A fusion elevation random forest fy-4a water vapor correction method

Abstract

The application provides a fusion height random forest FY-4A water vapor correction method, comprising the following steps: FY-4A LPW data is matched with the geographic coordinates of observation stations to obtain the FY-4A PWV corresponding to the observation station, observation data, broadcast ephemeris files, precise ephemeris files and clock error files are imported into RTKLIB to obtain the value of wet delay. The calculated wet delay is multiplied by the water vapor conversion factor ∏ to calculate the PWV above the observation station. A random forest water vapor model is constructed; the inverse distance weighting method is used to calculate and determine the ground height corresponding to the FY-4A LPW product pixel, the elevation information is used to calculate the elevation coordinates of each pixel point, and then the random forest water vapor model is brought in for correction to obtain the FY-4A PWV change. The application provides a higher-precision water vapor product for three-dimensional water vapor tomography and improves the purpose of improving the inversion three-dimensional water vapor precision.

Description

Technical Field

[0001] This invention provides a random forest FY-4A water vapor correction method that integrates elevation, belonging to the field of atmospheric physics technology. Background Technology

[0002] Although water vapor accounts for only 0.1% to 3% of the atmosphere, its spatiotemporal variations on a large scale affect the vertical stability of the atmosphere. In atmospheric physics, water vapor content is one of the most important parameters of the Earth's atmosphere. Water vapor content and its spatiotemporal variations play a crucial role in conventional weather forecasting, monitoring, analysis, and forecasting of severe weather such as heavy rainfall, global or regional climate change, air quality monitoring, hydrology, and oceanography. It is also closely related to weather events such as floods, droughts, and heavy rainfall. Currently, the main methods for obtaining atmospheric water vapor information include radiosonde, ground-based GNSS inversion, microwave radiometers, solar photometers, and satellite remote sensing. Among these, satellite remote sensing has advantages such as high spatiotemporal resolution, wide coverage, and high speed and efficiency, and has broad application prospects. However, it is also constrained by factors such as satellite orbit, cloud interference, and radiative transfer models, which affect the accuracy of the obtained atmospheric water vapor content. GNSS features all-weather operation, high precision, immunity to weather conditions, near real-time operation, wide coverage, and low cost. Furthermore, real-time, high-precision two-dimensional water vapor inversion based on GNSS technology is already quite mature. Related scholars have also used a combination of the four major global satellite navigation systems to perform real-time water vapor inversion, achieving an inversion accuracy of 1–3 mm.

[0003] By utilizing a combination of the four major global navigation satellite systems for real-time water vapor retrieval, potential gross errors are avoided, resulting in highly reliable water vapor products suitable for correcting water vapor products obtained through remote sensing methods, thus further improving the quality of remote sensing water vapor products. For example, for MODIS water vapor correction, water vapor data obtained through GNSS inversion is used to establish water vapor correction models using first-order polynomials, second-order polynomials, exponential functions, logarithmic functions, and power functions. MODIS is a medium-resolution imaging spectrometer; although it has high spatial resolution, it is affected by uncertainties such as precipitation, cloud cover, and surface reflectance spectra, resulting in relatively large errors, generally exceeding 10%.

[0004] Currently, there is limited research on FY-4A water vapor correction. However, using GNSS water vapor data to correct FY-4A water vapor data can improve the accuracy of FY-4A water vapor data and can be applied to research requiring high-precision water vapor products. Summary of the Invention

[0005] In many studies or experiments, many water vapor products cannot meet the accuracy required for applications. To overcome this deficiency, FY-4A water vapor data is introduced. Using GNSS water vapor data, a suitable correction model is analyzed, and elevation data is fused. A random forest algorithm is introduced to correct the FY-4A water vapor data.

[0006] There is limited research on FY-4A water vapor correction. Therefore, this invention analyzes a suitable correction model, integrates elevation data, introduces a random forest algorithm, and uses the PWV obtained from GNSS inversion to correct the FY-4A PWV, in order to obtain water vapor products with higher accuracy.

[0007] The specific technical solution is as follows:

[0008] A random forest FY-4A water vapor correction method incorporating elevation includes the following steps:

[0009] S1. Obtain observational and measured meteorological data from the observation station, and download broadcast ephemeris files, precise ephemeris files, and clock error files. Download the stratified water vapor product LPW from the Fengyun satellite remote sensing data network.

[0010] By matching the S2 and FY-4A LPW data with the geographic coordinates of the observation station, the corresponding FY-4A PWV of the station can be obtained.

[0011] S3. Import the observation data, broadcast ephemeris file, precise ephemeris file, and clock difference file into RTKLIB for data processing. Estimate the absolute tropospheric delay (ZTD) value using the Precise Point Positioning (PPP) technique based on non-differential observations.

