Deep learning based method for extracting inner tidal signals from swot satellite observations

Abstract

The application provides a SWOT satellite observation internal tide signal extraction method based on deep learning, and belongs to the field of ocean satellite remote sensing and artificial intelligence information processing. Based on the temperature, salinity and total sea surface height anomaly data of high-resolution ocean numerical simulation, the training sample of total sea surface height anomaly-internal tide induced sea surface height anomaly corresponding to space and time is constructed; the related correction items in the SWOT three-level product are recovered, and the actual observation input field consistent with the training input physical definition is constructed; the successive observation orbit geometric boundary is constructed according to the effective longitude and latitude of the actual SWOT observation, the tilt wide data is converted into a regular two-dimensional matrix through cross-track normalization; the missing position is temporarily filled, the internal tide induced sea surface height anomaly is extracted by using the trained deep learning model, then the original missing mask is recovered and inversely mapped to the original geographic coordinates, so that the internal tide induced sea surface height anomaly field corresponding to a single SWOT wide observation is obtained.

Description

Technical Field

[0001] This application relates to the fields of marine satellite remote sensing and artificial intelligence information processing, specifically to a deep learning-based SWOT satellite observation intidal signal extraction method. Background Technology

[0002] Internal tides (ITs) are baroclinic tidal movements generated by the interaction of barotropic currents and complex seafloor topography in a stratified ocean. They can form sea surface height anomalies (SSHAs) on the order of centimeters that can be observed by satellite altimeters.

[0003] Subsurface altimeters primarily provide one-dimensional along-orbit observations. Since satellite revisit periods are typically longer than the major diurnal and semi-diurnal tide cycles, existing methods for extracting internal tides usually rely on multi-year repeated orbital observations, using methods such as harmonic analysis and plane wave fitting to obtain long-term phase-locked internal tide components. While these methods can extract long-term stable, phase-locked internal tide components, they require long time series and are difficult to directly process from single satellite observations.

[0004] Surface Water and Ocean Topography (SWOT) satellite mission can Two-dimensional wide-swath sea surface height observations were obtained using a Ka-band radar interferometer, providing a data foundation for analyzing the two-dimensional internal tide spatial structure in a single observation. However, a single SWOT sea surface height anomaly simultaneously includes background circulation, mesoscale eddies, tides, and other multi-scale dynamic signals, while the internal tide-induced signal is relatively weak. Furthermore, SWOT has only been providing sea surface height observations since July 2023, resulting in a relatively short data accumulation period, and the repeat orbit observation cycle is approximately 21 days, making it difficult to directly separate the corresponding internal tide signal from a single observation using existing harmonic analysis. Therefore, existing spatial filtering techniques also struggle to extract the internal tide-induced sea surface height anomaly while preserving other dynamic signals with similar spatial scales to the internal tide.

[0005] Previous studies have disclosed methods for processing internal tide signals from wide-swath satellite altimeters using convolutional neural networks. For example, the paper "Filtering Internal Tides From Wide-Swath Altimeter Data Using Convolutional Neural Networks" uses convolutional neural networks to process simulated wide-swath altimeter sea surface height data to estimate and filter out the influence of internal tides. This method demonstrates that convolutional neural networks can learn spatial characteristics related to internal tides in the wide-swath sea surface height field, but its technical objectives and processing procedures are completely different from those of this application. Furthermore, deep learning has also been applied to the identification and separation of other marine dynamic signals. For example, the previously published paper "A Deep Learning Approach to Extract Balanced Motions From Sea Surface Height Snapshot," authored by Zhanwen Gao et al., published in the journal Geophysical Research Letters, 2024, Vol. 51, No. 7, e2023GL106623 (11 pages), utilizes deep learning to extract balanced motions from a single sea surface height anomaly snapshot. It discloses a methodology using high-resolution numerical simulation (LLC4320 / MITgcm) to construct labels, single-frame SSH→image-to-image regression, and a U-Net / CNN encoder-decoder model for SWOT wide-swath observations. Existing literature, including the aforementioned paper, demonstrates that neural networks can learn spatial structure features in two-dimensional ocean fields, but a complete internal tide extraction process for practical SWOT tilt wide-swath observations has not yet been established.

[0006] In summary, the main technical shortcomings of existing technologies include: First, satellite internal tide extraction methods typically rely on repetitive orbital observations over many years for harmonic analysis. However, SWOT data accumulation is currently limited, and the revisit cycle is long, making it difficult to directly process single SWOT satellite observations to obtain the corresponding internal tide signals. Second, single SWOT sea surface height anomalies contain background circulation, mesoscale eddies, tides, and other multi-scale dynamic signals. Internal tide-induced signals are relatively weak, and harmonic analysis focuses on extracting long-term phase-locked internal tide components, making it difficult to effectively distinguish different ocean dynamic signals with similar spatial scales. Third, SWOT observations exhibit a tilted double-width strip distribution and include nadir gaps, local missing measurements, and invalid pixels, which is inconsistent with the regular rectangular grids typically used in convolutional neural networks. Therefore, it is necessary to establish data preprocessing methods specifically for this observation geometry.

[0007] In view of the above, this application is hereby submitted. Summary of the Invention

[0008] This application proposes a deep learning-based method for extracting intidal signals from SWOT satellite observations. The method overcomes the problems of existing technologies by constructing intidal training samples based on high-resolution ocean numerical simulation data, extracting intidal-induced sea surface height anomalies using a deep learning model, restoring the extracted results to the original SWOT geographic coordinates using a regular grid, and then applying the trained model to actual SWOT observations. The aim is to achieve a solution capable of extracting instantaneous intidal-induced sea surface height anomalies (IT-induced SSHA) from a single two-dimensional SWOT wide-swath sea surface height anomaly observation.

[0009] To achieve the aforementioned objectives, the deep learning-based SWOT satellite observation inland tide signal extraction method constructs target data of inland tide-induced sea surface height anomalies corresponding to the spatiotemporal total sea surface height anomalies from three-dimensional temperature and salinity data of high-resolution ocean numerical simulations, forming an input-target data pair for deep learning model training. It reconstructs an actual sea surface height anomaly input field with a physical definition consistent with the numerical simulation training input based on relevant correction variables in the SWOT product. For each actual SWOT wide-swath observation, it constructs orbital geometric boundaries based on effective latitude and longitude coordinates, without using the absence of sea surface height anomalies at the corresponding locations as a boundary determination condition. It calculates the normalized coordinates of the transverse rails based on the local left and right boundaries, converting the tilted wide-swath observations into a regular two-dimensional matrix. The trained deep learning model extracts inland tide-induced sea surface height anomalies, temporarily filling in missing locations required for model calculation, and recovering the missing areas based on the original effective observation mask after model output. Finally, it restores the prediction results in the regular two-dimensional matrix to the original SWOT geographic coordinates.

