Satellite radiance data bias correction method and system based on constraint optimization

Abstract

The application discloses a satellite emissivity data bias correction method and system based on constraint optimization, which comprises the following steps: acquiring satellite emissivity observation data and corresponding numerical mode background field simulation values, and constructing observation minus background field data set; separating numerical prediction mode bias and observation bias based on unbiased reference observation and Jacobian matrix of the radiation transmission mode, estimating prior bias values of each satellite instrument channel, and constructing a configurable prediction factor set; the prediction factor set comprises air mass dependent prediction factor, scanning position polynomial fitting factor and scanning angle polynomial fitting factor; on the basis of basic quality control, statistical abnormal observation is removed by applying a double-weight robust estimation method, and a target function based on prior constraint is constructed; the relative weights of observation fitting terms and prior constraint terms are controlled through a regularization parameter, optimal bias correction coefficients are obtained by solving a linear equation set, a bias correction coefficient file is output, and a revised effect evaluation is generated.

Description

Technical Field

[0001] This invention relates to the field of satellite radiance data bias correction, and more particularly to a method and system for satellite radiance data bias correction based on constraint optimization. Background Technology

[0002] In numerical weather prediction, the assimilation of satellite radiance data is crucial. However, systematic biases exist in observational data, radiative transfer models, and the background field of numerical models, including instrument calibration errors and model uncertainties. Therefore, it is usually necessary to perform parameterized corrections to satellite observation biases before or during assimilation.

[0003] Currently widely adopted adaptive variational bias correction methods jointly optimize the bias coefficients as control variables during the assimilation process, relying on unbiased anchored observations to separate the model from the observation bias. However, this method has significant limitations: in regions with insufficient anchored observations or significant numerical weather prediction model bias (such as the mid-to-upper stratosphere), it is difficult to effectively distinguish between the two types of bias, causing the observation bias correction value to drift towards the numerical weather prediction model bias, severely affecting the quality of the analysis field and the forecast performance. Furthermore, this method is deeply coupled with a specific variational assimilation system, resulting in poor code portability and reusability, and a high barrier to entry for users without a complete assimilation system.

[0004] To address bias drift, some operational systems opt to disable bias correction for channels heavily influenced by numerical weather prediction model biases. However, this ignores the true bias structure, such as scan angle and air mass dependence, and lacks systematic theoretical support. Therefore, there is an urgent need to develop a new method that can suppress bias drift, effectively correct true observation biases, and possess high flexibility and portability to adapt to the needs of different satellite instruments and operational scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for correcting satellite radiance data bias based on constraint optimization.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0007] This invention includes the following steps:

[0008] Acquire satellite radiance observation data and corresponding numerical model background field simulation values, and construct an observation-subtracted background field dataset;

[0009] Based on the Jacobian matrix of unbiased reference observations combined with the radiative transfer model, the numerical prediction model bias and observation bias are separated, the prior bias values ​​of each satellite instrument channel are estimated, and a set of configurable prediction factors is constructed. The set of prediction factors includes air mass-dependent prediction factors, scan position polynomial fitting factors, and scan angle polynomial fitting factors.

[0010] Based on basic quality control, a dual-weight robust estimation method is applied to eliminate statistically abnormal observations and construct an objective function based on prior constraints.

[0011] The relative weights of the observation fitting term and the prior constraint term are controlled by the regularization parameter. The optimal deviation correction coefficient is obtained by solving the linear equation system, the deviation correction coefficient file is output, and an evaluation of the correction effect is generated.

[0012] Furthermore, the method for constructing the observation-subtracted background field dataset includes:

[0013] Brightness temperature was simulated using the fast radiative transfer model based on the forecast field of the numerical weather prediction model. The numerical model provides atmospheric profiles including temperature, humidity, and ozone, as well as surface parameters, as inputs to the radiative transfer model.

[0014] The satellite observations are spatiotemporally matched with the model background field to calculate the observation-to-background field residual; at the same time, the forecast factor values ​​required for bias correction are extracted, including air mass parameters such as atmospheric thickness.

[0015] The observation background field dataset is organized according to satellite instruments and channels and stored in a standardized data format.

[0016] Furthermore, the method for estimating the prior bias value includes:

[0017] By combining unbiased reference observations with the Jacobian matrix of the radiative transfer model, the numerical prediction model bias and observation bias are separated, and the prior bias values ​​of each satellite instrument channel are estimated.

