Dual-frequency scatterometer wind field joint inversion method based on deep learning
By employing a deep learning-based dual-frequency scatterometer wind field joint inversion method, and utilizing multi-angle, multi-polarization data from the Ku and C bands, a dual-branch deep learning model is constructed. This solves the problems of large decoupling errors between wind speed and wind direction and insufficient single-frequency observations in existing technologies, and achieves high-precision wind field prediction under complex sea conditions.
Patent Information
- Application Number
- CN202511879932.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-05-05
AI Technical Summary
Existing technologies for sea surface wind field inversion suffer from large decoupling errors between wind speed and wind direction, insufficient single-frequency observation information, and difficulties in handling complex sea conditions due to reliance on background fields, which affect the accuracy and adaptability of inversion.
A deep learning-based dual-frequency scatterometer wind field joint inversion method is adopted. By constructing a dual-branch deep learning model and sharing the feature extraction backbone network, multi-angle and multi-polarization data of Ku and C bands are used to realize the simultaneous inversion of wind speed and wind direction. Combined with physical enhancement features and adaptive interaction layer, error propagation is reduced.
It achieves high-precision inversion of wind speed and direction simultaneously, improves robustness and adaptability under complex sea conditions, reduces dependence on external background fields, and significantly improves wind field prediction performance.
Smart Images
Figure CN121980905A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a deep learning-based method for joint inversion of wind fields using dual-frequency scatterometers, specifically a deep learning-based method for joint inversion of sea surface wind speed and direction based on Ku and C dual-frequency scatterometer observation data, belonging to the field of space remote sensing and marine environmental monitoring technology. Background Technology
[0002] Sea surface wind field is a core physical quantity characterizing air-sea interaction and has important applications in ocean dynamics research, wind energy resource assessment, and disaster monitoring. Spaceborne scatterometers are one of the key payloads for remote sensing of ocean wind fields, primarily operating in the microwave band (such as the Ku and C bands). Their basic principle is to transmit microwave pulses to the sea surface and receive backscattered echo signals, achieving sea surface wind field inversion by establishing a geophysical model function (GMF) between the backscattering coefficient and the sea surface wind vectors (wind speed and direction). However, traditional methods have the following shortcomings: 1. Large decoupling error between wind speed and wind direction: There is a strong correlation between the two, which usually requires a prior background field for constraint and is prone to introducing cumulative deviation.
[0003] 2. Insufficient information from single-frequency observations: It is difficult to fully characterize the roughness features of the sea surface at multiple scales, and the inversion performance deteriorates under low wind speeds and complex sea conditions.
[0004] 3. Reliance on background field and iterative optimization: The processing flow is complex and the computational load is large, which may lead to misjudgment in sudden ocean-atmosphere events.
[0005] In summary, existing traditional inversion methods are limited in areas such as multi-band fusion and handling of complex sea conditions, which affects the accuracy of wind field inversion. There is an urgent need to develop new intelligent inversion technologies that can integrate multi-frequency information and improve the utilization of scatterometer data. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and solve the problems of insufficient accuracy and poor adaptability of the existing inversion technology.
[0007] The objective of this invention is achieved through the following technical solutions: A deep learning-based method for joint inversion of wind fields using dual-frequency scatterometers includes: Simulation of multi-angle measurement data from a dual-band scatterometer on a spaceborne satellite was conducted to obtain simulation data; physical enhancement features were constructed based on the sea surface scattering mechanism of the scatterometer. A dual-branch deep learning wind field inversion model is established, including a full wind speed model and a wind speed sub-model, sharing a feature extraction backbone network, with two output branches to realize wind speed and wind direction inversion respectively; Based on simulation data, the full wind speed model and the wind speed component model were trained. Wind field inversion is performed based on the trained full-wind speed model and wind speed component model.
