SAR (Synthetic Aperture Radar) sea surface significant wave high-depth learning inversion method fused with multi-source data

Abstract

The invention provides a multi-source data fused SAR sea surface significant wave height deep learning inversion method, and belongs to the technical field of remote sensing ocean, and the method specifically comprises the steps: preprocessing buoy data, and obtaining sea wave significant wave height data observed by a buoy; acquiring dual-polarization single-view complex SAR data collected in an interference wide-width mode; preprocessing the SAR data, and taking significant wave height data obtained by buoy observation as a label of the SAR data; acquiring auxiliary data corresponding to the SAR data, wherein the auxiliary data comprises wind speed, wind direction data, rainfall data and OCN data; and constructing a multi-source data-fused SAR significant wave high-depth learning inversion model, training the multi-source data-fused SAR significant wave high-depth learning inversion model, and testing and verifying the trained model by using the test set and the verification set to obtain a final significant wave height inversion result. According to the method, the SAR data inversion capability of the model is improved.

Description

Technical Field

[0001] This invention belongs to the technical field of marine remote sensing, and in particular relates to a deep learning inversion method for SAR sea surface effective wave height fused from multi-source data. Background Technology

[0002] Ocean waves have a significant impact on marine engineering, marine transportation, coastal port construction, marine fisheries, and the lives of coastal residents. Accurately acquiring ocean wave information is a crucial research topic in the field of marine remote sensing. Synthetic Aperture Radar (SAR) technology has wide applications in marine monitoring, and acquiring a large amount of sea surface SAR data lays the data foundation for inverting ocean wave information. Traditional SAR methods for inverting the effective wave height of the sea surface are mainly divided into physical methods and empirical models. The first type uses SAR data to invert the wave direction spectrum, thereby obtaining ocean wave information. Although the wave spectrum can provide relatively complete ocean wave information, the nonlinear mapping relationship between the SAR data spectrum and the wave spectrum is complex, making it difficult to obtain an accurate wave spectrum, especially the wind wave spectrum, from SAR data inversion. The second type is based on fitting methods to establish empirical relationship models between SAR data parameters and ocean wave information. For example, the CWAVE model (C-band Wave Retrieval Algorithm) establishes a relationship between wave height and SAR image spectrum through nonlinear fitting; parameterized models use features such as the average brightness and image variance of SAR data to invert wave height. These empirical relational models use few parameters and have insufficient nonlinear fitting capabilities, resulting in low accuracy in retrieving significant wave height. Furthermore, numerical simulation methods combined with SAR data are used to retrieve wave height. By adjusting model parameters, the model output is made to match the characteristics of the SAR data. However, wave physics models are complex, computationally intensive, and demand high-performance computing capabilities. Furthermore, model parameters are difficult to obtain precisely, leading to uncertainties in the retrieval results.

[0003] In recent years, the development of deep learning technology has provided a new solution for SAR wave significant height (SWH) inversion. Deep learning models can automatically extract multi-level features from large amounts of data and directly establish a mapping from the preprocessed SAR image spectrum to the SWH, without relying on complex physical assumptions or empirical corrections in traditional models. Typical CNN convolutional modules can extract local features but cannot extract long-range dependencies between different features. RNN networks can infer feature relationships between multi-source data but have shortcomings in multi-source data fusion. Because the generation of SAR images is affected by sea state, there is a certain degree of error in the calculated SWH. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a deep learning inversion method for SAR sea surface significant wave height that integrates multi-source data, thereby solving the problems of insufficient feature extraction from multi-source data and insufficient fusion of multi-source and multi-scale features when using deep learning to invert significant wave height.

[0005] Technical solution: This invention discloses a deep learning inversion method for SAR sea surface effective wave height fused from multi-source data, specifically as follows:

[0006] Preprocess the buoy data to obtain the significant wave height data of the ocean waves observed by the buoy;

[0007] Acquire dual-polarized single-look complex SAR data collected in interferometric wide-swath mode;

[0008] Preprocess the SAR data, using the effective wave height data as the label for the SAR data; construct training, testing, and validation sets.

[0009] Acquire auxiliary data corresponding to SAR data, including wind speed, wind direction, precipitation, and OCN data; construct corresponding training, testing, and validation sets.

[0010] A deep learning inversion model for SAR effective wave height fusion based on multi-source data was constructed. The model was trained using SAR data training set and auxiliary data training set. The trained model was then tested and validated using test set and validation set to obtain the final effective wave height inversion result.

[0011] Furthermore, the buoy data preprocessing is as follows: First, check the maximum, minimum, and median values ​​in each data set and remove outliers; then, if the change in observed data between adjacent times or locations exceeds a preset range, it is considered data anomaly and the abnormal data is removed. The wave class is divided into three levels, with the first, second, and third wave height intervals set. Waves in the first wave height interval are considered small waves, those in the second wave height interval are considered medium waves, and those in the third wave height interval are considered large waves; if the wave class observed in the buoy data does not match the actual wave height, it is considered abnormal data and removed; the buoy data after removing abnormal data is normalized.

[0012] Further preprocessing of SAR data includes: orbit correction, radiometric calibration, pulse band correction, polarization matrix, multi-look, terrain correction, land and sea masking, as well as cropping and channel synthesis;

[0013] The orbit correction specifically involves downloading precise orbit parameter values ​​from the orbit file server and replacing the orbit parameter values ​​that are included in the SAR data.

[0014] The radiometric calibration specifically involves establishing the relationship between SAR data and the backscattering coefficients of ground targets through absolute calibration, using the following formula:

[0015] ;

[0016] in, Let i be the radar backscattering coefficient, and i and j be the pixel coordinates in the SAR data, respectively. Grayscale value These are calibration parameters;

[0017] The polarization matrix is ​​generated specifically by using the following formula to generate the covariance matrix from the SAR data of the VV and VH polarization channels. : ;

[0018] in, , , , E(.) represents the expected value. , These represent SAR data matrices with VV and VH polarizations, respectively. , These represent the SAR backscattering intensities of the VV-polarized and VH-polarized channels, respectively. and This is the intermediate parameter matrix;

[0019] Terrain correction specifically involves using the digital elevation model of the Space Shuttle Radar Terrain Survey Mission to calculate the true geographical location of each pixel in the SAR data based on the three-dimensional information of the Earth's surface. This eliminates geometric distortion and radiometric distortion caused by the undulations of the Earth's surface, ensuring that the SAR data is accurately aligned with the geographical location.

[0020] The cropping and channel synthesis process involves cropping a 256×256 sub-block from the SAR data, centered on the buoy's latitude and longitude. Each sub-block corresponds to only one preprocessed buoy data point. Based on this sub-block, the matrix is ​​then synthesized. and The incident angle data are used as the R, G, and B channels of the RGB image, respectively; and these three channels are merged into an RGB image to obtain the final preprocessed SAR data.

