A land-sea boundary recognition and matching method based on LSGAN

Abstract

This invention discloses a land-sea boundary identification and matching method based on LSGAN, comprising the following steps: Step 1: Constructing a generator and discriminator based on a convolutional neural network; Step 2: Constructing a loss function; Step 3: Model training; Step 4: Constructing a classifier; and Step 5: Land-sea boundary matching. The disclosed method utilizes adversarial training between the generator and the discriminator, using the trained discriminator's convolutional layer as a feature extractor to construct a land-sea clutter classification model. This model exhibits stronger generalization capabilities than traditional deep learning classification models, and the constructed generator can address the challenges of difficult data labeling and poor data diversity.

Description

Technical Field

[0001] The invention belongs to the field of radar data processing, and in particular relates to a method for identifying and matching land-sea boundaries by utilizing skywave radar backscattered fixed-frequency data in the field. Background Art

[0002] As a key strategic / long-range early warning system, skywave over-the-horizon radar (THR) utilizes the reflection of electromagnetic waves between the ionosphere and the land / sea surface to transmit high-frequency energy. Its range is unrestricted by the Earth's curvature, making it the most economical and cost-effective long-range early warning method. The ionosphere serves as the transmission medium for THR, and accurate assessment of ionospheric properties is a key technology for achieving precise detection. Currently, three ionospheric detection methods—vertical, oblique, and backscatter—are used to acquire data for inversion and reconstruction, ultimately determining the ionospheric characteristics of the detection range. Deploying detection stations in distant seas is difficult and costly. Therefore, consideration is being given to using backscatter fixed-frequency detection to identify the land-sea boundary, creating a "quasi-oblique" detection system. This is equivalent to adding oblique detection stations, improving the accuracy of ionospheric reconstruction and, consequently, radar coordinate registration.

[0003] Deep learning-based land-sea boundary identification methods can automatically extract and classify land and sea clutter features. While fixed-frequency return sounding data is readily available in large quantities, the complex characteristics of the ionosphere make these data features extremely complex. Deep learning methods have the ability to automatically extract high-level, abstract features from large amounts of data. Practice has proven that deep learning methods can achieve high-precision classification of land and sea clutter.

[0004] However, deep learning-based land-sea boundary recognition methods require a large number of calibrated land-sea clutter samples, and manual sample calibration is time-consuming and laborious. Furthermore, the small number of labeled samples and their imbalanced distribution lead to poor accuracy and robustness in the land-sea boundary identifier. Generative adversarial networks, as implicit probability density generative models, can generate a large number of land-sea clutter samples to assist in sample calibration and, on the other hand, classify land-sea clutter samples based on a small number of calibrated samples. Therefore, how to utilize a small number of calibrated land-sea clutter samples to develop a robust land-sea boundary identifier based on a generative adversarial network, and how to rapidly align the recognition results with the prior terrain contours and extract the coordinate registration parameters, are both of practical engineering significance. Summary of the Invention

[0005] In order to overcome the problems of small and unbalanced manual calibration samples and improve the robustness and accuracy of the land-sea boundary identifier, the present invention provides a land-sea boundary identification and matching method based on LSGAN.

[0006] The present invention adopts the following technical solutions:

[0007] A land-sea boundary recognition and matching method based on LSGAN is improved in that it includes the following steps:

[0008] Step 1: Build a generator and discriminator based on convolutional neural network:

[0009] Step 1A, build a generator based on convolutional neural network:

[0010] The generator uses the data fitting capability of the convolutional network to learn the mathematical distribution of the backscattered fixed-frequency data by establishing a mapping relationship between the one-dimensional noise distribution and the real data distribution, ultimately achieving data generation by the model.

[0011] The generator is represented by a differentiable function G, the input z is a random noise signal that obeys Gaussian distribution or uniform distribution, and the output is the land and sea clutter generated data G(z);

[0012] The generator uses a deep convolutional neural network as the basic component of the network, and uses four convolutional layers and four upsampling layers as the basic network structure of the generator. The four convolutional layers are used to reconstruct the detailed features of the data; the four upsampling layers are used to expand the dimension of the feature vector.

