A sar sea surface high wind speed inversion method fusing double attention mechanisms

Abstract

The present application relates to the technical field of marine remote sensing, in particular to a SAR sea surface high wind speed inversion method fusing a double attention mechanism: SAR remote sensing image data under a typhoon is acquired; the SAR remote sensing image data is preprocessed to obtain a plurality of SAR slice data; the plurality of SAR slice data is input into a trained SAR sea surface high wind speed inversion model, the SAR sea surface high wind speed inversion model extracts features from the plurality of SAR slice data through a weight file; the extracted features are subjected to channel compression regression to obtain an inversion wind speed result. The present application is based on the complex texture features and high-dimensional frequency domain features of the radar backscattering intensity in SAR data, constructs a complex nonlinear mapping model weight on the mechanism of nonlinear response of sea surface wind speed, and realizes sea surface wind speed inversion under a typhoon. The present application significantly improves the problem of sea surface wind speed inversion distortion caused by SAR signal saturation under a typhoon, and greatly improves the inversion accuracy.

Description

Technical Field

[0001] This invention relates to the field of marine remote sensing technology, and in particular to a SAR method for inverting high wind speeds on the sea surface that integrates a dual attention mechanism. Background Technology

[0002] Typhoons are intense tropical cyclones with immense destructive power, often accompanied by strong winds, torrential rains, and storm surges, causing severe damage to coastal areas. Therefore, timely and accurate monitoring and forecasting of typhoon tracks, intensity, and wind speeds are crucial for mitigating disaster impacts and protecting people's lives and property. Wind speed is a key indicator of typhoon intensity and hazard level; accurate wind speed retrieval helps meteorologists assess typhoon intensity, predict its development and track, thus providing fundamental data support for disaster prevention and mitigation.

[0003] SAR data boasts high precision and all-weather, all-time imaging capabilities, enabling continuous data acquisition and real-time monitoring even under severe weather conditions during typhoons. Wind speed retrieval under typhoon conditions based on SAR data not only provides a new technical means for typhoon monitoring and assessment but also holds significant practical importance for safeguarding people's lives and property and promoting sustainable socio-economic development.

[0004] Existing methods, such as the CMOD series, which rely on empirical models based on backscattering coefficients, radar incident angles, and wind speeds, are not suitable for sea surface wind speed inversion in high-wind-speed ranges. This is often due to the gradual saturation of radar backscattered signals at high wind speeds. While the saturation problem persists after using cross-polarized data, it is significantly improved compared to data with the same polarization. In recent years, machine learning methods have become increasingly common in SAR wind speed extraction, but they all heavily rely on external multi-source data input. Further discussion is needed regarding the exploration of the complex high-dimensional relationship between backscattering coefficients and wind speeds, resulting in low model inversion accuracy and insufficient generalization ability. Therefore, it is necessary to improve current deep learning-based methods for typhoon-related sea surface wind speed inversion to address the issues of high dependence on multi-source external data and low accuracy in sea surface wind speed inversion. Summary of the Invention

[0005] This invention provides a SAR high wind speed inversion method for sea surface based on a dual attention mechanism to solve the above-mentioned problems.

[0006] This invention provides a SAR high wind speed inversion method for sea surface based on a dual attention mechanism, comprising the following steps: Step 1: Acquire SAR remote sensing image data under typhoon conditions; Step 2: Preprocess the SAR remote sensing image data to obtain several SAR slice data; Step 3: Input several SAR slice data into the trained SAR high wind speed inversion model. The SAR high wind speed inversion model extracts features from several SAR slice data through a weight file. The extracted features are then subjected to channel compression regression to obtain the inversion wind speed results.

[0007] As a preferred technical solution, the SAR remote sensing image data preprocessing process in step two is as follows: First, the SAR remote sensing image data is subjected to radiometric calibration, speckle noise removal, and geographic correction in sequence. Then, the processed SAR remote sensing image data is spatiotemporally matched with the corresponding typhoon event SFMR data; Finally, the spatiotemporally matched data is sliced ​​to obtain several SAR slice data.

