A method for wind and wave response modeling and ocean current inversion based on machine learning

Abstract

A method for wind and wave response modeling and ocean current inversion based on machine learning, relating to the fields of artificial intelligence and marine information services, includes the following steps: (1) data acquisition; (2) data processing; (3) constructing a dataset; (4) establishing a sea surface current velocity inversion model using a particle swarm optimization machine learning algorithm; and (5) obtaining the optimal sea surface current velocity inversion model through testing and case analysis. Targeting the oceanographic and Doppler frequency shift characteristics of the sea surface, this invention proposes for the first time a particle swarm optimization XGBoost model for predicting wind and wave Doppler contributions, which can effectively improve the accuracy and speed of wind and wave Doppler contribution estimation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and marine information service technology, specifically to a method for wind and wave response modeling and ocean current inversion based on machine learning. Background Technology

[0002] Ocean currents are a core element of the ocean dynamic system, playing a decisive role in maintaining the Earth's energy balance and climate stability by transporting heat and salinity on a global scale. Accurate ocean current observations are not only crucial for a deeper understanding of the coupling processes between the ocean and the atmosphere, but also significantly improve the accuracy of oil spill dispersion prediction, optimize shipping routes, and provide key support for fisheries resource management and maritime search and rescue operations. Ocean current measurements can be obtained through buoys, shore-based or shipborne radar, but these methods have high maintenance costs, limited spatial coverage, and difficulty in achieving continuous monitoring on a global scale. Satellite remote sensing, due to its ability to conduct large-scale repeated observations, has become an important means of obtaining global sea surface current field information. Spaceborne synthetic aperture radar (SAR) is currently the only satellite remote sensing method that can directly measure sea surface currents.

[0003] There are generally two methods for ocean current inversion using spaceborne SAR: Along Track Interferometry (ATI) and Doppler Centroid Anomaly (DCA). The ATI method mainly uses two antennas with short intervals along the satellite's orbit to acquire two SAR images of the same area, and then uses the phase difference between the two images to invert the sea surface velocity in the radar's line of sight, but its data is difficult to obtain. The DCA method inverts ocean currents mainly by correcting the Doppler contributions unrelated to ocean currents in the SAR Doppler frequency shift, thereby obtaining the Doppler frequency shift of the ocean current, and finally inverting the ocean current. Therefore, the accurate removal of the Doppler frequency shift caused by non-ocean currents is the key to the accurate ocean current inversion of the DCA method.

[0004] The Doppler effect refers to the phenomenon where, when there is relative motion between a wave source and an observer, the frequency of the wave received by the observer differs from the frequency actually emitted by the wave source. Spaceborne SAR can transmit electromagnetic waves of a specific frequency to the sea surface and receive their echo signals. By comparing the phase difference between the transmitted and received electromagnetic waves, the Doppler frequency shift caused by the relative motion between the sea surface and the satellite can be calculated. However, this frequency shift does not solely rely on the sea surface motion; the satellite's own motion and the Earth's rotation also affect the observed Doppler frequency shift during the satellite's observation of the sea surface. Considering that the Doppler frequency shift of OCN data products is obtained by processing SLC data, the Doppler frequency shift it provides... It typically includes the following contributions:

[0005] (01);

[0006] (02);

[0007] (03);

[0008] in, and These are the geophysical contribution and the non-geophysical contribution, which together constitute the total Doppler shift of SAR observations. ; It is the Doppler frequency shift caused by sea surface movement. The Doppler shift is caused by changes in satellite attitude, orbit, and antenna orientation, as well as the relative motion between the satellite and the Earth. Mainly composed of and Composition, in which The Doppler frequency shift caused by sea surface currents The Doppler shift is caused by wind and waves; Depend on , , and Composition, in which It is the Doppler frequency shift caused by changes in satellite orbital altitude and satellite attitude. The Doppler frequency shift is caused by the misalignment of the electromagnetic waves emitted by the SAR antenna from the center of the sea surface element. The Doppler frequency shift is caused by the sector error resulting from the movement of the antenna in the TOPSAR scanning mode of the SAR satellite. The Doppler shift caused by other unknown deviations.

