Rapid temperature and humidity profile inversion method for multi-frequency atmospheric microwave radiometer

Abstract

The invention belongs to the technical field of atmospheric microwave remote sensing and monitoring, and relates to a rapid temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer. Acquiring a brightness temperature data set of the multi-frequency atmospheric microwave radiometer, performing space-time matching on the brightness temperature data set and corresponding ERA5 reanalysis or sounding balloon temperature and humidity profile data, constructing a matched data set, and dividing the matched data set into a test set, a training set and a verification set; establishing an atmosphere temperature and humidity profile inversion model based on a CNN-LSTM combined neural network; and training the model by using the training set, guiding the model optimization process by using the verification set, finally inputting multi-frequency brightness temperature data measured by the multi-frequency microwave radiometer into the trained model, and carrying out inversion to obtain an atmosphere temperature and humidity profile. According to observation data of the multi-frequency atmosphere microwave radiometer, regional atmosphere temperature and humidity profile data are rapidly and accurately obtained, and efficient inversion of the temperature and humidity profile is achieved; the vertical detection advantage of the multi-frequency microwave radiometer is exerted, the inversion efficiency is improved, and the method has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of atmospheric microwave remote sensing and monitoring, and particularly relates to a fast temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer. BACKGROUND

[0002] Atmospheric temperature and humidity are important parameters affecting meteorological phenomena and atmospheric dynamics. Temperature changes at different altitudes directly affect atmospheric stability, wind speed, cloud formation and precipitation, and humidity changes affect atmospheric energy transmission and convection, thereby affecting climate and weather. Atmospheric temperature and humidity profiles refer to the distribution curves of the atmosphere at different altitudes, which show the temperature and humidity conditions of the atmosphere with altitude. Atmospheric temperature and humidity profiles have important application value in weather forecasting, climate analysis, climate change detection, atmospheric science research and aerospace fields.

[0003] Currently, visible and infrared band remote sensing cannot penetrate clouds and bad weather in the atmospheric temperature and humidity profile inversion method. Visible light is easily blocked by clouds, and cannot be monitored continuously all day long. Although the infrared band has a certain penetration, it is still disturbed by clouds and fog, resulting in a decline in image quality or information loss, and the water vapor absorption characteristics are significant, affecting the accuracy of humidity parameter inversion. Compared with visible and infrared band atmospheric remote sensing, microwave remote sensing can penetrate clouds and rain, and can be observed all day long without light restriction. Therefore, carrying a microwave radiometer has become an important and mainstream means of meteorological observation and prediction. Based on this, the present application establishes a fast atmospheric temperature and humidity profile inversion model suitable for atmospheric microwave radiometers. SUMMARY

[0004] The purpose of the present application is to solve the problem that the prior art cannot adapt to the use of different frequency atmospheric microwave radiometers for all-weather, non-light-limited monitoring. A fast temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer is proposed. The present application is based on the multi-frequency spectral characteristics of atmospheric microwave radiometers, and uses a deep learning method to train an atmospheric temperature and humidity profile inversion model, achieving efficient and accurate parameter inversion, and having the advantages of high pertinence, high efficiency and high accuracy.

[0005] The technical scheme of the present application is as follows: A fast temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer, characterized in that it comprises the following steps: S1, data acquisition and matching: obtaining a brightness temperature data set of a multi-frequency atmospheric microwave radiometer in a target area, and performing spatio-temporal matching with corresponding ERA5 reanalysis data or sounding balloon temperature and humidity profile data to construct a matching data set; dividing the matching data set into a training set, a validation set and a test set; S2, model construction: based on the CNN-LSTM joint neural network architecture, an atmospheric temperature and humidity profile inversion model is established; the CNN-LSTM joint neural network includes a convolutional neural network (CNN) for extracting spectral features of multi-frequency brightness temperature data, a single-layer bidirectional long short-term memory network (LSTM) for processing the height correlation characteristics of the atmospheric temperature and humidity profile between brightness temperatures, and a fully connected network (FC) for outputting the final inversion value; S3, model training and optimization: the atmospheric temperature and humidity profile inversion model constructed in step S2 is trained using the training set, the validation set is used to monitor the training process and guide the model hyperparameter optimization to prevent overfitting; S4, profile inversion: input the current multi-frequency atmospheric microwave radiometer brightness temperature data into the atmospheric temperature and humidity profile inversion model trained and optimized in step S3, and directly output the corresponding atmospheric temperature and humidity profile.

