the observation brightness temperature model and optimize the parameters of the aperture brightness temperature model; Step 3-2-2: Fix the parameters of the aperture brightness temperature model and optimize the parameters of the observation brightness temperature model; Step 3-2-3: Calculate whether the total residual of the aperture brightness temperature model and the observed brightness temperature model meets the set stop optimization condition. If not, repeat steps 3-2-1 and 3-2-2 until the set stop optimization condition is met.

[0011] As an improvement of the above method, step 3-2-1 includes: Define the objective function groundStageObjective. Combine the ground thermal vacuum test data and on-orbit observation data to adjust the parameters of the aperture brightness temperature model, so that the combined residual groundStageLoss of the ground and on-orbit data is minimized; The objective function groundStageObjective is expressed as:

[0012] where, m represents the data volume of the dataset groundDataTa obtained when the cryogenic calibration variable temperature source is stabilized at the set temperature and the observed variable temperature source is heated or cooled within the range of Tl K to T2 K with a step size of S1; n represents the data volume of the dataset groundDataTc obtained when the observed variable temperature source is stabilized at the set temperature and the cryogenic calibration variable temperature source is heated or cooled within the range of Tl K to T2 K with a step size of S2; represents the i true value of the aperture brightness temperature of the th observation branch; i represents the predicted value of the aperture brightness temperature of the th observation branch; j represents the true value of the aperture brightness temperature of the th cryogenic calibration branch; j represents the predicted value of the aperture brightness temperature of the th cryogenic calibration branch;

[0013] where, Loss groundDataTa and Loss groundDataTc respectively represent the loss functions of the T A curve and the T C curve during the ground thermal vacuum test; Loss orbitTrain represents the loss function of the on-orbit training set data; Use a global optimization algorithm to globally optimize the parameters of the aperture brightness temperature model; Use an optimization algorithm with strong constraint conditions to locally optimize the parameters of the aperture brightness temperature model: for the groundDataVAVH, groundDataVAVC, and groundDataVHVC datasets, the following strong constraint conditions need to be satisfied respectively: a 1 × T A +a 2 × T aw ≈ (1 - a 3 )× T as a 1 × T A + a 2 × T aw + a 3 × T as ≈ c 1 × T C + c 2 × T cw + c 3 × T cs c 1 × T C + c 2 × T cw ≈ (1 - c 3 )× T cs wherein, T A represents the aperture brightness temperature; T C represents the temperature of the low-temperature calibration source.

[0014] As an improvement to the above method, step 3-2-2 includes: Define the objective function nnStageObjective, and generate the predicted value of the aperture brightness temperature of the observation antenna by using the latest aperture brightness temperature model parameters , and calculate the joint residual nnStageLoss of the ground and on-orbit data by combining the observed brightness temperature model parameters and the on-orbit observation data; Use the global optimization algorithm to train the observed brightness temperature model parameters to make the predicted value of the observed brightness temperature approach the true value; use the local optimization algorithm to perform local optimization on the observed brightness temperature model parameters; The objective function nnStageObjective is expressed as:

[0015] Among them, L represents the amount of observation data in the on-orbit training set; represents the true value of the observed brightness temperature in orbit; represents the predicted value of the observed brightness temperature of the radiometer in orbit.

[0016] As an improvement of the above method, step 3-2-3 includes: Calculate the total residual of the aperture brightness temperature model and the observed brightness temperature model, and end the optimization when the set number of cycles is reached or the convergence threshold of the total residual reaches the set threshold.

[0017] As an improvement of the above method, it further includes: Step 4: Use the optimized parameters to construct the final aperture brightness temperature model and the observed brightness temperature model, test them on an independent test set, and evaluate the accuracy and generalization ability of the models.

[0018] Compared with the prior art, the advantages of this application are as follows: 1. For the aperture brightness temperature, a physical calibration model is adopted, and for the observed brightness temperature, a neural network model is used for calibration. The physical model provides prior knowledge and physical constraints, while the neural network model learns the complex non-linear relationship during the on-orbit operation of the radiometer in a data-driven manner. This fusion method provides physical constraints for the data-driven model, not only retaining the interpretability of the model but also improving the calibration accuracy and robustness of the model.

[0019] 2. During the calibration process of the aperture brightness temperature, the introduction of constraint conditions ensures the rationality of physical parameters. This constraint mechanism not only improves the stability of the model but also further enhances the optimization accuracy.