[0021] Furthermore, precipitation data with a temporal resolution of 0.5 h, a spatial resolution of 0.1°×0.1°, and a time difference of less than 0.5 h from buoy data were acquired. Wind speed and direction spectra from the OCN Level 2 products were obtained. Using the spatiotemporal data of the cropped SAR data as a reference, the wind speed and direction spectra were cropped and generated as 10×10 grayscale images. The wind component u in the u direction and the wind component v in the v direction at 10 m were acquired, with a temporal resolution of 1 h and a spatial resolution of 0.25°×0.25°. Wind composites were performed on u and v to obtain the sea surface wind speed and direction data at 10 m. The specific wind composite formula is as follows:

[0022] ;

[0023] in, The wind direction data is based on due north, which is the final synthesized wind direction data. (-1) represents the reverse direction. Represents wind direction data based on due east: .

[0024] Furthermore, the deep learning inversion model for SAR effective wave height, which integrates multi-source data, inputs the preprocessed SAR data into the first 3D convolutional layer. A dual-branch information enhancement attention module is introduced into the first 3D convolutional layer. The OCN data is then input into the second 3D convolutional layer. The outputs of the dual-branch information enhancement attention module and the second 3D convolutional layer are input into an interactive attention Transformer for feature fusion. The output of the interactive attention Transformer is then input into the feature relation inference module. Auxiliary data is passed through a fully connected layer and then input into the feature relation inference module. The feature relation inference module outputs the final effective wave height inversion result.

[0025] Furthermore, the dual-branch information enhancement attention module divides the input into G sub-features along the channel dimension, and performs horizontal global average pooling and vertical global average pooling on these G sub-features respectively.

[0026] ;

[0027] ;

[0028] in, This is the tensor obtained by horizontal global average pooling. Here, H is the input height and W is the input width, obtained by vertical global average pooling. Let g be the g-th sub-feature; g = 0, 1, ..., G-1;

[0029] Will and The input is concatenated and fed into a 1×1 convolution. The output of the 1×1 convolution is decomposed into two vectors, which are then fed into two non-linear sigmoid functions. The outputs of the two non-linear sigmoid functions are reweighted with G sub-features to alleviate overfitting. The reweighted result is then subjected to group normalization and global average pooling. The G sub-features are fed into a 3×3 convolution, and the result of the 3×3 convolution is subjected to global average pooling. The results of the two global average pooling operations are multiplied, and the result of the multiplication is subjected to softmax normalization. The result of the 3×3 convolution and the result of the softmax normalization are then multiplied together. Finally, the input is passed through a sigmoid activation function, and the output of the sigmoid activation function is reweighted with the G sub-features. The final reweighted result is then multiplied with the original input.

[0030] Furthermore, the interactive attention Transformer has a two-branch structure, including a SAR branch and an OCN branch. Each branch contains an embedding layer, an interactive attention Transformer module, and a regular Transformer module. The embedding layer scales the input and divides it into multiple small blocks according to a predetermined size. A flattening layer flattens these blocks into one-dimensional vectors and adds positional encoding to the heads of all vectors. The two embedding layers scale the two inputs to the same size. The interactive attention Transformer module includes an interactive attention multi-head self-attention module and a multilayer perceptron. The interactive attention multi-head self-attention module maps the output of the embedding layer to a query Q, a key K, and a value V vector. The key K vector in the SAR branch... Sum of vectors V The key K vector in the OCN branch serves as the input to the interactive attention Transformer module of the OCN branch. Sum of vectors V As input to the interactive attention Transformer module in the SAR branch, specifically:

[0031] ;

[0032] ;

[0033] in, This is the output of the interactive attention multi-head self-attention module in the SAR branch. This is the output of the interactive attention multi-head self-attention module in the OCN branch. This indicates a high level of attention from multiple parties. For the query Q vector in the SAR branch, This refers to the query Q-vector in the OCN branch;

[0034] The output of the embedding layer and the output of the interactive attention multi-head self-attention module are multiplied, normalized by the layer, and then fed into the multilayer perceptron. The output of the multilayer perceptron is multiplied with the normalized result, normalized again, and then fed into the regular Transformer module.

[0035] Furthermore, the feature relationship reasoning module includes a first and third two-layer bidirectional long short-term memory network. The output of the SAR branch of the interactive attention Transformer is input to the first two-layer bidirectional long short-term memory network, and the interactive attention Transformer... The output of the branch is input to the second two-layer bidirectional long short-term memory network, and the auxiliary data is input to the third two-layer bidirectional long short-term memory network through the fully connected layer. The two bidirectional long short-term memory networks process the data simultaneously in both the forward and reverse directions of the input sequence. The output of the feature relationship inference module is then passed through the fully connected layer to obtain the final effective wave height inversion result.

[0036] Furthermore, the Sparrow Optimization algorithm is used to globally optimize the number of hidden layer neurons, the number of segments in the embedding layer, the number of self-attention heads in the interactive attention multi-head self-attention module, the initial learning rate, and the learning rate adjustment factor in the bidirectional long short-term memory network.

[0037] Furthermore, firstly, a refraction-based reverse learning strategy is used to initialize individual sparrows in the optimization algorithm, generating a set of model parameters and constructing the model. Secondly, a sine and cosine strategy is used to update the position of the discoverer.

[0038] ;

[0039] in, This represents the position information of the i-th sparrow in the (t+1)-th iteration. This represents the position information of the i-th sparrow in the t-th iteration. It is a random number. This is the current optimal solution;

[0040] Finally, the Cauchy mutation strategy was used to calculate the follower positions during group foraging:

[0041] ;

[0042] in, This represents the j-th parameter information of the i-th sparrow in the (t+1)-th iteration; Let be the optimal solution for the parameters of the i-th sparrow in the t-th iteration. This is the standard Cauchy distribution function.

[0043] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0044] (1) An information-enhanced attention module is proposed to integrate spatial features into channel attention through feature extraction. The introduction of this module enables the model to better learn the interdependencies between different channels of preprocessed SAR data, thereby improving the model's ability to invert SAR data.

[0045] (2) Use the interactive attention Transformer module to improve the common Transformer model, so that while extracting data features, the model can fuse SAR data and OCN (Level-2 Ocean Geophysical Product) data features to form a mutually conditional attention feature, thereby optimizing the fusion effect of multi-scale features.

[0046] (3) Due to the azimuth truncation effect of SAR data, the obtained spectrum ignores small-scale wave energy. Introducing multi-source data to supplement the missing information not only enriches the types of inversion data but also improves the accuracy of effective wave height inversion.