[0009] To invert ocean currents using SAR Doppler shift, it is necessary to analyze the total Doppler shift observed by SAR. Corrections were made to remove Doppler contributions from ocean currents. Irrelevant Doppler contributions; in non-geophysical terms middle, The data has already been calculated and stored in the product during SAR satellite observations; It is periodic noise, which can be removed using frequency domain filtering. and Land cover data from the same observation orbit can be used for estimation; A wind-wave Doppler contribution estimation model needs to be established to eliminate it. Although the CDOP series of models based on BPNN perform well in estimating wind-wave Doppler frequency shift, BPNN relies on gradient descent iterative optimization, which has high computational complexity. It requires fine adjustment of the number of layers, neurons, and learning rate, is prone to getting trapped in local optima, and cannot automatically handle missing values. On the other hand, decision tree-based ensemble models, such as XGBoost, support parallel computing, and their training speed is usually 5-10 times faster than BPNN. Their parameters (such as tree depth and learning rate) are more robust, and they have built-in missing value handling and feature binning. Therefore, it is necessary to explore the application of invented and efficient machine learning models such as XGBoost in wind-wave Doppler frequency shift contribution modeling. Summary of the Invention

[0010] This invention provides a method for wind and wave response modeling and ocean current inversion based on machine learning. The purpose is to address the problems existing in the prior art by providing a new method that can improve the accuracy and speed of wind and wave Doppler contribution estimation.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows:

[0012] A method for wind and wave response modeling and ocean current inversion based on machine learning includes the following steps: (1) data acquisition; (2) data processing; (3) constructing a dataset; (4) establishing a sea surface current velocity inversion model using a particle swarm optimization machine learning algorithm; and (5) obtaining the optimal sea surface current velocity inversion model through testing and case analysis.

[0013] Preferably, step (1) includes: observing SAR Doppler frequency shift data through Sentinel-1 satellite; acquiring ocean wave data through WAVEYS; acquiring ocean current data through HYCOM; and acquiring ocean current data through HFRNet.

[0014] Preferably, step (2) includes: (21) non-geophysical Doppler frequency shift correction; and (22) spatiotemporal matching of the data.

[0015] Preferably, step (21) includes: satellite attitude Doppler contribution correction; sector error Doppler contribution correction; satellite antenna mispointing Doppler contribution and residual Doppler contribution correction; step (22) includes: combining the data corrected in step (21) with wave data, ocean current data and ECMWF wind field data for spatiotemporal matching, wherein the wind field data refers to wind direction and wind speed.

[0016] Preferably, in step (3), the spatiotemporal matching data is converted by coordinate transformation and ocean current Doppler conversion to construct a multi-parameter dataset of the sea surface.

[0017] Preferably, in step (4), the SAR incident angle, SAR radar ground distance wind speed and direction, and ground distance wave speed and direction are used as input parameters to establish the relationship with the wind-wave Doppler frequency shift.

[0018] Preferably, in step (4), the incident angle is selected. Ground distance wind speed ,wind direction Earth distance and directional wave orbital velocity , wave direction As a model input parameter, the wind-wave induced Doppler frequency shift is expressed as:

[0019] (1);

[0020] Converting the ECMWF wind direction to the angle relative to the SAR ground distance is expressed as:

[0021] (2);

[0022] In the formula, Indicates the azimuth angle of the SAR satellite orbit. This indicates the wind direction relative to true north; after transformation, 0° and 180° represent the windward (wind blowing towards the radar) and the tailwind (wind blowing away from the radar), respectively, and 90° and 270° represent the azimuth wind (wind along the satellite orbit); SAR ground-to-ground wind speed is expressed as:

[0023] (3);

[0024] In the formula, u represents the ECMWF wind speed in the matched dataset; the mean wave direction is transformed to the direction relative to the SAR ground-to-ground direction, as follows:

[0025] (4);

[0026] In the formula, The average wave direction is indicated relative to true north; the wave trajectory velocity is calculated using WAVEYS significant wave height and average wave period.

[0027] (5);

[0028] In the formula, Indicates the significant wave height. Indicates the average angular frequency. The average wave period is represented by: The SAR range-direction wave trajectory velocity is represented as:

[0029] (6);

[0030] Then, formula (1) is modeled based on two machine learning algorithms, BPNN and XGBoost.

