A method for suppressing noise of total polarized microwave radiometer T3 / T4 brightness temperature

Abstract

This invention discloses a method for suppressing brightness temperature noise in the T3 / T4 channels of a fully polarimetric microwave radiometer, belonging to the field of satellite microwave ocean remote sensing technology. The method first selects a pure ocean scene from the data products of the spaceborne fully polarimetric microwave radiometer, extracting multiple original brightness temperature images from the T3 / T4 channels and the corresponding observation geometric parameters for each image. Second, based on each original brightness temperature image, simulated brightness temperature images are generated, and then subjected to diffusion, noise addition according to quantized noise characteristics, and normalization. The processed simulated brightness temperature images and their corresponding unprocessed simulated brightness temperature images are combined to form training sample pairs, creating a simulated brightness temperature dataset. Finally, a brightness temperature suppression network model is established and trained using the simulated brightness temperature dataset. This invention solves the problem of significant noise in the brightness temperature products of the T3 / T4 channels of a spaceborne fully polarimetric microwave radiometer, resulting in higher-quality observed brightness temperature data.

Description

Technical Field

[0001] This invention relates to a method for suppressing brightness temperature noise in a fully polarized microwave radiometer T3 / T4, belonging to the field of satellite microwave ocean remote sensing technology. Background Technology

[0002] Sea surface wind fields are crucial parameters of the atmosphere and ocean, serving not only as the direct driving force behind sea wave generation but also as a major source of power for global and regional ocean circulation. They play a critical role in natural dynamic processes such as air-sea interaction, the marine environment, marine ecosystems, and regional and global weather change, making them essential for maritime weather warnings, marine disaster monitoring, and meteorological forecasting. For example, in weather forecasting, changes in sea surface wind intensity indicate changes in air pressure gradients, allowing for the prediction of many weather phenomena that originate at sea and then move to land. Sea surface wind field data is also an indispensable input variable in the early warning and monitoring of marine disasters such as storm surges and tsunamis. Furthermore, many human activities require sea surface wind field data, such as navigation and fisheries. Therefore, global sea surface wind field data is one of the fundamental data sources affecting efficient research on marine activities.

[0003] Traditional methods for acquiring sea surface wind fields include survey vessels, fixed oceanographic stations, and ocean buoys. While these methods provide high-precision wind field data, they suffer from limited observation range, short timescales, insufficient data acquisition, and uneven spatial distribution, failing to meet operational needs and lacking reliability under extreme weather conditions. With advancements in science and technology, satellite remote sensing has become a primary means of acquiring global marine environmental data. Among various satellite remote sensing methods, microwave radiometers, with their all-weather, all-time high-precision data acquisition capabilities, are one of the main methods for sea surface wind field remote sensing.

[0004] WindSat, the world's first spaceborne fully polarimetric microwave radiometer, demonstrated the feasibility of using the third and fourth Stokes parameters (T3 / T4) for remote sensing of sea surface wind fields. However, due to the relatively low T3 / T4 brightness temperatures, instrument noise has a greater impact on the observed data, resulting in visible random noise in T3 / T4 brightness temperature images across multiple frequency bands. This noise not only reduces the accuracy of brightness temperature data but is also a source of error in sea surface wind field data products that rely on high-precision brightness temperature input, thus limiting the accuracy of sea surface wind field inversion and the reliability of its data applications. Therefore, noise suppression processing of T3 / T4 brightness temperatures is of significant practical importance. The key to this problem is: (1) In current research, there is no clear frequency domain index to distinguish between signals and noise in the observed brightness temperature. Therefore, traditional methods (such as median filtering, Gaussian filtering, etc.) are prone to the loss of effective information; (2) The sea surface wind field inversion algorithm has high requirements for the accuracy of brightness temperature data. Conventional linear filtering or statistical denoising methods are prone to losing brightness temperature features and even introducing pseudo features; (3) Traditional denoising algorithms process data based on fixed thresholds or statistical characteristics. They have limited adaptability to different types of noise and image features and cannot adaptively adjust the denoising-related parameters. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a method for suppressing brightness temperature noise in the T3 / T4 channel of a fully polarized microwave radiometer. This method solves the problem of large noise in the brightness temperature products of the T3 / T4 channel of the spaceborne fully polarized microwave radiometer and obtains higher quality observation brightness temperature data.

