A joint correction method for microwave radiometer temperature and humidity profiles based on all-weather weather background
By introducing a multi-layer perceived convolutional length short-term memory network deep learning model and consistency loss function in the temperature and humidity profile correction of microwave radiometers, combined with weather background factors, the problem of incomplete consideration of factors in the existing technology is solved, and the more Accurate temperature and humidity profile inversion.
Patent Information
- Application Number
- CN202411068531.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-06
AI Technical Summary
The existing microwave radiometer temperature and humidity profile correction method fails to fully consider various factors such as weather phenomena and geographical elevation, resulting in errors in the measurement results, and traditional methods are difficult to provide accurate inversion results under different weather conditions.
Using a multi-layer perceived convolutional long short-term memory network deep learning model based on all-weather background, combining microwave radiometer measurement data and radiosonde detection data, the fusion model is constructed to correct the temperature and humidity profile by introducing a consistency loss function. , considering the physical consistency characteristics of temperature and humidity.
Improves correction accuracy and robustness, and can provide more comprehensive, accurate and efficient temperature and humidity profile inversion results under different weather conditions and geographical environments.
Smart Images

Figure CN118604919B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a microwave radiometer temperature and humidity profile correction method, in particular to a microwave radiometer temperature and humidity profile joint correction method based on all-weather weather background. Background Art
[0002] In recent years, microwave radiometers, as one of the important instruments for remote sensing measurement of atmospheric profiles, are often used to invert the vertical thermodynamic characteristics of the upper atmosphere. Microwave radiometers receive microwaves emitted by certain atmospheric components at specific frequencies and then convert them into the distribution of meteorological parameters in the atmosphere, so they can provide vertical profiles of temperature and humidity with high temporal and spatial resolution. However, on the one hand, microwave radiometer inversion is always affected by various factors, such as instrument noise, signal processing algorithms, conversion processes, etc.; on the other hand, atmospheric conditions are complex and changeable, and the measurement results of microwave radiometers are often affected by different weather backgrounds, resulting in errors in the measurement results. In the prior art, the temperature and humidity profile correction method of microwave radiometers is mainly based on empirical formulas or traditional statistical methods, but fails to fully consider various related factors such as weather phenomena and geographic elevation. Summary of the invention
[0003] Purpose of the invention: The purpose of the present invention is to provide a joint correction method for the temperature and humidity profile of a microwave radiometer based on the all-weather weather background, which can obtain more comprehensive, accurate and efficient temperature and humidity profile inversion results.
[0004] Technical solution: The combined correction method of microwave radiometer temperature and humidity profile based on all-weather weather background of the present invention comprises the following steps:
[0005] S1. Obtain microwave radiometer measurement data and radiosonde detection data, perform data cleaning and standardization preprocessing, and divide the obtained standardized microwave radiometer measurement data and radiosonde detection data matrix into training set and validation set respectively;
[0006] S2, marking and encoding the characteristics of the all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, generating a static feature vector matrix and a dynamic feature vector matrix, and performing normalized scale matching processing on the static feature vector matrix and the dynamic feature vector matrix respectively, to form a dimensionless all-weather weather background factor library, and corresponding to the data set division scheme in step S1, the dimensionless all-weather weather background factor library is also divided into a training set and a verification set;
[0007] S3. A joint loss function that introduces consistency loss is used to construct a multi-layer perceptual convolutional long short-term memory network deep learning model that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background features, and the training set and the validation set are used to carry out model training; wherein the multi-layer perceptual convolutional long short-term memory network deep learning model includes an input module, a multi-layer perceptual module, a convolutional network module, a long short-term memory module, and an output module; the input module is used to integrate the standardized microwave radiometer measurement data and radiosonde detection data obtained in step S1, and the dimensionless all-weather weather background factor library obtained in step S2 into input static features and input dynamic features, the multi-layer perceptual module is used to process the input static features, the convolutional network module is used to refine the input dynamic features, the long short-term memory module is used to generate time series features, and the output module converts the time series feature vector generated by the long short-term memory module into the final correction result, i.e., the corrected standardized temperature and humidity vertical profiles;
[0008] The joint loss function is used to balance the consistency characteristics of temperature and humidity during model training;
[0009] S4. According to the multi-layer perceptual convolutional long short-term memory network deep learning model trained in step S3, for the target time, based on microwave radiometer measurements and all-weather weather background characteristics, generate a corrected temperature and humidity profile inversion result.