[0006] Further, in step S1, the brightness temperature data set observed by the atmospheric microwave radiometer in the target area is obtained; According to the space-time resolution of the brightness temperature data, select the matching ERA5 reanalysis data or sounding balloon temperature and humidity profile data; Set appropriate time and space window size, and perform space-time matching on the brightness temperature data and ERA5 reanalysis data or sounding balloon temperature and humidity profile data to ensure one-to-one correspondence in observation time and observation spatial position, and construct a matched data set.

[0007] Further, in step S2, the specific structure of the CNN-LSTM joint neural network is: The convolutional neural network (CNN) receives multi-frequency brightness temperature data as input, extracts local features and abstract spectral features of brightness temperature data in the frequency dimension through convolutional layers and pooling layers; The atmospheric temperature or humidity profile sequence extracted by CNN and the multi-frequency brightness temperature are input into the single-layer bidirectional long short-term memory network (LSTM), and the LSTM network learns the dynamic characteristics of the atmospheric temperature and humidity profile changes with frequency (implicitly related to atmospheric height) contained in the multi-frequency brightness temperature; The final state output of the LSTM network is input into a fully connected network (FC) for mapping the learned features to the final atmospheric temperature and humidity profile inversion value.

[0008] Further, the model training process in step S3 is based on the error back propagation algorithm, including two stages of forward propagation and back propagation: The forward propagation is to pass the input brightness temperature from the input layer to the output layer layer by layer to obtain the inversion atmospheric temperature and humidity profile value; The backpropagation is performed by calculating the error between the inverted value and the true label value (from the matching dataset), and then using gradient descent to adjust and update the connection weights and bias terms of each layer in the neural network.

[0009] Furthermore, during the forward propagation process, the value of the node j reaching the hidden layer... and the value of output layer node l The following formula is used for calculation: (1) (2) in: : The input value of the j-th node in the hidden layer; Output layer l The input values ​​of each node; : The output value of the i-th node in the input layer; OH j : The output value of the j-th node in the hidden layer after activation; : The connection weight from the i-th node in the input layer to the j-th node in the hidden layer; : The distance from the j-th node in the hidden layer to the j-th node in the output layer l The connection weight of each node; : The bias of the j-th node in the hidden layer; Output layer l The bias of each node; ni: The total number of nodes in the input layer; nh: Total number of hidden layer nodes.

[0010] Furthermore, during the backpropagation process, the error calculation and weight / bias update formulas are as follows: Overall error calculation: (3) Weight update: (4) Bias update: (5) in: The model for the k-th sample l The inversion value of each output node; The k-th sample lThe actual label value of each output node; η: Learning rate, the step size for controlling parameter updates; Weights before the update; Updated weights; : The bias before the update; : Updated bias; Error E with respect to weight parameters The partial derivatives (gradient); Error E relative to bias parameter The partial derivatives (gradient); no: Total number of output layer nodes (the number of discrete height layers corresponding to the atmospheric temperature and humidity profile).

[0011] The beneficial effects of this invention are: The present invention provides a rapid temperature and humidity profile inversion method for multi-frequency atmospheric microwave radiometers. First, it acquires microwave radiometer brightness temperature data and ERA5 reanalysis or sounding balloon temperature and humidity profile data for the target area, constructing a matching dataset divided into training, validation, and test sets. Next, it employs a CNN-LSTM joint neural network to establish an atmospheric temperature and humidity profile inversion model. This model combines the local feature and abstract spectral feature extraction capabilities of convolutional networks in the frequency dimension with the dynamic feature analysis advantages of long short-term memory networks in analyzing atmospheric temperature and humidity profiles as a function of frequency (implying atmospheric height correlation). The model inputs the atmospheric temperature or humidity profile sequence extracted by the CNN into a single-layer bidirectional long short-term memory network (LSTM), and inputs the final output of the LSTM network into a single-layer fully connected network (FC). Finally, it inputs the multi-frequency brightness temperature data measured by the multi-frequency microwave radiometer into the trained model to invert the atmospheric temperature and humidity profile.