[0047] (4) Introduce the sparrow optimization algorithm to optimize the key parameters in the model, and use Cauchy mutation and sine / cosine strategies to optimize the model parameters and enhance the global optimal search of the algorithm. Attached Figure Description

[0048] Figure 1 This is an overall flowchart of the present invention;

[0049] Figure 2 This is a flowchart of the SAR data preprocessing process of the present invention;

[0050] Figure 3 This is a structural diagram of the SAR effective wave height deep learning inversion model that integrates multi-source data according to the present invention;

[0051] Figure 4 This is a structural diagram of the Information Enhancement Attention (IEAM) module in the inversion model of this invention;

[0052] Figure 5 This is a structural diagram of the feature fusion module based on interactive attention Transformer in the inversion model of the present invention. Detailed Implementation

[0053] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0054] The SAR effective wave height inversion method based on fused multi-source data described in this invention includes the following steps:

[0055] (1) Obtain the effective wave height data of the sea waves observed by the buoy in the Southeast Sea from January 2015 to June 2023, with a time resolution of 1h, and clean and preprocess the data to create a labeled dataset.

[0056] (2) Acquire dual-polarized (VV+VH) single-look complex (SLC) SAR data collected by Sentinel-1 in Interferometric Wide swath (IW) mode.

[0057] (3) Perform orbit correction, radiometric calibration, pulse zone correction (Deburst), polarization matrix, multi-look, terrain correction, and land-sea masking on the SAR data (i.e., SAR images), and select the SAR data corresponding to the buoy data to form a dataset. Then divide the preprocessed SAR dataset into training set, test set and validation set according to 8:1:1.

[0058] (4) Obtain the ERA5 (ECMWF Reanalysis Version 5) atmospheric reanalysis product from the official website of the European Centre for Medium-Range Weather Forecasts (ECMWF). Select 10m wind speed and direction data with a time resolution of 1h and a spatial resolution of 0.25°×0.25°. Obtain precipitation data from the IMERG (Integrated Multi-satellite Retrievals for GPM) global precipitation measurement mission provided by NASA, with a time resolution of 0.5h and a spatial resolution of 0.1°×0.1°. Clean and preprocess the data, and construct an auxiliary dataset with the buoy data according to the rule of closest time and space, and filter the data with a time difference of less than 0.5h. Obtain wind speed and wind direction images from the OCN Level 2 product of Sentinel-1, and select data that corresponds to the spatiotemporal data of the processed SAR data as auxiliary data. Divide the auxiliary dataset into training set, test set and validation set according to 8:1:1.

[0059] (5) Construct and train a deep learning inversion model for SAR effective wave height that integrates multi-source data. Use the training set to train the model to obtain the sea surface effective wave height inversion model, select the test set to test the trained inversion model, and obtain the inverted effective wave height results.

[0060] In step (1), in order to improve the reliability and accuracy of the effective wave height data of the buoy, outliers, missing values ​​and noise data are removed, and the preprocessed buoy data is normalized.

[0061] Step (3) includes the following steps:

[0062] The specific SAR preprocessing steps are as follows: First, orbit correction. Single-look complex SAR data contains orbit-related parameters recorded during imaging. Accurate orbit files are typically calculated two weeks after image generation. By downloading orbit file values ​​from the ESA Sentinel-1 orbit file server, the parameter values ​​describing the orbit in the SAR data are replaced, reducing errors in subsequent processing. Next, radiometric calibration. The orbit-corrected SAR data is converted to normalized radar backscattering coefficients for radiometric calibration. Through radiometric calibration, the pixel values ​​of the original SAR image are converted into physically meaningful scattering coefficients to better characterize the sea surface scattering characteristics. Then, pulse band correction is performed. In the original SAR image, adjacent pulse bands have a small dark band, corresponding to the non-signal portion of continuously transmitted pulse signals. Pulse band correction removes the dark band portion with pulse bands, merging all effective signal portions of the pulse bands. Next, polarization matrix transformation is performed. SAR data from the VV and VH polarization channels are fused to generate a covariance matrix C, facilitating subsequent processing. Finally, multi-look processing is completed. SAR (Synthetic Radiation Array) is a coherent imaging system. Echo signals from multiple scatterers on the ground interfere with each other, resulting in speckle noise in calm sea surface images, which affects the inversion results. Multi-view averaging can reduce speckle noise in the image and compress the data volume to some extent, reducing subsequent data processing. Next, terrain correction is performed. SAR data can be geometrically distorted by terrain undulations. This step, by incorporating a digital elevation model (DEM), eliminates geometric distortion and radiometric distortion caused by surface topography, thus attributing the radar signal to the scattering characteristics of the ground objects themselves, further optimizing the inversion results. Finally, a land-sea mask is applied. Non-oceanic areas are excluded to ensure the model focuses on the sea surface data.

[0063] For the preprocessed single-look complex SAR data, the SAR backscattering intensity of the VV polarization and VH polarization channels is obtained after processing. and The incident angle and the RGB three channels are used to synthesize an RGB image. The synthesized image is then cropped into a 256×256 pixel sub-image centered on the buoy coordinates, and spatiotemporally matched with the preprocessed buoy data to obtain the model input dataset.

[0064] In step (4), ERA5 is the fifth-generation atmospheric reanalysis product launched by ECMWF, which contains various types of datasets. Single-layer hourly ERA5 data is applied, including 2-meter air temperature, 2-meter dew point temperature, total precipitation, mean sea level pressure, surface pressure, 10-meter u-wind component and 10-meter v-wind component, wind direction, water depth, boundary layer height, and sea surface wave height. The 10-meter mu-wind component and 10-meter v-wind component data of ERA5 are used and wind synthesis is performed to obtain the sea surface 10-meter wind speed and wind direction as reference values; IMERG precipitation data are obtained using a three-level multi-satellite precipitation algorithm, which combines intermittent precipitation estimates from all constellation microwave sensors, infrared observation data from geostationary satellites, and monthly precipitation data from rain gauges. The latest data version includes three product types: IMERG Early Run (IMERG-E, a real-time product with a delay of approximately 4 hours), IMERG Late Run (IMERG-L, a near real-time processing product with a delay of approximately 14 hours), and IMERG Final Run (IMERG-F, the final running product). Precipitation data from the IMERG Final Run (IMERG-F) product is used. The data is cleaned and preprocessed, and auxiliary datasets are constructed by combining the data with buoy data according to the closest temporal and spatial rules, filtering data with a time difference within 0.5 hours. Wind speed and wind direction imagery from Sentinel-1's OCN Level 2 product is obtained, and spatiotemporal matching is performed with the buoy data. The data is then cropped to obtain corresponding auxiliary data, which is then divided into training, testing, and validation sets in an 8:1:1 ratio.

