An internal solitary wave identification method based on synthetic aperture radar altimeter data

Abstract

This invention relates to the field of marine remote sensing technology and discloses a method for identifying internal solitary waves based on synthetic aperture radar (SAR) altimeter data. The method includes the following steps: acquiring multidimensional feature components observed by the Sentinel-3 satellite SAR altimeter and preprocessing these components; sequentially performing frequency domain bandpass filtering and adaptive sliding standard deviation calculations on the preprocessed multidimensional feature components, and constructing an input tensor composed of four feature components after normalization; constructing a deep learning model ISW-DetSRALNet and training the model; inputting the input tensor into the successfully trained model, which automatically identifies internal solitary wave information. The method disclosed in this invention can effectively suppress marine background noise interference, significantly improve the accuracy and robustness of internal solitary wave detection, and provide technical support for marine dynamic environment monitoring, internal solitary wave disaster early warning, and marine engineering safety.

Description

Technical Field

[0001] This invention relates to the field of marine remote sensing technology, and in particular to a method for identifying internal solitary waves based on synthetic aperture radar altimeter data. Background Technology

[0002] Internal solitary waves are a significant phenomenon affecting ocean mixing, nutrient transport, and the safety of marine engineering. With advancements in ocean observation technology, accurate observation and parameter inversion of internal solitary waves have become crucial issues in marine scientific 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 performance and vertical information inversion capabilities for internal solitary waves, making it difficult to meet the current demands for refined ocean observation.

[0003] In recent years, satellite altimeter technology has developed rapidly, especially the Sentinel-3 satellite launched in 2016. Its Synthetic Aperture Radar Altimeter (SRAL) has several technological advantages, making it highly promising for internal solitary wave observation applications: First, it has high along-orbit spatial resolution (about 300 m), which can capture small- to medium-scale sea surface changes and is suitable for fine detection of solitary waves in local areas; Second, it can directly measure sea surface height anomalies and other related geophysical parameters, providing a direct means of observing the vertical motion of internal solitary waves.

[0004] However, Sentinel-3 satellite still faces several technical challenges in practical applications: First, the signals of internal solitary waves in SRAL data are usually weak and easily interfered with by background noise, so 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 single variables or manual experience to set parameter thresholds, which are difficult to adapt to complex marine environments and variable internal solitary wave characteristics, thus 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 feature 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 feature components and introducing deep learning algorithms, a high-precision and robust automatic detection and identification method for internal solitary waves can be achieved, thereby improving the accuracy and reliability of internal solitary wave information acquisition and expanding the application capabilities of Sentinel-3 satellite in the field of internal solitary wave observation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for identifying internal solitary waves based on synthetic aperture radar altimeter data, achieving high-precision automatic identification and extraction of internal solitary waves.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] An internal solitary wave identification method based on synthetic aperture radar altimeter data includes the following steps:

[0009] Step 1: Collect multidimensional feature components from the Sentinel-3 satellite synthetic aperture radar altimeter observations, including sea surface height anomaly, backscattering coefficient, significant wave height, and echo waveform power, and preprocess the multidimensional feature components.

[0010] Step 2: After processing the preprocessed multidimensional feature components, frequency domain bandpass filtering and adaptive sliding standard deviation calculation are performed sequentially, and after normalization, an input tensor consisting of four feature components is constructed.

[0011] Step 3: Construct the deep learning model ISW-DetSRALNet and train the model; the model includes an input module, a convolutional feature extraction module, a temporal modeling module, a residual temporal fusion module, a frequency domain-temporal joint attention mechanism module, and a classification decision output module;

[0012] Step four: Input the input tensor into the successfully trained model, and the model will automatically identify the internal solitary wave information.

[0013] In the above scheme, the convolutional feature extraction module uses a one-dimensional convolutional neural network to extract local perturbation features. It contains four layers with a kernel size of 3. The number of output channels is 16, 32, 64, and 128 respectively, all using the ReLU activation function. The final output tensor of the convolutional feature extraction module is... ,in, The length of the time series used in the model.