[0010] The method includes the following implementation steps: Phase (1): Data Construction and Preprocessing; Step 1.1: Construct training samples for total sea surface height anomaly and intidal-induced sea surface height anomaly; Training samples were constructed using high-resolution ocean numerical simulation data that can continuously output temperature, salinity, and total sea surface height over time and analyze tidal dynamic processes. Step 1.2: Establish a SWOT wide-span observation input field consistent with the physical definition of the numerical simulation training input; Read the filtered sea surface height anomalies, mean dynamic topography, barotropic tides, internal tides, and dynamic atmospheric correction variables from the SWOT Level 3 Ka-band radar interferometer low-rate sea surface height expert product. Step 1.3: Construct the SWOT orbital geometry boundary based on geographic coordinates; Effective latitude processing ranges were set for different SWOT orbits, and total sea surface height anomalies and intidal-induced sea surface height anomalies for corresponding wide areas were extracted from high-resolution ocean numerical simulation data based on the actual geographical location of each orbit. Region extraction, boundary construction, regularization and model training were performed separately for each orbit, and the geometric boundaries of other orbits were not shared. Step 1.4: Convert the tilted wide-span observations into a regular rectangular grid; Step 1.5: Calculate the normalization parameters and process invalid pixels; Based on all valid pixels in each orbit training set, calculate the mean and standard deviation of the total sea surface height anomaly input data and the target data of the sea surface height anomaly induced by the internal tide, excluding null values ​​during the calculation process; Phase (II): Model building, training, and practical SWOT observation application; Step 2.1: Establish a deep learning-based internal flow extraction model; The deep learning internal moisture extraction model described above employs U-Net, residual U-Net, attention U-Net, encoder-decoder convolutional network, or a network combining convolution and attention mechanisms. The encoder-decoder of the model has a 4-level, 5-level, or 6-level structure. Step 2.2: Train the model; The model is trained using L1 loss, MSE loss, or a gradient boosting composite loss function. Step 2.3: Training the model and parameters; The AdamW optimizer is used, with learning rates of constant learning rate, step decay learning rate, or cosine annealing learning rate. Step 2.4: Extraction of internal tides from actual SWOT satellite observations; For each SWOT track to be processed, perform the following operations: Read the SWOT Level 3 product for the corresponding orbit; add back the mean dynamic topography, barotropic tide, internal tide, and dynamic atmospheric correction terms to construct the sea surface height anomaly input field; convert the orbital geometric boundary observed for the corresponding orbit into a regular rectangular grid with a preset fixed number of transverse rail grid points; save the original valid observation mask; fill the missing values ​​in the regular grid using the nearest valid pixel extrapolation method; normalize using the mean and standard deviation of the corresponding orbital training set; input the normalization result into the deep learning internal tide extraction model trained for the corresponding orbit; inverse normalize the model output using the mean and standard deviation of the target internal tide training set; restore the original invalid observation positions to missing values; output the instantaneous internal tide-induced sea surface height anomaly results under the regular rectangular grid; Then, based on the saved original geographic information, for each original valid observation point, interpolation is performed in the prediction results of the regular rectangular grid according to the bin of latitude and the normalized coordinates of the horizontal rail to obtain the predicted value of the sea surface height anomaly induced by the internal tide corresponding to the observation point. Finally, the predicted values ​​of sea surface height anomalies induced by internal tides are mapped one-to-one with the original longitude and latitude to form the output results under the original SWOT tilted wide-span geographic coordinates; no predicted values ​​are output for the original missing measurement points, thus restoring the output results to the original SWOT tilted wide-span geographic coordinates.

[0011] Step 1.1, using the LLC4320 total sea surface height anomaly as model input, calculates seawater density according to the TEOS-10 seawater thermodynamic standard using temperature and salinity data, and calculates the density perturbation of instantaneous density relative to time-averaged density:

[0012] in, Represents instantaneous density. Represents the average density over time; Calculation of baroclinic pressure anomaly based on density perturbation under static approximation:

[0013]

[0014] in, Represents gravitational acceleration; , These represent the vertical coordinates of the sea surface and the seabed, respectively. This indicates the local water depth. Represents the vertical integral coordinates; Through depth-independent surface pressure terms The vertical average value of the baroclinic pressure anomaly is eliminated; then the baroclinic pressure anomaly at sea surface is converted into a baroclinic sea surface height anomaly. :

[0015] in, It is a fixed reference density; Using Butterworth bandpass filters, respectively for Conduct diurnal tidal frequency band and semi-diurnal tidal zone The bandpass filter was applied, and the two parts were added together to obtain the following sea surface height anomaly induced by intidal currents. :

[0016] This results in a one-to-one correspondence in time and space between training data pairs of “total sea surface height anomaly – sea surface height anomaly induced by internal tides”; The samples are divided into training set, validation set and test set in chronological order; the training set is used for model parameter updates and normalization parameter calculation, the validation set is used for monitoring the training process and early termination, and the test set is used for comparing different model configurations and evaluating model performance.

[0017] Step 1.2 involves constructing the actual SWOT sea surface height anomaly input field corresponding to the physical definition of the numerical simulation training input according to the following formula. :

[0018] in, The sea level in the SWOT analysis is abnormal. Average dynamic terrain; This is a positive pressure tidal correction term; For internal moisture correction; This is a dynamic atmospheric correction item.

[0019] In step 1.3, for each SWOT wide-span observation to be processed, the observation is divided according to a preset latitude interval (1 / 48°). Only observation points with valid longitude and latitude are used to construct the orbital geometric boundary of the corresponding orbital observation. The validity of the sea surface height anomaly at the corresponding location is not used as the boundary determination condition. For each latitude bin, calculate the minimum and maximum longitude of all valid geographic coordinates within that bin, which will serve as the local left and right boundaries of that latitude location. For latitude bins whose boundaries cannot be directly obtained due to observation gaps, linear interpolation is performed using the boundary values ​​of adjacent valid bins to fill in the corresponding left and right boundaries.

[0020] In step 1.4, after completing the geometric boundary construction, valid sea surface height data that simultaneously meets the following conditions is extracted: both longitude and latitude are valid; the sea surface height anomaly input value is valid; and it is located within the preset latitude range of the corresponding orbit. For any valid data point, obtain the corresponding local left and right boundaries based on its latitude binning, and calculate the normalized coordinates of the horizontal track according to the following formula. :

[0021] in, Let be the longitude of the i-th observation point; and These are the left and right boundaries of the bin at the j-th latitude where the i-th observation point is located; Let i be the normalized position of the horizontal rail within the local wide range of observation points; When the local left and right boundaries are invalid or the right boundary is not greater than the left boundary, the corresponding mapping point is discarded. Using the center of the latitude bin as the coordinate along the track, the normalized coordinates of the horizontal track are uniformly mapped to 128 fixed horizontal track grid points. The number of grid cells along the track direction is determined by the number of latitude bins, and is padded to an even number. Subsequently, scattered linear interpolation is used to interpolate the effective observations onto a regular rectangular grid, resulting in a two-dimensional input matrix suitable for convolutional neural network processing. At the same time, the longitude, latitude, and sea surface height information of the original effective data points are preserved as the basis for subsequent mapping of the prediction results to the actual geographic coordinates.

