A land-sea clutter classification method for multi-scale data

Abstract

The present invention discloses a method for classifying ground and sea clutter that can be used for multi-scale data. The method comprises the following steps: Step 1: constructing a multi-scale ground and sea clutter database and forming a training dataset and a test dataset for the model; Step 2: constructing a multi-scale ground and sea clutter classification model based on a bifurcating residual neural network; Step 3: training the multi-scale ground and sea clutter classification model based on the bifurcating residual neural network constructed in Step 2 using the training dataset; Step 4: testing the trained multi-scale ground and sea clutter classification model based on the bifurcating residual neural network using the test dataset. The multi-scale ground and sea clutter classification model based on the bifurcating residual neural network that has been tested and meets the project accuracy requirements can be used to classify ground and sea clutter in multi-scale data. The method solves the current problems of multi-scale ground and sea clutter classification in skywave radars, as well as the inability to train corresponding classifiers due to insufficient numbers of ground and sea clutter at certain scales.

Description

Technical Field

[0001] The present invention belongs to the field of spectrum data processing in the field of radar data processing, and relates to a ground and sea clutter classification method that can be used for multi-scale data. Technical Background

[0002] Skywave over-the-horizon radar (OTHR) uses electromagnetic waves reflected in the ionosphere to transmit high-frequency energy for target detection. Its effective range is not limited by the curvature of the earth and can achieve long-range early warning of high-value targets such as stealth fighters, intercontinental missiles, cruise missiles, and large and medium-sized ships. The propagation of electromagnetic waves through the ionosphere will cause a coordinate system conversion problem. The target position is defined in the geographic coordinate system (i.e., longitude and latitude), and the measurement signal received by the skywave radar is defined in the radar coordinate system (i.e., slant range and azimuth). In order to obtain the actual position of the detected target, it is necessary to convert the ray distance and azimuth measurement values ​​into the longitude and latitude coordinate values ​​of the ground. This is the coordinate registration (CR) process. The coordinate registration process can be divided into two sub-processes: (1) selecting the correct ionospheric propagation mode for each measured target echo estimate; (2) converting the state estimate of the measured target from the radar coordinate system to the earth coordinate system. However, due to the influence of the skywave radar working mechanism and its radio wave environment, it is difficult to perform coordinate registration by correctly matching the target echo-propagation mode.

[0003] The skywave radar system primarily consists of a main radar and an ionospheric sounding subsystem. The ionospheric sounding subsystem provides the main radar with ionospheric propagation condition assessment parameters and coordinate registration parameters. However, the sounding subsystem cannot be deployed in certain critical areas, such as offshore or in hostile areas, where accurate coordinate registration coefficients cannot be obtained. Furthermore, because the ionospheric sounding subsystem operates independently of the main radar, the coordinate registration parameters it provides are inconsistent with those of the main radar.

[0004] Another method is to use beacons for coordinate registration. Beacons can be divided into active beacons and passive beacons. Active beacons are signal transmitters that can actively transmit signals, while passive beacons are islands in offshore areas or near-shore coastlines. When using active beacons for coordinate registration, a skywave radar receives the signal sent by the beacon, outputs the ray distance and azimuth measurement value at its position in the radar coordinate system, and compares it with the position in the geodetic coordinate system provided by GPS to obtain the coordinate registration coefficient. However, the deployment area of ​​active beacons is limited and the beacons need to be maintained. When using passive beacons for coordinate registration, the coordinate registration coefficients can be extracted using geographic information or ground and sea clutter. In our engineering project, ground and sea clutter is huge and easy to obtain, so the present invention chooses to use ground and sea clutter to extract the coordinate registration coefficients. Due to the different operating parameters and work tasks of the radar, ground and sea clutter will have different coherent accumulation points and signal time widths, so ground and sea clutter has multi-scale characteristics. Skywave radar ground and sea clutter of different scales have similar characteristics. Sea clutter has Bragg peaks on both sides of zero frequency, and ground clutter has a peak at zero frequency. However, ground and sea clutter of different scales have differences in amplitude, Bragg peak spacing and other characteristics. Examples of ground and sea clutter of different scales are as follows: Figures 1 to 6 .

[0005] Coordinate registration is performed using land and sea clutter recognition. First, the land and sea conditions at each range and azimuth resolution unit are identified by analyzing the land and sea clutter. Then, the identified land and sea contours are matched with the prior map contour information, and the deviation of the matching results is used as the coordinate registration coefficient.