[0012] This invention enables the rapid and accurate acquisition of regional atmospheric temperature and humidity profile data based on observational data from a multi-frequency atmospheric microwave radiometer, achieving efficient inversion of these profiles. It leverages the vertical detection advantage of microwave radiometers, improving the efficiency of atmospheric temperature and humidity profile inversion and possessing broad application prospects. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the rapid temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer provided by the present invention. Figure 2 This is a graph showing the atmospheric temperature profile obtained using the method described in Example 1 of this invention. Figure 3This is a graph showing the atmospheric humidity profile obtained using the method described in Example 1 of this invention. Figure 4 This is a graph showing the atmospheric temperature profile obtained using the method described in Example 2 of this invention. Figure 5 This is a graph showing the atmospheric humidity profile obtained using the method described in Example 2 of this invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] To further understand the present invention, it will be further described in conjunction with the accompanying drawings and embodiments.

[0016] like Figure 1 As shown, this invention provides a rapid temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer, comprising the following steps: S1. Data Acquisition and Matching: Obtain the brightness temperature dataset of the multi-frequency atmospheric microwave radiometer in the target area. Based on the spatiotemporal resolution of the brightness temperature data, select matching ERA5 reanalysis or weather balloon atmospheric temperature and humidity profile data. Set appropriate time and spatial window sizes and perform spatiotemporal matching between the brightness temperature data and the corresponding ERA5 reanalysis data or weather balloon atmospheric temperature and humidity profile data to ensure that the two correspond one-to-one in observation time and observation spatial location, and construct a spatiotemporal matching dataset. Based on a linear weighted matching method, observed data is matched with ERA5 reanalysis data. This method calculates a comprehensive score by setting a time window (1800 seconds) and a spatial window (0.25°), and assigning weights to time and spatial distances. For each observation point, the ERA5 candidate data points within its window are traversed, and the normalized weighted sum of time and spatial distances is calculated, i.e., the comprehensive score. Finally, the ERA5 data point with the lowest comprehensive score is selected as the best match for that observation point.

[0017] The specific formula is shown below, where , .

[0018] .

[0019] The matching dataset is divided into a training dataset, a validation dataset, and a test dataset. The training dataset is used for training and improving the model, the validation dataset is used to adjust the model's hyperparameters and to initially evaluate the model's accuracy, and the test dataset is used to evaluate the accuracy of the final model.

[0020] S2, Model Building: A macroscopic temperature and humidity profile inversion model is established based on a CNN-LSTM joint neural network architecture.

[0021] The CNN-LSTM joint neural network includes: a convolutional neural network (CNN) for extracting spectral features of multi-frequency brightness temperature data; a single-layer bidirectional long short-term memory network (LSTM) for processing the high correlation features between atmospheric temperature and humidity profiles between brightness and temperature; and a fully connected network (FC) for outputting the final inversion value.

[0022] The CNN-LSTM joint network can effectively capture multidimensional features in microwave brightness temperature data and adapt to the complex nonlinear relationship between microwave brightness temperature and temperature and humidity profiles. Specifically, the CNN network is responsible for extracting the spectral features of the brightness temperature data; it receives multi-frequency brightness temperature data as input and extracts local and abstract spectral features in the frequency dimension through convolutional and pooling layers.

[0023] The atmospheric temperature or humidity profile sequence extracted by CNN and the multi-frequency brightness temperature are input into LSTM. The LSTM network part processes the dynamic features of the atmospheric temperature and humidity profile that change with frequency (implying atmospheric height correlation) contained in the feature sequence. Through forward propagation and backward propagation, the model parameters are continuously adjusted using the gradient descent method to improve the accuracy of inversion.

[0024] The final output of the LSTM network is input into a layer of FC network to map the learned features to the final atmospheric temperature and humidity profile inversion values.

[0025] S3. Model Training and Optimization: The atmospheric temperature and humidity profile inversion model (deep learning CNN-LSTM model) constructed in step S2 is trained using the training dataset. The training process is monitored using the validation dataset to guide the optimization of model hyperparameters and prevent overfitting.

[0026] Model training consists of two processes: forward propagation and backward propagation. Forward propagation passes the input brightness temperature from the input layer to the output layer layer by layer to obtain the inverted atmospheric temperature and humidity profile values. Backpropagation calculates the error between the inverted values ​​and the true label values ​​(from the matching dataset) and uses gradient descent to adjust and update the connection weights and bias terms of each layer in the neural network.