[0008] In one embodiment of the present invention, the simulation of multi-angle measurement data on a dual-band scatterometer is performed. The backscattering coefficient data of the Ku-band VV and VH polarizations are generated using the NSCAT-4DS model, the backscattering coefficient of the C-band VV is generated using the CMOD7 model, and the backscattering coefficient of the C-band VH channel is generated using the "H14S" model. The simulation data correspond to the sea surface scattering characteristics under different wind direction conditions.
[0009] In one embodiment of the present invention, the physical features include: Relative azimuth sine and cosine characteristics: reflect the wind direction modulation effect; Second and third cubic polynomial characteristics of wind speed: expressing nonlinear wind field response; Polarization difference characteristics: characterizing the sensitivity of Bragg scattering to polarization; Ku and C channel differential features: Enhanced expression of surface roughness and frequency dependence.
[0010] In one embodiment of the present invention, in the dual-branch deep learning wind field inversion model: Shared feature layer: Two-layer ReLU fully connected network with Dropout introduced; Adaptive interaction layer: Enhances feature fusion capabilities through learnable weights; Wind speed regression branch: Outputs non-negative continuous wind speed prediction values; Wind direction vector branch: Outputs a normalized wind direction cosine / sine vector to avoid angular periodicity issues.
[0011] In one embodiment of the present invention, training the full wind speed model includes: Based on simulation data, after standardization, the data is input into the full wind speed model. The full data is divided into training and test sets in an 8:2 ratio, and then a validation set is drawn from the training set. The test set is used to test noise robustness by selecting whether to inject independent random noise according to the experimental scheme. The model was trained on a training set containing multi-scan angle Ku and C dual-frequency backscattering observation data. During the training phase, a random mixed noise enhancement strategy was adopted to improve the stability of the model in complex sea surface environments.
[0012] In one embodiment of the present invention, training the wind speed model includes: Based on simulation data and according to the statistical law of wind field observation by scatterometer, the wind speed value is divided into multiple intervals. The data of each wind speed interval is divided into training set and test set in an 8:2 ratio. Then, a validation set is divided from the training set to train an independent model. Customized optimization is carried out for the differences in scattering coefficient response characteristics in different wind speed segments. The model was trained on a training set containing multi-scan angle Ku and C dual-frequency backscattering observation data. During the training phase, a random mixed noise enhancement strategy was adopted to improve the stability of the model in complex sea surface environments.
[0013] In one embodiment of the present invention, wind field inversion includes the following: Complete the extraction and standardization of physical enhancement features for the test set; Sub-models that meet the minimum sample size threshold are selected based on wind speed segments; During the inversion process, the full wind speed model is executed first to perform wind speed inversion, and then the corresponding sub-wind speed model is called to complete the inversion. The final output is the wind speed and wind direction angle in true dimensions.
[0014] In one embodiment of the present invention, the inversion method further includes an inversion performance evaluation, specifically including: S51. Perform error-free inversion; S52. Conduct a noise robustness test; S53. Performance and physical interpretability analysis by wind speed segment.
[0015] In one embodiment of the present invention, in S52, a noise robustness test is conducted by adding random errors to the test set.
[0016] In one embodiment of the present invention, in S53, the sensitivity of the backscattering coefficient to the wind direction is calculated for the test sample according to the wind speed range, which is used for the performance and physical interpretability analysis of the wind speed range.
[0017] Compared with the prior art, the present invention has the following advantages: (1) This invention constructs an end-to-end deep learning inversion model, which uses backscattered data under different polarization and observation geometry conditions with Ku and C frequencies to simultaneously infer sea surface wind speed and direction. For joint inversion of wind speed and direction, a multi-task learning architecture is used to share the underlying feature representation, avoiding the error amplification caused by mathematical coupling in traditional step-by-step inversion. For the collaborative fusion of dual-frequency band features, the physical advantages of Ku band being sensitive to shortwave rough structures and C band being more robust to low wind and rain areas are combined to improve the adaptability to the entire wind speed range. For the background field-free mechanism, the model directly inputs the backscattered original observation data, without the need for external reanalysis fields, numerical models or empirical wind fields, reducing additional error sources and achieving fully data-driven inversion.
