GNSS-based tropospheric delay modeling method for desert, Gobi and arid drought regions

Abstract

To fully consider the characteristics of an arid region, thus improving the accuracy and reliability of a troposphere delay model in an extreme region.SOLUTION: A modeling method incudes the steps of: calculating a troposphere zenith hydrostatic delay ZHD and a troposphere zenith wet delay ZWD by analyzing a GPT-3 model, and summing to obtain a troposphere zenith total delay ZTD; inversely analyzing the ZTD by utilizing a ground-based GNSS, and obtaining a high-precision ZWD inverted by the GNSS by calculating a difference between the ZTD and the troposphere zenith hydrostatic delay ZHD obtained by a Saastamonien model; calculating an atmospheric weighted average temperature T'm by inversely analyzing an ERA5 data set, and performing linear correction to an estimated prediction value Tm of a global pressure and temperature model to obtain a corrected temperature Trm; calculating a priori value ZWD0 of the troposphere zenith wet delay utilizing an Askne model; correcting ZWD0 by utilizing a bidirectional long and short term memory network BiLSTM, and obtaining a high-precision ZTD value corrected by adding the corrected ZWD0 to the ZHD obtained in the GPT-3 model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to the field of navigation positioning, and in particular to a tropospheric delay modeling method for desert, Gobi and arid drought areas based on GNSS. [Background technology]

[0002] Tropospheric zenith delay (Zenith Total Delay, ZTD) is one of the major error sources in high-precision positioning using Global Navigation Satellite Systems (GNSS) and is an important data source for studying the GNSS spatial environment. In GNSS positioning, tropospheric zenith delay is the delay that occurs when a satellite signal passes through the atmosphere and reaches the receiving antenna. This delay is caused by factors such as water vapor, dry air, and temperature gradients in the atmosphere. Therefore, the tropospheric zenith delay varies depending on weather conditions. Especially at low elevation angles such as a cutoff angle of 5°, the signal passes through a thicker atmosphere, resulting in a high slant total delay of 20 meters, which is a significant error source for high-precision positioning.

[0003] To achieve high-precision GNSS positioning, it is crucial to develop accurate tropospheric delay correction models. These correction models require estimating the tropospheric zenith delay based on GNSS observation data, taking into account factors such as temperature, humidity, and pressure in the atmosphere. Accurate measurement and correction of the tropospheric zenith delay is also crucial for meteorological and climate research, as it provides data on the distribution and dynamic changes of atmospheric water vapor across the globe, contributing to research in areas such as weather forecasting, climate simulation, and natural disaster monitoring. Therefore, developing accurate tropospheric delay correction models is crucial.

[0004] Tropospheric delay correction models can be divided into two types: models based on meteorological parameters and models independent of meteorological parameters. Currently, models based on observed meteorological parameters are widely used, including the Saastamoinen model, Hopfield model, and Black model. However, these models require accurately measured meteorological parameters as input data, which can limit their applicability. Currently, models independent of meteorological parameters are being actively researched and used, including the Global Pressure and Temperature (GPT) model, which is widely used. The third-generation Global Pressure and Temperature (GPT-3) model has improved stability and accuracy compared to previous generations.

[0005] Droughts are indeed a relatively complex situation, due to the extremely low humidity in the troposphere in such environments, which significantly differs from standard empirical models. Therefore, it is necessary to modify the tropospheric zenith delay model to meet the needs of drought regions. GNSS has become a novel method for obtaining the tropospheric zenith delay with the advantages of high accuracy and low cost, and it allows real-time inversion of tropospheric precipitable water vapor (PWV) based on GNSS ground observation data. Summary of the Invention [Problem to be solved by the invention]

[0006] The technical problem of the present invention is that the accuracy and reliability of conventional tropospheric models are low in drought regions.