[0027] Specifically, during the forward propagation process, the value of the node j reaching the hidden layer. and the value of output layer node l The following formula is used for calculation: (1) (2) in: : The input value of the j-th node in the hidden layer; Output layer l The input values ​​of each node; : The output value of the i-th node in the input layer; OH j : The output value of the j-th node in the hidden layer after activation; : The connection weight from the i-th node in the input layer to the j-th node in the hidden layer; : The distance from the j-th node in the hidden layer to the j-th node in the output layer l The connection weight of each node; : The bias of the j-th node in the hidden layer; Output layer l The bias of each node; ni: The total number of nodes in the input layer; nh: Total number of hidden layer nodes.

[0028] During backpropagation, the error calculation and the formulas for updating weights and biases are as follows: Overall error calculation: (3) Weight update: (4) Bias update: (5) in: The model for the k-th sample l The inversion value of each output node; The k-th sample l The actual label value of each output node; η: Learning rate, the step size for controlling parameter updates; Weights before the update; Updated weights; : The bias before the update; : Updated bias; Error E with respect to weight parameters The partial derivatives (gradient); Error E relative to bias parameter The partial derivatives (gradient); no: Total number of output layer nodes (the number of discrete height layers corresponding to the atmospheric temperature and humidity profile).

[0029] S4, Profile Inversion: The brightness temperature data observed by the multi-frequency atmospheric microwave radiometer is input into the trained and optimized atmospheric temperature and humidity profile inversion model to retrieve the corresponding atmospheric temperature and humidity profile.

[0030] Example 1 This embodiment uses brightness temperature data from 18 channels across the following frequency bands: 89 GHz, 118.75 ± 0.08 GHz, 118.75 ± 0.2 GHz, 118.75 ± 0.4 GHz, 118.75 ± 0.8 GHz, 118.75 ± 1.1 GHz, 118.75 ± 1.6 GHz, 118.75 ± 2.1 GHz, 118.75 ± 3.0 GHz, 118.75 ± 3.6 GHz, 118.75 ± 4.3 GHz, 118.75 ± 5.0 GHz, 166 GHz, 183.31 ± 1 GHz, 183.31 ± 1.8 GHz, 183.31 ± 3 GHz, 183.31 ± 4.5 GHz, and 183.31 ± 7 GHz. The oxygen absorption line (118.75 GHz) and surrounding frequency bands are highly sensitive to atmospheric temperature profiles. The water vapor absorption line (183.31 GHz) and surrounding frequency bands are extremely sensitive to atmospheric water vapor content and can accurately detect humidity distribution at different altitudes.

[0031] This embodiment provides a rapid temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer, comprising the following steps: First, a brightness temperature dataset from a multi-frequency atmospheric microwave radiometer is selected. Based on the spatiotemporal resolution of the brightness temperature data from the atmospheric microwave radiometer, an appropriate time and spatial window size is selected from the ERA5 reanalysis or sounding balloon temperature and humidity profile data. Spatiotemporal matching is then performed to ensure that the brightness temperature data from the microwave radiometer corresponds to the ERA5 reanalysis or sounding balloon temperature and humidity profile data in time and space, thus obtaining a matched dataset.

[0032] Secondly, the matching dataset is divided proportionally into training dataset, validation dataset, and test dataset.

[0033] The construction and training process of the CNN-LSTM joint network for inverting atmospheric temperature and humidity profiles is as follows: The brightness temperature training dataset from the multi-frequency atmospheric microwave radiometer observations at frequencies of 89 GHz, 118.75±0.08 GHz, 118.75±0.2 GHz, 118.75±0.4 GHz, 118.75±0.8 GHz, 118.75±1.1 GHz, 118.75±1.6 GHz, 118.75±2.1 GHz, 118.75±3.0 GHz, 118.75±3.6 GHz, 118.75±4.3 GHz, 118.75±5.0 GHz, 166 GHz, 183.31±1 GHz, 183.31±1.8 GHz, 183.31±3 GHz, 183.31±4.5 GHz, and 183.31±7 GHz was input into a convolutional neural network (CNN). The convolutional neural network uses a 5×1 kernel with a stride of 1. Max pooling is performed using a 2×1 kernel with a stride of 2. After two convolutional layers, a fully connected layer with 100 neurons receives feature information and integrates it for regression, outputting 18 layers of atmospheric temperature or humidity. The ReLU function is chosen as the activation function for the hidden layers, and the linear function is used as the activation function for the output layer. The Adam optimizer is used.