[0013] The LeakyReLU function is used as the activation function in each convolutional layer. The LeakyReLU function is defined as:

[0014]

[0015] In the above formula, x is the real data, negative_slope is the weight coefficient;

[0016] The Tanh function is used as the output layer. The Tanh function is defined as:

[0017]

[0018] In the above formula, x is real data and e is a natural constant;

[0019] Step 1B, build a discriminator based on convolutional neural network:

[0020] The discriminator is represented by D. The input of the discriminator is the generated data G(z) and the real data x, and the output is the probability of judging whether the input data is the real data. The process consists of two scenarios. In the first scenario, the real data x is sampled from the real training data as the input of the discriminator D. The discriminator D outputs a probability value between 0 and 1, indicating the probability that the real data x belongs to the real sample. The closer the probability value is to 1, the higher the probability of the real data, and vice versa. The higher the probability of the generated data; in the second scenario, the input z is sampled from a prior distribution, and G(z) is used as the input of the discriminator D. In this scenario, both the discriminator D and the generator G are involved. The goal of the discriminator D is to make the probability value of the output D(G(z)) close to 0 and the probability value of D(x) close to 1;

[0021] A deep convolutional neural network is used as the basic framework of the discriminator. The basic network structure of the discriminator is an input layer-three or more convolution layers-pooling layer stacking structure-output layer. The three or more convolution layers are used to fully extract multi-level features. The LeakyReLU function is used as the activation function between the convolution layers. The pooling layer reduces the dimension of the feature vector, reduces the amount of calculation, corrects overfitting, and extracts high-level features. The output layer is a fully connected layer. This layer uses the Sigmoid activation function to realize the two functions of distinguishing true from false. The Sigmoid activation function can map the output to the interval [0,1] as the true or false output.

[0022]

[0023] In the above formula, x is real data and e is a natural constant;

[0024] Step 2: Construct the loss function as follows:

[0025]

[0026]

[0027] In the above formula, x is the true distribution p data (x) is the real data, z is the data that follows the Gaussian distribution p z (z) is the random noise, a and b represent the labels judged as false and true respectively;

[0028] Step 3, model training:

[0029] Step 3A, land and sea clutter sample preprocessing:

[0030] First, normalize the real data to obtain the maximum value max(x) and minimum value min(x) of the real data, and map the data to the range [-1,1] using the following formula:

[0031]

[0032] In the above formula, x * is the normalized data, x is the real data,

[0033] Step 3B, network parameter configuration:

[0034] The batch size is 2 n , n is a natural number; the learning rate is 10 -6 ~1; Adma optimizer is used to update the weights and biases of the generator and discriminator;

[0035] Step 3C, network training:

[0036] The network is trained according to the loss function constructed in step 2:

[0037] The discriminator inputs real data to make the true and false identification results as close to 1 as possible, calculates the loss function of the discriminator, and backpropagates;

[0038] The discriminator inputs generated data to make the true and false identification results as close to 0 as possible, calculates the discriminator's loss function, and backpropagates;

[0039] The generator inputs random noise to obtain generated data. The generated data passes through the discriminator to obtain the authenticity identification result, making it as close to 1 as possible. The loss function of the generator is calculated and back-propagated;

[0040] Let the random noise be z, the generator function be G, its input is z, and the parameter to be learned is θ (G) ; The function of the discriminator is D, its input is the real data x, and the parameter to be learned is θ (D) , D needs to update θ (D) Minimize its objective function J (D) , G needs to update θ (G) Minimize its objective function J (G) , the two networks compete with each other, and the loss functions of both networks depend on each other. The goal of network training is to achieve Nash equilibrium, which refers to a pair of parameters θ (D) ,θ (G) , so that θ (D) It's J (D) A minimum point of θ (G) It's J (G) A minimum point of ;

[0041] Step 4: Build a classifier:

[0042] Step 4A, build feature extractor:

[0043] Train the LSGAN land and sea clutter sample generation model from step 3 until the loss is stable. Separate the trained discriminator and remove the last layer of the discriminator. Only the trained convolutional layer is separated and used as the feature extractor for the land and sea clutter classifier.