[0010] Furthermore, in step S1, the microwave radiometer measurement data and the radiosonde detection data are obtained, and data cleaning and standardization preprocessing are performed. The specific method is: firstly, incomplete, irrelevant, and incorrect parts are identified and eliminated, and then the above data are respectively standardized and preprocessed. The calculation method is:
[0011] ;
[0012] ;
[0013] ;
[0014] ;
[0015] in, and are matrices containing the original microwave radiometer measurement data and radiosonde detection data, respectively. and are the temperature and humidity measured by the microwave radiometer, and are the temperature and humidity detected by the radiosonde, and They are Mean and standard deviation, and They are The mean and standard deviation of and They are the standardized microwave radiometer measurement data and radiosonde detection data matrices, respectively.
[0016] Furthermore, in step S2, the all-weather weather background factor characteristics that affect the measurement accuracy of the microwave radiometer are annotated and encoded to generate a static feature vector matrix and a dynamic feature vector matrix, and the static feature vector matrix and the dynamic feature vector matrix are respectively normalized and scale-matched to form a dimensionless all-weather weather background factor library, specifically:
[0017] S21. Label the characteristics of all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, including time tags, weather phenomena, location information and geographic elevation data. The labeling method is:
[0018] ;
[0019] ;
[0020] ;
[0021] ;
[0022] ;
[0023] in, and They are the moment information and date information of the time label factor, For the current moment, is the current date, is the weather phenomenon factor, and are the longitude and latitude information of the location factor, is the longitude, is the latitude, and Represents geographic elevation data;
[0024] S22, integrate the feature annotation codes of the all-weather weather background factors obtained in step S21 into a static feature vector matrix and a dynamic eigenvector matrix , the method is:
[0025] ;
[0026] ;
[0027] S23, the static eigenvector matrix of the all-weather weather background factor obtained in step S22 and the dynamic eigenvector matrix Normalized scale matching processing is performed separately to form a dimensionless all-weather weather background factor library to ensure that the characteristics of different weather background factors are at the same scale and are applied to subsequent model training. The calculation method is:
[0028] ;
[0029] ;
[0030] in, and They are The maximum and minimum values of is the static eigenvector matrix of all-weather weather background factors after normalization; and They are The maximum and minimum values of is the dynamic eigenvector matrix of all-weather weather background factors after normalization; and Form a dimensionless all-weather weather background factor library.
[0031] Furthermore, the joint loss function introducing the consistency loss in step S3 is:
[0032] ;
[0033] ;
[0034] ;
[0035] ;
[0036] in, is the model joint loss function, is the evaluation sample size in model training, and The radiosonde detection Sample temperature and humidity, and After correction, they are The inversion results of sample temperature and humidity are: Loss of physical consistency due to temperature and humidity, is a balancing factor used to adjust the weight of consistency loss in the total loss, and Respectively The corrected and true water vapor pressure for each sample.
[0037] Furthermore, the multi-layer perceptron convolutional long short-term memory network deep learning model constructed in step S3 that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background characteristics includes an input module, a multi-layer perceptron module, a convolutional network module, a long short-term memory module, and an output module, and its specific form is:
[0038] Input module: used to read the standardized microwave radiometer measurement data With radiosonde data , and dimensionless all-weather weather background factor library and , which are integrated into the input static features in the input module Dynamic features with input ; Its specific form is:
[0039] ;
[0040] ;
[0041] Multi-layer perception module: contains multiple fully connected layers to process input static features ; Its specific form is:
[0042] ;
[0043] ;
[0044] in, is the initial static feature , Indicates activation operation. For the The weight matrix of the fully connected layer, For the The bias vector of the fully connected layer, For the The output feature vector of the fully connected layer is For the The output feature vector of the fully connected layer; finally, the multi-layer perception module outputs a high-level static feature vector ;
[0045] Convolutional network module: Contains multiple convolutional layers and pooling layers stacked alternately to refine the input dynamic features ; The convolution layer is used to extract local spatial features, and the pooling layer performs downsampling to reduce feature dimensions; its specific form is:
[0046] ;
[0047] ;
[0048] ;
[0049] in, is the initial dynamic feature , It is convolution kernel weight matrix, represents the convolution operation, For the The bias vector of the convolution kernel, Represents the maximum pooling operation; Indicates The output feature vector of the convolutional layer, Indicates The output feature vector of the pooling layer is Indicates The output feature vector of the pooling layer,; finally, the convolutional network module outputs the extracted refined dynamic feature vector ;
[0050] Long short-term memory module: convert high-level static feature vectors and refined dynamic feature vector The fusion feature vector is formed by splicing, and then input into the long short-term memory network model to process the above features, and obtain the time series feature vector generated by the long short-term memory module; its specific form is:
[0051] ;
[0052] ;
[0053] in, is the fusion feature vector, is the vector concatenation operation, Processing operations for the LSTM network model, for The output of the network update at each moment is the temporal feature vector generated by the long short-term memory module;
[0054] Output module: contains a fully connected layer that converts the time series feature vector generated by the long short-term memory module Converted into the final correction result, that is, the corrected standardized temperature and humidity vertical profile; its specific form is:
[0055] ;
[0056] in, is the weight matrix of the fully connected layer in the output module, is the bias vector of the fully connected layer in the output module, and are the corrected standardized temperature and humidity profile feature vectors, respectively.
