A method and system for radar altimeter echo classification based on transformer

Abstract

This application relates to the field of marine satellite altimetry technology, and particularly to a radar altimeter echo classification method and system based on Transformer. The method includes: modeling and simulating various echo waveforms from traditional pulse-limited and synthetic aperture radar altimeters; superimposing noise of different intensities to construct multi-class, multi-mode echo datasets; serializing and position-encoding the echo signals, inputting them into a Transformer encoder model, and using a multi-head self-attention mechanism to complete waveform feature extraction and classification; optimizing model parameters through dataset training to obtain a trained classification model; finally, inputting the echoes to be classified into the model and outputting the classification results. This invention can generate simulation data in batches, reducing data annotation costs, effectively extracting multi-scale waveform features, and significantly improving echo classification accuracy.

Description

Technical Field

[0001] This application relates to the field of marine satellite altimetry technology, and in particular to a radar altimeter echo classification method and system based on Transformer. Background Technology

[0002] As one of the most important marine remote sensing instruments, spaceborne radar altimeters can measure the global sea surface height field with centimeter-level accuracy, and simultaneously obtain information on significant wave height and wind speed. Altimeters play a crucial role in scientific research, national defense, and economic life. Nearshore waters (including waters surrounding islands and reefs) are the most vulnerable to the effects of global climate change and sea-level rise, making them of paramount research importance and thus a top priority in current satellite altimetry technology.

[0003] Traditional radar altimeters operate in a pulse-finite mode, with an on-orbit resolution on the order of kilometers. These are known as Low Resolution Mode (LRM) altimeters. Synthetic Aperture Radar (SAR) altimeters, as a new type of satellite radar altimeter, offer an order of magnitude higher resolution than traditional altimeters. These are known as High Resolution Mode (HRM) altimeters and are more suitable for near-shore applications. In recent years, Fully-Focusing Synthetic Aperture Radar (FFSAR) technology has been developed, further improving resolution.

[0004] In the complex environment of nearshore waters, neither traditional pulse-finite radar altimeters nor the widely used synthetic aperture radar altimeters can accurately process all types of echo signals using a single re-tracking algorithm. The current mainstream solution is to first classify and preprocess the echo signals, and then apply appropriate re-tracking strategies based on different types. Therefore, the accuracy of classification directly determines the reliability of subsequent re-tracking results. With the development of related technologies, echo classification methods have become increasingly diverse, and can be summarized into three categories: simple classification strategies based on feature extraction, classification methods based on traditional machine learning, and classification methods based on deep learning.

[0005] Simple classification methods based on feature extraction are relatively easy to implement, but the decision boundaries between categories are not clear enough, resulting in limited distinguishability between categories, which in turn restricts classification accuracy.

[0006] Traditional machine learning-based classification methods typically require the design of dedicated classifiers for traditional pulse finite echo and synthetic aperture radar altimeter echo, which not only increases the complexity of implementation but also leaves room for improvement in classification accuracy. Summary of the Invention

[0007] The purpose of this application is to overcome the above-mentioned defects of the prior art and provide a radar altimeter echo classification method and system based on Transformer, which is particularly suitable for situations where radar altimeter echoes of any type need to be classified.

[0008] To address the aforementioned technical problems, the present application provides a Transformer-based radar altimeter echo classification method, comprising: Step 1: Model and simulate various echo waveforms of radar altimeters of different systems. By superimposing noise of different intensity levels, generate radar altimeter echo data of multiple categories, thereby constructing a radar altimeter echo dataset covering multiple systems and multiple forms. Step 2: Serialize and position-encode the radar altimeter echo signals in the radar altimeter echo dataset to construct an input sequence. Input the input sequence into a radar altimeter echo classification model based on a Transformer encoder structure. Extract features and perform classification through a multi-head self-attention mechanism. Step 3: Train the radar altimeter echo classification model based on the Transformer encoder structure based on the radar altimeter echo dataset. Optimize the parameters in the radar altimeter echo classification model based on the Transformer encoder structure by setting the loss function and training hyperparameters to obtain the trained radar altimeter echo classification model. Step 4: Input the radar altimeter echo signal to be classified into the trained radar altimeter echo classification model, and output the corresponding radar altimeter echo classification result.

[0009] As an improvement to the above technical solution, in step 1, the radar altimeters of different systems include: conventional pulse-finite radar altimeters and synthetic aperture radar altimeters; the noise includes: Gaussian noise and Rayleigh noise; the multi-category radar altimeter echo data includes: standard ocean waveform, standard cone waveform, anomalous peak waveform, trailing edge noise waveform, multi-peak waveform, and cluttered signal waveform.