[0018] Observational data with independent calibration systems and reliable accuracy are selected as unbiased references. The atmospheric profiles of the unbiased reference observations are spatiotemporally matched with the background field of the numerical model. The matching criteria include time windows and spatial distance thresholds. The unbiased reference observations need to meet the following conditions: they are independent of the satellite observations to be corrected, have sufficient accuracy in the target atmospheric region, and have reasonable spatiotemporal coverage.

[0019] Calculate the deviation between the atmospheric profile of the numerical model and the profile of the unbiased reference observation:

[0020] ;

[0021] in This is the atmospheric profile for the numerical model. For unbiased reference observations of the atmospheric profile, The numerical forecast model bias for each pressure layer;

[0022] The numerical prediction model bias in the state space is mapped to the brightness-temperature bias in the observation space using the Jacobian matrix of the radiative transfer model. The expression is as follows:

[0023] ;

[0024] in The Jacobian matrix of the radiative transfer model with respect to the atmospheric profile represents the sensitivity of atmospheric state changes at each pressure layer to the channel brightness temperature. Numerical prediction model bias mapped to the observation space, This is the statistical average of the differences between the atmospheric profiles from the numerical model and the atmospheric profiles from the unbiased reference observations.

[0025] Based on the assumption that the observed subtracted background mean includes both observation bias and numerical weather prediction model bias, the observation bias is separated as a prior bias estimate by subtracting the numerical weather prediction model bias. The expression is as follows:

[0026] ;

[0027] in To observe the statistical mean of the background field, This is an estimate of the separated observation bias, i.e., the prior bias;

[0028] Perform the above steps on the temperature detection channels of each satellite instrument to generate the prior bias value for each channel. ;

[0029] Assess the uncertainty of the prior bias estimate and set the regularization parameter. The sources of uncertainty include occultation observation errors, radiative transfer mode errors, and representativeness errors of the matched samples.

[0030] Furthermore, the method for constructing a configurable set of forecast factors includes:

[0031] Satellite radiance data bias involves a variety of air mass forecasting factors related to atmospheric conditions, including atmospheric thickness between different pressure layers, surface temperature, total water vapor content of the atmospheric column, 10-meter wind speed, and the path of liquid water in the atmospheric column cloud.

[0032] Using linear combination form for channels deviation Parameterization is performed, and the expression is:

[0033] ;

[0034] in For channel The constant deviation term, The number of air mass forecasting factors, For satellite scanning angle, For scan points, For channel The air mass forecast factor for the i-th bias term, For channel No. m The coefficients of the scan point polynomial of order 1, For scan points m Power of 1 For satellite scan angle l Power of 1 For channel No. l coefficients of the scan point polynomial Let be the order of the scan point polynomial. Let be the order of the scan angle polynomial;

[0035] Forecast factors are configured according to channel characteristics: the selection of forecast factors is matched with the weight function characteristics of the channel, and the forecast factor combination is configured independently for each channel of each instrument;

[0036] All forecast factors are standardized, and the expression is:

[0037] ;

[0038] Where p is the current value of the forecast factor, The mean of the forecast factors, denoted as the standard deviation of the forecast factor.

[0039] Furthermore, the quality control method includes:

[0040] Conventional quality control methods were used to remove observation samples affected by non-atmospheric factors. Various anomalies were marked by quality control flags. Only observation samples that passed all checks were retained and used for subsequent deviation correction coefficient calculations. Conventional quality control methods included cloud detection, precipitation detection, land surface type identification, scan edge inspection, and gross error inspection.

[0041] A two-weighted robust estimation method is used for anomaly detection, with the median as the location estimate and the median absolute deviation as the scaling estimate. The standardized score for each sample is calculated.

[0042] ;

[0043] in The median of the sample. This represents the absolute deviation of the median. To adjust the parameters, To subtract the background field from the observation of the k-th sample, The standardized score of the k-th sample;

[0044] when When the threshold is exceeded, the observation is identified as an outlier and removed.

[0045] Furthermore, the method for constructing an objective function based on prior constraints includes:

[0046] Construct the objective function, with the expression:

[0047] ;

[0048] in For the observation value of the k-th sample, This represents the k-th simulated observation corresponding to the model's background field. For the sample size, This is the observation fitting term, which measures the consistency between the bias-corrected observations and the background field. As a priori constraint, the constraint bias correction value should not deviate too far from the prior estimate; For prior estimation of satellite instrument bias, This is a regularization parameter used to control prior bias. posterior bias The relative importance between them; The posterior bias estimate for the k-th sample is calculated by minimizing the objective function based on the current cumulative sample.

