A sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer

Abstract

The present invention discloses a method for inverting sea surface wind field based on AWPSO-CNN and HY-2C microwave scatterometer, comprising: obtaining a data set, the data set including L2A data of China Ocean Satellite HY-2C Microwave Scatterometer, data of European Centre for Medium-Range Weather Forecasts ECMWF and data of National Data Buoy Center NDBC; preprocessing the data set; performing data spatiotemporal matching on L2A data of China Ocean Satellite HY-2C Microwave Scatterometer and data of European Centre for Medium-Range Weather Forecasts ECMWF to obtain matching data, and dividing the matching data into a training set and a test set; using the data in the training set to train an AWPSO-CNN model; obtaining a wind speed inversion module and a wind direction inversion module. Independent NDBC data is used to evaluate the performance of the wind speed and wind direction models. Among them, the problem that the convolutional neural network has a slow convergence speed during back propagation and is prone to falling into local extreme values ​​is solved, and the inversion accuracy is improved. The improved support vector machine algorithm can solve the problem that multiple solutions of inverted wind direction often show 180° wind direction ambiguity, and finally obtain a relatively accurate sea surface wind direction.

Description

Technical Field

[0001] The present invention relates to the technical field of microwave scatterometer sea surface wind field remote sensing, and in particular to a sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer. Background Art

[0002] The ocean covers approximately 71% of the Earth's surface and contains abundant energy. The effective detection and utilization of marine resources has always been a research focus. The sea surface wind field is an important factor in the physical state information of the ocean. It drives sea waves and is closely related to the exchange of matter and energy between the ocean and the atmosphere. The effective prediction of the sea surface wind field has a great impact on the development of global weather forecasts and has far-reaching significance for human life and social progress. Traditional methods for obtaining sea surface wind fields include buoys and ships. With the increasing number of buoys, the density of measured data on global sea surface wind fields has gradually increased. However, due to the vast area of ​​the ocean, the measured data is gradually unable to meet the needs of scientific research in terms of temporal continuity and spatial resolution. Ocean satellite microwave remote sensing observation technology can achieve long-term and dynamic observations, and the signal can cover the global ocean. With the continuous development of sea surface wind field satellite remote sensing detection technology, the demand for regional and global sea surface wind field observations has been met.

[0003] Currently, the main methods for retrieving sea surface wind speed and direction are empirical inversion and semi-empirical inversion algorithms. Semi-empirical inversion methods utilize the inversion of geophysical model functions, which are quantitative functional relationships between the normalized radar backscatter cross section and sea surface wind speed, wind direction, radar observation parameters, and environmental parameters. The maximum likelihood estimation (MLE) method is generally used to invert the scatterometer wind field. The maximum likelihood cost function requires that the function values ​​be calculated point by point at a certain search interval throughout the entire two-dimensional wind speed-direction space. These values ​​are then compared to find local extreme points. The wind vector corresponding to each extreme point is called the fuzzy wind vector solution. In the fuzzy wind vector solution obtained, the wind speeds vary minimally and there is no ambiguity. The ambiguity mainly refers to wind direction ambiguity. Currently, the circular median filter method is commonly used to remove the ambiguity in the wind direction solution. By improving the maximum likelihood estimation method, a multi-solution inversion algorithm (MSS) for scatterometer wind fields is proposed. The MSS algorithm retains all fuzzy solutions corresponding to the maximum likelihood cost function and calculates the probability of each fuzzy solution being the "true solution" using a probabilistic model. The probabilistic model parameters are derived using empirical statistical methods. Finally, using all fuzzy solutions and their corresponding probabilities, and using the numerical forecast model wind field as the background field, a two-dimensional variational method is used to determine the inverted wind vector. The empirical inversion algorithm utilizes machine learning. It uses autonomous learning to find complex, abstract functional relationships between multiple independent variables and dependent variables, addressing the low accuracy inherent in physical empirical models due to simple functional relationships. The machine learning algorithm directly links the radar backscatter cross section, observational, and environmental parameters to the sea surface wind vector to invert the sea surface wind field. The wind speed inversion module outputs only wind speed. The wind direction inversion module divides wind direction from 0° to 360° into 36 intervals with a step size of 10°, with a wind direction of 0° indicating due north. It outputs a label indicating that the wind direction belongs to a certain wind direction interval, and then calculates the predicted wind direction based on this label.

[0004] A disadvantage of semi-empirical inversion methods is that their accuracy depends on the accuracy of the geophysical model function, but the empirical function relationship does not fully consider the influence of various factors on the sea surface wind vector. During the function solution process, the calculation of the objective function is relatively complex, resulting in a relatively high computational load for the entire inversion process. A disadvantage of machine learning algorithms is that hyperparameters significantly influence model performance. Empirical parameters used in traditional research may only be applicable to specific scenarios. Manually adjusting algorithm parameters is time-consuming. Gradient descent methods are also prone to falling into local minima in neural networks, preventing them from achieving a global optimum. Using a support vector machine to invert sea surface wind direction requires dividing the wind direction from 0° to 360° into 36 intervals with a step size of 10°. A wind direction of 0° indicates a northerly wind direction. The output is a label indicating that the wind direction belongs to a certain wind direction interval. The predicted wind direction value is then calculated based on this label. However, this wind direction inversion method is complex and computationally intensive in dealing with fuzzy wind direction solutions. Summary of the Invention