[0031] Preferably, in step (4), the key parameters in the BPNN and XGBoost models are optimized by using the Particle Swarm Optimization (PSO) algorithm: for the BPNN model, the optimized parameters include the learning rate, the number of iterations, and the number of batch samples; for the XGBoost model, the optimized parameters include the number of base models, the learning rate, the maximum tree depth, and the minimum leaf weight. The optimized BPNN model and the optimized XGBoost model are obtained through optimization.

[0032] Preferably, in step (5), the optimized XGBoost model obtained through testing and case analysis is the optimal sea surface current velocity inversion model.

[0033] An apparatus for wind and wave response modeling and ocean current inversion based on machine learning includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement a method for wind and wave response modeling and ocean current inversion based on machine learning when the computer program is executed.

[0034] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for wind and wave response modeling and ocean current inversion based on machine learning.

[0035] The method for wind and wave response modeling and ocean current inversion based on machine learning in this invention has the following beneficial effects:

[0036] Existing BPNN-based CDOP models often require large amounts of data and are time-consuming due to their inherent characteristics, resulting in slow convergence and poor current retrieval performance, particularly exhibiting significant errors in the core current region. This invention constructs a novel machine learning-based current retrieval algorithm using SAR Doppler shift and oceanographic data from satellite remote sensing. Its specific benefits are as follows:

[0037] (1) In view of the oceanographic and Doppler frequency shift characteristics of the sea surface, this invention proposes for the first time a particle swarm optimization XGBoost model wind-wave Doppler contribution prediction method; For the Doppler data of global observation by Sentinel-1 satellite, there are multiple Doppler contributions, and accurate estimation of wind-wave Doppler contributions is the key to ocean current inversion. The existing CDOP model based on BPNN has low prediction accuracy and slow convergence speed; How to estimate wind-wave Doppler contributions more accurately and quickly is a difficult problem in ocean current inversion; This invention proposes a new method that can improve the accuracy and speed of wind-wave Doppler contribution estimation.

[0038] (2) The new method proposed in this invention will provide new ideas and scientific understanding for SAR ocean current inversion and its assessment of primary marine productivity. Attached Figure Description

[0039] Figure 1 This is a distribution map of Sentinel-1 SAR data in the research area of ​​the algorithm of this invention.

[0040] Figure 2 The chart shows the trend of loss value as a function of iteration number during the training of the two models.

[0041] Figure 3 In the middle: (a) is a scatter plot of the wind and wave Doppler contribution estimation results of BPNN with particle swarm optimization; (b) is a scatter plot of the wind and wave Doppler contribution estimation results of XGBoost with particle swarm optimization.

[0042] Figure 4 A scatter plot comparing the Doppler shift of wind waves predicted by BPNN with SAR observations for different wind speed ranges.

[0043] Figure 5 A scatter plot comparing the Doppler shift of wind waves predicted by XGBoost with SAR observations for different wind speed ranges.

[0044] Figure 6 In the example: (a) is a scatter plot comparing the ocean current velocity retrieved by BPNN with that of HYCOM in a specific case; (b) is a scatter plot comparing the ocean current velocity retrieved by XGBoost with that of HYCOM in a specific case.

[0045] Figure 7 This is a feature map of the HYCOM current.

[0046] Figure 8 This is a BPNN-inverted ocean current feature map.

[0047] Figure 9 This is a feature map of ocean currents retrieved by XGBoost.

[0048] Figure 10 Scatter plot comparing the inverted ocean current velocity with the HF ocean current velocity (left: BPNN, right: XGBoost).

[0049] Figure 11 This is a technical roadmap for a method for wind and wave response modeling and ocean current inversion based on machine learning proposed in this invention. Detailed Implementation

[0050] The following description provides a detailed explanation of the embodiments of the present invention in a step-by-step manner. This description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0051] In the description of this invention, it should be noted that the terms "upper," "lower," "left," "right," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or a specific orientational structure and operation. Therefore, they should not be construed as limiting this invention.