[0049] When unbiased reference observations are lacking, if the satellite instruments themselves are well calibrated and have small observational biases, then... Setting it to 0 limits the absolute magnitude of the deviation correction value; when When the value is large, the bias correction trust prior estimate The posterior bias will approach the prior estimate. ;when When the value is small, the statistical characteristics of the current observed sample are trusted, and the posterior bias is mainly determined by the observation minus the background field statistics; when When the value is 0, the constraint term becomes invalid, and the constrained deviation correction method degenerates into the traditional unconstrained deviation correction method.

[0050] Furthermore, the method for obtaining the optimal deviation correction coefficient by solving a system of linear equations includes:

[0051] Substitute the deviation parameterization model into the objective function, calculate the partial derivatives of each deviation correction coefficient, and set the derivatives to zero according to the extremum condition to derive a system of linear equations about the coefficients.

[0052] Let the forecast factor matrix be... Let the observation bias vector be... The coefficient vector to be determined is The predictor factor matrix has a dimension of N×n, where N is the sample size and n is the number of predictor factors. Each column of the matrix corresponds to the standardized observation value of each factor.

[0053] The system of linear equations is transformed into normal equations in matrix form, expressed as:

[0054] ;

[0055] in Characterizing the autocorrelation of forecast factors, Characterize the cross-correlation between forecast factors and observation bias;

[0056] By analyzing the matrix By inverting the equation and multiplying it on the left by the right-hand side, we obtain the analytical solution formula for the optimal deviation correction coefficient:

[0057] ;

[0058] in The optimal deviation correction coefficient is output.

[0059] Secondly, a constrained optimization-based satellite radiance data bias correction system includes:

[0060] Dataset acquisition and construction module: used to acquire satellite radiance observation data and corresponding numerical model background field simulation values, and construct the observation subtraction background field dataset;

[0061] Bias separation and forecast factor construction module: used to separate numerical forecast model bias from observation bias based on unbiased reference observations combined with the Jacobian matrix of the radiative transfer model, estimate the prior bias values ​​of each satellite instrument channel, and construct a configurable set of forecast factors; the set of forecast factors includes air mass-dependent forecast factors, scan position polynomial fitting factors, and scan angle polynomial fitting factors.

[0062] Quality control and function construction module: This module is used to apply a two-weighted robust estimation method to eliminate statistically abnormal observations and construct an objective function based on prior constraints, on the basis of basic quality control.

[0063] The control weights and solution module is used to control the relative weights of the observation fitting term and the prior constraint term through regularization parameters, obtain the optimal deviation correction coefficients by solving the linear equation system, output the deviation correction coefficients file, and generate an evaluation of the correction effect.

[0064] The beneficial effects of this invention are:

[0065] This invention is a constrained optimization-based method and system for correcting satellite radiance data bias. Compared with existing technologies, this invention has the following technical advantages:

[0066] This invention effectively prevents bias correction values ​​from drifting into numerical weather prediction model bias by introducing prior constraints based on satellite instrument calibration uncertainties. This overcomes the problem of observation bias corrections drifting into numerical weather prediction model bias in areas with large numerical weather prediction model bias and insufficient anchored observations, which is a problem of traditional variational bias correction methods. At the same time, it adopts a flexible regularization parameter design, which can adjust the constraint strength according to the characteristics of different channels, thus preventing overcorrection while retaining the ability to identify the true observation bias structure. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating the steps of a constrained optimization-based satellite radiance data bias correction method and system of the present invention.

[0068] Figure 2 This is a schematic diagram of prior bias estimated based on inversion data in an embodiment of the present invention;

[0069] Figure 3 This is a comparison chart showing the evolution of bias correction values ​​over time in the cyclic assimilation experiment of this invention. Detailed Implementation

[0070] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0071] This invention discloses a satellite radiance data bias correction method and system based on constraint optimization, comprising the following steps:

[0072] like Figure 1 As shown, this embodiment includes the following steps:

[0073] Acquire satellite radiance observation data and corresponding numerical model background field simulation values, and construct an observation-subtracted background field dataset;

[0074] Based on the Jacobian matrix of unbiased reference observations combined with the radiative transfer model, the numerical prediction model bias and observation bias are separated, the prior bias values ​​of each satellite instrument channel are estimated, and a set of configurable prediction factors is constructed. The set of prediction factors includes air mass-dependent prediction factors, scan position polynomial fitting factors, and scan angle polynomial fitting factors.