[0014] In the above scheme, the temporal modeling module is based on the output of the convolutional feature extraction module from the bidirectional LSTM module. Bidirectional temporal modeling is performed, with the temporal modeling module having 64 hidden layer units and an output dimension of 128.

[0015] The timing modeling module simultaneously incorporates two independent LSTM units: forward and backward.

[0016] Forward LSTM unit: Input sequence Output ; This represents the feature vector at time step L after processing by the convolutional feature extraction module. The hidden state vector is calculated by the feedforward LSTM unit at time step t based on the current and previous input features;

[0017] Backward LSTM unit: Input sequence Output ; The hidden state vector is calculated by the backward LSTM unit at time step t based on the current and previous input features;

[0018] The outputs from the two directions are concatenated to obtain: H=64 is the number of hidden layer units in each direction, and the total output dimension is 128; This represents the context feature vector at time step t, which integrates forward and backward temporal information.

[0019] The output tensor of the entire time series is: , This is the bidirectional contextual feature vector at the last time step 𝐿.

[0020] In the above scheme, the residual temporal fusion module converts the output of the convolutional feature extraction module into a single output. Output of the timing modeling module By adding the element-level residuals, we obtain .

[0021] In the above scheme, the frequency-time domain joint attention mechanism module includes a frequency domain branch, a time domain branch, and a joint attention calculation module;

[0022] The frequency domain branch pairs obtained Perform a fast Fourier transform to obtain After taking the amplitude, Dimensionality reduced to 32, frequency domain attention components are obtained. ;in, For Fast Fourier Transform, This represents a complex amplitude. Represents convolution processing;

[0023] The time-domain branch will obtain Perform a linear mapping to obtain the time-domain query vector. ,in, For trainable weight matrix, It is a trainable bias vector;

[0024] The joint attention calculation module will use the time-domain query vector With frequency domain attention components After addition, the attention weight vector at each time step is obtained by Softmax normalization. And output the weighted context vector. ;in, Representative Attention weights for each sampling point Indicates the first A 128-dimensional feature vector at each time step.

[0025] In the above scheme, the classification decision output module outputs a weighted context vector. The input is fed into a fully connected layer for dimensionality reduction, the output dimension is 1, and the 1st dimension is output through a Sigmoid activation function. Classification probability of each sampling point :

[0026] ;

[0027] in, This is the weight matrix of the fully connected layer, used to map the 128-dimensional context vector to the scalar output; For bias terms of fully connected layers; For the Sigmoid activation function:

[0028] ;

[0029] like > ,but This corresponds to the presence of an internal isolated wave perturbation in the current sequence. The threshold value is used.

[0030] In the above scheme, step one specifically includes:

[0031] Step 1.1: Extract the average value of the echo waveform power between the 25th and 45th echo gates as the echo waveform power;

[0032] Step 1.2: Extract the backscattering coefficients of the Ku-band and C-band from the original dataset. For each time point, calculate the difference mean square slope between the Ku-band and C-band. :

[0033] ;

[0034] in, This is the C-band NRCS calibration value. and These are constants, taken as 0.427 and 0.617 respectively; and These represent the backscattering coefficients for the 20Hz Ku band and the 20Hz C band, respectively.

[0035] Will As a new one-dimensional variable, it is added to the original feature set, and together with sea surface height anomaly, significant wave height, and echo waveform power, it constitutes the final multidimensional feature component.

[0036] Step 1.3 involves quality control and invalid point removal of the extracted multidimensional feature vector samples, specifically including the following three sub-processes:

[0037] (1) Based on the surface classification markers in the synthetic aperture radar altimeter products of Sentinel-3 satellite, open sea samples with a value of 0 were screened out, and land, water vapor mixing zones and inland water bodies were removed;

[0038] (2) Use the rainfall markers provided by the synthetic aperture radar altimeter of the Sentinel-3 satellite to determine the sample area affected by strong convection or rainfall. Use the linear interpolation method to interpolate the rainfall marker data from 1Hz resolution to 20Hz, and delete the observation points that are non-zero in the interpolation results.