[0006] The technical solution of this invention is:

[0007] A method for suppressing T3 / T4 brightness temperature noise in a fully polarized microwave radiometer includes:

[0008] S1: Select a pure ocean scene from the data products of the spaceborne fully polarized microwave radiometer, and extract multiple original observation brightness temperature images from the T3 / T4 channels and the observation geometric parameters corresponding to each observation brightness temperature image;

[0009] S2: Based on each original observed brightness temperature image, generate simulated brightness temperature images respectively;

[0010] S3: Perform noise suppression processing on each original observed brightness temperature image to obtain a denoised observed brightness temperature image; calculate the mean and standard deviation of the brightness temperature difference based on each original observed brightness temperature image and the corresponding denoised observed brightness temperature image, as the quantized noise characteristics.

[0011] S4: For each simulated brightness temperature image, perform diffusion, add noise according to the quantized noise characteristics, and normalize in sequence; combine each processed simulated brightness temperature image and the corresponding unprocessed simulated brightness temperature image to form a training sample pair, thus forming a simulated brightness temperature dataset.

[0012] S5: Establish a brightness temperature suppression network model to generate denoised observed brightness temperature images; train the brightness temperature suppression network model using a simulated brightness temperature dataset;

[0013] S6: Process the data from the spaceborne fully polarized microwave radiometer according to S1 to obtain the observed brightness temperature image to be analyzed; input the observed brightness temperature image to be analyzed into the trained brightness temperature suppression network model to obtain the denoised observed brightness temperature image.

[0014] Furthermore, the observation geometric parameters of the T3 / T4 channels are extracted. These parameters include the geographic coordinates of each observation point pixel, the zenith angle and azimuth angle when the satellite observes each observation point.

[0015] Furthermore, based on the original observed brightness temperature image, a simulated brightness temperature image is generated. Specific methods include:

[0016] Download ERA5 meteorological reanalysis data with the same observation time as the spaceborne fully polarized microwave radiometer;

[0017] Atmospheric and sea surface parameters from ERA5 meteorological reanalysis data are used as background field data and interpolated onto the original observed brightness temperature image to achieve spatiotemporal data matching.

[0018] Based on the observed geometric parameters and background field data of the original observed brightness temperature image, the simulated brightness temperature image was calculated using the RTTOV radiative transfer model.

[0019] Furthermore, based on the observation time of the spaceborne fully polarized microwave radiometer, ERA5 meteorological reanalysis data from the same adjacent days as the observation time are downloaded to form multiple background field data. Based on the observation geometric parameters of the original brightness temperature image and each background field data, a simulated brightness temperature image is generated, thus obtaining multiple simulated brightness temperature images that conform to the actual physical process.

[0020] Furthermore, the method for extracting the observation geometric parameters of the T3 / T4 channels to obtain the original observation brightness temperature image is as follows: For spaceborne fully polarimetric microwave radiometer data of pure ocean scenes, the observation geometric parameters of the T3 / T4 channels are extracted. According to the geographic coordinates of each observation point pixel in the observation geometric parameters, the observation image of the spaceborne fully polarimetric microwave radiometer of pure ocean scenes is cropped to obtain the required observation brightness temperature image.

[0021] Furthermore, diffusion processing is performed on each simulated brightness temperature image, specifically including:

[0022] Each pixel in the simulated brightness temperature image is marked as either cloudy or clear. If the proportion of pixels marked as clear exceeds a set threshold, the pixels marked as cloudy are removed, forming an invalid area, and the remaining area is the valid area. If the proportion of pixels marked as clear is less than a set threshold, the simulated brightness temperature image is discarded.

[0023] After filling the pixel illumination temperature value of the invalid region with 0, a Gaussian filter is used to smooth the boundary between the invalid region and the valid region, thus completing the diffusion process.

[0024] Furthermore, each pixel in the simulated brightness temperature image is marked as either cloudy or clear sky, with the marking principle being: if the following formula holds true, then it is cloudy; otherwise, it is clear sky.

