Internal solitary wave identification method based on synthetic aperture radar altimeter data

Abstract

The invention relates to the technical field of ocean remote sensing, and discloses an internal solitary wave identification method based on synthetic aperture radar altimeter data, which comprises the following steps of: acquiring a multi-dimensional feature component observed by a Sentinel-3 satellite synthetic aperture radar altimeter, and preprocessing the multi-dimensional feature component; sequentially performing frequency domain band-pass filtering and adaptive sliding standard deviation operation on the preprocessed multi-dimensional feature components, and constructing an input tensor composed of four feature components after normalization processing; a deep learning model ISW-DetSRALNet is constructed, and the model is trained; and inputting the input tensor into the model which is qualified in training, and automatically identifying internal solitary wave information by the model. According to the method disclosed by the invention, ocean background noise interference can be effectively inhibited, the accuracy and robustness of internal solitary wave detection are remarkably improved, and technical support is provided for ocean dynamic environment monitoring, internal solitary wave disaster early warning and ocean engineering safety.

Description

Technical Field

[0001] The present invention relates to the field of ocean remote sensing technology, and in particular to an internal solitary wave identification method based on synthetic aperture radar altimeter data. Background Art

[0002] Internal solitary waves are a significant phenomenon that influences ocean mixing, nutrient transport, and marine engineering safety. With advances in ocean observation technology, the precise observation and parameter inversion of internal solitary waves have become key issues in marine science research and practical applications. Existing remote sensing methods, such as optical remote sensing and synthetic aperture radar, have significant limitations in their three-dimensional detection of internal solitary waves and their ability to invert vertical information, making them inadequate for current demands for refined ocean observations.

[0003] In recent years, satellite altimeter technology has developed rapidly, especially the Sentinel-3 satellite launched in 2016. The synthetic aperture radar altimeter (SRAL) on board has several technical advantages, making it show great potential in internal solitary wave observation applications: first, it has a high along-track spatial resolution (about 300 m), which can capture small and medium-scale sea surface changes and is suitable for the fine detection of solitary waves in local areas; second, it can directly measure sea surface height anomalies and other relevant geophysical parameters, providing a direct means of observing the vertical motion of internal solitary waves.

[0004] However, the Sentinel-3 satellite still faces several technical challenges in practical applications: First, the signals of internal solitary waves are usually weak in SRAL data and easily interfered with by background noise. There is an urgent need to develop an internal solitary wave signal extraction algorithm suitable for the characteristics of SRAL data. Second, most existing internal solitary wave detection methods rely on a single variable or artificial experience to set parameter thresholds, which makes it difficult to adapt to complex ocean environments and changeable internal solitary wave characteristics, limiting the adaptability and generalization ability of the detection algorithm. Third, there is a lack of an internal solitary wave detection model that can fully utilize multi-dimensional characteristic components, which cannot effectively integrate the coupling information between different physical quantities, resulting in insufficient detection accuracy and robustness.

[0005] Therefore, it is urgent to propose an internal solitary wave detection method based on Sentinel-3 SRAL data. By integrating multi-dimensional characteristic components and introducing deep learning algorithms, high-precision and high-robustness automatic detection and identification of internal solitary waves can be achieved, the accuracy and reliability of internal solitary wave information acquisition can be improved, and the application capabilities of Sentinel-3 satellites in the field of internal solitary wave observations can be expanded. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides an internal solitary wave identification method based on synthetic aperture radar altimeter data, which realizes high-precision automatic identification and extraction of internal solitary waves.

[0007] To achieve the above object, the technical solution of the present invention is as follows: A method for identifying internal solitary waves based on synthetic aperture radar altimeter data comprises the following steps: Step 1: Collect the multidimensional characteristic components of the Sentinel-3 satellite synthetic aperture radar altimeter observations, including sea surface height anomaly, backscatter coefficient, significant wave height, and echo waveform power, and preprocess the multidimensional characteristic components; Step 2: The pre-processed multi-dimensional feature components are subjected to frequency domain bandpass filtering and adaptive sliding standard deviation calculation in sequence, and then normalized to construct an input tensor consisting of four feature components; Step 3: Build and train the deep learning model ISW-DetSRALNet, which includes an input module, a convolutional feature extraction module, a time series modeling module, a residual time series fusion module, a frequency-time joint attention mechanism module, and a classification decision output module. Step 4: Input the input tensor into the trained model, and the model automatically identifies the internal solitary wave information.