[0065] Step (5) is as follows: The SAR effective wave height deep learning inversion model that integrates multi-source data consists of a convolutional feature extraction module, a feature fusion module, a feature relationship inference module, and a parameter optimization module. The convolutional feature extraction module consists of a 3D convolutional layer and an information enhanced attention module (IEAM); the feature fusion module includes a Transformer module based on interactive attention; the feature relationship inference module includes multi-source data combination, a bidirectional long short-term memory network (Bi-LSTM), and a fully connected layer. Among them, the convolutional feature extraction module embeds spatial information into the channel attention to extract the overall spatial and channel features of the preprocessed SAR data; the feature fusion module uses the interactive attention Transformer module to fuse SAR data and OCN data features, and then further learns the complex feature relationships in the fused data through a conventional Transformer coding block to obtain the final feature outputs; the feature relationship inference module fuses multi-source data through a Bi-LSTM network, infers the mutual relationships of multi-source features, and uses a fully connected layer to obtain the inverted effective wave height value; then the parameter optimization module optimizes the various parameters in the model to enhance the inversion capability of the model.

[0066] Convolutional Feature Extraction Module: In the convolutional feature extraction stage, 3D convolutional layers are used to extract multidimensional features from image data, and an Information Enhancement Attention (IEAM) module is introduced to emphasize inter-channel feature relationships, overcoming the limitations of conventional convolutional neural networks in capturing global information. First, 3D convolutional layers are used to extract spatial and channel-specific global features from preprocessed single-view complex SAR and OCN data. Then, a two-branch IEAM is introduced to embed the spatial information of the single-view complex SAR data into the channel attention. The IEAM module consists of two branches. Initially, the input is segmented into multiple sub-features along the channel dimension to capture different semantics while reducing computational load. One branch uses two one-dimensional global average pooling operations to encode channel information along two spatial directions, obtaining features in two different directions. Then, the two feature tensors are concatenated, and a 1×1 convolution is used to capture local cross-channel interactions. After decomposing the output of the 1×1 convolution into two vectors, two sigmoid activation functions are used to process the data, obtaining the channel attention weights. The two results are then multiplied and reweighted using a new input to mitigate overfitting. Group normalization is then applied to the reweighted result to alleviate gradient issues. After group normalization, a 2D global average pooling layer is used to capture spatial features from the multiplication result. In another branch, a 3×3 convolution is used to capture global features, and a 2D global average pooling layer is used to embed these global features into channel attention. The features from both branches are then combined and the original image data is introduced to further mitigate overfitting, ultimately yielding the output features.

[0067] Feature Fusion Module: In this stage, the single-view complex SAR feature data and OCN feature data obtained from the convolutional feature extraction module are input into the feature fusion module for feature fusion to overcome the limitations of conventional convolutional neural networks in multi-scale data fusion. This module has a two-branch structure, where each branch contains an image block encoding Embedding layer, an interactive attention Transformer module, and a regular Transformer module. First, the Embedding layer scales the received SAR and OCN data features to the same size and divides them into multiple small blocks according to a predetermined size. Then, a Flatten layer flattens these small blocks into one-dimensional vectors, adding positional encoding to the beginning of all vectors while preserving spatial information. The encoded SAR and OCN feature sequences are then input into the two sets of interactive attention Transformer modules respectively. Each interactive attention Transformer consists of an interactive attention multi-head self-attention layer and a multilayer perceptron (MLP) layer. In the multi-head self-attention layer, SAR and OCN feature sequences are mapped to query (Q), key (K), and sum (V) vectors to compute global dependencies between patches. Unlike typical multi-head self-attention layers, this module, when computing the query, key, and sum matrix, passes the keys and values ​​corresponding to the SAR feature sequences as input to the multi-head attention layer of the OCN data in another branch, and the keys and values ​​corresponding to the OCN feature sequences as input to the multi-head attention layer of the SAR data in another branch. This allows the attention module to generate mutually conditional attention features. Subsequently, a series of fully connected layers and activation functions in the MLP are used to capture the correlations between features in the obtained mutual attention features, resulting in fused features. After passing through the interactive attention Transformer, the data from the two branches are input into a regular Transformer module, which contains a multi-head attention layer and a multilayer perceptron layer. This module further learns the complex feature relationships in the fused data, yielding the final feature outputs.

[0068] Feature Relationship Inference Module: This module combines SAR features, OCN features, and other auxiliary data obtained from the feature fusion module and inputs them into a two-layer BiLSTM network. By processing the data simultaneously in both the forward and reverse directions of the input sequence, it can more comprehensively capture the features and contextual relationships of the sequence, better fusing multi-source features to obtain a fused feature sequence. Subsequently, the feature sequence is introduced into a fully connected layer, and dropout is used to alleviate overfitting, yielding the final inversion result.

[0069] Parameter optimization module (Sparrow Optimization Algorithm SCSSA): SCSSA is used to optimize parameters such as the number of neurons in the BiLSTM hidden layer, the number of patches in the Embedding layer, the number of self-attention heads in the Transformer, the initial learning rate, and the learning rate adjustment factor, thereby improving the model's inversion capability.

[0070] First, a refraction-based back-learning strategy is used to initialize a set of model parameters to construct the model. This strategy expands the search range by calculating the reverse solution of the current solution, thereby finding better alternative solutions. Second, a sine / cosine strategy is used to replace the discoverer position update formula in the original sparrow algorithm to generate new model parameters. When the model performance corresponding to the discoverer is locally optimal, a large number of followers will flood into that position, causing the discoverer and the entire group to stagnate, resulting in a loss of population position diversity. Therefore, the introduction of the sine / cosine strategy can effectively balance the population's global search and local exploitation capabilities. Finally, a Cauchy mutation strategy is used to replace the follower position update formula in the original sparrow algorithm, generating new model parameters based on the model parameters corresponding to the locally optimal discoverer. By using Cauchy mutation to perturb the sparrow positions, the search scale of the algorithm is expanded, thereby improving the ability to escape local optima.

[0071] One embodiment of the present invention, such as Figure 1 As shown, this embodiment includes the following steps:

[0072] Buoy data preprocessing: The acquired buoy data has a time resolution of 1 hour and includes significant wave height and wave magnitude. Outliers are removed by examining the maximum, minimum, and median values ​​in each data set. A gradient test is performed to ensure the continuity of oceanographic observation data within a certain time range; that is, changes in adjacent observation elements at similar times or locations should fall within a certain range. Otherwise, the data is considered abnormal. The test formula is as follows:

[0073] ;

[0074] Among them, the current observation value is ,and The previous correct value in the adjacent time period is , The gradient test parameter is set to 1m.