[0022] In step 1.5, the input data is normalized according to the following formula:

[0023] The sea surface height anomaly induced by the target internal tide is normalized according to the following formula:

[0024] The mean and standard deviation are determined only by the training set, while the validation set, test set, and actual SWOT observations all use the same set of training set statistical parameters for the corresponding orbits. During the training phase, an effective ocean mask is established based on whether the target field of sea surface height anomalies induced by inland tides is null.

[0025] Null values ​​in the training input and target data are temporarily replaced with preset values ​​before model calculation, and invalid positions are not included in the loss function calculation.

[0026] Step 2.1 uses a six-level encoder-decoder U-Net model, with the regularized and normalized two-dimensional sea surface height anomaly input field as a single-channel input and the corresponding two-dimensional intidal-induced sea surface height anomaly as a single-channel output. The encoder first converts the number of input channels from 1 to 64 through a double convolution module, and then successively doubles and expands the number of feature channels through six downsampling modules. Each dual convolutional module consists of two consecutive 3×3 convolutional layers; after each convolutional layer, a batch normalization layer and a linear rectified unit activation layer are set in sequence; the convolution is padded with 1 pixel to keep the spatial size consistent before and after the convolution; each downsampling module first uses 2×2 max pooling, and then performs dual convolution operation. The decoder includes six upsampling modules, which restore the feature channels by halving them sequentially; Each upsampling module first uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to restore the spatial resolution, then concatenates it with the feature map of the corresponding scale of the encoder, and fuses the features through a dual convolution module; When there is a difference in spatial size between the upsampled features and the encoder features, the upsampled features are symmetrically padded to make their sizes consistent. Finally, the 64-channel features are converted into a single-channel intidal-induced sea surface height anomaly output through a 1×1 convolutional layer.

[0027] In step 2.2, the mask gradient enhancement composite loss function is used to calculate the data error term using the following formula. :

[0028] in, Extract predicted values ​​for the model; The truth value of the target; For effective marine cover; To prevent the use of tiny positive numbers with a denominator of zero, this embodiment uses 10. -8 ; Then, the spatial gradients of the prediction field and the target field are calculated using the horizontal and vertical Sobel operators, respectively; The operator used is:

[0029]

[0030] The gradient error term is:

[0031] The total loss function is:

[0032] Where α is the weighting coefficient of the gradient error term, with a value of 0.001; The loss function described above is calculated only on valid ocean pixels, and the gradient error term is used to constrain the reconstruction of tidal wave crests, troughs, and spatial gradient structures.

[0033] In step 2.3, the learning rate is cosine annealing, gradually decreasing from the initial learning rate; after each training round, the composite loss is calculated using the validation set, and the current model parameters are saved when the validation set loss is lower than the previous best value. Training is terminated early if the validation set loss exceeds the preset number of rounds without improvement; a fixed random seed is used during training to improve the reproducibility of the model training results.

[0034] In summary, this application has the following advantages and beneficial effects compared with the prior art: 1. This application does not rely on the time-averaged field as a training target. Instead, it utilizes three-dimensional temperature and salinity data from the LLC4320 high-resolution numerical simulation to calculate density according to the seawater state equation. Further calculations are then made of density perturbations, baroclinic pressure anomalies, and baroclinic sea surface height anomalies. Bandpass filtering is applied in both diurnal and semi-diurnal tide bands to directly construct an intidal-induced sea surface height anomaly that corresponds one-to-one with the total sea surface height anomaly in time and space, serving as a supervisory label. Therefore, this application establishes a direct mapping between "total sea surface height anomaly and intidal-induced sea surface height anomaly." This allows for the instantaneous extraction of intidal-induced sea surface height anomalies at a given moment using a single two-dimensional SWOT sea surface height anomaly observation. It possesses a clear physical definition of intidal currents, providing a technical approach for extracting the spatial structure of intidal currents based on single two-dimensional wide-swath satellite observations.

[0035] 2. Based on actual SWOT two-dimensional wide-span observation, this application establishes an integrated technical process including training label construction with clear physical definition of internal tide, reconstruction of physical quantities of actual products, construction of geometric boundaries of successive observation orbits, tilt wide-span regularization, missing measurement processing, direct prediction of internal tide, and restoration of original geographic coordinates.

[0036] 3. This application utilizes temperature, salinity, and total sea surface height anomalies from high-resolution ocean numerical simulation data to construct spatiotemporally corresponding physical training labels. It also restores relevant physical items that have been deducted from the actual SWOT product, ensuring consistent physical definitions between the actual observation input and the numerical simulation training input. This helps reduce model application bias caused by inconsistencies between the physical definitions of the numerical simulation training input and the actual satellite observation input. For each actual SWOT observation, the left and right geometric boundaries of the orbit are redefined only based on the effective latitude and longitude coordinates, without using the absence of sea surface height anomalies at the corresponding locations as a boundary determination condition. Subsequently, the normalized coordinates of the transverse rail are calculated based on the local left and right boundaries of each latitude bin, uniformly mapping the tilted wide-swath observations into a fixed 128-column regular two-dimensional matrix. For missing locations after regularization, temporary filling is performed using the nearest effective pixel extrapolation method before model inference. After inference, the missing areas are restored based on the saved original effective pixel mask. Finally, the model results are restored to the actual SWOT tilted wide-swath geographic coordinates based on the original latitude and longitude and the normalized transverse rail coordinates. Therefore, this application forms a complete closed loop from actual irregular wide-swath observations to regular neural network inputs, and then from regular prediction results to recover the actual satellite observation coordinates.

[0037] 4. This application transforms the tilted wide-swath observations into a regular two-dimensional matrix suitable for deep learning model processing by using horizontal track normalization coordinate transformation, while preserving the original geographic mapping relationship. This allows the model output to be restored to the actual SWOT latitude and longitude coordinates. It directly outputs the sea surface height anomaly induced by the internal tide itself to obtain the two-dimensional internal tide signal corresponding to a single satellite observation time. The output results can then be directly used to analyze the spatial structure, propagation characteristics, and instantaneous changes of the internal tide, without needing to further infer the internal tide component from the sea surface height field after the tide has receded.

[0038] 5. In practical applications, this application reads the low-rate sea surface height expert product from a SWOT Level 3 Ka-band radar interferometer and reinstates the mean dynamic topography, barotropic tides, internal tide correction, and dynamic atmospheric correction terms that were previously deducted during product processing. This constructs a total sea surface height anomaly field with a physically defined match to the numerical simulation training input. This processing reduces the input discrepancies caused by inconsistencies between the numerical simulation training data and the actual satellite product's physical definitions, enabling the model trained from the simulation data to be used for actual SWOT observations.