[0025]

[0026] in, These refer to the 37.0 GHz vertical polarization brightness temperature of the pixel calculated by the RTTOV radiative transfer model in both all-sky and clear-sky modes.

[0027] Furthermore, the original observed brightness temperature image undergoes noise suppression processing, specifically including:

[0028] The observation-simulation difference analysis method was used to compare and analyze the brightness temperature values ​​of the original observed brightness temperature images and the corresponding simulated brightness temperature images under the T3 / T4 channels of each frequency band, and to quantify the noise characteristics of each channel:

[0029] For each channel, select several pairs of images whose original observed brightness temperature image and corresponding simulated brightness temperature image values ​​are closest, and calculate the mean and standard deviation of the brightness temperature difference as the noise characteristics of the quantization.

[0030] Furthermore, after performing diffusion processing on the simulated brightness temperature image, Gaussian noise is generated based on the obtained mean and standard deviation, and the Gaussian noise is superimposed on the simulated brightness temperature image.

[0031] Furthermore, a brightness temperature suppression network model based on the U-Net deep learning architecture is established, with the mean squared error (RMSE) loss function. The brightness temperature suppression network model is trained using a simulated brightness temperature dataset: the processed simulated brightness temperature image is input into the model, and the loss function value is calculated based on the denoised brightness temperature image output by the model and the unprocessed simulated brightness temperature image. The optimal brightness temperature suppression network model is obtained when the loss function value no longer decreases.

[0032] The advantages of this invention compared to the prior art are:

[0033] (1) This invention constructs a brightness temperature suppression network model to suppress instrument noise in the brightness temperature images of the T3 / T4 channels, solving the problem of decreased inversion accuracy caused by instrument noise in existing inversion techniques. It achieves good denoising effects in low-latitude, mid-latitude, and high-latitude regions, effectively improving the accuracy of T3 / T4 input data, and thus reducing wind speed and wind direction inversion errors. In some scenarios, the maximum improvement in wind direction error can reach 5°.

[0034] (2) Enhance the accuracy and stability of wind field inversion: This invention uses the RTTOV transmission model to simulate brightness temperature and artificially add noise to construct a dataset for training neural networks. This solves the problems of difficulty in obtaining data and difficulty in ensuring quality during model training, and realizes the independence of model training from observation data of real instruments.

[0035] (3) Good universality: This invention trains the U-Net model by constructing training data of different geographical regions and different meteorological conditions, which solves the problem of poor generalization ability of traditional methods for different geographical locations and meteorological conditions, and achieves a denoising effect with strong generalization ability.

[0036] (4) Provide end-to-end automated processing capability: Based on deep learning end-to-end modeling capability, this invention provides end-to-end automated processing capability. Compared with traditional denoising algorithms that rely on fixed thresholds or frequency domain statistics, it can adaptively identify and suppress noise features, reduce the need for human intervention, and fully preserve the structural information in the original brightness temperature image, making it suitable for automated processing of large-scale remote sensing data products. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0038] Figure 1 This is a flowchart of the brightness temperature noise suppression method for the T3 / T4 fully polarized microwave radiometer of the present invention;

[0039] Figure 2 The following is a comparison of the standard deviation of the observed brightness temperature of low-latitude scenes before and after processing by the noise suppression method in this embodiment of the invention: (a) 10.7GHz T3; (b) 10.7GHz T4; (c) 18.7GHz T3; (d) 18.7GHz T4; (e) 37.0GHz T3; (f) 37.0GHz T4;

[0040] Figure 3The following is a comparison of the standard deviation of the observed brightness temperature of mid-latitude scenes before and after processing by the noise suppression method in this embodiment of the invention: (a) 10.7GHz T3; (b) 10.7GHz T4; (c) 18.7GHz T3; (d) 18.7GHz T4; (e) 37.0GHz T3; (f) 37.0GHz T4;