[0006] Deep learning, a key branch of machine learning, is widely used in various fields of radar signal processing, such as clutter identification, target detection, and target classification, due to its powerful feature learning capabilities. Currently, deep learning is also gaining popularity in processing skywave radar spectrum data. This paper constructs a multi-layer residual neural network, extracting different feature layers to identify ground and sea clutter of different scales. The low-level feature layers identify small-scale ground and sea clutter, while the high-level feature layers identify large-scale ground and sea clutter.

[0007] Skywave radar spectrum data contains an uneven number of ground and sea clutter at different scales. The number of ground and sea clutter at one scale may be much smaller than at other scales, and training a model solely for that scale cannot achieve the required accuracy. This paper utilizes transfer learning. When a ground and sea clutter dataset for a specific scale is insufficient, the model is pre-trained using ground and sea clutter at a sufficient scale. The trained model is then fine-tuned or frozen, and all or part of the model layers are retrained using a small amount of new dataset, achieving better accuracy for classification tasks. Summary of the Invention

[0008] The purpose of the present invention is to provide a ground and sea clutter classification method that can be used for multi-scale data, so as to solve the current problem of multi-scale ground and sea clutter classification in skywave radar and the problem that the number of ground and sea clutter at a certain scale is insufficient, resulting in the inability to train the corresponding classifier.

[0009] The technical solution of the present invention is: a ground and sea clutter classification method applicable to multi-scale data, comprising the following steps:

[0010] Step 1: Construct a multi-scale land and sea clutter database and form a training dataset and a test dataset for the model;

[0011] Step 2: Construct a multi-scale land and sea clutter classification model based on a bifurcated residual neural network;

[0012] Step 3: using the training data set to train the multi-scale land and sea clutter classification model based on the bifurcated residual neural network constructed in step 2;

[0013] Step 4: Use the test data set to test the multi-scale ground and sea clutter classification model based on the bifurcated residual neural network that has been trained. The multi-scale ground and sea clutter classification model based on the bifurcated residual neural network that has been tested and meets the project accuracy requirements can be used to classify the ground and sea clutter of the multi-scale data.

[0014] Furthermore, the specific content of step one is:

[0015] Step S1.1, analyzing the spectrum data obtained by skywave radar detection;

[0016] Step S1.2, selecting multi-scale ground and sea clutter data that meet the project requirements, wherein the ground clutter zero-frequency peak and the sea clutter zero-frequency Bragge peaks on both sides of the peak are obvious;

[0017] Step S1.3, extracting the frequency range [-5 Hz, 5 Hz] in the feature set from the ground and sea clutter data selected in step S1.2;

[0018] Step S1.4: Randomly select 70% of the multi-scale land and sea clutter data processed as above as a training data set and 30% as a test data set.

[0019] Furthermore, in step 3, the specific content of using the training data set to train the multi-scale land and sea clutter classification model based on the bifurcated residual neural network is as follows:

[0020] Step S3.1: Using a sufficient number of scaled ground and sea clutter, the multi-scale ground and sea clutter classification model based on a bifurcated residual neural network constructed in Step 2 is trained. The parameters of all layers of the output channel n corresponding to the ground and sea clutter with scale label n (n≠N-1) are frozen. If n=N-1, the parameters of the N-1th feature extraction layer and all layers of the N-1th output channel are frozen, and only the remaining layers are trained.

[0021] Step S3.2: Use all scales of ground and sea clutter to train the multi-scale ground and sea clutter classification model based on a bifurcated residual neural network processed in step S3.1. Set all the parameters frozen in step S3.1 to the first group of parameters, and set the parameters of the remaining layers of the model to the second group. Set different training weights for different groups, set a smaller learning rate for fine-tuning the second group, and the learning rate of the first group is 5 times that of the second group.

