Light-weight spaceborne SAR (Synthetic Aperture Radar) echo interference type identification method

Abstract

The invention discloses a lightweight spaceborne SAR (Synthetic Aperture Radar) echo interference type identification method, which comprises the following steps of: firstly, modeling an interference signal, and constructing a spaceborne SAR interference data set on the basis of the modeling; then, building a lightweight interference type identification network; adding a sliding window multi-head attention mechanism in the deepest layer of the network; injecting the data set into the network for end-to-end training, and verifying the model training effect; and finally, an actually measured data processing stage: accurately identifying the types of interference signals contained in the interference echo matrix. According to the invention, the time cost of model training is greatly reduced through the lightweight network design, and the extraction capability of the network to the time-frequency local features is enhanced by using the sliding window attention mechanism. Different interference types including typical ground direct interference and inter-satellite scattering mutual interference in different scattering scenes can be accurately identified from an interference-containing echo matrix.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a lightweight method for identifying spaceborne SAR echo interference types. Background Technology

[0002] With the explosive growth of space radio traffic, spaceborne Synthetic Aperture Radar (SAR) systems inevitably suffer from radio frequency interference (RFI) from other electromagnetic radiation sources in the same frequency band during operation. This interference directly leads to image quality degradation, phase distortion, and decreased coherence. Depending on the source of the interference, common RFI experienced by spaceborne SAR systems can be categorized into direct ground interference (RDI) and mutual ground-scattered interference (MTSI). The main source of direct ground interference is high-power military and civilian radio equipment shared in the same frequency band on the ground / sea surface, such as airport surveillance radar and communication network equipment. MTSI, on the other hand, originates from interference information generated by ground scattering from different radar remote sensing satellites in scenarios where time, space, and frequency overlap. In most real-world scenarios, spaceborne SAR systems face heterogeneous interference environments from different waveforms. Most SAR anti-jamming technologies target only a single type of interference, often proving ineffective against other types. Therefore, to achieve better suppression, it is necessary to identify the types of interference on spaceborne SAR before employing anti-jamming techniques. Current research on radar interference signal type identification mainly focuses on direct ground-based interference, lacking research on the identification of inter-satellite scattering interference.

