Physical-statistical network wind speed retrieval method based on spaceborne microwave active and passive combination
By constructing a physical statistical network method based on joint active and passive microwave data from spaceborne microwave, and combining microwave radiometer and scatterometer data, and using residual neural networks for training, the problem of low wind speed inversion accuracy under high wind speed and heavy rainfall conditions in microwave remote sensing was solved, and high-precision inversion was achieved across the entire wind speed range.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SANYA INST OF OCEANOGRAPHY OCEAN UNIV OF CHINA
- Filing Date
- 2025-12-19
- Publication Date
- 2026-06-26
AI Technical Summary
In existing technologies, microwave remote sensing has low wind speed inversion accuracy under high wind speed and heavy rainfall conditions, and deep learning algorithms lack physical mechanisms, making it difficult to maintain high accuracy and interpretability across the entire wind speed range.
A physical statistical network method based on spaceborne microwave active and passive joint was adopted. By constructing a deep learning inversion algorithm model with physical statistical constraints, and using microwave radiometer and scatterometer data, a residual neural network and physical layer method were used to extract multi-frequency data of the whole sea area and the physical statistical constraints of the deep learning inversion algorithm model. Combined with microwave radiometer and scatterometer data, a sample dataset was constructed and trained using a residual neural network to generate a wind speed inversion model.
It solves the technical problem of wind speed inversion under complex weather conditions, improves the inversion accuracy across the entire wind speed range, and provides high-quality and reliable wind speed products.
Smart Images
Figure CN121365605B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of marine remote sensing technology and relates to a method for full wind speed inversion based on a physical statistical network using a spaceborne microwave active-passive joint approach. Background Technology
[0002] Sea surface wind (SSW) is a crucial physical parameter describing air-sea interaction and a key factor influencing ocean circulation, climate change, and weather system evolution. Due to the scarcity and limited coverage of marine observation points, satellite microwave remote sensing technology has become the primary means of acquiring global sea surface wind speed. Currently, microwave remote sensing payloads commonly used for wind speed retrieval mainly include microwave scatterometers (SCAT) and microwave radiometers (SMR).
[0003] A microwave scatterometer is an active microwave sensor that transmits microwave signals to the sea surface and receives the backscattered signals. By combining the incident angle, azimuth angle, and polarization information, it uses a geophysical model function (GMF) to invert sea surface wind vectors. Scatterometers typically operate in the Ku-band or C-band, and their inversion results have high accuracy under low to medium wind speed conditions. However, when sea surface wind speeds exceed 25 m / s, the backscattered signal saturates due to the effects of foam and wave breakup, significantly reducing the accuracy of scatterometers under high wind speed conditions. Unlike scatterometers, microwave radiometers are passive observation instruments that invert wind speed by measuring the brightness temperature (TB) of the sea surface in multiple microwave frequency bands. Changes in sea surface wind speed cause changes in sea surface roughness, which in turn affects microwave radiation characteristics. Higher wind speeds result in a rougher sea surface, increased microwave emissivity, and consequently, changes in brightness temperature. Microwave radiometers typically operate in multiple frequency bands: the low-frequency channel (6–10 GHz) is more sensitive to sea surface temperature and humidity, while the high-frequency channel (30–40 GHz) is more responsive to sea surface roughness and rainfall. Therefore, by combining multi-frequency, multi-polarization brightness temperature information with a radiative transfer model, sea surface wind speed can be estimated relatively accurately.
[0004] The Radiative Transfer Model (RTM) is used to establish the physical relationship between brightness temperature and wind speed. Its calculations require comprehensive consideration of various factors, including atmospheric absorption and scattering, sea surface reflection, and changes in emissivity. Under extreme weather conditions such as tropical cyclones, the simultaneous presence of heavy rainfall and high wind speeds causes multiple interferences to the microwave signal, significantly reducing the accuracy of wind speed retrieval. Decreased atmospheric transmittance due to rainfall and increased emissivity due to wave breaking can both lead to overestimation of wind speed. Therefore, distinguishing the effects of wind speed and rainfall on brightness temperature signals under complex meteorological conditions is a key challenge in microwave radiometer wind speed retrieval.