[0057] Furthermore, when the constructed multi-layer perceptual convolutional long short-term memory network deep learning model is trained in step S3, the correction effect is evaluated on the validation set, and the model parameters are iteratively updated using the gradient descent method to minimize the loss function and determine the optimal model configuration; its specific form is:
[0058] ;
[0059] in, and are the parameter values before and after the model parameter update, is the learning rate, is the joint loss function of the model.
[0060] Furthermore, step S4 is specifically as follows:
[0061] S41, bringing the microwave radiometer measurement and all-weather weather background characteristics within the target time into the multi-layer perceptron convolutional long short-term memory network deep learning model trained in step S3, and outputting the corrected standardized temperature and humidity profile feature vector;
[0062] S42, corresponding to the standardized inverse process of step S1, substituting the corrected standardized temperature and humidity profile feature vectors to generate a corrected temperature and humidity profile inversion result.
[0063] Based on the same inventive concept, in one embodiment of the present invention, a microwave radiometer temperature and humidity profile joint correction system based on all-weather weather background includes:
[0064] A data preprocessing unit is used to obtain microwave radiometer measurement data and radiosonde detection data, perform data cleaning and standardization preprocessing, and divide the obtained standardized microwave radiometer measurement data and radiosonde detection data matrix into a training set and a validation set respectively;
[0065] A background factor library generation unit is used to annotate and encode the characteristics of all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, generate a static feature vector matrix and a dynamic feature vector matrix, and perform normalized scale matching processing on the static feature vector matrix and the dynamic feature vector matrix respectively to form a dimensionless all-weather weather background factor library, and divide the dimensionless all-weather weather background factor library into a training set and a validation set in accordance with the data set division scheme in the data preprocessing unit;
[0066] A model building unit is used to adopt a joint loss function that introduces consistency loss to build a multi-layer perceptron convolutional long short-term memory network deep learning model that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background features, and use a training set and a validation set to carry out model training; wherein, the multi-layer perceptron convolutional long short-term memory network deep learning model includes an input module, a multi-layer perceptron module, a convolutional network module, a long short-term memory module, and an output module; the input module is used to integrate the standardized microwave radiometer measurement data and the radiosonde detection data, as well as the dimensionless all-weather weather background factor library into input static features and input dynamic features, the multi-layer perceptron module is used to process the input static features, the convolutional network module is used to refine the input dynamic features, the long short-term memory module is used to generate time series features, and the output module converts the time series feature vector generated by the long short-term memory module into a final correction result, i.e., the corrected standardized temperature and humidity vertical profiles;
[0067] The joint loss function is used to balance the consistency characteristics of temperature and humidity during model training;
[0068] The test unit is used to generate the corrected temperature and humidity profile inversion results based on the trained multi-layer perceptron convolutional long short-term memory network deep learning model, for the target time, based on microwave radiometer measurements and all-weather weather background characteristics.
[0069] Based on the same inventive concept, in one embodiment of the present invention, an electronic device includes:
[0070] A memory storing executable program code;
[0071] a processor coupled to the memory;
[0072] The processor calls the executable program code stored in the memory to execute the steps of the above-mentioned microwave radiometer temperature and humidity profile joint correction method based on all-weather weather background.
[0073] Based on the same inventive concept, in one embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are called, the steps of the combined correction method of the temperature and humidity profile of the microwave radiometer based on the all-weather weather background are executed as described above.
