CN113608239A - GNSS occultation troposphere parameter correction method based on BP neural network - Google Patents

Abstract

The invention relates to the field of atmospheric scientific research, in particular to a method for revising GNSS occultation troposphere parameters based on a BP neural network, the method comprising: receiving tropopause parameter product data obtained by acquisition and inversion by a GNSS occultation detector; The product data is preprocessed; the preprocessed data is input into the pre-established and trained correction model to obtain the corrected tropopause height and tropopause temperature; the correction model adopts a BP neural network. The invention uses the BP neural network method to correct the GNSS occultation tropopause parameter products for the first time, especially for the high latitude region, the error improvement effect is the most obvious, the use model is simple and efficient, the calculation is economical, and the high latitude region parameters of the GNSS occultation tropopause product can be effectively corrected. error, which improves the quality of GNSS occultation tropopause parameter products.

Description

Correction method of GNSS occultation tropospheric parameters based on BP neural network

technical field

The invention relates to the field of atmospheric scientific research, in particular to a method for correcting GNSS occultation tropospheric parameters based on a BP neural network.

Background technique

The tropopause is a hot spot in atmospheric climate research. GNSS occultation detection technology has the characteristics of high global coverage and high vertical resolution. The tropopause parameters retrieved from GNSS occultation detection data, hereinafter referred to as GNSS occultation tropopause parameters; tropopause parameters mainly include tropopause height and tropopause temperature).

The judgment method of the tropopause usually uses the temperature lapse rate judgment method defined by the World Meteorological Organization (WMO) in 1957, that is: the lowest point that satisfies the temperature lapse rate with the increase of altitude is greater than -2K/km, and this point The temperature lapse rate between any point within 2km above this point is greater than -2K/km.

At present, the four-dimensional variational data provided by the ECMWF business archives are model data with high accuracy, but the tropopause products obtained from the inversion of the occultation temperature profile and the results obtained from the model profile have obvious differences in some areas.

my country's Fengyun-3C (FY3C) GNSS Occultation Detector (GNOS) has been operating normally for 7 years, and the number of atmospheric temperature profiles provided is 400-500 per day. Compared with the ECMWF four-dimensional variational data, the tropopause height obtained from the FY3C satellite GNOS occultation data has a large negative deviation in high latitudes.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the defects of the prior art, and propose a method for correcting GNSS occultation tropospheric parameters based on BP neural network.

In order to achieve the above object, the present invention proposes a method for revising GNSS occultation tropospheric parameters based on BP neural network, the method comprising:

Receive the tropopause parameter product data acquired and retrieved by the GNSS occultation detector;

Preprocessing the tropopause parameter product data;

Input the preprocessed data into the pre-established and trained correction model to obtain the corrected tropopause height and tropopause temperature;

The correction model adopts BP neural network.

As an improvement of the above method, the tropopause parameter product data includes: tropopause temperature, tropopause height, profile latitude and longitude, and the corresponding collection date and time calculated according to the GNSS occultation stem temperature profile.

As an improvement of the above method, the preprocessing specifically includes:

Normalize the latitude and longitude of the profile and the corresponding collection date and time.

As an improvement of the above method, the input of the correction model is the preprocessed tropopause parameter product data, and the output is the corrected tropopause height and tropopause temperature. The adopted BP neural network includes 5 fully connected layers, Among them, 3 layers are hidden layers, each hidden layer has 10 neurons, and each neuron adopts the Relu activation function.

As an improvement of the above method, the method further includes a training step of revising the model, specifically including:

Select the FY3C occultation data and the temperature profile data of the ECMWF business archives, match them according to time, longitude and latitude, and calculate the tropopause height and tropopause temperature of the matched temperature profile through the temperature lapse rate algorithm to obtain the original sample set;

The data of the original sample set is screened, and the invalid data that cannot be determined due to the temperature profile of the tropopause are eliminated, and then preprocessed based on the sampling algorithm and the normalization algorithm, and the preprocessed sample set data is randomly assigned to the training according to a certain proportion. set and test set;

Initialize the weight matrix of the network, and set the summation weight of each pair of neurons connected in pairs;

The training set data is input into the BP neural network in turn, and the corresponding tropopause temperature and tropopause height are calculated according to the forward propagation. The matching ECMWF tropopause temperature and tropopause height are used as the reference true values, and the MSE is calculated from the calculation results and reference true values. For the loss function, the dynamic learning rate is adaptively adjusted by Adam, and the weight matrix is iteratively updated until the loss function converges and meets the preset maximum number of iterations, thereby obtaining a pre-trained correction model;

Input the test set data into the pre-trained correction model in turn, and judge whether the output result meets the evaluation requirements.