[0022] The present invention has the following beneficial effects: To address the problem of recognizing multi-scale ground and sea clutter data using skywave over-the-horizon radar, the present invention proposes an improved residual neural network model—a bifurcated residual neural network. This model, based on the idea of ​​using feature maps at different levels in the SSD model to detect targets of different scales, uses feature maps at different levels to predict ground and sea clutter data of different scales, enabling the model to recognize multi-scale ground and sea clutter data. Separate output channels are designed for data of different scales, and the model is trained step by step based on a transfer learning mechanism. Parameters of different parts of the model are divided into different parameter groups, and different learning rates are set. When the amount of data on a particular scale is insufficient, the model parameters obtained from training with multi-scale data can be used as pre-training parameters to improve the recognition accuracy of data on a smaller scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is the ground clutter waveform diagram with 256 coherent accumulation points mentioned in the background technology;

[0024] Figure 2 This is the sea clutter waveform diagram with 256 coherent accumulation points mentioned in the background technology;

[0025] Figure 3 This is the ground clutter waveform diagram with 512 coherent accumulation points mentioned in the background technology;

[0026] Figure 4 This is the sea clutter waveform diagram with 512 coherent accumulation points mentioned in the background technology;

[0027] Figure 5 This is the ground clutter waveform diagram with 1024 coherent accumulation points mentioned in the background technology;

[0028] Figure 6 This is the sea clutter waveform diagram with 1024 coherent accumulation points mentioned in the background technology;

[0029] Figure 7 Schematic diagram of the structure of a multi-scale land and sea clutter classification model based on a bifurcated residual neural network constructed in a land and sea clutter classification method that can be used for multi-scale data of the present invention;

[0030] Figure 8 This is a graph showing the change pattern of loss values ​​at different scales during the first step of training the network in an embodiment of the present invention;

[0031] Figure 9 This is a graph showing the variation of classification accuracy at different scales during the first step of network training in an embodiment of the present invention;

[0032] Figure 10 This is a graph showing the change pattern of loss values ​​at different scales during the second step of training the network in an embodiment of the present invention;

[0033] Figure 11 This is a diagram showing the variation of classification accuracy at different scales during the second step of network training in an embodiment of the present invention. Specific implementation plan

[0034] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] The present invention provides a ground-sea clutter classification method applicable to multi-scale data, comprising the following contents:

[0036] Step 1: Construct a multi-scale land and sea clutter database and form a training dataset and a test dataset for the model;

[0037] Step S1.1: Analyze the spectrum data obtained by skywave radar detection:

[0038] Skywave radar continuously detects during operation. To save storage space, the spectrum data corresponding to different times, sub-areas, distance units, and detection frequencies of the skywave radar are compressed and stored. Before use, the compressed data needs to be decompressed and the spectrum data of different scales needs to be saved separately.

[0039] Step S1.2: Select multi-scale ground and sea clutter data that meet the project requirements. The ground clutter zero-frequency peak and the Bragge peaks on both sides of the sea clutter zero-frequency that meet the project requirements are obvious.

[0040] During the operation of skywave radar, due to reasons such as ionospheric instability or abnormal radar working status, the spectrum data may experience frequency shift, broadening, radio frequency interference, and be submerged by clutter. These abnormal spectrum data cannot be used as training data for land and sea clutter classification experiments. It is necessary to select data with obvious land and sea clutter characteristics as a multi-scale land and sea clutter dataset.

[0041] Step S1.3: Extract the frequency range [-5 Hz, 5 Hz] in the feature set from the ground and sea clutter data selected in step S1.2:

[0042] The main identification features of ground and sea clutter are the peak of ground clutter at zero frequency and the Bragg peaks of sea clutter on both sides of zero frequency. They account for a relatively small proportion of the entire spectrum data. The present invention selects spectrum data in the frequency interval [-5 Hz, +5 Hz], which helps reduce computational costs and prevents overfitting to a certain extent.

[0043] Step S1.4: The multi-scale ground and sea clutter processed as described above is processed into a training data set and a test data set for the model. The data set includes spectrum data, a ground and sea clutter classification label corresponding to each spectrum data (0 indicates that the spectrum data is sea clutter, and 1 indicates that the spectrum data is ground clutter), and a coherent accumulation point classification label corresponding to each spectrum data. 70% of the data is randomly selected as training data, and 30% of the data is randomly selected as test data.

[0044] Step 2: Construct a multi-scale land and sea clutter classification model based on a bifurcated residual neural network:

[0045] Aiming at the problem of multi-scale land and sea clutter data classification, the present invention proposes an improved residual neural network - a bifurcated residual neural network model. Step 2 is a specific method for constructing the bifurcated residual neural network model.