[0007] To solve the above problem, the present invention aims to provide a GNSS-based tropospheric delay modeling method for desert, Gobi, and desert drought regions, which uses the tropospheric delay actually observed at a regional GNSS station and ERA5 meteorological data to correct the atmospheric mean weighted temperature in the model, obtain a highly accurate prior value of the tropospheric zenith moist delay, and then use a BiLSTM neural network to correct the prior value of the tropospheric zenith moist delay to obtain a corresponding corrected model, thereby constructing a highly accurate regional tropospheric delay model for drought situations. [Means for solving the problem]

[0008] The technical means of the present invention is Using the global temperature and pressure model, atmospheric pressure P, atmospheric weighted mean temperature Tm, and water vapor pressure lapse rate λ are calculated based on the geographical location and time of the observation station. m Step 1: Obtain the water vapor pressure e and analyze it based on the meteorological parameters obtained from the global temperature and pressure model to calculate the tropospheric zenith hydrostatic delay ZHD and the tropospheric zenith moist delay ZWD, and add ZHD and ZWD to obtain the tropospheric zenith delay ZTD; Step 2: Calculate ZTD by inversion using ground-based GNSS, input the pressure of the Global Climate 5th Generation Atmospheric Reanalysis dataset into the Saastamonien model to obtain the tropospheric zenith hydrostatic delay ZHD, calculate the difference between ZTD and ZHD, and obtain the highly accurate tropospheric zenith wet delay ZWD by GNSS inversion; Inversion of atmospheric weighted mean temperature T' based on water vapor pressure and temperature from the Global Climate 5th Generation Atmospheric Reanalysis Dataset m Calculate T′ m T calculated by the global temperature and pressure model in step 1 using m The corrected temperature T is linearly corrected for r m Step 3 to get T obtained in step 3 r m Step 4 calculates the a priori value ZWD0 of the tropospheric zenith moist delay using the Askne model. A bidirectional long short-term memory network (BiLSTM) is used to construct a tropospheric zenith moist delay correction model. The a priori value ZWD0 of the tropospheric zenith moist delay obtained in step 4 is used as input data for the correction model. The tropospheric zenith moist delay ZWD obtained by the GNSS inversion analysis in step 2 is used as the output result of the correction model, which is the corrected tropospheric zenith moist delay ZWD. m Step 5: construct a training dataset for the tropospheric zenith wet delay correction model as the reference data, and set a loss function to constantly optimize the BiLSTM; Tropospheric zenith moist delay correction value ZWD obtained by the tropospheric zenith moist delay correction model in Step 5 m and step 6 of calculating a corrected high-precision tropospheric zenith delay ZTD by summing it based on the ZHD calculated by the global temperature and pressure model in step 1.

[0009] Preferably, the global temperature and pressure model is the third generation global temperature and pressure model GPT-3.

[0010] Furthermore, in step 1, the calculation formulas for the tropospheric zenith wet delay ZWD and the tropospheric zenith hydrostatic delay ZHD are as follows:

number

[0011] Preferably, in step 3, the atmospheric weighted mean temperature T' is calculated by inversion based on the water vapor pressure and temperature of the Global Climate Fifth Generation Atmospheric Reanalysis Dataset. m Calculate T′ m The inverse analysis model is as follows:

number

[0012] Furthermore, in step 4, the formula for obtaining ZWD0 using the Askne model is as follows:

number

[0013] Furthermore, the BiLSTM model is as follows: h t =f(w1X t +w2h t-1 ) h t ′=f(w3X t +w5h t+1 ) o t =g(w4h t +w6h t ) However, h t , h t ' indicates the output of the forward layer and backward layer at time t, respectively, and o t indicates the output of the BiLSTM model at time t, f( ) and g( ) are activation functions, w1, w2, w3, w4, w5, and w6 are weight parameters, and h t-1 , h t+1 indicates the output at the previous time and the next time, respectively, and X t indicates the input of the BiLSTM model at time t. [Effects of the Invention]