[0041] Figure 4 The following is a comparison of the standard deviation of the observed brightness temperature of high-latitude scenes before and after processing by the noise suppression method in this embodiment of the invention: (a) 10.7GHz T3; (b) 10.7GHz T4; (c) 18.7GHz T3; (d) 18.7GHz T4; (e) 37.0GHz T3; (f) 37.0GHz T4;

[0042] Figure 5 The image effect comparison of the observed brightness temperature of a single scene before and after the noise suppression method in the embodiment of the present invention is shown below: the numbers af correspond to the 10.7GHz T3, 10.7GHz T4, 18.7GHz T3, 18.7GHz T4, 37.0GHz T3, and 37.0GHz T4 channels respectively; where (x1) represents the observed brightness temperature, (x2) represents the denoised brightness temperature, and (x3) represents the difference brightness temperature;

[0043] Figure 6 Error analysis of the true value-inverted wind speed in this embodiment of the invention: (a) absolute error; (b) relative error;

[0044] Figure 7 This is an RMSE error analysis of the true value-inverted wind direction in this embodiment of the invention. Detailed Implementation

[0045] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0046] This invention proposes a method for suppressing T3 / T4 brightness temperature noise in a fully polarized microwave radiometer, applicable to satellite systems carrying out sea surface radiation brightness temperature detection and inversion. This embodiment uses WindSat satellite radiometer L1c level data as an example for detailed explanation:

[0047] This method is as follows Figure 1 As shown, it includes the following steps:

[0048] S1: Acquire observed brightness temperature data and simulated brightness temperature data

[0049] From the WindSat satellite radiometer L1c-level data product, a pure ocean scene was selected, and the observation brightness temperature image and observation geometric parameters were acquired simultaneously. The observation geometric parameters include the geographic coordinates (latitude and longitude) corresponding to each observation point pixel, as well as the zenith angle and azimuth angle when the satellite observed that point.

[0050] Preprocessing of observation geometric parameters: Extract the required T3 / T4 channel observation geometric parameters from the observation geometric data, and crop the observation geometry according to latitude and longitude to select a fixed-size area;

[0051] ERA5 meteorological reanalysis data, downloaded from WindSat satellite radiometer observation times, was used to provide the necessary atmospheric and sea surface parameters for RTTOV radiative transfer model calculations. These parameters included temperature, humidity, ice-water cloud content, surface temperature, surface wind speed, and surface type at each altitude. Interpolation algorithms were then used to interpolate the ERA5 atmospheric and sea surface parameters to the geographic coordinates of the observed pixels.

[0052] The preprocessed observational geometric parameters, along with the atmospheric and sea surface parameters from ERA5, were input into the RTTOV radiative transfer model to calculate the simulated brightness temperature image.

[0053] In this implementation, L1c-level data products from 2020 were selected. To generate a sufficient number of simulated brightness temperature images that conform to the actual physical process, when using the RTTOV radiative transfer model, the observed geometric parameters (pixel geographic coordinates, instrument zenith angle, instrument azimuth angle) at a certain moment were combined with background field data (ERA5 atmospheric and sea surface parameters) at different times to generate multiple simulated brightness temperatures. For example, the observed geometric parameters at 9:00 AM on January 1, 2025, combined with ERA5 data at 9:00 AM on January 1, 2025, 9:00 AM on January 2, 2025, 9:00 AM on January 3, 2025, and 9:00 AM on January 4, 2025, respectively, can generate four simulated brightness temperatures that conform to the actual physical process.

[0054] S2: Noise suppression of observed brightness temperature data

[0055] The observed T3 / T4 brightness temperature images are input into a noise suppression method to obtain denoised brightness temperature images.

[0056] The noise characteristics in WindSat T3 / T4 observed brightness temperature images were quantitatively evaluated using two methods: observation-to-simulation (OB) difference analysis and windless scene analysis. The average standard deviation obtained from both methods was used as the noise level index in T3 / T4 polarization. In the OB difference analysis method, the observed brightness temperature and the corresponding simulated brightness temperature under each frequency band T3 / T4 channel were compared and analyzed to quantify the noise characteristics of each channel, providing a data foundation for subsequent network training. The specific quantification method is as follows: for the observed brightness temperature of each channel, simulated brightness temperature was generated using RTTOV. Based on the correlation coefficient between observation and simulation, the ten scenes with the closest correlation were selected (to ensure the accuracy of RTTOV simulation, i.e., to represent the observed brightness temperature without noise to the greatest extent). The mean and standard deviation of the difference brightness temperature were calculated. The resulting mean, standard deviation, and probability density distribution histogram are the quantified noise characteristics, and the mean and standard deviation will be used as the noise addition standard when constructing the simulation data, thus providing a data foundation for network training.

