A sea surface wave height inversion method based on deep learning

Abstract

This invention provides a deep learning-based method for retrieving sea surface wave height, comprising: establishing a radar image dataset for a specific radar and the corresponding navigational marine environment; dividing the radar image dataset into a training set, a test set, and a validation set; training the training set of the radar image dataset using a deep learning method, wherein the deep learning model adopts an encoder-decoder architecture, with the radar image as input and the wave surface height corresponding to each pixel of the radar image as output; testing the trained model on the validation set, selecting the best-performing model, and evaluating the effectiveness of the radar image wave surface inversion method on the test set. This invention fully utilizes the nonlinear fitting capability of deep learning methods, which can effectively improve the observation accuracy of sea surface waves and has significant application value.

Description

Technical Field

[0001] This invention relates to the field of marine remote sensing technology, and more specifically, to a deep learning-based method for retrieving sea surface wave height. Background Technology

[0002] Ocean wave measurement is of great significance in marine engineering. Accurate marine environmental perception can assist in ship navigation, offshore construction operations, and marine disaster forecasting. Traditional wave measurement methods mainly rely on single-point contact methods such as buoys. With the development of technology, regional wave measurement methods, such as radar, lidar, and binocular vision, have emerged. Radar-based wave height measurement can measure the wave height over a large area, and since the radar equipment used is the ship's own navigation and observation radar, it has considerable application value.

[0003] Electromagnetic waves emitted by radar towards the sea surface undergo Bragg scattering with capillary waves on the surface, generating backscattered signals that form a "sea clutter" image on the radar. Wave height inversion is achieved by processing this signal. Currently, the main method for inverting wave height from incoherent radar images is through three-dimensional Fourier transform. This method first performs a Fourier transform on the radar image to obtain the image spectrum; then, it filters the image based on the wave dispersion relationship; next, a modulation transfer function is introduced to convert the radar image spectrum into a wave spectrum. Since the modulation process in radar imaging is complex and lacks a precise theoretical formula, empirical formulas are generally used for the modulation function. Finally, an inverse Fourier transform is performed to obtain the wave height distribution image. Because the radar image only provides grayscale information and cannot directly correspond to the wave height on the sea surface, a meaningful wave height needs to be calculated to calibrate the wave height distribution image to the true height. In this method, the dispersion filtering is based on certain wave model assumptions, the modulation transfer function uses empirical formulas, and the calibration process for the meaningful wave height is relatively cumbersome, thus having certain limitations. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a deep learning-based sea surface wave height inversion method that can effectively improve the accuracy of sea surface wave observation.

[0005] To solve the above problems, the technical solution of the present invention is as follows:

[0006] A deep learning-based method for retrieving sea surface wave height includes the following steps:

[0007] Establish radar image datasets for specific radars and corresponding marine navigation environments;

[0008] The radar image dataset is divided into a training set, a test set, and a validation set. The training set in the radar image dataset is trained using a deep learning method. The deep learning model adopts an encoder-decoder architecture, with the radar image as the input and the wavefront height corresponding to each pixel of the radar image as the output.

[0009] The trained models are tested on a validation set to select the best performing model, and the effectiveness of the radar image wavefront inversion method is evaluated on the test set.

[0010] Preferably, the step of establishing a radar image dataset for a specific radar and the corresponding marine navigation environment specifically includes:

[0011] Determine the radar parameters and the marine environmental statistics for ship navigation;

[0012] By sampling using established marine environmental statistical parameters, wave surface heights were obtained through numerical simulation using a wave model.

[0013] The wavefront height is trimmed to conform to the shape requirements of the radar image;

[0014] The cropped wavefront height is subjected to shadow modulation and tilt modulation to obtain a numerically simulated radar image;

[0015] The radar images are normalized to obtain a numerically simulated radar image dataset.

[0016] Preferably, in the step of sampling using determined marine environmental statistical parameters and obtaining the numerically simulated wave surface height using a wave model, the wave model is a linear wave model:

[0017]

[0018] Where, ω i k represents the frequency of a sine wave over time. i The wave number θ represents a sine wave. j Represents wave direction, ε ij Represents a random initial phase.