[0014] Compared with the prior art, the present invention includes the following beneficial effects: 1) The present invention has the advantages of high accuracy and fast analytical calculation. It uses ERA5 meteorological data to correct the atmospheric mean weighted temperature to obtain a highly accurate a priori value of the tropospheric zenith moist delay, and then corrects it using the tropospheric zenith moist delay value obtained by GNSS inversion analysis of the corresponding area. This can fully take into account the characteristics of drought areas and improve the reliability of the tropospheric delay model in extreme areas. 2) The present invention uses a bidirectional long short-term memory network (BiLSTM) to construct a correction model for the tropospheric zenith moist delay, sets a loss function, and performs optimization training on the tropospheric zenith moist delay correction model using the tropospheric zenith moist delay values ​​from past GNSS inversion analysis and the tropospheric zenith moist delay values ​​obtained by the Askne model as training data, thereby further improving the accuracy of the tropospheric zenith moist delay. 3) The present invention does not require additional investment, does not require the establishment of a new observation station, but can utilize existing GNSS observation stations, and does not require the addition of additional data acquisition equipment. Simply by performing data processing training, it is possible to construct a tropospheric zenith delay correction model and obtain highly accurate tropospheric zenith delay. [Brief explanation of the drawings]

[0015] The invention will now be further described with reference to the figures and examples.

[0016] [Figure 1] 1 is a schematic diagram of a tropospheric delay modeling method for desert, Gobi, and arid drought regions based on GNSS according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0017] As shown in Figure 1, the GNSS-based tropospheric delay modeling method for desert, Gobi, and arid drought regions specifically includes the following steps 1) to 6).

[0018] 1) Using the GPT-3 model, atmospheric pressure P, atmospheric weighted mean temperature Tm, and water vapor pressure lapse rate λ are calculated based on the geographical location of the observation station, time, and other information. m , the water vapor pressure e is obtained as follows:

number

[0019] Based on the obtained meteorological parameters, the Saastamonien model is used to calculate the tropospheric zenith hydrostatic delay ZHD, and the Askne model is used to obtain the tropospheric zenith moist delay ZWD. The two are then added together to obtain ZTD.

number

[0020] By linking with historical data, we obtain past ZTD values, and correct ZWD because the water vapor content in drought-affected areas is extremely low, which can cause the ZWD calculated by the above model to have a very large error compared to the actual value. This step 1) is completed by computer.

[0021] 2) Calculate ZTD by inverse analysis using ground-based GNSS.

[0022] GNSS signals are affected by tropospheric refraction during propagation, and tropospheric delays are usually corrected using parameter estimation methods. Precise Point Positioning (PPP) performs parameter estimation for ZTD, enabling inversion accuracy to reach millimeter levels.

[0023] The method of calculating ZTD by inverse analysis using PPP technology eliminates the effects of satellite orbit and clock error by using precise satellite orbit and clock error data provided by IGS. During the analytical calculation process, tropospheric delay is included as an unknown parameter in the analytical calculation of the observation equation. The principle is to create and solve the observation equation by ionospheric-free coupling of the GNSS observation station pseudorange and carrier phase observation equation. The observation equation is as follows:

number

number

number

number

[0024] The parameters to be estimated in the observation equation are

number

number

number

[0025] Since the tropospheric delay includes the tropospheric zenith hydrostatic delay ZHD and the tropospheric zenith moist delay ZWD, ZWD = ZTD - ZHD.

[0026] The pressure data from the ERA5 dataset is substituted into the Saastamonien model to obtain the ZHD, and finally the ZWD is obtained by high-precision GNSS inversion. This step 2) is completed on a computer.

[0027] 3) Atmospheric weighted mean temperature T' calculated by inversion based on water vapor pressure and temperature from the ERA5 meteorological dataset mis calculated as shown in the following formula:

number

[0028] The atmospheric weighted mean temperature T obtained at different times m The atmospheric weighted mean temperature series T m、S Configure T m、S =(T m、1 ,T m、2 ,T m、3 ,…,T m、i ,…,T m、n ) However, T m、i where i=1, 2...n indicates the atmospheric weighted average temperature at the i-th time, and n indicates the total number of atmospheric weighted average temperature data at different times.