[0010] As an improvement to the above technical solution, the expression for the simulation model of the standard ocean waveform of the traditional pulse finite radar altimeter is... for: ; in, Let be the envelope function of the echo. For the shape function of the echo body, Let be the deformation function caused by the non-Gaussian probability density of the wave height. For the thermal noise of the radar system, Let time be the independent variable; where, The envelope function of the echo for: ; in, Echo scaling factor This refers to the antenna's misdirection angle. This is the antenna beam effect factor. To account for the mixed beamwidth and attitude angle standard deviation of the beam height, is the width parameter of the error function; where, The antenna beam effect factor for: ;in, This represents the 3dB beamwidth of the radar antenna. The width parameter of the error function for: ,in, This represents the time delay corresponding to the mean sea level. The mixed beam height standard deviation taking into account beamwidth and attitude angle for: ; in, For the cosine correction term of the misdirection angle, The standard deviation of the mixed wave height in the dimension of time. This is the sine correction term for the misdirection angle; where, , The speed of light in a vacuum. To determine the orbital height after curvature correction. , This is the satellite's nominal orbital altitude. The radius of the Earth; ; The echo body shape function for: ; in, Let be the error function. The skewness of the wave height probability density. The standard deviation of sea surface wave height is expressed in terms of time. The standard deviation of the mixed wave height in the dimension of time. The deformation function caused by the non-Gaussian probability density of the wave height for: .

[0011] As an improvement to the above technical solution, the expression for the simulation model of the standard cone-shaped waveform of the traditional pulse finite radar altimeter is... for: ; Where * denotes convolution operation, It is the impulse response and probability density function of a flat sea surface. The convolution; where, ; in, Echo scaling factor As the independent variable for time, The parameter is determined by the root mean square slope of the sea surface. This represents the time delay corresponding to the mean sea level. It is the thermal noise of the radar system.

[0012] As an improvement to the above technical solution, the expression of the simulation model of the echo waveform of the synthetic aperture radar altimeter is... for: ; in, Indicates the distance cell index of the echo model. This represents the azimuth Doppler frequency index of the echo model. This represents the peak echo power. This is the echo cell index corresponding to the echo mean sea level. For the effective wave height, The normalized standard deviation of sea surface wave height. For the first The distance unit and the Fresnel reflection coefficients of the sea surface at Doppler frequencies in each azimuth direction. Indicates the echo model with respect to k The zeroth-order expansion term, Indicates the echo model with respect to First-order expansion term, Physical quantities related to the antenna pattern; These are mixed parameters determined by the resolution of the radar altimeter; ; in, This represents the normalized width of the satellite's response to a point target along its orbit, approximated by a Gaussian approximation. This represents the normalized width of the point target response to the Gaussian approximation of the track intersection. Indicates the satellite's resolution along its orbit. Indicates the cross-track resolution. The normalized standard deviation of sea surface wave height; ; in, Indicates the pulse repetition frequency. This represents the number of pulses within a Burst. Indicates the Doppler frequencies corresponding to different sub-beams. , Indicates satellite speed. Indicates the carrier wavelength. Indicates the Doppler angles corresponding to different sub-beams. , This represents the number of views processed by the model in multi-view processing. It is the angular spacing between adjacent sub-beams. , For satellite speed, For burst repetition interval, This represents the satellite orbital altitude after curvature correction.

[0013] As an improvement to the above technical solution, the expression of the simulation model of the abnormal peak waveform is... for: ; in, It is the amplitude of Gaussian noise; It is the location of Gaussian noise; It is the standard deviation of Gaussian noise. Time is the independent variable. It is a standard echo waveform; It is the standard deviation of Rayleigh noise. Represents a random variable with a Rayleigh distribution; The expression of the simulation model of the trailing edge noise waveform for: ; in, For Gaussian noise index, For the first The amplitude of a Gaussian noise; For the first The location of Gaussian noise; For the first One standard deviation of Gaussian noise; The total number of Gaussian noises; This represents the overall amplitude of the trailing edge noise. The expression of the simulation model of the multi-peak waveform for: .

[0014] As an improvement to the above technical solution, step 2 specifically includes: Step 2.1: Divide the radar altimeter echo signal into multiple continuous and non-overlapping signal segments, and map them into a vector sequence through linear projection; Step 2.2: Perform position encoding processing on the vector sequence to characterize the timing information of the radar altimeter echo signal; Step 2.3: Input the position-encoded vector sequence as the embedding sequence into the radar altimeter echo classification model based on the Transformer encoder structure, extract features through a multi-head self-attention mechanism and output the classification result.

[0015] As an improvement to the above technical solution, step 2.2 is encoded using the following positional encoding method: ; ; in, Represents the positional encoding for even-numbered dimensions. Represents the positional encoding for odd-dimensional elements; Will and Add the vector sequence to the corresponding dimension to obtain the encoded vector sequence; In step 2.3, the radar altimeter echo classification model based on the Transformer encoder structure includes multiple stacked Transformer encoders, and the calculation of each Transformer encoder layer is as follows: ; ; in, For the layer index of the Transformer encoder, For the first Temporary sequences of layer Transformer encoders, For the first The output sequence of the layer Transformer encoder, For the first The output sequence of the layer Transformer encoder, This represents a multi-head self-attention function. Indicates the layer regularization function, Indicates a feedforward network; The radar altimeter echo classification model based on the Transformer encoder structure extracts features through a multi-head self-attention mechanism. For each head of the multi-head self-attention mechanism... Self-attention for: ; in, , and These are the query, key, and value matrices obtained by linear transformation of the embedding sequence from the Transformer encoder, with superscripts... T It is a matrix transpose operation. It is the dimension of the key vector; The radar altimeter echo classification model based on the Transformer encoder structure uses a feedforward neural network. Perform nonlinear mapping: ; in, This is the learnable weight matrix of the first layer of the feedforward network. This is the learnable bias vector of the first layer of the feedforward network. This is the learnable weight matrix for the second layer of the feedforward network. This is the learnable bias vector for the second layer of the feedforward network. It is the Gaussian error linear unit activation function. For embedded sequences; The radar altimeter echo classification model based on the Transformer encoder structure outputs classification results through a classification layer: ; in, It is the learnable weight matrix of the classification layer; It is the learnable bias vector of the classification layer. The encoder outputs the feature vector of the first position in the sequence. The total number of categories of radar altimeter echo data.