[0008] In the above scheme, the convolution feature extraction module uses a one-dimensional convolutional neural network to extract local perturbation features. It contains four layers, the convolution kernel size is set to 3, the number of convolution output channels is 16, 32, 64 and 128 respectively, and the ReLU activation function is used. The final output tensor of the convolution feature extraction module is ,in, Represents the length of the time series used by the model.

[0009] In the above scheme, the time series modeling module is based on the output of the convolution feature extraction module by the bidirectional LSTM module. Perform bidirectional time series modeling, where the number of hidden layer units in the time series modeling module is 64 and the output dimension is 128; The time series modeling module introduces two independent LSTM units, forward and reverse: Forward LSTM unit: input sequence , output ; Represents the feature vector at time step L after being processed by the convolutional feature extraction module; is the hidden state vector calculated by the forward LSTM unit based on the current and previous input features at time step t; Backward LSTM unit: input sequence , output ; is the hidden state vector calculated by the backward LSTM unit based on the current and previous input features at time step t; The outputs from the two directions are spliced ​​together to obtain: , H=64 is the number of hidden layer units in each direction, and the total output dimension is 128; Represents the context feature vector that integrates forward and reverse temporal information at time step t; The output tensor for the entire time series is: , is the bidirectional context feature vector at the last time step 𝐿.

[0010] In the above scheme, the residual time series fusion module combines the output of the convolution feature extraction module Output from the timing modeling module Add the element-wise residuals and get .

[0011] In the above scheme, the frequency domain-time domain joint attention mechanism module includes a frequency domain branch, a time domain branch and a joint attention calculation module; The frequency domain branch pair is obtained Perform fast Fourier transform and get , after taking the amplitude Reduce the dimension to 32 and get the frequency domain attention component ;in, is the fast Fourier transform, represents the complex amplitude, represents convolution processing; The time domain branch will obtain Perform linear mapping to obtain the time domain query vector ,in, is the trainable weight matrix, is the trainable bias vector; The joint attention calculation module converts the time domain query vector and frequency domain attention component After addition, the time-step attention weight vector is obtained by Softmax normalization ; and output the weighted context vector ;in, Representative The attention weight of the sampling points, Indicates the 128-dimensional feature vector for each time step.

[0012] In the above scheme, the classification decision output module outputs the weighted context vector Input to the fully connected layer for dimensionality reduction, the output dimension is 1, and the Sigmoid activation function is used to output the The classification probability of the sampling points : ; in, is the weight matrix of the fully connected layer, used to map the 128-dimensional context vector to a scalar output; is the bias term of the fully connected layer; For the Sigmoid activation function: ; like > ,but , which corresponds to the existence of internal solitary wave disturbance in the current sequence, is the threshold.

[0013] In the above scheme, step one specifically includes: Step 1.1, extracting the average value of the echo waveform power between the 25th and 45th echo gates as the echo waveform power; Step 1.2: Extract the backscatter coefficients of Ku band and C band from the original data set. For each time point, calculate the differential mean square slope between Ku band and C band. : ; in, is the C-band NRCS calibration value, and are constants, taking them as 0.427 and 0.617 respectively; and represent the backscatter coefficients of 20 Hz Ku band and 20 Hz C band respectively; Will It is added to the original feature set as a new one-dimensional variable, and together with the sea surface height anomaly, significant wave height, and echo waveform power, it constitutes the final multi-dimensional feature component; Step 1.3: Perform quality control and invalid point elimination on the extracted multi-dimensional feature vector samples, which specifically includes the following three sub-processes: (1) Based on the surface classification flags in the synthetic aperture radar altimeter product of the Sentinel-3 satellite, open sea samples with a value of 0 were screened, and land, water vapor mixing areas, and inland water bodies were eliminated; (2) The rainfall markers provided by the synthetic aperture radar altimeter of the Sentinel-3 satellite are used to determine the sample areas affected by severe convection or rainfall. The rainfall marker data are interpolated from 1 Hz resolution to 20 Hz using a linear interpolation method, and non-zero observation points in the interpolation results are deleted. (3) Perform shoreline masking, calculate the distance of each sample point from the shore, and eliminate data samples with a distance from the shore less than 5 km.

