A direct assimilation method for ground-based microwave radiometers

Abstract

The present invention discloses a direct assimilation method for ground-based microwave radiometers, which relates to the technical field of meteorological observations and includes: calling the RTTOV-gb fast radiative transfer model, corresponding the derived types in the WRFDA system with the derived types of RTTOV-gb, coupling RTTOV, RTTOV-gb and the WRFDA system simultaneously, and converting the background field of the numerical weather prediction model into simulated brightness temperature data; considering the source of the deviation and the non-linear relationship between the source and the deviation, calculating the deviation correction prediction factor according to the original observed brightness temperature and the background field, and performing deviation correction on the original observed brightness temperature to obtain the corrected observed brightness temperature; based on the corrected observed brightness temperature data, performing three-dimensional variational assimilation processing on the background field to obtain the analysis field and outputting the diagnostic file. The present invention can achieve a good deviation correction effect, is convenient to adjust according to the requirements of actual applications, RTTOV-gb and RTTOV can be used simultaneously, and joint direct assimilation of various ground-based microwave radiometers and spaceborne microwave radiometers can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological observation, and in particular, to a direct assimilation method for a ground-based microwave radiometer. Background Art

[0002] Numerical weather prediction relies on the efficient fusion of multi-source observation data by a data assimilation system. Among them, the ground-based microwave radiometer highlights its unique value in short-term forecasting due to its minute-level high-frequency observation ability and the advantage of finely capturing the atmospheric boundary layer. Currently, this type of data assimilation technology is mainly indirect assimilation, that is, data fusion is achieved by inverting temperature and humidity parameters, but there are three limitations: uncertainties and large errors will be introduced during the inversion process, thus limiting the effectiveness of assimilation; the inversion process makes error estimation complex, and it is difficult to quantify and process these errors through the error covariance matrix; secondary assimilation leads to repeated utilization of background information and reduces the independence of observations. In contrast, direct assimilation avoids inversion errors by processing the original radiance temperature data and has now become the preferred solution for operational applications.

[0003] The observation operator is the core component for realizing the direct assimilation of the ground-based microwave radiometer. Its function is to convert variables such as temperature and water vapor in the numerical model into radiance data equivalent to radiation observations. The direct assimilation technology of spaceborne microwave radiometers has been relatively mature, and usually, a fast radiative transfer model is used as the observation operator, such as RTTOV (Radiative Transfer for the Television Infrared Observation Satellite Operational Vertical Sounder). However, due to the essential difference in the bottom-up observation path between the ground-based microwave radiometer and the spaceborne platform, the spaceborne radiative transfer model cannot be directly applied. In recent years, the fast radiative transfer model RTTOV-gb (ground-based version of RTTOV) applicable to the ground-based microwave radiometer has been developed, providing strong support for the direct assimilation of the ground-based microwave radiometer.

[0004] However, in the prior art, the direct assimilation technology of ground-based microwave radiometer data does not perform non-linear bias correction. The assimilation method assumes that the observation error and background error follow an unbiased Gaussian distribution. However, in practice, the observation instrument error, the limitation of the radiation transfer model, and the numerical model background error may lead to a systematic bias between the observation and the simulated radiation, thus affecting the assimilation quality. Bias correction is one of the key technologies for the direct assimilation of ground-based microwave radiometers, aiming to identify and correct the bias between the observation and the background to improve the assimilation accuracy. When using variational assimilation in the prior art, bias correction is not performed, ignoring the systematic bias between the observation and the background; when using ensemble Kalman filtering for assimilation in the prior art, linear bias correction is adopted, but the source of the bias and the non-linear bias component are not considered.

[0005] The terrain and landform in the southwest region are complex. Affected by large terrains such as plateaus and the Hengduan Mountains, the model background error often has biases and is very complex. In addition, the applicability of RTTOV-gb has limitations in these areas because its training data mainly comes from global profile data rather than being specifically for the plateau region. Therefore, when performing direct assimilation of ground-based microwave radiometers in these areas, there may be complex non-linear biases between the observation and the simulation. If appropriate bias correction is not performed during direct assimilation, it will greatly limit the assimilation effect and reduce the benefit of the observation. Summary of the Invention

[0006] Aiming at complex terrains such as plateaus and the Hengduan Mountains, the present invention provides a direct assimilation method for ground-based microwave radiometers, aiming to effectively correct the bias of OMB (observation minus background, the observed brightness temperature minus the simulated brightness temperature) through bias correction. Considering the observation characteristics of ground-based microwave radiometers, the source of the bias (the predictor of bias correction) and the non-linear relationship between the source and the bias are considered, and the contribution of each predictor in the correction process is evaluated.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A direct assimilation method for ground-based microwave radiometers, comprising:

[0009] Reading the original radiation observation data of the ground-based microwave radiometer and performing preprocessing to obtain the original observed brightness temperature data;

[0010] Call the RTTOV-gb fast radiative transfer model, match the derived types in the WRFDA (weather research and forecasting model data assimilation system) system with the RTTOV-gb derived types, so that the RTTOV, RTTOV-gb and WRFDA systems are coupled simultaneously, and then the background field of the numerical weather forecast model is converted into simulated brightness temperature data;

[0011] Based on the machine learning algorithm, the source of the deviation and the nonlinear relationship between the source and the deviation are considered. The deviation correction prediction factor is calculated according to the original observed brightness temperature data and the background field. The deviation of the original observed brightness temperature data is corrected based on the deviation correction prediction factor to obtain the corrected observed brightness temperature data.

[0012] Calculate observational residual data based on simulated brightness temperature data and corrected observed brightness temperature data, and select corrected observed brightness temperature data required for assimilation based on observed residual data;

[0013] Using the assimilation module of the WRFDA system, the background field is subjected to three-dimensional variational assimilation according to the screened and corrected observed brightness temperature data. The analysis field is obtained by iterating the minimization of the cost function and the diagnostic file is output.

[0014] As a further description of the above technical solution, the preprocessing method includes:

[0015] (a) Time zone conversion, unifying observation data into Coordinated Universal Time;

[0016] (b) Check the temporal and spatial consistency, compare with the background field, and select the observation data required for assimilation time and region;

[0017] (c) Cloud and rain detection: cloud area detection is performed on the selected observation data using the AGRI (Advanced Geostationary Radiation Image) cloud classification product of the Fengyun-4B satellite, and rain area detection is performed on the selected observation data using the rain gauge data of the microwave radiometer;

[0018] (d) Observation data output: The observation data after cloud and rain detection are classified and summarized according to the instrument type, and the required files for assimilation are output one by one to obtain the original observation brightness temperature data.

[0019] As a further description of the above technical solution, the method for converting the background field of the numerical weather prediction model into simulated brightness temperature data specifically includes: reading the coefficient file of the ground-based microwave radiometer, reading the meteorological element data from the background field, performing horizontal and vertical spatial interpolation, and respectively assigning them to the derived arrays corresponding to the RTTOV-gb model, and calculating the simulated brightness temperature data of the ground-based microwave radiometer through RTTOV-gb.

[0020] As a further description of the above technical solution, when introducing RTTOV-gb into the WRFDA system and constructing a direct assimilation module for the ground-based microwave radiometer, keyword replacement needs to be performed on the RTTOV-gb source code, the file names related to the code itself need to be replaced, and the definitions and call names of relevant interfaces, modules, and libraries in the code also need to be replaced to avoid conflicts with RTTOV and achieve the coexistence of the two transmission modes.

[0021] As a further description of the above technical solution, the method for obtaining the corrected observed brightness temperature data based on the machine learning algorithm specifically includes:

[0022] S31, the WRFDA system calculates the bias correction prediction factors based on the original observed brightness temperature data and the background field;

[0023] S32, based on the bias correction prediction factors output in step S31, using the pre-trained bias correction model to predict the bias of OMB (observed brightness temperature minus simulated brightness temperature). The bias correction model is constructed based on the random forest algorithm, which can be used in the variational assimilation framework, and considering the observation characteristics and terrain location of the ground-based microwave radiometer, it takes into account the source of the bias and the non-linear relationship between the source of the bias and the bias, and can evaluate the importance of the source of the bias. It is separately trained for different instruments and different channels during training;

[0024] S33, perform bias correction on the original observed brightness temperature data according to the bias of OMB predicted in step S32 to obtain the corrected observed brightness temperature data.

[0025] As a further description of the above technical solution, the bias correction prediction factors selected in step S32 include the atmospheric thickness from 1000 - 700 hPa, the atmospheric thickness from 700 - 500 hPa, the atmospheric thickness from 500 - 300 hPa, the atmospheric thickness from 1000 - 300 hPa, the atmospheric thickness from 200 - 50 hPa, the temperature at 2 meters, the humidity at 2 meters, the meridional wind at 10 meters, the zonal wind at 10 meters, the surface pressure, the surface temperature, the precipitable water in the atmosphere, the observed brightness temperature, longitude, and latitude.