[0046] Skywave radar may obtain N scales of land and sea clutter data due to different working tasks and working parameters. The scale labels 0, 1, 2, ..., N-1 are set for the N scales of data from small to large. The length of land and sea clutter data of scale n (n = 0, 1, 2, ..., N-1) is T n (n=0,1,2,…,N-1). Since the neural network requires the input data to be of the same scale, we keep the scale N-1 data unchanged and fill the scale 0,1,2,…,N-2 data with zeros so that their length is T N-1 .

[0047] Inspired by the idea of ​​using feature maps of different levels based on the SSD network model to detect targets of different scales, the present invention designs a bifurcated residual neural network residual network model, which uses feature maps of different levels to predict ground and sea clutter data of different scales. Low-level feature maps predict small-scale ground and sea clutter data, and high-level feature maps predict large-scale ground and sea clutter data. At the same time, ground and sea clutter data of N scales are identified, and the model structure is adjusted according to the number of scales N that need to be identified. The input of the multi-scale ground and sea clutter classification model based on the bifurcated residual neural network is a ground and sea clutter data and the scale label corresponding to the data, and the output is the category probability that the ground and sea clutter data belongs to the two categories of ground clutter and sea clutter. The multi-scale ground and sea clutter classification model based on the bifurcated residual neural network for identifying ground and sea clutter data of N scales consists of 1 feature extractor, N-1 feature extraction layers, and N data output channels. The structure of the multi-scale ground and sea clutter classification model based on the bifurcated residual neural network is as follows: Figure 7 shown.

[0048] Currently, widely used residual neural networks include ResNet-10, 18, 34, 50, and 101 (the numbers 10, 18, 34, 50, and 101 represent the sum of the number of convolutional and fully connected layers in the model). However, the land and sea clutter data used in this invention is one-dimensional and has limited feature complexity. Excessive convolutional layers can easily lead to overfitting. Therefore, this invention uses the portion of the ResNet-10 network, minus the fully connected layers, as the feature extractor. N-1 feature extraction layers, numbered 1, 2, ..., N-1, are added sequentially after the feature extractor. The feature maps corresponding to the last layer of the feature extractor and all feature extraction layers each predict data at a specific scale. The corresponding data scales are arranged from small to large according to the hierarchy of the feature maps. A data output channel is added after the last layer of the feature extractor and after each feature extraction layer. The data output channels are numbered 0, 1, 2, ..., N-1, starting from the output channel connected to the feature extractor. The land and sea clutter to be processed is processed using the Bif-Resnet model based on the data scale label. The ground and sea clutter data with a scale label of 0 is processed by the feature extractor and then outputted through the scale 0 data output channel as the model's recognition result for the data; the ground and sea clutter data with a scale label of n (n = 1, 2, ..., N-1)) is processed by the feature extractor and the 1, 2, ..., n feature extraction layers and then outputted through the scale n data output channel as the model's recognition result for the scale data.

[0049] In Bif-ResNet, the feature map scale of the last layer of the feature extractor is K0, which corresponds to the recognition of scale 0 data. The scales of the feature extraction layers numbered 1, 2, ..., N-1 are K n(n=1,2,...,N-1), used to identify data of scale 1,2,...,N-1. Since data of scale 0,1,2,...,N-2 may be zero-filled, the influence of 0 values ​​needs to be removed during recognition. Currently, the scale of the feature map corresponding to each scale is K n (n=0,2,...,N-2). Select a sequence from the zero-filled sequence As the final recognition feature map data corresponding to each scale, the final recognition feature map is input into the output channel corresponding to the scale to classify the data.

[0050] The feature extractor consists of a convolutional layer with a filter size of 1x3 and a stride of 2, a max pooling layer with a filter size of 1x3 and a stride of 2, and four residual blocks. Each residual block includes a convolutional layer with a filter size of 1x3 and a stride of 2, a convolutional layer with a filter size of 1x3 and a stride of 1, and a shortcut connection. The feature extraction layer is a convolutional layer with a filter size of 1x3 and a stride of 2. The data output channel consists of a convolutional layer with a filter size of 1x3 and a stride of 2, a pooling layer, and a fully connected layer. Each convolutional layer in the model is followed by a normalization layer and a RELU layer. The normalization layer further mitigates overfitting, and the RELU activation function fully utilizes gradient information to ensure continuous model convergence.