[0057] Since the true value of the observed brightness temperature after denoising is lacking, the same difference analysis method is used to perform difference analysis between the original observed brightness temperature and the observed brightness temperature after denoising, and the mean, standard deviation and probability density histogram are obtained. These are compared with the mean, standard deviation and probability density histogram obtained in the above observation-simulation difference analysis. If they are similar, the reliability of denoising can be considered.

[0058] Wind field inversion was performed using both raw brightness temperature data and noise-suppressed brightness temperature data. The inversion results were then compared with ground truth data from multiple perspectives to verify the noise reduction effect. The inversion algorithm used was the LM algorithm, which is employed for wind field inversion from multiple marine remote sensing satellites. The ground truth data used was the ERA5 meteorological reanalysis data, currently recognized as having the highest accuracy in the remote sensing field.

[0059] S3: Construct a simulated brightness temperature dataset

[0060] A simulated brightness temperature dataset for network training is constructed using a simulated brightness temperature generation method, specifically including:

[0061] For the simulated brightness temperature data obtained from S1, considering that the RTTOV radiative transfer model is relatively insufficient in fitting the brightness temperature of cloud areas, when the pixels of type clear sky (effective area) account for more than 90% of the brightness temperature image, cloud points (invalid area) will be removed, and then the scene will be retained as an element in the simulated dataset.

[0062] In this scenario, each pixel is categorized as either cloudy or clear sky, with the following criteria used to determine whether a point is cloudy:

[0063]

[0064] These refer to the 37.0 GHz vertical polarization brightness temperature calculated by the RTTOV radiative transfer model in both all-sky and clear-sky modes, respectively.

[0065] Then, the simulated brightness temperature is subjected to data preprocessing operations such as diffusion, parameterized noise addition, normalization, and overall uplift to generate a simulated brightness temperature dataset.

[0066] The diffusion method is as follows: after padding the removed cloud areas with zeros, a Gaussian filter is used to smooth the boundary between the invalid area (removed cloud areas) and the valid area (clear sky area), and the smoothing result is only applied to the invalid area, while the value of the valid area is not changed.

[0067] The parameterized noise addition method is to add Gaussian noise to the image based on the mean and standard deviation obtained above.

[0068] The overall lifting method is to add a bias to avoid negative values.

[0069] S4: Constructing the T3 / T4 brightness temperature suppression network model

[0070] A deep convolutional neural network model based on the U-Net deep learning architecture, namely the T3 / T4 brightness temperature suppression network model, was established. The input data of the model is a noisy brightness temperature image, and the output data is the denoised brightness temperature image. The original simulated brightness temperature data filled_data for each channel represents the ideal observed brightness temperature and is used as the data label, while the noisy_data represents the noisy observed brightness temperature. The mean squared error (RMSE) function was selected as the loss function to reflect the performance of noise suppression in the entire scene.

[0071] The network model was trained using a simulated brightness temperature dataset. During this process, key hyperparameters such as the number of training iterations and the learning rate were continuously adjusted based on the performance of the network model. The optimal network model was obtained when the loss function value stopped decreasing and began to fluctuate.

[0072] S5: Input the observed brightness temperatures T3 / T4 into the optimal network model to obtain the noise-suppressed brightness temperatures.

