Joint prediction method for sea clutter amplitude distribution type and parameters

Abstract

The invention belongs to the technical field of electromagnetic scattering signal analysis, and particularly relates to a sea clutter amplitude distribution type and parameter joint prediction method, which comprises the following steps of: selecting long sequence features and statistical features from original sea clutter data through feature engineering based on actually measured X-waveband sea clutter data; the method comprises the following steps of: firstly, carrying out parameter estimation on four distributions including Pareto distribution, Log-normal distribution, Weibull distribution and K distribution through a maximum likelihood estimation algorithm, then checking optimal fitting distribution, selecting the optimal distribution from the four basic distributions to mark data, constructing a data set, and finally, selecting the optimal distribution from the four basic distributions to mark the data. A deep learning model for accurate prediction of amplitude distribution types and parameters thereof is established in combination with an MT1dCNN model, the problem that a single amplitude distribution characteristic cannot cover actual complex sea conditions is solved, accurate and reliable prediction of the amplitude distribution types and parameters is realized, and important technical support is provided for design and optimization of a radar system.

Description

Technical Field

[0001] The invention belongs to the technical field of electromagnetic scattering signal analysis, and in particular relates to a joint prediction method for sea clutter amplitude distribution types and parameters. Background Art

[0002] When marine radar detects targets, it will inevitably be affected by the scattered echoes from the sea surface, namely sea clutter. The profound influence of the marine environment and meteorological dynamic changes shows the complex characteristics of non-Gaussian distribution and non-stationarity, which constitute a major obstacle for the radar system to accurately detect targets. Therefore, in-depth understanding and mastery of the characteristics of sea clutter has become a necessary prerequisite for improving the radar's target detection performance in a background rich in sea clutter. Among them, the analysis and modeling of the amplitude distribution of sea clutter is recognized as one of the core characteristics, which is directly related to whether the target signal can be effectively distinguished from the background clutter, thereby affecting the overall performance of the radar system.

[0003] The two key factors that affect the accuracy of sea surface target detection based on sea clutter amplitude distribution are the amplitude distribution type and the parameter estimation algorithm. In the early days, due to the relatively low radar resolution and the presence of many scatterers in a receiving unit, the sea clutter amplitude was assumed to follow a Rayleigh distribution with one parameter. Rayleigh distribution is suitable for describing scenarios when the number of scatterers is very large and the contributions of each scatterer are independent of each other. With the improvement of radar resolution, multi-parameter lognormal distribution, K distribution, Pareto distribution, KK distribution, etc. have been proposed. Among the parameter estimation methods, the commonly used maximum likelihood estimation provides consistent and effective parameter estimation, but the solution process may be more complicated, especially when dealing with complex distribution models. Moment estimation is simple to calculate, but is easily affected by outliers. In addition, there are binary point estimation and parameter estimation based on intelligent evolutionary algorithms. However, under different sea conditions, the most suitable sea clutter amplitude distribution model is different. A single distribution model can no longer fully cover the complex diversity of the marine environment, and different estimation methods have different accuracy when estimating parameters of different amplitude distributions. Therefore, there is an urgent need for a method to accurately predict the amplitude distribution type and parameters of sea clutter in a dynamically changing ocean environment. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a joint prediction method for sea clutter amplitude distribution type and parameters, which uses four distributions, namely Pareto distribution, Log-normal distribution, Weibull distribution and K distribution, as the basic framework for prediction, and comprehensively and meticulously describes various sea conditions through diversified distribution types, thereby achieving effective prediction of sea clutter amplitude characteristics in a complex and changeable sea environment, solving the problem that a single distribution model is difficult to cover complex sea conditions, and can be applied to the prediction of sea clutter amplitude distribution type and parameters, and can improve the performance of marine radar target detection.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A joint prediction method for sea clutter amplitude distribution type and parameters comprises the following steps:

[0007] S1. Constructing a data set: Collect X-band sea clutter data, and remove ships and buoys to obtain pure sea clutter data. Extract long sequence features and statistical features from the pure sea clutter data through feature engineering. Estimate the parameters of Pareto distribution, Log-normal distribution, Weibull distribution, and K distribution through the maximum likelihood estimation algorithm. Select the best fitting distribution type through the ETIC test standard, and annotate the distribution type and corresponding parameters as the labels of the measured sea clutter data to obtain a data set. The data set is divided into a training set, a validation set, and a test set. The training set and the validation set are used to construct a joint prediction model of amplitude distribution type and parameters (MT1dCNN joint prediction model), and the test set is used to test the prediction performance of the model.