[0051] Land and sea clutter classification is a binary classification problem, and the PyTorch cross entropy loss function is used to calculate the loss. The loss is calculated for land and sea clutter at different scales, and the output channel corresponding to each scale outputs the model's prediction of the probability that the input data at that scale belongs to sea clutter. i1 and the probability of ground clutter s i2 , if s i1 >s i2 , then the clutter is considered to be sea clutter, otherwise it is ground clutter. The cross entropy loss function first converts the output result s i1 and s i2 Perform LogSoftmax operation

[0052]

[0053] Then perform the following cross entropy calculation to obtain the loss value of the scale data. y is the correct label of ground and sea clutter. The data label 0 indicates that the spectrum data is sea clutter, and 1 indicates that the spectrum data is ground clutter.

[0054] L i =-[(1-y)*S1+y*S2]

[0055] In step 2, based on the idea that the SSD model uses feature maps of different scales to detect targets of different scales, a multi-scale land and sea clutter classification model based on a bifurcated residual neural network is designed. Feature maps at different levels are extracted to predict land and sea clutter data of different scales. This enables the model to simultaneously identify land and sea clutter data of multiple scales, and output channels are designed separately for data of different scales.

[0056] Step 3: Use the training data set to train the multi-scale land and sea clutter classification model based on the bifurcated residual neural network constructed in step 2:

[0057] Step S3.1: Use sufficient number of scaled ground and sea clutter to train the multi-scale ground and sea clutter classification model based on the bifurcated residual neural network constructed in step 2. Freeze the parameters of all layers of the output channel n corresponding to the ground and sea clutter with scale label n (n≠N-1). If n=N-1, freeze the parameters of the N-1th feature extraction layer and all layers of the N-1th output channel, and only train the remaining layers.

[0058] In skywave radar spectrum data, the number of ground and sea clutter at different scales is uneven. The number of ground and sea clutter at a certain scale may be much smaller than that at other scales. Training a model for this scale data alone cannot achieve the required accuracy. This scale data is considered insufficient scale ground and sea clutter, while the number of ground and sea clutter at this scale is sufficient.

[0059] Step S3.2: Use all scales of ground and sea clutter to train the multi-scale ground and sea clutter classification model based on a bifurcated residual neural network processed in step S3.1. Set all the parameters frozen in step S3.1 to the first group of parameters, and set the parameters of the remaining layers of the model to the second group. Set different training weights for different groups, set a smaller learning rate for fine-tuning the second group, and the learning rate of the first group is 5 times that of the second group.

[0060] To address the problem of insufficient data of individual scales in multi-scale land and sea clutter recognition, this paper combines transfer learning theory and deep learning framework to propose a step-by-step training method for a bifurcated residual neural network model based on transfer learning.

[0061] Assuming that the scale label for the insufficient ground and sea clutter is n, and the scale labels for the sufficient ground and sea clutter are 0, 1, 2, …, n-1, n+1, …, N-1, we use the ground and sea clutter of scale n as the source domain, and the ground and sea clutter of scales 0, 1, 2, …, n-1, n+1, …, N-1 as the target domain. Using the feature extractor as the knowledge carrier, we leverage the knowledge learned by the feature extractor in the source domain to improve the model's classification ability for ground and sea clutter in the target domain. The ResNet-10 convolutional module, which serves as the feature extractor, possesses powerful feature extraction capabilities. The low-level convolutional layers of the feature extractor can extract general features of the data. As the number of convolutional layers increases, the features extracted become more specific, effectively extracting subtle features of the data, accelerating task learning and reducing the task's dependence on samples from the target domain.

[0062] Model training is a two-step process. The first step involves training the model using source domain data. The parameters of all layers in output channel n, corresponding to land and sea clutter at scale n (n≠N-1), are frozen. If n=N-1, the parameters of the N-1 feature extraction layer and all layers in output channel N-1 are frozen, and only the remaining layers are trained. In the first step, data at N-1 scales is used for training, resulting in N-1 output channels with output values ​​and N-1 loss values. These N-1 loss values ​​are summed to obtain the total model loss.

[0063]

[0064] The second step is training, using data of N scales for training, and a total of N loss values ​​are obtained. Let the number of sea and land clutter at each scale in the source domain be m i (i=0,1,2,n-1,n+1,...,N-1), the land / ocean clutter in the target domain is m n The inverse ratio of the mean number of sea and land clutter at each scale in the target domain and the source domain is taken as the coefficient of the source domain data loss value, that is,

[0065]

[0066] When training the model in step 3, data of different scales are used to train the model in steps, and land and sea clutter data of a large number of scales are used as pre-training for land and sea clutter data of a small number of scales to improve the model's classification ability for land and sea clutter data of a small number of scales.