[0008] As a preferred technical solution, the construction process of the SAR sea surface high wind speed inversion model in step three is as follows: Step 31: Perform dB conversion on the SAR slice data, resample the data, and augment the data; Step 32: Calculate the polarization ratio data for the data-enhanced SAR slice data, and concatenate and stitch the SAR slice data and polarization ratio data in the channel dimension; Polarization ratio calculation:

[0009] in, SAR slice data after resampling under the VH polarization channel; SAR slice data after resampling under the VH polarization channel; Step 33: After cascading and stitching, the SAR slice data is used to invert the wind speed using the SAR sea surface wind speed model; The cascaded and stitched SAR slice data is denoted as... The wind speed was finally obtained by inversion as follows: ;

[0010] in, Model predicts wind speed; Input: SAR slice data after cascaded stitching; : Bi-branch manifold mapping operator; Prior estimation of intermediate wind speed; Dynamic gain operator; and Adama product and nonlinear gating; Eventual regression operator; Attention mechanisms; Step 34: Compare the inverted wind speed prediction values ​​with the training data. The difference from the true value is used to construct a SAR high wind speed inversion model for sea surface based on a SAR high wind speed inversion model framework that integrates dual attention mechanism, by iterating the mapping weight between SAR slice data and target wind speed.

[0011] As a preferred technical solution, in step 31, dB conversion: converting the values ​​of each point in the SAR slice data into dB format:

[0012] in, SAR slice data after conversion to dB format; SAR slice data; Data resampling: Resample each point of the dB SAR image data to the interval [0, 225].

[0013] in, This is SAR slice data after resampling; : Maximum value of SAR slice data in dB format; : minimum value of SAR slice data in dB format; SAR slice data in dB format; Data augmentation: Perform random horizontal flipping, vertical flipping, and 15° random rotation on the SAR slice data.

[0014] As a preferred technical solution, in step 34, the weight iterative optimization is achieved by the loss function formulas (3-2) to (3-4):

[0015]

[0016]

[0017] Where TotalLoss: the total loss function; : The main loss function, calculated according to MSELoss, is the loss function for the processing of features after data augmentation; : Auxiliary loss function, calculated according to SmoothL1Loss, is the loss function for the processing part of the predicted intermediate wind speed characteristics; N: Total number; : Represents the actual value and the predicted value, respectively; : Represents the actual value and the predicted value, respectively.

[0018] As a preferred technical solution, in step 34, the SAR sea surface high wind speed inversion model framework integrating dual attention mechanism includes a frequency domain attention module, a dual-branch ConvNeXT backbone network, and a wind speed guided attention module. The frequency domain attention module maps SAR slice data from the spatial domain X to the frequency domain K through a two-dimensional discrete Fourier transform, performs adaptive filtering in the complex space, and then restores the data.

[0019] in, Input SAR remote sensing image intensity: Spatial frequency coordinates; and Complex filter kernel weights; Adama accumulation; The dual-branch ConvNeXT backbone network performs nonlinear fusion of the restored frequency domain features and spatial domain features through two independent network architectures:

[0020] in, Features after fusion; Spatial domain branch network processes and outputs spatial domain features; Frequency domain branch network output frequency domain characteristics; : Fusion ratio coefficient, default value is 0.8; The wind speed-guided attention module includes feature prediction and feature modulation for dynamically scaling the gain of feature channels. Through dimensionality reduction mapping Extract core features and calculate an intermediate scalar wind speed prediction. The predicted value of the intermediate wind speed scalar During training, the model is supervised by real wind speeds, representing its initial physical assessment of the predicted wind speed in the current area.

[0021] in, and Intermediate prediction weights and biases; Features after fusion; (·): Activation function; Then, an attention mechanism is introduced into this module:

[0022] Where A: feature mask vector, which records the features that should be amplified and the features that should be suppressed during regression; : Dimensional upscaling; Features after fusion; Nonlinear gating; : Predicted scalar value of intermediate wind speed.

[0023] As a preferred technical solution, the training of the SAR high wind speed inversion model is as follows: repeat steps 33 and 34 to iteratively train the SAR high wind speed inversion model, continuously correct the model parameters using the gradient descent algorithm until the loss function reaches the order of 1e-3 on the validation set and remains stable, then stop training and obtain the trained SAR high wind speed inversion model and obtain the weight file.