[0044] Step 4B, build the classifier:

[0045] Connect the feature extractor to the classification layer used to classify land and sea clutter samples to build a land and sea clutter classifier. Finally, fix the parameters of the convolutional layer and fine-tune the classification layer using real land and sea clutter samples.

[0046] Step 5: Land-sea boundary matching:

[0047] After using the land-sea clutter classifier to identify the land and sea attributes of the return scattered fixed-frequency data, the land-sea boundary recognition results and the prior geographic information map are used as inputs to the matching algorithm. The land-sea boundary recognition results are traversed in the prior geographic information map and the matching error is calculated at the same time. Finally, the land-sea boundary matching result is output.

[0048] Assume that the land-sea boundary recognition result is T(m,n), m is the number of distance units of the land-sea boundary recognition result, n is the number of orientation units of the land-sea boundary recognition result, and the prior geographic information map is S(w,h), w is the number of distance units of the prior geographic information map, h is the number of orientation units of the prior geographic information map, and the sub-area where the land-sea boundary recognition result and the prior geographic information map overlap is defined as S i,j , i, j are the prior geographic information map coordinates, and the following formula is used to measure T(m,n) and S i,j Similarity D(i,j):

[0049]

[0050] Normalize it and get the correlation coefficient R(i,j) of template matching:

[0051]

[0052] When the land-sea boundary identification result is the same as the sub-region, the correlation coefficient R(i,j) = 1. After completing all searches in the prior geographic information map, find the maximum value of the correlation coefficient R max (i, j), its corresponding sub-region S i,j This is the matching result.

[0053] The beneficial effects of the present invention are:

[0054] The method disclosed in this paper uses adversarial training between a generator and a discriminator, using the trained discriminator's convolutional layer as a feature extractor to construct a land and sea clutter classification model. This model has stronger generalization capabilities than traditional deep learning classification models, and the constructed generator can address the challenges of difficult data labeling and poor data diversity.

[0055] The method disclosed in this paper improves the loss function of traditional generative adversarial networks. Land and sea clutter samples are backscattered constant-frequency data with a high degree of randomness, which makes the training process of traditional generative adversarial networks very unstable. This is largely due to the loss function, especially the possibility of vanishing gradients when minimizing the loss function, making it difficult to update the generator. This paper proposes an LSGAN based on the least squares loss function to update network parameters, which effectively solves the training instability problem.

[0056] The method disclosed in the present invention, based on the above-mentioned land-sea boundary recognition result, spatially aligns the foreground map formed by the recognition result and the prior geographic information map, so as to determine the relative translation process between the two images and obtain the translation parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is an overall flow chart of the method disclosed in the present invention;

[0058] Figure 2 is a schematic diagram of the generated land clutter samples;

[0059] Figure 3 is a schematic diagram of the generated sea clutter samples;

[0060] Figure 4 This is a schematic diagram of the land-sea boundary matching results. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0062] Example 1. This example discloses a land-sea boundary identification and matching method based on LSGAN. This method addresses the robustness and low accuracy issues of land-sea boundary identification. Based on a least-squares loss function, an LSGAN land-sea clutter sample generation model is constructed. Based on the trained LSGAN model, the discriminator feature layer is split to serve as a land-sea clutter sample feature extractor. This feature extractor is then connected to the classification layer and fine-tuned to obtain a land-sea clutter sample classifier, which is then used for land-sea boundary identification. Template matching is used to match the land-sea boundary identification results with prior geographic information. Finally, coordinate registration parameters are extracted based on the matching results.

[0063] The specific steps include:

[0064] Step 1: Build a generator and discriminator based on convolutional neural network:

[0065] Step 1A, build a generator based on convolutional neural network:

[0066] The generator leverages the powerful data fitting capabilities of convolutional networks to learn the mathematical distribution of backscattered fixed-frequency data by establishing a mapping relationship between one-dimensional noise distribution and real data distribution. This ultimately enables the model to generate large amounts of reliable data.