[0005] In recent years, wind speed inversion methods based on neural networks and deep learning have made significant progress. By learning the nonlinear mapping relationship between brightness temperature and wind speed, the problem of insufficient parameterization in traditional physical models can be alleviated to some extent. However, purely data-driven models lack physical constraints and are prone to producing unstable or non-physical inversion results in extreme wind speed ranges. To improve the accuracy of wind speed inversion, researchers have proposed a collaborative inversion approach using active and passive microwave data. Scattermeters provide highly sensitive low-wind-speed structure information, while radiometers have stronger penetration capabilities under high-wind-speed and rainy conditions. By combining the advantages of both, the dynamic range of wind speed inversion can be significantly broadened, enabling continuous wind speed inversion from calm sea states to super typhoons. For example, existing studies have used the scattermeter and scanning microwave radiometer (SMR) carried on the HY-2A satellite to jointly invert typhoon wind fields, achieving good results. Currently, most active-passive collaborative algorithms still rely on empirical models or simple neural network structures, lacking explicit constraints on microwave physical processes, making it difficult to balance model interpretability and inversion accuracy. Under extreme conditions of heavy rainfall and high wind speeds, the model still suffers from problems such as overestimation of wind speed and insufficient spatial continuity. To address these issues, it is imperative to design a comprehensive algorithm that retains the radiative transfer model while incorporating more advanced inversion algorithms. Summary of the Invention
[0006] This invention overcomes the above-mentioned defects and provides a full wind speed inversion method based on a physical statistical network of spaceborne microwave active and passive combined methods. It solves the technical problems in the prior art, such as the lack of physical mechanism in deep learning algorithms and the low accuracy of high wind speed inversion algorithms in traditional microwave remote sensing (combined microwave radiometer and scatterometer).
[0007] To achieve the above objectives, the present invention provides a full wind speed inversion method based on a spaceborne microwave active-passive joint physical statistical network, comprising the following steps:
[0008] S1. Acquire microwave radiometer data, microwave scatterometer data, and wind speed data from different products, and perform data quality control.
[0009] S2. Extract multi-band brightness temperature data and backscattering coefficients for the entire sea area, and construct a sample dataset using the wind speed data described in S1.
[0010] S3. Load the deep learning inversion algorithm model with physical statistical constraints, and train the model using the sample dataset to obtain the deep learning wind speed inversion algorithm model with physical statistical constraints.
[0011] S4. Extract the required full-ocean-area microwave radiometer multi-band brightness temperature data and microwave scatterometer multi-band backscattering coefficients. Input the brightness temperature data and backscattering coefficients to be inverted into the deep learning wind speed inversion algorithm model with physical statistical constraints, and output the wind speed of the full-ocean-area at the current time.
[0012] Furthermore, in step S2, the input samples of the sample dataset are brightness temperature data scanned by the spaceborne microwave radiometer and backscattering coefficient scanned by the spaceborne microwave scatterometer, and the output samples are wind speed data.
[0013] Furthermore, in step S3, the deep learning inversion algorithm model of physical statistical constraints includes: an input layer, a network layer, a physical layer, and an output layer; the network layer is composed of a residual neural network; the physical layer is composed of a microwave statistical matrix P matrix, which is formed by adding the Log matrix formula of the microwave scatterometer to the physical layer based on only the microwave radiometer network D matrix.