[0073] Figure 2 , Figure 3 , Figure 4This paper presents a comparison of STD (Standard Deviation) values ​​before and after noise suppression processing for scenes at various latitudes. The horizontal axis represents the test sample sequence number, and the vertical axis represents the STD value (unit: K). The std_obs and std_deNoise lines represent the standard deviations of the observed brightness temperature and the denoised brightness temperature, respectively, calculated using the MAD local statistical method. Observing the line graphs, although the std_obs values ​​fluctuate differently across different frequency bands and polarization modes, they remain at a relatively high level overall. It can be seen that the std_deNoise values ​​in the images are all lower than std_obs, especially in the special 37.0GHz T4 channel, where the STD value of the processed data drops to near zero. This comparison indicates that the noise suppression algorithm effectively reduces the dispersion of the data and improves data stability. Notably, regardless of the fluctuation trend of the original std_obs, std_deNoise consistently remains at a lower level, demonstrating the algorithm's noise suppression capability.

[0074] Table 1 Comparison of brightness temperature statistics before and after noise reduction for each channel (low latitude)

[0075]

[0076] Table 2 Comparison of brightness temperature statistics before and after noise reduction for each channel (Zhongwei)

[0077]

[0078] Table 3 Comparison of brightness temperature statistics before and after noise reduction for each channel (Gao Wei)

[0079]

[0080]

[0081] Furthermore, Tables 1, 2, and 3 summarize the average STD statistics for all scenes in the datasets of each latitude. The data shows that the mean of the brightness temperature difference (diffTb) (observeTb - deNoiseTb) for each channel is close to zero, indicating that the brightness temperature after denoising has not shifted at the overall level, and the algorithm has preserved the original characteristics of the data well. Meanwhile, the reduction rate of the standard deviation of brightness temperature after noise suppression generally ranges from 35% to 85%, further quantifying the denoising efficiency of the algorithm. The above analysis results collectively demonstrate that the denoising process performs well in reducing data dispersion, verifying the effectiveness of the algorithm in suppressing noise in brightness temperature data across various latitudes.

[0082] After completing the numerical analysis, since the results across different dimensions are similar, this section presents the before-and-after effects of the algorithm processing on a specific scene within the dataset. Figure 5Taking [example image] as an example, in images with different frequency bands and polarizations, noise of varying sizes can be observed in all observed brightness temperatures. After brightness temperature noise suppression processing, the image trend becomes smoother, and the similarity between obsTb and deNoiseTb is extremely high under visual observation, demonstrating excellent fitting results. The diffTb image, representing the difference between the two absolute values, also exhibits typical visual distribution characteristics of random noise. This demonstrates the effectiveness of the noise suppression algorithm in visualized images.

[0083] The wind field inversion experiment selected observation scenes covering all latitudes globally. Through data integration, a comprehensive dataset containing 35,960 data points was constructed for inversion. The LM algorithm, used for wind field inversion from multiple ocean remote sensing satellites, was employed. The experiment used observed brightness temperature data and noise-suppressed brightness temperature data from the WindSat L1c data product for wind field inversion, and then compared the inversion results with the ERA5 reference values ​​from multiple perspectives.

[0084] Figure 6 The results of absolute and relative error analysis for wind speed inversion are presented. It can be seen that in the low wind speed range (2–4 m / s), Ret_DeNoise outperforms Ret_Obs in terms of absolute error. Similarly, in the relative error plot, the highest relative error of the inversion from the original brightness temperature data is close to 1.4, while the error after noise suppression is effectively controlled within 1.1, indicating that the proposed algorithm has stronger denoising capabilities under low wind speed conditions. As the actual wind speed increases, the errors of each method tend to converge, because the inversion algorithm itself has strong stability in high wind speed scenarios. However, Ret_deNoise still maintains a slightly better performance, further verifying the universality of the algorithm under different wind speed conditions.

[0085] Regarding standard deviation, the trends of the dashed curves for Ret_DeNoise and Ret_Obs are basically consistent in both the absolute error plot and the relative error plot, indicating no significant difference before and after noise suppression. Based on the physical meaning of the standard deviation index, this demonstrates that noise suppression improves the average accuracy while maintaining the overall consistency between the inversion results and the true value.

[0086] To further evaluate the actual effect of the brightness temperature noise suppression algorithm in wind direction inversion, this invention analyzes the root mean square error variation trend between the inverted wind direction and the true ERA5 value under the condition of real wind speed binning. Figure 7 The curve trend of RMSE with wind speed is shown, and Table 4 lists the specific numerical results for each wind speed range.

