Ocean current inversion method combining SAR Doppler information and deep learning

Abstract

The invention discloses a synthetic aperture radar (SAR) Doppler information and deep learning combined ocean current inversion method, which belongs to the technical field of laser radar data processing, is used for ocean current inversion, and comprises the following steps of: performing radio frequency interference correction by using a YOLO network based on a YOLOv8 architecture, inputting a to-be-processed tensor into a radio frequency interference subspace, and obtaining a to-be-processed tensor; obtaining a projection component of the to-be-processed tensor in the radio frequency interference subspace, and subtracting the projection component from the to-be-processed tensor to obtain a signal after radio frequency interference correction; and respectively calculating the components of the Doppler center frequency shift along the distance direction and the azimuth direction to obtain the Doppler frequency shift and invert the radial velocity of the sea surface flow field. According to the method, radio frequency interference is suppressed through the deep learning model, and the precision of ocean current inversion is indirectly improved; two components of Doppler center frequency shift are accurately calculated, the calculation precision of Doppler frequency shift is improved, and the radial velocity of the sea surface flow field is accurately predicted.

Description

Technical Field

[0001] The invention discloses an ocean current inversion method combining SAR Doppler information with deep learning, and belongs to the technical field of laser radar data processing. Background Art

[0002] SAR echo signals contain Doppler frequency shift components caused by the relative motion between the satellite and the ground and Doppler frequency shift components caused by the motion of the sea surface itself. That is to say, there is a linear relationship between the Doppler frequency shift anomaly in the SAR echo signal and the apparent motion speed of the sea surface. Based on this principle, ocean current inversion can be performed through SAR echo signals. However, when SAR is used for ocean current inversion, if radio frequency interference covers the ocean echo signal, it will seriously affect the echo spectrum of the radar network, resulting in errors in the inversion results of the ocean surface radial flow, which in turn affects the detection performance of the radar system and ultimately affects the ocean current inversion results. Summary of the invention

[0003] The purpose of the present invention is to provide a current inversion method combining SAR Doppler information with deep learning, so as to solve the problem in the prior art that radio frequency interference affects the current inversion results.

[0004] A current inversion method combining SAR Doppler information and deep learning is proposed. The YOLO network is used for radio frequency interference correction, and the components of the Doppler center frequency shift along the range and azimuth directions are calculated respectively to obtain the Doppler frequency shift and invert the radial velocity of the current field.

[0005] The YOLO network is based on the YOLOv8 architecture and includes a backbone network, a neck network, and a head network.

[0006] The backbone network sequentially includes an initial convolution layer and four sparse convolution layers, the four sparse convolution layers are connected to a spatial pyramid pooling layer, and each sparse convolution layer sequentially includes a submanifold sparse convolution layer and a convolution layer.

[0007] The neck network includes extracting the output features of the second and third sparse convolutional layers and the output features of the spatial pyramid pooling layer, fusing the three output features, and outputting the three fused features.

[0008] The neck network uses the FPN feature fusion module to fuse the three output features.

[0009] The head network imports the fused features into three convolutional layers for decoding.

[0010] Using the YOLO network for RFI correction includes: S1. Perform two fast Fourier transforms on the radar echo data, use the transformed radar data RDS to mark the interference object, and make a data set to input into the YOLO network for training; S2, re-collect radar echo data and input it into the trained YOLO network, output the time when RF interference exists and RDS, and construct the tensor to be processed; S3, inputting the tensor to be processed into the radio frequency interference subspace to obtain the projection component of the tensor to be processed in the radio frequency interference subspace; S4. Subtract the projection component from the tensor to be processed to obtain a signal after radio frequency interference correction.

[0011] Calculating the components of the Doppler center frequency shift along the range direction includes: ; In the formula, is the component of the Doppler center frequency shift along the range direction, is the initial transmission frequency, is the equivalent mean tilt viewing angle, is the initial distance between the radar and the target, is the real-time distance between the radar and the target, is the radar altitude.