[0029] Because the modeling data source for the GPT-3 model is monthly average data and contains only annual and semiannual cycle terms, T m However, it cannot reflect information that changes with height, limiting its accuracy.

[0030] High-precision T' calculated by inverse analysis m The atmospheric weighted mean temperature T obtained by the GPT-3 model is m A linear correction to takes into account the effect of height changes on the value.

[0031] Specifically, T m The formula for linear correction to is as follows: T r m =aT m +b However, T r m denotes the corrected atmospheric weighted mean temperature, and a and b are the parameters to be estimated, calculated using the least squares method. This step 3) is completed by computer.

[0032] 4) Using the formula in step 3), the corrected atmospheric weighted mean temperature T r m We can obtain a new ZWD value using the Askne model, denoted as ZWD0 as a highly accurate prior value,

number

[0033] 5) A bidirectional long short-term memory network (BiLSTM) is used to construct a tropospheric zenith moist delay correction model, and the ZWD obtained by the past GNSS inversion analysis in step 2) is used as high-precision reference data for the tropospheric zenith moist delay correction model. The ZWD0 obtained in step 4) is used as input data for the tropospheric zenith moist delay correction model. The input data for the tropospheric zenith moist delay correction model further includes, but is not limited to, the longitude Lon, latitude Lat, pressure P, temperature T, and altitude H of the observation point.

[0034] Feature sequence matrix B f Create a

number

[0035] A BiLSTM dataset is constructed using the ZWD obtained by GNSS inversion in step 2) and the ZWD0 obtained in step 4), and is divided into a training set, a validation set, and a test set.

[0036] Because the number of features is too large, in order to accelerate the model training efficiency, the input data is normalized to ensure the training stability of the neural network.

[0037] The partial autocorrelation function (PACF) is used to calculate the correlation and reconstruct the input variables of the feature series. The PACF map of each feature is analyzed to observe whether there are significant peaks at different feature lag orders. Significant peaks indicate that this lag order has a large influence on the current feature. These lag orders are considered to be important time points related to the current feature.

[0038] Based on the PACF results, features with significant correlation with the target variable are selected. These features are believed to make an important contribution to the construction of the tropospheric zenith moist delay correction model, and are incorporated into the correction model to reduce the subjectivity of artificially setting the input variables of the tropospheric zenith moist delay correction model and improve the prediction accuracy. At the same time, series samples are generated, and the time series data is divided into windows of a specific length and used as the input for the BiLSTM model.

[0039] After completing the correlation reduction of the feature sequence, the weight matrix is ​​randomly initialized to break the symmetry, which improves the learning ability of the BiLSTM model for different features and prevents the problems of gradient explosion and gradient vanishing.

[0040] We constructed a BiLSTM model, and the main expression of BiLSTM is as follows: h t =f(w1X t +w2h t-1 ) h t ′=f(w3X t +w5h t+1 ) o t =g(w4ht +w6h t ) However, X t is the input at time t, and w1, w2, w3, w4, w5, and w6 are all weight values ​​that must be initialized before training. Random initialization is the process of randomly generating weight values ​​within a given range, and then training continuously adjusts these values. h t-1 is the output at the previous time, and h t is the output of the forward layer at time t, and h t+1 is the output at the next time, and h t ′ is the output of the backward layer at time t, and o t is the output at time t. By performing forward series fitting on the data and then performing backward fitting, the data set can be made more resistant to interference and the ZWD correction can have a more favorable processing effect.

[0041] The hyperparameters involved in the BiLSTM model are mainly four terms: the number of hidden layers, the number of neurons, the size of the batch processing sample, and the number of training iterations. Because these hyperparameters have a significant impact on the prediction accuracy of the BiLSTM model, it is necessary to optimize the hyperparameters during training to ensure that the subsequent prediction effect can reach the expected level.

