SAR wind field automatic inversion method and device based on machine learning
Through machine learning-based methods, wind speed and wind direction are automatically inverted from SAR images, solving the problem of inability to effectively invert high wind speed during tropical cyclones in the prior art, improving the wind field inversion accuracy, and realizing the automatic inversion of the sea surface wind field of the autonomous SAR satellite.
Patent Information
- Application Number
- CN202510047204.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-13
- Publication Date
- 2025-05-09
AI Technical Summary
The prior art cannot invert high wind speed through homopolarized SAR images during tropical cyclones due to the problem of synoptic NRCS saturation when the wind speed exceeds 30 m/s, and the cross-polarized SAR echo signal is weak, making it difficult to effectively invert the wind field.
Using a machine learning-based method, the wind speed and wind direction are automatically inverted from SAR images. By training machine learning models such as XGBoost, RF and LightGBM, the input variables such as HH polarized NRCS, HV polarized NRCS, and machine learning inverted wind direction and azimuth angle are achieved to achieve automatic inversion of wind speed and wind direction.
The accuracy of wind field inversion of SAR data under normal sea conditions is improved, and the automatic inversion of the sea surface wind field of autonomous SAR satellites is realized, especially during tropical cyclones to effectively invert high wind speeds.
Smart Images
Figure CN119962684A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of remote sensing image inversion, and in particular to a SAR wind field automatic inversion method and device based on machine learning. Background Art
[0002] Satellites with microwave frequency sensors are designed for sea surface monitoring, such as scatterometers and microwave radiometers. Remotely sensed wind speed products from scatterometers and microwave radiometers have been well calibrated and widely used in the marine science community over the past two decades. However, the limitation of these wind speed products is their relatively coarse spatial resolution of about 10 kilometers. Synthetic aperture radar (SAR) can obtain sea surface information with higher resolution, such as China's Gaofen-3 (GF-3) and its successor 1mC-SAR. Studies have shown that the co-polarization echo of SAR is positively correlated with wind speed, but fluctuates with wind direction.
[0003] With the accumulation of SAR data, the co-polarization ground model function (GMF) for wind speed inversion has been developed in the prior art. When using the co-polarization GMF for wind speed inversion, it is necessary to know the wind direction information in advance. This problem is usually solved by using the results of atmospheric numerical models, polarization correlation coefficients (PCCs) and texture features. However, during tropical cyclones (TCs), high wind speeds cannot be inverted through co-polarization SAR images because the co-polarization NRCS will be saturated when the wind speed exceeds 30m / s. Compared with the co-polarization NRCS, the cross-polarization (vertical-horizontal VH and horizontal-vertical HV) SAR echo signal is weaker because its noise equivalent zero (NESZ) is lower. Therefore, tropical cyclone wind speed is usually inverted through cross-polarization SAR images, and it is necessary to provide a SAR wind field automatic inversion method based on machine learning. Summary of the invention
[0004] In view of the defects in the prior art, the purpose of the present invention is to provide a SAR wind field automatic inversion method and device based on machine learning. The method of the present invention can improve the accuracy of wind field inversion using SAR data.
[0005] To solve the above problems, the technical solution of the present invention is:
[0006] A SAR wind field automatic inversion method based on machine learning includes the following steps:
[0007] Obtain original SAR remote sensing images, corresponding ERA-5 reanalysis wind field data, HY-2 and ASCAT scatterometer data;
[0008] Construct a SAR wind direction inversion method based on machine learning;
[0009] Construct a SAR wind speed inversion method based on machine learning;
[0010] Verify the accuracy of wind field inversion results.
[0011] Preferably, the step of obtaining the original SAR remote sensing image, the corresponding ERA-5 reanalysis wind field data, HY-2 and ASCAT scatterometer data specifically includes: collecting 1mC-SAR remote sensing image data from 2023 to 2024, ERA-5 reanalysis wind field data with a time resolution of 1h and a spatial resolution of 0.1°, and HY-2 and ASCAT scatterometer data covering the world with a spatial resolution of 0.25°.