[0067] The generator is represented by a differentiable function G, the input z is a 100-dimensional Gaussian noise sequence, and the output is the land and sea clutter generated data G(z);

[0068] The generator uses a deep convolutional neural network as the basic component of the network, and adopts four convolutional layers and four upsampling layers as the basic network structure of the generator. The four convolutional layers are used to reconstruct the detailed features of the data; the four upsampling layers are used to expand the dimension of the feature vector to increase the feature map.

[0069] The LeakyReLU function is used as the activation function in each convolutional layer. The LeakyReLU function is defined as:

[0070]

[0071] In the above formula, x is the real data, negative_slope is the weight coefficient, which is set to 0.2;

[0072] The Tanh function is used as the output layer. The Tanh function is defined as:

[0073]

[0074] In the above formula, x is real data and e is a natural constant;

[0075] Step 1B, build a discriminator based on convolutional neural network:

[0076] The discriminator is represented by D. The input of the discriminator is the generated data G(z) and the real data x, and the output is the probability of judging whether the input data is the real data. The process consists of two scenarios. In the first scenario, the real data x is sampled from the real training data as the input of the discriminator D. The discriminator D outputs a probability value between 0 and 1, indicating the probability that the real data x belongs to the real sample. The closer the probability value is to 1, the higher the probability of the real data, and vice versa. The higher the probability of the generated data; in the second scenario, the input z is sampled from a prior distribution, and G(z) is used as the input of the discriminator D. In this scenario, both the discriminator D and the generator G are involved. The goal of the discriminator D is to make the probability value of the output D(G(z)) close to 0 and the probability value of D(x) close to 1;

[0077] The discriminator D plays a two-player zero-sum game with the generator G, and the main goal of D is to determine whether the input is a real sample and provide an update feedback mechanism for G.

[0078] A deep convolutional neural network is used as the basic framework of the discriminator. The basic network structure of the discriminator is an input layer-three or more convolution layers-pooling layer stacking structure-output layer. The three or more convolution layers are used to fully extract multi-level features. The LeakyReLU function is used as the activation function between the convolution layers. The pooling layer reduces the dimension of the feature vector, reduces the amount of calculation, corrects overfitting, and extracts high-level features. The output layer is a fully connected layer. This layer uses the Sigmoid activation function to realize the two functions of distinguishing true from false. The Sigmoid activation function can map the output to the interval [0,1] as the true or false output.

[0079]

[0080] In the above formula, x is real data and e is a natural constant;

[0081] Step 2: Construct the loss function as follows:

[0082] Land and sea clutter samples are backscattered constant-frequency data with a high degree of randomness, making the training process of traditional generative adversarial networks very unstable. This is largely due to the fact that the loss function, especially when minimizing the loss function, may experience vanishing gradients, making it difficult to update the generator.

[0083] LSGAN uses the least squares loss to replace the original Sigmoid loss, and the loss function is defined as follows:

[0084]

[0085]

[0086] In the above formula, x is the true distribution p data (x) is the real data, z is the data that follows the Gaussian distribution p z (z) Random noise, a = 0 and b = 1 represent the labels judged as false and true, respectively;

[0087] Due to the unique characteristics of L2 (least squares) distance, when land and sea clutter sample data deviates from the target, a penalty proportional to the deviation distance is applied, which then brings it back to the target, preventing the data from deviating further and further. This means that the generated land and sea clutter samples are closer to the decision boundary. This makes the network training process more stable and allows the generator and discriminator to be updated in a direction with more favorable performance.

[0088] Step 3, model training:

[0089] Step 3A, land and sea clutter sample preprocessing:

[0090] During the training process of the two aforementioned models, the generator continuously generates fake samples that approximate the real data distribution to deceive the discriminator. The discriminator distinguishes between the generated fake samples and real samples during the learning process. A small number of land and sea clutter samples are sampled and calibrated from the land and sea clutter sample database, with 10,000 sea clutter samples and 10,000 land clutter samples. 70% of these samples, or 7,000 sea clutter samples and 7,000 land clutter samples, are used to train the LSGAN. The remaining 30% of these samples, or 3,000 sea clutter samples and 3,000 land clutter samples, are used to fine-tune the classification layer of the classifier.