[0016] As an improvement to the above technical solution, in step 3, the loss function includes: cross-entropy loss function. : ; in, It is the total number of radar altimeter echo data within a batch. For indexing radar altimeter echo data, It is the total number of categories of radar altimeter echo data. It is a category index. It is the first The radar altimeter echo data corresponds to the first... One-hot tags for categories, It is the first The radar altimeter echo data belongs to the first... The probability distribution of the categories.

[0017] To achieve another objective of the present invention, the present invention also provides a radar altimeter echo classification system based on Transformer, comprising: The dataset construction module is used to model and simulate various echo waveforms of radar altimeters of different systems. By superimposing noise of different intensity levels, it generates radar altimeter echo data of multiple categories, thereby constructing a radar altimeter echo dataset covering multiple systems and multiple forms. The signal processing module is used to serialize and encode the radar altimeter echo signals in the radar altimeter echo dataset, construct an input sequence, and input the input sequence into a radar altimeter echo classification model based on a Transformer encoder structure, where feature extraction and classification are performed through a multi-head self-attention mechanism. The model training and optimization module is used to train the radar altimeter echo classification model based on the Transformer encoder structure based on the radar altimeter echo dataset. It optimizes the parameters of the radar altimeter echo classification model by setting the loss function and training hyperparameters to obtain the trained radar altimeter echo classification model. The classification module is used to input the radar altimeter echo signal to be classified into the trained radar altimeter echo classification model and output the corresponding radar altimeter echo classification result. The advantages of this application are that the radar altimeter echo classification method and system based on Transformer disclosed in this application simulates various forms of echoes from radar altimeters of different systems during the dataset preparation stage, generates training data in batches, and significantly reduces the cost of data annotation. In the classifier design and training process, the multi-head attention mechanism in the Transformer model is introduced to identify waveform features at different scales and improve the waveform classification accuracy. Attached Figure Description

[0018] Figure 1 A flowchart of the radar altimeter echo classification method based on Transformer provided in Embodiment 1 of the present invention; Figure 2(a) is a schematic diagram of the SAR ocean standard waveform simulation in the dataset; Figure 2(b) is a schematic diagram of the SAR cone standard waveform simulation in the dataset; Figure 2(c) is a schematic diagram of LRM ocean standard waveform simulation in the dataset; Figure 2(d) is a schematic diagram of the LRM cone standard waveform simulation in the dataset; Figure 2(e) is a schematic diagram of the LRM ocean anomaly peak simulation in the dataset; Figure 2(f) is a schematic diagram of the simulation of the LRM cone anomaly peak in the dataset; Figure 2(g) is a schematic diagram of the SAR ocean anomaly peak simulation in the dataset; Figure 2(h) is a schematic diagram of the simulation of SAR cone anomaly peaks in the dataset; Figure 2(i) is a schematic diagram of LRM ocean trailing edge noise simulation in the dataset; Figure 2(j) is a schematic diagram of the LRM tip trailing edge noise simulation in the dataset; Figure 2(k) is a schematic diagram of the SAR ocean trailing edge noise simulation in the dataset; Figure 2(l) is a schematic diagram of the simulated noise at the trailing edge of the SAR cone in the dataset; Figure 2 (m) is a schematic diagram of LRM ocean multi-peak simulation in the dataset; Figure 2(n) is a schematic diagram of LRM cone multi-peak simulation in the dataset; Figure 2(o) is a schematic diagram of SAR ocean multi-peak simulation in the dataset; Figure 2(p) is a schematic diagram of SAR cone multi-peak simulation in the dataset; Figure 2(q) is a schematic diagram of a pure cluttered signal simulation in the dataset; Figure 3 This is a schematic diagram of the Wave-Transformer model. Figure 4 This is a schematic diagram of the Transformer encoder structure; Figure 5 A schematic diagram of multi-head attention and scaled dot product attention; Figure 6(a) shows the loss curve during the model training process; Figure 6(b) shows the accuracy curve during model training. Detailed Implementation

[0019] The technical solutions provided in this application are further illustrated below with reference to the embodiments.

[0020] Example 1 To address the issue of complex waveforms in near-shore altimeters and the tendency for direct retracking to fail, this embodiment proposes a Transformer-based radar altimeter echo classification method, comprising the following steps: Step A: Simulate the echo waveforms of various radar altimeters, including those from traditional pulse-finite radar altimeters and synthetic aperture radar altimeters. Then, superimpose noise of different intensity levels to simulate real echoes and construct a dataset.