[0012] Preferably, the step of constructing a SAR wind direction inversion method based on machine learning specifically includes: the wind direction is related to the polarization coherence coefficient, and the coherence coefficient PCC between polarizations is calculated by the following formula:
[0013]
[0014] Where PCC is a complex number representing correlation and phase difference, S PP and S QQ is the decomposed amplitude and phase, and the symbol * represents the complex conjugate.
[0015] Preferably, in the step of constructing a SAR wind direction inversion method based on machine learning, three machine learning models XGBoost, RF and LightGBM are trained to invert wind direction, the input variables include peak slope, PCC features, incident angle and azimuth obtained from the intensity spectrum, the target variable is ERA-5 wind direction, and automatic inversion of SAR wind direction is achieved through the machine learning model.
[0016] Preferably, the step of constructing a SAR wind speed inversion method based on machine learning specifically includes: training three machine learning models XGBoost, RF and LightGBM to invert wind speed, the input variables include HH polarization NRCS, HV polarization NRCS, machine learning SAR inversion wind direction and azimuth, the target variable is ERA-5 wind speed, and the automatic inversion of SAR wind speed is realized through the machine learning model.
[0017] Preferably, the step of verifying the accuracy of the wind field inversion results specifically includes: verifying the 1mC-SAR wind direction and wind speed simulation results through HY-2 and ASCAT scatterometer data, matching the two types of data in time and space, and screening by using time less than half an hour and spatial distance less than 0.5 kilometers as conditions to evaluate the accuracy of 1mC-SAR wind field inversion based on machine learning.
[0018] Furthermore, the present invention also provides a SAR wind field automatic inversion device based on machine learning, the device comprising a processor and a memory for storing executable instructions of the processor, the processor being configured to execute the SAR wind field automatic inversion method based on machine learning as described above by executing the executable instructions.
[0019] Compared with the prior art, the present invention adopts a machine learning method to automatically invert wind speed and wind direction from SAR images, realizes the automatic inversion of the sea surface wind field by autonomous SAR satellites, and improves the accuracy of wind field inversion using SAR data under normal sea conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:
[0021] Figure 1 This is a flowchart of the SAR wind field automatic inversion method based on machine learning of the present invention;
[0022] Figure 2a is the normalized radar cross section diagram under HH polarization; Figure 2b is the normalized radar cross section diagram under HV polarization;
[0023] Figure 3 This is an example of a wind field from the European Centre for Medium-Range Weather Forecasts reanalysis;
[0024] Figure 4a This is an example image of the Ocean 2C satellite; Figure 4b This is an example diagram of an advanced scatterometer;
[0025] Figure 5 Performance graphs of training wind directions for three machine learning methods;
[0026] Figure 6 This is a wind direction map obtained by combining SAR with three machine learning methods;
[0027] Figure 7 Performance graphs of training wind speed for three machine learning methods;
[0028] Figure 8 This is a case diagram of wind speed inversion;
[0029] Fig. 9 Comparison diagram of wind direction retrieved by SAR and measured data of HY-2 and ASCAT scatterometers for validation by three machine learning methods;
[0030] Fig.10 Comparison of wind speed retrieved by SAR and measured data from HY-2 and ASCAT scatterometers validated by three machine learning methods. DETAILED DESCRIPTION
[0031] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.