[0005] The present invention aims to at least partially address one of the technical problems encountered in the aforementioned technologies. To this end, the present invention proposes a method for inverting sea surface wind fields based on the AWPSO-CNN and HY-2C microwave scatterometer. Based on data from the HY-2C microwave scatterometer, an optimized sea surface wind speed inversion method is designed using an adaptive particle swarm optimization algorithm combined with a convolutional neural network. Furthermore, based on the inverted wind speed data, an improved support vector machine (SVM)-based sea surface wind direction inversion method is designed. The AWPSO algorithm is combined with a convolutional neural network to determine the initial parameters (weights, thresholds, and biases) suitable for the convolutional neural network model based on the input dataset. This algorithm addresses the slow convergence speed and tendency of convolutional neural networks to fall into local extrema during backpropagation, thereby improving inversion accuracy. The improved SVM algorithm can also address the frequent 180° wind direction ambiguity in inverted wind direction solutions, ultimately yielding a more accurate sea surface wind direction. Traditional particle swarm optimization algorithms utilize a large number of particles within a certain search space to find an optimal set of parameters that meet certain requirements. The inertia factor and acceleration coefficient are used to motivate particles to move to their individual and global optimal positions. The adaptive particle swarm optimization algorithm (AWPSO algorithm) adaptively adjusts the inertia factor and acceleration coefficient during the iteration process, thereby improving the algorithm's search capability and finding the global optimal solution more quickly.

[0006] To achieve the above objectives, the present invention proposes a method for inverting sea surface wind fields based on AWPSO-CNN and HY-2C microwave scatterometer, including:

[0007] Step S1: Acquire a data set, wherein the data set includes the China Ocean Satellite HY-2C Microwave Scatterometer L2A data, the European Centre for Medium-Range Weather Forecasts ECMWF data, and the National Data Buoy Center NDBC data;

[0008] Step S2: Dataset preprocessing;

[0009] Step S3: performing spatiotemporal matching on the L2A data of the China Ocean Satellite HY-2C Microwave Scatterometer and the European Centre for Medium-Range Weather Forecasts (ECMWF) data to obtain matching data, and dividing the matching data into a training set and a test set;

[0010] Step S4: Use the data in the training set and the ECMWF data as the true wind speed to train the AWPSO-CNN model;

[0011] Step S5: Design a convolutional neural network model structure. The basic structure includes an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer. The convolution layer and the pooling layer are arranged in an alternating manner.

[0012] Step S6: Initialize the AWPSO algorithm parameters, where the parameters in the convolutional neural network structure include weights, thresholds, and biases, which serve as the number of particle dimensions in the particle swarm algorithm and are arranged in order;

[0013] Step S7: randomly generate m particles, each particle is used as the initial parameter of the convolutional neural network structure, the training set in S3 is input into the convolutional neural network model, and the mean square error between the model predicted wind speed and the true wind speed is used as the fitness function of the particle swarm algorithm;

[0014] Step S8: Set the number of iterations. When the iteration is completed, the global optimal particle found by the AWPSO algorithm is the optimal initialization parameters of the model: weight, threshold and bias;

[0015] Step S9: assign the value of each dimension of the global optimal particle position to the weight, threshold and bias of the convolutional neural network model in sequence; input the training set data in S3 into the model, train the model, set the loss function to the mean root square error function, and judge whether it meets the required accuracy;

[0016] Step S10: Use the test set data in S3 to evaluate the performance of the AWPSO-CNN model;

[0017] Step S11: Matching the HY-2C microwave scatterometer L2A data and the NDBC data in terms of time and space, using the NDBC data as the wind speed reference true value, and evaluating the performance of the AWPSO-CNN inversion model to obtain a wind speed inversion module;

[0018] Step S12: designing an improved support vector machine (SVM) algorithm based on the wind speed inversion module, initializing SVM parameters, and selecting a kernel function and a penalty factor;

[0019] Step S13: Using the data in the training set and the ECMWF buoy data as the true wind direction, the optimal hyperparameters of the SVM model are determined by grid search, and the improved SVM model is trained using the hyperparameters;

[0020] Step S14: Use the test set data in S3 to evaluate the performance of the improved SVM model;

[0021] Step S15: The HY-2C microwave scatterometer L2A data and the NDBC data are matched in terms of time and space. The NDBC data is used as the true reference value of the wind direction, and the performance of the improved SVM model is evaluated to obtain a wind direction inversion module.

[0022] According to some embodiments of the present invention, step S3 specifically includes:

[0023] Step S3.1: spatially match the longitude and latitude of the mirror reflection point of the SCA data with the wind speed reference data;

[0024] Step S3.2: Time-match the SCA data with the wind speed reference data based on the observation time.