[0012] Calculation of the Doppler center frequency shift along the azimuth direction includes: ; In the formula, is the component of the Doppler center frequency shift along the azimuth direction, is the wavelength, For time, is the frequency constant term, is the relative speed.

[0013] Get Doppler shift include: ; In the formula, is the Doppler frequency shift component caused by the relative motion between the satellite and the ground, is the Doppler frequency shift component caused by the motion of the sea surface itself, and Combined to obtain.

[0014] Compared with the prior art, the present invention has the following beneficial effects: the present invention suppresses radio frequency interference through a deep learning model, thereby indirectly improving the accuracy of ocean current inversion; accurately calculates the two components of the Doppler center frequency shift, improves the calculation accuracy of the Doppler frequency shift, and accurately predicts the radial velocity of the sea surface flow field. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is the position of the ground target in real space; Figure 2 is the position of the ground target in the signal space; Figure 3 is the position of the ground target in the image space; Figure 4 It is a deep learning network structure; Figure 5 The distribution of ocean currents in the experimental area during the passage of a hurricane; Figure 6 To calculate the flow vector size in the experimental area based on Glob Current data; Figure 7 is an image presented by projection of the corresponding ocean current image; Figure 8 Based on Figure 6 and Figure 7 Calculated predicted radial velocity; Fig. 9 is the Doppler real-time radial velocity of the sea surface; Fig.10 is the Euclidean distance inversion result obtained by projecting onto the direction of satellite motion; Fig.11 is the best prediction of the radial velocity obtained in the best iteration; Fig.12 is the ocean current vector under the best iteration; Fig.13 is the azimuth under the best iteration; Fig.14 is the longitude component of the stream angle; Fig.15 is the latitude component of the stream angle. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] A current inversion method combining SAR Doppler information and deep learning is proposed. The YOLO network is used for radio frequency interference correction, and the components of the Doppler center frequency shift along the range and azimuth directions are calculated respectively to obtain the Doppler frequency shift and invert the radial velocity of the current field.

[0018] The YOLO network is based on the YOLOv8 architecture and includes a backbone network, a neck network, and a head network.

[0019] The backbone network sequentially includes an initial convolution layer and four sparse convolution layers, the four sparse convolution layers are connected to a spatial pyramid pooling layer, and each sparse convolution layer sequentially includes a submanifold sparse convolution layer and a convolution layer.

[0020] The neck network includes extracting the output features of the second and third sparse convolutional layers and the output features of the spatial pyramid pooling layer, fusing the three output features, and outputting the three fused features.

[0021] The neck network uses the FPN feature fusion module to fuse the three output features.

[0022] The head network imports the fused features into three convolutional layers for decoding.

[0023] Using the YOLO network for RFI correction includes: S1. Perform two fast Fourier transforms on the radar echo data, use the transformed radar data RDS to mark the interference object, and make a data set to input into the YOLO network for training; S2, re-collect radar echo data and input it into the trained YOLO network, output the time when RF interference exists and RDS, and construct the tensor to be processed; S3, inputting the tensor to be processed into the radio frequency interference subspace to obtain the projection component of the tensor to be processed in the radio frequency interference subspace; S4. Subtract the projection component from the tensor to be processed to obtain a signal after radio frequency interference correction.

[0024] Calculating the components of the Doppler center frequency shift along the range direction includes: ; In the formula, is the component of the Doppler center frequency shift along the range direction, is the initial transmission frequency, is the equivalent mean tilt viewing angle, is the initial distance between the radar and the target, is the real-time distance between the radar and the target, is the radar altitude.

[0025] Calculation of the Doppler center frequency shift along the azimuth direction includes: ; In the formula, is the component of the Doppler center frequency shift along the azimuth direction, is the wavelength, For time, is the frequency constant term, is the relative speed.

[0026] Get Doppler shift include: ; In the formula, is the Doppler frequency shift component caused by the relative motion between the satellite and the ground, is the Doppler frequency shift component caused by the motion of the sea surface itself, and Combined to obtain.