[0039] 6. This application employs a U-Net model with a multi-scale encoder-decoder and skip connection structure, enabling the model to learn the regionally differentiated characteristics of intidal generation and propagation within different orbital ranges. Simultaneously, an effective ocean mask is used during model training, and a composite loss is constructed using sea surface height error and spatial gradient error terms to simultaneously constrain the numerical amplitude of the predicted field and the crests, troughs, and spatial gradient structure of the intidal wave. Attached Figure Description

[0040] The present application will now be further described in conjunction with the following figures; Figure 1 This is a flowchart illustrating the deep learning-based SWOT satellite observation internal tide signal extraction method. Figure 2 A schematic diagram of the orbital geometry boundary construction, regular rectangular grid transformation, and inverse coordinate recovery for SWOT tilted double-width observations; in the figure, (a) is the original orbital observation map; (b) is the latitude binning and boundary determination map; (c) is the missing boundary interpolation map; (d) is the transverse track coordinate normalization map; (e) is the regular rectangular grid transformation map; (f) is the inverse coordinate recovery map of the prediction results; Figure 3 This is a diagram of a six-level U-Net network structure; solid arrows in the diagram represent main feature transfer, and dashed arrows represent same-scale skip connections from the encoder to the decoder; Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to specific examples. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0042] Example 1, such as Figures 1 to 3 As shown, the deep learning-based SWOT satellite observation inland tide signal extraction method establishes a closed-loop scheme based on the physical quantity definition, orbital geometry and missing measurement characteristics of actual SWOT tilted double-width observations. This scheme includes numerical simulation physical label construction, orbital geometry regularization, deep learning extraction and original geographic coordinate recovery. As a result, the model trained on a regular grid can process a single actual SWOT wide-width observation and output the inland tide-induced sea surface height anomaly field at the corresponding time.

[0043] Specifically, this method constructs target data of intidal-induced sea surface height anomalies corresponding to the spatiotemporal total sea surface height anomalies from three-dimensional temperature and salinity data of high-resolution ocean numerical simulation, forming an input-target data pair for deep learning model training; it reconstructs the actual sea surface height anomaly input field with a physical definition consistent with the numerical simulation training input based on relevant correction variables in the SWOT product; for each actual SWOT wide-swath observation, it constructs the orbital geometric boundary based on the effective latitude and longitude coordinates, without using whether the sea surface height anomaly at the corresponding location is missing as the boundary determination condition; it calculates the normalized coordinates of the transverse track based on the local left and right boundaries, converting the tilted wide-swath observation into a regular two-dimensional matrix; it uses the trained deep learning model to extract intidal-induced sea surface height anomalies, temporarily filling in the missing locations required for model calculation, and restores the missing areas based on the original effective observation mask after model output; finally, it restores the prediction results in the regular two-dimensional matrix to the original SWOT geographic coordinates.

[0044] This method comprises two phases: model building and practical SWOT observation application, and includes the following execution process: The model building phase includes: training label generation (including density perturbation, baroclinic pressure anomaly, baroclinic sea surface height anomaly, and tidal band filtering), extraction of simulated data samples for each orbit (including diurnal and semi-diurnal tidal bands), tilt width regularization, data normalization, effective ocean mask construction, and deep learning model training. The actual SWOT observation application stage includes: SWOT product physical quantity reconstruction, successive observation trajectory geometric boundary construction, regular grid conversion, temporary missing measurement filling, model inference, original missing measurement mask restoration and inverse coordinate mapping.

[0045] Phase (1): Data Construction and Preprocessing; Step 1.1: Construct training samples for total sea surface height anomaly and intidal-induced sea surface height anomaly; Training samples can be constructed using LLC4320 or other high-resolution ocean numerical simulation data that can continuously output temperature, salinity, and sea surface height and analyze tidal dynamic processes; this embodiment uses LLC4320, which has a horizontal grid spacing of approximately 1 / 48° and hourly output. This embodiment uses 200 consecutive days of anomaly data on temperature, salinity, and total sea surface height in a certain sea area; Taking the LLC4320 total sea surface height anomaly as the model input, the seawater density can be calculated using temperature and salinity data according to the TEOS-10 seawater thermodynamic standard, and the density perturbation of the instantaneous density relative to the time-averaged density can be calculated:

[0046] in, Represents instantaneous density. Represents the average density over time; Calculation of baroclinic pressure anomaly based on density perturbation under static approximation:

[0047]

[0048] in, Represents gravitational acceleration; , These represent the vertical coordinates of the sea surface and the seabed, respectively. This indicates the local water depth; Represents the vertical integral coordinates; Through depth-independent surface pressure terms The vertical average value of the baroclinic pressure anomaly is eliminated; then the baroclinic pressure anomaly at sea surface is converted into a baroclinic sea surface height anomaly. :

[0049] in, It is a fixed reference density; Using Butterworth bandpass filters, respectively for Conduct diurnal tidal frequency band and semi-diurnal tidal zone The bandpass filter was applied, and the two parts were added together to obtain the following sea surface height anomaly induced by intidal currents. :

[0050] This results in a one-to-one correspondence in time and space between training data pairs of “total sea surface height anomaly – sea surface height anomaly induced by internal tides”; To further mitigate the boundary effects caused by the initial and final states of filtering, the first 48 hours and the last 48 hours of the 200 consecutive days of data can be discarded, retaining 4704 hours of data. The 4704 samples are divided into training, validation, and test sets in chronological order, with a ratio of 80%:10%:10%, containing 3764, 470, and 470 samples respectively. The training set is used for model parameter updates and normalized parameter calculation, the validation set is used for monitoring the training process and early termination, and the test set is used for comparing different model configurations and evaluating model performance. Step 1.2: Establish a SWOT wide-span observation input field consistent with the physical definition of the numerical simulation training input; Read the filtered sea surface height anomalies, mean dynamic topography, barotropic tides, internal tides, and dynamic atmospheric correction variables from the SWOT Level 3 Ka-band radar interferometer low-rate sea surface height expert product. Construct the total sea surface height anomaly input field corresponding to the numerical simulation training input according to the following formula. :