[0091] First, normalize the real data and preprocess it to obtain the maximum value max(x) = 98.58 and the minimum value min(x) = -54.12 of the real data. Then map the data to the range [-1, 1] using the following formula:

[0092]

[0093] In the above formula, x * is the normalized data, x is the real data, max(x) is the maximum value in x, and min(x) is the minimum value in x;

[0094] Step 3B, network parameter configuration:

[0095] The batch size determines the direction of gradient descent. If the batch size is too small, the gradient update becomes more random, making it difficult for the network to converge. Taking the extreme case of a batch size of 1, the direction of each gradient correction is adjusted based on the direction of the gradient of the individual sample, resulting in severe gradient update fluctuations and difficulty in convergence. Furthermore, a small batch size reduces memory utilization, leading to a very low training rate for adversarial networks. Increasing the batch size can improve the accuracy of gradient descent and alleviate gradient update fluctuations. However, beyond a certain size, the direction of gradient descent becomes largely unchanged. Furthermore, an excessively large batch size consumes more memory, which can easily lead to memory overflows and training failures when the land / sea clutter sample size is large. Therefore, considering the above factors, a batch size of 64 was chosen as a good choice.

[0096] The learning rate determines the convergence and convergence rate of the loss function. If the learning rate is too high, the loss gradient may oscillate violently around the minimum, making it difficult for the network to converge. Reducing the learning rate helps stabilize training, but too low a learning rate will slow the rate of descent of the loss gradient and increase convergence time. A learning rate of 0.0002 has been found to work well.

[0097] The momentum-based Adma algorithm introduces first-order and second-order gradient moment estimates and uses historical gradient information to dynamically adjust the parameter learning rate. After bias correction, the learning rate remains within a defined range, thus stabilizing parameter learning. The Adma optimizer is proposed to update the weights and biases of the generator and discriminator. The exponential decay rates of the first-order and second-order moment estimates are set to 0.5 and 0.999, respectively.

[0098] Step 3C, network training:

[0099] The network is trained according to the loss function constructed in step 2:

[0100] The discriminator inputs real data to make the true and false identification results as close to 1 as possible, calculates the loss function of the discriminator, and backpropagates;

[0101] The discriminator inputs generated data to make the true and false identification results as close to 0 as possible, calculates the discriminator's loss function, and backpropagates;

[0102] The generator inputs random noise to obtain generated data. The generated data passes through the discriminator to obtain the authenticity identification result, making it as close to 1 as possible. The loss function of the generator is calculated and back-propagated;

[0103] The generator function is G, its input is a 100-dimensional Gaussian noise sequence, and the parameter to be learned is θ (G) ; The function of the discriminator is D, its input is the real data x, and the parameter to be learned is θ (D) , D needs to update θ(D) Minimize its objective function J (D) , G needs to update θ (G) Minimize its objective function J (G) , the two networks compete with each other, and the loss functions of both networks depend on each other. The goal of network training is to achieve Nash equilibrium, which refers to a pair of parameters θ (D) ,θ (G) , so that θ (D) It's J (D) A minimum point of θ (G) It's J (G) A minimum point of ;

[0104] Step 4: Build a classifier:

[0105] Step 4A, build feature extractor:

[0106] Train the LSGAN land and sea clutter sample generation model from step 3 until the loss stabilizes. Separate the trained discriminator and remove the last layer of the discriminator, i.e., the classification layer that distinguishes true from false. Only the trained convolutional layer is separated and used as the feature extractor for the land and sea clutter classifier.

[0107] Step 4B, build the classifier:

[0108] Connect the feature extractor to the classification layer used to classify land and sea clutter samples to build a land and sea clutter classifier. Finally, fix the parameters of the convolutional layer and fine-tune the classification layer using real land and sea clutter samples.

[0109] Step 5: Land-sea boundary matching:

[0110] After using the land-sea clutter classifier to identify the land and sea attributes of the return scattered fixed-frequency data, the land-sea boundary recognition result (foreground map) and the prior geographic information map (background map) are used as inputs to the matching algorithm. The land-sea boundary recognition result is traversed in the prior geographic information map and the matching error is calculated at the same time. Finally, the land-sea boundary matching result is output.