[0012] S4. The dry delay (static delay, ZHD) can be calculated by combining the Saastamoinen model with the actual meteorological data from the station. The wet delay value can be obtained based on the relationship between the dry delay, wet delay (ZWD) and tropospheric delay.

[0013] S5. Multiply the calculated wet delay by the water vapor conversion factor ∏, and calculate the atmospheric precipitable water volume (PWV) above each station using the following formula.

[0014]

[0015] In the formula, k2' = 16.48 K * hPa -1 k3 = 3.75 x 10⁵ K² * hPa -l Both represent the atmospheric refractive index constant, R. v This indicates that the water vapor gas constant is 461.53 J / (kg*K), T m This is the atmospheric weighted average temperature.

[0016] S6. After obtaining the GNSS PWV and FY-4A PWV, a correlation analysis was first performed between the two, and a water vapor model calibration was constructed. First-order polynomials, second-order polynomials, exponential functions, logarithmic functions, and power functions were proposed to establish the water vapor calibration model. 70% of all sample data was randomly extracted as the training set, and the remaining 30% as the validation set. A comparative analysis of the fitting results obtained from the training set data showed that the calibration effects of the five models were generally average, and the graph results of the other functions showed that they all approached the first-order polynomial. The calculated R² (coefficient of determination, R0) was... 2 The values ​​are also very close, indicating that these correction models have a consistent fit. However, some studies have found that the correlation coefficient is negatively correlated with altitude. Therefore, we consider introducing an elevation system and constructing a multiple regression correction model based on a first-order polynomial, as shown in the following equation:

[0017] GNSS pwv =a0·FY4A pwv +a1·H+b

[0018] In the formula, GNSS pwv This represents the PWV, FY4A obtained through GNSS inversion. pwv H represents the atmospheric precipitable water obtained from FY-4A, and H represents the elevation.

[0019] We considered introducing random forest to explore the nonlinear relationship between two types of water vapor data and elevation. Experimental comparisons clearly show that using random forest to correct FY-4A water vapor data is better than a general linear model, although the overall bias is slightly higher. Adding elevation data as an influencing factor to the model reduces RMSE and MAE after correction, improving accuracy. Therefore, the random forest correction model incorporating elevation data is superior to the random forest model without elevation data, and also superior to linear models such as quadratic polynomials and power functions. It better reflects the relationship between GNSS PWV and FY-4A PWV, and this water vapor correction model can be applied to FY-4ALPW products to improve water vapor data.

[0020] S7. Using the above water vapor correction model, all pixel values ​​within the study area of ​​the FY-4A water vapor product can be corrected. To determine the elevation corresponding to these pixels, the inverse distance weighting method is used to calculate the ground height corresponding to the pixels in the FY-4A LPW product. The elevation coordinates of each pixel point are calculated using the elevation information and then substituted into the random forest water vapor model for correction to obtain the FY-4A PWV variation. Detailed Implementation

[0021] The specific technical solutions of the present invention will be described with reference to the embodiments.

[0022] This embodiment uses continuous observation data from the SatRef system in Hong Kong in June 2023 for research. GNSS inversion water vapor data is calculated by estimating the delay of satellite signal lines in the atmosphere and combining it with an empirical model to calculate atmospheric precipitable water.

[0023] The specific technical process involves estimating the absolute tropospheric delay (ZTD) value using precise point positioning (PPP) techniques based on non-differential observations. Then, the dry delay (static delay, ZHD) is calculated using the Saastamoinen model combined with measured meteorological data from the stations. Based on the relationship between dry delay, wet delay (ZWD), and tropospheric delay, the wet delay value is obtained. Finally, multiplying by the water vapor conversion factor π, the PWV over each station in the Hong Kong area is calculated using the following formula.

[0024]

[0025] In Equation 1, k2' = 16.48 K * hPa -1 k3 = 3.75 x 10⁵ K² * hPa -l Both represent the atmospheric refractive index constant, R. v This indicates that the water vapor gas constant is 461.53 J / (kg*K), T m The atmospheric weighted average temperature is used. For ease of calculation, Chen Yongqi's weighted temperature model is directly adopted, as shown in the following formula:

[0026] Tm = 106.7 + 0.605t s

[0027] In the formula t s This indicates the ground temperature, expressed in Kelvin (K).

[0028] The remote sensing water vapor product being calibrated is the stratified water vapor product LPW provided by the FY-4A satellite. This product is provided by the Advanced Geostationary Radiometric Imager (AGRI) carried by the next-generation geostationary meteorological satellite Fengyun-4A (FY-4A), one of the world's most advanced next-generation geostationary meteorological satellites. The LPW product's water vapor unit is g / kg, which can be assimilated to cm, with a temporal resolution of up to 15 minutes and a spatial resolution of 4 km. Since the Fengyun satellite data network does not directly provide geolocation files, but only latitude and longitude lookup tables at different resolutions and using data from the Chinese region, it is necessary to convert the row and column numbers in the product to full-disk row and column numbers, and then refer to the latitude and longitude lookup table to find the pixel that matches the latitude and longitude of the GNSS station. The value of this pixel is the FY-4A PWV corresponding to the station.