[0018] (2) This invention, while ensuring the model's universality and generalization ability, introduces physical prior feature enhancement, wind speed segmentation (speed-bin) fine-tuning strategy, and trend regularization constraints to improve the accuracy of wind speed and wind direction inversion. Compared with the traditional GMF iterative method, this invention has significant innovations and advantages as follows: It adopts a multi-task learning structure to reduce error propagation and improve wind direction stability, and wind speed and wind direction can be inverted simultaneously; the dual-frequency fusion mechanism enhances the inversion sensitivity of high-wind and low-wind areas and can handle multi-frequency scattering characteristics; it is a fully data-driven inversion, which improves the reliability and autonomy under sudden extreme weather conditions and does not rely on the background field.
[0019] (3) This method adopts an end-to-end dual-task learning model to achieve simultaneous inversion of wind speed and wind direction. By integrating backscattering characteristics under different frequency bands and polarizations, and introducing a physical enhancement mechanism and a piecewise adaptive strategy for wind speed, it maintains high accuracy and robustness under complex sea surface conditions. Compared with traditional iterative algorithms based on geophysical model functions, the method of this invention does not rely on the external background wind field, can avoid decoupling error amplification, significantly improves wind field prediction performance, and has important engineering application prospects. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the variation curves of the Ku and C dual-frequency backscattering coefficients generated by the simulation of this invention with wind direction, showing the observed response of different polarizations under a wind speed of about 10 m / s.
[0021] Figure 2 This is a schematic diagram of the loss convergence curve during the model training process, including the changes in training error and validation error of wind speed and direction with each training round.
[0022] Figure 3 The schematic diagram shows the wind direction inversion results after applying the method of this invention with a 0.2dB error added to the test set, demonstrating the improved robustness.
[0023] Figure 4 The schematic diagram shows the wind direction inversion results after applying the method of this invention with a 0.2dB error added to the test set, demonstrating the improved robustness.
[0024] Figure 5 This is a comparison chart showing the changes in RMSE (Relative Strength and Root Cause Analysis) based on wind direction inversion in various wind speed ranges according to the present invention. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0026] A deep learning-based method for joint inversion of wind fields using dual-frequency scatterometers includes: (1) Data preprocessing and physical feature construction Simulations were performed using multi-angle measurements from a dual-band scatterometer onboard aerospace. The Ku-band VV and VH polarization backscattering coefficients were generated using the NSCAT-4DS model, the C-band VV backscattering coefficients using the CMOD7 model, and the C-band VH channel backscattering coefficients using the "H14S" model. The simulation data correspond to sea surface scattering characteristics under different wind directions. The relationship between the Ku and C-band dual-frequency fully polarized responses and wind direction at a typical wind speed of 10 m / s is shown below. Figure 1 As shown, different frequency bands exhibit significant differences in their sensitivity to sea surface roughness and wind direction modulation, making them suitable for dual-frequency joint inversion.
[0027] Simulation data containing backscattering coefficients and relative azimuth information of different polarizations and scanning beams were used to train a subsequent bi-branch deep learning wind field inversion model. The original features used included Ku-band VV, Ku-band HH, C-band VV, and C-band VH polarization channels, as well as the observed azimuth angle φ of the corresponding scanning sequence. The ground truth wind direction was converted into cosine and sine forms as wind direction monitoring labels.
[0028] Physical enhancement features were constructed based on the sea surface scattering mechanism of the scatterometer, including: Relative azimuth sine and cosine characteristics: reflect the wind direction modulation effect; Second and third cubic polynomial characteristics of wind speed: expressing nonlinear wind field response; Polarization differential characteristics (PR): characterize the sensitivity of Bragg scattering to polarization; Ku and C channel differential features: Enhanced expression of surface roughness and frequency dependence.