[0027] The present invention first introduces the principle of Doppler current inversion. SAR is a high-resolution pulse Doppler imaging radar that uses Doppler frequency shift theory to achieve target resolution. The radar uses Doppler frequency shift to improve azimuth resolution in the synthetic aperture process, while pulse compression is used to improve distance resolution. In data processing, the target position is determined by zero Doppler line and zero Doppler moment. Doppler center estimation technology is crucial to the success of imaging. Doppler synthesis is used to explain the synthetic aperture principle, and a virtual antenna is formed by continuous beamforming for high-resolution imaging. When the main flap of the radar beam approaches, is orthogonal to, and moves away from the observed target point A, the Doppler frequency shift of the return signal gradually changes from positive to zero and then to negative. The Doppler bandwidth determines the resolution of the SAR in azimuth, and the influence of the Doppler effect on the SAR imaging process is reflected here. The distance direction mainly depends on the delay time of the echo pulse. In the distance direction, two points on the ground are distinguishable, so the minimum time difference between the echoes of the two points is equal to the width of the transmitted pulse.

[0028] In SAR target observation, the distance and azimuth directions are always perpendicular. When the radar transmits a continuous beam, it illuminates the ground to form a surface rather than a point or line. At each moment, the beam must have a projection line directly on the ground, called the zero Doppler line. The ground fixed target is closest to the satellite and the frequency shift is zero, which is the zero Doppler moment. The zero Doppler line and the vertical azimuth line of sight constitute the zero Doppler plane. Figure 1 , Figure 2 , Figure 3 Respectively represent the ground targets in real space , signal space and Image Space In the position relationship, A, B, C, and D are the four points to be measured. The radar antenna measures at the target interval, usually not perpendicular to the azimuth direction, and points to the target position along the beam centerline. Zero Doppler compression causes the signal line to focus on the zero Doppler moment, and the target becomes a point again. When located in the zero Doppler direction, the moment is independent of the slant angle, and the final image target position does not depend on the slant angle, eliminating the need for slant range conversion.

[0029] The development of Doppler center imaging technology has made SAR the only available equipment for high-precision ocean current observation. In the embodiment of the present invention, Doppler information is first collected. The Sentinel-1A and Sentinel-1B satellites launched in 2014 and 2016 provide ocean surface radial current data in combination with Doppler center imaging technology. The Sentinel-1 satellite is C-band, providing multi-mode data at an altitude of 700 kilometers and a revisit period of 12 days. The resolution of the IW mode Doppler pixel used for flow field inversion is 1km×1km, and the Doppler pixel is obtained from the satellite raw data through the Doppler center estimation technology.

[0030] Sentinel-1 products are divided into three levels: Level 0 is used for raw data, Level 1 is used for single view complex image (SLC) and ground range detection (GRD) products, and Level 2 is used for ocean data products (OCN), which include surface surge spectrum (OSW), surface wind field (OWI) and surface radial flow (RVL) products. The dataset used in the present invention mainly includes metadata from level 1 SLC data and level 2 OWI and RVL products. When downloading data, first download the OCN data, and then retrieve the corresponding SLC product metadata. The Doppler center parameters in the SLC metadata are converted to a common data format for subsequent processing and interpolation to generate images. The parameters are concentrated into a common data format for subsequent processing and interpolation to generate images. The RVL data of the OCN product contains latitude and longitude information that can be used to calibrate the Doppler image. In addition, the OWI product of the OCN product can be combined with the Doppler image for joint inversion of ocean currents.

[0031] The time-varying sea surface has only a single observation orbit, so a full geometric analysis of this single ascending or descending orbit is required in order to find a projection analysis scheme suitable for this single orbit. For different orbits and side-view states, the flow in certain directions has different projection angles, so a specific analysis needs to be given for specific conditions. And for specific side-view and orbital states, the flow vector on the actual sea surface can be oriented in any direction within the plane, which can also make the analysis situation extremely complicated due to the existence of projection. For example, in the case of right-up descending orbit right-view, the projection may lead to a multi-resolution situation due to the angular difference between the satellite's flight direction and the flow. However, the actual satellite reception of Doppler shift is complex, and the distribution of the Doppler center measured by SAR is characterized by linear changes in the azimuth and range directions.