[0075] The wave class is divided into three levels: S is a minor wave (0~1.5m), M is a medium wave (1.5~2.5m), and L is a major wave (greater than 2.5m). The effective wave height is compared with the wave class, and outliers that do not match the wave class are filtered out.

[0076] The float data, after filtering out outliers, is normalized using the following formula:

[0077] ;

[0078] in, For normalized buoy data, This is the original buoy data. These represent the maximum and minimum values ​​of the buoy data, respectively.

[0079] The final preprocessed buoy surface effective wave height is used as the label dataset.

[0080] SAR Data Acquisition: We acquired dual-polarized single-look complex SAR data collected by Sentinel-1 in its interferometric wide-swath mode. The IW mode yielded the largest amount of SAR data and offered excellent resolution and swath width. Dual polarization (VV+VH) not only enhances sensitivity to sea surface scattering characteristics, making its analysis of complex sea conditions more accurate, but also improves the identification of sea surface roughness variations through the cross-polarization (VH) channel, thereby enhancing the robustness of the inversion. Single-look complex images, composed of the real and imaginary parts of the radar signal, retain complete amplitude and phase information, better reflecting the specific conditions of the sea surface.

[0081] SAR data preprocessing is extremely important for model inversion. Effective preprocessing can greatly suppress noise interference in SAR data, enhance the accuracy of SAR data, and improve inversion accuracy. For example... Figure 2 As shown, this embodiment uses SNAP software for preprocessing, and the specific preprocessing workflow is as follows: orbit correction, radiometric calibration, pulse zone correction (Deburst), polarization matrix, multi-look, terrain correction, and land-sea masking. SAR data corresponding to the buoy data are selected to construct the dataset. The specific steps are as follows:

[0082] Orbit correction: The metadata of single-look complex SAR images contains orbit-related parameters, which are recorded during imaging. Accurate orbit files are typically calculated only two weeks after image generation. By downloading orbit file values ​​from the ESA Sentinel-1 orbit file server, the parameter values ​​describing the orbit in the image metadata are replaced, reducing errors in subsequent processing.

[0083] Radiometric calibration: Absolute calibration establishes the relationship between single-look complex SAR data and the backscattering coefficients of ground targets. This converts the original complex or amplitude values ​​recorded for each pixel in the SAR data into standardized backscattering coefficients with clear physical meaning and directly related to ground features, thus better reproducing the roughness of the sea surface. The specific formula is as follows:

[0084] ;

[0085] in, Let be the radar backscattering coefficient, and i, j be the pixel coordinates in the image. The value is a grayscale value, which is the original complex number or amplitude value in the SAR data. These are calibration parameters.

[0086] Pulse band correction: In the original SAR data, adjacent pulse bands have a small dark band, which corresponds to the non-signal part of the continuously transmitted pulse signal. The pulse band correction operation can remove the dark band part of the image that contains pulse bands, so that all the effective signal parts of the pulse bands can be merged.

[0087] Polarization Matrix Generation: Single-look complex SAR data from VV and VH polarization channels are fused to generate a covariance matrix, facilitating subsequent processing. The formula for calculating the covariance matrix is ​​as follows:

[0088] ;

[0089] in, These represent SAR data matrices with VV and VH polarizations, respectively. ; ; ; E(.) represents the expected value. , These represent the SAR backscattering intensities of the VV-polarized and VH-polarized channels, respectively. and This is the intermediate parameter matrix.

[0090] Multi-view (after pulse band correction) and (Processing): SAR is a coherent imaging system. Echo signals from multiple scatterers of ground objects interfere with each other, resulting in speckle noise on calm sea surface images. This noise affects the inversion effect. Multi-look averaging can obtain approximately square pixels, reduce speckle noise in the image, and compress the data volume to some extent, reducing subsequent data volume. Range and azimuth views are important parameters in multi-look processing of SAR data. Range and azimuth views refer to the number of pixels processed in the range and azimuth directions. The number of views can be determined by calculating the pixel interval in the range and azimuth directions and combining it with the downsampling ratio during multi-look processing. The formula for calculating the ground resolution in the range direction is as follows:

[0091] ;

[0092] Where Ground Range Resolution is the ground resolution in the range direction; Range Pixel Spacing is the distance spacing between pixels in the image matrix in the range direction. , Matrix pixel distance spacing, The incident angle can be found in the single-look complex SAR data metadata file. The formulas for calculating the range and azimuth views are as follows:

[0093] ;

[0094] ;

[0095] in, , These represent the range and azimuth views, respectively; Desired Resolution is the set target resolution; Azimuth Resolution is the azimuth resolution, which can be found in the single-look complex SAR data metadata file. , The matrix performs multi-view processing on the corresponding number of views in the range and azimuth directions, that is, it averages each corresponding number of pixels in the range and azimuth directions into one pixel, thereby reducing noise while preserving the target signal.

[0096] Terrain Correction: SAR data may suffer from geometric distortion due to terrain undulations. By combining the Chinese digital elevation model with a resolution of 3 arcseconds from SRTM (Shuttle Radar Topography Mission), the true geographical location of each pixel in the SAR data is calculated using the three-dimensional information of the ground surface. This eliminates the geometric distortion and radiometric distortion of the image caused by the terrain undulations, thereby attributing the radar signal to the scattering characteristics of the ground objects themselves, ensuring accurate alignment between the image and the geographical location, and further optimizing the inversion effect.

[0097] Land-sea masking: The backscattering characteristics of SAR data differ significantly between land and ocean areas, which can introduce errors in wave height retrieval. This step excludes non-oceanic areas from the imagery, ensuring that the model focuses on the sea surface shadow data.

[0098] Cropping and Channel Synthesis: Each raw SAR data point covers one or more buoys. Centered on the buoy's latitude and longitude, a 256×256 image sub-block is cropped from the preprocessed SAR image, matched with the buoy observations, and then combined with the preprocessed data. , The incident angle data, along with the R, G, and B channels of the RGB image, were used to synthesize an RGB image using SNAP software, combining the three channels with 256×256 image sub-blocks. The significant wave height value in the buoy data is the indicative wave height, calculated by averaging the first third of wave heights within a certain range, arranged from largest to smallest; this is the most representative value in engineering. The 256×256 crop size provides sufficient resolution while covering a suitable area around the buoy point, ensuring a good spatial correspondence with the buoy data. , The matrix contains amplitude and phase information from single-view complex SAR data, while the incident angle data provides additional geometric information from the radar. By combining these three elements into an RGB image, the model can utilize all three types of information for analysis simultaneously.