[0034] The aforementioned multi-frequency atmospheric microwave radiometer observation brightness temperature training dataset, along with the 18-layer atmospheric temperature or humidity profiles output from the convolutional neural network, are input into a single-layer bidirectional long short-term memory (LSTM) network to output the final inverted 18-layer atmospheric temperature or humidity profiles. The Tanh function is chosen as the activation function for the output layer, and the Adam optimizer is selected.

[0035] Finally, in this embodiment, a deep learning method is used to train an atmospheric temperature and humidity profile inversion model, achieving high-efficiency parameter inversion. The obtained atmospheric temperature and humidity profile results are as follows: Figure 2 and Figure 3 As shown.

[0036] Example 2 This embodiment uses brightness temperature data from 17 channels in the 22-32 GHz and 50-60 GHz frequency bands, specifically at the following frequencies: 22.24 GHz, 23.04 GHz, 23.84 GHz, 25.44 GHz, 26.24 GHz, 27.84 GHz, 31.40 GHz, 50.3 GHz, 51.76 GHz, 52.8 GHz, 53.326 GHz, 53.711 GHz, 54.029 GHz, 54.4 GHz, 54.94 GHz, 55.5 GHz, and 57.29 GHz. The 50-60 GHz band contains a strong oxygen absorption band, and multiple channels from 50.3 GHz to 57.29 GHz constitute a complete oxygen absorption band detection system. This frequency band is extremely sensitive to atmospheric temperature profiles, with different frequencies corresponding to temperature information at different altitudes. The 22-32 GHz band includes the 22.24 GHz water vapor absorption line and its surrounding frequencies. These frequencies are highly sensitive to atmospheric water vapor distribution and can effectively invert humidity profiles at different altitudes.

[0037] This embodiment provides a rapid temperature and humidity profile inversion method for a multi-frequency atmospheric microwave radiometer, comprising the following steps: First, a brightness temperature dataset from a multi-frequency atmospheric microwave radiometer is selected. Based on the spatiotemporal resolution of the brightness temperature data from the atmospheric microwave radiometer, an appropriate time and spatial window size is selected from the ERA5 reanalysis or sounding balloon temperature and humidity profile data. Spatiotemporal matching is then performed to ensure that the brightness temperature data from the microwave radiometer corresponds to the ERA5 reanalysis or sounding balloon temperature and humidity profile data in time and space, thus obtaining a matched dataset.

[0038] Secondly, the matching dataset is divided proportionally into training dataset, validation dataset, and test dataset.

[0039] The construction and training process of the CNN-LSTM joint network for inverting atmospheric temperature and humidity profiles is as follows: The brightness temperature training dataset from multi-frequency atmospheric microwave radiometer observations at 22.24 GHz, 23.04 GHz, 23.84 GHz, 25.44 GHz, 26.24 GHz, 27.84 GHz, 31.40 GHz, 50.3 GHz, 51.76 GHz, 52.8 GHz, 53.326 GHz, 53.711 GHz, 54.029 GHz, 54.4 GHz, 54.94 GHz, 55.5 GHz, and 57.29 GHz was input into a convolutional neural network (CNN). The CNN used a 5×1 kernel with a stride of 1; max pooling was performed using a 2×1 kernel with a stride of 2. After two convolutional layers, a fully connected layer with 100 neurons received the feature information and integrated it for regression, outputting 17 layers of atmospheric temperature or humidity. The ReLU function was chosen as the activation function for the hidden layers, the linear function was chosen as the activation function for the output layers, and the Adam optimizer was chosen.

[0040] The aforementioned multi-frequency atmospheric microwave radiometer observation brightness temperature training dataset, along with the 17-layer atmospheric temperature or humidity profiles output from the convolutional neural network, are input into a single-layer bidirectional long short-term memory (LSTM) network to output the final inverted 17-layer atmospheric temperature or humidity profiles. The Tanh function is chosen as the activation function for the output layer, and the Adam optimizer is selected.