[0020] 3. The global optimization algorithm can effectively explore the solution space and avoid local optima; the optimization with strong constraint conditions and the local optimization algorithm can perform fine searches near the optimal solution found by the global optimization algorithm, further improving the accuracy of the solution. This combined optimization strategy significantly improves the accuracy and efficiency of optimization while maintaining the global optimization ability.

[0021] 4. The alternating optimization strategy decomposes the complex joint optimization problem, not only reducing the optimization difficulty and improving the calculation efficiency but also making the convergence more stable. Description of the Drawings

[0022] Figure 1 It shows a schematic diagram of the front-end structure of a microwave radiometer receiver; Figure 2 It shows a flow chart of on-orbit calibration data processing for a microwave radiometer; Figure 3 The probability density map of the on-orbit calibration results of MWR at 18.7 GHz in the test set is shown (green: pure physical model, red: physical + neural network model). Detailed implementation manners

[0023] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0024] The present invention aims to provide an on-orbit joint calibration method for a spaceborne microwave radiometer based on the fusion of a physical model and a neural network. Among them, the aperture brightness temperature is modeled by a physical model; the observed brightness temperature is modeled by a neural network; by alternately optimizing the physical model parameters and the neural network model parameters, combining the strong constraints of the ground thermal vacuum test data and the soft constraints of the on-orbit observation data, the global optimum of the model parameters is realized to calibrate the radiometer on orbit, and at the same time, the aperture brightness temperature T A and the observed brightness temperature T B are obtained.

[0025] As Figure 2 shown, the present invention provides an on-orbit joint calibration method for a spaceborne microwave radiometer based on the fusion of a physical model and a neural network. The method is as follows: Step 1: Modeling the aperture brightness temperature of the microwave radiometer, specifically including: As Figure 1 shown, the microwave radiometer is calibrated by a periodic two-point calibration method. Through the periodic switching of the internal switch of the receiver, the observed signal and the high- and low-temperature calibration source signals enter the receiver through different transmission paths respectively. The transmission path usually includes microwave components such as a feed, a connecting waveguide, and a switch. For the observation branch and the low-temperature calibration branch, there are respectively: (1) (2) where and respectively represent the aperture calibration equations established based on the physical model for the observation branch and the low-temperature calibration branch of the microwave radiometer; and respectively represent the input vectors of the calibration equations in the observation branch and the low-temperature calibration branch (including the physical temperatures and output voltages of each microwave component, etc.); and respectively represent the predicted aperture brightness temperature values of the observation branch and the low-temperature calibration branch.

[0026] During the ground thermal vacuum test, the observed variable temperature source and the low-temperature calibration source of the microwave radiometer (with known radiation characteristics, controllable and variable temperature, and the temperature change range from Tl K to T2 K) are used to simulate the on-orbit observation target and the low-temperature calibration reference source respectively. During the test, specific temperature test points are usually set for the operating temperature of the microwave radiometer. At each temperature test point, when the low-temperature calibration variable temperature source is stable at the preset temperature, the observed variable temperature source is heated or cooled within the range of Tl K to T2 K with a step size of S1 to obtain the data set groundDataTa; when the observed variable temperature source is stable at the preset temperature, the low-temperature calibration variable temperature source is heated or cooled within the range of Tl K to T2 K with a step size of S2 to obtain the data set groundDataTc. At the same time, the data acquisition unit of the microwave radiometer will collect the temperature measurement data of the observed variable temperature source, the high- and low-temperature variable temperature sources, the corresponding output voltages, and each attenuation component on each transmission path. In addition, during the cycle, by setting the voltage equal points: V A =V H 、V A =V C and V H =V C the corresponding data sets groundDataVAVH, groundDataVAVC, and groundDataVHVC are obtained respectively. Among them, V A 、V H and V C represent the output voltages of the receivers on the observed branch and the high- and low-temperature calibration branches respectively.

[0027] For the data sets groundDataTa and groundDataTc, they need to satisfy equations (1) and (2) respectively. For the data sets groundDataVAVH, groundDataVAVC, and groundDataVHVC, the non-linear error on the observed branch or the low-temperature calibration branch of the microwave radiometer is minimized (almost zero). At this time, the microwave radiometer system is almost in a linear state, and there is no need to correct its non-linear error. At this time, the predicted values of the aperture brightness temperature of the observed branch and the low-temperature calibration branch are as follows: (3) (4) Among them, and represent the aperture calibration equations without non-linear correction established based on the physical model for the observed branch and the low-temperature calibration branch of the microwave radiometer respectively.