[0039] (3) Perform shoreline masking, calculate the distance from the shore for each sample point, and remove data samples with a distance from the shore of less than 5 km.

[0040] In the above scheme, step two involves performing frequency domain bandpass filtering on the preprocessed multidimensional feature components as follows:

[0041] (1) Perform fast Fourier transform on the preprocessed multidimensional feature components to map the signal from the time domain to the frequency domain;

[0042] (2) Filter the obtained frequency domain data to retain only the frequency components within the 2-20 km scale;

[0043] (3) Perform an 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 multidimensional feature components.

[0044] In the above scheme, the specific method for performing adaptive sliding standard deviation calculation on the multidimensional feature components after frequency domain bandpass filtering in step two is as follows:

[0045] (1) Set multiple sliding window sizes W to 3 to 21 points to construct a multi-scale sliding window group;

[0046] (2) For the characteristic component sequence after frequency domain bandpass filtering, calculate the local mean in each sliding window W:

[0047] ;

[0048] Where W is the size of the sliding window. The sequence of characteristic components includes significant wave height, sea surface height anomaly, differential mean square slope of backscattering coefficient, and echo waveform power. This refers to the i-th data point within the window. This represents the local mean of the feature component sequence within window W.

[0049] Through local mean Calculate the local standard deviation of the characteristic component sequence:

[0050] ;

[0051] in, This represents the local standard deviation under the sliding window W;

[0052] (3) For each obtained local standard deviation Furthermore, its global variance value is calculated as a metric to assess the significance of the perturbation structure at this window scale;

[0053] (4) Select the standard deviation data with the largest global variance from the calculation results of all sliding windows W. The final data with the largest global variance and standard deviation of the four characteristic components, namely sea surface height anomaly, difference mean square slope of backscattering coefficients of Ku band and C band, significant wave height, and echo waveform power, are as follows: , , , ;

[0054] (5) The obtained , , , Perform min-max normalization on each value and map them to... The intervals ultimately yield four feature components. , , , Construct a unified input tensor .

[0055] In the above scheme, during 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 acquired. The presence of internal solitary waves in the images is confirmed by manual visual interpretation, internal solitary wave labels are generated, and the corresponding positions and data in the multidimensional feature components are located by combining the latitude and longitude information of the internal solitary waves in the images. The multidimensional feature components at the corresponding positions are input into the model, and the internal solitary wave labels are used as outputs to train the model.

[0056] The internal solitary wave identification method based on synthetic aperture radar altimeter data provided by the above technical solution has the following beneficial effects:

[0057] 1. Based on the multi-source physical component data provided by the Sentinel-3 Synthetic Aperture Radar Altimeter (SRAL), this invention utilizes four key feature components—sea surface height anomaly, normalized radar backscattering coefficient, significant wave height, and echo waveform power—combined with targeted data preprocessing and disturbance feature highlighting strategies, to characterize the disturbance features of internal solitary waves on the sea surface in multiple dimensions, thereby improving the detection sensitivity and generalization performance of the model.

[0058] 2. This invention employs a feature extraction method combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM). It simultaneously acquires the local spatial patterns and long-term temporal dependencies of internal solitary wave perturbations. A residual temporal fusion mechanism is introduced to alleviate gradient decay and information loss problems during deep network training, achieving efficient fusion of multi-layer features across time scales, thereby enhancing the model's adaptability to different waveform morphologies. Furthermore, a frequency-temporal joint attention mechanism is applied to adaptively allocate feature weights at typical wavelength scales, highlighting the characteristics of internal solitary wave perturbations and improving detection accuracy and stability.