[0021] Step B: Construct a Transformer-based radar altimeter echo classifier. This classifier consists of three parts: signal serialization and position encoding, a Transformer encoder, and a classification head. The signal serialization and position encoding part divides an echo signal into 9 patches and maps each patch to a vector space using a linear projection layer. The encoder part computes multi-head attention in parallel on the encoded patches and connects the multi-head outputs to a feedforward neural network. The classification head part calculates the probability vector of the classification result using the softmax function (normalized exponential function) on the output of the feedforward neural network.

[0022] Step C: Set the loss function and training hyperparameters for the classifier, and train it using the simulation waveform dataset generated in Step A to obtain the optimal network parameter model.

[0023] Step D: Load the network parameter model, input the waveform to be classified into the classifier, and obtain the classification label.

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] The solution provided in this embodiment consists of two parts: first, generating waveform datasets of radar altimeters for various systems; then, constructing a classifier based on the Transformer model; subsequently, setting the loss function and training hyperparameters, and training the network using the dataset to obtain the optimal model. The flowchart is as follows... Figure 1 As shown.

[0026] Step 1: Creating a dataset Step A: To prevent model overfitting, this invention independently constructed a dataset containing 35,409 waveforms, which adopts a combination of model simulation and real data. Specifically, it includes 17 categories such as standard ocean / cone waveforms from traditional pulse finite altimeters and synthetic aperture radar altimeters, ocean / cone anomalous peaks, ocean / cone trailing edge noise, ocean / cone multi-peaks, and pure cluttered signals.

[0027] Step A1: LRM Standard Ocean Waveform Simulation The echo model, using the first-order exponential approximation of the Brown model, is used to simulate the "standard ocean" waveform of a traditional altimeter. : (1) in, Let the envelope function of the echo be (defined as in equation (2)). The shape function of the echo body (defined as in equation (9)). , is the deformation function caused by the non-Gaussian probability density of wave height (defined as in equation (10)). For the thermal noise of the radar system, t The independent variable is time.

[0028] envelope function of the echo Defined as: (2) in, Echo scaling factor This refers to the antenna's misdirection angle. , is the antenna beam effect factor (defined as in equation (3)). To take into account the mixed beamwidth and attitude angle standard deviation of the beam height (defined as in equation (4)). This is the width parameter of the error function.

[0029] (2) In the formula, the antenna beam effect factor for: (3) In equation (3), This represents the 3dB beamwidth of the radar antenna.

[0030] (2) In the formula, the standard deviation d of the mixed beam height, taking into account the beamwidth and attitude angle, is: (4) in, For the cosine correction term of the misdirection angle, The standard deviation of the mixed wave height in the dimension of time. This is the sine correction term for the misdirection angle.

[0031] (4) In the formula, the cosine correction term for the misdirection angle for: (5) in, The speed of light in a vacuum. For the orbital height after curvature correction: (6) in, This is the satellite's nominal orbital altitude. The radius is the Earth's radius.

[0032] (4) In the formula, the sine correction term for the misdirection angle for: (7) (2) In the formula, the width parameter of the error function for: ; (8) in, This represents the time delay corresponding to the mean sea level.

[0033] (1) In the formula, the echo body shape function The definition is as follows: (9) in, Let be the error function. The skewness of the wave height probability density. The standard deviation of sea surface wave height is expressed in terms of time. The standard deviation of the mixed wave height in the dimension of time. In equation (1), the deformation function caused by the non-Gaussian probability density of the wave height is... The definition is as follows: (10) This model is widely used for ground data retracking of Jason series satellite altimeters.

[0034] Step A2: LRM Standard Conical Waveform Simulation An adaptive model is used to simulate the "standard cone-shaped" waveform of a traditional altimeter. The model's echo power... Defined as the convolution of the Sea Surface Response (SSR) and the Radar System Point Target Response (PTR): (11) Where * represents convolution operation. It is the convolution of the Flat Sea Surface Response (FSSR) and the probability density function (PDF), and its formula is as follows: (12) in, Echo scaling factor As the independent variable for time, The parameter is determined by the root mean square slope of the sea surface. This represents the time delay corresponding to the mean sea level. This refers to the thermal noise of the radar system; all other parameters are defined in the same way as the model defined in step A1. This model enhances its adaptability to different types of echo signals.

[0035] Step A3: Simulation of SAR Standard Ocean Waveform and Standard Cone Waveform The official SAMOSA (SAR Altimetry Mode Studies and Applications) model of the Sentinel-3 / 6 altimeter was used to simulate standard ocean and standard cone waveforms of the synthetic aperture radar altimeter. An echo model was established by integrating the radar equations and the probability distribution of the sea surface effective wave height. And it is simplified to the corresponding parsing form, the expression of which is as follows: (13) in, and These represent the range cell index and the azimuth Doppler frequency index of the echo model, respectively. This represents the peak echo power. This is the echo cell index corresponding to the echo mean sea level. For the effective wave height, The normalized standard deviation of sea surface wave height. For the first The distance unit, the first Fresnel reflection coefficients of the sea surface at Doppler frequencies in each azimuth direction. Indicates the echo model with respect to k The zeroth-order expansion term, Indicates the echo model with respect to First-order expansion term, These are physical quantities related to the antenna pattern. The mixed parameters are determined by the resolution of the radar altimeter: (14) in, This represents the normalized width of the satellite's response to a point target along its orbit, approximated by a Gaussian approximation. This represents the normalized width of the point target response to the Gaussian approximation of the track intersection. Indicates the satellite's resolution along its orbit. Indicates the cross-track resolution. The standard deviation of sea surface wave height is the normalized value.