[0019] Preferably, in the step of sampling using determined marine environmental statistical parameters and obtaining the numerically simulated wave surface height using a wave model, the wave model is a higher-order spectral method model, and the kinematic and dynamic boundary conditions of the free surface are:

[0020]

[0021]

[0022] Where η is the wave height, φ SLet be the velocity potential of the free surface.

[0023] Preferably, in the step of performing shadow modulation and tilt modulation on the cropped wavefront height to obtain a numerically simulated radar image, the formula for shadow modulation is:

[0024] The grazing angle is: α(r,θ)=arctan(H-ηr(r,θ)), where H is the radar altitude and η(r,θ) is the wave height at the polar coordinate (r,θ).

[0025] Preferably, in the step of performing shadow modulation and tilt modulation on the cropped wavefront height to obtain a numerically simulated radar image, the tilt modulation process is as follows:

[0026]

[0027] The formula for tilt modulation is: The unit normal vector of the sea surface element is: The unit vector of the incident direction is: In the formula, r0 and H are the radial and vertical coordinates of the radar antenna.

[0028] Preferably, the step of establishing a radar image dataset for a specific radar and the corresponding marine navigation environment specifically includes:

[0029] Determine the radar parameters and the marine environmental statistics for ship navigation;

[0030] By sampling using established marine environmental statistical parameters, wave surface heights were obtained through numerical simulation using a wave model.

[0031] The wavefront height is trimmed to conform to the shape requirements of the radar image;

[0032] The cropped wavefront height is subjected to shadow modulation and tilt modulation to obtain a numerically simulated radar image;

[0033] The radar images are normalized to obtain a numerically simulated radar image dataset, which is then combined with radar images and buoy wave height datasets obtained from field measurements to obtain the final dataset.

[0034] Preferably, the step of establishing a radar image dataset for a specific radar and the corresponding marine navigation environment specifically includes:

[0035] Determine the radar parameters and the marine environmental statistics for ship navigation;

[0036] By sampling using established marine environmental statistical parameters, wave surface heights were obtained through numerical simulation using a wave model.

[0037] The wavefront height is trimmed to conform to the shape requirements of the radar image;

[0038] The cropped wavefront height is subjected to shadow modulation and tilt modulation to obtain a numerically simulated radar image;

[0039] The radar images are normalized to obtain a numerically simulated radar image dataset. Measured radar image data is obtained, and wave height is inverted using a three-dimensional Fourier transform algorithm. Regions with high inversion accuracy are selected, and wavefront height samples from these regions are added to the dataset to form a measured dataset. The normalized numerically simulated radar image dataset and the measured dataset are combined to obtain the final dataset.

[0040] Preferably, in the step of dividing the radar image dataset into a training set, a test set, and a validation set, and training the training set of the radar image dataset using a deep learning method, wherein the deep learning model adopts an encoder-decoder architecture, with the input being the radar image and the output being the wavefront height corresponding to each pixel of the radar image, the encoder-decoder structure in the deep learning model adopts a U-NET, a U-shaped convolutional neural network model, which includes a convolutional module, a pooling module, and a non-linear activation layer, and the training loss function is:

[0041] Where, σ simulation η represents 1 if the sample is a numerical simulation sample, and 0 otherwise; S is the number of wave height labels in the numerical simulation sample; η represents the number of wave height labels in the numerical simulation sample. i and These represent the predicted wave height and the actual wave height, respectively; σ experiment This indicates that if the sample is a measured sample, it is 1, otherwise it is 0; P is the number of wave height labels in the measured sample; λ is the specified weight coefficient.

[0042] Preferably, in the step of testing the trained model on a validation set, selecting the best-performing model, and evaluating the effectiveness of the radar image wavefront inversion method on the test set, the evaluation metrics are the correlation coefficient between the reconstructed wave height and the actual wave height, and the dimensionless mean square error, as shown in the formula:

[0043]

[0044]

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. This invention is based on deep learning methods. The model is an end-to-end model. The input is a radar image and the output is the inverted wavefront height. There are no intermediate processes. Compared with classical methods that contain multiple theoretical model assumptions and empirical formulas, the steps are simple and avoid errors in intermediate processes.