[0003] Traditional radar jamming type identification methods mainly include two approaches: maximum likelihood criterion-based and feature extraction-based. Maximum likelihood criterion-based methods require prior knowledge and have limited applicability. Feature extraction-based methods rely on manual selection of feature parameters; deviations in the selected features or classifier design can lead to a decrease in recognition rate. In recent years, data-driven methods such as deep learning have developed rapidly. These methods achieve end-to-end feature learning and have made significant progress in radar jamming identification. However, deep learning-based methods often have a large number of network model parameters, resulting in high computational complexity and significant time costs. Reducing the number of model layers, parameters, and computational load can lead to a decrease in model accuracy. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a lightweight method for identifying spaceborne SAR echo interference types. First, the interference signal is modeled, and a spaceborne SAR interference dataset is constructed based on this model. Then, a lightweight interference type identification network is built. This network utilizes a four-dimensional tensor processing module based on convolution and pooling operations, which can directly process four-dimensional tensors, avoiding the frequent and time-consuming tensor reshaping operations in traditional transformer networks, significantly reducing the time cost required for network training. A sliding window multi-head attention mechanism is added to the deepest layer of the network, accurately capturing local features of the time-frequency map while further reducing computational complexity. Next, the dataset is injected into the network for end-to-end training, and its model training effect is verified. Finally, in the experimental data processing stage, the types of interference signals contained in the interference echo matrix are accurately identified. This invention significantly reduces the time cost of model training through lightweight network design and enhances the network's ability to extract local time-frequency features using the sliding window attention mechanism, enabling accurate identification of different interference types from the interference echo matrix, including typical direct ground interference and inter-satellite scattering interference under different scattering scenarios.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Modeling the interference signal; Ground-based direct interference signal model and inter-satellite scattering mutual interference signal model are constructed respectively; based on the characteristic differences of the signals in the time and frequency domains, ground-based direct interference is further subdivided into narrowband interference, linear frequency modulation interference and sinusoidal modulation interference; Step 2: Dataset Construction; MTSI-containing measured data from different scenarios were selected as the data source, and MTSI-containing pulses were extracted to construct a dataset. Echo pulses corresponding to slow times at any azimuth on the original interference-free satellite echo matrix were selected, and these pulses were superimposed with three types of artificially simulated ground-based direct interference signals to obtain interference-containing pulses. Interference signal parameters such as amplitude, frequency modulation, and pulse width were randomly set. Short-time Fourier transforms were performed on the measured and simulated data to obtain a time-spectrum diagram, which is a training sample. Each training sample corresponds to a label for the interference type. Before network training began, all training samples were standardized in size and randomly divided into training, validation, and test sets according to a set ratio. Step 3: Interference type identification network construction; A lightweight interference type recognition network was constructed, whose overall architecture includes a shallow feature extraction module, a first-stage feature processing module, a second-stage feature processing module, a third-stage feature processing module, and a fourth-stage feature processing module. The input to the lightweight interference type identification network is of size [size missing]. The image shows B as the batch size. Indicates the number of image channels. Indicates the image height. This represents the image width; the input image is first composed of two steps of 2. The shallow feature extraction module, composed of convolutional layers, performs convolutional processing to obtain a value of [size missing]. The initial four-dimensional feature tensor, This indicates the number of image channels after processing by the shallow feature extraction module; the processed features are then input into the first-stage feature processing module. The first-stage feature processing module consists of three consecutive four-dimensional tensor processing modules and an embedding layer. The four-dimensional tensor processing modules first use a pooling mixer to extract low-level features. These low-level features then pass through a multilayer perceptron layer, and their output is superimposed with the output of the pooling mixer to obtain the final output features of the four-dimensional tensor processing module. The embedding layer downsamples the output features of the four-dimensional tensor processing module to... , This indicates the number of image channels after processing by the first-stage feature processing module; The second-stage feature processing module consists of two consecutive four-dimensional tensor processing modules and an embedding layer. This stage downsamples the features to... , This indicates the number of image channels after processing by the second-stage feature processing module; The third-stage feature processing module consists of six consecutive four-dimensional tensor processing modules and an embedding layer, and its output feature size is... , This indicates the number of image channels after processing by the third-stage feature processing module; The first three layers of the fourth-stage feature processing module are four-dimensional tensor processing modules, while the last layer switches to a three-dimensional tensor processing module. This three-dimensional tensor processing module consists of an attention mechanism and a linear multilayer perceptron module, which processes the three-dimensional tensors. The first three layers process the four-dimensional feature tensors, and the last layer triggers a dimension transformation, ultimately outputting the size of the feature. The three-dimensional tensor; the classification head normalizes and performs global average pooling on the output of the fourth-stage feature processing module to obtain the final output; The deepest layer of the lightweight interference type recognition network uses a sliding window attention mechanism. This mechanism divides the image into non-overlapping local windows, allowing for independent computation of self-attention within each window rather than calculating attention relationships between all locations, thus capturing local information in the image. The calculation formula for the sliding window attention mechanism is as follows: (1) Among them, relative position offset , These represent the query matrix, key matrix, and value matrix, respectively. Indicates the dimension of the query / key. This indicates the number of patches in the window; The sliding window attention mechanism regularly shifts the window position between consecutive Transformer layers, allowing the new window to cover different areas of the previous layer, thereby establishing cross-window connections. Step 4: Network training and testing; The training set and validation set constructed in step 2 are respectively input into the lightweight interference type recognition network built in step 3 for training and validation. When the loss rate and accuracy of the validation set converge, the lightweight interference type recognition network under the corresponding training round is selected and saved. The test set is input into the trained lightweight interference type recognition network for forward inference. After obtaining the corresponding predicted label, it is compared with the real label to test the training effect. Step 5: Processing the measured data; Step 5-1: Based on the interference pulse energy and median absolute deviation, filter out the interference pulses in the echo matrix, and then perform a short-time Fourier transform on them to generate the corresponding time-frequency spectrum. The pulse energy is calculated by the following formula: (2) in and They represent the first The pulse and the first The pulse energy of each pulse and These represent the fast time for distance and the slow time for direction, respectively. Indicates the total number of sampling points in the distance direction; If the energy of the pulse is higher than the median absolute deviation, the pulse is determined to be an interference pulse; Step 5-2: Input the time-frequency graph into the trained lightweight interference type recognition network to obtain the corresponding interference type identification results; Step 5.3: Perform cluster analysis on the interference type identification results of all interference pulse time-frequency diagrams to obtain the final interference type; if there are multiple cluster centers, it indicates that the echo matrix contains multiple different types of interference signals.