[0008] S2. Build a joint prediction model for amplitude distribution type and parameters: The constructed joint prediction model for amplitude distribution type and parameters (MT1dCNN joint prediction model) includes an input layer, a shared feature extraction layer (one-dimensional convolution layer), a task private layer (fully connected layer) and an output layer. The shared feature extraction layer uses a one-dimensional convolution layer. The task private layer includes a classification task private layer and a regression task private layer. Both use a fully connected layer. The one-dimensional convolution layer, the classification task private layer and the regression task private layer all contain hidden layers. The fully connected layers are composed of multiple neurons. The neurons in the layer are not connected, and the neurons between the layers are connected. The forward-backward unidirectional connection is used. The result of the classification task is passed to the input of the regression task. The activation function of the two adjacent hidden layers is the rectified linear unit. The activation function between the last hidden layer and the output layer in the classification task private layer uses the Softmax activation function, and the regression task private layer uses the Linear activation function. The one-dimensional convolution layer takes three long sequence features as input and outputs one high-level feature after convolution. The high-level feature is then combined with 10 statistical features as the input of the classification task private layer. The regression task private layer additionally adds the output result of the classification task as input.

[0009] S3, training amplitude distribution type and parameter joint prediction model: based on the data set of step S1, use the training set and the validation set to train the amplitude distribution type and parameter joint prediction model, wherein the loss function of the classification task is the cross entropy loss, and the loss function of the regression task is the mean absolute error (MAE), and use the stochastic gradient descent algorithm to train the parameters in the model to obtain the trained amplitude distribution type and parameter joint prediction model;

[0010] S4. Test the model and optimize the network structure: Based on the data set in step S1, use the test set to optimize the number of shared layers, independent layers, convolution kernel size, and the number of neural networks corresponding to each hidden layer in the deep neural network in the amplitude distribution type and parameter joint prediction model trained in step S3, so that the various indicators of the amplitude distribution type and parameter joint prediction model test set are optimal, where the indicators of the classification task include accuracy, precision, recall rate, and F1 score, and the indicators of the regression task include MAE, root mean square error (RMSE), R 2 value.

[0011] As a further technical solution of the present invention, the X-band sea clutter data in step S1 includes T1 and T2 pulse data of HH polarization and VV polarization, the long sequence features include histogram, probability density function (PDF), complementary cumulative distribution function (CCDF), and the statistical features include minimum value, maximum value, variance, standard deviation, kurtosis coefficient, skewness coefficient, 25th percentile, 50th percentile, 75th percentile and 90th percentile.

[0012] As a further technical solution of the present invention, the gradient of the rectified linear unit (ReLU) in step S2 is 1 when the eigenvalue is relatively large.

[0013] As a further technical solution of the present invention, the optimal model obtained in step S4 is: the one-dimensional convolution layer contains 5 hidden layers, the convolution kernel sizes are 11, 11, 7, 7, and 1, respectively, the number of channels are 256, 128, 64, 32, and 1, respectively, and the size of the average pooling convolution kernel is 3; the classification task private layer and the regression task private layer have 2 hidden layers, and the number of neurons in each hidden layer is 128.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] The present invention uses measured sea clutter data, combined with long and short sequence features, and utilizes the MT1dCNN model to jointly predict the amplitude distribution type and parameters. After optimization and parameter adjustment, the F1 score of the HH polarization classification task reaches 97.4%, the RMSE of the regression task is 0.746, and the F1 score of the VV polarization is 96.74%, and the RMSE is 1.071, which achieves accurate and reliable prediction of the amplitude distribution type and parameters, providing important technical support for the design and optimization of radar systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention provides a flow chart for the joint prediction of sea clutter amplitude distribution types and parameters.

[0017] Figure 2 This is a schematic diagram of the overall framework of the MT1dCNN model provided by the present invention.

[0018] Figure 3 The present invention provides a schematic diagram of the one-dimensional convolution shared feature extraction layer framework of the MT1dCNN model, taking the PDF feature as an example.