[0014] Compared to the D matrix, the P matrix can fully extract the radiative transfer brightness temperature information from the microwave radiometer, automatically applying it to wind speed inversion under extreme sea conditions such as high rainfall and high water vapor. Simultaneously, it extracts the backscattering coefficient information from the microwave scatterometer, automatically improving the inversion accuracy in the low-to-medium wind speed range. The P matrix provides deep learning with the physical information needed for simultaneous wind speed inversion from both the microwave radiometer and scatterometer, making the physical mechanism transparent when deep learning retrieves wind speed from microwave sensors. This method can ensure high-accuracy inversion at high wind speeds while simultaneously improving high-accuracy inversion at low-to-medium wind speeds throughout the entire inversion process.
[0015] Furthermore, in step S3, the physical statistical constraint deep learning inversion algorithm model is trained using the sample dataset, including:
[0016] S31. Divide the sample dataset into training set, validation set and test set according to the preset ratio;
[0017] S32. Using the training set, train the physical statistical constraint deep learning inversion algorithm model, and use the validation set to monitor the training results in real time to obtain an intermediate physical statistical constraint deep learning inversion algorithm model.
[0018] S33. Using the test set, the intermediate physical statistical constraint deep learning inversion algorithm model is tested, and the intermediate physical statistical constraint deep learning inversion algorithm model that has completed the test is determined as a deep learning wind speed inversion algorithm model with physical statistical constraints.
[0019] Furthermore, the physical statistical constraint deep learning inversion algorithm model is trained using the training set, including:
[0020] S321. Network layer calculation, including fully connected calculation and residual calculation; the last layer of the network has l=5 and k=0~14, and the results TBPs and SiPs are obtained. The results TBPs and SiPs are the coefficients of the P matrix required by the physical layer.
[0021] S322. Physical layer calculation: The P-matrix coefficients TBPs and SiPs, along with the multi-band brightness temperature data and multi-band backscattering coefficients of the input layer, are used to calculate the wind speed through the P-matrix equation. Here, the P-matrix equation serves as a physical statistical constraint on the entire model. The calculation formula is as follows:
[0022] ,
[0023] ,
[0024] ,
[0025] ,
[0026] ,
[0027] ,
[0028] Meanwhile, the physical statistical function can also be expressed by the following formula:
[0029] ,
[0030] in,
[0031] ;
[0032] S323, Loss Calculation: Calculate the loss based on the wind speed calculated from the P matrix and the wind speed in the training set.
[0033] S324, Reverse Callback Step: Based on the inversion loss, perform reverse callback on the parameters of the physical statistical constraint deep learning inversion algorithm model, and repeat the execution until the inversion loss no longer decreases or the number of repetitions reaches 5000 and the loop stops. The model obtained after stopping is determined as the intermediate physical statistical constraint deep learning inversion algorithm model.
[0034] Furthermore, the input data of the test set is fed into the intermediate physical statistical constraint deep learning inversion algorithm model to obtain wind speed data, and the error is calculated with the output data of the test set.
[0035] If the error value is within the set range, a deep learning wind speed inversion algorithm model with physical statistical constraints is obtained; if the error value exceeds the set range, the network parameters are readjusted and the model is trained.
[0036] Furthermore, the error value is set within a range of 0-1 m / s.
[0037] Furthermore, in step S3, the input layer needs to skip to the physical layer while connecting to the network layer.
[0038] Skip connections from the input layer to the physical layer can directly pass the original input information to the physical layer, which is useful for P-matrix requirements. Skip connections in the network layers avoid the problem of information from higher layers gradually disappearing as the number of network layers increases, fully transferring information from higher layers to the required positions, and mitigating gradient decay during network training.
[0039] Furthermore, in step S4, after extracting the microwave radiometer brightness temperature data and scatterometer backscattering coefficients for the entire sea area, the data to be inverted is input into the full wind speed inversion algorithm model with physical statistical constraints to obtain the wind speed and matrix coefficients of the P matrix over the entire range. The matrix coefficients of the P matrix are used for physical feature analysis during the full wind speed inversion process.