[0032] Specifically, the present invention provides a SAR wind field automatic inversion method based on machine learning, such as Figure 1 As shown, the method comprises the following steps:
[0033] S1: Acquire original SAR remote sensing images, corresponding ERA-5 reanalysis wind field data, HY-2 and ASCAT scatterometer data;
[0034] Specifically, 1-meter C-SAR remote sensing image data from 2023 to 2024, ERA-5 reanalysis wind field data with a temporal resolution of 1h and a spatial resolution of 0.1°, and HY-2 and ASCAT scatterometer data covering the globe with a spatial resolution of 0.25° were collected. Figure 2a is the normalized radar cross section diagram under HH polarization, Figure 2b It is the normalized radar cross section diagram under HV polarization. Figure 3 This is an example of a wind field from the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis (ERA-5). Figure 4a This is an example image of the Ocean 2C satellite (HY-2C). Figure 4b This is an example image of the Advanced Scatterometer (ASCAT), where the black rectangle indicates the geographic location of the image in Figure 2.
[0035] S2: Construct a SAR wind direction inversion method based on machine learning;
[0036] Specifically, polarimetric synthetic aperture radar (SAR) measures the amplitude and phase of a composite scattering matrix, which reflects the structure on the sea surface. Scattering theory explains the phenomenon that short-period sea surface waves modulated by wind produce periodic streaks in SAR images. This phenomenon allows the wind direction to be inferred from single-polarization SAR images. The wind direction is implicitly related to the polarization coherence coefficients (PCCs), which is due to the echo characteristics of SAR in co-polarization and cross-polarization. The PCC between polarizations can be calculated by the following formula:
[0037]
[0038] Where PCC is a complex number, representing the correlation and phase difference, S PP and S QQ is the decomposed amplitude and phase, and the symbol * represents the complex conjugate.
[0039] The advantage of machine learning is that it can handle extremely complex linear and nonlinear relationships. Using the synergistic data in dual polarization (HH and HV), three machine learning methods, namely extreme gradient boosting (XGBoost), random forest (RF) and lightweight gradient boosting machine (LightGBM), were trained to invert wind direction without relying on any external information. The input variables include peak slope, PCC characteristics, incident angle and azimuth derived from the intensity spectrum, and the target variable is the ERA-5 wind direction. The training results of the three machine learning methods are shown in Figure 2. Figure 5 As shown in the figure, the RMSE of the wind direction inverted by XGBoost is 20°, while that of RF and LightGBM is about 30° after 200 iterations of convergence. As an example, Figure 6 The wind direction obtained by SAR at 10:16 UTC on June 20, 2024, is shown, with measurements from HY-2C and ASCAT superimposed. It is observed that the SAR-retrieved wind direction is generally consistent with the wind track. The SAR wind direction obtained by three machine learning methods is paired with the scatterometer wind direction, and the temporal and spatial differences between them are within 1.5 hours and 5 kilometers.
[0040] S3: Construct a SAR wind speed inversion method based on machine learning;
[0041] Specifically, the advantage of machine learning is that it can handle extremely complex linear and nonlinear relationships. The input variables include HH polarization NRCS, HV polarization NRCS, machine learning SAR inversion wind direction and azimuth, and the target variable is ERA-5 wind speed. The automatic inversion of SAR wind speed is achieved through the machine learning model. The performance of the three machine learning methods for training wind speed is as follows: Figure 7 As shown, namely (a) XGBoost, (b) Random Forest (RF) and (c) LightGBM. The study found that XGBoost outperformed RF and LightGBM. Three machine learning models and GMF CSARMOD-GF were used to invert wind speed using wind direction inverted by SAR. Figure 8 The figure shows the wind speed inversion cases at 10:16 UTC on June 20, 2024, namely (a) the map obtained by GMF CSARMOD-GF, (b) the map obtained by XGBoost, (c) the map obtained by RF, and (d) the map obtained by LightGBM. The color points represent the measurement data from HY-2 and ASCAT.
[0042] S4: Verify the accuracy of wind field inversion results.
[0043] Specifically, the HY-2 and ASCAT scatterometer data were used to verify the 1mC-SAR wind direction and wind speed simulation results. The two types of data were matched in time and space, and the time less than half an hour and the spatial distance less than 0.5 kilometers were used as conditions for screening to evaluate the accuracy of 1mC-SAR wind field inversion based on machine learning.