[0052] A machine learning-based method for wind and wave response modeling and ocean current inversion: First, ocean currents are inverted using SAR Doppler frequency shift, which requires analyzing the total Doppler frequency shift observed by SAR. Corrections were made to remove Doppler contributions from ocean currents. Irrelevant Doppler contributions; in non-geophysical terms middle, The data has already been calculated and stored in the product during SAR satellite observations; It is periodic noise, which can be removed using frequency domain filtering. and Estimation can be performed using land cover data from the same observation orbit; based on land Assuming the average Doppler shift is 0 Hz, the non-geophysical terms are approximated as:

[0053] (04);

[0054] From the total Doppler frequency shift Remove non-geophysical components The steps are as follows: (1) Select an observation point with 100% land coverage. At this time, the total Doppler frequency shift of the SAR observation is... (2) Using the Doppler frequency shift of the land observation area minus get (3) Use frequency domain filtering to remove periodic sector error. (4) The residual Doppler frequency shift of the land area is linearly fitted using the least squares method and modeled as a function of the incident angle; (5) The fitted model is then calibrated. .

[0055] This study performs spatiotemporal matching of Sentinel-1A OCN (Doppler shift, ECMWF wind field), WAVEYS wave data (significant wave height, mean wave direction, and mean wave period), and HYCOM current data (surface meridional velocity and surface zonal velocity). First, the ECMWF wind field data (wind speed and direction) provided by OCN is spatially interpolated onto the latitude and longitude grid corresponding to the Doppler shift data. Then, the OCN Doppler shift and wind field data are spatiotemporally matched with the WAVEYS wave data and HYCOM current data. Specifically, using OCN time (UTC) as the reference time, the wave data is... The time units for both RYS wave data and HYCOM current data were unified to OCN time units (accumulated hours starting from 00:00:00 on January 1, 2000). Time matching was performed on WAVERYS and HYCOM data within a ±30 minute time window. Spatial matching was performed using bilinear interpolation, interpolating all data to a Doppler frequency shift grid. In addition, the above matched datasets were spatiotemporally matched with high-frequency radar observation data. Using HF time and latitude and longitude as the reference, a ±30 minute time window and bilinear interpolation were selected. Finally, a total of 3404 pairs of samples were successfully matched.

[0056] There is a complex nonlinear coupling relationship between Doppler frequency shift and sea surface wind and wave fields. Among existing artificial intelligence methods, deep learning models represented by Back Propagation Neural Networks (BPNN) and ensemble learning models represented by extreme gradient boosting (XGBoost) have been widely used due to their excellent performance. This invention selects these two representative intelligent algorithms to establish a relationship model between wind-wave induced Doppler frequency shift and sea surface wind and wave parameters, and evaluates their applicability by comparison to determine a better algorithm scheme.

[0057] Select the angle of incidence Ground distance wind speed ,wind direction Earth distance and directional wave orbital velocity , wave direction As a model input parameter, the wind-wave induced Doppler frequency shift is expressed as:

[0058] (1);

[0059] Converting the ECMWF wind direction to the angle relative to the SAR ground distance is expressed as:

[0060] (2);

[0061] In the formula, Indicates the azimuth angle of the SAR satellite orbit. This indicates the wind direction relative to true north; after transformation, 0° and 180° represent the windward (wind blowing towards the radar) and the tailwind (wind blowing away from the radar), respectively, and 90° and 270° represent the azimuth wind (wind along the satellite orbit); SAR ground-to-ground wind speed is expressed as:

[0062] (3);

[0063] In the formula, u represents the ECMWF wind speed in the matched dataset; the mean wave direction is transformed to the direction relative to the SAR ground-to-ground direction, as follows:

[0064] (4);

[0065] In the formula, The average wave direction is indicated relative to true north; the wave trajectory velocity is calculated using WAVEYS significant wave height and average wave period.

[0066] (5);

[0067] In the formula, Indicates the significant wave height. Indicates the average angular frequency. The average wave period is represented by: The SAR range-direction wave trajectory velocity is represented as:

[0068] (6);

[0069] Formula (1) is modeled using two machine learning algorithms, BPNN and XGBoost.