[0025] According to some embodiments of the present invention, step S6 specifically includes:

[0026] Step S6.1: The number of trainable particles in each convolutional layer is (Input × CSize + 1) × Output; where Input is the number of input neurons, CSize is the size of the convolution kernel, 1 is the threshold number, each layer has one threshold, and Output is the number of output neurons;

[0027] Step S6.2: Initial weight range:

[0028]

[0029] Among them, Node in and Node out Represents the number of input and output nodes in each layer respectively;

[0030] Step S6.3: Adopt an adaptive weighting strategy to adaptively control the acceleration coefficient in each iteration to find the optimal solution as quickly as possible and accelerate the global convergence speed;

[0031] Step S6.4: The selection of parameters in the AWPSO algorithm is:

[0032]

[0033] Among them, w is the inertia weight, w max and w min are the initial and final inertia weights, which are generally set to 0.9 and 0.1, maxiter is the maximum number of iterations, and j is the current number of iterations;

[0034] The adaptive weighting strategy includes:

[0035] C1=F(g1),C2=F(g2)

[0036]

[0037]

[0038] Where C1 and C2 represent acceleration coefficients, g1 and g2 represent the distances between the d-th dimension of particle i and the individual and global optimal values ​​in the t-th iteration; F(·) represents the adaptive weighted update function; a is a constant representing the steepness of the curve, and b, c, and d are 0.5, 0, and 1.5, respectively; pbest represents the individual optimal value of the d-th dimension of particle i in the t-th iteration, and gbest represents the global optimal value of the d-th dimension of particle i in the t-th iteration; F(D) is the specific expression of F(·), which represents the adaptive weighted update function.

[0039] According to some embodiments of the present invention, step S7 includes:

[0040] Step S7.1: Assume that the search space is n-dimensional, and the particle swarm X=(X1,…X i ,…,X m ) contains m particles, the position of the i-th particle Xi=(x i1 ,x i2 ,…,x in ) T , speed Vi=(v i1 ,v i2 ,…,v in ) T ;

[0041] Step S7.2: Use the mean square error function as the fitness function:

[0042]

[0043] Where N represents the number of iterations, y i and y ture represent the model predicted wind speed and the true reference wind speed, respectively.

[0044] According to some embodiments of the present invention, step S8 includes:

[0045] Step S8.1: In each iteration, once the particle has found its individual and global optimal positions, it can update its own velocity and position information:

[0046]

[0047]

[0048] Among them, w is the inertia weight, c1 and c2 are acceleration coefficients, r1 and r2 are random numbers between 0 and 1, and denote the velocity and position of the dth dimension of the i-th particle in the t-th iteration respectively;

[0049] Step S8.2: The specific steps of the AWPSO algorithm are:

[0050] (1) Randomly set the initial velocity and position of the particles, set the number of particle swarms, and initialize the parameters to be adjusted;

[0051] (2) Input the particles into the convolutional neural network model in sequence, input the training set data, obtain the model-predicted wind speed value, and calculate the fitness of each particle according to the fitness formula;

[0052] (3) Compare the fitness of each particle's current position with the fitness of the individual optimal position. If it is smaller, update the individual optimal position; otherwise, keep it unchanged.

[0053] (4) Compare the fitness of each particle's current position with the fitness of the global optimal position. If it is smaller, update the global optimal position; otherwise, keep it unchanged.

[0054] (5) Update the parameter value to be adjusted according to the formula;

[0055] (6) Update the speed and position of each particle according to the formula;

[0056] (7) If the set termination condition is not met, return to step (2); if the termination condition is met, end the loop and output the global optimal particle.

[0057] According to some embodiments of the present invention, step S9 specifically includes:

[0058] Step S9.1: The output global optimal particles are assigned to the positions corresponding to the convolutional neural network parameters in order;

[0059] Step S9.2: Input the training set data. When the result output by the convolutional neural network does not match our expected value, the back propagation process is performed; the error between the result and the expected value is calculated. The data will be lost in the process of being transmitted between layers. The error value caused by each layer is different, so the error is returned layer by layer, the error of each layer is calculated, and then the weight is updated.

[0060] According to some embodiments of the present invention, step S10 specifically includes:

[0061] Step S10.1: Evaluate the trained model on the test set; use the RMSE function as an indicator to analyze the accuracy of the model:

[0062]

[0063] Where N is the number of samples, f i and f true represent the wind speed output by the model and the true reference wind speed, respectively.

[0064] According to some embodiments of the present invention, step S12 specifically includes:

[0065] Step S12.1: The nonlinear support vector machine f(z) obtained by solution is as follows:

[0066]

[0067]

[0068] Among them, {(x1,y1),(x2,y2),…,(x N ,y N )} is the training data set, x i and y i represents the characteristic parameters and wind direction interval label of the i-th sample vector in the training dataset; z represents the input sample vector; γ is the only hyperparameter of the Gaussian kernel function, which represents the modulus of the sample vector; a i * represents the first coefficient of the i-th sample vector, b * represents the second coefficient; K(z,x i ) represents the Gaussian kernel function; N is the number of samples in the training data set; e is a natural constant.