[0075] Based on basic quality control, a dual-weight robust estimation method is applied to eliminate statistically abnormal observations and construct an objective function based on prior constraints.

[0076] The relative weights of the observation fitting term and the prior constraint term are controlled by the regularization parameter. The optimal deviation correction coefficient is obtained by solving the linear equation system, the deviation correction coefficient file is output, and an evaluation of the correction effect is generated.

[0077] In practical evaluations, AMSU-A is a microwave temperature sounding instrument carried on NOAA series polar-orbiting meteorological satellites and MetOp series satellites, with a total of 15 channels. It is mainly used for atmospheric temperature profile detection, with channels 9-14 being sensitive to the temperature of the upper and middle stratosphere. ATMS is a new generation microwave sounding instrument carried on US Suomi-NPP and NOAA-20 series satellites, with a total of 22 channels, covering the functions of AMSU-A and MHS. MWTS-3 is a microwave temperature sounding instrument carried on China's Fengyun-3 E / F / G satellites, with a total of 17 channels, and its performance is comparable to ATMS. GPS is the global satellite navigation system established by the United States. COSMIC-2 is a GPS radio occultation observation constellation composed of 6 low-Earth orbit satellites. By receiving information on the refraction and delay of GPS signals as they pass through the atmosphere, it retrieves high-precision atmospheric temperature and humidity profiles, providing good coverage in tropical regions.

[0078] In this embodiment, the method for constructing the observation-subtracted background field dataset includes:

[0079] Brightness temperature was simulated using the fast radiative transfer model based on the forecast field of the numerical weather prediction model. The numerical weather prediction model provides atmospheric profiles including temperature, humidity, and ozone, as well as surface parameters, as inputs to the radiative transfer model.

[0080] The satellite observations are spatiotemporally matched with the model background field to calculate the observation-to-background field residual; at the same time, the forecast factor values ​​required for bias correction are extracted, including air mass parameters such as atmospheric thickness.

[0081] The background field data set is organized according to satellite instruments and channels and stored in a standardized data format;

[0082] In actual evaluation, taking spaceborne passive microwave detectors such as AMSU-A, MWTS-3, and ATMS as examples, satellite radiance observation data includes brightness temperature observation values ​​for each channel, as well as auxiliary information such as observation location, observation time, scanning location, and satellite zenith angle.

[0083] In this embodiment, the method for estimating the prior bias value includes:

[0084] By combining unbiased reference observations with the Jacobian matrix of the radiative transfer model, the numerical prediction model bias and observation bias are separated, and the prior bias values ​​of each satellite instrument channel are estimated.

[0085] Observational data with independent calibration systems and reliable accuracy are selected as unbiased references. The atmospheric profiles of the unbiased reference observations are spatiotemporally matched with the background field of the numerical model. The matching criteria include time windows and spatial distance thresholds. The unbiased reference observations need to meet the following conditions: they are independent of the satellite observations to be corrected, have sufficient accuracy in the target atmospheric region, and have reasonable spatiotemporal coverage.

[0086] Calculate the deviation between the atmospheric profile of the numerical model and the profile of the unbiased reference observation:

[0087]

[0088] in This is the atmospheric profile for the numerical model. For unbiased reference observations of the atmospheric profile, The numerical forecast model bias for each pressure layer;

[0089] The numerical prediction model bias in the state space is mapped to the brightness-temperature bias in the observation space using the Jacobian matrix of the radiative transfer model. The expression is as follows:

[0090]

[0091] in The Jacobian matrix of the radiative transfer model with respect to the atmospheric profile represents the sensitivity of atmospheric state changes at each pressure layer to the channel brightness temperature. Numerical prediction model bias mapped to the observation space, This is the statistical average of the differences between the atmospheric profiles from the numerical model and the atmospheric profiles from the unbiased reference observations.

[0092] Based on the assumption that the observed subtracted background mean includes both observation bias and numerical weather prediction model bias, the observation bias is separated as a prior bias estimate by subtracting the numerical weather prediction model bias. The expression is as follows:

[0093]

[0094] in To observe the statistical mean of the background field, This is an estimate of the separated observation bias, i.e., the prior bias;

[0095] Perform the above steps on the temperature detection channels of each satellite instrument to generate the prior bias value for each channel. ;

[0096] Assess the uncertainty of the prior bias estimate and set the regularization parameter. The sources of uncertainty include occultation observation errors, radiative transfer mode errors, and representativeness errors of the matched samples.