[0028] During the ground thermal vacuum test, the observations of the microwave radiometer and the cryogenic feed are respectively aligned with the corresponding variable temperature sources. Therefore, the observed values with known radiation characteristics and the brightness temperature values of the cryogenic calibration source can be directly used as the "true values" of the aperture brightness temperature for each.

[0029] Step 2: Model the observed brightness temperature of the microwave radiometer, specifically including: Step 2-1: Calculate the along-track observed brightness temperature: Use the along-track simulated brightness temperature of the microwave radiometer generated based on the radiative transfer model (or, based on the star-cross comparison method, the observed brightness temperature at the cross-matching point with the reference payload) as the "true value" of the on-orbit observed brightness temperature of the radiometer to be calibrated.

[0030] Step 2-2: Control data quality: Collect the raw data and clean the data (e.g., handle missing values, detect outliers, process duplicate data, etc.). In addition, since the observation accuracy of the radiative transfer model (or the reference payload) is easily affected by factors such as surface type (land, sea, or ice surface), cloud thickness, and precipitation. Therefore, it is necessary to further screen out "stable" data as the data set.

[0031] Step 2-3: Divide the data set: Randomly (or in a fixed ratio sequence, etc.) divide the data after quality control into a training set orbitTrain, a validation set orbitValidation, and a test set orbitTest according to the ratio of m:n:k. Among them, the training set is used to solve the calibration model parameters, the validation set is used to adjust the model parameters, and the test set is used to evaluate the model.

[0032] Step 2-4: Construct the model input and output: Generally, the aperture brightness temperature of the microwave radiometer is a linear weighted sum of the antenna efficiency and the effective brightness temperature of the corresponding part. Different from explicitly calculating the contributions of each part in the aperture brightness temperature, the neural network mainly establishes the relationship between the observed brightness temperature and the aperture brightness temperature through the main features of the samples. Therefore, the input of the observed brightness temperature model is data related to the effective brightness temperature of each part, such as: the aperture brightness temperature of the antenna, the latitude of the satellite, and the physical temperature of the parabolic reflector, etc. By calculating the correlation between each feature and the observed brightness temperature, select the features with higher correlation as the input. The finally selected input can be T A 、T A 2 、Tref (physical temperature of the reflector surface), latitude lat, satellite cabin panel temperature, sea surface temperature, etc. The output of the model is the predicted value of the observed brightness temperature of the microwave radiometer.

[0033] Step 2-5: Normalization processing: The input data may have different dimensions or ranges. To avoid the "gradient explosion phenomenon" caused by too large input values, it is usually necessary to normalize the input and output data of the network so that each index is on the same dimension. Common normalization methods include: Z-score normalization method, Min-Max normalization method, etc.

[0034] Step 2-6: Select the network model: For the observation data of the microwave radiometer, a feedforward neural network (or convolutional neural network, recurrent neural network, and its variants such as LSTM, GRU, etc.) can be selected to fit the data. At the same time, the corresponding network structure parameters need to be set (such as the number of hidden layers, the number of neurons, and the activation function in the feedforward neural network. The number of convolutional kernels, size, stride, and pooling method in the convolutional neural network, etc.).

[0035] Step 3: Joint solution of alternating optimization of physical model parameters and neural network parameters, specifically including: Step 3-1: Initialize the physical model parameters and neural network parameters and set the upper and lower limits of each parameter. By establishing a brightness temperature radiation transfer model for the physical parameters of each microwave component of the radiometer (such as insertion loss, voltage reflection coefficient at the input and output ends, etc.), calculate the upper and lower limits of each parameter in the physical model to ensure the rationality of each parameter during the optimization process. The neural network parameters mainly refer to the parameters that need to be set in the model network structure.

[0036] Step 3-2: Execute the alternating optimization strategy: Alternately optimize the physical model parameters and neural network model parameters, specifically including the following steps: Step 3-2-1: Fix the neural network parameters and optimize the physical model parameters: By defining the objective function groundStageObjective, combining the ground thermal vacuum test data and on-orbit observation data, and adjusting the physical model parameters, the joint residual groundStageLoss of the ground and on-orbit data is minimized.