[0067] Step 4: Use the test data set to test the multi-scale ground and sea clutter classification model based on the bifurcated residual neural network that has been trained. The multi-scale ground and sea clutter classification model based on the bifurcated residual neural network that has been tested and meets the project accuracy requirements can be used to classify the ground and sea clutter of the multi-scale data.

[0068] In order to verify the classification performance of the model, the evaluation method adopted in this paper is accuracy:

[0069]

[0070] Among them, TP is the number of positive classes predicted as positive classes; FN is the number of positive classes predicted as negative classes; FP is the number of negative classes predicted as positive classes; TN is the number of negative classes predicted as negative classes.

[0071] Example

[0072] 1. Construct a multi-scale land and sea clutter database and form the training and test datasets for the model:

[0073] In this embodiment, three groups (N=3) of ground and sea clutter data with coherent accumulation points of 256, 512, and 1024, respectively, are prepared. These groups are designated as Group A (containing 2000 ground and sea clutter), Group B (containing 49402 ground and sea clutter), and Group C (containing 74174 ground and sea clutter). Each group of data includes spectrum data, a ground and sea clutter classification label corresponding to the spectrum data (0 indicates that the spectrum data is sea clutter, and 1 indicates that the spectrum data is ground clutter), and a coherent accumulation point classification label corresponding to the spectrum data (0 indicates that the spectrum data is spectrum data with a coherent accumulation point number of 256, 1 indicates that the spectrum data is spectrum data with a coherent accumulation point number of 512, and 2 indicates that the spectrum data is spectrum data with a coherent accumulation point number of 1024).

[0074] 2. Construct a bifurcated residual neural network model:

[0075] This training uses land and sea clutter data of three scales (N=3). The model consists of a feature extractor, two feature extraction layers, and three data output channels, connected sequentially according to the model structure. The model has a total of 17 convolutional and fully connected layers. Therefore, this experiment constructed a Bif-ResNet-17 (a bifurcated residual neural network with a total of 17 convolutional and fully connected layers). The first convolutional layer of the feature extractor has 32 filters, and each residual block has two convolutional layers with the same number of filters. The four residual blocks have 64, 128, 256, and 512 convolutional filters, respectively. The convolutional layers in the two feature extraction layers and the data output channels all have 512 convolutional filters. 70% of the data was used as training data, and 30% as testing data.

[0076] 3. Use the training data set to train a multi-scale land and sea clutter classification model based on a bifurcated residual neural network:

[0077] In this example's multi-scale land and sea clutter classification experiment, the hyperparameters required for model training were set as follows: The model was trained using mini-batch samples, with a sample size of 64. The SGD algorithm was used to optimize the loss function. The model was implemented using the PyTorch deep learning framework and trained on a GeForce RTX 3090 GPU.

[0078] The Bif-ResNet-17 network was trained in two steps. The first step used only the data from Groups B and C. The second step used Group A, with the same number of ground and sea clutter intercepted from Groups B and C as from Group A. The initial learning rate for the first step was set to 0.001. In the second step, the initial learning rate for the first group was set to 0.0005, and the initial learning rate for the second group was set to 0.002.

[0079] In this embodiment, the network is trained for 20 cycles in both the first and second training steps. Figure 8 This means that the model can converge quickly when training the data of Group B and Group C in the first step. Figure 10 It means that the model can also gradually converge during the second step of training. Figure 9 This means that the training accuracy of the first step of training for group B and group C data can be rapidly improved during the training process. Figure 11 This indicates that during the second step of training, the model's classification accuracy for group B and group C data is relatively stable, and the model's classification accuracy for group A data is gradually improving.

[0080] 4. Use the test data set to test the trained multi-scale land and sea clutter classification model based on the bifurcated residual neural network:

[0081] This example compares the ability of Bif-ResNet-17 (a bifurcated residual neural network with 17 convolutional and fully connected layers) to classify multi-scale land and sea clutter data simultaneously with the ability of multiple groups of ResNet network models to classify multi-scale land and sea clutter data separately. For residual neural networks, equivalent pathways for data of different scales were constructed in Bif-ResNet-17 (a bifurcated residual neural network with a total of 17 convolutional and fully connected layers) for comparison. Ground and sea clutter with 256 coherent accumulation points undergoes a feature extractor and one data output channel in Bif-ResNet-17 (a bifurcated residual neural network with a total of 17 convolutional and fully connected layers). A Resnet-11 network model with the same channel structure was constructed to process ground and sea clutter with 256 coherent accumulation points. Similarly, Resnet-12 and Resnet-13 network models were constructed to process ground and sea clutter with 512 and 1024 coherent accumulation points, respectively.