[0070] (1) BPNN:

[0071] BPNN employs a classic four-layer fully connected neural network architecture, consisting of an input layer, two hidden layers, and an output layer. The specific configuration of this network topology is as follows: The input layer contains 5 neurons, corresponding to input feature dimensions such as incident angle, wind speed, wind direction, wave speed, and wave direction; the first hidden layer has 16 neurons, responsible for primary feature extraction and nonlinear transformation; the second hidden layer has 8 neurons, used for advanced feature abstraction and dimensionality compression; the output layer uses a single neuron for regression prediction; the Adaptive Moment Estimation (Adam) is selected as the weight optimizer, combining the advantages of momentum and RMSprop to effectively adjust the learning rate; the loss function uses the root mean square error (RMSE) to quantify the deviation between the predicted and true values; each hidden layer uses the Corrected Linear Unit (ReLU) as the activation function. This nonlinear activation mechanism alleviates the gradient vanishing problem and enhances the model's ability to represent complex nonlinear relationships; BPNN calculates the predicted output through forward propagation and then iterates based on the backpropagation algorithm. Its mathematical model is expressed as:

[0072] (7);

[0073] In the formula, Let w be the weight matrix for each layer and b be the bias vector for each layer; Input parameter vector Defined as:

[0074] (8).

[0075] (2) XGBoost:

[0076] XGBoost is an ensemble learning algorithm based on the gradient boosting framework. Its core idea is to gradually optimize the model's prediction performance by iteratively adding decision trees. Compared with traditional gradient boosting methods, XGBoost introduces second-order Taylor expansion and regularization terms, which significantly improves the model's accuracy and generalization ability. This second-order approximation can not only more accurately approximate various complex loss functions (including custom loss functions), but also provide more reliable curvature information for model optimization. From the model architecture perspective, XGBoost adopts an additive training strategy, consisting of n base models (decision trees) connected in series, where the optimization process of the tree model in the t-th iteration directly depends on the cumulative results of the previous t-1 iterations.

[0077] Assume the first The tree model to be trained in the next iteration is , Indicates the first Then the model's training sample, the first training sample... The objective function is expressed as:

[0078] (9);

[0079] In the formula, For the sample size, Indicates the first The loss function term of the wheel, The regularization term for the model is as follows:

[0080] (10);

[0081] Where γ and λ are user-defined parameters; performing a second-order Taylor expansion on equation (9), we obtain:

[0082] (11);

[0083] Where g is the first derivative and h is the second derivative; as shown in the following equation: (12);

[0084] (13);

[0085] Substituting equations (10), (12), and (13) into equation (11) and taking the derivative, we obtain the following solution:

[0086] (14);

[0087] (15);

[0088] in, This represents the value of the loss function; the smaller the value, the better the tree structure. It is the weight.

[0089] The performance of BPNN and XGBoost models largely depends on the selection of hyperparameters. Although these models have preset default parameter values ​​suitable for most problems, these parameters may not be suitable for all datasets and application scenarios. Therefore, adjusting hyperparameters for specific datasets is a key step in optimizing model performance. In this invention, Particle Swarm Optimization (PSO) is used to optimize key parameters in BPNN and XGBoost models. Specifically, for the BPNN model, the optimized parameters include the learning rate, number of iterations, and batch size; for the XGBoost model, the optimized parameters include the number of base models, learning rate, maximum tree depth, and minimum leaf weight.

[0090] First, based on the characteristics of the parameters to be optimized, the dimensions of the search space and the velocities of the particles are determined. The position of each particle in the search space is a multi-dimensional vector, with each dimension corresponding to different BPNN or XGBoost parameters; therefore, the initialization range for each dimension is different. Taking XGBoost as an example, the... The particle in the first The position vector for the next iteration is represented as:

[0091] (16);

[0092] Since all particles move in the same search space, when At time 1, the particle velocity in each dimension is randomly initialized to a value within the range (0,1); The particle in the first The velocity vector for the next iteration is represented as:

[0093] (17);

[0094] Then, the position vector is assigned to the corresponding parameters of the model, and the performance on the training set is used as the fitness. This invention uses the mean squared error (RMSE) of the model's predictions as the optimization objective, so the fitness is set to the mean squared error of the model. The particle in the first The fitness of the next iteration is:

[0095] (18);

[0096] For a single particle, it is in front The historical best position in the next iteration, i.e., the position with the minimum fitness, is represented as:

[0097] (19);

[0098] All particles are in front. The historical best position, i.e., the global best position, in the next iteration is represented as:

[0099] (20).