[0014] In the above scheme, in step 2, the method of performing frequency domain bandpass filtering on the pre-processed multi-dimensional feature components is as follows: (1) Perform fast Fourier transform on the preprocessed multidimensional feature components to map the signal from the time domain to the frequency domain; (2) Filter the obtained frequency domain data to retain only the frequency components within the 2-20 km scale; (3) Perform inverse fast Fourier transform on the filtered frequency domain data to restore the data to the time domain and obtain the data after bandpass filtering of the multi-dimensional feature components.

[0015] In the above scheme, in step 2, the specific method of performing adaptive sliding standard deviation calculation on the multi-dimensional feature components after frequency domain bandpass filtering is as follows: (1) Setting multiple sliding window sizes W to 3 to 21 points to construct a multi-scale sliding window group; (2) For the characteristic component sequence after frequency domain bandpass filtering, calculate the local mean under each sliding window W: ; Where W is the sliding window size, is a sequence of characteristic components, including significant wave height, sea surface height anomaly, differential mean square slope of backscatter coefficient, and echo waveform power; is the i-th data point in the window; is the local mean of the feature component sequence under window W; Through the local mean Compute the local standard deviation of the eigenvalue sequence: ; in, Represents the local standard deviation under the sliding window W; (3) For each local standard deviation obtained , further calculate its global variance value as a metric to evaluate the significance of the disturbance structure at the window scale; (4) The standard deviation data with the largest global variance is selected from the calculation results of all sliding windows W. The final global variance maximum standard deviation data of the four characteristic components of sea surface height anomaly, differential mean square slope of backscatter coefficient between Ku band and C band, significant wave height and echo waveform power are obtained as follows: 、 、 、 ; (5) What you will get 、 、 、 Perform minimum-maximum normalization processing respectively and map its value to interval, and finally obtain four characteristic components 、 、 、 , construct a unified input tensor .

[0016] In the above scheme, in step three, when training the model, ocean and land colorimeter images synchronized with the orbit of the Sentinel-3 satellite synthetic aperture radar altimeter are obtained. The images are manually visually interpreted to confirm whether internal solitary waves exist, and internal solitary wave labels are generated. The corresponding positions and data in the multidimensional feature components are located based on the longitude and latitude information of the internal solitary waves in the images. The multidimensional feature components at the corresponding positions are input into the model, and the model is trained using the internal solitary wave labels as output.