[0037] In the echo model, the Doppler frequency index It is through the Doppler frequencies corresponding to different sub-views in the model. Calculated: (15) in, This represents the pulse repetition frequency. This represents the number of pulses within a Burst. This indicates the Doppler frequency corresponding to different sub-beams.

[0038] The Doppler frequencies corresponding to different sub-beams are: (16) in, Indicates satellite speed. Indicates the carrier wavelength. Indicates the Doppler angles corresponding to different sub-beams. Defined as: (17) in, This represents the number of views processed by the model in multi-view processing. It is the angular spacing between adjacent sub-beams, determined by the satellite velocity. burst repeat interval The satellite orbital altitude after curvature correction (Definition as shown in equation (6)) The expression is as follows: (18) Step A4: Abnormal peak waveform simulation: (19) in, This represents the final output signal value of the abnormal peak echo waveform. It is the amplitude of Gaussian noise; It is the location of Gaussian noise; It is the standard deviation of Gaussian noise. Time is the independent variable. It is a standard echo waveform; It is the standard deviation of Rayleigh noise. Let represent a random variable with a Rayleigh distribution.

[0039] Step A5: Echo Simulation with Trailing Edge Noise To simulate trailing edge peaks, multiple Gaussian and Rayleigh noises of varying intensities are superimposed on various standard waveforms. The calculation formula is as follows: (20) in, The final output signal value, including the trailing edge noise echo waveform, is superimposed with a total of n Gaussian noise. For Gaussian noise index, For the first The amplitude of a Gaussian noise; For the first The location of Gaussian noise; For the first One standard deviation of Gaussian noise; The total number of Gaussian noises; This represents the overall amplitude of the trailing edge noise (to control the waveform height from being too high). This is the standard echo waveform; Given the standard deviation of Rayleigh noise, the required values ​​are: The spacing between them is less than 20 gates.

[0040] Step A6: Echo Simulation with Multiple Peaks To simulate abnormal peaks, multiple Gaussian and Rayleigh noises of varying intensities are superimposed on various standard waveforms. The calculation formula is as follows: (twenty one) in, The final output signal value of the multi-peak echo waveform is superimposed with n Gaussian noises. For the first The amplitude of a Gaussian noise; For the first The location of Gaussian noise; For the first One standard deviation of Gaussian noise; The total number of Gaussian noises; This is the standard echo waveform; Given the standard deviation of Rayleigh noise, the required values ​​are: The distance between them must be far enough. They do not overlap.

[0041] The total data volume consists of 17 subcategories of echo data, with a total of 35,409 echo data. The waveform diagrams of the entire dataset are shown in Figures 2(a) to 2(q).

[0042] Step 2: Construct a Transformer-based waveform classifier Step B: This invention employs an Encoder-only structure to extract the "semantic information" of the waveforms. Since the purpose of this invention is to extract features between waveforms for classification, and it does not involve generative results, the Decoder process is omitted. Furthermore, the special mechanism of multi-head attention is insensitive to positional information, and the matrix composed of tokens lacks "temporal" information. Therefore, a positional encoding is added before vectorization, allowing the Transformer to remember the temporal information, such as... Figure 3 As shown.

[0043] Step B1: Signal Serialization and Position Encoding The original one-dimensional radar echo signal is denoted as ,in, This represents the length of the signal. (The signal is then...) Divided into There are 3 consecutive, non-overlapping patches. Each patch has a length of 1. Subsequently, each patch is mapped to a trainable linear projection layer. A dimensional vector space.

[0044] The formula for this process is expressed as follows: (twenty two) in, Representing the One signal block, For signal block index. It is a linear projection matrix. It is the embedding of the obtained serialized signal.

[0045] To preserve the sequence's order information, a positional code with the same dimension as the signal embedding is generated. And add it to the signal embedding: (twenty three) in It is the embedding sequence input to the Transformer encoder. The initial value. Position coding is typically calculated using sine and cosine functions of different frequencies: (twenty four) (25) in It is a position index. It is a dimensional index. Represents the positional encoding for even-numbered dimensions. This represents the positional encoding for odd-dimensional elements.

[0046] Step B2: Transformer Encoder Initial embedding It is fed into a Transformer encoder consisting of 8 stacked layers, as shown in the diagram below. Each layer... The calculation is as follows: (26) (27) in, For the first Temporary sequence of layers, For the first Layer output sequence, For the first The output sequence of the layer Transformer encoder, This represents a multi-head self-attention function. Indicates the layer regularization function, This represents a feedforward network, which is Figure 4 The specific implementation of MLP (Multilayer Perceptron) is defined in the form of equation (29).