[0042] The tropospheric zenith moist delay correction model is used to calculate the corrected value of the tropospheric zenith moist delay ZWD, and the ZWD value at time T is fitted using the data set from the previous time T-1 to perform a fitting prediction. The data from time T is then incorporated into the data set to predict the ZWD value at time T+1, and this is repeated in order to obtain the ZWD value for the entire series. For the ZWD value obtained using the BiLSTM model, the loss function is as follows:

number

[0043] Optimize the BiLSTM model parameters by calculating the loss function and updating the parameters through backpropagation, adapt the BiLSTM model to the mode of the data, and improve the accuracy of the BiLSTM model.

[0044] Set a threshold to evaluate the accuracy of ZWD values,

number

[0045] The ZWD values ​​obtained after reaching the accuracy requirement are denormalized to obtain the final ZWD values, and then the training is carried out again in reverse.

[0046] This step 5) is completed by computer.

[0047] 6) By simply adding the ZWD value obtained in step 5) and the ZHD calculated by the GPT-3 model, a corrected, highly accurate ZTD value can be obtained, completing the construction of a corrected model for tropospheric zenith delay ZTD in drought areas. ZTD m =ZWD m +ZHD However, ZTD m indicates the corrected high-precision tropospheric zenith delay, and ZWD m indicates the corrected high-precision tropospheric zenith wet delay. This step 6) is completed by computer.

[0048] In the embodiment, five IGS observation stations were selected to conduct the simulation experiment, among which WIND00NAM, ADIS00ETH, KITG00UZB, URUM00CHN and ALIC00AUS are drought area IGS observation stations selected within the global range, covering different regions of the Northern and Southern Hemispheres, as shown in Table 1.

[0049] Based on the observation data from the observation station from March to September 2022, the mean absolute deviation MAE and root mean square error RMS value of the tropospheric zenith delay ZTD calculated by the present invention are compared with the MAE and RMS value of the tropospheric zenith delay ZTD calculated by the GPT3 model, as shown in Table 2.

[0050] Table 1. Distribution of IGS observation stations [Table 1]

[0051] Table 2. Error table of ZTD values ​​for different models [Table 2]

[0052] Comparing the MAE and RMS values ​​of the tropospheric zenith delay ZTD obtained by the present invention with the MAE and RMS values ​​of the tropospheric zenith delay ZTD obtained by the GPT-3 model, the accuracy of the ZTD correction model of the present invention is improved to a certain extent compared to the GPT-3 model, and the modeling effect in the event of a drought is more favorable.

[0053] This invention uses high-precision GNSS data and ERA5 meteorological data to correct the GPT-3 tropospheric delay model, first fitting the Tm value to obtain a highly accurate a priori ZWD, and then combining it with the ZWD obtained from GNSS inversion analysis to correct the ZWD using the BiLSTM model to obtain a highly accurate regional tropospheric zenith delay model.

[0054] (Addendum) (Appendix 1) Using the global temperature and pressure model, atmospheric pressure P and atmospheric weighted mean temperature T are calculated based on the geographic location and time of the observation station. m , water vapor pressure decrement rate λ m Step 1: Obtain the water vapor pressure e and analyze it based on the meteorological parameters obtained from the global temperature and pressure model to calculate the tropospheric zenith hydrostatic delay ZHD and the tropospheric zenith moist delay ZWD, and add ZHD and ZWD to obtain the tropospheric zenith delay ZTD; Step 2: Calculate ZTD by inversion using ground-based GNSS, input the pressure of the Global Climate 5th Generation Atmospheric Reanalysis dataset into the Saastamonien model to obtain the tropospheric zenith hydrostatic delay ZHD, calculate the difference between ZTD and ZHD, and obtain the highly accurate tropospheric zenith wet delay ZWD by GNSS inversion; Inversion of atmospheric weighted mean temperature T' based on water vapor pressure and temperature from the Global Climate 5th Generation Atmospheric Reanalysis Dataset m Calculate T′ m T calculated by the global temperature and pressure model in step 1 using m The corrected temperature T is linearly corrected for r m Step 3 to get T obtained in step 3 r m Step 4 calculates the a priori value ZWD0 of the tropospheric zenith moist delay using the Askne model. A bidirectional long short-term memory network (BiLSTM) is used to construct a tropospheric zenith moist delay correction model. The a priori value ZWD0 of the tropospheric zenith moist delay obtained in step 4 is used as input data for the correction model. The tropospheric zenith moist delay ZWD obtained by the GNSS inversion analysis in step 2 is used as the output result of the correction model, which is the corrected tropospheric zenith moist delay ZWD. m Step 5: construct a training dataset for the tropospheric zenith wet delay correction model as the reference data, and set a loss function to constantly optimize the BiLSTM; Tropospheric zenith moist delay correction value ZWD obtained by the tropospheric zenith moist delay correction model in Step 5 m and step 6 of calculating a corrected high-precision tropospheric zenith delay ZTD by summing it based on the ZHD calculated by the global temperature and pressure model in step 1.