[0026] As a further description of the above technical solution, for the corrected observed brightness temperature data, when it meets any one of the rejection conditions, it is rejected; otherwise, it is the corrected observed brightness temperature data required for assimilation. The rejection conditions include: the corresponding observed residual data is greater than 20K; the corresponding observed residual data is greater than 3 times the error; it is the observation data in a cloud and rain environment; the integrated cloud water in the whole layer of the observation point is greater than 0.2g / kg. As a further description of the above technical solution, three-dimensional variational assimilation processing is performed based on minimization iteration, and the objective cost function of the minimization iteration is defined as:

[0027] ;

[0028] where, and represent the analysis field and the background field respectively; is the corrected observed brightness temperature data; and represent the background error covariance matrix and the observation error covariance matrix respectively; represents the observation operator.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] 1) In previous studies on direct assimilation of ground-based microwave radiometers, variational assimilation did not adopt bias correction, and the bias correction method in ensemble assimilation did not consider the source of bias. The method of the present invention fills the gap in the bias correction technology of ground-based microwave radiometers under variational assimilation and can be extended to be used in ensemble assimilation.

[0031] 2) The present invention more carefully considers the source of bias and the non-linear relationship between the source of bias and the bias, which is beneficial to obtaining better bias correction effects and is convenient for adjustment according to the needs of actual applications.

[0032] 3) The present invention establishes a direct assimilation module for ground-based microwave radiometers in WRFDA based on the modified RTTOV-gb, which can realize direct assimilation of various ground-based microwave radiometers and sets a general-purpose interface, which is beneficial to the subsequent realization of direct assimilation of more types of ground-based microwave radiometers.

[0033] 4) The present invention solves the coupling problem between the radiation transfer models RTTOV-gb and RTTOV required by ground-based microwave radiometers and spaceborne microwave radiometers in the data assimilation system WRFDA, enabling the two to be used simultaneously. This setting is beneficial to the subsequent joint direct assimilation of spaceborne and ground-based microwave radiometers.

[0034] 5) The present invention can evaluate the contribution of prediction factors in bias correction, which is beneficial to discovering the source of bias, fundamentally solving these biases, and also providing a reference for other bias correction methods to consider the source of bias.

[0035] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, specific embodiments of the present invention are hereinafter given, and in conjunction with the accompanying drawings, the detailed description is as follows. Description of the Drawings

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0037] Figure 1 is the flowchart of the direct assimilation method described in the embodiment;

[0038] Figure 2 is the flowchart of the RTTOV-gb code processing in the embodiment;

[0039] Figure 3 is the training and evaluation process of the bias correction model in the embodiment;

[0040] Figure 4 is the OMB bias and standard deviation of each channel of HATPRO (the Humidity And Temperature PROfiler) in the embodiment;

[0041] Figure 5 is the contribution diagram of the predictors in the embodiment, that is, the importance of the predictors used in the bias correction model. Detailed Embodiments

[0042] In order to make the objects, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments.

[0043] Please refer to Figure 1 , the embodiment of the present invention provides a direct assimilation method for a ground-based microwave radiometer, including:

[0044] (1) Read the original radiation observation data of the ground-based microwave radiometer and perform preprocessing to obtain the original observed brightness temperature data.

[0045] This step reads the original radiation observation data of the ground-based microwave radiometer based on the Python language and preprocesses the original data:

[0046] (a) Time zone conversion to unify the observation data to Coordinated Universal Time;

[0047] (b) Spatiotemporal consistency check, compare with the background field, and screen the observation data required for the assimilation time and area;

[0048] (c) Cloud and rain detection, use the AGRI cloud classification product of FY-4B satellite to detect cloud areas, and use the rain gauge data of the microwave radiometer itself to detect rain areas;

[0049] (d) Observation data output, classify and summarize the multi-site data according to the instrument type, and output the files required for assimilation for each type one by one, that is, obtain the original observed brightness temperature data.

[0050] (2) Call the RTTOV-gb fast radiative transfer model to convert the background field of the numerical weather prediction model into simulated brightness temperature data.