[0017] Through the above technical solution, the internal solitary wave identification method based on synthetic aperture radar altimeter data provided by the present invention has the following beneficial effects: 1. Based on multi-source physical component data provided by the Sentinel-3 Synthetic Aperture Radar Altimeter (SRAL), this method utilizes four key characteristic components—sea surface height anomaly, normalized radar backscatter coefficient, significant wave height, and echo waveform power—combined with targeted data preprocessing and disturbance feature highlighting strategies to characterize the disturbance characteristics of internal solitary waves on the sea surface in multiple dimensions, thereby improving the detection sensitivity and generalization performance of the model. 2. This invention uses a feature extraction method that combines a convolutional neural network (CNN) with a bidirectional long short-term memory network (BiLSTM) to simultaneously capture the local spatial patterns and long-term dependencies of ISW disturbances. It introduces a residual temporal fusion mechanism to alleviate the gradient decay and information loss problems in deep network training, enabling efficient fusion of multiple layers of features across time scales, thereby enhancing the model's adaptability to diverse waveform morphologies. Furthermore, a frequency-domain and time-domain joint attention mechanism is applied to adaptively assign feature weights at typical wavelength scales, highlighting the characteristics of ISW disturbances and improving detection accuracy and stability. 3. Through targeted data processing methods and dedicated deep learning models, this paper realizes the automated detection of internal solitary waves in Sentinel-3 satellite data, can extract and analyze their modulation characteristics, and provide reliable data support for marine engineering construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0019] Figure 1 A flow chart of an internal solitary wave identification method based on synthetic aperture radar altimeter data disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the deep learning model ISW-DetSRALNet disclosed in the present invention; Figure 3 Schematic diagram of the position of the internal solitary wave in the acquired OICL image; Figure 4 Schematic diagram of feature components input to the model for training; Figure 5 Schematic diagram of the detected internal solitary wave position output by the model. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] The present invention provides a method for identifying internal solitary waves based on synthetic aperture radar altimeter data. Figure 1 As shown, the following steps are included: Step 1: Collect multi-dimensional characteristic components of Sentinel-3 satellite synthetic aperture radar altimeter (SRAL) observations, including sea surface height anomaly (SSHA), backscatter coefficient (including Ku-band and C-band), significant wave height (SWH), and echo waveform power (WAFM), and pre-process the multi-dimensional characteristic components; Specifically include: Step 1.1: To capture the influence of internal solitary wave modulation on radar echo signal, the average value of echo waveform power between the 25th and 45th echo gates is extracted as the echo waveform power; In step 1.2, in order to address the weak modulation effect of internal solitary wave disturbances on sea surface roughness, the differential mean square slope of the Ku-band and C-band backscatter coefficients (NRCS) is further introduced as an enhancement feature to improve the sensitivity to perturbation signals. The specific processing flow is as follows: Extract raw NRCS data: Extract the Ku-band and C-band backscatter coefficients from the SRAL dataset (fields sig0_ku and sig0_c, respectively). Both variables have a 20 Hz resolution and are in dB.

[0022] Calculate the NRCS differential mean square slope: For each time point, calculate the differential mean square slope between the Ku band and the C band. The expression is: ;

[0023] in, is the C-band NRCS calibration value, and are constants, taking them as 0.427 and 0.617 respectively; and represent the backscatter coefficients of 20 Hz Ku band and 20 Hz C band respectively; After the difference operation, we get a one-dimensional sequence , whose length is consistent with the original NRCS variable, is used to represent the difference in the frequency response of internal solitary waves in terms of sea surface roughness modulation. It is added to the original feature set as a new one-dimensional variable, and together with the sea surface height anomaly (SSHA), significant wave height (SWH) and wave echo waveform power (WAFM) constitute the final multidimensional feature component.

[0024] Step 1.3: Perform quality control and invalid point elimination on the extracted multi-dimensional feature vector samples, which specifically includes the following three sub-processes: (1) Based on the surface classification flag (surf_class_20_ku) in the synthetic aperture radar altimeter product of the Sentinel-3 satellite, open ocean samples with a value of 0 were screened and land, water vapor mixing areas, and inland water bodies were excluded; (2) The rainfall flag (rain_flag_01_ku) provided by the synthetic aperture radar altimeter of the Sentinel-3 satellite is used to determine the sample area affected by severe convection or rainfall. The rainfall flag data is interpolated from 1 Hz resolution to 20 Hz using the linear interpolation method, and the observation points with non-zero values ​​in the interpolation results are deleted. (3) Perform shoreline masking, calculate the distance of each sample point from the shore, and eliminate data samples with a distance from the shore less than 5 km.

[0025] In step 2, the pre-processed multi-dimensional feature components are subjected to frequency domain bandpass filtering and adaptive sliding standard deviation calculation in sequence, and then normalized to construct an input tensor consisting of four feature components.

[0026] The pre-processed multi-dimensional characteristic components are subjected to frequency domain bandpass filtering to enhance the disturbance signal within the scale range of the inner solitary wave. The specific method is as follows: (1) Perform fast Fourier transform on the preprocessed multidimensional feature components to map the signal from the time domain to the frequency domain; (2) Filter the obtained frequency domain data to retain only the frequency components within the 2-20 km scale; (3) Perform inverse fast Fourier transform on the filtered frequency domain data to restore the data to the time domain and obtain the data after bandpass filtering of the multi-dimensional feature components.