[0097] In actual evaluation, GPS radio occultation observation is used as an example. Occultation observation has the characteristics of high vertical resolution, self-calibration and global coverage. Its temperature profile inversion has high accuracy from the upper troposphere to the middle stratosphere. Occultation profile data such as COSMIC-2 are selected, and the matching criteria are set to a time window of ±1 hour and a spatial distance of within 50 kilometers.

[0098] For the upper stratospheric temperature channel, the numerical prediction model has a large bias, but this method can effectively separate out the relatively small observation bias.

[0099] In this embodiment, the method for constructing a configurable set of forecast factors includes:

[0100] Satellite radiance data bias involves a variety of air mass forecasting factors related to atmospheric conditions, including atmospheric thickness between different pressure layers, surface temperature, total water vapor content of the atmospheric column, 10-meter wind speed, and the path of liquid water in the atmospheric column cloud.

[0101] Using linear combination form for channels deviation Parameterization is performed, and the expression is:

[0102]

[0103] in For channel The constant deviation term, The number of air mass forecasting factors, For satellite scanning angle, For scan points, For channel The air mass forecast factor for the i-th bias term, For channel No. m The coefficients of the scan point polynomial of order 1, For scan points m Power of 1 For satellite scan angle l Power of 1 For channel No. l The coefficients of the scan point polynomial of order 1, Let be the order of the scan point polynomial. Let be the order of the scan angle polynomial;

[0104] Forecast factors are configured according to channel characteristics: the selection of forecast factors is matched with the weight function characteristics of the channel, and the forecast factor combination is configured independently for each channel of each instrument;

[0105] All forecast factors are standardized, and the expression is:

[0106]

[0107] in The current value of the forecast factor. The mean of the forecast factors, The standard deviation of the forecast factor;

[0108] In actual evaluation, users can add other forecast factors in the configuration file according to business needs. The standard deviation of the forecast factors can be obtained from historical sample statistics or specified in the configuration file.

[0109] Multiple types of forecast factors are designed to parameterize the physical structure of the bias and support flexible configuration by instrument and by channel; atmospheric thickness includes 1000-300hPa, 200-50hPa, and 50-5hPa; the total water vapor content of the atmospheric column is the precipitable water.

[0110] Configure forecast factors according to channel characteristics: Taking AMSU-A as an example, the peak of the weight function of the lower channel is located in the lower troposphere, with high atmospheric transmittance and sensitivity to surface conditions. It is appropriate to configure forecast factors such as surface temperature and surface wind speed. The peak of the weight function of the upper channel is located in the stratosphere and is not sensitive to the surface. It is appropriate to configure the atmospheric thickness of the pressure layer corresponding to its weight function as a forecast factor.

[0111] Used to correct atmospheric state-dependent biases; Used to eliminate asymmetric deviations along the scan line direction; To correct for radiative transfer bias caused by changes in the observation angle tilt, higher-order terms for the scanning point and scanning angle are introduced to more precisely fit the nonlinear systematic bias in the instrument's scanning geometry. All coefficients in the formula... , , , These are all undetermined parameters that will be solved in subsequent steps;

[0112] In this embodiment, the quality control method includes:

[0113] Conventional quality control methods were used to remove observation samples affected by non-atmospheric factors. Various anomalies were marked by quality control flags. Only observation samples that passed all checks were retained and used for subsequent deviation correction coefficient calculations. Conventional quality control methods included cloud detection, precipitation detection, land surface type identification, scan edge inspection, and gross error inspection.

[0114] A two-weighted robust estimation method is used for anomaly detection, with the median as the location estimate and the median absolute deviation as the scaling estimate. The standardized score for each sample is calculated.

[0115]

[0116] in The median of the sample. This represents the absolute deviation of the median. To adjust the parameters, To subtract the background field from the observation of the k-th sample, The standardized score of the k-th sample;

[0117] when When the threshold is exceeded, the observation is identified as an outlier and removed.

[0118] In the actual assessment, basic quality control and a two-weighted robust estimation method are used to eliminate observations and statistical outliers contaminated by factors such as clouds, precipitation, and surface conditions, ensuring that the sample used for the calculation of the bias correction coefficient has good representativeness.

[0119] Since the median and median absolute deviation are naturally resistant to systematic bias, quality control can accurately identify true outliers even when bias exists, thus improving the robustness of bias estimation.