[0029] (2) Dual-branch deep learning wind field inversion model A dual-branch deep learning wind field inversion model based on PyTorch is constructed, including a full wind speed model and a wind speed-specific model. They share a core feature extraction network, and wind speed and direction inversion are achieved by two output branches respectively. Shared feature layer: Two-layer ReLU fully connected network with Dropout introduced; Adaptive interaction layer: Enhances feature fusion capabilities through learnable weights; Wind speed regression branch: Outputs non-negative continuous wind speed prediction values; Wind direction vector branch: Outputs a normalized wind direction cosine / sine vector to avoid angular periodicity issues.
[0030] (3) Model training Full wind speed model: The simulation data obtained in step 1 is standardized and then input into the dual-branch deep learning model constructed in step 2. The entire dataset is divided into a training set and a test set in an 8:2 ratio, and a validation set is then derived from the training set. The test set may be injected with independent random noise, depending on the experimental design, for noise robustness testing.
[0031] Wind speed model: Based on the simulation data obtained in step 1, and according to the statistical laws of wind field observation by the scatterometer, the wind speed values are divided into multiple intervals (speed-bin). The data of each wind speed interval are divided into training and test sets in an 8:2 ratio, and a validation set is then drawn from the training set to train an independent model. Customized optimization is performed to address the differences in scattering coefficient response characteristics in different wind speed segments.
[0032] The model was trained on a training set containing multi-scan angle Ku and C dual-frequency backscattering observations. During training, a random mixed noise enhancement strategy was employed to improve the model's stability in complex sea surface environments. The training loss curve and validation set RMSE variation of the full wind speed model are shown below. Figure 2 As shown, the wind speed and wind direction inversion results both converge stably.
[0033] (4) Wind field inversion Based on the trained dual-branch deep learning wind field inversion model, wind field inversion is performed on the test set. The specific steps are as follows: Complete the extraction and standardization of physical enhancement features for the test set; Sub-models that meet the minimum sample size threshold are selected based on wind speed segments; During the inversion process, the full wind speed model is executed first to perform wind speed inversion, and then the corresponding sub-wind speed model is called to complete the inversion. The final output is the wind speed and wind direction angle in true dimensions.
[0034] Preferably, after step (4), the method of the present invention further includes step (5) wind field inversion performance evaluation, as follows: S51, Error-Free Inversion Wind field inversion was performed on a noise-free test set, and the inversion results were evaluated. The RMSE of wind speed was 0.034 m / s, and the RMSE of wind direction was 7.15°. The results show that the bi-branch deep learning wind field inversion model constructed in this invention can achieve high accuracy and stability under noise-free conditions. S52, Noise Robustness Test Adding a 0.2 dB random error to the test set, the results show that the wind speed inversion error is 0.324 m / s and the wind direction inversion error is 12.496° (see wind speed inversion results). Figure 3 The wind direction inversion results are shown below. Figure 4The results show that the bi-branch deep learning wind field inversion model constructed in this invention has an increased wind field inversion error under noisy conditions compared to the noise-free conditions, but it can still achieve good inversion accuracy. S53, Performance and Physical Interpretability Analysis by Wind Speed Section The sensitivity of wind direction (|dσ) to the backscattering coefficient was calculated for the test samples according to wind speed range. 0 The data obtained are shown in Table 1. Table 1 shows that the sensitivity decreases significantly in the 14~16 m / s wind speed range. This phenomenon corresponds to the shift in the sea surface scattering mechanism from Bragg dominance to foam and multiple scattering dominance, resulting in a reduction in wind direction information and a sudden increase in wind direction error in this range (see...). Figure 5 ).
[0035] Table 1. Statistics on the variation of backscattering sensitivity to wind direction with wind speed (Ku / C three-channel)
[0036] The contents not described in detail in this specification are common knowledge to those skilled in the art.