[0019] Figure 4 The scatter density map of the confusion matrix heat map of the classification task of the MT1dCNN model provided by the present invention and the comparison between the true value and the model prediction value of the regression task, wherein (a) is the HH polarization classification confusion matrix heat map, (b) is the VV polarization classification confusion matrix heat map, (c) is the HH polarization regression task shape factor scatter density map, (d) is the VV polarization regression task shape factor scatter density map, (e) is the HH polarization regression task scale factor scatter density map, and (f) is the VV polarization regression task scale factor scatter density map. DETAILED DESCRIPTION

[0020] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0021] Example:

[0022] like Figure 1 As shown, this embodiment provides a joint prediction method for sea clutter amplitude distribution type and parameters. By feature engineering and selecting four amplitude distribution types for annotation to construct a data set, combined with the MT1dCNN model, a deep learning model is established that can accurately predict the amplitude distribution type and parameters of sea clutter in the marine environment, solving the problem that a single amplitude distribution characteristic cannot cover the actual complex sea conditions. The specific implementation includes the following steps:

[0023] S1. Build a dataset:

[0024] First, the data is cleaned. In this embodiment, the measured X-band sea clutter data of a certain city in 2022 is collected, including T1 and T2 pulse data of HH polarization and VV polarization, and the influence of targets such as ships and buoys is eliminated to obtain pure sea clutter data. Then, feature engineering is started to extract three long sequence features, namely histogram, probability density function (PDF), and complementary cumulative distribution function (CCDF), from the pure sea clutter data. The number of segments is set to 100, as well as some statistical features. The statistical features include minimum value, maximum value, variance, standard deviation, kurtosis coefficient, skewness coefficient, 25th percentile, 50th percentile, 75th percentile and 90th percentile, a total of 10 short features. Then, the data is labeled, and the Pareto distribution, Log distribution, and probability density function (PDF) are respectively analyzed through the maximum likelihood estimation algorithm. The parameters of four distributions, namely, Log-normal distribution, Weibull distribution, and K distribution, are estimated, and the best fitting distribution type is selected through the ETIC test standard. The distribution type and the corresponding parameters are used as labels for the measured sea clutter data to annotate the HH polarization and VV polarization data sets suitable for joint prediction, which contain about 135,000 samples respectively. The data sets are divided into training set, validation set, and test set. According to the TEIC test standard, in the optimal distribution of HH polarization, Log-normal distribution accounts for 13.3%, Weibull distribution accounts for 18.4%, K distribution accounts for 5%, Pareto distribution accounts for 63.3%, and the proportions of VV polarization are 6.2%, 18%, 22.9%, and 52.9%, respectively.

[0025] S2. Build a joint prediction model of amplitude distribution type and parameters:

[0026] The amplitude distribution type and parameter joint prediction model constructed is the MT1dCNN joint prediction model. Figure 2 As shown in the figure, it includes an input layer, a shared feature extraction layer (one-dimensional convolution layer), a task-private layer (fully connected layer) and an output layer. The input data includes three long sequence features and 10 statistical features. The one-dimensional convolution layer only takes three long sequence features as input. The three long sequence features do not interfere with each other and are independent of each other in the one-dimensional convolution layer. After convolution, each long sequence feature outputs one convolution feature, as shown in Figure 3As shown in the figure, these three convolutional features are concatenated to form a high-level feature, which is then combined with 10 statistical features as the input of the private layer for the classification task. The private layer for the regression task additionally adds the output result of the classification task as the input; the one-dimensional convolutional layer contains one batch normalization, the convolution kernel changes from large to small, and finally the average pooling layer is used to reduce the spatial dimension of the feature. The fully connected layer of each independent task is composed of multiple neurons. The neurons in the layer are not connected, and the neurons between the layers are unidirectionally connected from front to back. The result of the classification task is also passed to the input of the regression task; the activation function of the two adjacent hidden layers in the model selects the rectified linear unit (ReLU). When the eigenvalue of ReLU is relatively large, the gradient is 1, so it will not cause the problem of gradient disappearance. Compared with Sigmoid and hyperbolic tangent activation functions, ReLU can speed up the convergence of the model.

[0027]

[0028] S3. Training amplitude distribution type and parameter joint prediction model:

[0029] Based on the data set in step S1, the amplitude distribution type and parameter joint prediction model are trained through the training set and the validation set. The loss function of the classification task is the cross entropy loss, and the loss function of the regression task is the mean absolute error (MAE). Because the loss of the classification task and the loss of the regression task are not the same in magnitude, different weights are designed for different tasks when designing the overall loss function. Finally, the parameters in the model are trained using the stochastic gradient descent algorithm.

[0030] S4. Test the model and optimize the network structure:

[0031] Based on the test set of step S1, the parameters such as the number of shared layers, the number of independent layers, the size of convolution kernels, and the number of neural networks corresponding to each hidden layer of the deep neural network of the amplitude distribution type and parameter joint prediction model are optimized to optimize the various indicators of the joint prediction model test set, wherein the one-dimensional convolution layer contains 5 hidden layers, the sizes of convolution kernels are 11, 11, 7, 7, and 1, respectively, the number of channels are 256, 128, 64, 32, and 1, respectively, the size of the average pooling convolution kernel is 3, and the independent layers of the classification task and regression task have 2 hidden layers, and the number of neurons in each hidden layer is 128. The final result is that the F1 score of the HH polarization classification task is 97.4%, the RMSE of the regression task is 0.746, the VV polarization F1 score is 96.74%, and the RMSE is 1.071; the comparison diagram of the true value and the predicted value of the sea clutter amplitude distribution type and parameter joint prediction model obtained in this embodiment is as follows Figure 4 shown.