[0120] In this embodiment, the method for constructing an objective function based on prior constraints includes:

[0121] Construct the objective function, with the expression:

[0122]

[0123] in For the observation value of the k-th sample, This represents the k-th simulated observation corresponding to the background field of the numerical weather prediction model. For the sample size, This is the observation fitting term, which measures the consistency between the bias-corrected observations and the background field. As a priori constraint, the constraint bias correction value should not deviate too far from the prior estimate; For prior estimation of satellite instrument bias, This is a regularization parameter used to control prior bias. posterior bias The relative importance between them; The posterior bias estimate for the k-th sample is calculated by minimizing the objective function based on the current cumulative sample.

[0124] When unbiased reference observations are lacking, if the satellite instruments themselves are well calibrated and have small observational biases, then... Setting it to 0 limits the absolute magnitude of the deviation correction value; when When the value is large, the bias correction trust prior estimate The posterior bias will approach the prior estimate. ;when When the value is small, the statistical characteristics of the current observed sample are trusted, and the posterior bias is mainly determined by the observation minus the background field statistics; when When the value is 0, the constraint term becomes invalid, and the constrained deviation correction method degenerates into the traditional unconstrained deviation correction method.

[0125] In actual evaluation, the posterior bias estimate is calculated by minimizing the objective function based on the current cumulative sample. The posterior bias estimate integrates observational statistics and prior constraint information.

[0126] In this embodiment, the method for obtaining the optimal deviation correction coefficient by solving a system of linear equations includes:

[0127] Substitute the deviation parameterization model into the objective function, calculate the partial derivatives of each deviation correction coefficient, and set the derivatives to zero according to the extremum condition to derive a system of linear equations about the coefficients.

[0128] Let the forecast factor matrix be... Let the observation bias vector be... The coefficient vector to be determined is The predictor factor matrix has a dimension of N×n, where N is the sample size and n is the number of predictor factors. Each column of the matrix corresponds to the standardized observation value of each factor.

[0129] The system of linear equations is transformed into normal equations in matrix form, expressed as:

[0130]

[0131] in Characterizing the autocorrelation of forecast factors, Characterize the cross-correlation between forecast factors and observation bias;

[0132] By analyzing the matrix By inverting the equation and multiplying it on the left by the right-hand side, we obtain the analytical solution formula for the optimal deviation correction coefficient:

[0133]

[0134] in The optimal deviation correction coefficient is output.

[0135] In practical evaluation, the obtained bias correction coefficients are output and applied to the data assimilation system, and the bias correction coefficients of each channel are used. Standardized parameters of forecast factors (mean) and standard deviation ), and prior bias The information is output as a coefficient file in a standard format.

[0136] Secondly, a constrained optimization-based satellite radiance data bias correction system includes:

[0137] Dataset acquisition and construction module: used to acquire satellite radiance observation data and corresponding numerical model background field simulation values, and construct the observation subtraction background field dataset;

[0138] Bias separation and forecast factor construction module: used to separate numerical forecast model bias from observation bias based on unbiased reference observations combined with the Jacobian matrix of the radiative transfer model, estimate the prior bias values ​​of each satellite instrument channel, and construct a configurable set of forecast factors; the set of forecast factors includes air mass-dependent forecast factors, scan position polynomial fitting factors, and scan angle polynomial fitting factors.

[0139] Quality control and function construction module: This module is used to apply a two-weighted robust estimation method to eliminate statistically abnormal observations and construct an objective function based on prior constraints, on the basis of basic quality control.

[0140] The control weights and solution module is used to control the relative weights of the observation fitting term and the prior constraint term through regularization parameters, obtain the optimal deviation correction coefficients by solving the linear equation system, output the deviation correction coefficients file, and generate an evaluation of the correction effect.

[0141] Figure 2 This is a schematic diagram of the AMSU-A instrument prior bias estimated based on COSMIC-2 occultation inversion data according to an embodiment of the present invention, wherein (a) is the bias estimate of each channel of NOAA-19 AMSU-A, and (b) is the bias estimate of each channel of NOAA-18 AMSU-A; the solid curve is the OB mean, the short dashed curve is the numerical prediction model bias estimated using occultation data, and the short dashed curve is the separated observation prior bias;