[0069] According to some embodiments of the present invention, step S13 specifically includes:

[0070] Step S13.1: The SVM input parameters include the backscatter coefficient measured by the scatterometer, the incident angle, azimuth, longitude and latitude of the antenna beam, the signal-to-noise ratio, and the wind speed information obtained by the wind speed inversion module. The output parameter is the label of the wind direction interval to which the sea surface wind direction belongs. The mapping relationship WD between the sea surface wind direction and each characteristic parameter is as follows:

[0071] WD=f(WS,θ,μ,lat,lon,NBRCS,SNR)

[0072] Among them, the characteristic parameters are the wind speed information WS obtained by the wind speed inversion module, the antenna beam incidence angle θ, the azimuth angle μ, the longitude lon, the latitude lat, the backscatter coefficient NBRCS and the signal-to-noise ratio SNR.

[0073] According to some embodiments of the present invention, step S14 specifically includes:

[0074] Step S14.1: Evaluate the trained improved SVM model on the test set, using the RMSE function as an indicator to analyze the model:

[0075]

[0076] Where N is the number of samples, d true represents the true value of wind direction in the ECMWF reanalysis dataset; d iIndicates the middle value of the predicted wind direction range.

[0077] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0078] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0080] Figure 1 This is a flow chart of a method for inverting sea surface wind fields based on AWPSO-CNN and HY-2C microwave scatterometer according to one embodiment of the present invention;

[0081] Figure 2 is a flow chart of a wind speed inversion module according to one embodiment of the present invention;

[0082] Figure 3 4 is a flow chart of a wind direction inversion module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0083] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0084] According to the attached Figure 1-3 A sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer is described.

[0085] To achieve the above objectives, the present invention proposes a method for inverting sea surface wind fields based on AWPSO-CNN and HY-2C microwave scatterometer, including:

[0086] Step S1: Acquire a data set, wherein the data set includes the China Ocean Satellite HY-2C Microwave Scatterometer L2A data, the European Centre for Medium-Range Weather Forecasts ECMWF data, and the National Data Buoy Center NDBC data;

[0087] Step S2: Dataset preprocessing;

[0088] Step S3: performing spatiotemporal matching on the L2A data of the China Ocean Satellite HY-2C Microwave Scatterometer and the European Centre for Medium-Range Weather Forecasts (ECMWF) data to obtain matching data, and dividing the matching data into a training set and a test set;

[0089] Step S4: Use the data in the training set and the ECMWF data as the true wind speed to train the AWPSO-CNN model;

[0090] Step S5: Design a convolutional neural network model structure. The basic structure includes an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer. The convolution layer and the pooling layer are arranged in an alternating manner.

[0091] Step S6: Initialize the AWPSO algorithm parameters, where the parameters in the convolutional neural network structure include weights, thresholds, and biases, which serve as the number of particle dimensions in the particle swarm algorithm and are arranged in order;

[0092] Step S7: randomly generate m particles, each particle is used as the initial parameter of the convolutional neural network structure, the training set in S3 is input into the convolutional neural network model, and the mean square error between the model predicted wind speed and the true wind speed is used as the fitness function of the particle swarm algorithm;

[0093] Step S8: Set the number of iterations. When the iteration is completed, the global optimal particle found by the AWPSO algorithm is the optimal initialization parameters of the model: weight, threshold and bias;

[0094] Step S9: assign the value of each dimension of the global optimal particle position to the weight, threshold and bias of the convolutional neural network model in sequence; input the training set data in S3 into the model, train the model, set the loss function to the mean root square error function, and judge whether it meets the required accuracy;

[0095] Step S10: Use the test set data in S3 to evaluate the performance of the AWPSO-CNN model;

[0096] Step S11: Matching the HY-2C microwave scatterometer L2A data and the NDBC data in terms of time and space, using the NDBC data as the wind speed reference true value, and evaluating the performance of the AWPSO-CNN inversion model to obtain a wind speed inversion module;

[0097] Step S12: designing an improved support vector machine (SVM) algorithm based on the wind speed inversion module, initializing SVM parameters, and selecting a kernel function and a penalty factor;

[0098] Step S13: Using the data in the training set and the ECMWF buoy data as the true wind direction, the optimal hyperparameters of the SVM model are determined by grid search, and the improved SVM model is trained using the hyperparameters;

[0099] Step S14: Use the test set data in S3 to evaluate the performance of the improved SVM model;

[0100] Step S15: The HY-2C microwave scatterometer L2A data and the NDBC data are matched in terms of time and space. The NDBC data is used as the true reference value of the wind direction, and the performance of the improved SVM model is evaluated to obtain a wind direction inversion module.

[0101] The above technical solution has the following beneficial effects: Based on data from the HY-2C microwave scatterometer, an optimized sea surface wind speed inversion method using an adaptive particle swarm optimization algorithm combined with a convolutional neural network was designed. Furthermore, based on the inverted wind speed data, an improved support vector machine (SVM)-based sea surface wind direction inversion method was designed. The AWPSO algorithm was combined with the convolutional neural network to determine the appropriate initial parameters for the convolutional neural network model: weights, thresholds, and biases, based on the input dataset. This algorithm addresses the slow convergence speed and local extrema of the convolutional neural network during backpropagation, improving inversion accuracy. The improved SVM algorithm also addresses the 180° ambiguity often seen in multiple solutions of inverted wind direction, ultimately yielding a more accurate sea surface wind direction. Traditional particle swarm algorithms utilize a large number of particles to find an optimal set of parameters within a certain search space. The inertia factor and acceleration coefficient are used to motivate particles to move to individual and global optimal positions. The adaptive particle swarm algorithm (AWPSO) adaptively adjusts the inertia factor and acceleration coefficient during the iteration process, thereby improving the algorithm's search capability and accelerating the search for the global optimal solution.