[0041] Finally, in this embodiment, a deep learning method is used to train the atmospheric temperature and humidity profile inversion model, achieving high-efficiency and high-precision parameter inversion. An example of the obtained atmospheric temperature and humidity profile results is shown below. Figure 4 and Figure 5 As shown.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, alterations, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fast temperature and humidity profile retrieval method for a multi-frequency atmospheric microwave radiometer, characterized in that, The method comprises the following steps: S1, data acquisition and matching: obtaining a brightness temperature data set of a multi-frequency atmospheric microwave radiometer in a target area, and performing time and space matching with corresponding ERA5 reanalysis data or sounding balloon temperature and humidity profile data to construct a matched data set; dividing the matched data set into a training set, a validation set and a test set; S2, model construction: based on a CNN-LSTM combined neural network architecture, an atmospheric temperature and humidity profile inversion model is established; the CNN-LSTM combined neural network comprises a convolutional neural network for extracting spectral features of multi-frequency brightness temperature data, a single-layer bidirectional long short-term memory network for processing atmospheric temperature and humidity profile height correlation characteristics between brightness temperatures, and a fully connected network for outputting final inversion values; S3, model training and optimization: the atmospheric temperature and humidity profile inversion model constructed in step S2 is trained using the training set, and the validation set is used to monitor the training process and guide the optimization of model hyperparameters to prevent overfitting; S4, profile inversion: inputting current multi-frequency atmospheric microwave radiometer brightness temperature data into the atmospheric temperature and humidity profile inversion model trained and optimized in step S3, and directly outputting the corresponding atmospheric temperature and humidity profile.

2. The fast temperature and humidity profile retrieval method for multi-frequency atmospheric microwave radiometer according to claim 1, characterized in that, The step S1 specifically comprises: obtaining a brightness temperature data set observed by an atmospheric microwave radiometer in a target area; selecting matching ERA5 reanalysis data or sounding balloon temperature and humidity profile data according to the time and space resolution of the brightness temperature data; setting appropriate time and space window sizes to perform time and space matching on the brightness temperature data and the ERA5 reanalysis data or the sounding balloon temperature and humidity profile data, ensuring one-to-one correspondence between the two in terms of observation time and observation spatial position, and constructing a matched data set.

3. The fast temperature and humidity profile retrieval method for multi-frequency atmospheric microwave radiometer of claim 1, wherein, The specific structure of the CNN-LSTM combined neural network in step S2 is as follows: The convolutional neural network CNN receives multi-frequency brightness temperature data as input, extracts local features and abstract spectral features of the brightness temperature data in the frequency dimension through convolutional layers and pooling layers; the atmospheric temperature or humidity profile sequence extracted by CNN and the multi-frequency brightness temperature are input into the single-layer bidirectional long short-term memory network LSTM, and the LSTM network learns the dynamic characteristics of the atmospheric temperature and humidity profile with respect to frequency changes contained in the multi-frequency brightness temperature; the final state output of the LSTM network is input into a fully connected network FC for mapping the learned features to the final atmospheric temperature and humidity profile inversion values.

4. The fast temperature and humidity profile retrieval method for multi-frequency atmospheric microwave radiometer of claim 1, wherein, The model training process in step S3 is based on the error backpropagation algorithm, including two stages of forward propagation and backpropagation: The forward propagation is to pass the input brightness temperature from the input layer to the output layer layer by layer to obtain the inversion atmospheric temperature and humidity profile value; The backpropagation is to calculate the error between the inversion value and the true label value from the matched data set, and to update the connection weights and bias terms of each layer in the neural network by using the gradient descent method.

5. The fast temperature and humidity profile retrieval method for multi-frequency atmospheric microwave radiometer of claim 4, wherein, The value reaching the hidden layer node j in the forward propagation process and the value of the output layer node l is calculated using the following formula: (1) (2) Wherein: : input value of the jth node of the hidden layer; : output layer first l node input value; : output value of the input layer i-th node; OH j : activation output value of the jth node of the hidden layer : input layer i-th node to hidden layer j-th node connection weight; : connection weight of the jth node of the hidden layer to the ith node of the output layer l : connection weight of the jth node of the hidden layer to the ith node of the output layer : bias of the jth node of the hidden layer; : output layer bias for the l node. ni: total number of input layer nodes; nh: total number of hidden layer nodes.

6. The fast temperature and humidity profile retrieval method for multi-frequency atmospheric microwave radiometer of claim 4, wherein, In the backpropagation process, the error calculation and weight and bias update formulas are as follows: Overall error calculation: (3) Weight update: (4) Bias update: (5) Wherein: : model's inverse value for the kth sample at the jth output node l : model's inverse value for the kth sample at the jth output node : true label value of the kth sample l output node; η: learning rate, controlling parameter update step size; : weight before update; : updated weight; : bias before update; : updated bias; : partial derivative of error E with respect to weight parameter ; : partial derivative of error E with respect to bias parameter ​ no: total number of output layer nodes.

Images