[0040] The matrix coefficients of the P matrix constrain the impact of extreme weather factors such as rainfall on wind speed inversion. In the network inversion process, the scatterometer automation plays a significant role in low wind speeds, while the radiometer automation plays a significant role in high wind speeds.
[0041] Compared with the prior art, the advantages of the present invention are as follows:
[0042] (1) This invention provides a full wind speed inversion method based on a residual network with physical statistical constraints. By using the brightness temperature of the microwave radiometer and the backscattering coefficient of the scatterer, and by combining the microwave statistical matrix P matrix composed of the Log formula of the microwave radiometer and the Log formula of the microwave scatterer with a neural network, a network model with physical information can be constructed. This can effectively reduce the influence of changes in sea surface brightness temperature received by the spaceborne microwave radiometer and the influence of the attenuation of the backscattering coefficient of the microwave scatterer caused by changes in sea surface emissivity and atmospheric transmittance caused by rainfall and wave breaking, thereby obtaining more accurate wind speed information for the entire wind speed range.
[0043] (2) The dynamic P matrix coefficients in this invention fully explain the physical information of the ocean atmosphere (including microwave radiometer and microwave scatterometer) used by deep learning in the process of microwave remote sensing sea surface wind speed inversion.
[0044] The P-matrix formula also fully constrains the training of deep learning, effectively solving the problem of lack of physical support in the process of deep learning in ocean parameter inversion, and semi-transparently making the learning and training process of deep learning, providing high-quality and reliable wind speed products applicable to the entire wind speed range. Attached Figure Description
[0045] Figure 1 This is a flowchart of the full wind speed inversion method based on a physical statistical network using a spaceborne microwave active-passive joint approach, as described in this invention.
[0046] Figure 2 This is a schematic diagram of the physical statistical constraint deep learning wind speed inversion algorithm of the present invention. Detailed Implementation
[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0048] like Figure 1 As shown, this invention proposes a full wind speed inversion method based on a spaceborne microwave active-passive joint physical statistical network, specifically including:
[0049] S1. Acquire microwave radiometer data, microwave scatterometer data, and wind speed data from different products, and perform data quality control.
[0050] Specifically, microwave radiometer and microwave scatterometer data are sourced from my country's domestically built Haiyang-2B (HY-2B) satellite. Wind speed data includes reanalysis wind speed data from the ECMWF and wind speed data from the SMAP satellite. The accuracy of the wind speed data across each wind speed range is crucial to ensure sufficient precision for subsequent model training. The low-to-medium wind speed range of the ECMWF reanalysis data is widely recognized as high-precision, while the high wind speed range of the SMAP satellite is also considered high-precision. Data quality control utilizes corresponding data quality labels for screening, such as land-sea labels and quality labels.
[0051] S2. Extract multi-band brightness temperature data and backscattering coefficients for the entire sea area, and construct a sample dataset using the wind speed data described in S1.
[0052] Specifically, the input samples for the sample dataset are the brightness temperature (TB, unit: K) of the TC level data of the HY-2B radiometer and the L2A level backscattering cross section (backscattering coefficient) of the HYSCAT scatterometer. (Unit: dB), the output sample is constructed by ECMWF and SMAP (ECMWF<20m / s, SMAP>=20m / s), the three data are spatiotemporally matched, spatial range ±25 km, spatial range ±0.5 hours, other required data for the dataset are shown in Table 1.
[0053] Table 1 Dataset
[0054]
[0055] S3. Load the deep learning inversion algorithm model with physical statistical constraints, and train the model using the sample dataset to obtain the deep learning wind speed inversion algorithm model with physical statistical constraints.
[0056] Specifically, the deep learning inversion algorithm model with physical statistical constraints includes an input layer, a network layer, a physical layer, and an output layer. The network layer is constructed from a residual neural network; the physical layer consists of a microwave statistical matrix (P matrix) formed by the Log formulas of the microwave radiometer and the microwave scattering coefficient. Unlike a typical physical layer network, this model adds the Log formula of the microwave scattering coefficient to the physical layer, which only contains the D matrix of the microwave radiometer network. The resulting P matrix is constructed from the Log matrices of brightness temperature and backscattering coefficients. Its construction form is as follows: Figure 2 As shown.