[0074] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows: (1) Based on the fusion-improved deep learning model, the nonlinear relationship between microwave radiometer measurement data, radiosonde detection data and all-weather weather background characteristics is effectively extracted, which reflects the actual environmental conditions more comprehensively than the traditional scheme, is suitable for different weather conditions and geographical environments, and improves the correction accuracy; (2) The constructed multi-layer perception convolutional long short-term memory network deep learning model adds multi-layer perception and convolutional network modules on the basis of the traditional long short-term memory network, processes static, dynamic and time series characteristics in modules, captures the input data relationship at multiple levels and angles, and enhances the robustness and generalization ability of the model; (3) For the correction of microwave radiometer temperature and humidity profile based on all-weather weather background, a new joint loss function introducing consistency loss is established, considering the physical consistency characteristics of temperature and humidity, and jointly correcting the microwave radiometer temperature and humidity profile, and finally obtaining a more comprehensive, accurate and efficient temperature and humidity profile inversion result. BRIEF DESCRIPTION OF THE DRAWINGS
[0075] Figure 1 is a flow chart of the method of the present invention;
[0076] Figure 2 This is a schematic diagram of the multi-layer perceptron convolutional long short-term memory network deep learning model;
[0077] Figure 3 It is a structural diagram of the multi-layer perception module;
[0078] Figure 4 This is the structure diagram of the convolutional network module. DETAILED DESCRIPTION
[0079] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0080] The present invention provides a combined correction method for microwave radiometer temperature and humidity profiles based on all-weather weather background, the process is as follows: Figure 1 As shown, the specific steps include:
[0081] S1. Obtain microwave radiometer measurement data and radiosonde detection data, and perform data cleaning and standardization preprocessing. The specific method is to first identify and remove incomplete, irrelevant, and incorrect parts, and then perform standardization preprocessing on the above data respectively. The calculation method is:
[0082] ;
[0083] ;
[0084] ;
[0085] ;
[0086] in, and are matrices containing the original microwave radiometer measurement data and radiosonde detection data, respectively. and are the temperature and humidity measured by the microwave radiometer, and are the temperature and humidity detected by the radiosonde, and They are The mean and standard deviation of and They are The mean and standard deviation of and They are the standardized microwave radiometer measurement data and radiosonde detection data matrices, respectively.
[0087] Subsequently, the obtained standardized microwave radiometer measurement data and radiosonde detection data matrix are divided into training set and validation set respectively according to a certain ratio.
[0088] S2, marking and encoding the characteristics of all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, including time tags, weather phenomena, location information, and geographic elevation data, generating corresponding feature vectors, forming a dimensionless all-weather weather background factor library, and corresponding to the data set division scheme in step S1, dividing the dimensionless all-weather weather background factor library into a training set and a verification set; the specific steps are as follows;
[0089] S21. Label the characteristics of all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, including time tags (time, date), weather phenomena (clear sky / cloudy / precipitation / other), location information (latitude and longitude), and geographic elevation data. The labeling method is as follows:
[0090] ;
[0091] ;
[0092] ;
[0093] ;
[0094] ;
[0095] in, and are the moment information and date information of the time label factor, respectively. For the current moment, is the current date is the weather phenomenon factor, and are the longitude and latitude information of the location factor, is the longitude, is the latitude, and Represents geographic elevation data.
[0096] S22, after obtaining the feature annotation code of the all-weather weather background factor in S21, integrate it into a static feature vector matrix and a dynamic eigenvector matrix , as follows:
[0097] ;
[0098] ;
[0099] S23, the static eigenvector matrix of all-weather weather background factors obtained in S22 and the dynamic eigenvector matrix Normalized scale matching processing is performed separately to form a dimensionless all-weather weather background factor library to ensure that the characteristics of different weather background factors are at the same scale and are applied to subsequent model training. The calculation method is:
[0100] ;
[0101] ;
[0102] in, and They are The maximum and minimum values of is the static eigenvector matrix of all-weather weather background factors after normalization; and They are The maximum and minimum values of is the dynamic eigenvector matrix of all-weather weather background factors after normalization; and Form a dimensionless all-weather weather background factor library.
[0103] S24. Corresponding to the data set division scheme in step S1, the dimensionless all-weather weather background factor library obtained in S23 is also divided into a training set and a validation set.
[0104] S3. A joint loss function that introduces consistency loss is adopted, and the consistency characteristics of temperature and humidity are considered to construct a multi-layer perceptual convolutional long short-term memory network deep learning model that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background characteristics, and carry out model training; wherein the multi-layer perceptual convolutional long short-term memory network deep learning model mainly includes an input module, a multi-layer perceptual module, a convolutional network module, a long short-term memory module, and an output module; the input module is used to obtain and allocate the data features obtained in steps S1 and S2, that is, to integrate the standardized microwave radiometer measurement data and radiosonde detection data obtained in step S1, and the dimensionless all-weather weather background factor library obtained in step S2 as input static features. Dynamic features with input ; The multi-layer perception module is used to process the input static features, the convolutional network module is used to refine the input dynamic features, the long short-term memory module is used to generate time series features, and the output module converts the time series feature vector generated by the long short-term memory module into the final correction result, that is, the corrected standardized temperature and humidity vertical profiles; The specific steps are as follows:
[0105] S31. Establish a joint loss function that introduces consistency loss, consider the consistency characteristics of temperature and humidity, and jointly correct the microwave radiometer temperature and humidity profile to obtain more comprehensive, accurate and efficient temperature and humidity profile inversion results. The specific form is:
[0106] ;
[0107] in, is the model joint loss function, is the evaluation sample size in model training, and The radiosonde detection Sample temperature and humidity, and After correction, they are The inversion results of sample temperature and humidity are: Loss of physical consistency due to temperature and humidity, is a balancing factor used to adjust the weight of consistency loss in the total loss, and Respectively The loss function can balance the consistency characteristics of temperature and humidity during the training process, so that the training model can achieve the best effect in terms of both accuracy and consistency, and realize the joint correction of the microwave radiometer temperature and humidity profile.