As an improvement of the above method, the data latitude distributions of the training set and the test set are kept consistent.

A correction system for GNSS occultation tropospheric parameters based on BP neural network, the system includes: a correction model, a receiving module, a preprocessing module and an output module; wherein,

The receiving module is used to receive the tropopause parameter product data obtained by the GNSS occultation detector acquisition and inversion;

The preprocessing module is used to preprocess the tropopause parameter product data;

The output module is used for inputting the preprocessed data into a pre-established and trained correction model to obtain the corrected tropopause height and tropopause temperature;

The correction model adopts BP neural network.

Compared with the prior art, the advantages of the present invention are:

The invention uses the BP neural network method to correct the GNSS occultation tropopause parameter products for the first time, especially for the high latitude region, the error improvement effect is the most obvious, the use model is simple and efficient, the calculation is economical, and the high latitude region parameters of the GNSS occultation tropopause product can be effectively corrected. error, which improves the quality of GNSS occultation tropopause parameter products.

Description of drawings

Fig. 1 is the overall flow chart of the correction method of GNSS occultation tropospheric parameters based on BP neural network of the present invention;

Fig. 2 is the BP neural network training flow schematic diagram of the present invention;

Figure 3 is an example of training set and test set data, including latitude and seasonal distribution;

Fig. 4 is a correction effect diagram using the method of the present invention, wherein Fig. 4(a) is the correction effect of the tropopause height, and Fig. 4(b) is the correction effect of the tropopause temperature.

Detailed ways

After investigation, it is found that the tropospheric parameters obtained from the GNSS occultation data of FY3C and other satellites have a large deviation, especially the tropopause height obtained from the GNSS occultation data has a large negative deviation compared with the ECMWF four-dimensional variational data in high latitude regions. , so try to correct the negative deviation of the GNSS tropopause parameter product in the high latitude area to achieve the purpose of improving the product accuracy.

Since the tropopause products are mostly grid-averaged or latitude-averaged, they are statistical results. In this regard, the BP neural network method is chosen to correct the GNSS occultation tropopause parameter products.

The purpose of the present invention is to provide a GNSS occultation tropopause parameter correction method based on BP neural network, which has the advantages of high calculation efficiency and good regression correction effect, and can effectively improve the quality of GNSS occultation tropopause parameter products. Negative bias for high latitude parameter product quality.

In order to use the BP neural network algorithm to improve the accuracy of GNSS occultation tropopause parameter products in high latitude areas, the technical solution mainly includes the following five steps, as shown in Figure 1:

The first step: the construction of the original data sample set. Select a large number of GNSS occultation temperature profile data and ECMWF four-dimensional variational model temperature profile data for space-time matching, calculate the tropopause height and temperature of the space-time matched GNSS occultation products and ECMWF data, and obtain the original sample;

Step 2: Generate training set and test set. First, the original sample data is screened to eliminate some invalid values caused by the inability of the temperature profile to determine the tropopause; after screening, the problem of uneven distribution of the data set and inconsistent dimensions is solved based on the sampling algorithm and the normalization algorithm, and then according to the 6.5:3.5 Divide the preprocessed sample set into training set and test set.

Step 3: Build a BP neural network model. Firstly, the number of neurons in the input layer and output layer of the model is determined according to the input and output characteristics, and then the number of hidden layers and neurons in the hidden layer of the neural network are determined according to continuous experiments. The normalized time, date, longitude and latitude are the input parameters, and the corrected tropopause height and temperature are the output parameters of the BP neural network model.

Step 4: BP neural network model training. The specific process is shown in Figure 2. First, optimize the model structure. After testing, the neural network uses a 5-layer fully connected neural network, which has 3 hidden layers, and each hidden layer has 10 neurons. The activation function of each neuron is then assigned, and the Relu function is used in this model. The weight matrix of the network is then initialized to give the summed weights of each pair of neurons connected in pairs. Taking the matched ECMWF tropopause temperature and tropopause height as reference values, set the loss function as MSE and the error back-propagation iterator algorithm as Adam, which uses an adaptive dynamic learning rate, starting from a smaller learning rate η Gradually improve the training accuracy. Subsequently, the number of epochs for training is determined based on the validation set error. The weight matrix is updated by training multiple times until the loss function converges or the maximum number of iterations is satisfied, and the number of iterations is 25,000.