[0032] The algorithms, network structures, etc. not described in detail in the present invention are all common technologies in the field.

[0033] It should be noted that the purpose of publishing the embodiments is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A joint prediction method for sea clutter amplitude distribution type and parameters, characterized in that: The specific steps include: S1. Constructing a data set: Collect X-band sea clutter data, and remove ships and buoys to obtain pure sea clutter data. Extract long sequence features and statistical features from the pure sea clutter data through feature engineering. Estimate the parameters of Pareto distribution, Log-normal distribution, Weibull distribution, and K distribution through the maximum likelihood estimation algorithm. Select the best fitting distribution type through the ETIC test standard, and annotate the distribution type and corresponding parameters as the labels of the measured sea clutter data to obtain a data set. The data set is divided into a training set, a validation set, and a test set. The training set and the validation set are used to construct a joint prediction model of amplitude distribution type and parameters, and the test set is used to test the prediction performance of the model. S2. Build a joint prediction model for amplitude distribution type and parameters: The constructed joint prediction model for amplitude distribution type and parameters includes an input layer, a shared feature extraction layer, a task private layer and an output layer, wherein the shared feature extraction layer uses a one-dimensional convolution layer, the task private layer includes a classification task private layer and a regression task private layer, both of which use a fully connected layer, the one-dimensional convolution layer, the classification task private layer and the regression task private layer all contain hidden layers, the fully connected layers are composed of multiple neurons, the neurons within the layer are not connected, the neurons between the layers are unidirectionally connected from front to back, the results of the classification task are passed to the input of the regression task, the activation function of the two adjacent hidden layers selects the rectified linear unit, the activation function between the last hidden layer in the classification task private layer and the output layer uses the Softmax activation function, and the regression task private layer uses the Linear activation function, the one-dimensional convolution layer takes 3 long sequence features as input, and outputs 1 high-level feature after convolution, the high-level feature is combined with 10 statistical features as the input of the classification task private layer, and the regression task private layer additionally adds the output result of the classification task as input; S3, training amplitude distribution type and parameter joint prediction model: based on the data set of step S1, use the training set and the validation set to train the amplitude distribution type and parameter joint prediction model, wherein the loss function of the classification task is the cross entropy loss, and the loss function of the regression task is the mean absolute error. The parameters in the training model are trained using the stochastic gradient descent algorithm to obtain the trained amplitude distribution type and parameter joint prediction model; S4. Test the model and optimize the network structure: Based on the data set of step S1, use the test set to optimize the number of shared layers, independent layers, convolution kernel size, and the number of neural networks corresponding to each hidden layer in the deep neural network in the amplitude distribution type and parameter joint prediction model trained in step S3, so that the various indicators of the amplitude distribution type and parameter joint prediction model test set are optimal, where the indicators of the classification task include accuracy, precision, recall rate, and F1 score, and the indicators of the regression task include MAE, root mean square error, R 2 value.

2. The joint prediction method of sea clutter amplitude distribution type and parameters according to claim 1, characterized in that: In step S1, the X-band sea clutter data includes T1 and T2 pulse data of HH polarization and VV polarization, the long sequence features include histogram, probability density function, complementary cumulative distribution function, and the statistical features include minimum value, maximum value, variance, standard deviation, kurtosis coefficient, skewness coefficient, 25th percentile, 50th percentile, 75th percentile and 90th percentile.

3. The joint prediction method of sea clutter amplitude distribution type and parameters according to claim 2, characterized in that: In step S2, the gradient of the corrected linear unit is 1 when the eigenvalue is relatively large.

4. The joint prediction method of sea clutter amplitude distribution type and parameters according to claim 3 is characterized in that: The optimal model obtained in step S4 is: the one-dimensional convolution layer contains 5 hidden layers, the convolution kernel sizes are 11, 11, 7, 7, and 1 respectively, the number of channels are 256, 128, 64, 32, and 1 respectively, and the size of the average pooling convolution kernel is 3; the classification task private layer and the regression task private layer have 2 hidden layers, and the number of neurons in each hidden layer is 128.

Images