[0047] The core of multi-head self-attention is scaled dot product attention. For each head... Self-attention for: (28) in, , and It is an embedding sequence of the Transformer encoder (Iteration quantity) The query, key, and value matrix obtained through linear transformation, with superscript... T It is a matrix transpose operation. This refers to the dimension of the key vector. The outputs of multiple heads are concatenated and then subjected to another linear transformation, such as... Figure 4 and Figure 5 As shown.

[0048] Feedforward network For a simple two-layer neural network: (29) in, This is the learnable weight matrix of the first layer of the feedforward network. This is the learnable bias vector of the first layer of the feedforward network. This is the learnable weight matrix for the second layer of the feedforward network. This is the learnable bias vector for the second layer of the feedforward network. The Gaussian Error Linear Unit (Gaussian Error Linear Unit) is the activation function for Gaussian Error Linear Units. Compared to the traditional ReLU (Rectified Linear Unit) activation function, this function is smoother, has no "dead neurons," and incorporates probability regularization, which can significantly alleviate problems such as vanishing / exploding gradients and improve the performance of deep network training.

[0049] Step B3: Sorting Head Extracting the feature vector of the first position in the output sequence of the Transformer encoder The data is then fed into a linear classifier, and finally the class probability distribution is obtained through the Softmax function. (30) in, It is the learnable weight matrix of the classification layer. ; It is the learnable bias vector of the classification layer. , The encoder outputs the feature vector of the first position in the sequence. This is the total number of categories. Final prediction. It is a probability vector.

[0050] in and These are the parameters of the classification layer. This is the total number of categories. Final prediction. It is a probability vector.

[0051] Step 3: Training the network Step C: Set the loss function and training hyperparameters for the waveform classifier, and train it using the waveform simulation dataset generated in Step A to obtain the optimal network parameter model.

[0052] Step C1: The model uses the cross-entropy loss function: (31) in, It is the number of samples within a batch. For sample index, It is the total number of categories. It is a category index. It is the first The sample corresponds to the first One-hot tags for categories, It is the first The sample belongs to the first The probability distribution of the categories.

[0053] Step C2: Update network parameters using the Adam optimizer, setting the learning rate (lr) to 1e-4 and the batch size to 32. Initialize training parameters, load the training and validation sets, create the network structure, and train the network to obtain the optimal parameter model for the hyperparameter prediction network.

[0054] The training results show that the accuracy of the test set is 89.16%, as shown in Figure 6(a), Figure 6(b) and Table 1.

[0055] The F1 score for all categories is 0.78 or higher, especially for the classification results of LRM standard ocean waveforms, SAR standard ocean waveforms, LRM anomalous peaks, SAR anomalous peaks, LRM trailing edge noise, and SAR trailing edge noise, which reached over 0.90. The high precision, low recall, and low F1 score for cone-shaped waveforms indicate that most of the predicted positive results are correct, but many true positive samples are not being detected. Future work will focus on improving the model's ability to identify positive samples through data augmentation (simulating cone-shaped waveforms with different noise levels).

[0056] Precision represents the proportion of samples that the model predicts to belong to a certain class but actually belong to that class; Recall represents the proportion of samples that actually belong to a certain class but were correctly predicted by the model; and F1-score is the harmonic mean of precision and recall.

[0057] Step 4: Classification of Real Waveforms Step D: Load the network parameter model, input the waveform to be classified into the classifier, and obtain the classification label. Echoes were obtained using Jason-3 SGDR, Sentinel-3, and Sentinel-6HR radar altimeters for classification statistics. The classification results vary with distance from the shore, as shown in Tables 2, 3, and 4. Table 1: Precision, Recall, and F1 Score Evaluation Table

[0058] Table 2: LRM Waveform Classification Percentage Statistics (Unit: %)

[0059] Table 3: SAR Waveform Classification Percentage Statistics (Unit: %)

[0060] Example 2 This embodiment provides a Transformer-based radar altimeter echo classification system to implement the Transformer-based radar altimeter echo classification method provided in Embodiment 1. This system is a unified classification model—the Wave-Transformer classifier—capable of simultaneously processing traditional pulsed finite echoes and synthetic aperture radar altimeter echoes, and possessing the ability to identify inter-class differences and intra-class details. The system includes: The dataset construction module is used to model and simulate various echo waveforms of radar altimeters of different systems. By superimposing noise of different intensity levels, it generates radar altimeter echo data of multiple categories, thereby constructing a radar altimeter echo dataset covering multiple systems and multiple forms. The signal processing module is used to serialize and encode the radar altimeter echo signals in the radar altimeter echo dataset, construct an input sequence, and input the input sequence into a radar altimeter echo classification model based on a Transformer encoder structure, where feature extraction and classification are performed through a multi-head self-attention mechanism. The model training and optimization module is used to train the radar altimeter echo classification model based on the Transformer encoder structure based on the radar altimeter echo dataset. It optimizes the parameters of the radar altimeter echo classification model by setting the loss function and training hyperparameters to obtain the trained radar altimeter echo classification model. The classification module is used to input the radar altimeter echo signal to be classified into the trained radar altimeter echo classification model and output the corresponding radar altimeter echo classification result.