[0100] This invention uses 80% of the data as the training set to train the model, and the remaining 20% ​​as the test set to evaluate the model's performance. The BPNN and XGBoost models are optimized on the training set respectively. The final results show that the BPNN has a learning rate of 0.01, 512 batches, and 100 iterations; the XGBoost has 152 base models, a learning rate of 0.04, a maximum tree depth of 7, and a minimum leaf weight of 10. Figure 2The training processes of the two models are presented. It can be seen that the mean squared error of BPNN decreases rapidly in the early stage of training and basically converges at the 80th iteration, with the mean squared error converging to around 7.0. The mean squared error of XGBoost decreases even faster in the early stage, and the mean squared error has converged to around 4.0 at the 40th iteration. It can be seen that XGBoost has a stronger fitting ability than BPNN.

[0101] The BPNN and XGBoost models were evaluated on the test set to estimate the wind-wave-induced Doppler shift. Performance; such as Figure 3 As shown, the prediction results of the BPNN model are relatively scattered, with a root mean square error (RMSE) of approximately 6.941 Hz and a correlation coefficient of approximately 0.903. In contrast, the XGBoost model exhibits better prediction performance, with its prediction points densely distributed near the 1:1 reference line, a lower RMSE of 4.043 Hz, and an increased correlation coefficient of 0.971. Compared to the BPNN model, the XGBoost model has a lower mean square error of 2.898 Hz, demonstrating higher prediction accuracy.

[0102] The differences in prediction performance between BPNN and XGBoost models across different wind speed ranges were further analyzed; for example... Figure 4 , Figure 5 As shown, the BPNN model has an RMSE of 9.224 Hz in the low wind speed range (0-5 m / s), which drops to 6.767 Hz in the low-to-medium wind speed range (5-10 m / s), and further decreases to 5.848 Hz in the medium-to-high wind speed range (greater than 10 m / s). In contrast, the XGBoost model shows better prediction stability in all wind speed ranges, with RMSEs of 4.835 Hz (0-5 m / s), 3.771 Hz (5-10 m / s), and 3.494 Hz (greater than 10 m / s) in the same wind speed range.

[0103] To further verify the regional effectiveness of the new algorithm proposed in this invention, the flow velocity obtained by the new algorithm based on Sentinel-1A SAR Doppler frequency shift observation data is compared with the HYCOM flow velocity. Figure 6Scatter plots comparing the ocean current velocities retrieved using BPNN and XGBoost models with the HYCOM ocean current velocities in the validation region are presented. The plots show that the XGBoost model significantly outperforms the BPNN model. Specifically, the scatter plots of the XGBoost model's retrieval results are more closely distributed around the 1:1 baseline, with an RMSE of 0.252 m / s and a MAE of 0.202 m / s, representing reductions of 25% and 26% respectively compared to the BPNN model (RMSE 0.336 m / s, MAE 0.271 m / s). The mean error (ME) decreased from 0.127 m / s in the BPNN model to 0.027 m / s in the XGBoost model (an improvement of 79%), and the correlation coefficient R also increased from 0.71 to 0.787 (an improvement of 11%).

[0104] Figures 7-9 The spatial distribution maps of the HYCOM current velocity, the BPNN model-inverted velocity, and the XGBoost model-inverted velocity in the validation area are shown. Comparative analysis reveals that the spatial distribution of the velocities inverted by the BPNN and XGBoost models is generally consistent with the spatial characteristics of the HYCOM current velocity. However, the BPNN model-inverted velocity exhibits significant overestimation in some regions (see...). Figure 8 (The area marked with a Chinese box). In contrast, the spatial characteristics of the current retrieved by the XGBoost model better overcome the overestimation problem of the BPNN model and show a higher degree of agreement with the HYCOM current in terms of ocean current details.