[0087] Table 4. RMSE Error Table of True Value - Inverted Wind Direction

[0088]

[0089] The trend in the figure shows that the wind direction inversion error gradually decreases with increasing actual wind speed, consistent with the characteristic that microwave remote sensing wind direction inversion is more sensitive to low wind speeds. In the low wind speed range (2–4 m / s), the wind direction inversion error is greatest due to the weak wind speed signal and low radiation brightness temperature signal-to-noise ratio. At this point, the RMSE of the original data is as high as 31.81°, while the noise-suppressed data reduces it to 29.14°, a decrease of approximately 2.67°, demonstrating the robustness of the noise suppression algorithm under low signal-to-noise ratio conditions.

[0090] In the low to medium wind speed range (4–6 m / s), the RMSE before and after denoising were 29.29° and 27.49°, respectively, a decrease of 1.8°. In the medium to high wind speed range (6–10 m / s), the improvement effect of the noise suppression algorithm was more significant: the RMSE decreased from 20.87° and 17.41° to 18.03° and 14.07°, respectively, an improvement of 2–3°. Observing the image, the binning error in this region decreased by more than 5°. This indicates that under conditions of relatively stable wind field and high brightness-temperature signal-to-noise ratio, noise suppression processing can more fully exert its improvement capabilities.

[0091] In the high wind speed range of 10–12 m / s, the RMSE decreased from 16.63° to 14.47°, maintaining a robust performance. Especially in the range where the wind speed exceeds 12 m / s, the error decreased from 4.31° to 2.29°, indicating that the brightness temperature noise suppression still has high reliability in extreme scenarios of wind field inversion.

[0092] comprehensive Figure 7 As can be seen from the results in Table 4, the proposed brightness temperature noise suppression algorithm exhibits consistent and comprehensive advantages in wind direction inversion tasks. The RMSE of the noise suppression data is lower than that of the original data in each wind speed range. Although the improvement varies with wind speed, the overall trend is clear: the noise suppression algorithm has a certain anti-interference ability under low wind speed conditions, further enhances accuracy under medium and high wind speed conditions, and remains stable in the high wind speed range.

[0093] In summary, it is clear that the noise suppression algorithm obtained by this invention is technically effective in WindSat observations T3 / T4. It not only verifies the universality of the algorithm in multi-regional scenarios, but also demonstrates its noise suppression capability and data accuracy improvement effect in wind field inversion applications, providing a reference for remote sensing data processing of WindSat radiometers.

[0094] The method proposed in this invention can effectively suppress T3 / T4 brightness temperature noise, guiding the development of a fully polarized microwave radiometer on a marine microwave remote sensing satellite.

[0095] The embodiments described above are merely preferred embodiments of the present invention. Ordinary variations and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for suppressing T3 / T4 brightness temperature noise in a fully polarized microwave radiometer, characterized in that, include: S1: Select a pure ocean scene from the data products of the spaceborne fully polarized microwave radiometer, and extract multiple original observation brightness temperature images from the T3 / T4 channels and the observation geometric parameters corresponding to each observation brightness temperature image; S2: Based on each original observed brightness temperature image, generate simulated brightness temperature images respectively; S3: Perform noise suppression processing on each original observed brightness temperature image to obtain a denoised observed brightness temperature image; calculate the mean and standard deviation of the brightness temperature difference based on each original observed brightness temperature image and the corresponding denoised observed brightness temperature image, as the quantized noise characteristics. S4: For each simulated brightness temperature image, perform diffusion, add noise according to the quantized noise characteristics, and normalize in sequence; combine each processed simulated brightness temperature image and the corresponding unprocessed simulated brightness temperature image to form a training sample pair, thus forming a simulated brightness temperature dataset. S5: Establish a brightness temperature suppression network model to generate denoised observed brightness temperature images; train the brightness temperature suppression network model using a simulated brightness temperature dataset; S6: Process the data from the spaceborne fully polarized microwave radiometer according to S1 to obtain the observed brightness temperature image to be analyzed; input the observed brightness temperature image to be analyzed into the trained brightness temperature suppression network model to obtain the denoised observed brightness temperature image.

2. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 1, characterized in that, Extract the observation geometric parameters of the T3 / T4 channels. The observation geometric parameters include the geographic coordinates of each observation point pixel, the zenith angle and azimuth angle when the satellite observes each observation point.

3. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 2, characterized in that, Based on the original observed brightness temperature image, a simulated brightness temperature image is generated. Specific methods include: Download ERA5 meteorological reanalysis data with the same observation time as the spaceborne fully polarized microwave radiometer; Atmospheric and sea surface parameters from ERA5 meteorological reanalysis data are used as background field data and interpolated onto the original observed brightness temperature image to achieve spatiotemporal data matching. Based on the observed geometric parameters and background field data of the original observed brightness temperature image, the simulated brightness temperature image was calculated using the RTTOV radiative transfer model.

4. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 3, characterized in that, Based on the observation time of the spaceborne fully polarized microwave radiometer, ERA5 meteorological reanalysis data from the same period of observation are downloaded to form multiple background field data. Based on the observation geometric parameters of the original brightness temperature image and each background field data, a simulated brightness temperature image is generated, thus obtaining multiple simulated brightness temperature images that conform to the actual physical process.

5. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 2, characterized in that, The method for extracting the observation geometric parameters of the T3 / T4 channels to obtain the original observation brightness temperature image is as follows: For spaceborne fully polarized microwave radiometer data of pure ocean scenes, the observation geometric parameters of the T3 / T4 channels are extracted. According to the geographic coordinates of each observation point pixel in the observation geometric parameters, the observation image of the spaceborne fully polarized microwave radiometer of pure ocean scenes is cropped to obtain the required observation brightness temperature image.

6. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 3, characterized in that, Each simulated brightness temperature image undergoes diffusion processing, specifically including: Each pixel in the simulated brightness temperature image is marked as either cloudy or clear. If the proportion of pixels marked as clear exceeds a set threshold, the pixels marked as cloudy are removed, and the invalid area formed by these pixels is removed. The remaining area is the valid area. If the proportion of pixels marked as clear is less than a set threshold, the simulated brightness temperature image is discarded. After filling the pixel illumination temperature value of the invalid region with 0, a Gaussian filter is used to smooth the boundary between the invalid region and the valid region, thus completing the diffusion process.

7. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 6, characterized in that, Each pixel in the simulated brightness temperature image is marked as either cloudy or clear sky. The marking principle is: if the following formula is true, it is cloudy; otherwise, it is clear sky. in, These refer to the 37.0 GHz vertical polarization brightness temperature of the pixel calculated by the RTTOV radiative transfer model in both all-sky and clear-sky modes.

8. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 1, characterized in that, The original observed brightness temperature image undergoes noise suppression processing. Specific methods include: The observation-simulation difference analysis method was used to compare and analyze the brightness temperature values ​​of the original observed brightness temperature images and the corresponding simulated brightness temperature images under the T3 / T4 channels of each frequency band, and to quantify the noise characteristics of each channel: For each channel, select several pairs of images whose original observed brightness temperature image and corresponding simulated brightness temperature image values ​​are closest, and calculate the mean and standard deviation of the brightness temperature difference as the noise characteristics of the quantization.

9. The method for suppressing brightness temperature noise of a fully polarized microwave radiometer T3 / T4 according to claim 8, characterized in that, After diffusion processing of the simulated brightness temperature image, Gaussian noise is generated based on the obtained mean and standard deviation, and then superimposed on the simulated brightness temperature image.

10. A method for suppressing brightness temperature noise in a fully polarized microwave radiometer T3 / T4 according to claim 8, characterized in that, A brightness temperature suppression network model based on the U-Net deep learning architecture is established, with the mean squared error (RMSE) loss function. The brightness temperature suppression network model is trained using a simulated brightness temperature dataset: the processed simulated brightness temperature image is input into the model, and the loss function value is calculated based on the denoised brightness temperature image and the unprocessed simulated brightness temperature image output by the model. The optimal brightness temperature suppression network model is obtained when the loss function value no longer decreases.

Images