[0059] 3. This invention achieves automated detection of internal solitary waves in Sentinel-3 satellite data through targeted data processing methods and a dedicated deep learning model. It can extract and analyze the modulation characteristics, providing reliable data support for marine engineering construction and other projects. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0061] Figure 1 This is a schematic flowchart of an internal solitary wave identification method based on synthetic aperture radar altimeter data disclosed in an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram of the structure of the deep learning model ISW-DetSRALNet disclosed in this invention;

[0063] Figure 3 A schematic diagram showing the location of internal solitary waves in the acquired OICL image;

[0064] Figure 4 A schematic diagram of the feature components input to the model for training;

[0065] Figure 5 This is a schematic diagram of the detected internal solitary wave locations output by the model. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0067] This invention provides a method for identifying internal solitary waves based on synthetic aperture radar altimeter data, such as... Figure 1 As shown, it includes the following steps:

[0068] Step 1: Collect multidimensional feature components from the Sentinel-3 satellite synthetic aperture radar altimeter (SRAL), including sea surface height anomaly (SSHA), backscattering coefficient (including Ku-band and C-band), significant wave height (SWH), and echo waveform power (WAFM), and preprocess the multidimensional feature components.

[0069] Specifically include:

[0070] Step 1.1: In order to capture the effect 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 echo waveform power.

[0071] Step 1.2: To address the weak modulation effect of internal solitary wave disturbances on sea surface roughness, the differential mean square slope of the backscattering coefficients (NRCS) of the Ku-band and C-band is further introduced as an enhancement feature to improve sensitivity to the disturbance signal. The specific processing flow is as follows:

[0072] Extracting raw NRCS data: Extract the backscattering coefficients of the Ku-band and C-band from the SRAL dataset (fields sig0_ku and sig0_c, respectively). Both variables are 20Hz resolution and are in dB.

[0073] 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, expressed as follows:

[0074] ;

[0075] in, This is the C-band NRCS calibration value. and These are constants, taken as 0.427 and 0.617 respectively; and These represent the backscattering coefficients for the 20Hz Ku band and the 20Hz C band, respectively.

[0076] The difference operation yields a one-dimensional sequence. Its length is consistent with the original NRCS variable, and this feature is used to represent the frequency response difference of internal solitary waves in terms of sea surface roughness modulation. As a new one-dimensional variable, it is added to the original feature set, and together with sea surface height anomaly (SSHA), significant wave height (SWH), and echo waveform power (WAFM), it constitutes the final multidimensional feature component.

[0077] Step 1.3 involves quality control and invalid point removal of the extracted multidimensional feature vector samples, specifically including the following three sub-processes:

[0078] (1) Based on the surface classification marker (surf_class_20_ku) in the synthetic aperture radar altimeter product of Sentinel-3 satellite, open sea samples with a value of 0 were screened out, and land, water vapor mixing zone and inland water body were removed;

[0079] (2) Use the rainfall flag (rain_flag_01_ku) provided by the synthetic aperture radar altimeter of Sentinel-3 satellite to determine the sample area affected by strong convection or rainfall. Use linear interpolation to interpolate the rainfall flag data from 1Hz resolution to 20Hz and delete the observation points that are non-zero in the interpolation results.

[0080] (3) Perform shoreline masking, calculate the distance from the shore for each sample point, and remove data samples with a distance from the shore of less than 5 km.

[0081] Step 2: After preprocessing, the multidimensional feature components are sequentially subjected to frequency domain bandpass filtering and adaptive sliding standard deviation calculation, and then normalized to construct an input tensor consisting of four feature components.

[0082] The preprocessed multidimensional feature components are subjected to frequency domain bandpass filtering to enhance the disturbance signal within the internal solitary wave scale range. The specific method is as follows:

[0083] (1) Perform fast Fourier transform on the preprocessed multidimensional feature components to map the signal from the time domain to the frequency domain;

[0084] (2) Filter the obtained frequency domain data to retain only the frequency components within the 2-20 km scale;

[0085] (3) Perform an 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 multidimensional feature components.