[0055] (Appendix 2) In step 1, the global temperature and pressure model is as follows: Atmospheric pressure P and atmospheric weighted mean temperature T are calculated based on the geographic location and time of the observation station.m , water vapor pressure decrement rate λ m , calculate the water vapor pressure e,

number

[0056] (Appendix 3) In step 1, the calculation formulas for the tropospheric zenith wet delay ZWD and the tropospheric zenith hydrostatic delay ZHD are as follows:

number

[0057] (Appendix 4) In step 2, ZTD is calculated by inverse analysis using ground GNSS, and the GNSS observation equation is as follows:

number

number

number

number

[0058] (Appendix 5) In step 3, the atmospheric weighted mean temperature T' is calculated by inversion based on the water vapor pressure and temperature of the Global Climate 5th Generation Atmospheric Reanalysis dataset. m Calculate T′ m The inverse analysis calculation model is as follows:

number

[0059] (Appendix 6) In step 3, T′ m T calculated by the global temperature and pressure model in step 1 using m The formula for linear correction is as follows: T rm =aT m +b where a is a proportionality coefficient, b is a constant parameter, and T r m denotes the corrected atmospheric weighted mean temperature, and T m A GNSS-based tropospheric delay modeling method for desert, Gobi, and arid drought regions described in Appendix 5, characterized in that indicates the atmospheric weighted average temperature calculated by a global temperature and pressure model.

[0060] (Appendix 7) In step 4, the calculation formula for obtaining ZWD0 using the Askne model is as follows:

number

[0061] (Appendix 8) In step 5, the BiLSTM model is as follows: h t =f(w1X t +w2h t-1 ) h t ′=f(w3X t +w5h t+1 ) o t =g(w4h t +w6h t ) However, h t , h t ' indicates the output of the forward layer and backward layer at time t, respectively, and o tindicates the output of the BiLSTM model at time t, f( ) and g( ) are activation functions, w1, w2, w3, w4, w5, and w6 are weight parameters, and h t-1 , h t+1 indicates the output at the previous time and the next time, respectively, and X t The GNSS-based tropospheric delay modeling method for desert, Gobi, and arid drought regions described in Appendix 7, characterized in that indicates the input of the BiLSTM model at time t.

[0062] (Appendix 9) In step 5, the loss function of the BiLSTM model is:

number

[0063] (Appendix 10) We set a threshold value and evaluated the accuracy of the tropospheric zenith wet delay correction value ZWDm output from the BiLSTM model.