[0051] in, The sea level in the SWOT analysis is abnormal. Average dynamic terrain; This is a positive pressure tidal correction term; For internal moisture correction; For dynamic atmospheric correction items; Through the above processing, the relevant physical terms that have been deducted in the SWOT product processing are reintroduced, so that the actual SWOT input field and the numerical simulation training input field have matching physical definitions; the internal tide correction term is the correction amount deducted in the product processing based on the external empirical tide product (HRET), and is not the observed true value; Step 1.3: Construct the SWOT orbital geometry boundary based on geographic coordinates; Effective latitude processing ranges were set for different SWOT orbits, and total sea surface height anomalies and intidal-induced sea surface height anomalies were extracted from the high-resolution ocean numerical simulation data (LLC4320) based on the actual geographical location of each orbit. Region extraction, boundary construction, regularization and model training were performed separately for each orbit, and the geometric boundaries of other orbits were not shared. Since the SWOT observation strips are skewed relative to the latitude and longitude grid, this application divides each wide area into 1 / 48° latitude intervals, and uses only observation points with valid longitude and latitude to construct the orbital geometric boundary, without considering whether the sea surface height data at the corresponding location is missing; for each latitude bin, the minimum and maximum longitudes of all valid geographic coordinates in the bin are calculated as the local left and local right boundaries of that latitude location; For latitudinal bins whose boundaries cannot be directly obtained due to observation gaps, linear interpolation is performed using the boundary values ​​of adjacent valid bins to fill in the corresponding left and right boundaries. The above boundary calculation only uses the latitude and longitude information of the observation points and does not use the validity of the sea surface height anomaly at the corresponding location as a boundary determination condition, thus enabling the construction of the orbital geometric boundary and the subsequent selection of valid pixels for the sea surface height anomaly to be performed separately. For actual SWOT observations, the above orbital left and right geometric boundaries are recalculated for each observation to be processed based on the valid latitude and longitude coordinates of that observation, without directly reusing the boundary values ​​obtained from other observation times.

[0052] Step 1.4: Convert the tilted wide-span observations into a regular rectangular grid; After completing the geometric boundary construction, extract the valid sea surface height data that simultaneously meets the following conditions: Longitude and latitude are valid; Sea surface height anomaly input value is valid; Located within the latitude range preset by the corresponding orbit; For any valid data point, obtain the corresponding local left and right boundaries based on its latitude binning, and calculate the normalized coordinates of the horizontal track according to the following formula. :

[0053] in, Let be the longitude of the i-th observation point; and These are the left and right boundaries of the j-th latitude bin, respectively; This represents the normalized position of the horizontal rail within a local wide area. When the local left and right boundaries are invalid or the right boundary is not greater than the left boundary, the corresponding mapping point is discarded. Using the center of the latitude bin as the coordinate along the track, the normalized coordinates of the horizontal track are uniformly mapped to 128 fixed horizontal track grid points. The number of grids along the track direction is determined by the number of latitude bins, and may be padded to an even number if necessary; Subsequently, scattered linear interpolation is used to interpolate the effective observations onto a regular rectangular grid, resulting in a two-dimensional input matrix suitable for convolutional neural network processing. At the same time, the longitude, latitude, and sea surface height information of the original effective data points are preserved as the basis for subsequent mapping of the prediction results to the actual geographic coordinates.

[0054] Step 1.5: Calculate the normalization parameters and process invalid pixels; Based on all valid pixels in each orbit training set, calculate the mean and standard deviation of the total sea surface height anomaly input data and the target data of the sea surface height anomaly induced by the internal tide, excluding null values ​​during the calculation process; The input data is normalized according to the following formula:

[0055] The sea surface height anomaly induced by the target internal tide is normalized according to the following formula:

[0056] The mean and standard deviation are determined only by the training set, while the validation set, test set, and actual SWOT observations all use the same set of training set statistical parameters for the corresponding orbits. During the training phase, an effective ocean mask is established based on whether the target field of sea surface height anomalies induced by inland tides is null.

[0057] In practice, null values ​​in the training input and target are temporarily replaced with 0, but invalid regions are not included in the loss function calculation. Invalid locations during the training phase mainly include locations outside the land and simulation areas. The null values ​​in the normalized input are replaced with 0, which is equivalent to using the mean of the training set as a temporary filler value, and the corresponding target locations are not included in the loss calculation. Phase (II): Model building, training, and practical SWOT observation application; Step 2.1: Establish a deep learning-based internal flow extraction model; The deep learning internal moisture extraction model described above employs U-Net, residual U-Net, attention U-Net, encoder-decoder convolutional network, or a network combining convolution and attention mechanisms. The encoder-decoder of the model has a 4-level, 5-level, or 6-level structure. This embodiment uses a six-level encoder-decoder U-Net model, with the regularized and normalized two-dimensional sea surface height anomaly input field as a single-channel input and the corresponding two-dimensional intidal-induced sea surface height anomaly as a single-channel output. The encoder first converts the number of input channels from 1 to 64 through a double convolution module, and then sequentially expands the number of feature channels by doubling them one by one through six downsampling modules, such as: 64→128→256→512→1024→2048→4096; Each dual convolutional module consists of two consecutive 3×3 convolutional layers; after each convolutional layer, a batch normalization layer and a linear rectified unit activation layer are set in sequence; the convolution is padded with 1 pixel to keep the spatial size consistent before and after the convolution; each downsampling module first uses 2×2 max pooling, and then performs dual convolution operation; The decoder includes six upsampling modules, and the number of feature channels is restored by halving them one by one, for example, the restored number of channels is: 4096→2048→1024→512→256→128→64. Each upsampling module first uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to restore the spatial resolution, then concatenates it with the feature map of the corresponding scale of the encoder, and fuses the features through a dual convolution module; When there is a difference in spatial dimensions between the upsampled features and the encoder features, the upsampled features are symmetrically padded to make their dimensions consistent. Finally, a 1×1 convolutional layer is used to convert the 64-channel features into a single-channel output of tidal-induced sea surface height anomalies. Step 2.2: Train the model; To simultaneously constrain the numerical amplitude and spatial gradient of sea surface height anomalies induced by internal tides, the model is trained using L1 loss, MSE loss, or a gradient enhancement composite loss function. This embodiment employs a mask gradient enhancement composite loss function, and calculates the data error term using the following formula. :

[0058] in, Extract predicted values ​​for the model; The true value of the target; For effective marine cover; To prevent the use of tiny positive numbers with a denominator of zero, this embodiment uses 10. -8 ; Then, the spatial gradients of the prediction field and the target field are calculated using the horizontal and vertical Sobel operators, respectively; The operator used is:

[0059]

[0060] The gradient error term is:

[0061] The total loss function is:

[0062] In this embodiment, the gradient error weighting coefficient α is taken as 0.001; The loss function described above is calculated only on valid ocean pixels, excluding land and invalid regions from model optimization; the gradient error term is used to constrain the reconstruction of tidal wave peaks, troughs, and spatial gradient structures. Step 2.3: Training the model and parameters; This embodiment uses the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 The weight decays to 1×10 -5 The maximum number of training rounds is 100; in the batch size sensitivity experiment, 32, 50 and 64 were used for comparison.