[0029] After obtaining the GNSS PWV and FY-4A PWV data, a correlation analysis was first performed between the two, and a water vapor model calibration was constructed. The proposed models used included first-order polynomials, second-order polynomials, exponential functions, logarithmic functions, and power functions. 70% of the data was randomly extracted from all samples as the training set, and the remaining 30% as the validation set. A comparative analysis of the fitting results obtained from the training set data showed that the calibration effects of all five models were generally average, and the graph results of the other functions tended to approximate the first-order polynomial. The calculated R² (coefficient of determination) was... 2 The values ​​are also very close, indicating that the fitting degree of these calibration models is consistent.

[0030] However, some studies have found that the correlation coefficient is negatively correlated with altitude. Therefore, we consider introducing an elevation system and constructing a multivariate regression correction model based on a first-order polynomial, as shown in the following formula.

[0031] GNSS pwv =a0·FY4A pwv +a1·H+b

[0032] In the formula, GNSS pwv This represents the PWV, FY4A obtained through GNSS inversion. pwv H represents the atmospheric precipitable water obtained from FY-4A, and H represents the elevation.

[0033] We considered introducing random forests to explore the nonlinear relationship between two types of water vapor data and elevation. A comparison clearly shows that using random forests to correct FY-4A water vapor data is more effective than a typical linear model, reducing RMSE by 16% and MAE by 20.5%, although the overall bias slightly increases. Adding elevation data as an influencing factor to the model further reduces RMSE to 1.926% and MAE to 1.21%, improving accuracy by 24%. Therefore, the random forest correction model incorporating elevation data is superior to the random forest model without elevation data, and also superior to linear models such as quadratic polynomials and power functions. It better reflects the relationship between GNSS PWV and FY-4APWV. This water vapor correction model can be applied to the Hong Kong region of the FY-4ALPW product to improve water vapor data.

[0034] The above-mentioned water vapor correction model can be used to correct all pixel values ​​of the FY-4A water vapor product within the Hong Kong region. To determine the elevation corresponding to these pixels, the inverse distance weighting method is used to calculate the ground height corresponding to the pixels in the FY-4ALPW product. The elevation coordinates of each pixel point are calculated using the elevation information of Hong Kong region, and then substituted into the random forest water vapor model for correction, so as to obtain the FY-4APWV change in Hong Kong region in June.

Claims

1. A random forest FY-4A water vapor correction method integrating elevation, characterized in that, Includes the following steps: S1. Obtain observational and measured meteorological data from the observation station, and download broadcast ephemeris files, precise ephemeris files, and clock error files; download stratified water vapor product LPW from the Fengyun satellite remote sensing data network; By matching the S2 and FY-4A LPW data with the geographic coordinates of the observation station, the FY-4A PWV corresponding to the station can be obtained. S3. Import the observation data, broadcast ephemeris file, precise ephemeris file and clock difference file into RTKLIB for data processing, and estimate the absolute tropospheric delay value ZTD through the precise single-point positioning (PPP) technology of non-differential observation. S4. The dry delay ZHD is then calculated using the Saastamoinen model combined with the actual meteorological data from the station. The wet delay value can be obtained based on the relationship between the dry delay, wet delay ZWD and tropospheric delay. S5. Multiply the calculated wet delay by the water vapor conversion factor ∏ to calculate the atmospheric precipitable water volume (PWV) above the station. ， In the formula =16.48 K*hPa -1 , =3.75X10 5 K 2 *hPa -l Both represent the atmospheric refractive index constant, R. v This indicates that the water vapor gas constant is 461.53 J / (kg*K), T m The atmospheric weighted average temperature; S6. After obtaining GNSS PWV and FY-4A PWV, perform correlation analysis between the two and construct a water vapor model correction. Introduce the elevation system, construct a multivariate regression correction model based on a first-order polynomial, and add the random forest algorithm to form a random forest water vapor correction model. The constructed multivariate regression correction model is shown in the following equation: ， In the formula, GNSS pwv This represents the PWV, FY4A obtained through GNSS inversion. pwv H represents the atmospheric precipitable water obtained from FY-4A, and H represents the elevation. S7. Use the inverse distance weighting method to calculate and determine the ground height corresponding to the pixels in the FY-4A LPW product. Calculate the elevation coordinates of each pixel using the elevation information, and then substitute them into the random forest water vapor model for correction to obtain the FY-4A PWV variation.

Images