[0099] We obtained the wind component data (u) at 10m in the u-direction and the wind component data (v) at 10m in the v-direction from the ERA5 atmospheric reanalysis product provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) website, with a temporal resolution of 1 hour and a spatial resolution of 0.25° × 0.25°. We then performed wind synthesis on these data to obtain reference values ​​for sea surface wind speed and direction at 10m. The specific wind synthesis formula is as follows:

[0100] ;

[0101] ;

[0102] D ;

[0103] Where atan2(u,v) is the arctangent value of the calculated component. is the scaling factor that converts radians to degrees, 90 is the angle correction that converts due east to due north, (-1) is the reversal of direction to conform to the usual meteorological conventions, Speed ​​represents wind speed, DirectionFromeast represents wind direction data based on due east, and Direction represents wind direction data based on due north, i.e., the final wind direction data.

[0104] We acquired IMERG global precipitation measurement data from NASA, with a temporal resolution of 0.5 h and a spatial resolution of 0.1° × 0.1°. An auxiliary dataset was constructed using buoy data based on the closest temporal and spatial proximity, filtering data with a time difference within 0.5 h. We also acquired wind speed and wind direction imagery from Sentinel-1's OCN Level 2 product. Since OCN data has already undergone a series of preprocessing steps, no further preprocessing is required. Using the spatiotemporal basis of the cropped single-look complex SAR data, we cropped the wind speed and wind direction spectra separately and generated 10×10 grayscale images to construct the auxiliary dataset.

[0105] A SAR significant wave height inversion model integrating multi-source data was constructed and trained. The model was trained using a training set to obtain sea surface significant wave height prediction data. The model parameters were optimized using a sparrow optimization algorithm. Finally, the model with the best accuracy in the validation set was selected for testing on the test set. Figure 3 As shown, the details are as follows:

[0106] Convolutional Feature Extraction Module: Two 3D convolutional layers are used to extract the spatial and channel-specific global features of the preprocessed single-view complex SAR and OCN data, respectively. Following this, a two-branch IEAM information enhancement attention module is introduced to embed the spatial information of the preprocessed single-view complex SAR data into the channel attention. The IEAM module is as follows: Figure 4 As shown, initially, the input is segmented into G sub-features along the channel dimension to capture different semantics while reducing computational load. These sub-features are represented as... ,in H is the image height; W is the image width; C is the number of image channels; and G is the number of sub-features. In one branch, two one-dimensional global average pooling operations are used to encode the channel information along two spatial directions, as shown below:

[0107] ;

[0108] ;

[0109] in, and Pay attention to information from different directions. It is obtained through one-dimensional global average pooling along the horizontal dimension (i.e., the transverse direction), effectively capturing positional information in the vertical dimension while maintaining accurate positional information in the horizontal direction. Similarly, it is obtained through one-dimensional global average pooling along the vertical dimension (i.e., the longitudinal direction). This process encodes positional information along the horizontal direction. Then, the two tensors are concatenated, and a 1×1 convolution is used to capture local cross-channel interactions. After decomposing the output of the 1×1 convolution into two vectors, two non-linear sigmoid functions are applied to each vector. Subsequently, the channel attention maps within each group are combined using simple multiplication, and G sub-features are introduced for reweighting to alleviate overfitting. The reweighted results are then group-normalized to mitigate gradient problems. Different cross-channel interaction features are created between the two parallel paths of the 1×1 branch. Then, two-dimensional global average pooling is performed in the 1×1 branch to capture spatial features.

[0110] ;

[0111] in, This represents the matrix before two-dimensional average pooling. This is the matrix after 2D average pooling. In this branch, spatial location information is used to model cross-channel dependencies. Furthermore, a single 3×3 kernel is used in the right-hand 3×3 branch to capture global features of the image data, and another 2D global average pooling operation is used to concentrate these global features into channel attention. Then, multiplication is used to fuse the feature information from the two branches, resulting in richer spatial and channel attention features. After that, a sigmoid activation function is applied to obtain the computed spatial attention features. Finally, this result is multiplied by the original input of the module to alleviate overfitting.

[0112] Feature fusion module: such as Figure 5 The diagram illustrates how SAR and OCN features obtained from the feature extraction module are input into an interactive attention Transformer for feature fusion. To better fuse single-look complex SAR and OCN features, an interactive attention multi-head attention layer (MHSA) is used instead of the conventional multi-head attention layer (MHSA) in the Transformer encoder. This module has a two-branch structure, where each branch contains an embedding layer, an interactive attention Transformer module, and a conventional Transformer module. First, the embedding layer scales the received SAR and OCN data features to the same size and divides them into multiple small blocks according to a predetermined size. Then, a flattening layer flattens these blocks into one-dimensional vectors, adding positional encoding at the beginning of all vectors to preserve the spatial information of the feature map. Subsequently, the encoded SAR and OCN feature sequences are input into the two sets of interactive attention Transformer modules, respectively. Figure 5 As shown, each interactive attention Transformer group contains an interactive attention multi-head self-attention layer and a multilayer perceptron (MLP) layer. In the multi-head self-attention layer, SAR and OCN feature sequences are mapped to query (Q), key (K), and value (V) vectors to compute global dependencies between patches. The multi-head self-attention mechanism is used to obtain the relationships between values ​​in time series data and to extract complete information from the time series, as shown in the following formula:

[0113] ;

[0114] ;

[0115] ;

[0116] in, This indicates the attention of the bulls. Indicates the first The attention of each head Indicates the first A query vector for each head. Indicates the first A key vector for each head. Indicates the first A vector of values ​​for each head. Representing dimension, This indicates the total number of heads. The final mapping weight matrix, Represents the join function. This represents the attention weight.

[0117] Unlike typical multi-head self-attention layers, interactive attention multi-head self-attention layers, when calculating the query, key, and value matrices, pass the keys and values ​​corresponding to the SAR feature sequences as input to the multi-head attention modules corresponding to the OCN feature sequences, and vice versa. This results in the multi-head attention modules generating mutually conditional attention features. The specific formula is as follows:

[0118] ;

[0119] ;

[0120] in, This is the output of the interactive attention multi-head self-attention module in the SAR branch. This is the output of the interactive attention multi-head self-attention module in the OCN branch. This indicates a high level of attention from multiple parties. For the query Q vector in the SAR branch, This is the query Q vector in the OCN branch.

[0121] Subsequently, an embedding layer is added to the image segmented from the data obtained from the multi-head attention module to alleviate overfitting, and layer normalization is performed. Then, a series of fully connected layers and activation functions in the MLP are used to capture the correlation between SAR and OCN data features in the obtained mutual attention features, and residual connections and layer normalization are performed again to obtain the feature-fused data.

[0122] After passing through the interactive attention Transformer, the features obtained from the two branches are input into a regular Transformer module, which contains a multi-head attention layer and a multi-layer perceptron layer. This module further learns the complex feature relationships in the fused data, and then outputs the fused features.