[0063] The learning rate mentioned is a constant learning rate, a stepped decay learning rate, or a cosine annealing learning rate; In this embodiment, the learning rate is cosine annealing, which gradually decreases from the initial learning rate. The cosine annealing cycle is 100 rounds, and the minimum learning rate is 1×10⁻⁶. -7 After each training round, the composite loss is calculated using the validation set. When the validation set loss is lower than the previous best value, the current model parameters are saved. Training is terminated early if the validation set loss does not improve for 15 consecutive rounds; the random seed is fixed at 42 during training to improve the reproducibility of the model training results. Step 2.4: Extraction of internal tides from actual SWOT satellite observations; For each SWOT track to be processed, perform the following operations: First, the original effective pixel mask of the SWOT observation is saved. For null values ​​in the input matrix after regularization, since the nadir gaps and local missing measurements in the SWOT observations are located within the ocean area covered by the convolution operation, a temporary filling method based on the nearest effective pixel extrapolation using Euclidean distance transformation is used to avoid propagating null values ​​during convolution calculation. This filling is only used to form the complete model calculation matrix. After model inference and inverse normalization are completed, the original missing measurement areas are reset to null values ​​based on the original effective pixel mask saved before inference, thus ensuring that the final result only retains the actual effective observation coverage area. Then, the SWOT Level 3 product of the corresponding orbit is read; mean dynamic topography, barotropic tide, internal tide, and dynamic atmospheric correction terms are added back to construct the sea surface height anomaly input field; the orbital geometry boundary of this observation is converted into a regular rectangular grid with a preset fixed number of horizontal rail grid points (128 columns); the original effective observation mask is saved; the missing values ​​in the regular grid are filled using the nearest effective pixel extrapolation method; normalization is performed using the mean and standard deviation of the corresponding orbital training set; the normalization result is input into the deep learning internal tide extraction model trained for the corresponding orbit; the model output is inversely normalized using the mean and standard deviation of the target internal tide training set; the original invalid observation positions are restored to missing values; the instantaneous internal tide-induced sea surface height anomaly results under the regular rectangular grid are output; Then, based on the saved original geographic information, for each original valid observation point, interpolation is performed in the prediction results of the regular rectangular grid according to the bin of latitude and the normalized coordinates of the horizontal rail to obtain the predicted value of the sea surface height anomaly induced by the internal tide corresponding to the observation point. Finally, the predicted values ​​of sea surface height anomalies induced by internal tides are mapped one-to-one with the original longitude and latitude to form the output results under the original SWOT tilted wide-span geographic coordinates; no predicted values ​​are output for the original missing measurement points, thus restoring the output results to the original SWOT tilted wide-span geographic coordinates.

[0064] As described above, after applying the trained model to actual SWOT wide-swath observations, the resulting sea surface height anomaly field induced by internal tides can reproduce the main large-scale stationary structures represented by the high-resolution empirical internal tide product (HRET), indicating that the method can handle actual satellite observations. The model results also retain more instantaneous spatial variations related to the ocean state in a single observation; however, since there is currently a lack of independent two-dimensional instantaneous internal tide observation ground truths, these additional variations are not considered direct proof for accurately extracting non-stationary internal tides.

[0065] To verify the effectiveness of this embodiment, this application was applied to nine different SWOT orbits in a certain sea area to conduct independent simulation data experiments, in order to verify the applicability of the method under different internal tide generation and propagation environments.

[0066] Specifically, based on the geographical coverage of the nine actual SWOT orbits, the total sea surface height anomaly and the intidal-induced sea surface height anomaly for the corresponding regions were extracted from the LLC4320 high-resolution numerical simulation. Data preprocessing and model training were performed separately for each orbit. Each orbit used 200 consecutive days of hourly data and the same number of training samples (3764 sets), validation samples (470 sets), and test samples (470 sets). The nine orbits were each trained with their own U-Net models, using the gradient-enhanced composite loss, cosine annealing learning rate, six-level encoder-decoder structure, and batch size of 32 determined in the aforementioned sensitivity experiments as a unified optimal configuration; hyperparameters were not individually adjusted for each orbit.

[0067] The results for the test data are as follows: orbital 243 R=0.8273, RMSE=0.98cm; orbital 521 R=0.7889, RMSE=1.24cm; orbital 215 R=0.8098, RMSE=1.36cm; orbital 493 R=0.7973, RMSE=1.53cm; orbital 187 R=0.7831, RMSE=1.68cm; orbital 465 R=0.8373, RMSE=1.62cm; orbital 159 R=0.8802, RMSE=1.76cm; orbital 437 R=0.8706, RMSE=1.78cm; orbital 131 R=0.9204, RMSE=1.78cm. The spatial correlation coefficients of all nine orbitals are higher than 0.78, and the root mean square error is lower than 1.8cm.

[0068] Considering the differences in the generation and propagation characteristics of internal tides at different locations in a certain sea area, this application establishes corresponding training regions based on 9 SWOT orbitals, and uses LLC4320 training samples within each orbital range to train the model and normalize the parameters, so that the model learns the spatial characteristics of internal tides in the corresponding region and obtains the corresponding model weights.

[0069] To determine a suitable model configuration for extracting sea surface height anomalies induced by internal tides, a hyperparameter sensitivity experiment was conducted using SWOT orbital 159 as a representative orbital. This orbital employed the same data partitioning method as described above: 3764 sets were used for training, 470 sets for validation, and 470 sets for model configuration comparison and performance testing. The sensitivity experiment used a controlled variable approach, changing only one parameter at a time while maintaining the other training conditions consistent with the baseline model. The baseline model employed an L1 loss function, a constant learning rate, a five-level encoder-decoder structure, and a batch size of 50. On the 470 test samples, the average spatial correlation coefficient R was 0.8560, and the root mean square error (RMSE) was 1.90 cm.

[0070] First, the loss functions were compared. Keeping other parameters constant, the L1 loss was replaced with mean squared error loss (MSE) and gradient boosting composite loss (Hybrid). Using MSE, R increased to 0.8631, and RMSE decreased to 1.86cm; using gradient boosting composite loss, R further increased to 0.8673, RMSE remained at 1.86cm, and it better preserved the tidal wave peaks, troughs, and local spatial gradient structure. Therefore, gradient boosting composite loss was chosen.

[0071] The learning rate scheduling methods were then compared. With other parameters remaining constant, the stepped decay learning rate yielded R=0.8693 and RMSE=1.86 cm, while the cosine annealing learning rate yielded R=0.8689 and RMSE=1.84 cm. Both were superior to the constant learning rate, with cosine annealing achieving a lower root mean square error and better overall reconstruction performance; therefore, the cosine annealing learning rate was chosen.

[0072] In the encoder-decoder layer experiment, the four-level structure achieved R=0.8556 and RMSE=1.92 cm, which is slightly lower than the five-level baseline structure; the six-level structure achieved R=0.8686 and RMSE=1.84 cm, indicating that increasing the multi-scale feature extraction range of the network is beneficial for reconstructing complex internal tidal space structures. Therefore, a six-level encoder-decoder structure was adopted.

[0073] In the batch size experiment, a batch size of 32 yielded R=0.8624 and RMSE=1.89 cm, while a batch size of 64 yielded R=0.8604 and RMSE=1.87 cm. Both settings enabled stable training, with batch size 32 exhibiting a higher spatial correlation coefficient. Considering the spatial structure correlation, amplitude error, and overall performance combined with other optimized parameters, batch size 32 was ultimately selected.