[0006] Preferably, the different scenarios include land scenarios, ocean scenarios, and land-sea boundary scenarios.

[0007] Preferably, the three types of ground-based direct interference signals include narrowband interference, linear frequency modulation interference, and sinusoidal modulation interference.

[0008] Preferably, the narrowband interference is constructed in the following manner:

[0009] in and These represent distance (fast time) and direction (slow time), respectively. Represents a rectangular window function. Indicates the number of interfering signals. and They represent the first The amplitude and frequency of the interference signal, , and These represent the SAR reception duration, the pulse width of the interference signal, and the delay of the interference signal, respectively.

[0010] Preferably, the linear frequency modulation interference is constructed in the following form:

[0011] in Indicates the first The frequency modulation of the interference signal.

[0012] Preferably, the sinusoidal modulation interference is constructed in the following form:

[0013] in , and The first The modulation index, modulation frequency, and initial phase of the interference signal.

[0014] Preferably, the regular offset is a rightward shift by half a window size or a downward shift by half a window size.

[0015] Preferably, in step 3, the lightweight interference type recognition network classification head is set to 6; the loss function is used to measure the difference between the network's predicted label and the true label, and the cross-entropy loss function is defined as follows: (3) in Let cross-entropy be the loss function. Indicates training samples, This represents the true label corresponding to the training sample. express Predicted values ​​after passing through a lightweight interference type identification network.

[0016] The beneficial effects of this invention are as follows: This invention significantly reduces the time cost of model training through a lightweight network design and enhances the network's ability to extract local time-frequency features by utilizing a sliding window attention mechanism. It can accurately identify different types of interference from the interference echo matrix, including typical direct ground interference and inter-satellite scattering interference under different scattering scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention.

[0018] Figure 2 shows a partial training sample diagram. Figure 2(a) Narrowband interference time-frequency diagram, Figure 2(b) Linear frequency modulation interference time-frequency diagram, Figure 2(c) Sine modulation interference time-frequency diagram, Figure 2(d) MTSI time-frequency diagram (land scene), Figure 2(e) MTSI time-frequency diagram (ocean scene), Figure 2(f) MTSI time-frequency diagram (land-sea boundary scene).

[0019] Figure 3 This is a schematic diagram of the network model used in this invention.

[0020] Figure 4 This is a schematic diagram of the four-dimensional tensor processing module.

[0021] Figure 5 This is a schematic diagram of the three-dimensional tensor processing module.

[0022] Figure 6 This is the loss rate curve.

[0023] Figure 7 This is the accuracy curve.

[0024] Figure 8 This is a quick view of the measured data. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Based on the differentiated characteristic distribution of different RFIs in the time-frequency domain, this invention proposes a lightweight method for identifying interference types in spaceborne SAR echoes. This invention significantly reduces time costs while improving the ability to extract the time-frequency characteristics of interference signals. It can accurately identify the type of interference signal contained in the original interference echo based on the time-spectrum diagram of the interference pulse, thus providing reliable prior information support for the formulation of subsequent anti-interference strategies.

[0027] The basic idea of ​​this invention is as follows: First, the interference signal is modeled, and a spaceborne SAR interference dataset is constructed based on this model. Then, a lightweight interference type identification network is built. This network utilizes a four-dimensional tensor processing module based on convolution and pooling operations, which can directly process four-dimensional tensors, avoiding the frequent and time-consuming tensor reshaping operations in traditional transformer networks, significantly reducing the time cost required for network training. A sliding window multi-head attention mechanism is added to the deepest layer of the network to accurately capture local features of the time-frequency map while further reducing computational complexity. Next, the dataset is injected into the network for end-to-end training, and its model training effect is verified. Finally, in the experimental data processing stage, the types of interference signals contained in the interference echo matrix are accurately identified.

[0028] To achieve the above objectives, the technical solution of the present invention is: a lightweight spaceborne SAR echo interference type identification method, mainly comprising the following steps: Step 1: Modeling the interference signal.