[0051] In the embodiment of the present invention, an observation operator is constructed using the fast radiative transfer model, and the radiation quantity is directly assimilated without introducing an inversion process, making the assimilation observation accuracy higher and the error more controllable, and more effective in improving the initial value in the numerical weather model. Based on the RTTOV-gb fast radiative transfer model, the direct assimilation of the ground-based microwave radiometer can be realized. For single-frequency radiation, the atmospheric transmission model is calculated as follows:

[0052] ;

[0053] Where is the radiation received by the ground; is the atmospheric temperature; is the Planck radiation quantity at a specific temperature; is the transmittance from the top of the atmosphere to the ground; is the microwave cosmic background temperature, set to 2.728K. RTTOV-gb calculates the optical thickness and transmittance of the atmosphere using a linear relationship based on the given instrument elevation angle, atmospheric temperature and humidity profiles, and air pressure, so as to obtain the simulated radiation quantity (brightness temperature).

[0054] In the existing WRFDA system, the access of RTTOV-gb and the construction of the ground-based microwave radiometer assimilation module have not been realized. The embodiment of the present invention compares the coupling status of WRFDA and RTTOV, and parallelly establishes a coupling framework of WRFDA and RTTOV-gb, and avoids the conflict of simultaneous coupling of RTTOV and RTTOV-gb.

[0055] Simultaneous coupling of RTTOV, RTTOV-gb and WRFDA:

[0056] The WRFDA system uses RTTOV to construct an observation operator to achieve satellite data assimilation. This invention is based on the WRFDA system and uses RTTOV-gb to form an observation operator to achieve the direct assimilation of ground-based microwave radiometers. RTTOV-gb is developed based on RTTOV. The scenarios of their simultaneous use are not considered. The function module names are exactly the same, and the keywords of the call interfaces are both rttov, but the actual calculations are different. If we want to achieve the direct assimilation of ground-based microwave radiometers without affecting satellite data assimilation, we need to avoid conflicts with the existing RTTOV when introducing RTTOV-gb into WRFDA. This invention takes the approach of replacing keywords in all the codes related to interfaces, modules, and libraries in the RTTOV-gb source code. We need to replace not only the file names of the related codes themselves, but also the definitions and call names of the relevant interfaces, modules, and libraries in the codes. The specific process is as Figure 2 shown:

[0057] 1) Compile RTTOV-gb normally;

[0058] 2) Obtain key files: Traverse the.h suffix files in the src directory, the.F90 suffix files in the src directory, the.o suffix files in the obj directory, the.mod suffix files in the mod directory, and the.interface suffix files in the include directory;

[0059] 3) Obtain keywords: Remove the path and suffix from the file names of the key files;

[0060] 4) Prepare the uncompiled RTTOV-gb code;

[0061] 5) Obtain specific files: Traverse the.F90,.h,.f suffix files and Makefile files in the src / main, src / test, src / coef_io, src / other, and src / parallel directories;

[0062] 6) Replace the file names of the specific files: Add _gb to the main name. For example, replace Y.f90 with Y_gb.f90;

[0063] 7) Replace the file content of the specific files: Replace the keywords in the files. If there are no letters and underscores before and after the keyword X, add _gb after the keyword. For example, replace X with X_gb.

[0064] After the replacement is completed, start compiling RTTOV-gb. At this time, the keywords of the RTTOV-gb interface, library, and module are all rttov_gb, which are independent of the keywords of the RTTOV-related interface, library, and module, and the two do not affect each other. Comparing the current coupling situation between WRFDA and RTTOV, a coupling framework of WRFDA and RTTOV-gb is established in parallel based on the modified and compiled RTTOV-gb to achieve the collaborative coupling of the assimilation system with two sets of radiative transfer models, namely satellite-borne and ground-based. Combining RTTOV-gb, establish the observation operator and its tangent linear adjoint required for the direct variational assimilation of the brightness temperature observations of the microwave radiometer, complete the mutual conversion between the brightness temperature observations and the model variables such as temperature, humidity, and wind field, and establish a direct assimilation module for the ground-based microwave radiometer, including the reading of observations, assimilation, and the output of diagnostic files.

[0065] In the embodiment of the present invention, an independent control structure for the ground-based microwave radiometer is established in WRFDA, and the direct assimilation of the ground-based microwave radiometer is realized without affecting the assimilation of other types of data. First, the derived types in the WRFDA system are corresponded to the derived types of RTTOV-gb, and the coefficient file of the ground-based microwave radiometer is read; meteorological elements such as temperature, humidity, and air pressure are read from the background field, and horizontal and vertical spatial interpolations are performed and assigned to the derived arrays corresponding to the RTTOV-gb model respectively; the simulated brightness temperature of the ground-based microwave radiometer is calculated through RTTOV-gb.