[0123] Feature Relationship Inference Module: SAR features, OCN features, and other auxiliary data obtained from the feature fusion module are input into a two-layer Bi-LSTM network. The Bi-LSTM network processes data simultaneously in both forward and reverse directions of the input sequence. With three two-layer Bi-LSTM networks, it can more comprehensively capture the features and contextual relationships of the sequence, better fusing multi-source features to obtain a fused feature sequence. Subsequently, the feature sequence is introduced into a fully connected layer to obtain the final effective wave height inversion result.

[0124] Parameter optimization module: The Sparrow Optimization Algorithm (SCSSA) is used to globally optimize parameters such as the number of neurons in the Bi-LSTM hidden layer, the number of patches in the Embedding layer, the number of self-attention heads in the Transformer, the initial learning rate, and the learning rate adjustment factor, thereby improving the model's inversion capability.

[0125] First, a refraction-based back-learning strategy is used to initialize individual sparrow algorithm instances, generating a set of model parameters to construct the models. This strategy expands the search range by calculating the reverse solution of the current solution, thereby finding better alternative solutions. Second, a sine / cosine strategy is used to replace the discoverer position update formula of the original sparrow algorithm, generating new model parameters based on the performance of the initial model. The specific formula is as follows:

[0126] ;

[0127] in, This represents the position information of the i-th sparrow in the (t+1)-th iteration. This represents the position information of the i-th sparrow in the t-th iteration. It is a random number. This is the current optimal solution.

[0128] When the food sought by the discoverer is in a locally optimal location, a large number of followers will flood into that location. At this point, the discoverer and the entire population stagnate, resulting in a loss of population location diversity. Therefore, the introduction of a sine / cosine strategy can effectively balance the population's global search and local exploitation capabilities.

[0129] Finally, during group foraging, when individual followers search for food around the optimal forager, they may voluntarily transform into new foragers due to resource competition. To avoid prematurely getting trapped in local optima, a Cauchy mutation strategy is introduced to enhance the overall global optimization capability. The Cauchy mutation strategy replaces the follower position update formula in the original sparrow algorithm. The specific formula is as follows:

[0130] ;

[0131] in: This refers to the j-th parameter information of the i-th sparrow in the (t+1)-th iteration; Let be the optimal solution for the parameters of the i-th sparrow in the t-th iteration. The standard Cauchy distribution function is used. Cauchy mutation is used to perturb the sparrow's position, expanding the algorithm's search scale and thus improving its ability to escape local optima.

[0132] A total of 10 iterations were performed, and the optimal solution was selected as the final model parameters.

[0133] The effective wave height inversion model is trained using a training set. A test set is then used to test the trained inversion model, and the prediction results are inversely normalized to obtain the predicted effective wave height value. The inverse normalization formula is as follows:

[0134]

[0135] in, This represents the predicted data after inverse normalization. For the maximum value of the data, The minimum value of the data. For model prediction data.

[0136] The obtained significant wave height prediction values ​​are compared with the significant wave height labels of the buoys to obtain an evaluation of the inversion accuracy. The specific evaluation indicators are as follows: three indicators are selected: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Correlation Coefficient (CC). The formulas for each indicator are as follows:

[0137] ; ;

[0138] ;

[0139] in, These are model predictions. It is the effective wave height value observed by the buoy. It is the average value of buoy observations. is the predicted average, and n is the number of samples.

[0140] To further analyze the inversion capability of the SAR significant wave height inversion model that integrates multi-source data, we compared our model with the CNN-BiLSTM-Attention (Convolutional Neural Network - Bidirectional Long Short-Term Memory - Attention Mechanism) and ViT (Vision Transformer) models. A dataset constructed from SAR images of the Southeast China Sea from 2015 to 2023 was used for inversion testing, and the results are shown in Table 1. Experimental results demonstrate that the deep learning inversion method for SAR significant wave height inversion, which integrates multi-source data, achieves the lowest RMSE and MAE, and the highest correlation coefficient (CC), fully showcasing its excellent significant wave height inversion performance.

[0141] Table 1. Experimental results of different models on the test set.

[0142]

[0143] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for deep learning-based inversion of SAR sea surface effective wave height by fusing multi-source data, characterized in that, Specifically: Preprocess the buoy data to obtain the significant wave height data of the ocean waves observed by the buoy; Acquire dual-polarized single-look complex SAR data collected in interferometric wide-swath mode; Preprocess the SAR data, using the effective wave height data as the label for the SAR data; construct training, testing, and validation sets. Acquire auxiliary data corresponding to SAR data, including wind speed, wind direction, precipitation, and OCN data; construct corresponding training, testing, and validation sets. A deep learning inversion model for SAR effective wave height fusion based on multi-source data was constructed. The model was trained using SAR data training set and auxiliary data training set. The trained model was then tested and validated using test set and validation set to obtain the final effective wave height inversion result.

2. The SAR sea surface effective wave height deep learning inversion method based on multi-source data as described in claim 1, characterized in that, The buoy data preprocessing is as follows: First, check the maximum, minimum, and median values ​​in each data set and remove outliers. Then, if the change in observed data between adjacent times or locations exceeds a preset range, it is considered an abnormal data set and removed. Wave levels are divided into three levels, with first, second, and third wave height intervals set. Waves in the first wave height interval are considered small waves, those in the second wave height interval are considered medium waves, and those in the third wave height interval are considered large waves. If the observed wave level in the buoy data does not match the actual wave height, it is considered an abnormal data set and removed. Normalize the buoy data after removing outliers.

3. The SAR sea surface effective wave height deep learning inversion method based on multi-source data as described in claim 1, characterized in that, Preprocessing of SAR data includes: orbit correction, radiometric calibration, pulse band correction, polarization matrix, multi-look, terrain correction, land and sea masking, as well as cropping and channel synthesis; The orbit correction specifically involves downloading precise orbit parameter values ​​from the orbit file server and replacing the orbit parameter values ​​that are included in the SAR data. The radiometric calibration specifically involves establishing the relationship between SAR data and the backscattering coefficients of ground targets through absolute calibration, using the following formula: ； in, Let i be the radar backscattering coefficient, and i and j be the pixel coordinates in the SAR data, respectively. Grayscale value These are calibration parameters; The polarization matrix is ​​generated specifically by using the following formula to generate the covariance matrix from the SAR data of the VV and VH polarization channels. : ; in, , , , E(.) represents the expected value. , These represent SAR data matrices with VV and VH polarizations, respectively. , These represent the SAR backscattering intensities of the VV-polarized and VH-polarized channels, respectively. and This is the intermediate parameter matrix; Terrain correction specifically involves using the digital elevation model of the Space Shuttle Radar Terrain Survey Mission to calculate the true geographical location of each pixel in the SAR data based on the three-dimensional information of the Earth's surface. This eliminates geometric distortion and radiometric distortion caused by the undulations of the Earth's surface, ensuring that the SAR data is accurately aligned with the geographical location. The cropping and channel synthesis process involves cropping a 256×256 sub-block from the SAR data, centered on the buoy's latitude and longitude. Each sub-block corresponds to only one preprocessed buoy data point. Based on this sub-block, the matrix is ​​then synthesized. and The incident angle data are used as the R, G, and B channels of the RGB image, respectively; and these three channels are merged into an RGB image to obtain the final preprocessed SAR data.