[0057] In this embodiment of the invention, the model is trained using the sample dataset from step S2, including the following steps:
[0058] (1) Use the data from 2019 and 2020 as training and validation sets (70% for training and 30% for validation), and the data from 2021 and 2022 as independent test datasets;
[0059] (2) The training set is used to train the physical statistical constraint deep learning inversion algorithm model, and the training results are monitored in real time using the validation set to obtain the intermediate physical statistical constraint deep learning inversion algorithm model.
[0060] (3) Using the test set, the intermediate physical statistical constraint deep learning inversion algorithm model is tested, and the intermediate physical statistical constraint deep learning inversion algorithm model that has completed the test is determined as the deep learning wind speed inversion algorithm model with physical statistical constraints.
[0061] Specifically, the training set contains multiple sample data, and step (2) also includes the following network construction and training steps:
[0062] The sample dataset contains three types of input-output data: HY-2B TBs, HY-2B And label wind speeds (Labels SSWs) jointly labeled by ECMWF and SAMP; input data includes 9 TB bands (TB6.9H, TB6.9V, TB10.7H, TB10.7V, TB18.7H, TB18.7V, TB23.8V, TB37H, and TB37V) and 4 frequency band ( , , and The N×9 and N×4 matrices are formed respectively, and their representations are as follows:
[0063] ,
[0064] ,
[0065] Where N represents the total number of observed samples in the dataset, and the output data is the label wind speed Y, which is represented as an N×1 matrix as follows:
[0066] ,
[0067] These input parameters are fed into the designed network structure, where calculations are performed, including those for fully connected layers, residual connections, and activation functions (AF). Before obtaining the physical statistics, the final activation function is applied, as follows:
[0068] ,
[0069] Here, Linear(o) represents the linear activation function (Linear). To improve the high-wind-speed inversion capability through physical statistics, the hyperbolic tangent function (Tanh) was modified to form the Tanha activation function, whose expression is as follows:
[0070] ,
[0071] These physical statistical coefficients will be related to the initial brightness temperature band set TBs and the backscattering coefficient set 𝜎 0 Together with s, we calculate the wind speed values needed for the inversion.
[0072] The corresponding formula is as follows:
[0073] ,
[0074] ,
[0075] ,
[0076] ,
[0077] ,
[0078] ,
[0079] in, The wind speeds output by the model are represented by Log(TBs) and Log( ) respectively represent TBs and The input is a constructed logarithmic matrix, where TBPs, SiPs, and P0s represent the coefficient matrices in the form of physical functions in RS-PSNet. Specifically, TBPs and SiPs represent the coefficients of TBPs and SiPs, respectively. The physical coefficients are given by P0s, where P0s is the physical coefficient of the automatic control deviation. This method achieves nonlinear full-range wind speed inversion through statistical formulas.
[0080] Meanwhile, the statistical functions of physics can also be expressed using the following formula:
[0081] ,
[0082] in,
[0083] ,
[0084] Finally, the model uses the initial output conditions to determine the optimal weights to minimize the defined loss function. The model continuously optimizes the parameters to reduce the loss function, resulting in the final trained model. Therefore, this model is not only applicable to inversion across the entire wind speed range but also incorporates constraints imposed by physical statistical functions.
[0085] The loss calculation steps involve using the wind speed calculated from the physical statistics matrix and the labeled wind speed (i.e., the output wind speed of the dataset) to calculate the loss. The loss function is the root mean square error (MSE).
[0086] ,
[0087] in, Let represent the labeled wind speed of the i-th sample. This represents the model output wind speed for the i-th sample.