[0066] (3) Based on the machine learning algorithm, considering the source of the bias and the non-linear relationship between the source and the bias, calculate the bias correction prediction factors according to the original observed brightness temperature data and the background field, and perform bias correction on the original observed brightness temperature data based on the bias correction prediction factors to obtain the corrected observed brightness temperature data, specifically as follows:

[0067] S31, the WRFDA system calculates the bias correction prediction factors according to the original observed brightness temperature and the background field.

[0068] In this embodiment, by constructing relevant calculation codes within the WRFDA system, the bias correction prediction factors are calculated based on the original observed brightness temperature and the background field.

[0069] S32, based on the bias correction prediction factors output in step S31, use the pre-trained bias correction model to predict the bias of the current OMB.

[0070] In this embodiment, based on the output diagnostic information, select the following elements as the bias correction prediction factors:

[0071] 1000 - 700hPa atmospheric thickness, 700 - 500hPa atmospheric thickness, 500 - 300hPa atmospheric thickness, 1000 - 300hPa atmospheric thickness, 200 - 50hPa atmospheric thickness, 2 - meter temperature, 2 - meter humidity, 10 - meter meridional wind, 10 - meter zonal wind, surface pressure, surface temperature, atmospheric precipitable water, observed brightness temperature, longitude, latitude.

[0072] The variational assimilation assumes that the errors of both observations and background follow unbiased Gaussian distributions. However, due to instrument errors, limitations of radiation transfer models, and background errors of numerical weather prediction models, the observed brightness temperature and simulated brightness temperature contain errors, and their errors may be biased distributions. Bias correction is a key process in radiative data assimilation, aiming to identify and eliminate these biases. In the real atmosphere, the observed brightness temperature and simulated brightness temperature are regarded as the true value T plus their respective differences, as shown in the following formula:

[0073] ;

[0074] where, is the observed brightness temperature, is the simulated brightness temperature, The statistical expected value of can represent the systematic bias. Therefore, it is crucial to evaluate the characteristics of the systematic bias through OMB and correct it.

[0075] For some regions, affected by the plateau and complex terrain, the bias of OMB is very obvious, and it is impossible to meet the needs of assimilation without bias correction or simple bias correction. Therefore, based on machine learning, the present invention proposes a bias correction method for direct assimilation of ground - based microwave radiometers. This method is based on the random forest algorithm in machine learning and aims to correct the OMB bias, with the following characteristics:

[0076] 1) It can be used in the variational assimilation framework;

[0077] 2) Considering the observation characteristics and terrain location of ground - based microwave radiometers, taking into account the sources of bias, that is, the predictors for bias correction;

[0078] 3) Considering the non - linear relationship between predictors and bias;

[0079] 4) It can evaluate the importance of predictors;

[0080] 5) Separately train for different instruments and different channels.

[0081] To evaluate the reliability of this method, it is necessary to train and evaluate the bias correction model. In this embodiment, a three-month experiment was conducted to obtain a large number of samples. The WRF (Weather Research and Forecasting Model) was used as the numerical weather prediction model, and the NCEP-GFS (the National Centers for Environmental Prediction Global Forecast System) data was used to initialize it every 6 hours. The WRFDA was run in monitoring mode (without minimizing iteration) every hour to obtain a large number of samples.

[0082] The specific process is as Figure 3 shown, including:

[0083] 1) Preprocessing of the original radiation observations of multi-site and multi-model ground-based microwave radiometers. This preprocessing process is the same as the previous preprocessing method.

[0084] 2) Simulated brightness temperature calculation

[0085] Based on RTTOV-gb, an observation operator is constructed to complete the conversion from model space variables to observation space variables. Based on the model background pressure, temperature, humidity profiles, and the given elevation angle of the observation instrument, the observation operator calculates the optical thickness and transmittance of the simulated atmosphere, thereby obtaining the simulated radiation amount, that is, the simulated brightness temperature.

[0086] 3) Sample statistics and segmentation

[0087] a. Sample information acquisition

[0088] To obtain training and evaluation samples, in this step, based on the direct assimilation module of the ground-based microwave radiometer constructed in WRFDA, the predicted variables and predictors are calculated and output to form samples of multiple sites and multiple times.

[0089] b. Training set acquisition

[0090] To train the bias correction model, 70% of the data in the sample information is randomly selected as the training sample, that is, the training set is constructed.

[0091] c. Test set acquisition

[0092] To evaluate the bias correction effect, the remaining 30% of the data in the sample information is selected as the test sample, that is, the test set is constructed.