[0047] 2. This invention fully utilizes the nonlinear fitting capability of deep learning to better fit the nonlinear modulation effect of radar signals, thereby enabling more accurate wavefront height inversion. This helps to obtain accurate wavefront height measurement results, and more accurate marine environmental perception has significant application value for ship navigation, offshore construction operations, and marine disaster forecasting. Attached Figure Description

[0048] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0049] Figure 1 A flowchart illustrating the deep learning-based sea surface wave height inversion method provided in this embodiment of the invention;

[0050] Figure 2 This is a diagram illustrating the deep learning model algorithm framework provided in an embodiment of the present invention.

[0051] Figure 3a , 3b Comparison of the effects of the deep learning-based method for retrieving sea surface wave height from marine radar images provided in this embodiment of the invention. Detailed Implementation

[0052] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention and will make the technical solution and beneficial effects of the present invention readily apparent.

[0053] Example 1

[0054] Figure 1 A flowchart of a deep learning-based sea surface wave height inversion method provided in an embodiment of the present invention is shown below. Figure 1 As shown, the method includes the following steps:

[0055] S1: Establish radar image datasets for specific radars and corresponding navigation marine environments;

[0056] Specifically, in step S1, obtaining a specific radar and its corresponding navigation marine environment parameters is to establish a corresponding dataset for a specific radar and its environment, wherein each sample contains a radar image and the corresponding wavefront height label.

[0057] Step S1 includes the following steps:

[0058] Step 11: Determine the radar parameters and the marine environmental statistics for ship navigation;

[0059] The parameters for determining the radar include the measurement range, blind zone range, and radar altitude. The marine environmental statistical parameters for ship navigation include the wave direction spectrum, the wave's meaningful wave height-spectral peak period probability distribution map, and the wave spectrum.

[0060] Step 12: Sample using the determined marine environmental statistical parameters and obtain the wave surface height of the numerical simulation using a wave model;

[0061] Using established marine environmental statistical parameters, a large number of wave height, wave direction, spectral peak periods, and random phase combinations are sampled. Wave height is then numerically simulated using a wave model. The wave model used is a linear wave model.

[0062]

[0063] Where, ω i k represents the frequency of a sine wave over time. i The wave number θ represents a sine wave. j Represents wave direction, ε ij Represents a random initial phase.

[0064] Step 13: Trim the wavefront height to make it conform to the shape requirements of the radar image;

[0065] Step 14: Apply shadow modulation and tilt modulation to the cropped wavefront height to obtain a numerically simulated radar image;

[0066] Specifically, the formula for shadow modulation is:

[0067]

[0068] The glancing angle is: In the formula, H is the radar altitude, and η(r,θ) is the wave height at the polar coordinate (r,θ).

[0069] The tilt modulation process is as follows: The formula for tilt modulation is: The unit normal vector of the sea surface element is:

[0070]

[0071] The unit vector of the incident direction is:

[0072]

[0073] In the formula, r0 and H are the radial and vertical coordinates of the radar antenna.

[0074] Step 15: Normalize the radar images to obtain a numerical simulation radar image dataset.

[0075] The radar image is normalized so that its value is within the range of [0, 255] to obtain a numerically simulated radar image dataset.

[0076] S2: Divide the radar image dataset into a training set, a test set, and a validation set. Train the training set of the radar image dataset using a deep learning method. The deep learning model adopts an encoder-decoder architecture. The input is the radar image, and the output is the wavefront height corresponding to each pixel of the radar image.

[0077] Specifically, the radar image dataset is divided into a training set, a test set, and a validation set, with a ratio of 8:1:1.

[0078] like Figure 2 As shown, the deep learning model adopts an encoder-decoder architecture, and the encoder-decoder structure adopts U-NET, a U-shaped convolutional neural network model. This model consists of a convolution module, a pooling module, and a non-linear activation layer. The input is a radar image, and the output is the wavefront height corresponding to each pixel in the radar image.