[0088] The above steps are repeated cyclically. When the loss remains constant or the number of repetitions reaches 5000, the loop stops, and the obtained model is regarded as an intermediate physical statistical constraint deep learning inversion algorithm model. Then proceed to step (3).
[0089] S4. After obtaining the brightness temperature data of the tropical cyclone to be inverted, the brightness temperature data of the current tropical cyclone is input into the tested deep learning wind speed inversion algorithm model with physical statistical constraints to perform real-time inversion of the wind speed of the marine tropical cyclone and obtain the wind speed information of the current tropical cyclone. Physical analysis is then performed on the inversion process using the P coefficient.
[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, alterations, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A full wind speed inversion method based on a spaceborne microwave active-passive joint physical statistical network, characterized in that, Includes the following steps: S1. Acquire microwave radiometer data, microwave scatterometer data, and wind speed data from different products, and perform data quality control. S2. Extract multi-band brightness temperature data and backscattering coefficients for the entire sea area, and construct a sample dataset using the wind speed data described in S1. S3. Load a deep learning inversion algorithm model with physical statistical constraints, and train the model using a sample dataset to obtain a deep learning wind speed inversion algorithm model with physical statistical constraints. The deep learning inversion algorithm model with physical statistical constraints includes: an input layer, a network layer, a physical layer, and an output layer. The physical layer is composed of a microwave statistical matrix P. The P matrix is composed of the brightness temperature data of the microwave radiometer and the Log matrix of the backscattering coefficient of the microwave scatterer. In step S3, the physical statistical constraint deep learning inversion algorithm model is trained using the sample dataset, including: S31. Divide the sample dataset into training set, validation set and test set according to the preset ratio; S32. Using the training set, train the physical statistical constraint deep learning inversion algorithm model, and use the validation set to monitor the training results in real time to obtain an intermediate physical statistical constraint deep learning inversion algorithm model. S33. Using the test set, the intermediate physical statistical constraint deep learning inversion algorithm model is tested, and the intermediate physical statistical constraint deep learning inversion algorithm model that has completed the test is determined as a deep learning wind speed inversion algorithm model with physical statistical constraints. In step S32, the physical statistical constraint deep learning inversion algorithm model is trained using the training set, including: S321. Network layer calculation, including fully connected calculation and residual calculation; the last layer of the network has l=5 and k=0~14, and the results TBPs and SiPs are obtained. The results TBPs and SiPs are the coefficients of the P matrix required by the physical layer. S322. Physical layer calculation: The P-matrix coefficients TBPs and SiPs, along with the multi-band brightness temperature data and multi-band backscattering coefficients of the input layer, are used to calculate the wind speed through the P-matrix equation. Here, the P-matrix equation serves as a physical statistical constraint on the entire model. The calculation formula is as follows: ； Where TB represents brightness temperature. The backscattering coefficient is represented by TBs, which represents the set of brightness temperature frequency bands. Let Log(TBs) represent the set of backscattering coefficients. Represented by TBs and The input is a constructed logarithmic matrix, where TBPs, SiPs, and P0s represent the coefficient matrices in physical function form, and TBPs and SiPs represent the values of TBPs and P0s, respectively. The physical coefficient matrix, P0s is the physical coefficient matrix of the automatic adjustment deviation; ， Where N represents the total number of samples, This represents the brightness temperature coefficient of the i-th sample calculated by the model at a frequency of 6.9 GHz and H polarization. This represents the brightness temperature coefficient of the i-th sample at a frequency of 6.9 GHz and polarization V. This represents the brightness temperature coefficient of the i-th sample at a frequency of 37 GHz and polarization V. ， in, This represents the brightness temperature input data of the i-th sample at a frequency of 6.9 GHz and H polarization. This represents the brightness temperature input data of the i-th sample at a frequency of 6.9 GHz and a polarization of V. This represents the brightness temperature input data of the i-th sample at a frequency of 37 GHz and polarization