[0086] The specific method for adaptively calculating the moving standard deviation of the multidimensional feature components after frequency domain bandpass filtering is as follows:

[0087] (1) Set multiple sliding window sizes W to 3 to 21 points to construct a multi-scale sliding window group;

[0088] (2) For the characteristic component sequence after frequency domain bandpass filtering, calculate the local mean in each sliding window W:

[0089] ;

[0090] Where W is the size of the sliding window. The sequence of characteristic components includes significant wave height, sea surface height anomaly, differential mean square slope of backscattering coefficient, and echo waveform power. This refers to the i-th data point within the window. This represents the local mean of the feature component sequence within window W.

[0091] Through local mean Calculate the local standard deviation of the characteristic component sequence:

[0092] ;

[0093] in, This represents the local standard deviation under the sliding window W;

[0094] (3) For each obtained local standard deviation Furthermore, its global variance value is calculated as a metric to assess the significance of the perturbation structure at this window scale;

[0095] (4) Select the standard deviation data with the largest global variance from the calculation results of all sliding windows W. The final data with the largest global variance and standard deviation of the four characteristic components, namely sea surface height anomaly, difference mean square slope of backscattering coefficients of Ku band and C band, significant wave height, and echo waveform power, are as follows: , , , ;

[0096] (5) The obtained , , , Perform min-max normalization on each value and map them to... The intervals ultimately yield four feature components. , , , Construct a unified input tensor .

[0097] Step 3: Construct the deep learning model ISW-DetSRALNet to automatically identify and extract internal solitary wave signals, and train the model.

[0098] I. Model Building:

[0099] The model integrates convolutional neural networks and bidirectional long short-term memory networks, and introduces residual-temporal fusion (RTF) and frequency-temporal joint attention mechanism (FTJA) to improve the ability to capture sparse and weak signals of internal isolated waves without increasing the burden of additional hyperparameter tuning. Specifically, the model includes an input module, a convolutional feature extraction module, a temporal modeling module, a residual-temporal fusion module, a frequency-temporal joint attention mechanism module, and a classification decision output module. The functions of each module are as follows:

[0100] (1) Input module

[0101] The model input consists of time-series samples composed of the four feature components processed in step two. ,in, This represents the length of the time series used in the model, which defaults to 200. , representing the input feature components, i.e. .

[0102] (2) Convolutional Feature Extraction Module

[0103] The convolutional feature extraction module uses a one-dimensional convolutional neural network to extract local perturbation features. It contains four layers with a kernel size of 3 and output channels of 16, 32, 64 and 128 respectively, all using the ReLU activation function.

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] The formula for the ReLU activation function is: .

[0109] The final output tensor of the convolutional feature extraction module is ,in, The length of the time series used in the model.

[0110] (3) Temporal modeling module

[0111] The temporal modeling module uses the bidirectional LSTM module to process the output of the convolutional feature extraction module. Two-way temporal modeling is performed. The number of hidden layer units in the temporal 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 regions of the sea surface modulated by internal solitary waves.

[0112] The timing modeling module incorporates two independent LSTM units, one forward and one backward:

[0113] Forward LSTM unit: Input sequence Output ; This represents the feature vector at time step L after processing by the convolutional feature extraction module. The hidden state vector is calculated by the feedforward LSTM unit at time step t based on the current and previous input features;

[0114] Backward LSTM unit: Input sequence Output ; The hidden state vector is calculated by the backward LSTM unit at time step t based on the current and previous input features;

[0115] The outputs from the two directions are concatenated to obtain: H=64 is the number of hidden layer units in each direction, and the total output dimension is 128; This represents the context feature vector at time step t, which integrates forward and backward temporal information.

[0116] The output tensor of the entire time series is: , This is the bidirectional contextual feature vector at the last time step 𝐿.

[0117] (4) Residual time series fusion module

[0118] To enhance the coupling between local perturbations and global temporal sequences, residual connection fusion is introduced. The residual temporal fusion module integrates the output of the convolutional feature extraction module. Output of the timing modeling module By adding the element-level residuals, we obtain The residual It can alleviate deep gradient attenuation and enhance local-global feature coupling, which is beneficial for highlighting internal solitary wave disturbance signals.