[0102] According to some embodiments of the present invention, step S3 specifically includes:

[0103] Step S3.1: spatially match the longitude and latitude of the mirror reflection point of the SCA data with the wind speed reference data;

[0104] Step S3.2: Time-match the SCA data with the wind speed reference data based on the observation time.

[0105] The beneficial effect of the above technical solution is: based on spatial matching and temporal matching, the accuracy of the obtained matching data is guaranteed.

[0106] SCA data refers to the L2A data of the microwave scatterometer on the China Ocean Satellite HY-2C. Wind speed reference data refers to the data of the European Centre for Medium-Range Weather Forecasts (ECMWF).

[0107] According to some embodiments of the present invention, step S6 specifically includes:

[0108] Step S6.1: The number of trainable particles in each convolutional layer is (Input × CSize + 1) × Output; where Input is the number of input neurons, CSize is the size of the convolution kernel, 1 is the threshold number, each layer has one threshold, and Output is the number of output neurons;

[0109] Step S6.2: Initial weight range:

[0110]

[0111] Among them, Node in and Node out Represents the number of input and output nodes in each layer respectively;

[0112] Step S6.3: Adopt an adaptive weighting strategy to adaptively control the acceleration coefficient in each iteration to find the optimal solution as quickly as possible and accelerate the global convergence speed;

[0113] Step S6.4: The selection of parameters in the AWPSO algorithm is:

[0114]

[0115] Among them, w is the inertia weight, w max and w min are the initial and final inertia weights, which are generally set to 0.9 and 0.1, maxiter is the maximum number of iterations, and j is the current number of iterations;

[0116] The adaptive weighting strategy includes:

[0117] C1=F(g1),C2=F(g2)

[0118]

[0119]

[0120] Where C1 and C2 represent acceleration coefficients, g1 and g2 represent the distances between the d-th dimension of particle i and the individual and global optimal values ​​in the t-th iteration; F(·) represents the adaptive weighted update function; a is a constant representing the steepness of the curve, and b, c, and d are 0.5, 0, and 1.5, respectively; pbest represents the individual optimal value of the d-th dimension of particle i in the t-th iteration, and gbest represents the global optimal value of the d-th dimension of particle i in the t-th iteration; F(D) is the specific expression of F(·), which represents the adaptive weighted update function.

[0121] Beneficial effects of the above technical solution: Adaptive Particle Swarm Optimization (AWPSO algorithm) adaptively adjusts the inertia factor and acceleration coefficient during the iteration process, thereby improving the algorithm's search capability and finding the global optimal solution more quickly.

[0122] According to some embodiments of the present invention, step S7 includes:

[0123] Step S7.1: Assume that the search space is n-dimensional, and the particle swarm X=(X1,…X i ,…,X m ) contains m particles, the position of the i-th particle Xi=(x i1 ,x i2 ,…,x in ) T , speed Vi=(v i1 ,v i2 ,…,v in ) T ;

[0124] Step S7.2: Use the mean square error function as the fitness function:

[0125]

[0126] Where N represents the number of iterations, y i and y ture represent the model predicted wind speed and the true reference wind speed, respectively.

[0127] The beneficial effect of the above technical solution is to obtain an accurate fitness function of the particle swarm algorithm.

[0128] According to some embodiments of the present invention, step S8 includes:

[0129] Step S8.1: In each iteration, once the particle has found its individual and global optimal positions, it can update its own velocity and position information:

[0130]

[0131]

[0132] Among them, w is the inertia weight, c1 and c2 are acceleration coefficients, r1 and r2 are random numbers between 0 and 1, and denote the velocity and position of the dth dimension of the i-th particle in the t-th iteration respectively;

[0133] Step S8.2: The specific steps of the AWPSO algorithm are:

[0134] (1) Randomly set the initial velocity and position of the particles, set the number of particle swarms, and initialize the parameters to be adjusted;

[0135] (2) Input the particles into the convolutional neural network model in sequence, input the training set data, obtain the model-predicted wind speed value, and calculate the fitness of each particle according to the fitness formula;

[0136] (3) Compare the fitness of each particle's current position with the fitness of the individual optimal position. If it is smaller, update the individual optimal position; otherwise, keep it unchanged.

[0137] (4) Compare the fitness of each particle's current position with the fitness of the global optimal position. If it is smaller, update the global optimal position; otherwise, keep it unchanged.

[0138] (5) Update the parameter value to be adjusted according to the formula;

[0139] (6) Update the speed and position of each particle according to the formula;

[0140] (7) If the set termination condition is not met, return to step (2); if the termination condition is met, end the loop and output the global optimal particle.

[0141] The beneficial effect of the above technical solution is: based on the iterative calculation of the AWPSO algorithm, the global optimal particle found is the optimal initialization parameter of the model.