V. The coefficients of the logarithmic matrix corresponding to each brightness temperature input data are represented. ， in, This represents the physical coefficient corresponding to the backscattering coefficient of the i-th sample under forward-looking HH polarization with an incident angle of 41°, calculated by the model. This represents the physical coefficient corresponding to the backscattering coefficient of the i-th sample under an incident angle of 41° and back-view HH polarization; This represents the physical coefficient corresponding to the backscattering coefficient of the i-th sample under forward-looking VV polarization with an incident angle of 48°. This represents the physical coefficient corresponding to the backscattering coefficient of the i-th sample under an incident angle of 48° and backsight VV polarization; ， in, This represents the input data for the backscattering coefficient of the i-th sample under forward-looking HH polarization with an incident angle of 41°. This represents the input data for the backscattering coefficient of the i-th sample under an incident angle of 41° and backsight HH polarization; This represents the input data for the backscattering coefficient of the i-th sample under forward-looking VV polarization with an incident angle of 48°. This represents the input data for the backscattering coefficient of the i-th sample under an incident angle of 48° and backsight VV polarization; This represents the coefficients in the logarithmic matrix corresponding to each backscattering coefficient input data; ， in, This represents the bias coefficient of the i-th sample; Meanwhile, the physical statistical function also uses the following formula: ， in, ， S323, Loss Calculation: Calculate the loss based on the wind speed calculated from the P matrix and the wind speed in the training set. S324, Reverse Callback Step: Based on the inversion loss, perform reverse callback on the parameters of the physical statistical constraint deep learning inversion algorithm model, and repeat until the inversion loss no longer decreases or the number of repetitions reaches 5000 and the loop stops. The model obtained after stopping is determined as the intermediate physical statistical constraint deep learning inversion algorithm model. S4. Extract the required full-ocean-area microwave radiometer multi-band brightness temperature data and microwave scatterometer multi-band backscattering coefficients. Input the brightness temperature data and backscattering coefficients to be inverted into the deep learning wind speed inversion algorithm model with physical statistical constraints, and output the wind speed of the full-ocean-area at the current time.
2. The full wind speed inversion method based on a physical statistical network using a spaceborne microwave active-passive joint approach, as described in claim 1, is characterized in that... In step S2, the input samples of the sample dataset are brightness temperature data scanned by the spaceborne microwave radiometer and backscattering coefficient scanned by the spaceborne microwave scatterometer, and the output samples are wind speed data.
3. The full wind speed inversion method based on a physical statistical network using a spaceborne microwave active-passive joint approach, as described in claim 1, is characterized in that... The input data of the test set is fed into the intermediate physical statistical constraint deep learning inversion algorithm model to obtain wind speed data, and the error is calculated with the output data of the test set. If the error value is within the set range, a deep learning wind speed inversion algorithm model with physical statistical constraints is obtained; if the error value exceeds the set range, the network parameters are readjusted and the model is trained.
4. The full wind speed inversion method based on a physical statistical network using a spaceborne microwave active-passive joint approach, as described in claim 3, is characterized in that... The error value is set within the range of 0-1 m / s.
5. The full wind speed inversion method based on a physical statistical network using a spaceborne microwave active-passive joint approach, as described in claim 1, is characterized in that... In step S3, the input layer needs to skip to the physical layer while connecting to the network layer.
6. The full wind speed inversion method based on a physical statistical network using a spaceborne microwave active-passive joint approach, as described in claim 1, is characterized in that... In step S4, after extracting the microwave radiometer brightness temperature data and scatterometer backscattering coefficients for the entire sea area, the data to be inverted is input into the full wind speed inversion algorithm model with physical statistical constraints to obtain the wind speed and matrix coefficients of the P matrix over the entire range. The matrix coefficients of the P matrix are used for physical feature analysis during the full wind speed inversion process.
Citation Information
Patent Citations
CN119886224A