[0024] As a preferred technical solution, the hyperparameter settings for the SAR sea surface high wind speed inversion model are as follows: the AdamW optimizer is used, the initial learning rate is set to 1e-5, the minimum learning rate is set to 1e-7, the learning rate scheduler is CosineAnnealingLR, the weight decay is 1e-6, and the learning patience rounds are set to 50, that is, training stops when the validation loss and validation RMSE have not decreased for 50 consecutive rounds.

[0025] As a preferred technical solution, spatiotemporal matching: the matching time window for SFMR data of typhoon events is determined by 1 hour before and after the imaging time of SAR remote sensing image data, and the effective sample points of SFMR data of typhoon events within the imaging range of SAR remote sensing image data are obtained by matching latitude and longitude.

[0026] As a preferred technical solution, the spatiotemporally matched data is sliced: based on the location of the valid sample points of the typhoon event SFMR data, corresponding slices are made at 224x224 pixels with the points as the center, and SAR sub-images are cropped out. The sub-images are labeled in the CSV file according to the file name; then the corresponding wind speed values ​​are read from the typhoon event SFMR data according to the sws field and recorded.

[0027] This invention provides a SAR sea surface high wind speed inversion method that integrates a dual attention mechanism. Based on the complex texture features and high-dimensional frequency domain features of radar backscattering intensity (roughness) in SAR data, a complex nonlinear mapping model weight is constructed on the mechanism of nonlinear response of sea surface wind speed, realizing sea surface wind speed inversion under typhoon conditions. This significantly improves the problem of sea surface wind speed inversion distortion caused by SAR signal saturation under typhoon conditions and greatly enhances the inversion accuracy. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of the overall architecture of the SAR high wind speed inversion method for sea surface fusion with dual attention mechanism of the present invention; Figure 2 This is a schematic diagram of the frequency domain attention module architecture; Figure 3 This is a schematic diagram of the architecture of the dual-branch ConvNeXT backbone network; Figure 4 A schematic diagram of the architecture of the wind speed-guided attention module; Figure 5 This is a schematic diagram illustrating the convergence of the training loss. Figure 6 The scatter plot and error distribution diagram of the RMSE index of the SAR sea surface high wind speed inversion model on the test set are shown. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0032] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0033] See Figure 1 This invention provides a SAR high wind speed inversion method for sea surface based on a dual attention mechanism, comprising the following steps: Step 1: Acquire SAR remote sensing image data under typhoon conditions; Step 2: Preprocess the SAR remote sensing image data to obtain several SAR slice data; Step 3: Input several SAR slice data into the trained SAR high wind speed inversion model. The SAR high wind speed inversion model extracts features from several SAR slice data through a weight file. The extracted features are then subjected to channel compression regression to obtain the inversion wind speed results.

[0034] In step two, the preprocessing of SAR remote sensing image data is as follows: First, the SAR remote sensing image data is subjected to radiometric calibration, speckle noise removal, and geographic correction in sequence.

[0035] Then, the processed SAR remote sensing image data is spatiotemporally matched with the corresponding typhoon event SFMR data.

[0036] The matching time window for SFMR data of typhoon events is determined by 1 hour before and after the imaging time of SAR remote sensing image data. The effective sample points of SFMR data of typhoon events within the imaging range of SAR remote sensing image data are obtained by matching latitude and longitude.

[0037] Finally, the spatiotemporally matched data is sliced ​​to obtain several SAR slice data.

[0038] Based on the location of the valid sample points in the typhoon event SFMR data, corresponding slices of 224x224 pixels are made with these points as the center, and SAR sub-images are cropped out. The sub-images are labeled with file names in the CSV file. Then, the corresponding wind speed values ​​are read from the typhoon event SFMR data by the sws field and recorded.

[0039] In step three, the construction and training process of the SAR sea surface high wind speed inversion model is as follows: Step 31: Perform dB conversion on the SAR slice data, resample the data, and augment the data; dB conversion: Converting the values ​​of each point in the SAR slice data to dB (decibels):

[0040] in, SAR slice data after conversion to dB format; SAR slice data; Data resampling: Resample each point of the dB SAR image data to the interval [0, 225].