4. The SAR sea surface effective wave height deep learning inversion method based on multi-source data as described in claim 3, characterized in that, Acquire precipitation data with a temporal resolution of 0.5 h, a spatial resolution of 0.1° × 0.1°, and a time difference of less than 0.5 h with buoy data; The wind speed and wind direction spectra from the OCN Level 2 product were obtained. Using the spatiotemporal data of the cropped SAR data as a reference, the wind speed and wind direction spectra were cropped and generated as 10×10 grayscale images. The wind component u in the u direction and the wind component v in the v direction at 10m were obtained, with a temporal resolution of 1 hour and a spatial resolution of 0.25°×0.25°. The u and v components were then combined to obtain the wind speed and direction data at 10m above sea level. The specific wind combination formula is as follows: ； in, The wind direction data is based on due north, which is the final synthesized wind direction data. (-1) represents the reverse direction. Represents wind direction data based on due east: .

5. The SAR sea surface effective wave height deep learning inversion method according to claim 1, characterized in that, The SAR effective wave height deep learning inversion model that integrates multi-source data inputs preprocessed SAR data into a first 3D convolutional layer. A dual-branch information enhancement attention module is introduced into the first 3D convolutional layer. OCN data is input into a second 3D convolutional layer. The outputs of the dual-branch information enhancement attention module and the second 3D convolutional layer are input into an interactive attention Transformer for feature fusion. The output of the interactive attention Transformer is input into a feature relation inference module. Auxiliary data is then input into the feature relation inference module after passing through a fully connected layer. The feature relation inference module outputs the final effective wave height inversion result.

6. The SAR sea surface effective wave height deep learning inversion method based on multi-source data as described in claim 5, characterized in that, The dual-branch information enhancement attention module divides the input into G sub-features along the channel dimension, and performs horizontal global average pooling and vertical global average pooling on these G sub-features respectively. ； ； in, This is the tensor obtained by horizontal global average pooling. Here, H is the input height and W is the input width, obtained by vertical global average pooling. Let g be the g-th sub-feature; g = 0, 1, ..., G-1; Will and The input is concatenated and fed into a 1×1 convolution. The output of the 1×1 convolution is decomposed into two vectors, which are then fed into two non-linear sigmoid functions. The outputs of the two non-linear sigmoid functions are reweighted with G sub-features to alleviate overfitting. The reweighted result is then subjected to group normalization and global average pooling. The G sub-features are fed into a 3×3 convolution, and the result of the 3×3 convolution is subjected to global average pooling. The results of the two global average pooling operations are multiplied, and the result of the multiplication is subjected to softmax normalization. The result of the 3×3 convolution and the result of the softmax normalization are then multiplied together. Finally, the input is passed through a sigmoid activation function, and the output of the sigmoid activation function is reweighted with the G sub-features. The final reweighted result is then multiplied with the original input.

7. The SAR sea surface effective wave height deep learning inversion method based on multi-source data as described in claim 5, characterized in that, The interactive attention Transformer has a two-branch structure, including the SAR branch and the OCN branch. Each branch contains an embedding layer, an interactive attention Transformer module, and a regular Transformer module. The embedding layer scales the input and divides it into multiple small blocks according to a predetermined size. The flattening layer flattens these small blocks into one-dimensional vectors and adds position encoding to the head of all vectors. The two embedding layers scale the two inputs to the same size; the interactive attention Transformer module includes an interactive attention multi-head self-attention module and a multilayer perceptron, the interactive attention multi-head self-attention module maps the output of the embedding layers to query Q, key K, and value V vectors; the key K vector in the SAR branch Sum of vectors V The key K vector in the OCN branch serves as the input to the interactive attention Transformer module of the OCN branch. Sum of vectors V As input to the interactive attention Transformer module in the SAR branch, specifically: ； ； in, This is the output of the interactive attention multi-head self-attention module in the SAR branch. This is the output of the interactive attention multi-head self-attention module in the OCN branch. Indicates multi-headed attention; For the query Q vector in the SAR branch, This refers to the query Q-vector in the OCN branch; The output of the embedding layer and the output of the interactive attention multi-head self-attention module are multiplied, normalized by the layer, and then fed into the multilayer perceptron. The output of the multilayer perceptron is multiplied with the normalized result, normalized again, and then fed into the regular Transformer module.

8. The SAR sea surface effective wave height deep learning inversion method according to claim 7, characterized in that, The feature relationship reasoning module includes a first and third two-layer bidirectional long short-term memory network. The output of the SAR branch of the interactive attention Transformer is input to the first two-layer bidirectional long short-term memory network. The output of the branch is input to the second two-layer bidirectional long short-term memory network, and the auxiliary data is input to the third two-layer bidirectional long short-term memory network through the fully connected layer. The two bidirectional long short-term memory networks process the data simultaneously in both the forward and reverse directions of the input sequence. The output of the feature relationship inference module is then passed through the fully connected layer to obtain the final effective wave height inversion result.

9. The SAR sea surface effective wave height deep learning inversion method according to claim 8, characterized in that, The Sparrow Optimization Algorithm is used to globally optimize the number of neurons in the hidden layer, the number of segments in the embedding layer, the number of self-attention heads in the interactive attention multi-head self-attention module, the initial learning rate, and the learning rate adjustment factor in the bidirectional long short-term memory network.

10. The SAR sea surface effective wave height deep learning inversion method according to claim 9, characterized in that, First, a refraction-based reverse learning strategy is used to initialize individual sparrows in the optimization algorithm, generating a set of model parameters and constructing the model. Second, a sine and cosine strategy is used to update the position of the discoverer. ； in, This represents the position information of the i-th sparrow in the (t+1)-th iteration. This represents the position information of the i-th sparrow in the t-th iteration. It is a random number. This is the current optimal solution; Finally, the Cauchy mutation strategy was used to calculate the follower positions during group foraging: ； in, This represents the j-th parameter information of the i-th sparrow in the (t+1)-th iteration; Let be the optimal solution for the parameters of the i-th sparrow in the t-th iteration. This is the standard Cauchy distribution function.

Images