[0142] According to some embodiments of the present invention, step S9 specifically includes:

[0143] Step S9.1: The output global optimal particles are assigned to the positions corresponding to the convolutional neural network parameters in order;

[0144] Step S9.2: Input the training set data. When the result output by the convolutional neural network does not match our expected value, the back propagation process is performed; the error between the result and the expected value is calculated. The data will be lost in the process of being transmitted between layers. The error value caused by each layer is different, so the error is returned layer by layer, the error of each layer is calculated, and then the weight is updated.

[0145] The beneficial effects of the above technical solution are: the values ​​of each dimension of the global optimal particle position are sequentially assigned to the weights, thresholds and biases of the convolutional neural network model, and the accuracy of the model is trained based on the training set data.

[0146] According to some embodiments of the present invention, step S10 specifically includes:

[0147] Step S10.1: Evaluate the trained model on the test set; use the RMSE function as an indicator to analyze the accuracy of the model:

[0148]

[0149] Where N is the number of samples, f i and f true represent the wind speed output by the model and the true reference wind speed, respectively.

[0150] The beneficial effects of the above technical solution are: the performance of the AWPSO-CNN model is evaluated based on the test set data, and the parameters of the AWPSO-CNN model are verified and corrected.

[0151] According to some embodiments of the present invention, step S12 specifically includes:

[0152] Step S12.1: The nonlinear support vector machine f(z) obtained by solution is as follows:

[0153]

[0154]

[0155] Among them, {(x1,y1),(x2,y2),…,(x N ,y N )} is the training data set, x i and y i represents the characteristic parameters and wind direction interval label of the i-th sample vector in the training dataset; z represents the input sample vector; γ is the only hyperparameter of the Gaussian kernel function, which represents the modulus of the sample vector; a i * represents the first coefficient of the i-th sample vector, b * represents the second coefficient; K(z,x i ) represents the Gaussian kernel function; N is the number of samples in the training data set; e is a natural constant.

[0156] Support vector machines have obvious advantages in solving nonlinear classification problems and high-dimensional feature space classification problems. The key to the SVM algorithm is to find the optimal feature space separation hyperplane.

[0157] The beneficial effects of the above technical solution are: accurately establishing an improved support vector machine (SVM) algorithm, initializing SVM parameters, selecting kernel functions and penalty factors, and facilitating obtaining an improved SVM model.

[0158] According to some embodiments of the present invention, step S13 specifically includes:

[0159] Step S13.1: The SVM input parameters include the backscatter coefficient measured by the scatterometer, the incident angle, azimuth, longitude and latitude of the antenna beam, the signal-to-noise ratio, and the wind speed information obtained by the wind speed inversion module. The output parameter is the label of the wind direction interval to which the sea surface wind direction belongs. The mapping relationship WD between the sea surface wind direction and each characteristic parameter is as follows:

[0160] WD=f(WS,θ,μ,lat,lon,NBRCS,SNR)

[0161] Among them, the characteristic parameters are the wind speed information WS obtained by the wind speed inversion module, the antenna beam incidence angle θ, the azimuth angle μ, the longitude lon, the latitude lat, the backscatter coefficient NBRCS and the signal-to-noise ratio SNR.

[0162] The beneficial effects of the above technical solution are: obtaining an accurate improved SVM model, using the wind speed obtained by inverting the wind speed module as the input parameter of the SVM model, which can improve the accuracy of wind direction inversion, and the improved support vector machine algorithm can solve the problem that the inverted wind direction multiple solutions often show 180° wind direction ambiguity.

[0163] According to some embodiments of the present invention, step S14 specifically includes:

[0164] Step S14.1: Evaluate the trained improved SVM model on the test set, using the RMSE function as an indicator to analyze the model:

[0165]

[0166] Where N is the number of samples, d true represents the true value of wind direction in the ECMWF reanalysis dataset; d i Indicates the median value of the predicted wind direction interval. For example, the median value of the interval labeled 1 is 5°, and the median value of the interval labeled 2 is 10°.

[0167] The beneficial effects of the above technical solution are: the performance of the improved SVM model is evaluated based on the test set data, and the parameters of the improved SVM model are verified and corrected.

[0168] According to some embodiments of the present invention, step S5 specifically includes:

[0169] Step S5.1: Use Sigmoid as the activation function in the output layer, and use ReLU activation function in the input layer, convolutional layer, pooling layer, and fully connected layer.

[0170] According to some embodiments of the invention, the matching data comprises scatterometer data away from a coastline.

[0171] According to some embodiments of the present invention, step S1 includes:

[0172] Step S1.1: Batch obtain the datasets of the specified date through the China Ocean Satellite Data Service System, the European Centre for Medium-Range Weather Forecasts, and the National Buoy Data Center;

[0173] Step S1.2: Download and save in HDF5 format.

[0174] According to some embodiments of the present invention, step S2 specifically includes:

[0175] Step S2.1: Perform quality control and filtering on the dataset: remove NaN values.

[0176] The beneficial effects of the above technical solution are: improving the effective utilization and accuracy of data sets.

[0177] Step S4 specifically includes: Step S4.1: the input data is the backscatter coefficient, and the true value, i.e., the reference wind speed, is the corresponding ECMWF data.