number

Claims

1. Using the global temperature and pressure model, atmospheric pressure P and atmospheric weighted mean temperature T are calculated based on the geographical location and time of the observation station. ｍ , water vapor pressure decrement rate λ ｍ and water vapor pressure e, and calculates a first ZHD, which is a tropospheric zenith hydrostatic delay, and a first ZWD, which is a tropospheric zenith wet delay, based on meteorological parameters, which are the atmospheric pressure P, the atmospheric weighted mean temperature T m , the water vapor pressure lapse rate λ m , and the water vapor pressure e, obtained by a global temperature and pressure model, and adds the first ZHD and the first ZWD to obtain a first ZTD, which is a tropospheric zenith delay; Step 2: calculating the second ZTD, which is the tropospheric zenith delay, by inverse analysis using GNSS observation stations; inputting the atmospheric pressure of the ERA5 dataset into the Saastamonien model to obtain the second ZHD, which is the tropospheric zenith hydrostatic delay; and calculating the difference between the second ZTD and the second ZHD to obtain the second ZWD, which is the highly accurate tropospheric zenith wet delay, by GNSS inverse analysis. The atmospheric weighted mean temperature T' was calculated by inverse analysis based on the water vapor pressure and temperature of the ERA5 data set. ｍ Calculate T calculated by the global temperature and pressure model in step 1. ｍ The temperature T is corrected by linear correction to ｒ ｍ Step 3 to obtain T obtained in step 3 ｒ ｍ The Askne model uses the a priori value of the tropospheric zenith moist delay ZWD ０ Step 4: calculating The bidirectional long short-term memory network (BiLSTM) is used to construct a tropospheric zenith moist delay correction model, and the a priori value of the tropospheric zenith moist delay (ZWD) obtained in step 4 is used. ０ is used as input data for the correction model, and the second ZWD, which is the tropospheric zenith wet delay obtained by the GNSS inverse analysis in step 2, is used as the output result of the correction model, which is the tropospheric zenith wet delay correction value ZWD ｍ Step 5: constructing a training dataset of a tropospheric zenith wet delay correction model as the reference data of the BiLSTM, and setting a loss function to continuously optimize the BiLSTM; The tropospheric zenith moist delay correction value ZWD obtained by the tropospheric zenith moist delay correction model in Step 5 ｍ and step 6 of calculating a third ZTD, which is a corrected high-precision tropospheric zenith delay, by summing the first ZHD calculated by the global temperature and pressure model in step 1.

2. In step 1, the global temperature and pressure model is as follows: Based on the geographical location and time of the observation station, atmospheric pressure P and atmospheric weighted mean temperature T ｍ , water vapor pressure decrement rate λ ｍ , calculate the water vapor pressure e, [Equation 1] However, A ０ is the least squares mean, and A １ , B １ are all annual cycle parameters, and A ２ , B ２ The GNSS-based tropospheric delay modeling method for desert, Gobi and arid drought regions according to claim 1, characterized in that: are all semi-annual periodic parameters, doy is the day of the month, π is the ratio of the circumference of a circle to its circumference, and r() represents a function for estimating meteorological parameters.

3. In step 1, the calculation formulas for the first ZWD, which is the tropospheric zenith wet delay, and the first ZHD, which is the tropospheric zenith hydrostatic delay, are as follows: [Equation 2] where P is the atmospheric pressure at the observation station, φ is the latitude at the observation station, H is the ellipsoid height at the observation station, and k １ , k ２ are all atmospheric refraction constants, and R ｄ denotes the gas constant, and λ ｍ is the vapor pressure decrement rate, and T ｍ is the atmospheric weighted mean temperature, e is the water vapor pressure, and g ｍ The tropospheric delay modeling method for desert, Gobi and arid drought areas based on GNSS as claimed in claim 2, characterized in that:

4. In step 2, the second ZTD is calculated by inverse analysis using the GNSS observation station, and the GNSS observation equation is as follows: [Equation 3] however, [Equation 4] is the unit vector between the receiver at the observation station and the satellite, [Equation 5] is the coordinate component correction number vector, and T ｓ ｒ is the zenith tropospheric delay, λ is the carrier wavelength, and N ｓ，ｊ ＩＦ，ｒ is the ambiguity parameter, [Equation 6] denotes the receiver clock error parameter including the pseudorange hardware delay, and ε ｓ ｒ，Ｐ denotes the pseudorange-related noise, and ε ｓ ｒ，Ｌ denotes the carrier phase related noise, and P ＩＦ denotes the pseudorange observation value of the ionospheric free coupling after error correction, and L ＩＦ The GNSS-based tropospheric delay modeling method for desert, Gobi and desert drought regions according to claim 3, characterized in that: denotes the carrier phase observation value of the ionosphere-free coupling after error correction.