[0079] The training loss function is:

[0080]

[0081] Where, σ simulation η represents 1 if the sample is a numerical simulation sample, and 0 otherwise; S is the number of wave height labels in the numerical simulation sample; η represents the number of wave height labels in the numerical simulation sample. i and These represent the predicted wave height and the actual wave height, respectively; σ experiment This indicates that if the sample is a measured sample, it is 1, otherwise it is 0; P is the number of wave height labels in the measured sample; λ is the specified weight coefficient.

[0082] S3: Test the trained model on the validation set, select the best performing model, and evaluate the effectiveness of the radar image wavefront inversion method on the test set.

[0083] The evaluation metrics are the correlation coefficient and dimensionless mean square error between the reconstructed wave height and the actual wave height, and the specific formula is as follows:

[0084]

[0085]

[0086] This embodiment demonstrates numerical simulation data generated by the most basic linear wave model, using a technical approach without measured data. For example... Figure 3a , 3b The image shown is a distribution of the relative error of the final prediction. Figure 3a The error of the three-dimensional Fourier transform algorithm, Figure 3b The error of the algorithm of this invention is from Figure 3a and Figure 3b It can be seen that the algorithm in this invention is more accurate than the three-dimensional Fourier algorithm.

[0087] Example 2

[0088] Based on the method described in Embodiment 1, the wave model in step 12 of Embodiment 1 is replaced with a higher-order spectral method model, thereby generating simulation data that is closer to the real physical environment.

[0089] Specifically, in this embodiment, the wave model in step 12 adopts a higher-order spectral method model, and the kinematic and dynamic boundary conditions of the free surface are as follows:

[0090]

[0091]

[0092] Where η is the wave height, φ S Let be the velocity potential of the free surface.

[0093] Expanding on the velocity potential:

[0094]

[0095]

[0096] Finally, regarding φ S The evolution equations for η are:

[0097]

[0098] Example 3

[0099] Based on the method described in Embodiment 1, this embodiment considers adding measured buoy data samples. Building upon Embodiment 1, a measured dataset is added, namely, acquiring measured buoy data and corresponding radar image data. The radar images and the wave surface heights of several corresponding buoy locations are combined to form a sample, constituting the measured dataset. The numerical simulation radar image dataset from step 15 of Embodiment 1, along with the field-measured radar images and buoy wave height dataset, are combined to obtain the final dataset, which is then divided into a training set, a test set, and a validation set. Because this embodiment incorporates measured data from actual sea areas, the robustness of the model can be effectively improved.

[0100] Example 4

[0101] Building upon Example 1, this example acquires measured radar image data and uses the classic 3D Fourier transform algorithm to invert wave height. Regions with high inversion accuracy are selected, namely those facing the main wave direction and far from the radar's far-end and near-end blind zones. Wavefront height samples from these regions are added to the dataset, forming the measured dataset. The numerical simulation radar image dataset from step 15 of Example 1 is combined with the measured dataset to obtain the final dataset, which is then divided into training, testing, and validation sets. This example considers the scenario where radar images lack actual wave height labels, effectively overcoming application challenges by using regions with high accuracy from the classic method as labels.

[0102] In summary, the beneficial effects of this invention are:

[0103] 1. This invention is based on deep learning methods. The model is an end-to-end model. The input is a radar image and the output is the inverted wavefront height. There are no intermediate processes. Compared with classical methods that contain multiple theoretical model assumptions and empirical formulas, the steps are simple and avoid errors in intermediate processes.

[0104] 2. This invention fully utilizes the nonlinear fitting capability of deep learning to better fit the nonlinear modulation effect of radar signals, thereby enabling more accurate wavefront height inversion. This helps to obtain accurate wavefront height measurement results, and more accurate marine environmental perception has significant application value for ship navigation, offshore construction operations, and marine disaster forecasting.