[0093] 4) Bias correction model training

[0094] a. Hyperparameter tuning

[0095] A bias correction model is constructed using random forest. Its essence is to use the predictors as the feature variables of the model and the OMB as the predicted variable, fit the relationship between the predictors and the OMB, and predict the OMB bias through the predictors. The predictors include the atmospheric thickness from 1000 - 700 hPa, the atmospheric thickness from 700 - 500 hPa, the atmospheric thickness from 500 - 300 hPa, the atmospheric thickness from 1000 - 300 hPa, the atmospheric thickness from 200 - 50 hPa, the 2 - meter temperature, the 2 - meter humidity, the 10 - meter meridional wind, the 10 - meter zonal wind, the surface pressure, the surface temperature, the precipitable water in the atmosphere, the observed brightness temperature, the longitude, and the latitude.

[0096] There are two types of parameters in the bias correction model: model parameters and hyperparameters. The model parameters are initialized and updated during the learning process. The hyperparameters define the architecture of the model, cannot be updated during the model training process, and must be set before training. The setting of hyperparameters greatly affects the training time and prediction performance of the model. The process of determining the hyperparameter configuration and the ideal model architecture is called hyperparameter tuning. Hyperparameter tuning is a key element of an effective machine learning model. The random forest algorithm has four key hyperparameters: the number of trees in the forest, the maximum depth of the tree, the minimum number of samples required to split an internal node, and the minimum number of samples required at a leaf node. Set a certain range for these hyperparameters, combine them, and a hyperparameter grid can be formed, which contains all possible hyperparameters within the preset range. Based on the training set data, traverse the hyperparameter grid, train the model, and score it to obtain the score for each hyperparameter combination. Among them, for a certain hyperparameter setting, the 5 - fold cross - validation method is adopted for model training and scoring: after evenly dividing the data into 5 equal parts. Select the first part of the data as the test data, and the remaining 4 parts as the training data. Use the training data to train the random forest model, and use the test data to evaluate the effect of the trained model, and calculate the model score S 1 . By analogy, select the second, third, fourth, and fifth parts of the data as the test data in turn, and calculate the model scores S 2 、S 3 、S 4 、S 5 . Finally, average the 5 scores to obtain the hyperparameter score S.

[0097] b. Optimal hyperparameter setting

[0098] According to the scoring results, select the hyperparameter combination with the highest score S as the optimal hyperparameter.

[0099] c. Bias correction model

[0100] Based on the optimal hyperparameter setting, retrain the model using the training set. For each type of instrument and each item, train and save them locally respectively.

[0101] 5) Deviation correction model evaluation

[0102] The randomly selected test set data accounts for 30% of the total sample and does not participate in model training. This data can be used to test the effect of the deviation correction model.

[0103] a. Original OMB

[0104] Based on the test set data, calculate and statistically obtain the original OMB.

[0105] b. Predicted OMB

[0106] Based on the test set data, for each type of instrument and channel, load the corresponding deviation correction model, input the selected predictors, predict the OMB, and obtain the predicted OMB, denoted by .

[0107] c. Corrected OMB

[0108] Based on the original OMB and the predicted OMB, obtain the corrected OMB, denoted by:

[0109] ;

[0110] d. Comparative evaluation

[0111] Based on the original OMB and the corrected OMB, evaluate the model deviation correction effect and analyze the importance of predictors.

[0112] Figure 4 shows the deviation correction effect of each channel of HATPRO. Among them, Figure (a) shows the OMB deviation with and without deviation correction, and Figure (b) shows the standard deviation of OMB with and without deviation correction. Before deviation correction, the OMB deviation of HATPRO is between 0 and 2K, and the deviation in the K band (especially channels 4 to 7) is less than that in the V band. After deviation correction, the deviation of each channel is about 0K. In terms of the STD (Standard Deviation) of OMB, without deviation correction, its value is between 2 and 4K, and the values of channels 4 to 7 are less than those of other channels. After deviation correction, the STD of OMB fluctuates between 0.5 and 1.5K. After applying this deviation correction model, both the OMB deviation and the OMB standard deviation are significantly reduced, by 0.77K (i.e., a reduction of 99.8%) and 1.6K (i.e., a reduction of 65.0%) respectively. The distribution of the corrected OMB shows a Gaussian characteristic centered around zero, indicating that the systematic OMB deviation has been effectively alleviated.