[0061] As can be seen from the above detailed description of this application: 1. This invention simulates various forms of echoes from altimeters of different systems, and for the first time constructs a simulation dataset containing 17 categories, improving the diversity and completeness of the dataset.

[0062] 2. In the encoder, a multi-head attention mechanism is introduced to enhance the network's ability to extract weak features at different scales.

[0063] The application scenario of this invention is that radar altimeter echoes of any type can be input into the classifier of this invention, and the corresponding classification label can be output. This can expand the application scope of satellite altimetry data and enhance its contribution to operational applications and scientific research in the field of international oceanography.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A radar altimeter echo classification method based on Transformer, characterized in that, include: Step 1: Model and simulate various echo waveforms of radar altimeters of different systems. By superimposing noise of different intensity levels, generate radar altimeter echo data of multiple categories, thereby constructing a radar altimeter echo dataset covering multiple systems and multiple forms. Step 2: Serialize and position-encode the radar altimeter echo signals in the radar altimeter echo dataset to construct an input sequence. Input the input sequence into a radar altimeter echo classification model based on a Transformer encoder structure. Extract features and perform classification through a multi-head self-attention mechanism. Step 3: Train the radar altimeter echo classification model based on the Transformer encoder structure based on the radar altimeter echo dataset. Optimize the parameters in the radar altimeter echo classification model based on the Transformer encoder structure by setting the loss function and training hyperparameters to obtain the trained radar altimeter echo classification model. Step 4: Input the radar altimeter echo signal to be classified into the trained radar altimeter echo classification model, and output the corresponding radar altimeter echo classification result.

2. The radar altimeter echo classification method based on Transformer according to claim 1, characterized in that, In step 1, the different types of radar altimeters include: conventional pulse-finite radar altimeters and synthetic aperture radar altimeters; the noise includes: Gaussian noise and Rayleigh noise; the multi-category radar altimeter echo data includes: standard ocean waveform, standard cone waveform, anomalous peak waveform, trailing edge noise waveform, multi-peak waveform, and cluttered signal waveform.

3. The radar altimeter echo classification method based on Transformer according to claim 2, characterized in that, The expression for the simulation model of the standard ocean waveform of the conventional pulse finite radar altimeter. for: ； in, Let be the envelope function of the echo. For the shape function of the echo body, Let be the deformation function caused by the non-Gaussian probability density of the wave height. For the thermal noise of the radar system, Let time be the independent variable; where, The envelope function of the echo for: ； in, Echo scaling factor This refers to the antenna's misdirection angle. Antenna beam effect factor, To account for the mixed beamwidth and attitude angle standard deviation of the beam height, is the width parameter of the error function; where, The antenna beam effect factor for: ;in, This represents the 3dB beamwidth of the radar antenna. The width parameter of the error function for: ,in, This represents the time delay corresponding to the mean sea level. The mixed beam height standard deviation taking into account beamwidth and attitude angle for: ; in, For the cosine correction term of the misdirection angle, The standard deviation of the mixed wave height in the dimension of time. This is the sine correction term for the misdirection angle; where, , The speed of light in a vacuum. To determine the orbital height after curvature correction. , This is the satellite's nominal orbital altitude. The radius of the Earth; ； The echo body shape function for: ； in, Let be the error function. The skewness of the wave height probability density. The standard deviation of sea surface wave height is expressed in terms of time. The standard deviation of the mixed wave height in the dimension of time. The deformation function caused by the non-Gaussian probability density of the wave height for: 。 4. The radar altimeter echo classification method based on Transformer according to claim 3, characterized in that, The expression for the simulation model of the standard cone-shaped waveform of the conventional pulse finite radar altimeter. for: ； Where * denotes convolution operation, It is the impulse response and probability density function of a flat sea surface. The convolution; where, ； in, Echo scaling factor Time is the independent variable. The parameter is determined by the root mean square slope of the sea surface. This represents the time delay corresponding to the mean sea level. It is the thermal noise of the radar system.

5. The radar altimeter echo classification method based on Transformer according to claim 4, characterized in that, The expression of the simulation model of the echo waveform of the synthetic aperture radar altimeter. for: ； in, Indicates the distance cell index of the echo model. This represents the azimuth Doppler frequency index of the echo model. This represents the peak echo power. This is the echo cell index corresponding to the echo mean sea level. For the effective wave height, The normalized standard deviation of sea surface wave height. For the first The distance unit and the Fresnel reflection coefficients of the sea surface at Doppler frequencies in each azimuth direction. Indicates the echo model with respect to k The zeroth-order expansion term, Indicates the echo model with respect to First-order expansion term, Physical quantities related to the antenna pattern; These are mixed parameters determined by the resolution of the radar altimeter; ； in, This represents the normalized width of the satellite's response to a point target along its orbit, approximated by a Gaussian approximation. This represents the normalized width of the point target response to the Gaussian approximation of the track intersection. Indicates the satellite's resolution along its orbit. Indicates the cross-track resolution. The normalized standard deviation of sea surface wave height; ； in, Indicates the pulse repetition frequency. This represents the number of pulses within a Burst. Indicates the Doppler frequencies corresponding to different sub-beams. , Indicates satellite speed. Indicates the carrier wavelength. Indicates the Doppler angles corresponding to different sub-beams. , This represents the number of views processed by the model in multi-view processing. It is the angular spacing between adjacent sub-beams. , For satellite speed, For burst repetition interval, This is the satellite orbital altitude after curvature correction.