[0105] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0106] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for retrieving sea surface wave height based on deep learning, characterized in that, The method includes the following steps: Establish radar image datasets for specific radars and corresponding marine navigation environments; The radar image dataset is divided into a training set, a test set, and a validation set. The training set is trained using deep learning. The deep learning model employs an encoder-decoder architecture, with the radar image as input and the wavefront height corresponding to each pixel in the radar image as output. The encoder-decoder structure in the deep learning model uses a U-NET, a U-shaped convolutional neural network model, which includes convolutional modules, pooling modules, and non-linear activation layers. The training loss function is: ,in, This indicates that if the sample is a numerical simulation sample, the value is 1; otherwise, it is 0. This represents the number of wave height labels in the numerical simulation sample; and These are the predicted wave height and the actual wave height, respectively. This indicates that if the sample is a measured sample, it is 1; otherwise, it is 0. This represents the number of wave height labels in the measured sample; The specified weighting coefficient; The trained models are tested on a validation set to select the best performing model, and the effectiveness of the radar image wavefront inversion method is evaluated on the test set.

2. The deep learning-based sea surface wave height inversion method according to claim 1, characterized in that, The steps for establishing a radar image dataset for a specific radar and the corresponding marine navigation environment specifically include: Determine the radar parameters and the marine environmental statistics for ship navigation; By sampling using established marine environmental statistical parameters, wave surface heights were obtained through numerical simulation using a wave model. The wavefront height is trimmed to conform to the shape requirements of the radar image; The cropped wavefront height is subjected to shadow modulation and tilt modulation to obtain a numerically simulated radar image; The radar images are normalized to obtain a numerically simulated radar image dataset.

3. The deep learning based sea surface significant wave height retrieval method of claim 2, wherein, In the step of sampling using determined marine environmental statistical parameters and obtaining the numerically simulated wave surface height using a wave model, the wave model used is a linear wave model: in, This represents the frequency of a sine wave over time. The wave number represents a sine wave. Represents wave direction, Represents a random initial phase.

4. The deep learning based sea surface significant wave height retrieval method of claim 2, wherein, In the step of sampling using determined marine environmental statistical parameters and obtaining the numerically simulated wave surface height using a wave model, the wave model is a higher-order spectral method model, and the kinematic and dynamic boundary conditions of the free surface are: wherein, is the wave height, is the free surface velocity potential.

5. The deep learning based sea surface significant wave height retrieval method of claim 2, wherein, In the step of performing shadow modulation and tilt modulation on the clipped wavefront height to obtain a numerically simulated radar image, the formula for shadow modulation is: Wherein, the grazing angle is: In the formula, The altitude of the radar. polar coordinates Wave height at that location.

6. The deep learning based sea surface significant wave height retrieval method of claim 1, wherein, The steps for establishing a radar image dataset for a specific radar and the corresponding marine navigation environment specifically include: Determine the radar parameters and the marine environmental statistics for ship navigation; By sampling using established marine environmental statistical parameters, wave surface heights were obtained through numerical simulation using a wave model. The wavefront height is trimmed to conform to the shape requirements of the radar image; The cropped wavefront height is subjected to shadow modulation and tilt modulation to obtain a numerically simulated radar image; The radar images are normalized to obtain a numerically simulated radar image dataset, which is then combined with radar images and buoy wave height datasets obtained from field measurements to obtain the final dataset.

7. The deep learning-based sea surface significant wave height retrieval method of claim 1, wherein, The steps for establishing a radar image dataset for a specific radar and the corresponding marine navigation environment specifically include: Determine the radar parameters and the marine environmental statistics for ship navigation; By sampling using established marine environmental statistical parameters, wave surface heights were obtained through numerical simulation using a wave model. The wavefront height is trimmed to conform to the shape requirements of the radar image; The cropped wavefront height is subjected to shadow modulation and tilt modulation to obtain a numerically simulated radar image; The radar images are normalized to obtain a numerically simulated radar image dataset. Measured radar image data is obtained, and wave height is inverted using a three-dimensional Fourier transform algorithm. Regions with high inversion accuracy are selected, and wavefront height samples from these regions are added to the dataset to form a measured dataset. The normalized numerically simulated radar image dataset and the measured dataset are combined to obtain the final dataset.

8. The deep learning-based sea surface wave height inversion method according to claim 1, characterized in that, In the step of testing the trained model on a validation set, selecting the best-performing model, and evaluating the effectiveness of the radar image wavefront inversion method on the test set, the evaluation metrics are the correlation coefficient between the reconstructed wave height and the actual wave height, and the dimensionless mean square error, as shown in the formula: 。

Images