[0037] The objective function groundStageObjective is expressed as: (5) Among them, m represents the data volume of the dataset groundDataTa obtained when the low-temperature calibration variable temperature source is stabilized at the set temperature and the observed variable temperature source is heated or cooled within the range of Tl K to T2 K with a step size of S1; n represents the data volume of the dataset groundDataTc obtained when the observed variable temperature source is stabilized at the set temperature and the low-temperature calibration variable temperature source is heated or cooled within the range of Tl K to T2 K with a step size of S2; represents the true aperture brightness temperature of the i th observation branch; represents the predicted aperture brightness temperature value of the i th observation branch; represents the true aperture brightness temperature of the j th low-temperature calibration branch; represents the predicted aperture brightness temperature value of the j th low-temperature calibration branch. (6) Among them, Loss groundDataTa and Loss groundDataTc respectively represent the loss functions of the T A curve and the T C curve during the ground thermal vacuum test; Loss orbitTrain represents the loss function of the in-orbit training set data. The loss function can adopt the root mean square error or the average error, etc.

[0038] Use a global optimization algorithm (such as: whale optimization algorithm, genetic algorithm, etc.) to globally optimize the physical model parameters.

[0039] Use an optimization algorithm with strong constraint conditions to optimize the physical model parameters: for the groundDataVAVH, groundDataVAVC, and groundDataVHVC data sets, they need to satisfy the strong constraint conditions of the approximate equalities , and respectively: : a 1 × T A + a 2 × T aw ≈ (1 - a 3 )× T as (7) : a 1 × T A + a 2 × T aw + a 3 × T as ≈c 1 × T C + c 2 × T cw + c 3 × T cs (8) : c 1 × T C + c 2 × T cw ≈ (1 - c 3 )× T cs (9) Among them, T A represents the brightness temperature of the aperture plane; T C represents the temperature of the low-temperature calibration source. The approximate meaning of the above formula is that the difference between the values at both ends of the formula is within the set range.

[0040] Step 3-2-2: Fix the physical model parameters and optimize the neural network parameters: By defining the objective function nnStageObjective, generate the predicted value of the brightness temperature of the aperture plane of the observation antenna using the latest physical model parameters , and combine the network parameters and on-orbit observation data to calculate the joint residual nnStageLoss of the ground and on-orbit data.

[0041] The objective function nnStageObjective is expressed as: (10) Among them, L represents the amount of observation data in the on-orbit training set; represents the true value of the on-orbit observed brightness temperature; represents the predicted value of the on-orbit observed brightness temperature of the radiometer.

[0042] Use a global optimization algorithm (such as: whale optimization algorithm, genetic algorithm, ant colony algorithm, etc.) to train the neural network parameters so that the predicted value of the observed brightness temperature approaches the "true value".

[0043] Use a local optimization algorithm (such as gradient descent algorithm, Levenberg-Marquardt algorithm LM, etc.) to perform local optimization on the neural network parameters.

[0044] Step 3-2-3: Calculate the total residual totalLoss (groundStageLoss + nnStageLoss) and check for convergence. If the convergence condition is met, stop the optimization; otherwise, continue to the next iteration and continue to execute steps 3-2-1) to 3-2-3) until convergence.

[0045] Step 4: Model output inspection and evaluation. It means using the optimized parameters to construct the final model and obtaining the aperture brightness temperature and observed brightness temperature of the radiometer. The model parameters can be continuously adjusted on the validation set, and the performance of the model can be evaluated on the test set.

[0046] The on-orbit joint calibration method for spaceborne microwave radiometers based on the fusion of physical models and neural networks provided by this application has the following advantages: 1. The physical model is used to model the aperture brightness temperature of the microwave radiometer, and the neural network model is used to model the observed brightness temperature, breaking through the bottleneck of the calibration accuracy of traditional pure physical models and overcoming the problem of poor interpretability of pure neural network models. The fusion method of the present invention achieves a good balance between accuracy and interpretability.

[0047] 2. Bidirectional alternating optimization framework: An alternating iteration mechanism of "strong constraints on ground data, soft constraints on on-orbit data, and joint residual feedback" is proposed. By decomposing the complex joint optimization problem, it not only reduces the optimization difficulty and improves the calculation efficiency, but also enables more stable convergence.