Step 5: Correction of tropopause parameters. Use the trained BP neural network model to calculate the corrected tropopause height and temperature parameters to evaluate the correction effect.

The technical solutions of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

Example 1

Embodiment 1 of the present invention proposes a method for correcting GNSS occultation tropospheric parameters based on BP neural network.

Using the occultation observation data of the GNSS occultation detector (GNOS) on the FY3C satellite, the data is corrected by using the BP neural network based method of the present invention to correct the tropopause parameters in the high latitude area of the GNSS occultation of the FY3C satellite. Launched in September 2013, the FY3C satellite is a sun-synchronous orbit satellite with an orbital inclination of 98.8°, an average altitude of 836km, and an orbital period of 101.5 minutes. Its GNSS occultation receiver GNOS is compatible with Beidou Navigation Satellite System (BDS) signals and Global Positioning System (GPS) signals at the same time. The number of atmospheric temperature profiles provided during normal operational operations of FY3CGNOS is 400-500 per day.

The first step is to construct raw data samples: FY3C occultation data and ECMWF data are matched in space and time. In this example, the FY3C occultation data from June 2018 to August 2018 (JJA) and the temperature profile data of the ECMWF business archive are used for matching according to time, longitude and latitude, and the temperature after matching is calculated by the temperature decrement rate algorithm. The tropopause height and temperature of the profile were obtained to obtain the original sample set. Among them, 20187 sets of data in the 2018.6-2018.8 time period, and 10798 sets of data in the test set.

The second step is to generate a training set and a test set: screen the original sample set, remove the invalid data that cannot be determined due to the temperature profile of the tropopause, randomly select 65% as the training set, and the remaining 35% as the test set, to ensure two seasons of training sets , the latitude distribution of the test set data is consistent, and the specific latitude distribution is shown in Figure 3.

The third step is to build a BP neural network: choose to use a 5-layer BP neural network, including 3 hidden layers, choose the MSE function for the loss function, and use the Adam algorithm for error back propagation. Then, the evaluation index of the selected model is the root mean square error (RMSE), and a BP neural network model is built, which takes the parameters of the FY3C occultation tropopause before correction as the input and the tropopause parameters of the corrected FY3C occultation as the output.

Specifically include:

First build the model, which uses a 5-layer fully connected neural network with 3 hidden layers and 10 neurons in each hidden layer. The input data are FY3C tropopause height, temperature, and normalized time, date, longitude, and latitude. The output data is the corrected tropopause height and temperature.

After the activation of the Relu function, the dynamic learning rate is adaptively adjusted by Adam until the model obtains the output value, and the error between the output value and the target value is calculated, that is, the loss function. The network parameters are updated according to the error, and the cycle is repeated until the specified number of cycles is reached or the loss function meets the requirements. Train the model successfully, save the parameters.

The fourth step is to train the BP neural network model: initialize the weight matrix, input the relevant parameters of the FY3C occultation tropopause in the training set into the network, and take the corresponding ECMWF tropopause height and temperature of the training set as the target parameters. The weight matrix is updated through iterative training to make the MSE approach the minimum, and the number of iterations is 25,000.

The training set and the test set have the same latitude distribution, both at 80°N～80°S. The time is summer (2018.6-2018.8). The input data is the tropopause temperature, tropopause height, profile latitude and longitude, date, and time calculated from the FY3C satellite GNSS occultation stem temperature profile. The target data is the ECMWF four-dimensional variation that matches the FY3C satellite GNSS occultation stem temperature profile. Temperature profile calculated tropopause height and tropopause temperature. The latitude, longitude, date, and time data are normalized before feeding the training data into the machine learning model.

Step 5: Correction of tropopause parameters: Input the data to be measured into the BP neural network tropopause parameter correction model, obtain the corrected tropopause height and temperature parameters, and evaluate the correction of the model for the FY3C occultation tropopause parameters. On the whole, this method significantly reduces the deviation between the parameters of the FY3C occultation tropopause and the ECMWF results, and the specific correction is shown in Figure 4.