[0111] Assume that the land-sea boundary recognition result is T(m,n), where m=1, n=1760, and the prior geographic information map is S(w,h), where w=5000, h=5000. The sub-area where the land-sea boundary recognition result and the prior geographic information map overlap is defined as S i,j , i, j are the prior geographic information map coordinates, and the following formula is used to measure T(m,n) and S i,j Similarity D(i,j):

[0112]

[0113] Normalize it and get the correlation coefficient R(i,j) of template matching:

[0114]

[0115] When the land-sea boundary identification result is the same as the sub-region, the correlation coefficient R(i,j) = 1. After completing all searches in the prior geographic information map, find the maximum value of the correlation coefficient R max (i, j), its corresponding sub-region S i,j This is the matching result.

[0116] In order to verify the effectiveness of the present invention, the following experiments are conducted. The overall process of the land-sea boundary recognition and matching method based on LSGAN is as follows: Figure 1 As shown in the figure, according to the above steps, a small amount of existing backscattered fixed-frequency data is used as the input of the model, and the LSGAN is iteratively trained to obtain a generative model. The trained generator is split out and a 100-dimensional Gaussian noise sequence is input to generate a large number of land and sea clutter samples. The schematic diagrams of the generated land clutter samples and sea clutter samples are shown in Figure 1. Figure 2 and Figure 3 As shown. The convolution layer of the trained discriminator is split out as a feature extractor, and the classification layer is connected as a classifier. Then the convolution layer parameters are fixed, and the classification layer is fine-tuned with the calibrated backscattered fixed-frequency data to obtain the final classifier. The classification accuracy is 99.21% when tested with the calibrated test data. The classification accuracy of the classifier without fixed convolution layer parameters is 95.35% when trained and tested. Furthermore, the land-sea boundary formed by the recognition result is matched with the real geographic information through the land-sea boundary, and the average accuracy is 97.52%. Figure 4 This is a schematic diagram before and after matching the land-sea boundary. Through the schematic diagram, we can clearly see that the land-sea boundary is accurately matched to the real geographic information map.