[0142] Figure 3 This is a comparison chart showing the evolution of bias correction values ​​over time in a cyclic assimilation experiment provided by an embodiment of the present invention; the short dashed curve represents the unconstrained bias correction scheme, and the solid curve represents the constrained bias correction scheme; the bias correction value of the unconstrained scheme gradually drifts towards the bias of the numerical prediction model as the assimilation cycle progresses, while the constrained scheme can effectively suppress bias drift and maintain stability.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for correcting biases in satellite radiance data based on constrained optimization, characterized in that, Includes the following steps: Acquire satellite radiance observation data and corresponding numerical model background field simulation values, and construct an observation-subtracted background field dataset; Based on the Jacobian matrix of unbiased reference observations combined with the radiative transfer model, the numerical prediction model bias and observation bias are separated, the prior bias values ​​of each satellite instrument channel are estimated, and a set of configurable prediction factors is constructed. The set of prediction factors includes air mass-dependent prediction factors, scan position polynomial fitting factors, and scan angle polynomial fitting factors. Based on basic quality control, a dual-weight robust estimation method is applied to eliminate statistically abnormal observations and construct an objective function based on prior constraints. The relative weights of the observation fitting term and the prior constraint term are controlled by the regularization parameter. The optimal deviation correction coefficient is obtained by solving the linear equation system, the deviation correction coefficient file is output, and an evaluation of the correction effect is generated.

2. The satellite radiance data bias correction method based on constraint optimization according to claim 1, characterized in that, The method for constructing the observation subtraction background field dataset includes: Brightness temperature was simulated using the fast radiative transfer model based on the forecast field of the numerical weather prediction model. The numerical weather prediction model provides atmospheric profiles including temperature, humidity, and ozone, as well as surface parameters, as inputs to the radiative transfer model. The satellite observations are spatiotemporally matched with the model background field to calculate the observation-to-background field residual; at the same time, the forecast factor values ​​required for bias correction are extracted, including air mass parameters such as atmospheric thickness. The observation background field dataset is organized according to satellite instruments and channels and stored in a standardized data format.

3. The satellite radiance data bias correction method based on constrained optimization according to claim 1, characterized in that, The method for estimating the prior bias value includes: By combining unbiased reference observations with the Jacobian matrix of the radiative transfer model, the numerical prediction model bias and observation bias are separated, and the prior bias values ​​of each satellite instrument channel are estimated. Observational data with independent calibration systems and reliable accuracy are selected as unbiased references. The atmospheric profiles of the unbiased reference observations are spatiotemporally matched with the background field of the numerical model. The matching criteria include time windows and spatial distance thresholds. The unbiased reference observations need to meet the following conditions: they are independent of the satellite observations to be corrected, have sufficient accuracy in the target atmospheric region, and have reasonable spatiotemporal coverage. Calculate the deviation between the atmospheric profile of the numerical model and the profile of the unbiased reference observation: ； in This is the atmospheric profile for the numerical model. For unbiased reference observations of the atmospheric profile, The numerical forecast model bias for each pressure layer; The numerical prediction model bias in the state space is mapped to the brightness-temperature bias in the observation space using the Jacobian matrix of the radiative transfer model. The expression is as follows: ； in The Jacobian matrix of the radiative transfer model with respect to the atmospheric profile represents the sensitivity of atmospheric state changes at each pressure layer to the channel brightness temperature. Numerical prediction model bias mapped to the observation space, This is the statistical average of the differences between the atmospheric profiles from the numerical model and the atmospheric profiles from the unbiased reference observations. Based on the assumption that the observed subtracted background mean includes both observation bias and numerical weather prediction model bias, the observation bias is separated as a prior bias estimate by subtracting the numerical weather prediction model bias. The expression is as follows: ； in To observe the statistical mean of the background field, This is an estimate of the separated observation bias, i.e., the prior bias; Perform the above steps on the temperature detection channels of each satellite instrument to generate the prior bias value for each channel. ; Assess the uncertainty of the prior bias estimate and set the regularization parameter. The sources of uncertainty include occultation observation errors, radiative transfer mode errors, and representativeness errors of the matched samples.

4. The satellite radiance data bias correction method based on constrained optimization according to claim 1, characterized in that, The method for constructing a configurable set of forecast factors includes: Satellite radiance data bias involves a variety of air mass forecasting factors related to atmospheric conditions, including atmospheric thickness between different pressure layers, surface temperature, total water vapor content of the atmospheric column, 10-meter wind speed, and the path of liquid water in the atmospheric column cloud. Using linear combination form for channels deviation Parameterization is performed, and the expression is: ； in For channel The constant deviation term, The number of air mass forecasting factors, For satellite scanning angle, For scan points, For channel The air mass forecast factor for the i-th bias term, For channel No. m The coefficients of the scan point polynomial of order 1, For scan points m Power of 1 For satellite scan angle l Power of 1 For channel No. l coefficients of the scan point polynomial Let be the order of the scan point polynomial. Let be the order of the scan angle polynomial; Forecast factors are configured according to channel characteristics: the selection of forecast factors is matched with the weight function characteristics of the channel, and the forecast factor combination is configured independently for each channel of each instrument; All forecast factors are standardized, and the expression is: ； Where p is the current value of the forecast factor, The mean of the forecast factors, denoted as the standard deviation of the forecast factor.