[0032] To achieve radio frequency interference correction, the deep learning network structure used in the present invention is as follows Figure 4As shown, based on the YOLOv8 architecture, it includes a backbone network, a neck network and a head network. The backbone network includes an initial convolution layer and four sparse convolution layers in sequence, and the four sparse convolution layers are connected to a spatial pyramid pooling layer, and each sparse convolution layer includes a submanifold sparse convolution layer and a convolution layer in sequence. The neck network includes extracting the output features of the second and third sparse convolution layers and the output features of the spatial pyramid pooling layer, performing feature fusion on the three output features, and outputting three fused features. The neck network uses the FPN feature fusion module to perform feature fusion on the three output features, and the head network imports the fused features into three convolution layers for decoding. After collecting the data set, it is divided into a training set and a validation set, and the deep learning network training is performed first. After the training is completed, the data to be processed is sent to the deep learning network for processing to obtain a data set after radio frequency interference correction, and then the ocean current inversion is performed.

[0033] Complex sea surface current conditions cause multi-resolution situations in the projection, which brings great inconvenience to the current inversion. One solution to this problem is to give priority to the inversion of certain areas with better flow anisotropy, such as areas of geostrophic currents or coastal currents. For areas with large flow variations, such as mesoscale eddies or Ekman flow areas, the inversion method needs to be further explored and optimized. Figure 5 Shown is the flow distribution in the experimental area during a hurricane passage simulated by HYCOM.

[0034] The radial velocity field calculated based on the SAR Doppler center anomaly provides real-time sea surface information, while the predicted sea surface flow field can be provided from historical data, so iterative approximation between the two is a possible method to obtain reasonable sea current information. The historical data used is still Glob Current data, and the real-time sea surface radial motion information is still extracted from the metadata of Sentinel-1 SARSLC products and converted into radial Doppler velocity. Figure 6 is the flow vector size of the experimental area calculated based on GlobCurrent data, Figure 7 is the image presented by the projection of the corresponding flow image, Figure 8 Based on Figure 6 and Figure 7 The predicted radial velocity calculated, Fig. 9 is the Doppler real-time radial velocity of the sea surface. The imaging time is around 13:00 on January 20, 2023. For the convenience of iterative calculation, the vertical and horizontal coordinates are changed from longitude and latitude coordinates to rectangular coordinates.

[0035] The radial velocity predicted by the present invention Calculated as: ; In the formula, is the azimuth, is the relative speed. The results calculated from historical data are continuously corrected to be close enough to the results of SAR measurements, so that it can be considered that the inversion has produced a sea surface current field close to the real one. The key process of this algorithm is iterative correction: first, the magnitude and flow angle of the initial field current vector are calculated from the Glob Current data, and the first guess prediction is given from this, and then the difference between the image and the measured image is calculated. Euclidean distance to obtain the initial image distance : ; In the formula, and is the pixel value of each of the two images involved in the calculation, and is the pixel position in the vertical and horizontal coordinates of the image. The first judgment process begins, and then continuous iterative correction is performed by modifying and , so that The predicted value is getting closer The value is measured until it meets the condition of being small enough. Then, the direction angle and its inverse are obtained from the corrected cosine function and the corrected cosine function, which is the main process of the algorithm.

[0036] The inversion results obtained by projecting onto the direction of satellite motion are as follows Fig.10 The Euclidean distance between the predicted image and the measured image is shown as The number of iterations of correction changes as Therefore, the results of 11 iterations are taken to achieve the best iterative approximation. Fig.11 Shows The best prediction result obtained in the best iteration, Fig.12 is the ocean current vector under the best iteration, Fig.13 is the azimuth at the best iteration, which is the result of the last iteration, which means that their product is good enough to match the measured real-time value very well. The flow angle is derived, and thus the latitude and longitude components are calculated, Fig.14 is the longitude component of the stream angle, Fig.15 is the latitudinal component of the stream angle. As a validation, the correlation coefficient between the predicted and measured values ​​was calculated to be 0.9.