[0044] Fig. 9 Validation of SAR-derived wind direction against HY-2 and ASCAT scatterometer measurements using three machine learning methods: (a) XGBoost, (b) RF, and (c) LightGBM. Fig.10 The statistical analysis results of wind speed of GMF CSARMOD-GF and three machine learning models are shown, namely (a) the results of GMF CSARMOD-GF, (b) the results of XGBoost, (c) the results of RF and (d) the results of LightGBM. The study found that the wind speed RMSE of XGBoost was 1.37m / s, the correlation coefficient (Cor) was 0.87, and the standard deviation (SI) was 0.27, which were better than the results of RF and LightGBM. In addition, compared with the 1.87m / sRMSE, 0.79 correlation coefficient and 0.38 standard deviation of GMF CSARMOD-GF, XGBoost improved the accuracy of wind speed inversion by 0.5m / s.
[0045] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.
Claims
1. A SAR wind field automatic inversion method based on machine learning, characterized in that: The method comprises the following steps: Obtain original SAR remote sensing images, corresponding ERA-5 reanalysis wind field data, HY-2 and ASCAT scatterometer data; Construct a SAR wind direction inversion method based on machine learning; Construct a SAR wind speed inversion method based on machine learning; Verify the accuracy of wind field inversion results.
2. The SAR wind field automatic inversion method based on machine learning according to claim 1 is characterized in that: The steps of obtaining the original SAR remote sensing image, the corresponding ERA-5 reanalysis wind field data, HY-2 and ASCAT scatterometer data specifically include: collecting 1mC-SAR remote sensing image data from 2023 to 2024, ERA-5 reanalysis wind field data with a time resolution of 1h and a spatial resolution of 0.1°, and HY-2 and ASCAT scatterometer data covering the world with a spatial resolution of 0.25°.
3. The SAR wind field automatic inversion method based on machine learning according to claim 1 is characterized in that: The steps of constructing a SAR wind direction inversion method based on machine learning specifically include: the wind direction is related to the polarization coherence coefficient, and the coherence coefficient PCC between polarizations is calculated by the following formula: Where PCC is a complex number representing correlation and phase difference, S PP and S QQ is the decomposed amplitude and phase, and the symbol * represents the complex conjugate.
4. The SAR wind field automatic inversion method based on machine learning according to claim 3 is characterized in that: In the step of constructing a SAR wind direction inversion method based on machine learning, three machine learning models XGBoost, RF and LightGBM are trained to invert wind direction. The input variables include peak slope, PCC characteristics, incident angle and azimuth obtained from the intensity spectrum. The target variable is ERA-5 wind direction. Automatic inversion of SAR wind direction is achieved through the machine learning model.
5. The SAR wind field automatic inversion method based on machine learning according to claim 1 is characterized in that: The steps of constructing a SAR wind speed inversion method based on machine learning specifically include: training three machine learning models XGBoost, RF and LightGBM to invert wind speed, the input variables include HH polarization NRCS, HV polarization NRCS, machine learning SAR inversion wind direction and azimuth, the target variable is ERA-5 wind speed, and the automatic inversion of SAR wind speed is realized through the machine learning model.
6. The SAR wind field automatic inversion method based on machine learning according to claim 1 is characterized in that: The step of verifying the accuracy of the wind field inversion results specifically includes: verifying the 1mC-SAR wind direction and wind speed simulation results through HY-2 and ASCAT scatterometer data, matching the two types of data in time and space, and screening them by using time less than half an hour and spatial distance less than 0.5 kilometers as conditions to evaluate the accuracy of 1mC-SAR wind field inversion based on machine learning.
7. A SAR wind field automatic inversion device based on machine learning, characterized in that: The device includes a processor and a memory for storing executable instructions of the processor, and the processor is configured to perform the SAR wind field automatic inversion method based on machine learning as described in any one of claims 1 to 6 by executing the executable instructions.