[0027] The specific method of performing adaptive sliding standard deviation calculation on the multidimensional feature components after frequency domain bandpass filtering is as follows: (1) Setting the size W of multiple sliding windows to 3 to 21 points to construct a multi-scale sliding window group; (2) For the characteristic component sequence after frequency domain bandpass filtering, calculate the local mean under each sliding window W: ;

[0028] Where W is the sliding window size, is a sequence of characteristic components, including significant wave height, sea surface height anomaly, differential mean square slope of backscatter coefficient, and echo waveform power; is the i-th data point in the window; is the local mean of the feature component sequence under the window W; Through the local mean Compute the local standard deviation of the eigenvalue sequence: ;

[0029] in, Represents the local standard deviation under the sliding window W; (3) For each local standard deviation obtained , further calculate its global variance value as a metric to evaluate the significance of the disturbance structure at the window scale; (4) The standard deviation data with the largest global variance is selected from the calculation results of all sliding windows W. The final global variance maximum standard deviation data of the four characteristic components of sea surface height anomaly, differential mean square slope of backscatter coefficient between Ku band and C band, significant wave height and echo waveform power are obtained as follows: 、 、 、 ; (5) What you will get 、 、 、 Perform minimum-maximum normalization processing respectively and map its value to interval, and finally obtain four characteristic components 、 、 、 , construct a unified input tensor .

[0030] Step 3: Build a deep learning model ISW-DetSRALNet to automatically identify and extract internal solitary wave signals and train the model.

[0031] 1. Build the model: The model integrates convolutional neural networks and bidirectional long short-term memory networks, and introduces residual-temporal fusion (RTF) and frequency-time joint attention (FTJA) mechanisms. This improves the ability to capture sparse, weak signals such as internal solitary waves without adding the burden of additional hyperparameter tuning. Specifically, the model includes an input module, a convolutional feature extraction module, a time series modeling module, a residual-temporal fusion module, a frequency-time joint attention mechanism, and a classification decision output module. The functions of each module are as follows: (1) Input module The model input is a time series sample consisting of the four feature components processed in step 2. ,in, Represents the length of the time series used by the model, the default is 200. , represents the characteristic component of the input, that is .

[0032] (2) Convolutional feature extraction module The convolution feature extraction module uses a one-dimensional convolutional neural network to extract local perturbation features. It consists of four layers, with the convolution kernel size set to 3 and the number of convolution output channels being 16, 32, 64, and 128, respectively. All layers use the ReLU activation function. ; ; ; ; The ReLU activation function formula is: .

[0033] The final output tensor of the convolution feature extraction module is ,in, Represents the length of the time series used by the model.

[0034] (3) Timing modeling module The time series modeling module is based on the output of the convolutional feature extraction module by the bidirectional LSTM module. Bidirectional time series modeling is performed. The number of hidden layer units in the time series modeling module is 64 and the output dimension is 128. This module enhances the model's ability to capture the relative positions of the convergence and divergence zones of the sea surface modulated by internal solitary waves.

[0035] The time series modeling module introduces two independent LSTM units, forward and reverse: Forward LSTM unit: input sequence , output ; Represents the feature vector at time step L after being processed by the convolutional feature extraction module; is the hidden state vector calculated by the forward LSTM unit based on the current and previous input features at time step t; Backward LSTM unit: input sequence , output ; is the hidden state vector calculated by the backward LSTM unit based on the current and previous input features at time step t; The outputs from the two directions are spliced ​​together to obtain: , H=64 is the number of hidden layer units in each direction, and the total output dimension is 128; Represents the context feature vector that integrates forward and reverse temporal information at time step t; The output tensor for the entire time series is: , is the bidirectional context feature vector at the last time step 𝐿.

[0036] (4) Residual time series fusion module In order to strengthen the coupling between local perturbations and global time series, residual connection fusion is introduced. The residual time series fusion module combines the output of the convolutional feature extraction module Output from the timing modeling module Add the element-wise residuals and get The residual It can alleviate deep gradient attenuation and strengthen local-global feature coupling, which is beneficial for highlighting the internal solitary wave disturbance signal.

[0037] (5) Frequency-time joint attention mechanism module The frequency-time domain joint attention mechanism module includes a frequency domain branch, a time domain branch and a joint attention calculation module.