[0119] (5) Frequency-Time Joint Attention Mechanism Module

[0120] The frequency-time domain joint attention mechanism module includes a frequency domain branch, a time domain branch, and a joint attention computation module.

[0121] Frequency domain branch pairs obtained Perform a fast Fourier transform to obtain After taking the amplitude, Dimensionality reduced to 32, frequency domain attention components are obtained. ;in, For Fast Fourier Transform, This represents a complex amplitude. Represents convolution processing;

[0122] The time-domain branch will receive Perform a linear mapping to obtain the time-domain query vector. ,in, For trainable weight matrix, It is a trainable bias vector;

[0123] The joint attention computation module will use the time-domain query vector With frequency domain attention components After addition, the attention weight vector at each time step is obtained by Softmax normalization. And output the weighted context vector. ;in, Representing the Attention weights for each sampling point Indicates the first A 128-dimensional feature vector at each time step.

[0124] The FTJA constructed in this model enables attention to simultaneously perceive energy and waveform abrupt changes in the 2-20 km frequency band, significantly improving the sensitivity of the isolated wave response in sparse areas.

[0125] (6) Classification Decision Output Module

[0126] The classification decision output module outputs a weighted context vector. The input is fed into a fully connected layer for dimensionality reduction, the output dimension is 1, and the 1st dimension is output through a Sigmoid activation function. Classification probability of each sampling point :

[0127] ;

[0128] in, This is the weight matrix of the fully connected layer, used to map the 128-dimensional context vector to the scalar output; For bias terms of fully connected layers; For the Sigmoid activation function:

[0129] ;

[0130] like > ,but This corresponds to the presence of an internal isolated wave perturbation in the current sequence. The threshold value is used.

[0131] II. Model Training:

[0132] 1. Generate inner solitary wave labels for model training:

[0133] (1) Acquisition of synchronized OLCI images. Acquire OLCI Level-1B images within ±30 min of the SRAL observation time window, such as... Figure 3 As shown.

[0134] (2) Location of internal solitary waves. After radiometric and atmospheric corrections are applied to the OLCI imagery, the presence of internal solitary waves is confirmed by visual inspection. For images where internal solitary waves are identified, the latitude and longitude range of the internal solitary waves is recorded. For example... Figure 3 The red line shows the Sentinel-3 SRAL probe trajectory, which detected an internal solitary wave near 10.3°N, as shown by the blue line in the figure.

[0135] (3) Trajectory matching and sample extraction

[0136] Using the latitude and longitude range mentioned above, the four acquired feature components are cropped according to their latitude and longitude ranges. Data points within the latitude and longitude range are considered as "inner solitary wave samples"; the rest are "background samples".

[0137] (4) Label assignment

[0138] All “inner isolated wave samples” are assigned a positive label = 1, and “background samples” are assigned a negative label = 0, thus obtaining a supervised training dataset that corresponds one-to-one with the multidimensional feature parameters of SRAL.

[0139] 2. Training configuration:

[0140] Input sequence length: 200; Feature dimension: 4; Number of bidirectional LSTM hidden units: 64, single-layer structure, bidirectional mode; Attention mechanism intermediate layer dimension: 32; Number of output classes: 2; Optimizer: Adam, initial learning rate 0.001, L2 regularization term weight decay coefficient is 1e-5; Batch size: 32, maximum number of training epochs: 1000; Validation set ratio: 25%; Supports CUDA accelerated computation.

[0141] To suppress overfitting and improve training stability, the following strategies are introduced: Adaptive learning rate scheduling mechanism: When the validation set F1-score does not improve significantly within 10 consecutive rounds, the learning rate is automatically reduced to 1 / 10 of the original value; Early stopping mechanism: When the validation set F1-score does not improve within 25 consecutive rounds, training is terminated early.

[0142] 3. Model Training and Performance Evaluation

[0143] When training the model, Figure 4 The multidimensional feature components at the corresponding positions shown are input into the model, and the internal solitary wave labels are used as outputs to train the model.