[0029] Parametric modeling is performed for typical RFI scenarios faced by spaceborne SAR systems, constructing ground-based direct interference signal models and inter-satellite scattering interference signal models. Based on the differences in signal characteristics in the time and frequency domains, ground-based direct interference can be further subdivided into narrowband interference, linear frequency modulation interference, and sinusoidal modulation interference.

[0030] Step 2: Dataset construction.

[0031] Due to the terrain scattering process of MTSI signals, different scattering terrains lead to different time-frequency characteristics of MTSI. Therefore, measured MTSI data from different scenarios (including land, ocean, and land-sea boundary scenarios) were selected as the data source, and MTSI pulses were extracted to construct a dataset. To further expand the dataset types, the echo pulse corresponding to a certain azimuth slow time on the interference-free original echo matrix of the satellite was selected. This pulse was superimposed with three artificially simulated types of ground direct interference signals (including narrowband interference, linear frequency modulation interference, and sinusoidal modulation interference) to obtain interference-containing pulses. To simulate real interference signals, interference signal parameters such as amplitude, frequency modulation, and pulse width were randomly set. After performing short-time Fourier transform on the measured and simulated data, a time-spectrum diagram was obtained, which is a training sample. Each training sample corresponds to a label of interference type. Before network training begins, all training samples are standardized in size and randomly divided into training, validation, and test sets according to proportions.

[0032] Step 3: Interference type identification network construction.

[0033] A lightweight interference type recognition network was constructed, whose overall architecture includes a shallow feature extraction module, a first-stage feature processing module, a second-stage feature processing module, a third-stage feature processing module, and a fourth-stage feature processing module.

[0034] The input to this network is of size . The image (B is the batch size) is first composed of two images with a step size of 2. The shallow feature extraction module, composed of convolutional layers, performs convolutional processing to obtain a value of [size missing]. The initial four-dimensional feature tensor is obtained. The processed features are input into the first-stage feature processing module. The first-stage feature processing module consists of three consecutive four-dimensional tensor processing modules and an embedding layer. The four-dimensional tensor processing module first uses a simple pooling mixer to extract low-level features, which are then passed through a multilayer perceptron layer. The output of this layer is then superimposed with the output of the pooling mixer to obtain the output features of the four-dimensional tensor processing module. The use of this processing module avoids the frequent and time-consuming tensor reshaping operations in traditional transformer networks, significantly reducing time costs. The embedding layer downsamples the output features of the four-dimensional tensor processing module to... The second-stage feature processing module consists of two consecutive four-dimensional tensor processing modules and an embedding layer. This stage downsamples the features to... The extracted features are then further input into the third-stage feature processing module, which consists of six consecutive four-dimensional tensor processing modules and an embedding layer. The size of its output features is... The fourth-stage feature processing module consists of three layers: the first three layers are four-dimensional tensor processing modules, and the last layer switches to a three-dimensional tensor processing module. This three-dimensional tensor processing module comprises an attention mechanism and a linear multilayer perceptron module to process the three-dimensional tensors. The first three layers process the four-dimensional feature tensors, and the last layer triggers a dimension transformation, ultimately outputting the size feature. The three-dimensional tensor. The classification head normalizes and performs global average pooling on the output of the fourth-stage feature processing module to obtain the final output.

[0035] Traditional global multi-head self-attention mechanisms struggle to effectively capture local features, while interfering signals often occupy only a limited area in the time-frequency image. Therefore, this invention employs a sliding window attention mechanism in the deepest layer of the network. This mechanism divides the image into non-overlapping local windows, allowing for independent computation of self-attention within each window rather than calculating attention relationships between all locations, thus effectively capturing local information in the image. Furthermore, this attention mechanism is deployed only in the deepest layer of the network, further reducing computational complexity. The calculation formula for this attention mechanism is as follows: (1) Among them, relative position offset , Represents the query matrix, key matrix, and value matrix. Indicates the dimension of the query / key. This indicates the number of patches in the window. To enhance information exchange between windows, this mechanism regularly offsets the window position between consecutive Transformer layers (such as moving it to the right or down by half a window size), allowing the new window to cover different areas of the previous layer, thereby establishing cross-window connections.

[0036] Step 4: Network training and testing.