[0108] S32. Using the joint loss function obtained in S31, a multi-layer perception convolutional long short-term memory network deep learning model is constructed that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background characteristics. The model mainly includes an input module, a multi-layer perception module, a convolutional network module, a long short-term memory module, and an output module, such as Figure 2 As shown, its specific form is:
[0109] Input module: The model reads the standardized microwave radiometer measurement data obtained in step S1 With radiosonde data , and the dimensionless all-weather weather background factor library obtained in step S2 and , which are integrated into the input static features in the input module Dynamic features with input . Its specific form is:
[0110] ;
[0111] Multi-layer perception module: Figure 3 As shown, it contains multiple fully connected layers to process input static features. This module processes the input vector layer by layer through multiple fully connected layers and ReLU activation functions, extracting its more abstract and more representative high-level features, and providing richer information for subsequent convolutional network modules and long short-term memory modules. Its specific form is:
[0112] ;
[0113] ;
[0114] in, is the initial static feature , For the The output feature vector of the fully connected layer is To activate the operation, For the The weight matrix of the fully connected layer, For the The output feature vector of the fully connected layer is For the Finally, the multi-layer perception module outputs a high-level static feature vector .
[0115] Convolutional network module: such as Figure 4 As shown, it contains multiple convolutional layers and pooling layers stacked alternately to refine the input dynamic features. The convolution layer can extract local spatial features, and the pooling layer performs downsampling to reduce feature dimensions. Its specific form is:
[0116] ;
[0117] in, is the initial dynamic feature , Indicates The output feature vector of the convolutional layer, It is Convolution kernel weight matrix, represents the convolution operation, Indicates The output feature vector of the pooling layer is For the The bias vector of the convolution kernel, Indicates The output feature vector of the pooling layer is represents the maximum pooling operation; finally, the convolutional network module outputs the extracted refined dynamic feature vector .
[0118] Long short-term memory module: convert high-level static feature vectors and refined dynamic feature vector The fusion feature vector is formed by splicing, and then input into the long short-term memory network model to process the above features and obtain the time series feature vector generated by the long short-term memory module. Its specific form is:
[0119] ;
[0120] ;
[0121] in, is the fusion feature vector, is the vector concatenation operation, Processing operations for the LSTM network model, for The output of the network update at each moment is the temporal feature vector generated by the long short-term memory module.
[0122] Output module: contains a fully connected layer that converts the time series feature vector generated by the long short-term memory module Converted into the final correction result, that is, the corrected standardized temperature and humidity vertical profile. Its specific form is:
[0123] ;
[0124] in, is the weight matrix of the fully connected layer in the output module, is the bias vector of the fully connected layer in the output module, and is the corrected standardized temperature and humidity profile feature vector.
[0125] S33, train the model constructed in S32, evaluate the correction effect on the validation set, iteratively update the model parameters using the gradient descent method, minimize the loss function, and determine the optimal model configuration. The specific form is:
[0126] ;
[0127] in, and are the parameter values before and after the model parameter update, is the learning rate.
[0128] S4. According to the multi-layer perceptual convolutional long short-term memory network deep learning model trained in step S3, based on the microwave radiometer measurement and all-weather weather background characteristics, generate the corrected temperature and humidity profile inversion results for the target time; the specific steps are as follows:
[0129] S41, bringing the microwave radiometer measurement and all-weather weather background characteristics within the target time into the multi-layer perceptron convolutional long short-term memory network deep learning model trained in S3, and outputting the corrected standardized temperature and humidity profile feature vector;
[0130] S42, corresponding to the standardized inverse process of S1, substitutes the corrected standardized temperature and humidity profile feature vectors to generate the corrected temperature and humidity profile inversion results.