[0037] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.
Claims
1. A deep learning-based method for joint inversion of wind fields using dual-frequency scatterometers, characterized in that, include: Simulation of spaceborne multi-angle measurement data from a dual-band scatterometer was performed to obtain simulation data. Constructing physical enhancement features based on the sea surface scattering mechanism of scatterometers; A dual-branch deep learning wind field inversion model is established, including a full wind speed model and a wind speed sub-model, sharing a feature extraction backbone network, with two output branches to realize wind speed and wind direction inversion respectively; Based on simulation data, the full wind speed model and the wind speed component model were trained. Wind field inversion is performed based on the trained full-wind speed model and wind speed component model.
2. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, Simulation of spaceborne multi-angle measurement data from a dual-band scatterometer: Ku-band VV and VH polarized backscattering coefficient data were generated using the NSCAT-4DS model, C-band VV backscattering coefficient was generated using the CMOD7 model, and C-band VH channel backscattering coefficient was generated using the "H14S" model; the simulation data correspond to the sea surface scattering characteristics under different wind direction conditions.
3. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, Physical characteristics include: Relative azimuth sine and cosine characteristics: reflect the wind direction modulation effect; Second and third cubic polynomial characteristics of wind speed: expressing nonlinear wind field response; Polarization difference characteristics: characterizing the sensitivity of Bragg scattering to polarization; Ku and C channel differential features: Enhanced expression of surface roughness and frequency dependence.
4. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, In a two-branch deep learning wind field inversion model: Shared feature layer: Two-layer ReLU fully connected network with Dropout introduced; Adaptive interaction layer: Enhances feature fusion capabilities through learnable weights; Wind speed regression branch: Outputs non-negative continuous wind speed prediction values; Wind direction vector branch: Outputs a normalized wind direction cosine / sine vector to avoid angular periodicity issues.
5. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, Training the full wind speed model includes: Based on simulation data, after standardization, the data is input into the full wind speed model. The full data is divided into training and test sets in an 8:2 ratio, and then a validation set is drawn from the training set. The test set is used to test noise robustness by selecting whether to inject independent random noise according to the experimental scheme. The model was trained on a training set containing multi-scan angle Ku and C dual-frequency backscattering observation data. During the training phase, a random mixed noise enhancement strategy was adopted to improve the stability of the model in complex sea surface environments.
6. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, Training the wind speed model includes: Based on simulation data and according to the statistical law of wind field observation by scatterometer, the wind speed value is divided into multiple intervals. The data of each wind speed interval is divided into training set and test set in an 8:2 ratio. Then, a validation set is divided from the training set to train an independent model. Customized optimization is carried out for the differences in scattering coefficient response characteristics in different wind speed segments. The model was trained on a training set containing multi-scan angle Ku and C dual-frequency backscattering observation data. During the training phase, a random mixed noise enhancement strategy was adopted to improve the stability of the model in complex sea surface environments.
7. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, Wind field inversion includes the following: Complete the extraction and standardization of physical enhancement features for the test set; Sub-models that meet the minimum sample size threshold are selected based on wind speed segments; During the inversion process, the full wind speed model is executed first to perform wind speed inversion, and then the corresponding sub-wind speed model is called to complete the inversion. The final output is the wind speed and wind direction angle in true dimensions.
8. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 1, characterized in that, The inversion method also includes an inversion performance evaluation, specifically including: S51. Perform error-free inversion; S52. Conduct a noise robustness test; S53. Performance and physical interpretability analysis by wind speed segment.
9. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 8, characterized in that, In S52, noise robustness tests are conducted by adding random errors to the test set.
10. The method for joint inversion of wind field using a dual-frequency scatterometer according to claim 8, characterized in that, In S53, the sensitivity of the backscattering coefficient to wind direction is calculated for the test samples according to the wind speed range, which is used for the performance and physical interpretability analysis of each wind speed range.