[0113] It is crucial to diagnose the contribution of each predictor, Figure 5Shows the importance of predictors in HATPRO bias correction. The higher the score, the stronger the correlation between the predictor and the OMB bias. Observed brightness temperature, total precipitable water in the atmosphere, and surface pressure are important factors in K-band bias correction. For the V-band, observed brightness temperature, latitude, and surface pressure are the most influential predictors. Compared with other predictors, the contribution of the atmospheric thickness predictor is relatively small; however, the contribution of the 1,000 - 700 hPa thickness predictor is relatively large, which is related to the fact that the ground-based microwave radiometer mainly observes the radiation of the lower atmosphere. It should be noted that surface pressure plays a key role in the bias correction of the temperature channel, which is the reason for the positive OMB bias observed at plateau stations.

[0114] S33. Perform bias correction on the original observed brightness temperature data according to the bias of the OMB predicted in step S32 to obtain the corrected observed brightness temperature data.

[0115] Since the bias correction is implemented based on Python code and cannot be mixed-compiled with the code of WRFDA, the coupling between the two is not at the memory level but through file reading and writing. WRFDA outputs the variables required for bias correction by calculating diagnostic information such as predictors, and the bias correction model outputs the correction results. Since WRFDA can only read the observed information y, rather than the bias to be corrected, therefore, in the present invention, the bias to be corrected is implemented on the original observed brightness temperature y, according to the formula

[0116] ;

[0117] Calculate the corrected observed brightness temperature , and perform output of the corrected observed data. Classify and summarize the multi-site data according to the instrument type, and output the files required for assimilation one by one type, that is, obtain the corrected observed brightness temperature data.

[0118] (4) Three-dimensional variational assimilation

[0119] a. Calculate the observed residual data according to the simulated brightness temperature data and the corrected observed brightness temperature data.

[0120] ;

[0121] In the formula represents the observed residual, is the observed brightness temperature after bias correction, is the simulated brightness temperature.

[0122] b. Screen out the corrected observed brightness temperature data required for assimilation according to the observed residual data.

[0123] Due to the influence of instrument detection itself and the atmospheric environment, the observation data will inevitably have errors. Before entering the three-dimensional variational assimilation, certain technical means must be used to eliminate the data with large errors. The remaining data is the corrected observation brightness temperature data required for assimilation, which can ensure the quality of the observation data entering the assimilation module.

[0124] The quality control of the ground-based microwave radiometer in the embodiment of the present invention mainly includes the following:

[0125] Observation quality control code: included in the original data, provided by the instrument manufacturer and observers, to eliminate data with problems in observation quality.

[0126] Absolute Deviation Test: Elimination Observations greater than 20K.

[0127] Relative deviation test: Elimination Observations with errors greater than 3 times.

[0128] Observation cloud and rain area detection: Based on the cloud and rain area markings, observations under cloud and rain environments are eliminated.

[0129] Simulated cloud and rain area detection: Eliminate observations where the cloud water integral of the entire layer at the observation point is greater than 0.2g / kg in the elimination mode.

[0130] c. Using the assimilation module of the WRFDA system, based on the background error covariance matrix, the observation error covariance matrix, the screened corrected observation brightness temperature, the observation operator and its tangent adjoint, the background field is iterated by three-dimensional variational minimization to finally obtain the analysis field.

[0131] In this embodiment, the target cost function of the minimization iteration is defined as:

[0132] ;

[0133] in, For the target price, and denote the analysis field and the background field respectively; is the corrected observed brightness temperature data; and denote the background error covariance matrix and the observation error covariance matrix respectively; Represents the observation operator.

[0134] d. Diagnostic file output.

[0135] Based on the background field and analysis field , and the diagnostic files such as the mean and root mean square error of OMB and OMA (observation minus analysis) in the observation space are output, which can be used to further evaluate the assimilation effect.

[0136] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A direct assimilation method for ground-based microwave radiometers, characterized in that: include: Read the original radiation observation data of the ground-based microwave radiometer and pre-process it to obtain the original observation brightness temperature data; Call the RTTOV-gb fast radiative transfer model, correspond the derived types in the WRFDA system to the RTTOV-gb derived types, so that the RTTOV, RTTOV-gb fast radiative transfer models and the WRFDA system are coupled simultaneously, and then the background field of the numerical weather forecast model is converted into simulated brightness temperature data; Based on the machine learning algorithm, the source of the deviation and the nonlinear relationship between the source and the deviation are considered. The deviation correction prediction factor is calculated according to the original observed brightness temperature data and the background field. The deviation of the original observed brightness temperature data is corrected based on the deviation correction prediction factor to obtain the corrected observed brightness temperature data. Calculate observational residual data based on simulated brightness temperature data and corrected observed brightness temperature data, and select corrected observed brightness temperature data required for assimilation based on observed residual data; Using the assimilation module of the WRFDA system, the background field is subjected to three-dimensional variational assimilation processing based on the screened and corrected observational brightness temperature data to obtain the analysis field and output the diagnostic file.