[0131] In summary, the present invention is a method for joint correction of microwave radiometer temperature and humidity profiles based on all-weather weather background, which integrates historical microwave radiometer detection data with radiosonde detection data, and introduces a multi-layer perceptual convolutional long short-term memory network deep learning model, combined with all-weather weather background such as time tags (time, date), weather phenomena (clear sky / cloudy / precipitation / other), location information (latitude and longitude), and geographic elevation data for joint training and learning, which can more comprehensively capture the correlation and influence of the above-mentioned multiple factors on the microwave radiometer temperature and humidity profile, and adopts a new joint loss function that introduces consistency loss, considering the consistency characteristics of temperature and humidity, and jointly corrects the microwave radiometer temperature and humidity profile, and finally obtains a more comprehensive, accurate, and efficient temperature and humidity profile inversion result. The invention integrates the radiosonde detection data and the all-weather weather background characteristics, and adopts a new joint loss function, which improves the consistency of local temperature and humidity inversion, and can accurately and efficiently realize the joint correction of microwave radiometer temperature and humidity profile measurement inversion, and has extremely high application value.
[0132] Based on the same inventive concept, in one embodiment of the present invention, a microwave radiometer temperature and humidity profile joint correction system based on all-weather weather background includes:
[0133] A data preprocessing unit is used to obtain microwave radiometer measurement data and radiosonde detection data, perform data cleaning and standardization preprocessing, and divide the obtained standardized microwave radiometer measurement data and radiosonde detection data matrix into a training set and a validation set respectively;
[0134] The background factor library generation unit is used to annotate and encode the characteristics of the all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, generate corresponding feature vectors, form a dimensionless all-weather weather background factor library, and divide the dimensionless all-weather weather background factor library into a training set and a validation set in accordance with the data set division scheme in the data preprocessing unit;
[0135] A model building unit is used to adopt a joint loss function that introduces consistency loss to build a multi-layer perceptron convolutional long short-term memory network deep learning model that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background features, and to carry out model training; wherein the multi-layer perceptron convolutional long short-term memory network deep learning model includes an input module, a multi-layer perceptron module, a convolutional network module, a long short-term memory module, and an output module; the input module is used to obtain and distribute data features obtained by a data preprocessing unit and a background factor library generation unit, the multi-layer perceptron module is used to process input static features, the convolutional network module is used to refine input dynamic features, the long short-term memory module is used to generate time series features, and the output module converts the time series feature vector generated by the long short-term memory module into a final correction result, i.e., the corrected standardized temperature and humidity vertical profiles;
[0136] The joint loss function is used to balance the consistency characteristics of temperature and humidity during the model training process, so that the training model can achieve the best effect in terms of both accuracy and consistency, and realize the joint correction of the temperature and humidity profile of the microwave radiometer;
[0137] The test unit is used to generate the corrected temperature and humidity profile inversion results based on the trained multi-layer perceptron convolutional long short-term memory network deep learning model, for the target time, based on microwave radiometer measurements and all-weather weather background characteristics.
[0138] Based on the same inventive concept, in one embodiment of the present invention, an electronic device includes:
[0139] A memory storing executable program code;
[0140] a processor coupled to the memory;
[0141] The processor calls the executable program code stored in the memory to execute the steps of the above-mentioned microwave radiometer temperature and humidity profile joint correction method based on all-weather weather background.
[0142] Based on the same inventive concept, in one embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are called, the steps of the combined correction method of the temperature and humidity profile of the microwave radiometer based on the all-weather weather background are executed as described above.
Claims
1. A combined correction method for microwave radiometer temperature and humidity profiles based on all-weather weather background, characterized in that: The following steps are involved: S1. Obtain microwave radiometer measurement data and radiosonde detection data, and perform data cleaning and standardization preprocessing. The specific method is: first identify and remove incomplete, irrelevant, and incorrect parts, and then perform standardization preprocessing on the above data respectively. The calculation method is: ; ; ; ; in, and are matrices containing the original microwave radiometer measurement data and radiosonde detection data, respectively. and are the temperature and humidity measured by the microwave radiometer, and are the temperature and humidity detected by the radiosonde, and They are The mean and standard deviation of and They are The mean and standard deviation of and They are the standardized microwave radiometer measurement data and radiosonde detection data matrices respectively; The obtained standardized microwave radiometer measurement data and radiosonde detection data matrix are divided into training set and validation set respectively; S2, marking and encoding the characteristics of the all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, generating a static feature vector matrix and a dynamic feature vector matrix, and performing normalized scale matching processing on the static feature vector matrix and the dynamic feature vector matrix respectively, to form a dimensionless all-weather weather background factor library, and corresponding to the data set division scheme in step S1, the dimensionless all-weather weather background factor library is also divided into a training set and a verification set; S3. A joint loss function that introduces consistency loss is used to construct a multi-layer perceptual convolutional long short-term memory network deep learning model that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background features, and the training set and the validation set are used to carry out model training; wherein the multi-layer perceptual convolutional long short-term memory network deep learning model includes an input module, a multi-layer perceptual module, a convolutional network module, a long short-term memory module, and an output module; the input module is used to integrate the standardized microwave radiometer measurement data and radiosonde detection data obtained in step S1, and the dimensionless all-weather weather background factor library obtained in step S2 into input static features and input dynamic features, the multi-layer perceptual module is used to process the input static features, the convolutional network module is used to refine the input dynamic features, the long short-term memory module is used to generate time series features, and the output module converts the time series feature vector generated by the long short-term memory module into the final correction result, i.e., the corrected standardized temperature and humidity vertical profiles; The joint loss function is used to balance the consistency characteristics of temperature and humidity during model training; S4. According to the multi-layer perceptual convolutional long short-term memory network deep learning model trained in step S3, for the target time, based on microwave radiometer measurements and all-weather weather background characteristics, generate a corrected temperature and humidity profile inversion result.