[0048] 4. Hybrid optimization algorithm: The global optimization and non-linear algorithm with strong constraint conditions are used to optimize the ground data; the global optimization algorithm and local optimization algorithm are combined to optimize the neural network model. The global optimization algorithm can effectively explore the solution space and avoid local optima; the non-linear optimization algorithm with strong constraint conditions and the local optimization algorithm can perform fine search near the optimal solution found by the global optimization algorithm to further improve the accuracy of the solution. This combined optimization strategy significantly improves the accuracy and efficiency of optimization while maintaining the global optimization ability.

[0049] Simulation experiment: Assume that a three-frequency fixed-pointing microwave radiometer MWR operates on an orbit at an altitude of 970 km and an inclination of 66°. The observation frequencies of MWR are 18.7 GHz, 23.8 GHz, and 37 GHz. The data products of MWR are released in a cycle of 10 days (cycle), and the observation frequency of the L2-level data of MWR is 1 second. From Figure 2 it can be seen that the main steps of the on-orbit calibration of this microwave radiometer include: aperture brightness temperature modeling based on physical models, preprocessing / quality control / dataset partitioning / normalization processing of on-orbit data, observed brightness temperature modeling based on neural networks, and alternating optimization solution of calibration parameters.

[0050] Physical model-based aperture brightness temperature modeling: (1) Ideally, the receiver of the microwave radiometer is linear, and the relationship between the input brightness temperature and the output voltage of the receiver is as follows: (11) Wherein, V N represents the output voltage of the receiver; a and b are the calibration coefficients of the radiometer; T N ' represents the equivalent input brightness temperature of the receiver; N =( C , A , H ) represents different brightness temperature signals received by the receiver, where C represents the cold source, A represents the observation, H represents the heat source. Using the equivalent input brightness temperatures T H ' and T C ' of the high- and low-temperature calibration reference sources, as well as the corresponding output voltages V H and V C the calibration coefficients a and b of the radiometer can be obtained, and then the equivalent input brightness temperature T A ' of the observation branch can be obtained:

[0051] (12) However, in the actual working process, the non-linearity of the power response of the detection diode in the radiometer will bring non-linear errors to the system. For this non-linear error, the U coefficient method proposed by the Jet Propulsion Laboratory (JPL) of the United States can be used for correction, and the calculation formula is as follows: (13) Wherein, Q A represents the non-linear error of the aperture brightness temperature; u is used to represent the non-linearity of the radiometer system; G represents the system gain. For the observation branch, when the brightness temperature is T AThe input signal enters the feed and passes through microwave components such as waveguides, isolators, and switches. The equivalent input brightness temperature at the input port of the receiver is: (14) In summary, the predicted value of the aperture brightness temperature of the observation antenna is: (15) Similarly, for the low-temperature calibration branch, the predicted value of its aperture brightness temperature is: (16) where a 1 , a 2 , a 3 , c 1 , c 2 , c 3 and u are physical model parameters to be optimized; T aw and T cw represent the physical temperatures of the waveguides in the observation branch and the low-temperature calibration branch, respectively. T as and T cs represent the physical temperatures of the switches in the observation branch and the low-temperature calibration branch, respectively; Q C represents the non-linear error of the aperture brightness temperature in the low-temperature calibration branch.

[0052] (2) In the ground thermal vacuum test, temperature cycling is performed on the temperature-controllable and variable observation temperature source and low-temperature calibration source, respectively (for example: when the low-temperature calibration source stabilizes at 90 K, the observation temperature source changes from 90 K to 330 K; or when the observation temperature source stabilizes at 90 K, the low-temperature calibration source changes from 90 K to 330 K), and the data sets groundDataTa (T A curve) and groundDataTc (Tc curve) are obtained. In addition, during the temperature cycling process, voltage equal points (V A = V H , V C = V C or V H = V C ) will appear, and the voltage equal point data sets groundDataVAVH, groundDataVAVC, and groundDataVHVC can be obtained.

[0053] (3) The observations of the radiometer and the cryogenic feed are respectively aligned with the corresponding variable temperature sources. Therefore, the brightness temperature values of the observations and cryogenic calibration sources with known radiation characteristics can be directly used as the "true values" of the aperture brightness temperature for each.