[0074] Finally, the best-performing configurations from the above experiments—namely, gradient-enhanced composite loss, cosine annealing learning rate, a six-level encoder-decoder structure, and a batch size of 32—were trained. The final combined model achieved an average spatial correlation coefficient of 0.8802 and a root mean square error (RMSE) reduced to 1.76 cm on 470 test samples of track 159. Compared to the baseline model (R=0.8560, RMSE=1.90 cm), both spatial structure consistency and prediction error were improved. Therefore, this combination was chosen as the preferred configuration for training the subsequent nine SWOT track models in this embodiment. It should be noted that the optimal values ​​obtained from the individual parameter sensitivity experiments do not necessarily mean that their combination constitutes the globally optimal configuration among all possible combinations. The configuration used in this embodiment is a comprehensive optimal scheme formed based on the above sensitivity experiment results. The sensitivity comparison is shown in the table below.

[0075] Table 1 Comparison of Sensitivity Experiment Results The aforementioned nine orbits extend from the western propagation area of ​​a certain sea area to the nearby major intidal generation area, and their intidal signal amplitude, spatial regularity, and propagation environment show significant differences. After training and testing in the aforementioned different regions, high spatial correlation coefficients and low root mean square errors were achieved, indicating that the training label construction, orbit regularization, and deep learning extraction process established in this application is not only applicable to a single orbit or a single intidal spatial structure, but can stably extract the main intidal-induced sea surface height anomaly structures under different intidal dynamic environments. The simulation data test results are shown in the table below.

[0076] Table 2. Simulation data test results of nine SWOT orbits in a certain sea area Through the aforementioned preprocessing steps, including physical quantity reconstruction of the actual SWOT product, tilt and wide-swath regularization, deep learning model inference, and original geographic coordinate recovery, the sea surface height anomaly field induced by the internal tide induced by a single satellite observation can be obtained. The results can reproduce the main large-scale stable internal tide spatial structure represented by the high-resolution empirical internal tide product (HRET), indicating that the model trained from numerical simulation data can complete the processing of actual SWOT wide-swath observations and the extraction of internal tide signals. Since there is currently no actual two-dimensional instantaneous internal tide observation ground truth corresponding to the time of a single SWOT observation, HRET is only used as a reference product for the stable internal tide structure. Therefore, the above comparison is used to illustrate the feasibility of this application in practical applications and the consistency of the main spatial structure, and not as a direct verification of the accuracy of instantaneous internal tide extraction.

[0077] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based SWOT satellite observation internal tide signal extraction method, characterized in that: From the three-dimensional temperature and salinity data of high-resolution ocean numerical simulation, target data of intidal-induced sea surface height anomalies corresponding to the spatiotemporal total sea surface height anomalies are constructed to form input-target data pairs for deep learning model training; based on the relevant correction variables in the SWOT product, the actual sea surface height anomaly input field with a physical definition consistent with the numerical simulation training input is reconstructed. For each actual SWOT wide-swath observation, the orbital geometric boundary is constructed based on the effective latitude and longitude coordinates, without using the absence of sea surface height anomalies at the corresponding locations as a boundary determination condition; the normalized coordinates of the transverse track are calculated based on the local left and right boundaries, and the inclined wide-swath observation is converted into a regular two-dimensional matrix; the sea surface height anomaly induced by the internal tide is extracted using the trained deep learning model to temporarily fill in the missing locations required for model calculation, and the missing areas are restored based on the original effective observation mask after the model output; finally, the prediction results in the regular two-dimensional matrix are restored to the original SWOT geographic coordinates.

2. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 1, characterized in that: The implementation process includes the following steps: Phase (1): Data Construction and Preprocessing; Step 1.1: Construct training samples for total sea surface height anomaly and intidal-induced sea surface height anomaly; Training samples were constructed using high-resolution ocean numerical simulation data that can continuously output temperature, salinity, and total sea surface height over time and analyze tidal dynamic processes. Step 1.2: Establish a SWOT wide-span observation input field consistent with the physical definition of the numerical simulation training input; Read the filtered sea surface height anomalies, mean dynamic topography, barotropic tides, internal tides, and dynamic atmospheric correction variables from the SWOT Level 3 Ka-band radar interferometer low-rate sea surface height expert product. Step 1.3: Construct the SWOT orbital geometry boundary based on geographic coordinates; Effective latitude processing ranges were set for different SWOT orbits, and the total sea surface height anomaly and the intidal-induced sea surface height anomaly for the corresponding wide area were extracted from the high-resolution ocean numerical simulation data based on the actual geographical location of each orbit. Each track undergoes separate region extraction, boundary construction, regularization, and model training, without sharing the geometric boundaries of other tracks; Step 1.4: Convert the tilted wide-span observations into a regular rectangular grid; Step 1.5: Calculate the normalization parameters and process invalid pixels; Based on all valid pixels in each orbit training set, calculate the mean and standard deviation of the total sea surface height anomaly input data and the target data of the sea surface height anomaly induced by the internal tide, excluding null values ​​during the calculation process; Phase (II): Model building, training, and practical SWOT observation application; Step 2.1: Establish a deep learning-based internal flow extraction model; The deep learning internal moisture extraction model described above employs U-Net, residual U-Net, attention U-Net, encoder-decoder convolutional network, or a network combining convolution and attention mechanisms. The encoder-decoder of the model has a 4-level, 5-level, or 6-level structure. Step 2.2: Train the model; The model is trained using L1 loss, MSE loss, or a gradient boosting composite loss function. Step 2.3: Training the model and parameters; The AdamW optimizer is used, with learning rates of constant learning rate, step decay learning rate, or cosine annealing learning rate. Step 2.4: Extraction of internal tides from actual SWOT satellite observations; For each SWOT track to be processed, perform the following operations: Read the SWOT Level 3 product for the corresponding orbit; add back the mean dynamic topography, barotropic tide, internal tide, and dynamic atmospheric correction terms to construct the sea surface height anomaly input field; convert the orbital geometric boundary observed for the corresponding orbit into a regular rectangular grid with a preset fixed number of transverse rail grid points; save the original effective observation mask; fill in the missing values ​​in the regular grid using the nearest effective pixel extrapolation method; normalize using the mean and standard deviation of the corresponding orbital training set; The normalized result is input into the deep learning endocytic extraction model trained on the corresponding orbit; the mean and standard deviation of the target endocytic training set are used to inversely normalize the model output; the original invalid observation positions are restored to null values; and the instantaneous endocytic induced sea surface height anomaly results under the regular rectangular grid are output. Then, based on the saved original geographic information, for each original valid observation point, interpolation is performed in the prediction results of the regular rectangular grid according to the bin of latitude and the normalized coordinates of the horizontal rail to obtain the predicted value of the sea surface height anomaly induced by the internal tide corresponding to the observation point. Finally, the predicted values ​​of sea surface height anomalies induced by internal tides are mapped one-to-one with the original longitude and latitude to form the output results under the original SWOT tilted wide-span geographic coordinates; no predicted values ​​are output for the original missing measurement points, thus restoring the output results to the original SWOT tilted wide-span geographic coordinates.

3. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 2, characterized in that: Step 1.1, using the LLC4320 total sea surface height anomaly as model input, calculates seawater density according to the TEOS-10 seawater thermodynamic standard using temperature and salinity data, and calculates the density perturbation of instantaneous density relative to time-averaged density: in, Represents instantaneous density. Represents the average density over time; Calculation of baroclinic pressure anomaly based on density perturbation under static approximation: in, Represents gravitational acceleration; , These represent the vertical coordinates of the sea surface and the seabed, respectively. This indicates the local water depth; Represents the vertical integral coordinates; Through depth-independent surface pressure terms The vertical average value of the baroclinic pressure anomaly is eliminated; then the baroclinic pressure anomaly at sea surface is converted into a baroclinic sea surface height anomaly. : in, It is a fixed reference density; Using Butterworth bandpass filters, respectively for Conduct diurnal tidal frequency band and semi-diurnal tidal zone The bandpass filter was applied, and the two parts were added together to obtain the following sea surface height anomaly induced by intidal currents. : This results in a training data pair that corresponds one-to-one in time and space: "total sea surface height anomaly - sea surface height anomaly induced by internal tides"; The samples are divided into training set, validation set and test set in chronological order; the training set is used for model parameter updates and normalization parameter calculation, the validation set is used for monitoring the training process and early termination, and the test set is used for comparing different model configurations and evaluating model performance.

4. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 3, characterized in that: Step 1.2 involves constructing the actual SWOT sea surface height anomaly input field corresponding to the physical definition of the numerical simulation training input according to the following formula. : in, The sea level in the SWOT analysis is abnormal. Average dynamic terrain; This is a positive pressure tidal correction term; For internal moisture correction; This is a dynamic atmospheric correction item.

5. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 4, characterized in that: In step 1.3, for each SWOT wide-span observation to be processed, the observation is divided according to a preset latitude interval (1 / 48°). Only observation points with valid longitude and latitude are used to construct the orbital geometric boundary of the corresponding orbital observation. The validity of the sea surface height anomaly at the corresponding location is not used as the boundary determination condition. For each latitude bin, calculate the minimum and maximum longitude of all valid geographic coordinates within that bin, which will serve as the local left and right boundaries of that latitude location. For latitude bins whose boundaries cannot be directly obtained due to observation gaps, linear interpolation is performed using the boundary values ​​of adjacent valid bins to fill in the corresponding left and right boundaries.

6. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 5, characterized in that: In step 1.4, after completing the geometric boundary construction, valid sea surface height data that simultaneously meets the following conditions is extracted: both longitude and latitude are valid; the sea surface height anomaly input value is valid; and it is located within the preset latitude range of the corresponding orbit. For any valid data point, obtain the corresponding local left and right boundaries based on its latitude binning, and calculate the normalized coordinates of the horizontal track according to the following formula. : in, Let be the longitude of the i-th observation point; and These are the left and right boundaries of the bin at the j-th latitude where the i-th observation point is located; Let i be the normalized position of the horizontal rail within the local wide range of observation points; When the local left and right boundaries are invalid or the right boundary is not greater than the left boundary, the corresponding mapping point is discarded. Using the center of the latitude bin as the coordinate along the track, the normalized coordinates of the horizontal track are uniformly mapped to 128 fixed horizontal track grid points. The number of grid cells along the track direction is determined by the number of latitude bins, and is padded to an even number. Subsequently, scattered linear interpolation is used to interpolate the effective observations onto a regular rectangular grid, resulting in a two-dimensional input matrix suitable for convolutional neural network processing. At the same time, the longitude, latitude, and sea surface height information of the original effective data points are preserved as the basis for subsequent mapping of the prediction results to the actual geographic coordinates.

7. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 6, characterized in that: In step 1.5, the input data is normalized according to the following formula: The sea surface height anomaly induced by the target internal tide is normalized according to the following formula: The mean and standard deviation are determined only by the training set, while the validation set, test set, and actual SWOT observations all use the same set of training set statistical parameters for the corresponding orbits. During the training phase, an effective ocean mask is established based on whether the target field of sea surface height anomalies induced by inland tides is null. Null values ​​in the training input and target data are temporarily replaced with preset values ​​before model calculation, and invalid positions are not included in the loss function calculation.

8. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 2, characterized in that: Step 2.1 uses a six-level encoder-decoder U-Net model, with the regularized and normalized two-dimensional sea surface height anomaly input field as a single-channel input and the corresponding two-dimensional intidal-induced sea surface height anomaly as a single-channel output. The encoder first converts the number of input channels from 1 to 64 through a double convolution module, and then successively doubles and expands the number of feature channels through six downsampling modules. Each dual convolutional module consists of two consecutive 3×3 convolutional layers; after each convolutional layer, a batch normalization layer and a linear rectified unit activation layer are set in sequence; the convolution is padded with 1 pixel to keep the spatial size consistent before and after the convolution; each downsampling module first uses 2×2 max pooling, and then performs dual convolution operation. The decoder includes six upsampling modules, which restore the feature channels by halving them sequentially. Each upsampling module first uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to restore the spatial resolution, then concatenates it with the feature map of the corresponding scale of the encoder, and fuses the features through a dual convolution module; When there is a difference in spatial size between the upsampled features and the encoder features, the upsampled features are symmetrically padded to make their sizes consistent. Finally, the 64-channel features are converted into a single-channel intidal-induced sea surface height anomaly output through a 1×1 convolutional layer.

9. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 8, characterized in that: In step 2.2, the mask gradient enhancement composite loss function is used to calculate the data error term using the following formula. : in, Extract predicted values ​​for the model; The truth value of the target; For effective marine cover; To prevent the use of tiny positive numbers with a denominator of zero, this embodiment uses 10. -8 ; Then, the spatial gradients of the prediction field and the target field are calculated using the horizontal and vertical Sobel operators, respectively; The operator used is: The gradient error term is: The total loss function is: Where α is the weighting coefficient of the gradient error term, with a value of 0.001; The loss function described above is calculated only on valid ocean pixels, and the gradient error term is used to constrain the reconstruction of tidal wave crests, troughs, and spatial gradient structures.

10. The deep learning-based SWOT satellite observation intidal signal extraction method according to claim 9, characterized in that: In step 2.3, the learning rate is cosine annealing, gradually decreasing from the initial learning rate; after each training round, the composite loss is calculated using the validation set, and the current model parameters are saved when the validation set loss is lower than the previous best value. Training will be terminated early if the validation set loss exceeds a preset number of rounds without improvement. A fixed random seed is used during training to improve the reproducibility of model training results.

Images