[0178] Regarding the specific implementation, step S11 specifically includes:

[0179] Step S11.1: Because the wind speed reference values ​​used for both training and testing the AWPSO-CNN model are ECMWF data, it is necessary to evaluate the model's performance using independent wind speed measurement data. To this end, the HY-2C microwave scatterometer L2A data and the NDBC data are temporally and spatially matched, with the NDBC data used as the wind speed reference values ​​to generate the validation data set. This validation data set is then fed into the AWPSO-CNN model, and the RMSE function is used to evaluate the model's performance, resulting in the final wind speed retrieval module.

[0180] For a specific implementation, step S15 specifically includes:

[0181] Step S15.1: Because the wind direction reference values ​​used for training and testing the improved SVM model are all ECMWF data, it is necessary to evaluate the model performance using independent wind direction measurement data. To this end, the HY-2C microwave scatterometer L2A data and the NDBC data are matched temporally and spatially, and the NDBC wind direction data are used as the wind direction reference values ​​to obtain the validation data set. This validation data set is input into the improved SVM model to obtain the inverted wind direction results, and the RMSE function is used to evaluate the model performance.

[0182] The beneficial effects of the present invention are as follows: the adaptive particle swarm algorithm adaptively adjusts the inertia factor and acceleration coefficient during the iteration process. The search capability of the algorithm can be improved, and the global optimal solution can be found more quickly. The AWPSO algorithm is combined with the convolutional neural network. Through the AWPSO algorithm, the initial parameters suitable for the convolutional neural network model are found according to the input data set, which solves the problem that the convolutional neural network has a slow convergence speed during back propagation and is prone to falling into local extreme values, thereby improving the inversion accuracy. The wind speed obtained by the wind speed module inversion is used as the input parameter of the SVM model to improve the wind direction inversion accuracy, and the improved support vector machine algorithm can solve the problem that the inverted wind direction multiple solutions often show 180° wind direction ambiguity.

[0183] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer, characterized in that: include: Step S1: Acquire a data set, wherein the data set includes the China Ocean Satellite HY-2C Microwave Scatterometer L2A data, the European Centre for Medium-Range Weather Forecasts ECMWF data, and the National Data Buoy Center NDBC data; Step S2: Dataset preprocessing; Step S3: performing spatiotemporal matching on the L2A data of the China Ocean Satellite HY-2C Microwave Scatterometer and the ECMWF data to obtain matching data, and dividing the matching data into a training set and a test set; Step S4: Use the data in the training set and the ECMWF data as the true wind speed to train the AWPSO-CNN model; Step S5: Design a convolutional neural network model structure, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer; wherein the convolutional layer and the pooling layer are arranged in an alternating manner; Step S6: Initialize the AWPSO algorithm parameters, where the parameters in the convolutional neural network structure include weights, thresholds, and biases, which serve as the number of particle dimensions in the particle swarm algorithm and are arranged in order; Step S7: randomly generate m particles, each particle is used as the initial parameter of the convolutional neural network structure, the training set in S3 is input into the convolutional neural network model, and the mean square error between the model predicted wind speed and the true wind speed is used as the fitness function of the particle swarm algorithm; Step S8: Set the number of iterations. When the iteration is completed, the global optimal particle found by the AWPSO algorithm is the optimal initialization parameters of the model: weight, threshold and bias; Step S9: Assign the value of each dimension of the global optimal particle position to the weight, threshold, and bias of the convolutional neural network model in sequence; input the training set data in S3 into the model, train the model, set the loss function to the mean root square error function, and determine whether it meets the required accuracy; Step S10: Use the test set data in S3 to evaluate the performance of the AWPSO-CNN model; Step S11: Matching the HY-2C microwave scatterometer L2A data and the NDBC data in terms of time and space, using the NDBC data as the wind speed reference true value, and evaluating the performance of the AWPSO-CNN inversion model to obtain a wind speed inversion module; Step S12: designing an improved support vector machine (SVM) algorithm based on the wind speed inversion module, initializing SVM parameters, and selecting a kernel function and a penalty factor; Step S13: Using the data in the training set and the ECMWF buoy data as the true wind direction, the optimal hyperparameters of the SVM model are determined by grid search, and the improved SVM model is trained using the hyperparameters; Step S14: Use the test set data in S3 to evaluate the performance of the improved SVM model; Step S15: The HY-2C microwave scatterometer L2A data and the NDBC data are matched in terms of time and space. The NDBC data is used as the true reference value of the wind direction, and the performance of the improved SVM model is evaluated to obtain a wind direction inversion module.

2. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S3 specifically includes: Step S3.1: spatially match the longitude and latitude of the mirror reflection point of the SCA data with the wind speed reference data; Step S3.2: Time-match the SCA data with the wind speed reference data based on the observation time.

3. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S6 specifically includes: Step S6.1: The number of trainable particles in each convolutional layer is (Input × CSize + 1) × Output; where Input is the number of input neurons, CSize is the size of the convolution kernel, 1 is the threshold number, each layer has one threshold, and Output is the number of output neurons; Step S6.2: Initial weight range: Among them, Node in and Node out Represents the number of input and output nodes in each layer respectively; Step S6.3: Adopt an adaptive weighting strategy to adaptively control the acceleration coefficient in each iteration to find the optimal solution as quickly as possible and accelerate the global convergence speed; Step S6.4: The selection of parameters in the AWPSO algorithm is: Among them, w is the inertia weight, w max and w min are the initial and final inertia weights, with values ​​of 0.9 and 0.1, maxiter is the maximum number of iterations, and j is the current number of iterations; The adaptive weighting strategy includes: Where C1 and C2 represent acceleration coefficients, g1 and g2 represent the distances between the d-th dimension of particle i and the individual optimal value and the global optimal value in the t-th iteration; F(·) represents the adaptive weighted update function; a is a constant representing the steepness of the curve, and b, c, and d are 0.5, 0, and 1.5, respectively; pbest represents the individual optimal value of the d-th dimension of particle i in the t-th iteration, and gbest represents the global optimal value of the d-th dimension of particle i in the t-th iteration; F(D) is the specific expression of F(·), which represents the adaptive weighted update function.

4. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S7 includes: Step S7.1: Assume that the search space is n-dimensional, and the particle swarm X=(X1,…X i ,…,X m ) contains m particles, the position of the i-th particle Xi=(x i1 ,x i2 ,…,x in ) T , speed Vi=(v i1 ,v i2 ,…,v in ) T ; Step S7.2: Use the mean square error function as the fitness function: Where N represents the number of iterations, y i and y ture represent the model predicted wind speed and the true reference wind speed, respectively.

5. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S8 includes: Step S8.1: In each iteration, once the particle has found its individual and global optimal positions, it can update its own velocity and position information: Among them, w is the inertia weight, c1 and c2 are acceleration coefficients, r1 and r2 are random numbers between 0 and 1, and denote the velocity and position of the dth dimension of the i-th particle in the t-th iteration respectively; Step S8.2: The specific steps of the AWPSO algorithm are: (1) Randomly set the initial velocity and position of the particles, set the number of particle swarms, and initialize the parameters to be adjusted; (2) Input the particles into the convolutional neural network model in sequence, input the training set data, obtain the model-predicted wind speed value, and calculate the fitness of each particle according to the fitness formula; (3) Compare the fitness of each particle's current position with the fitness of the individual optimal position. If it is smaller, update the individual optimal position; otherwise, keep it unchanged. (4) Compare the fitness of each particle's current position with the fitness of the global optimal position. If it is smaller, update the global optimal position; otherwise, keep it unchanged. (5) Update the parameter value to be adjusted according to the formula; (6) Update the speed and position of each particle according to the formula; (7) If the set termination condition is not met, return to step (2); if the termination condition is met, end the loop and output the global optimal particle.

6. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S9 specifically includes: Step S9.1: The output global optimal particles are assigned to the positions corresponding to the convolutional neural network parameters in order; Step S9.2: Input the training set data. When the result output by the convolutional neural network does not match our expected value, the back propagation process is performed; the error between the result and the expected value is calculated. The data will be lost in the process of being transmitted between layers. The error value caused by each layer is different, so the error is returned layer by layer, the error of each layer is calculated, and then the weight is updated.

7. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S10 specifically includes: Step S10.1: Evaluate the trained model on the test set; use the RMSE function as an indicator to analyze the accuracy of the model: Where N is the number of samples, f i and f true represent the wind speed output by the model and the true reference wind speed, respectively.

8. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S12 specifically includes: Step S12.1: The nonlinear support vector machine f(z) obtained by solution is as follows: Among them, {(x1,y1),(x2,y2),…,(x N ,y N )} is the training data set, x i and y i represents the characteristic parameters and wind direction interval label of the i-th sample vector in the training dataset; z represents the input sample vector; γ is the only hyperparameter of the Gaussian kernel function, which represents the modulus of the sample vector; a i * represents the first coefficient of the i-th sample vector, b * represents the second coefficient; K(z,x i ) represents the Gaussian kernel function; N is the number of samples in the training data set; e is a natural constant.

9. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S13 specifically includes: Step S13.1: The SVM input parameters include the backscatter coefficient measured by the scatterometer, the incident angle of the antenna beam, azimuth, longitude and latitude, signal-to-noise ratio, and wind speed information obtained by the wind speed inversion module; the output parameter is the label of the wind direction interval to which the sea surface wind direction belongs; the mapping relationship between the sea surface wind direction and each characteristic parameter is: as follows: Among them, the characteristic parameters are the wind speed information WS obtained by the wind speed inversion module, the antenna beam incidence angle θ, the azimuth angle μ, the longitude lon, the latitude lat, the backscatter coefficient NBRCS and the signal-to-noise ratio SNR.

10. The sea surface wind field inversion method based on AWPSO-CNN and HY-2C microwave scatterometer according to claim 1, characterized in that: Step S14 specifically includes: Step S14.1: Evaluate the trained improved SVM model on the test set, using the RMSE function as an indicator to analyze the model: Where N is the number of samples, d true represents the true value of wind direction in the ECMWF reanalysis dataset; d i Indicates the middle value of the predicted wind direction range.

Images