[0054] Preprocess / quality control / dataset partitioning / normalization of on-orbit data: Data preprocessing: Using the on-orbit observation data of radiometer cycles 003, 013, 022, and 031 as the original dataset, perform missing value processing, outlier detection, and duplicate data processing on the data.

[0055] Quality control: To obtain "stable" on-orbit observation data, the data quality control indicators are as follows: 1) Use land-sea identification to remove observation points located on land; 2) Use sea ice identification to remove observation points on sea ice-covered areas; 3) The latitude of the observation points is restricted between 60°N and 60°S; 4) Remove observation points where the liquid water content in clouds is greater than 0.1 kg / m 2 of. 5) Remove data points where the deviation between the aperture brightness temperature of the radiometer and the simulated brightness temperature is greater than 20 K.

[0056] Dataset partitioning: Randomly partition the data after quality control into a training set and a test set according to a ratio of 2:1. Among them, the training set data and the thermal vacuum test data are used to solve the calibration model parameters. The test set is used to test the solved calibration model parameters.

[0057] Normalization processing: Use the Min-Max method to normalize the input and output data.

[0058] Observed brightness temperature modeling based on neural network: Input and output of the model: The input of the network contains 4 nodes: T A 、T A 2 、Tref (physical temperature of the reflector) and latitude lat. The output of the network is the predicted value of the radiometer observed brightness temperature. In this application, the fifth-generation reanalysis data product ERA-5 provided by ECMWF is used to simulate the along-track brightness temperature of the radiometer, and it is used as the "true value" of the MWR observed brightness temperature. Among them, the atmospheric millimeter-wave radiation transfer model MPM93 is used to calculate the atmospheric coefficients, and the fast microwave emissivity model FASTEM-5 is used to calculate the ocean surface emissivity.

[0059] Model parameter settings: In this application, a feedforward neural network is used to fit the MWR observation data. The relevant parameter settings are as follows: a 4-layer neural network (1 input layer, 2 hidden layers, and 1 output layer); the number of nodes in each layer is 4, 6, 3, and 1 respectively; the activation functions are: tansig (hyperbolic tangent), tansig, and linear; the initial weights and biases of the network are random numbers between 0 and 1.

[0060] Alternating optimization solution of calibration parameters: Parameter initialization and upper and lower limit settings: A bright temperature radiation transfer model is established for the physical parameters (insertion loss, voltage reflection coefficient at the input and output ends) on the MWR observation branch and the cryogenic calibration branch, and the upper and lower limits of the physical model parameters ( a 1 、 a 2 、 a 3 、 c 1 、 c 2 、 c 3 、 u ) are calculated. The number of alternating optimizations is set to 20, and the convergence threshold is set to 1e-4.

[0061] Stage 1: Fix the neural network model parameters, and optimize the physical model parameters successively through the whale optimization algorithm and the nonlinear optimization algorithm with constraints (the corresponding approximate equality constraints that the voltage equal point data set needs to satisfy), so that the loss value of the ground stage objective function groundStageObjective is minimized.

[0062] (17) Among them, i represents the data points collected during the process of the observed variable temperature source changing from 90K to 330K when the cryogenic calibration source is maintained at 90K (T A curve data points); j represents the data points collected during the process of the cryogenic calibration source changing from 90K to 330K when the observed variable temperature source is maintained at 90K (T C curve data points).

[0063] Stage 2: Fix the physical model parameters, generate using the latest physical model parameters, and train the parameters (weights and biases) of the neural network using the whale optimization algorithm and the LM local optimization algorithm, so that the observed bright temperature value predicted by the radiometer approaches the "true value".

[0064] Stage 3: Calculate the total residuals of Stage 1 and Stage 2 and check the convergence conditions.

[0065] Repeat steps 1 to 3 until the total residual converges or the maximum number of iterations is reached.

[0066] Model verification: Use the optimized parameters to construct the final output model (the aperture brightness temperature and the observed brightness temperature of the output MWR), and test it on an independent test set to evaluate the accuracy and generalization ability of the model.

[0067] From Figure 3 The probability density distribution curves of the observed brightness temperature deviation OMB of the MWR 18.7 GHz channel obtained by the pure physical model (green) and the observed brightness temperature TbMix (red) obtained by the hybrid model (physical + neural network) and the true value of the observed brightness temperature (simulated brightness temperature) are given. It can be seen from the figure that the average deviation of OMB after calibration by the hybrid model is smaller.