[0037] The training and validation sets constructed in step 2 are input into the network built in step 3 for training and validation, respectively. When the loss rate and accuracy of the validation set converge, the training model for the corresponding training epoch is selected and saved. The interference type identification network built in step 3 has 6 classification heads (three types of direct ground interference and inter-satellite scattering interference under three different scattering terrains). The optimizer is the core algorithm that guides the update of neural network parameters. Essentially, it calculates the gradient of the loss function and dynamically adjusts the weight matrix and bias vector of each layer according to a certain strategy to gradually reduce the model's prediction error. The loss function is used to measure the difference between the network's predicted label and the true label. The cross-entropy loss function is defined as follows: (2) in Let cross-entropy be the loss function. Indicates training samples, This represents the true label corresponding to the training sample. express The predicted value after passing through the network model.

[0038] The test set is input into the trained interference type identification model for forward inference. The corresponding predicted labels are then compared with the true labels to test the model training effect.

[0039] Step 5: Measured Data Processing. After completing network training, the measured data processing stage begins. The original interference-containing echo matrix is ​​processed to identify its interference type. The specific steps are as follows: Step 5.1: Based on the interference pulse energy and median absolute deviation, filter out the interference pulses in the echo matrix, and then perform a short-time Fourier transform on them to generate the corresponding time-frequency spectrum. The pulse energy can be calculated by the following formula: (3) in and They represent the first A pulse and its pulse energy, and These represent the fast time for distance and the slow time for direction, respectively. This indicates the total number of sampling points in the distance direction. If the energy of the pulse is higher than the median absolute deviation, the pulse is considered to contain interference.

[0040] Step 5.2: Input the time-frequency graph into the trained interference type identification model to obtain the corresponding interference type identification result.

[0041] Step 5.3: Perform cluster analysis on the interference type identification results of all interference pulse time-frequency diagrams to obtain the final determined interference type. If multiple cluster centers exist, it indicates that the echo matrix contains multiple different types of interference signals.

[0042] Example: Step 1: Modeling the interference signal.

[0043] Parametric modeling is performed for typical RFI scenarios faced by spaceborne SAR systems, constructing models for ground-based direct interference signals and inter-satellite scattering interference signals. Based on the differences in signal characteristics in the time and frequency domains, ground-based direct interference can be further subdivided into narrowband interference, linear frequency modulation interference, and sinusoidal modulation interference. In this example, the interference signal modeling is as follows: The narrowband interference is constructed as follows: (4) in and These represent distance (fast time) and direction (slow time), respectively. Represents a rectangular window function. Indicates the number of interfering signals. and They represent the first The amplitude and frequency of the interference signal, , and These represent the SAR reception duration, the pulse width of the interference signal, and the delay of the interference signal, respectively.

[0044] The linear frequency modulation interference is constructed as follows: (5) in Indicates the first The frequency modulation of the interference signal.

[0045] The sinusoidal modulation interference is constructed as follows: (6) in , and The first The modulation index, modulation frequency, and initial phase of the interference signal.

[0046] Since the MTSI source itself is a spatially adjacent heterogeneous radar remote sensing satellite, its transmitted signal, after being scattered by the ground, overlaps with the echo from the spaceborne SAR system, thus interfering with the spaceborne SAR system. The MTSI and target echo have similar bidirectional propagation paths, which can be modeled as follows: (7) in , and These represent the pulse width, modulation frequency, and carrier frequency of the MTSI source's transmitted signal, respectively. This indicates the amplitude of the MTSI signal. , Indicates the distance between the MTSI source and its irradiated area. This indicates the distance between the MTSI illumination area and the spaceborne SAR system.

[0047] Step 2: Dataset construction.

[0048] Due to the terrain scattering process of MTSI signals, different scattering terrains will cause MTSI to exhibit different time-frequency characteristics. Therefore, measured MTSI data from different scenarios (including land, ocean, and land-sea boundary scenarios) were selected as the data source, and MTSI pulses were extracted to construct a dataset. To further expand the dataset types, the echo pulse corresponding to a certain azimuth slow time on the interference-free original echo matrix of the satellite was selected. This pulse was superimposed with three artificially simulated types of ground direct interference signals (including narrowband interference, linear frequency modulation interference, and sinusoidal modulation interference) to obtain interference-containing pulses. To simulate real interference signals, interference signal parameters such as amplitude, frequency modulation, and pulse width were randomly set. The dataset constructed in this example contains 4038 training samples. Before the network training began, all training samples were standardized in size and randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training samples in this example are essentially single-channel pseudo-color images with an image size of [missing information]. Figure 2 shows some of the training samples.