Figure 4 shows the decrease in the absolute value of the error between the FY3C star GNSS occultation data product and the ECMWF model data in each geographic grid before and after correction (the grid marked with an inverted triangle represents the increase in error), of which Figure 4(a) represents the tropopause Altitude correction effect, Figure 4(b) shows the tropopause temperature correction effect. During the summer (June-August), the neural network has an obvious effect on the correction of the parameters of the FY3C star GNSS occultation tropopause in the high latitudes of the northern hemisphere. Only individual grid points are marked, but in the high latitudes of the southern hemisphere, although the overall There is a certain correction effect on the grid, but there are also some grid points with markers attached.

Example 2

Embodiment 2 of the present invention proposes a BP neural network-based GNSS occultation tropospheric parameter correction system, which is implemented based on the method of Embodiment 1, and the system includes: a correction model, a receiving module, a preprocessing module, and an output module; in,

The receiving module is used to receive the tropopause parameter product data obtained by the GNSS occultation detector acquisition and inversion;

The preprocessing module is used to preprocess the tropopause parameter product data;

The output module is used for inputting the preprocessed data into a pre-established and trained correction model to obtain the corrected tropopause height and tropopause temperature;

The correction model adopts BP neural network.

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

Claims (7)

1. a correction method for GNSS occultation tropospheric parameters based on BP neural network, the method comprises: Receive the tropopause parameter product data acquired and retrieved by the GNSS occultation detector; Preprocessing the tropopause parameter product data; Input the preprocessed data into the pre-established and trained correction model to obtain the corrected tropopause height and tropopause temperature; The correction model adopts BP neural network. 2. the correction method of the GNSS occultation troposphere parameter based on BP neural network according to claim 1, is characterized in that, described tropopause parameter product data comprises: the tropopause temperature obtained according to GNSS occultation stem temperature profile calculation , tropopause height, profile latitude and longitude, and the corresponding collection date and time. 3. the correction method of the GNSS occultation troposphere parameter based on BP neural network according to claim 2, is characterized in that, described preprocessing specifically comprises: Normalize the latitude and longitude of the profile and the corresponding collection date and time. 4. the correction method of the GNSS occultation troposphere parameter based on BP neural network according to claim 1, is characterized in that, the input of described correction model is the tropopause parameter product data after preprocessing, and the output is the troposphere after correction The top height and the top temperature of the troposphere, the BP neural network used includes 5 fully connected layers, 3 of which are hidden layers, each hidden layer has 10 neurons, and each neuron uses the Relu activation function. 5. the revising method of the GNSS occultation troposphere parameter based on BP neural network according to claim 4, is characterized in that, described method also comprises the training step of revising model, specifically comprises: Select the FY3C occultation data and the temperature profile data of the ECMWF business archives, match them according to time, longitude and latitude, and calculate the tropopause height and tropopause temperature of the matched temperature profile through the temperature lapse rate algorithm to obtain the original sample set; The data of the original sample set is screened, and the invalid data that cannot be determined due to the temperature profile of the tropopause are eliminated, and then preprocessed based on the sampling algorithm and the normalization algorithm, and the preprocessed sample set data is randomly assigned to the training according to a certain proportion. set and test set; Initialize the weight matrix of the network, and set the summation weight of each pair of neurons connected in pairs; The training set data is input into the BP neural network in turn, and the corresponding tropopause temperature and tropopause height are calculated according to the forward propagation. The matching ECMWF tropopause temperature and tropopause height are used as the reference true values, and the MSE is calculated from the calculation results and reference true values. For the loss function, the dynamic learning rate is adaptively adjusted by Adam, and the weight matrix is iteratively updated until the loss function converges and meets the preset maximum number of iterations, thereby obtaining a pre-trained correction model; Input the test set data into the pre-trained correction model in turn, and judge whether the output result meets the evaluation requirements. 6. The method for correcting GNSS occultation troposphere parameters based on BP neural network according to claim 3, wherein the data latitude distribution of the training set and the test set is consistent. 7. A correction system for GNSS occultation tropospheric parameters based on BP neural network, characterized in that the system comprises: a correction model, a receiving module, a preprocessing module and an output module; wherein, The receiving module is used to receive the tropopause parameter product data obtained by the GNSS occultation detector acquisition and inversion; The preprocessing module is used to preprocess the tropopause parameter product data; The output module is used for inputting the preprocessed data into a pre-established and trained correction model to obtain the corrected tropopause height and tropopause temperature; The correction model adopts BP neural network.

Images