Claims

1. A land-sea boundary recognition and matching method based on LSGAN, characterized by: The steps include: Step 1: Build a generator and discriminator based on convolutional neural network: Step 1A, build a generator based on convolutional neural network: The generator uses the data fitting capability of the convolutional network to learn the mathematical distribution of the backscattered fixed-frequency data by establishing a mapping relationship between the one-dimensional noise distribution and the real data distribution, ultimately achieving data generation by the model. The generator is represented by a differentiable function G, the input z is a random noise signal that obeys Gaussian distribution or uniform distribution, and the output is the land and sea clutter generated data G(z); The generator uses a deep convolutional neural network as the basic component of the network, and uses four convolutional layers and four upsampling layers as the basic network structure of the generator. The four convolutional layers are used to reconstruct the detailed features of the data; the four upsampling layers are used to expand the dimension of the feature vector. The LeakyReLU function is used as the activation function in each convolutional layer. The LeakyReLU function is defined as: In the above formula, x is the real data, negative_slope is the weight coefficient; The Tanh function is used as the output layer. The Tanh function is defined as: In the above formula, x is real data and e is a natural constant; Step 1B, build a discriminator based on convolutional neural network: The discriminator is represented by D. The input of the discriminator is the generated data G(z) and the real data x, and the output is the probability of judging whether the input data is the real data. The process consists of two scenarios. In the first scenario, the real data x is sampled from the real training data as the input of the discriminator D. The discriminator D outputs a probability value between 0 and 1, indicating the probability that the real data x belongs to the real sample. The closer the probability value is to 1, the higher the probability of the real data, and vice versa. The higher the probability of the generated data; in the second scenario, the input z is sampled from a prior distribution, and G(z) is used as the input of the discriminator D. In this scenario, both the discriminator D and the generator G are involved. The goal of the discriminator D is to make the probability value of the output D(G(z)) close to 0 and the probability value of D(x) close to 1; A deep convolutional neural network is used as the basic framework of the discriminator. The basic network structure of the discriminator is an input layer-three or more convolution layers-pooling layer stacking structure-output layer. The three or more convolution layers are used to fully extract multi-level features. The LeakyReLU function is used as the activation function between the convolution layers. The pooling layer reduces the dimension of the feature vector, reduces the amount of calculation, corrects overfitting, and extracts high-level features. The output layer is a fully connected layer. This layer uses the Sigmoid activation function to realize the two functions of distinguishing true from false. The Sigmoid activation function can map the output to the interval [0,1] as the true or false output. In the above formula, x is real data and e is a natural constant; Step 2: Construct the loss function as follows: In the above formula, x is the true distribution p data (x) is the real data, z is the data that follows the Gaussian distribution p z (z) is the random noise, a and b represent the labels judged as false and true respectively; Step 3, model training: Step 3A, land and sea clutter sample preprocessing: First, normalize the real data to obtain the maximum value max(x) and minimum value min(x) of the real data, and map the data to the range [-1,1] using the following formula: In the above formula, x * is the normalized data, x is the real data, Step 3B, network parameter configuration: The batch size is 2 n , n is a natural number; the learning rate is 10 -6 ~1; Adma optimizer is used to update the weights and biases of the generator and discriminator; Step 3C, network training: The network is trained according to the loss function constructed in step 2: The discriminator inputs real data to make the true and false identification results as close to 1 as possible, calculates the loss function of the discriminator, and backpropagates; The discriminator inputs generated data to make the true and false identification results as close to 0 as possible, calculates the discriminator's loss function, and backpropagates; The generator inputs random noise to obtain generated data. The generated data passes through the discriminator to obtain the authenticity identification result, making it as close to 1 as possible. The loss function of the generator is calculated and back-propagated; Let the random noise be z, the generator function be G, its input is z, and the parameter to be learned is θ (G) ; The function of the discriminator is D, its input is the real data x, and the parameter to be learned is θ (D) , D needs to update θ (D) Minimize its objective function J (D) , G needs to update θ (G) Minimize its objective function J (G) , the two networks compete with each other, and the loss functions of both networks depend on each other. The goal of network training is to achieve Nash equilibrium, which refers to a pair of parameters θ (D) ,θ (G) , so that θ (D) It's J (D) A minimum point of θ (G) It's J (G) A minimum point of ; Step 4: Build a classifier: Step 4A, build feature extractor: Train the LSGAN land and sea clutter sample generation model from step 3 until the loss is stable. Separate the trained discriminator and remove the last layer of the discriminator. Only the trained convolutional layer is separated and used as the feature extractor for the land and sea clutter classifier. Step 4B, build the classifier: Connect the feature extractor to the classification layer used to classify land and sea clutter samples to build a land and sea clutter classifier. Finally, fix the parameters of the convolutional layer and fine-tune the classification layer using real land and sea clutter samples. Step 5: Land-sea boundary matching: After using the land-sea clutter classifier to identify the land and sea attributes of the return scattered fixed-frequency data, the land-sea boundary recognition results and the prior geographic information map are used as inputs to the matching algorithm. The land-sea boundary recognition results are traversed in the prior geographic information map and the matching error is calculated at the same time. Finally, the land-sea boundary matching result is output. Assume that the land-sea boundary recognition result is T(m,n), m is the number of distance units of the land-sea boundary recognition result, n is the number of orientation units of the land-sea boundary recognition result, and the prior geographic information map is S(w,h), w is the number of distance units of the prior geographic information map, h is the number of orientation units of the prior geographic information map, and the sub-area where the land-sea boundary recognition result and the prior geographic information map overlap is defined as S i,j , i, j are the prior geographic information map coordinates, and the following formula is used to measure T(m,n) and S i,j Similarity D(i,j): Normalize it and get the correlation coefficient R(i,j) of template matching: When the land-sea boundary identification result is the same as the sub-region, the correlation coefficient R(i,j) = 1. After completing all searches in the prior geographic information map, find the maximum value of the correlation coefficient R max (i, j), its corresponding sub-region S i,j This is the matching result.

Images