6. The radar altimeter echo classification method based on Transformer according to claim 5, characterized in that, The expression of the simulation model of the abnormal peak waveform for: ； in, It is the amplitude of Gaussian noise; It is the location of Gaussian noise; It is the standard deviation of Gaussian noise. Time is the independent variable. It is a standard echo waveform; It is the standard deviation of Rayleigh noise. Represents a random variable with a Rayleigh distribution; The expression of the simulation model of the trailing edge noise waveform for: ； in, For Gaussian noise index, For the first The amplitude of a Gaussian noise; For the first The location of Gaussian noise; For the first One standard deviation of Gaussian noise; The total number of Gaussian noises; This represents the overall amplitude of the trailing edge noise. The expression of the simulation model of the multi-peak waveform for: 。 7. The radar altimeter echo classification method based on Transformer according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Divide the radar altimeter echo signal into multiple continuous and non-overlapping signal segments, and map them into a vector sequence through linear projection; Step 2.2: Perform position encoding processing on the vector sequence to characterize the timing information of the radar altimeter echo signal; Step 2.3: Input the position-encoded vector sequence as the embedding sequence into the radar altimeter echo classification model based on the Transformer encoder structure, extract features through a multi-head self-attention mechanism and output the classification result.

8. The radar altimeter echo classification method based on Transformer according to claim 7, characterized in that, In step 2.2, the encoding is performed using the following positional encoding method: ； ； in, Represents the positional encoding for even-numbered dimensions. Represents the positional encoding for odd-dimensional elements; Will and Add the vector sequence to the corresponding dimension to obtain the encoded vector sequence; In step 2.3, the radar altimeter echo classification model based on the Transformer encoder structure includes multiple stacked Transformer encoders, and the calculation of each Transformer encoder layer is as follows: ； ； in, For the layer index of the Transformer encoder, For the first Temporary sequences of layer Transformer encoders, For the first The output sequence of the layer Transformer encoder, For the first The output sequence of the layer Transformer encoder, This represents a multi-head self-attention function. Indicates the layer regularization function, Indicates a feedforward network; The radar altimeter echo classification model based on the Transformer encoder structure extracts features through a multi-head self-attention mechanism. For each head of the multi-head self-attention mechanism... Self-attention for: ； in, , and These are the query, key, and value matrices obtained by linear transformation of the embedding sequence from the Transformer encoder, with superscripts... T It is a matrix transpose operation. It is the dimension of the key vector; The radar altimeter echo classification model based on the Transformer encoder structure uses a feedforward neural network. Perform nonlinear mapping: ； in, This is the learnable weight matrix of the first layer of the feedforward network. This is the learnable bias vector of the first layer of the feedforward network. This is the learnable weight matrix for the second layer of the feedforward network. This is the learnable bias vector for the second layer of the feedforward network. It is the Gaussian error linear unit activation function. For embedded sequences; The radar altimeter echo classification model based on the Transformer encoder structure outputs classification results through a classification layer: ； in, It is the learnable weight matrix of the classification layer; It is the learnable bias vector of the classification layer. The encoder outputs the feature vector of the first position in the sequence. The total number of categories of radar altimeter echo data.

9. The radar altimeter echo classification method based on Transformer according to claim 1, characterized in that, In step 3, the loss function includes: cross-entropy loss function. : ； in, It is the total number of radar altimeter echo data within a batch. For indexing radar altimeter echo data, It is the total number of categories of radar altimeter echo data. It is a category index. It is the first The radar altimeter echo data corresponds to the first... One-hot tags for categories, It is the first The radar altimeter echo data belongs to the first... The probability distribution of the categories.

10. A radar altimeter echo classification system based on Transformer, comprising: The dataset construction module is used to model and simulate various echo waveforms of radar altimeters of different systems. By superimposing noise of different intensity levels, it generates radar altimeter echo data of multiple categories, thereby constructing a radar altimeter echo dataset covering multiple systems and multiple forms. The signal processing module is used to serialize and encode the radar altimeter echo signals in the radar altimeter echo dataset, construct an input sequence, and input the input sequence into a radar altimeter echo classification model based on a Transformer encoder structure, where feature extraction and classification are performed through a multi-head self-attention mechanism. The model training and optimization module is used to train the radar altimeter echo classification model based on the radar altimeter echo dataset. By setting the loss function and training hyperparameters, the module optimizes the parameters in the radar altimeter echo classification model based on the Transformer encoder structure to obtain the trained radar altimeter echo classification model. and The classification module is used to input the radar altimeter echo signal to be classified into the trained radar altimeter echo classification model and output the corresponding radar altimeter echo classification result.

Images