[0049] Step 3: Interference type identification network construction.

[0050] A lightweight interference type recognition network is constructed, whose overall architecture includes a shallow feature extraction module, a first-stage feature processing module, a second-stage feature processing module, a third-stage feature processing module, and a fourth-stage feature processing module, as follows: Figure 3 .

[0051] In this example, the input to the network is of size [size missing]. The image (32-bit batch size) is first composed of two images with a step size of 2. The shallow feature extraction module, composed of convolutional layers, performs convolutional processing to obtain a value of [size missing]. The initial four-dimensional feature tensor is obtained. The processed features are then input into the first-stage feature processing module. This module consists of three consecutive four-dimensional tensor processing modules and an embedding layer. Each module first uses a simple pooling mixer to extract low-level features. These low-level features then pass through a multilayer perceptron layer, and their output is superimposed with the pooling mixer's output to finally obtain the output features of the four-dimensional tensor processing module, such as... Figure 4 As shown. The use of this processing module avoids the frequent and time-consuming tensor reshaping operations in traditional transformer networks, significantly reducing time costs. The embedding layer downsamples the output features of the four-dimensional tensor processing module to... The second-stage feature processing module consists of two consecutive four-dimensional tensor processing modules and an embedding layer. This stage downsamples the features to... The extracted features are then further input into the third-stage feature processing module, which consists of six consecutive four-dimensional tensor processing modules and an embedding layer. The size of its output features is... The fourth-stage feature processing module consists of three layers: the first three layers are four-dimensional tensor processing modules, and the last layer switches to a three-dimensional tensor processing module. The three-dimensional tensor processing module comprises an attention mechanism and a linear multilayer perceptron module, which processes the three-dimensional tensors, such as... Figure 5 As shown. The first three layers process the four-dimensional feature tensor, and the last layer triggers a dimension transformation, ultimately outputting the feature as size. The three-dimensional tensor. The classification head normalizes and performs global average pooling on the output of the fourth-stage feature processing module to obtain the final output.

[0052] Traditional global multi-head self-attention mechanisms struggle to effectively capture local features, while interfering signals often occupy only a limited area in the time-frequency image. Therefore, this invention employs a sliding window attention mechanism in the deepest layer of the network. This mechanism divides the image into non-overlapping local windows, allowing for independent computation of self-attention within each window rather than calculating attention relationships between all locations, thus effectively capturing local information in the image. Furthermore, this attention mechanism is deployed only in the deepest layer of the network, further reducing computational complexity. The calculation formula for this attention mechanism is as follows: (8) Among them, relative position offset , Represents the query matrix, key matrix, and value matrix. Indicates the dimension of the query / key. This indicates the number of patches in the window. To enhance information exchange between windows, this mechanism regularly offsets the window position between consecutive Transformer layers (such as moving it to the right or down by half a window size), allowing the new window to cover different areas of the previous layer, thereby establishing cross-window connections.

[0053] Step 4: Network training.

[0054] The network model built in step 3 is trained using the training and validation sets. When the loss rate and accuracy of the validation set converge, the training model for the corresponding training epoch is selected and saved. This example uses the cross-entropy loss function as the loss function, and employs the AdamW optimizer and backpropagation method to optimize the network parameters. The weight decay coefficient is set to 5e-3, the initial learning rate is 2e-5, the batch size of both the training and validation sets is set to 32, the number of training epochs is 50, and the interference type identification network classification head is set to 6 (three types of direct ground interference and inter-satellite scattering interference under three different scattering terrains). The cross-entropy loss function is defined as follows: (9) in Let cross-entropy be the loss function. Indicates training samples, This represents the true label corresponding to the training sample. express The predicted value after passing through the network model.

[0055] Because this interference type identification network employs a lightweight architecture, the time cost required for network training is significantly reduced; in this example, a single training round takes only about 40 seconds. In contrast, under the same hardware platform and dataset conditions, a ResNet34 network requires over 600 seconds to complete a single training round. Table 1 shows the hardware platform parameters required for training in this example.