2. The combined correction method of microwave radiometer temperature and humidity profile based on all-weather weather background according to claim 1 is characterized in that: In step S2, the all-weather weather background factor characteristics that affect the measurement accuracy of the microwave radiometer are annotated and encoded to generate a static feature vector matrix and a dynamic feature vector matrix, and the static feature vector matrix and the dynamic feature vector matrix are respectively normalized and scale-matched to form a dimensionless all-weather weather background factor library, specifically: S21. Label the characteristics of all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, including time tags, weather phenomena, location information and geographic elevation data. The labeling method is: ; ; ; ; ; in, and They are the moment information and date information of the time label factor, For the current moment, is the current date, is the weather phenomenon factor, and are the longitude and latitude information of the location factor, is the longitude, is the latitude, and Represents geographic elevation data; S22, integrate the feature annotation codes of the all-weather weather background factors obtained in step S21 into a static feature vector matrix and a dynamic eigenvector matrix , the method is: ; ; S23, the static eigenvector matrix of the all-weather weather background factor obtained in step S22 and the dynamic eigenvector matrix Normalized scale matching processing is performed separately to form a dimensionless all-weather weather background factor library to ensure that the characteristics of different weather background factors are at the same scale and are applied to subsequent model training. The calculation method is: ; ; in, and They are The maximum and minimum values of is the static eigenvector matrix of all-weather weather background factors after normalization; and They are The maximum and minimum values of is the dynamic eigenvector matrix of all-weather weather background factors after normalization; and Form a dimensionless all-weather weather background factor library.
3. The combined correction method of microwave radiometer temperature and humidity profile based on all-weather weather background according to claim 1 is characterized in that: The joint loss function that introduces consistency loss in step S3 is: ; ; ; ; in, is the model joint loss function, is the evaluation sample size in model training, and The radiosonde detection Sample temperature and humidity, and After correction, they are The inversion results of sample temperature and humidity are: Loss of physical consistency due to temperature and humidity, is a balancing factor used to adjust the weight of consistency loss in the total loss, and Respectively The corrected and true water vapor pressure for each sample.
4. The combined correction method of microwave radiometer temperature and humidity profile based on all-weather weather background according to claim 1 is characterized in that: The multi-layer perceptual convolutional long short-term memory network deep learning model constructed in step S3 that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background characteristics is specifically in the form of: Input module: used to read the standardized microwave radiometer measurement data With radiosonde data , and dimensionless all-weather weather background factor library and , which are integrated into the input static features in the input module Dynamic features with input ; Its specific form is: ; ; Multi-layer perception module: contains multiple fully connected layers to process input static features ; Its specific form is: ; ; in, is the initial static feature , Indicates activation operation. For the The weight matrix of the fully connected layer, For the The bias vector of the fully connected layer, For the The output feature vector of the fully connected layer is For the The output feature vector of the fully connected layer; finally, the multi-layer perception module outputs a high-level static feature vector ; Convolutional network module: Contains multiple convolutional layers and pooling layers stacked alternately to refine the input dynamic features ; The convolution layer is used to extract local spatial features, and the pooling layer is used to downsample and reduce the feature dimension; its specific form is: ; ; ; in, is the initial dynamic feature , Indicates The output feature vector of the convolutional layer is It is convolution kernel weight matrix, represents the convolution operation, Indicates The output feature vector of the pooling layer is For the The bias vector of the convolution kernel, Indicates The output feature vector of the pooling layer is represents the maximum pooling operation; finally, the convolutional network module outputs the extracted refined dynamic feature vector ; Long short-term memory module: convert high-level static feature vectors and refined dynamic feature vector The fusion feature vector is formed by splicing, and then input into the long short-term memory network model to process the above features, and obtain the time series feature vector generated by the long short-term memory module; its specific form is: ; ; in, is the fusion feature vector, is the vector concatenation operation, Processing operations for the LSTM network model, for The output of the network update at each moment is the temporal feature vector generated by the long short-term memory module; Output module: contains a fully connected layer that converts the time series feature vector generated by the long short-term memory module Converted into the final correction result, that is, the corrected standardized temperature and humidity vertical profile; its specific form is: ; in, is the weight matrix of the fully connected layer in the output module, is the bias vector of the fully connected layer in the output module, and are the corrected standardized temperature and humidity profile feature vectors, respectively.