[0068] To quantitatively evaluate the observed brightness temperature after on-orbit calibration of MWR, the statistical results of the OMB deviation of the MWR 18.7 GHz channel in the training set and the test set are listed in Table 1. Among them, the mean absolute deviation MAD, the standard deviation STD, the root mean square error RMSE, and the goodness of fit R 2 are used to measure the average difference, the difference dispersion, the accuracy, and the fitting degree between the predicted value of the MWR observed brightness temperature and the along-track simulated brightness temperature T B,simu respectively. It can be seen from Table 1 that the OMB deviation of the MWR 18.7 GHz channel has decreased after calibration by the hybrid model, indicating that the systematic deviation between the observed brightness temperature and the simulated brightness temperature of MWR can be eliminated after on-orbit calibration by the hybrid model. After on-orbit hybrid model calibration, the MAD, STD, and RMSE values of OMB in the test set of the 18.7 GHz channel are 1.16 K, 1.49 K, and 1.48 K respectively, which are 31.4%, 3.2%, and 27.8% lower than the MAD, STD, and RMSE values of the OMB deviation between the observed brightness temperature and the simulated brightness temperature of MWR after calibration by the pure physical model. The above results show that the on-orbit calibration algorithm can improve the calibration accuracy and stability of the MWR on-orbit observed brightness temperature.

[0069] Table 1: Statistical results of OMB of the 18.7 GHz channel of MWR after on-orbit calibration in the training set and the test set (K) (TbPhy represents the observed brightness temperature after calibration by the pure physical model; TbMix represents the observed brightness temperature after calibration by the hybrid model of physical and neural network models)

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them. Although the present application has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present application does not depart from the spirit and scope of the technical solutions of the present application, and they should all be covered within the scope of the claims of the present application.

Claims

1. An on-orbit joint calibration method for a spaceborne microwave radiometer based on the fusion of a physical model and a neural network, comprising: Step 1: Use the physical model to model the brightness temperature of the microwave radiometer aperture; Step 2: Use neural network to model the brightness temperature observed by microwave radiometer; Step 3: Alternately optimize the parameters of the aperture brightness temperature model and the observation brightness temperature model until the set optimization conditions are met.

2. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 1 is characterized in that: The step 1 comprises: Establish the microwave radiometer aperture brightness temperature model: ； ； in, and They represent the brightness temperature of the aperture predicted by the observation branch and the low-temperature calibration branch respectively; a 1 , a 2 , a 3 , c 1 , c 2 , c 3 and u are the physical model parameters to be optimized; T aw and T cw denote the physical temperature of the waveguide in the observation branch and the cryogenic calibration branch respectively; T as and T cs They represent the physical temperatures of the switches in the observation branch and the low-temperature calibration branch respectively; Q A Represents the nonlinear error of the aperture brightness temperature; V A , V H and V C Respectively represent the output voltage of the receiver on the observation branch and the high and low temperature calibration branches; represents the equivalent input brightness temperature of the high temperature calibration reference source; represents the equivalent input brightness temperature of the low temperature calibration reference source; represents the equivalent input brightness temperature of the observation branch; Q C Indicates the nonlinear error of the brightness temperature of the aperture in the low temperature calibration branch; A ground thermal vacuum test was conducted, and multiple working temperature test points of microwave radiometers were set. At each working temperature test point, when the low-temperature calibration variable temperature source was stable at the set temperature, the observation variable temperature source was heated or cooled with a step length of S1 in the range of T1 K to T2 K, and the data set groundDataTa was obtained; when the observation variable temperature source was stable to the set temperature, the low-temperature calibration variable temperature source was heated or cooled with a step length of S2 in the range of T1 K to T2 K, and the data set groundDataTc was obtained; at the same time, the temperature measurement data of the observation variable temperature source and the high and low temperature variable temperature sources, the corresponding output voltages, and each attenuation component on each transmission branch were collected; when V A =V H 、V A =V C and V H =V C The corresponding data sets groundDataVAVH, groundDataVAVC and groundDataVHVC are obtained respectively.

3. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 2 is characterized in that: The step 3 comprises: Step 3-1: Initialize the parameters of the aperture brightness temperature model and the observation brightness temperature model and set the upper and lower limits of each parameter; Step 3-2: Alternately optimize the parameters of the surface brightness temperature model and the observation brightness temperature model, including: Step 3-2-1: Fix the parameters of the observation brightness temperature model and optimize the parameters of the aperture brightness temperature model; Step 3-2-2: Fix the parameters of the aperture brightness temperature model and optimize the parameters of the observation brightness temperature model; Step 3-2-3: Calculate whether the total residual of the aperture brightness temperature model and the observed brightness temperature model meets the set stop optimization condition. If not, repeat steps 3-2-1 and 3-2-2 until the set stop optimization condition is met.

4. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 3 is characterized in that: The step 3-2-1 includes: Define the objective function groundStageObjective, combine the ground thermal vacuum test data and the on-orbit observation data, adjust the aperture brightness temperature model parameters, and minimize the joint residual groundStageLoss of the ground and on-orbit data; The objective function groundStageObjective is expressed as: ； in, m It indicates the amount of data in the data set groundDataTa obtained when the low temperature calibration variable temperature source is stable at the set temperature and the variable temperature source is observed to be heated or cooled in the range of T1 K to T2 K with a step length of S1; n It indicates the amount of data in the data set groundDataTc obtained when the observed variable temperature source is stable at the set temperature and the low temperature calibration variable temperature source is heated or cooled in the range of T1 K to T2 K with a step size S2; Indicates i The true value of the brightness temperature of the observation branch; Indicates i The brightness temperature of the mouth predicted by each observation branch; Indicates j The true value of the brightness temperature of the mouth of the low temperature calibration branch; Indicates j The brightness temperature of the mouth predicted by the low temperature calibration branch; The joint residual groundStageLoss is expressed as: ； in, Loss groundDataTa and Loss groundDataTc They represent T A Curve and T C The loss function of the curve; Loss orbitTrain represents the loss function of the on-orbit training set data; The global optimization algorithm is used to globally optimize the parameters of the aperture brightness temperature model; An optimization algorithm with strong constraints is used to locally optimize the parameters of the surface brightness temperature model: For the groundDataVAVH, groundDataVAVC, and groundDataVHVC datasets, the following strong constraints need to be met respectively: a 1 × T A + a 2 × T aw ≈ (1 - a 3 )× T as ； a 1 × T A + a 2 × T aw + a 3 × T as ≈ c 1 × T C + c 2 × T cw + c 3 × T cs ； c 1 × T C + c 2 × T cw ≈ (1 - c 3 )× T cs ； in, T A Indicates the oral surface brightness temperature; T C Indicates the temperature of the cryogenic calibration source.

5. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 3 is characterized in that: The step 3-2-2 includes: Define the objective function nnStageObjective, and use the latest aperture brightness temperature model parameters to generate the aperture brightness temperature prediction value of the observation antenna , combining the observed brightness temperature model parameters and on-orbit observation data, calculate the joint residual nnStageLoss of ground and on-orbit data; The global optimization algorithm is used to train the observation brightness temperature model parameters so that the observed brightness temperature prediction value is close to the true value; the local optimization algorithm is used to locally optimize the observation brightness temperature model parameters; The objective function nnStageObjective is expressed as: ； in, L represents the amount of observation data in the on-orbit training set; represents the true value of brightness temperature observed on orbit; Represents the brightness temperature prediction value of the on-orbit radiometer observation.

6. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 3 is characterized in that: The step 3-2-3 includes: The total residual of the aperture brightness temperature model and the observed brightness temperature model is calculated, and the optimization ends when the set number of cycles is reached or the convergence threshold of the total residual reaches the set threshold.

7. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 1, characterized in that: The step 2 comprises: The on-orbit observation data of the microwave radiometer is used as the original data set, and the original data set is quality controlled and normalized to obtain the test data; Divide the experimental data set into training set and test set; Choose a feedforward neural network, a convolutional neural network, or a recurrent neural network and their variants to model the observed brightness temperature.

8. The on-orbit joint calibration method of a spaceborne microwave radiometer based on the fusion of a physical model and a neural network according to claim 1, characterized in that: Also includes: Step 4: Use the optimized parameters to construct the final aperture brightness temperature model and observation brightness temperature model, test them on an independent test set, and evaluate the accuracy and generalization ability of the model.

Images