2. The direct assimilation method for ground-based microwave radiometer according to claim 1, characterized in that: Preprocessing methods include: (a) Time zone conversion, unifying observation data into Coordinated Universal Time; (b) Check the temporal and spatial consistency, compare with the background field, and select the observation data required for assimilation time and region; (c) Cloud and rain detection: cloud area detection is performed on the selected observation data using the AGRI cloud classification product of the Fengyun-4B satellite, and rain area detection is performed on the selected observation data using the rain gauge data of the microwave radiometer; (d) Observation data output: The observation data after cloud and rain detection are classified and summarized according to the instrument type, and the required files for assimilation are output one by one to obtain the original observation brightness temperature data.

3. The direct assimilation method for ground-based microwave radiometer according to claim 1, characterized in that: The method for converting the background field of the numerical weather forecast model into simulated brightness temperature data specifically includes: reading the coefficient file of the ground-based microwave radiometer, reading the meteorological element data from the background field, performing horizontal and vertical spatial interpolation, and assigning them to the derived arrays corresponding to the RTTOV-gb model respectively, and calculating the simulated brightness temperature data of the ground-based microwave radiometer through RTTOV-gb.

4. The direct assimilation method for ground-based microwave radiometer according to claim 3, characterized in that: When introducing RTTOV-gb into the WRFDA system and building the ground-based microwave radiometer direct assimilation module, it is necessary to replace keywords in the RTTOV-gb source code, including replacing the file name of the code itself, as well as the definition and call names of the relevant interfaces, modules, and libraries in the code.

5. The direct assimilation method for ground-based microwave radiometer according to claim 1, characterized in that: The method for obtaining the corrected observation brightness temperature data based on the machine learning algorithm specifically includes: S31, WRFDA system calculates the bias correction prediction factor based on the original observed brightness temperature data and background field; S32, based on the bias correction prediction factor outputted from step S31, using the bias correction model obtained by pre-training to predict the bias of the observed brightness temperature minus the simulated brightness temperature, the bias correction model is constructed based on the random forest algorithm, which can be used in the variational assimilation framework, and takes into account the source of the bias and the nonlinear relationship between the bias source and the bias for the observation characteristics and terrain position of the ground-based microwave radiometer, and can evaluate the importance of the bias source, and is trained separately for different instruments and different channels during training; S33, performing deviation correction on the original observed brightness temperature data according to the deviation of the observed brightness temperature predicted in step S32 minus the simulated brightness temperature, to obtain the corrected observed brightness temperature data.

6. The direct assimilation method for ground-based microwave radiometer according to claim 5, characterized in that: The deviation correction prediction factors selected in step S32 include 1000-700hPa atmospheric thickness, 700-500hPa atmospheric thickness, 500-300hPa atmospheric thickness, 1000-300hPa atmospheric thickness, 200-50hPa atmospheric thickness, 2-meter temperature, 2-meter humidity, 10-meter meridional wind, 10-meter zonal wind, surface air pressure, surface temperature, atmospheric precipitable water, observed brightness temperature, longitude and latitude.

7. The direct assimilation method for ground-based microwave radiometer according to claim 1, characterized in that: For the corrected observational brightness temperature data, if it meets any of the exclusion conditions, it will be excluded, otherwise it is the corrected observational brightness temperature data required for assimilation; the exclusion conditions include: the corresponding observational residual data is greater than 20K; the corresponding observational residual data is greater than 3 times the error; it is the observation data under cloud and rain environment; the cloud water integral of the entire layer at the observation point is greater than 0.2g / kg.

8. The direct assimilation method for ground-based microwave radiometer according to claim 1, characterized in that: The three-dimensional variational assimilation processing is performed based on minimization iteration, and the objective cost function of the minimization iteration is defined as: ； in, For the target price, and denote the analysis field and the background field respectively; is the corrected observed brightness temperature data; and denote the background error covariance matrix and the observation error covariance matrix respectively; Represents the observation operator.

Images