[0037] Although the iterative algorithm can effectively invert the sea surface flow direction, there are still some problems to be overcome. One of the important challenges is to improve the algorithm and make its design more reasonable, because it can be found from the inverted image that there are still some areas with outliers. These abnormal areas are distributed in the form of stripes, and they are usually areas with zero values. The existence of zero points makes it difficult for the algorithm to make effective corrections near the anomalies during the iteration process, making the inversion results unreliable with respect to the anomalies and their neighborhoods on the image. Another core challenge is to evaluate the impact of the presence of sea surface wind on the reversal of airflow. This is because the Ekman flow in the wind direction will complicate the change of flow direction on the sea surface. Even in the experimental area, the change of flow direction caused by the presence of Ekman flow cannot be ignored. The third important issue is that since the sea surface is the interface where the atmosphere and the ocean interact, many small-scale physical processes will also lead to an imbalance in the flow distribution. The drastic change in flow direction can easily lead to zero values ​​during the projection calculation, affecting the effectiveness of the algorithm. Figure 4 The distribution characteristics of the mesoscale vortex can be clearly seen in the figure. The 360° change in flow direction caused by the mesoscale vortex is shown in Fig.11 Two abnormal bands appeared on the inversion image, which affected the inversion results.

[0038] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for ocean current inversion combining SAR Doppler information and deep learning, characterized in that: The YOLO network is used to correct radio frequency interference, and the components of the Doppler center frequency shift along the range and azimuth directions are calculated respectively to obtain the Doppler frequency shift and invert the radial velocity of the ocean current field.

2. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 1, characterized in that: The YOLO network is based on the YOLOv8 architecture and includes a backbone network, a neck network, and a head network.

3. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 2, characterized in that: The backbone network sequentially includes an initial convolution layer and four sparse convolution layers, the four sparse convolution layers are connected to a spatial pyramid pooling layer, and each sparse convolution layer sequentially includes a submanifold sparse convolution layer and a convolution layer.

4. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 3, characterized in that: The neck network includes extracting the output features of the second and third sparse convolutional layers and the output features of the spatial pyramid pooling layer, fusing the three output features, and outputting the three fused features.

5. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 4, characterized in that: The neck network uses the FPN feature fusion module to fuse the three output features.

6. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 5, characterized in that: The head network imports the fused features into three convolutional layers for decoding.

7. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 1, characterized in that: Using the YOLO network for RFI correction includes: S1. Perform two fast Fourier transforms on the radar echo data, use the transformed radar data RDS to mark the interference object, and make a data set to input into the YOLO network for training; S2, re-collect radar echo data and input it into the trained YOLO network, output the time when RF interference exists and RDS, and construct the tensor to be processed; S3, inputting the tensor to be processed into the radio frequency interference subspace to obtain the projection component of the tensor to be processed in the radio frequency interference subspace; S4. Subtract the projection component from the tensor to be processed to obtain a signal after radio frequency interference correction.

8. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 1, characterized in that: Calculating the components of the Doppler center frequency shift along the range direction includes: ； In the formula, is the component of the Doppler center frequency shift along the range direction, is the initial transmission frequency, is the equivalent mean tilt viewing angle, is the initial distance between the radar and the target, is the real-time distance between the radar and the target, is the radar altitude.

9. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 8, characterized in that: Calculation of the Doppler center frequency shift along the azimuth direction includes: ； In the formula, is the component of the Doppler center frequency shift along the azimuth direction, is the wavelength, For time, is the frequency constant term, is the relative speed.

10. The ocean current inversion method combining SAR Doppler information and deep learning according to claim 9, characterized in that: Get Doppler shift include: ； In the formula, is the Doppler frequency shift component caused by the relative motion between the satellite and the ground, is the Doppler frequency shift component caused by the motion of the sea surface itself, and Combined to obtain.

Images