[0082] In this embodiment, the training cycles for the Bif-ResNet-17, Resnet-11, Resnet-12, and Resnet-13 networks were all 20. Table 1 below shows their recognition results. Table 1 shows that the multi-scale ground and sea clutter classification model based on a bifurcated residual neural network of the present invention achieves a 13.54% higher recognition accuracy for Group A data than a bifurcated residual neural network model designed solely for that data. The classification accuracy for Groups B and C is essentially equal to that of the Resnet network. This shows that the multi-scale ground and sea clutter classification model based on a bifurcated residual neural network of the present invention can significantly improve the classification accuracy for small amounts of ground and sea clutter.

[0083] Table 1 Comparison of results of different models

[0084]

[0085] To improve the efficiency of coordinate registration for skywave over-the-horizon radar (OTHR), this paper proposes a bifurcated residual neural network model for classifying multi-scale ground and sea clutter data. This model utilizes feature maps at different levels to identify ground and sea clutter data of different scales, and separate output channels are designed for each scale. Due to the model's bifurcated structure, transfer learning can be used to train the model step by step. Using a model trained on a large number of ground and sea clutter scales as a pre-trained model improves the model's recognition accuracy for smaller amounts of ground and sea clutter, resulting in excellent classification accuracy.

Claims

1. A method for classifying land and sea clutter for multi-scale data, characterized in that: The following steps are involved: Step 1: Construct a multi-scale land and sea clutter database and form a training dataset and a test dataset for the model; Step 2: construct a multi-scale ground clutter classification model based on a bifurcated residual neural network; the multi-scale ground clutter classification model consists of 1 feature extractor, N-1 feature extraction layers, and N data output channels; Specifically, the part of the ResNet-10 network excluding the fully connected layer is used as the feature extractor. N-1 feature extraction layers are added sequentially after the feature extractor. The feature maps corresponding to the last layer of the feature extractor and all feature extraction layers respectively predict data of one scale. The corresponding data scales are arranged from small to large according to the level of the feature map. A data output channel is added after the last layer of the feature extractor and each feature extraction layer. Step 3: Use the training data set to train the multi-scale land and sea clutter classification model based on the bifurcated residual neural network constructed in step 2; the specific content is: Step S3.1: Use sufficient number of scaled ground and sea clutter to train the multi-scale ground and sea clutter classification model based on the bifurcated residual neural network constructed in step 2, freeze the parameters of all layers of the output channel n corresponding to the scale label of ground and sea clutter; if n is the scale, freeze the parameters of all layers of the feature extraction layer and the output channel n, and only train the remaining layers; where N represents the number of scales and n is the scale label; Step S3.2: Using all scales of ground and sea clutter to train the multi-scale ground and sea clutter classification model based on a bifurcated residual neural network processed in step S3.1, all parameters frozen in step S3.1 are set as the first group of parameters, and the remaining layer parameters are set as the second group. Different training weights are set for different groups, and a smaller learning rate is set for fine-tuning the second group. The learning rate of the first group is 5 times that of the second group. Step 4: Use the test data set to test the trained multi-scale ground and sea clutter classification model based on the bifurcated residual neural network. Use the tested multi-scale ground and sea clutter classification model based on the bifurcated residual neural network that meets the project accuracy requirements to classify the ground and sea clutter of the multi-scale data.

2. The method for classifying ground and sea clutter for multi-scale data according to claim 1, wherein: The specific content of step one is: Step S1.1, analyzing the spectrum data obtained by skywave radar detection; Step S1.2: Select multi-scale ground and sea clutter data that meet the project requirements. The peak at the zero frequency of the ground clutter and the Bragge peaks on both sides of the zero frequency of the sea clutter that meet the project requirements should have obvious characteristics. Step S1.3, extracting the frequency range [-5 Hz, 5 Hz] in the feature set from the ground and sea clutter data selected in step S1.2; Step S1.4: Randomly select 70% of the multi-scale land and sea clutter data processed as the training data set and 30% as the test data set.

Images