[0038] The frequency domain branch pair is obtained Perform fast Fourier transform and get , after taking the amplitude Reduce the dimension to 32 and get the frequency domain attention component ;in, is the fast Fourier transform, represents the complex amplitude, represents convolution processing; The time domain branch will get Perform linear mapping to obtain the time domain query vector ,in, is the trainable weight matrix, is the trainable bias vector; The joint attention calculation module converts the time domain query vector and frequency domain attention component After addition, the time-step attention weight vector is obtained by Softmax normalization ; and output the weighted context vector ;in, Representative The attention weight of the sampling points, Indicates the 128-dimensional feature vector for each time step.

[0039] The FTJA constructed by this model enables attention to perceive both the energy and waveform mutations in the 2-20 km frequency band, significantly improving the sensitivity of sparse internal solitary wave responses.

[0040] (6) Classification decision output module The classification decision output module will output the weighted context vector Input to the fully connected layer for dimensionality reduction, the output dimension is 1, and the Sigmoid activation function is used to output the The classification probability of the sampling points : ; in, is the weight matrix of the fully connected layer, used to map the 128-dimensional context vector to a scalar output; is the bias term of the fully connected layer; For the Sigmoid activation function: ; like > ,but , which corresponds to the existence of internal solitary wave disturbance in the current sequence, is the threshold.

[0041] 2. Model training: 1. Generate internal solitary wave labels for model training: (1) Synchronous OLCI image acquisition. Acquire OLCI Level-1B images within the SRAL observation time window of ±30 minutes, such as Figure 3 shown.

[0042] (2) Internal solitary wave location. After performing radiometric and atmospheric correction on the OLCI image, the presence of internal solitary waves is confirmed by manual visual interpretation. For images where internal solitary waves are detected, the latitude and longitude ranges where the internal solitary waves are detected are recorded. Figure 3The red line in the middle shows the Sentinel-3 SRAL detection track, which detected an internal solitary wave near 10.3°N, as shown by the blue line in the figure.

[0043] (3) Trajectory matching and sample extraction Using the above longitude and latitude ranges, the four acquired characteristic components are clipped by the longitude and latitude ranges, and the data points within the longitude and latitude ranges are regarded as “inner solitary wave samples”; the rest are “background samples”.

[0044] (4) Label assignment All “internal solitary wave samples” are assigned a positive label of label = 1, and all “background samples” are assigned a negative label of label = 0, and finally a supervised training dataset is obtained that corresponds one-to-one to the SRAL multidimensional feature parameters.

[0045] 2. Training configuration: Input sequence length: 200; feature dimension: 4; number of bidirectional LSTM hidden units: 64, single-layer structure, using bidirectional mode; dimension of the attention mechanism intermediate layer: 32; number of output categories: 2; optimizer: Adam, initial learning rate 0.001, L2 regularization term weight decay coefficient of 1e-5; batch size: 32, maximum number of training rounds is 1000; validation set ratio: 25%; support CUDA accelerated computing.

[0046] To suppress overfitting and improve training stability, the following strategies are introduced: an adaptive learning rate scheduling mechanism: when the F1-score of the validation set does not improve significantly within 10 consecutive rounds, the learning rate is automatically reduced to 1 / 10 of the original value; an early stopping mechanism: when the F1-score of the validation set does not improve within 25 consecutive rounds, training is terminated early.

[0047] 3. Model training and performance evaluation When training the model, Figure 4 The multi-dimensional feature components at the corresponding positions shown are input into the model, and the model is trained using the inner solitary wave labels as output.

[0048] During model training, real-time monitoring of indicators such as loss, accuracy, precision, recall, and F1-score during training and validation phases; The model output results are classified according to the true labels as follows: true positive (TP): the internal solitary wave actually exists and is correctly identified; false positive (FP): the internal solitary wave actually does not exist but is mistakenly identified as existing; true negative (TN): the internal solitary wave actually does not exist and is correctly identified; false negative (FN): the internal solitary wave actually exists but is mistakenly identified as not existing.