[0056] Table 1 Hardware Platform Parameters

[0057] The loss and accuracy curves for the training and validation sets are as follows: Figure 6 and Figure 7 When the epoch is greater than 30, the loss rate and accuracy of the validation set begin to converge.

[0058] The test set is input into the trained interference type identification model for forward inference. After obtaining the corresponding predicted labels, they are compared with the true labels to test the model's training effect. Table 2 shows the identification accuracy of some different types of interference.

[0059] Table 2. Identification accuracy of some different types of interference

[0060] Step 5: Processing the measured data. Figure 8 This is a quick view image corresponding to the measured data processed in this instance. After completing network training, the measured data processing stage begins, where the original interference-containing echo matrix is ​​processed to identify its interference type. The specific steps are as follows: Step 5.1: Based on the interference pulse energy and median absolute deviation, filter out the interference pulses in the echo matrix, and then perform a short-time Fourier transform on them to generate the corresponding time-frequency spectrum. The pulse energy can be calculated by the following formula: (10) in and They represent the first A pulse and its pulse energy, and These represent the fast time for distance and the slow time for direction, respectively. This indicates the total number of sampling points in the distance direction. If the energy of the pulse is higher than the median absolute deviation, the pulse is considered to contain interference.

[0061] Step 5.2: Input the time-frequency graph into the trained interference type identification model to obtain the corresponding interference type identification result.

[0062] Step 5.3: Perform cluster analysis on the interference type identification results of all interference pulse time-frequency maps to obtain the final determined interference type. If multiple cluster centers exist, it indicates that the echo matrix contains multiple different types of interference signals. In this example, there are 2 cluster centers, indicating that the echo matrix contains 2 different types of interference signals: inter-satellite scattering interference in the ocean scene and inter-satellite scattering interference in the ground scene.

Claims

1. A lightweight method for identifying spaceborne SAR echo interference types, characterized in that, Includes the following steps: Step 1: Modeling the interference signal; Ground-based direct interference signal model and inter-satellite scattering mutual interference signal model are constructed respectively; based on the characteristic differences of the signals in the time and frequency domains, ground-based direct interference is further subdivided into narrowband interference, linear frequency modulation interference and sinusoidal modulation interference; Step 2: Dataset Construction; We selected measured data containing MTSI from different scenarios as data sources, and extracted MTSI pulses from them to construct a dataset. The echo pulse corresponding to the slow time at any azimuth on the original interference-free satellite echo matrix is ​​selected, and this pulse is superimposed with three types of artificially simulated ground direct interference signals to obtain an interference-containing pulse; the amplitude, frequency modulation, and pulse width parameters of the interference signal are randomly set; the time spectrum diagram is obtained by performing short-time Fourier transform on the measured and simulated data, which is a training sample, and each training sample corresponds to an interference type label; before the network training begins, all training samples are standardized in size and randomly divided into training set, validation set, and test set according to proportion; Step 3: Interference type identification network construction; A lightweight interference type recognition network was constructed, whose overall architecture includes a shallow feature extraction module, a first-stage feature processing module, a second-stage feature processing module, a third-stage feature processing module, and a fourth-stage feature processing module. The input to the lightweight interference type identification network is of size [size missing]. The image shows B as the batch size. Indicates the number of image channels. Indicates the image height. This represents the image width; the input image is first composed of two steps of 2. The shallow feature extraction module, composed of convolutional layers, performs convolutional processing to obtain a value of [size missing]. The initial four-dimensional feature tensor, This indicates the number of image channels after processing by the shallow feature extraction module; the processed features are then input into the first-stage feature processing module. The first-stage feature processing module consists of three consecutive four-dimensional tensor processing modules and an embedding layer. The four-dimensional tensor processing module first uses a pooling mixer to extract low-level features, which are then passed through a multilayer perceptron layer. The output of this layer is then superimposed with the output of the pooling mixer to finally obtain the output features of the four-dimensional tensor processing module. The embedding layer downsamples the output features of the four-dimensional tensor processing module to , This indicates the number of image channels after processing by the first-stage feature processing module; The second-stage feature processing module consists of two consecutive four-dimensional tensor processing modules and an embedding layer. This stage downsamples the features to... , This indicates the number of image channels after processing by the second-stage feature processing module; The third-stage feature processing module consists of six consecutive four-dimensional tensor processing modules and an embedding layer, and its output feature size is... , This indicates the number of image channels after processing by the third-stage feature processing module; The first three layers of the fourth-stage feature processing module are four-dimensional tensor processing modules, while the last layer switches to a three-dimensional tensor processing module. This three-dimensional tensor processing module consists of an attention mechanism and a linear multilayer perceptron module, which processes the three-dimensional tensors. The first three layers process the four-dimensional feature tensors, and the last layer triggers a dimension transformation, ultimately outputting the size of the feature. The three-dimensional tensor; The classification head normalizes and performs global average pooling on the output of the fourth-stage feature processing module to obtain the final output. The deepest layer of the lightweight interference type recognition network uses a sliding window attention mechanism. This mechanism divides the image into non-overlapping local windows, allowing for independent computation of self-attention within each window rather than calculating attention relationships between all locations, thus capturing local information in the image. The calculation formula for the sliding window attention mechanism is as follows: （1） Among them, relative position offset , These represent the query matrix, key matrix, and value matrix, respectively. Indicates the dimension of the query / key. This indicates the number of patches in the window; The sliding window attention mechanism regularly shifts the window position between consecutive Transformer layers, allowing the new window to cover different areas of the previous layer, thereby establishing cross-window connections. Step 4: Network training and testing; The training set and validation set constructed in step 2 are respectively input into the lightweight interference type recognition network built in step 3 for training and validation. When the loss rate and accuracy of the validation set converge, the lightweight interference type recognition network under the corresponding training round is selected and saved. The test set is input into the trained lightweight interference type recognition network for forward inference. After obtaining the corresponding predicted label, it is compared with the real label to test the training effect. Step 5: Processing the measured data; Step 5-1: Based on the interference pulse energy and median absolute deviation, filter out the interference pulses in the echo matrix, and then perform a short-time Fourier transform on them to generate the corresponding time-frequency spectrum. The pulse energy is calculated by the following formula: （2） in and They represent the first The pulse and the first The pulse energy of each pulse and These represent the fast time for distance and the slow time for direction, respectively. Indicates the total number of sampling points in the distance direction; If the energy of the pulse is higher than the median absolute deviation, the pulse is determined to be an interference pulse; Step 5-2: Input the time-frequency graph into the trained lightweight interference type recognition network to obtain the corresponding interference type identification results; Step 5.3: Perform cluster analysis on the interference type identification results of all interference pulse time-frequency diagrams to obtain the final interference type; if there are multiple cluster centers, it indicates that the echo matrix contains multiple different types of interference signals.