[0144] During model training, metrics such as loss, accuracy, precision, recall, and F1 score are monitored in real time during the training and validation phases.

[0145] The model output results are classified according to the true labels as follows: True positive (TP): The internal isolated wave is present and correctly identified; False positive (FP): The internal isolated wave is not present but is incorrectly identified as present; True negative (TN): The internal isolated wave is not present but is correctly identified; False negative (FN): The internal isolated wave is present but is mistakenly identified as not present.

[0146] The following performance metrics were used to evaluate the model's detection capability: Precision: the proportion of true positive samples among those predicted as positive; Recall: the proportion of true positive samples correctly identified; F1 score: the harmonic mean of Precision and Recall, reflecting the overall detection performance of the model. Table 1 lists and compares the results of input training with different combinations of feature parameters. The results show that the present invention, by inputting four parameter vectors into the model together, achieves the highest prediction accuracy.

[0147] Table 1. Training results with different parameter combinations

[0148]

[0149] Step four: Input the input tensor into the successfully trained model, and the model will automatically identify the internal solitary wave information.

[0150] The internal solitary wave information identified in the embodiments of the present invention is as follows: Figure 5 As shown, from Figure 5 As can be seen from the data, the location of the internal solitary wave identified by the model (indicated by the black cross) is consistent with... Figure 3 The consistent positions of the internal solitary waves shown demonstrate the accuracy of the model.

[0151] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those 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 invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded 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, Includes the following steps: Step 1: Collect multidimensional feature components from the Sentinel-3 satellite synthetic aperture radar altimeter observations, including sea surface height anomaly, backscattering coefficient, significant wave height, and echo waveform power, and preprocess the multidimensional feature components. Step 2: After processing the preprocessed multidimensional feature components, frequency domain bandpass filtering and adaptive sliding standard deviation calculation are performed sequentially, and after normalization, an input tensor consisting of four feature components is constructed. Step 3: Construct the deep learning model ISW-DetSRALNet and train the model; the model includes an input module, a convolutional feature extraction module, a temporal modeling module, a residual temporal fusion module, a frequency domain-temporal joint attention mechanism module, and a classification decision output module; The residual temporal fusion module converts the output of the convolutional feature extraction module. Output of the timing modeling module By adding the element-level residuals, we obtain ; The frequency-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 pairs obtained Perform a fast Fourier transform to obtain After taking the amplitude, Dimensionality reduced to 32, frequency domain attention components are obtained. ;in, For Fast Fourier Transform, This represents a complex amplitude. Represents convolution processing; The time-domain branch will obtain Perform a linear mapping to obtain the time-domain query vector. ,in, For trainable weight matrix, It is a trainable bias vector; The joint attention calculation module will use the time-domain query vector With frequency domain attention components After addition, the attention weight vector at each time step is obtained by Softmax normalization. And output the weighted context vector. ;in, Representing the Attention weights for each sampling point Indicates the first A 128-dimensional feature vector at each time step; The classification decision output module outputs a weighted context vector. The input is fed into a fully connected layer for dimensionality reduction, the output dimension is 1, and the 1st dimension is output through a Sigmoid activation function. Classification probability of each sampling point : ； in, This is the weight matrix of the fully connected layer, used to map the 128-dimensional context vector to the scalar output; For bias terms of fully connected layers; For the Sigmoid activation function: ； like > ,but This corresponds to the presence of an internal isolated wave perturbation in the current sequence. For the threshold; Step four: Input the input tensor into the successfully trained model, and the model will automatically identify the internal solitary wave information.