5. In step 3, the atmospheric weighted mean temperature T' is calculated by inverse analysis based on the water vapor pressure and temperature of the ERA5 dataset. ｍ Calculate T' ｍ The inverse analysis calculation model is as follows: [Equation 7] However, p ｗ indicates the water vapor pressure, T indicates the temperature, z indicates the altitude, and Z ｗ denotes the water vapor compression coefficient, and H ０ The tropospheric delay modeling method for desert, Gobi and arid drought areas based on GNSS as claimed in claim 4, characterized in that σ denotes the altitude of the observation station and s denotes the distance to the ground.

6. In step 3, T calculated by the global temperature and pressure model in step 1 is ｍ The formula for linear correction is as follows: T ｒ ｍ =aT ｍ +b where a is a proportionality coefficient, b is a constant parameter, and T ｒ ｍ denotes the corrected atmospheric weighted mean temperature, and T ｍ The GNSS-based tropospheric delay modeling method for desert, Gobi and desert drought regions according to claim 5, characterized in that Θ represents the atmospheric weighted average temperature calculated by the global temperature and pressure model.

7. In step 4, the ZWD is calculated using the Askne model. ０ The formula to obtain is as follows: [Equation 8] However, T ｒ ｍ denotes the corrected atmospheric weighted mean temperature, and k １ , k ２ are all atmospheric refraction constants, and R ｄ denotes the gas constant, and λ ｍ is the vapor pressure decrement rate, and T ｍ is the atmospheric weighted mean temperature, e is the water vapor pressure, and g ｍ The tropospheric delay modeling method for desert, Gobi and arid drought areas based on GNSS as claimed in claim 6, characterized in that:

8. In step 5, the BiLSTM model is as follows: h ｔ =f(w １ X ｔ +w ２ h ｔ－１ ) h ｔ ′=f(w ３ X ｔ +w ５ h ｔ＋１ ) o ｔ =g(w ４ h ｔ +w ６ h ｔ ) However, h ｔ , h ｔ ' indicates the output of the forward layer and backward layer at time t, respectively, and o ｔ indicates the output of the BiLSTM model at time t, f() and g() are activation functions, and w １ , w ２ , w ３ , w ４ , w ５ , w ６ are all weight parameters, and h ｔ－１ , h ｔ＋１ indicates the output at the previous time and the next time, respectively, and X ｔ The GNSS-based tropospheric delay modeling method for desert, Gobi and arid drought regions according to claim 7, characterized in that: denotes the input of the BiLSTM model at time t.

9. In step 5, the loss function of the BiLSTM model is: [Equation 9] However, ZWD ｍ，ｉ indicates the ZWD value calculated by the model at the i-th time, and ZWD ＧＮＳＳ，ｉ indicates the ZWD value back-analyzed by GNSS at the i-th time, N indicates the total amount of samples, i indicates the time, and LOSS bias The GNSS-based tropospheric delay modeling method for desert, Gobi and arid drought regions according to claim 8, wherein: denotes a loss function value.

10. A threshold is set to evaluate the accuracy of the tropospheric zenith wet delay correction value ZWDm output from the BiLSTM model. [Equation 10] where threshold indicates the threshold value, Δ indicates the accuracy result of the calculation, and ZWD ｍ，ｉ is the ZWD value calculated by the BiLSTM model at the i-th time, and ZWD ＧＮＳＳ，ｉ The GNSS-based tropospheric delay modeling method for desert, Gobi and desert drought regions according to claim 9, characterized in that: is a ZWD value calculated by inverse analysis using GNSS at the i-th time.

Images