5. The combined correction method of microwave radiometer temperature and humidity profile based on all-weather weather background according to claim 1 is characterized in that: When training the constructed multi-layer perceptron convolutional long short-term memory network deep learning model in step S3, the correction effect is evaluated on the validation set, and the model parameters are iteratively updated using the gradient descent method to minimize the loss function and determine the optimal model configuration; its specific form is: ; in, and are the parameter values before and after the model parameter update, is the learning rate, is the joint loss function of the model.
6. The combined correction method of microwave radiometer temperature and humidity profile based on all-weather weather background according to claim 1 is characterized in that: Step S4 is specifically as follows: S41, bringing the microwave radiometer measurement and all-weather weather background characteristics within the target time into the multi-layer perceptron convolutional long short-term memory network deep learning model trained in step S3, and outputting the corrected standardized temperature and humidity profile feature vector; S42, corresponding to the standardized inverse process of step S1, substituting the corrected standardized temperature and humidity profile feature vectors to generate a corrected temperature and humidity profile inversion result.
7. The microwave radiometer temperature and humidity profile joint correction system based on all-weather weather background is characterized by: include: The data preprocessing unit is used to obtain microwave radiometer measurement data and radiosonde detection data, and perform data cleaning and standardization preprocessing. The specific method is: first identify and remove incomplete, irrelevant, and incorrect parts, and then perform standardization preprocessing on the above data respectively. The calculation method is: ; ; ; ; in, and are matrices containing the original microwave radiometer measurement data and radiosonde detection data, respectively. and are the temperature and humidity measured by the microwave radiometer, and are the temperature and humidity detected by the radiosonde, and They are The mean and standard deviation of and They are The mean and standard deviation of and They are the standardized microwave radiometer measurement data and radiosonde detection data matrices respectively; The obtained standardized microwave radiometer measurement data and radiosonde detection data matrix are divided into training set and validation set respectively; A background factor library generation unit is used to annotate and encode the characteristics of all-weather weather background factors that affect the measurement accuracy of the microwave radiometer, generate a static feature vector matrix and a dynamic feature vector matrix, and perform normalized scale matching processing on the static feature vector matrix and the dynamic feature vector matrix respectively to form a dimensionless all-weather weather background factor library, and divide the dimensionless all-weather weather background factor library into a training set and a validation set in accordance with the data set division scheme in the data preprocessing unit; A model building unit is used to adopt a joint loss function that introduces consistency loss to build a multi-layer perceptron convolutional long short-term memory network deep learning model that integrates microwave radiometer measurement data, radiosonde detection data, and all-weather weather background features, and use a training set and a validation set to carry out model training; wherein, the multi-layer perceptron convolutional long short-term memory network deep learning model includes an input module, a multi-layer perceptron module, a convolutional network module, a long short-term memory module, and an output module; the input module is used to integrate the standardized microwave radiometer measurement data and the radiosonde detection data, as well as the dimensionless all-weather weather background factor library into input static features and input dynamic features, the multi-layer perceptron module is used to process the input static features, the convolutional network module is used to refine the input dynamic features, the long short-term memory module is used to generate time series features, and the output module converts the time series feature vector generated by the long short-term memory module into a final correction result, i.e., the corrected standardized temperature and humidity vertical profiles; The joint loss function is used to balance the consistency characteristics of temperature and humidity during model training; The test unit is used to generate the corrected temperature and humidity profile inversion results based on the trained multi-layer perceptron convolutional long short-term memory network deep learning model, for the target time, based on microwave radiometer measurements and all-weather weather background characteristics.
8. An electronic device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the microwave radiometer temperature and humidity profile joint correction method based on all-weather weather background as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, which, when called, are used to execute the steps of the microwave radiometer temperature and humidity profile joint correction method based on all-weather weather background as described in any one of claims 1 to 6.