2. The lightweight spaceborne SAR echo interference type identification method according to claim 1, characterized in that, The different scenarios include land scenarios, ocean scenarios, and land-sea boundary scenarios.

3. The lightweight spaceborne SAR echo interference type identification method according to claim 1, characterized in that, The three types of ground-based direct interference signals include narrowband interference, linear frequency modulation interference, and sinusoidal modulation interference.

4. The lightweight spaceborne SAR echo interference type identification method according to claim 3, characterized in that, The narrowband interference is constructed in the following manner: in and These represent distance (fast time) and direction (slow time), respectively. Represents a rectangular window function. Indicates the number of interfering signals. and They represent the first The amplitude and frequency of the interference signal, , and These represent the SAR reception duration, the pulse width of the interference signal, and the delay of the interference signal, respectively.

5. The lightweight spaceborne SAR echo interference type identification method according to claim 3, characterized in that, The linear frequency modulation interference is constructed in the following manner: in Indicates the first The frequency modulation of the interference signal.

6. The lightweight spaceborne SAR echo interference type identification method according to claim 3, characterized in that, The sinusoidal modulation interference is constructed in the following manner: in , and The first The modulation index, modulation frequency, and initial phase of the interference signal.

7. The lightweight spaceborne SAR echo interference type identification method according to claim 1, characterized in that, The regular offset is either shifting to the right by half a window size or shifting down by half a window size.

8. The lightweight spaceborne SAR echo interference type identification method according to claim 1, characterized in that, In step 3, the lightweight interference type recognition network classification head is set to 6; the loss function is used to measure the difference between the network's predicted label and the true label, and the cross-entropy loss function is defined as follows: （3） in Let cross-entropy be the loss function. Indicates training samples, This represents the true label corresponding to the training sample. express Predicted values ​​after passing through a lightweight interference type identification network.

Images