[0105] Figure 10 The results show further model validation in the Gulf Stream region using a dataset matched with in-situ high-frequency radar. A comparison of BPNN and XGBoost validation results on HF data is presented. Figure 10 As shown in the figure, the left figure displays the BPNN inversion results, with root mean square error, mean relative error, and bias of 0.252 m / s, 0.182 m / s, and -0.147 m / s, respectively. The right figure shows the comparison between the XGBoost current inversion results and high-frequency radar data, where the root mean square error, mean relative error, and bias are reduced to 0.213 m / s, 0.167 m / s, and -0.073 m / s, respectively. The results indicate that the XGBoost current inversion accuracy is superior to BPNN, and it can achieve high-precision current inversion.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] like Figure 11 As shown, the present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes; although preferred embodiments of the invention have been described, those skilled in the art, once they learn the basic inventive concept, can make other changes and modifications to these embodiments; therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0110] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for wind and wave response modeling and ocean current inversion based on machine learning, characterized by: Includes the following steps: (1) Data acquisition; (2) Data processing; (3) Data set construction; (4) Establishing a sea surface velocity inversion model using particle swarm optimization machine learning algorithm; (5) Obtaining the optimal sea surface velocity inversion model through testing and case analysis; Step (1) includes: observing SAR Doppler frequency shift data through Sentinel-1 satellite; acquiring ocean wave data through WAVERYS; acquiring ocean current data through HYCOM; acquiring ocean current data through HFRNet; Step (2) includes: (21) non-geophysical Doppler frequency shift correction; (22) performing spatiotemporal matching on the data; In step (4), the SAR incident angle, SAR radar ground distance wind speed and direction, and ground distance wave speed and direction are used as input parameters to establish the relationship with the wind-wave Doppler frequency shift. In step (4), the incident angle θ and the ground distance wind speed u are selected. r ,wind direction Earth distance to directional wave orbital velocity u wr , wave direction As a model input parameter, the wind-wave induced Doppler frequency shift is expressed as: Converting the ECMWF wind direction to the angle relative to the SAR ground distance is expressed as: In the formula, Indicates the azimuth angle of the SAR satellite orbit. This indicates the wind direction relative to true north; after transformation, 0° and 180° represent the wind blowing towards the radar and the wind blowing away from the radar, respectively, while 90° and 270° represent the azimuth wind, i.e., the wind along the satellite orbit; the SAR ground-to-ground wind speed is expressed as: In the formula, u represents the ECMWF wind speed in the matched dataset; the mean wave direction is transformed to the direction relative to the SAR ground-to-ground direction, as follows: In the formula, The average wave direction is indicated relative to true north; the wave trajectory velocity is calculated using WAVEYS significant wave height and average wave period. he w =ω m ·H s (5) In the formula, H s Indicates the significant wave height, ω m =1 / T m T represents the average angular frequency. m The average wave period is represented by: The SAR range-direction wave trajectory velocity is represented as: Then, formula (1) is modeled based on two machine learning algorithms, BPNN and XGBoost. In step (4), the key parameters in the BPNN and XGBoost models are optimized by using the particle swarm optimization algorithm: for the BPNN model, the optimized parameters include the learning rate, the number of iterations and the number of batch samples; for the XGBoost model, the optimized parameters include the number of base models, the learning rate, the maximum tree depth and the minimum leaf weight. The optimized BPNN model and the optimized XGBoost model are obtained through optimization.

2. The method for wind and wave response modeling and ocean current inversion based on machine learning as described in claim 1, characterized in that: Step (21) includes: satellite attitude Doppler contribution correction; sector error Doppler contribution correction; satellite antenna mispointing Doppler contribution and residual Doppler contribution correction; Step (22) includes: combining the data corrected in step (21) with wave data, ocean current data and ECMWF wind field data for spatiotemporal matching, wherein wind field data refers to wind direction and wind speed.

3. The method for wind and wave response modeling and ocean current inversion based on machine learning as described in claim 2, characterized in that: In step (3), the spatiotemporal matching data is transformed by coordinate transformation and ocean current Doppler conversion to construct a multi-parameter dataset of the sea surface.

4. The method for wind and wave response modeling and ocean current inversion based on machine learning as described in claim 3, characterized in that: In step (5), the optimized XGBoost model obtained through testing and case analysis is the optimal sea surface current velocity inversion model.

5. A device for wind and wave response modeling and ocean current inversion based on machine learning, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, a method for wind and wave response modeling and ocean current inversion based on machine learning as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements a method for wind and wave response modeling and ocean current inversion based on machine learning as described in any one of claims 1-4.

Images