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

3. The method for identifying internal solitary waves based on synthetic aperture radar altimeter data according to claim 2, characterized in that, The temporal modeling module is based on the output of the convolutional feature extraction module from the bidirectional LSTM module. Bidirectional temporal modeling is performed, with the temporal modeling module having 64 hidden layer units and an output dimension of 128. The timing modeling module simultaneously incorporates two independent LSTM units: forward and backward. Forward LSTM unit: Input sequence Output ; This represents the feature vector at time step L after processing by the convolutional feature extraction module. The hidden state vector is calculated by the feedforward LSTM unit at time step t based on the current and previous input features; Backward LSTM unit: Input sequence Output ; The hidden state vector is calculated by the backward LSTM unit at time step t based on the current and previous input features; The outputs from the two directions are concatenated to obtain: H=64 is the number of hidden layer units in each direction, and the total output dimension is 128; This represents the context feature vector at time step t, which integrates forward and backward temporal information. The output tensor of the entire time series is: , This is the bidirectional contextual feature vector at the last time step 𝐿.

4. The method for identifying internal solitary waves based on synthetic aperture radar altimeter data according to claim 3, characterized in that, Step one specifically includes: Step 1.1: Extract 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 backscattering coefficients of the Ku-band and C-band from the original dataset. For each time point, calculate the difference mean square slope between the Ku-band and C-band. : ； in, This is the C-band NRCS calibration value. and These are constants, taken as 0.427 and 0.617 respectively; and These represent the backscattering coefficients for the 20Hz Ku band and the 20Hz C band, respectively. Will As a new one-dimensional variable, it is added to the original feature set, and together with sea surface height anomaly, significant wave height, and echo waveform power, it constitutes the final multidimensional feature component. Step 1.3 involves quality control and invalid point removal of the extracted multidimensional feature vector samples, specifically including the following three sub-processes: (1) Based on the surface classification markers in the synthetic aperture radar altimeter products of Sentinel-3 satellite, open sea samples with a value of 0 were screened out, and land, water vapor mixing zones and inland water bodies were removed; (2) Use the rainfall markers provided by the synthetic aperture radar altimeter of the Sentinel-3 satellite to determine the sample area affected by strong convection or rainfall. Use the linear interpolation method to interpolate the rainfall marker data from 1Hz resolution to 20Hz, and delete the observation points that are non-zero in the interpolation results. (3) Perform shoreline masking, calculate the distance from the shore for each sample point, and remove data samples with a distance from the shore of less than 5 km.

5. The method for identifying internal solitary waves based on synthetic aperture radar altimeter data according to claim 1, characterized in that, In step two, the method for performing frequency domain bandpass filtering on the preprocessed multidimensional 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 an 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 multidimensional feature components.

6. The method for identifying internal solitary waves based on synthetic aperture radar altimeter data according to claim 1, characterized in that, In step two, the specific method for performing adaptive sliding standard deviation calculation on the multidimensional feature components after frequency domain bandpass filtering is as follows: (1) Set 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 in each sliding window W: ； Where W is the size of the sliding window. The sequence of characteristic components includes significant wave height, sea surface height anomaly, differential mean square slope of backscattering coefficient, and echo waveform power. This refers to the i-th data point within the window. This represents the local mean of the feature component sequence within window W. Through local mean Calculate the local standard deviation of the characteristic component sequence: ； in, This represents the local standard deviation under the sliding window W; (3) For each obtained local standard deviation Furthermore, its global variance value is calculated as a metric to assess the significance of the perturbation structure at the W window scale; (4) Select the standard deviation data with the largest global variance from the calculation results of all sliding windows W. The final data with the largest global variance and standard deviation of the four characteristic components, namely sea surface height anomaly, difference mean square slope of backscattering coefficients of Ku band and C band, significant wave height, and echo waveform power, are as follows: , , , ; (5) The obtained , , , Perform min-max normalization on each value and map them to... The intervals ultimately yielded four feature components. , , , Construct a unified input tensor .

7. The method for identifying internal solitary waves 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 Sentinel-3 satellite synthetic aperture radar altimeter orbit are acquired. The presence of internal solitary waves in the images is confirmed by manual visual interpretation, internal solitary wave labels are generated, and the corresponding positions and data in the multidimensional feature components are located by combining the latitude and longitude information of the internal solitary waves in the images. The multidimensional feature components at the corresponding positions are input into the model, and the internal solitary wave labels are used as outputs to train the model.

Images