[0049] The following performance metrics are used to evaluate the model's detection capabilities: Precision, which indicates the proportion of predicted positive samples that are actually positive; Recall, which indicates the proportion of all true positive samples that are correctly identified; and F1-score, which is the harmonic mean of Precision and Recall and reflects the model's overall detection performance. Table 1 lists and compares the results of training with different feature parameter combinations. The results show that the proposed method achieves the highest prediction accuracy when all four parameter vectors are input into the model.

[0050] Table 1. Training results of different parameter combinations

[0051] Step 4: Input the input tensor into the trained model, and the model automatically identifies the internal solitary wave information.

[0052] The internal solitary wave information identified by the embodiment of the present invention is as follows: Figure 5 As shown, from Figure 5 It can be seen that the position of the internal solitary wave identified by the model (shown by the black cross) is consistent with Figure 3 The position of the internal solitary wave shown in is consistent with that in the figure, which proves the accuracy of the model.

[0053] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying internal solitary waves based on synthetic aperture radar altimeter data, characterized in that: The steps include: Step 1: Collect the multidimensional characteristic components of the Sentinel-3 satellite synthetic aperture radar altimeter observations, including sea surface height anomaly, backscatter coefficient, significant wave height, and echo waveform power, and preprocess the multidimensional characteristic components; Step 2: The pre-processed multi-dimensional feature components are subjected to frequency domain bandpass filtering and adaptive sliding standard deviation calculation in sequence, and then normalized to construct an input tensor consisting of four feature components; Step 3: Build and train the deep learning model ISW-DetSRALNet, which includes an input module, a convolutional feature extraction module, a time series modeling module, a residual time series fusion module, a frequency-time joint attention mechanism module, and a classification decision output module. Step 4: Input the input tensor into the trained model, and the model automatically identifies the internal solitary wave information.

2. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 1, characterized in that: The convolution feature extraction module uses a one-dimensional convolutional neural network to extract local perturbation features. It contains four layers, the convolution kernel size is set to 3, the number of convolution output channels is 16, 32, 64 and 128 respectively, and the ReLU activation function is used. The final output tensor of the convolution feature extraction module is ,in, Represents the length of the time series used by the model.

3. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 2, characterized in that: The time series modeling module is based on the output of the convolutional feature extraction module by the bidirectional LSTM module. Perform bidirectional time series modeling, where the number of hidden layer units in the time series modeling module is 64 and the output dimension is 128; The time series modeling module introduces two independent LSTM units, forward and reverse: Forward LSTM unit: input sequence , output ; Represents the feature vector at time step L after being processed by the convolutional feature extraction module; is the hidden state vector calculated by the forward LSTM unit based on the current and previous input features at time step t; Backward LSTM unit: input sequence , output ; is the hidden state vector calculated by the backward LSTM unit based on the current and previous input features at time step t; The outputs from the two directions are spliced ​​together to obtain: , H=64 is the number of hidden layer units in each direction, and the total output dimension is 128; Represents the context feature vector that integrates forward and reverse temporal information at time step t; The output tensor for the entire time series is: , is the bidirectional context feature vector at the last time step 𝐿.

4. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 3, characterized in that: The residual time series fusion module combines the output of the convolution feature extraction module Output from the timing modeling module Add the element-wise residuals and get .

5. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 4 is characterized in that: The frequency domain-time domain joint attention mechanism module includes a frequency domain branch, a time domain branch and a joint attention calculation module; The frequency domain branch pair is obtained Perform fast Fourier transform and get , after taking the amplitude Reduce the dimension to 32 and get the frequency domain attention component ;in, is the fast Fourier transform, represents the complex amplitude, represents convolution processing; The time domain branch will obtain Perform linear mapping to obtain the time domain query vector ,in, is the trainable weight matrix, is the trainable bias vector; The joint attention calculation module converts the time domain query vector and frequency domain attention component After addition, the time-step attention weight vector is obtained by Softmax normalization ; and output the weighted context vector ;in, Representative The attention weight of the sampling points, Indicates the 128-dimensional feature vector for each time step.

6. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 5, characterized in that: The classification decision output module outputs the weighted context vector Input to the fully connected layer for dimensionality reduction, the output dimension is 1, and the Sigmoid activation function is used to output the The classification probability of the sampling points : ； in, is the weight matrix of the fully connected layer, used to map the 128-dimensional context vector to a scalar output; is the bias term of the fully connected layer; For the Sigmoid activation function: ； like > ,but , which corresponds to the existence of internal solitary wave disturbance in the current sequence, is the threshold.

7. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 5, characterized in that: Step 1 specifically includes: Step 1.1, extracting the average value of the echo waveform power between the 25th and 45th echo gates as the echo waveform power; Step 1.2: Extract the backscatter coefficients of Ku band and C band from the original data set. For each time point, calculate the differential mean square slope between Ku band and C band. : ； in, is the C-band NRCS calibration value, and are constants, taking them as 0.427 and 0.617 respectively; and represent the backscatter coefficients of 20 Hz Ku band and 20 Hz C band respectively; Will It is added to the original feature set as a new one-dimensional variable, and together with the sea surface height anomaly, significant wave height, and echo waveform power, it constitutes the final multi-dimensional feature component; Step 1.3: Perform quality control and invalid point elimination on the extracted multi-dimensional feature vector samples, which specifically includes the following three sub-processes: (1) Based on the surface classification flags in the synthetic aperture radar altimeter product of the Sentinel-3 satellite, open sea samples with a value of 0 were screened, and land, water vapor mixing areas, and inland water bodies were eliminated; (2) The rainfall markers provided by the synthetic aperture radar altimeter of the Sentinel-3 satellite are used to determine the sample areas affected by severe convection or rainfall. The rainfall marker data are interpolated from 1 Hz resolution to 20 Hz using a linear interpolation method, and non-zero observation points in the interpolation results are deleted. (3) Perform shoreline masking, calculate the distance of each sample point from the shore, and eliminate data samples with a distance from the shore less than 5 km.

8. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 1, characterized in that: In step 2, the method for performing frequency domain bandpass filtering on the pre-processed multi-dimensional feature components is as follows: (1) Perform fast Fourier transform on the preprocessed multidimensional feature components to map the signal from the time domain to the frequency domain; (2) Filter the obtained frequency domain data to retain only the frequency components within the 2-20 km scale; (3) Perform inverse fast Fourier transform on the filtered frequency domain data to restore the data to the time domain and obtain the data after bandpass filtering of the multi-dimensional feature components.

9. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 1, characterized in that: In step 2, the specific method of performing adaptive sliding standard deviation calculation on the multi-dimensional feature components after frequency domain bandpass filtering is as follows: (1) Setting multiple sliding window sizes W to 3 to 21 points to construct a multi-scale sliding window group; (2) For the characteristic component sequence after frequency domain bandpass filtering, calculate the local mean under each sliding window W: ； Where W is the sliding window size, is a sequence of characteristic components, including significant wave height, sea surface height anomaly, differential mean square slope of backscatter coefficient, and echo waveform power; is the i-th data point in the window; is the local mean of the feature component sequence under window W; Through the local mean Compute the local standard deviation of the eigenvalue sequence: ； in, Represents the local standard deviation under the sliding window W; (3) For each local standard deviation obtained , further calculate its global variance value as a metric to evaluate the significance of the disturbance structure at the window scale; (4) The standard deviation data with the largest global variance is selected from the calculation results of all sliding windows W. The final global variance maximum standard deviation data of the four characteristic components of sea surface height anomaly, differential mean square slope of backscatter coefficient between Ku band and C band, significant wave height and echo waveform power are obtained as follows: 、 、 、 ; (5) What you will get 、 、 、 Perform minimum-maximum normalization processing respectively and map its value to interval, and finally obtain four characteristic components 、 、 、 , construct a unified input tensor .

10. The internal solitary wave identification method based on synthetic aperture radar altimeter data according to claim 1, characterized in that: In step three, when training the model, ocean and land colorimeter images synchronized with the orbit of the Sentinel-3 satellite synthetic aperture radar altimeter are obtained. The presence of internal solitary waves in the images is confirmed through manual visual interpretation, and internal solitary wave labels are generated. The corresponding positions and data in the multidimensional feature components are located by combining the longitude and latitude information of the internal solitary waves in the images. The multidimensional feature components at the corresponding positions are input into the model, and the model is trained using the internal solitary wave labels as output.

Images