5. The satellite radiance data bias correction method based on constrained optimization according to claim 1, characterized in that, The quality control method includes: Conventional quality control methods were used to remove observation samples affected by non-atmospheric factors. Various anomalies were marked by quality control flags. Only observation samples that passed all checks were retained and used for subsequent deviation correction coefficient calculations. Conventional quality control methods included cloud detection, precipitation detection, land surface type identification, scan edge inspection, and gross error inspection. A two-weighted robust estimation method is used for anomaly detection, with the median as the location estimate and the median absolute deviation as the scaling estimate. The standardized score for each sample is calculated. ； in The median of the sample. This represents the absolute deviation of the median. To adjust the parameters, To subtract the background field from the observation of the k-th sample, The standardized score of the k-th sample; when When the threshold is exceeded, the observation is identified as an outlier and removed.

6. The satellite radiance data bias correction method based on constraint optimization according to claim 1, characterized in that, The method for constructing an objective function based on prior constraints includes: Construct the objective function, with the expression: ； in For the observation value of the k-th sample, This represents the k-th simulated observation corresponding to the model's background field. For the sample size, This is the observation fitting term, which measures the consistency between the bias-corrected observations and the background field. As a priori constraint, the constraint bias correction value should not deviate too far from the prior estimate; For prior estimation of satellite instrument bias, This is a regularization parameter used to control prior bias. posterior bias The relative importance between them; The posterior bias estimate for the k-th sample is calculated by minimizing the objective function based on the current cumulative sample. When unbiased reference observations are lacking, if the satellite instruments themselves are well calibrated and have small observational biases, then... Setting it to 0 limits the absolute magnitude of the deviation correction value; when When the value is large, the trust prior estimate The posterior bias will approach the prior estimate. ;when When the value is small, the statistical characteristics of the current observed sample are trusted, and the posterior bias is mainly determined by the observation minus the background field statistics; when When the value is 0, the constraint term becomes invalid, and the constrained deviation correction method degenerates into the traditional unconstrained deviation correction method.

7. The satellite radiance data bias correction method based on constrained optimization according to claim 1, characterized in that, The method for obtaining the optimal deviation correction coefficient by solving a system of linear equations includes: Substitute the deviation parameterization model into the objective function, calculate the partial derivatives of each deviation correction coefficient, and set the derivatives to zero according to the extremum condition to derive a system of linear equations about the coefficients. Let the forecast factor matrix be... Let the observation bias vector be... The coefficient vector to be determined is The predictor factor matrix has a dimension of N×n, where N is the sample size and n is the number of predictor factors. Each column of the matrix corresponds to the standardized observation value of each factor. The system of linear equations is transformed into normal equations in matrix form, expressed as: ； in Characterizing the autocorrelation of forecast factors, Characterize the cross-correlation between forecast factors and observation bias; By analyzing the matrix By inverting the equation and multiplying it on the left by the right-hand side, we obtain the analytical solution formula for the optimal deviation correction coefficient: ； in The optimal deviation correction coefficient is output.

8. A constrained optimization-based satellite radiance data bias correction system, for performing the method according to any one of claims 1-7, characterized in that, include: Dataset acquisition and construction module: used to acquire satellite radiance observation data and corresponding numerical model background field simulation values, and construct the observation subtraction background field dataset; Bias separation and forecast factor construction module: used to separate numerical forecast model bias from observation bias based on unbiased reference observations combined with the Jacobian matrix of the radiative transfer model, estimate the prior bias values ​​of each satellite instrument channel, and construct a configurable set of forecast factors; the set of forecast factors includes air mass-dependent forecast factors, scan position polynomial fitting factors, and scan angle polynomial fitting factors. Quality control and function construction module: This module is used to apply a two-weighted robust estimation method to eliminate statistically abnormal observations and construct an objective function based on prior constraints, on the basis of basic quality control. The control weights and solution module is used to control the relative weights of the observation fitting term and the prior constraint term through regularization parameters, obtain the optimal deviation correction coefficients by solving the linear equation system, output the deviation correction coefficients file, and generate an evaluation of the correction effect.

Images