[0041] in, This is SAR slice data after resampling; : Maximum value of SAR slice data in dB format; : minimum value of dB SAR slice data; SAR slice data in dB format; Data augmentation: Perform random horizontal flipping, vertical flipping, and 15° random rotation on the SAR slice data; Step 32: Calculate the polarization ratio data for the data-enhanced SAR slice data, and concatenate and stitch the SAR slice data (VV VH polarization) and the polarization ratio data along the channel dimension; Polarization ratio calculation:

[0042] in, SAR slice data after resampling under the VH polarization channel; SAR slice data after resampling under the VH polarization channel; Step 33: After cascading and stitching, the SAR slice data is used to invert the wind speed using the SAR sea surface wind speed model; The cascaded and stitched SAR slice data is denoted as... The wind speed was finally obtained by inversion as follows: ;

[0043] in, Model predicts wind speed; Input: SAR slice data after cascaded stitching; : Bi-branch manifold mapping operator; Prior estimation of intermediate wind speed; Dynamic gain operator; and Adama product and nonlinear gating; Eventual regression operator; Attention mechanisms; Step 34: Compare the inverted wind speed prediction values ​​with the training data. The difference from the true value is used to construct a SAR high wind speed inversion model based on a SAR sea surface high wind speed inversion model framework that integrates dual attention mechanism, by iterating the mapping weight between SAR slice data and target wind speed. Its weight iterative optimization is achieved by the loss function formulas (3-2) to (3-4):

[0044]

[0045]

[0046] Where TotalLoss is the total loss function; : The main loss function, calculated according to MSELoss, is the loss function for the processing of features after data augmentation; : Auxiliary loss function, calculated according to SmoothL1Loss, is the loss function for the processing part of the predicted intermediate wind speed characteristics; N: Total number; : Represents the actual value and the predicted value, respectively; : Represents the actual value and the predicted value, respectively.

[0047] In step 34, the SAR sea surface high wind speed inversion model framework integrating dual attention mechanism includes a frequency domain attention module, a dual-branch ConvNeXT backbone network, and a wind speed guided attention module. In the model inference and inversion stage, the model of this invention extracts high-dimensional frequency domain features from the input SAR remote sensing image data through a frequency domain attention module. Its core task is to achieve noise suppression and feature highlighting at the signal source. In the SAR imaging mechanism, speckle noise is the main factor interfering with wind speed inversion, manifesting as random fluctuations in the high-frequency domain. This module transforms the image to the frequency domain through two-dimensional discrete Fourier transform (2D-FFT) and introduces a learnable complex filter.

[0048]

[0049] By adaptively controlling the gain of specific frequency components in the frequency domain, the model can act like a physical filter, suppressing noise while significantly enhancing the effective signal components related to Bragg scattering. This provides an initial flow field distribution with higher physical purity and more significant spectral features for subsequent feature extraction, ensuring a scientific benchmark for the inversion from the data source.

[0050] After obtaining the purified signal, the dual-branch ConvNeXt backbone network achieves a multi-dimensional and comprehensive representation of the complex state of the sea surface through parallel mapping in the spatial and frequency domains. The spatial branch focuses on capturing intuitive geometric textures such as wind streaks and boundary layer roll clouds, which directly reflect the dynamic structure of the near-surface atmosphere; while the frequency branch extracts the energy distribution patterns caused by changes in sea surface roughness through deep analysis of the preprocessed spectrum. These two branches are fused together using learnable weights. Nonlinear weighting is applied. This collaborative mode ensures that the model can simultaneously explain large-scale spatial morphology and microscopic scattering energy, thus solving the problem of insufficient expressive power of a single feature domain under different sea conditions.

[0051] Finally, the wind speed-guided attention module performs high-dimensional feature enhancement on the output of the dual-branch network. This module introduces a feedback adjustment mechanism similar to that in adaptive control theory. Before the feature flow is directed to the final regression head, this module first extracts an intermediate wind speed prediction value representing a "preliminary understanding of sea conditions" through intermediate mapping. Then, the model uses this intermediate cognition to construct a dynamic gain factor. This information is then incorporated into the attention mask generation process. This module simulates the nonlinear response of sea surface scattering to wind speed: when a high wind speed range is initially identified, the dynamic gain forcibly amplifies the polarization channel characteristics sensitive to high wind speeds and automatically suppresses invalid signals that tend to saturate under extreme conditions. Thus, through the sequential collaborative processing of these three modules, SAR sea surface high wind speed inversion under typhoon scenarios is achieved.

[0052] The frequency domain attention module has the following structure: Figure 2 As shown, SAR slice data is mapped from the spatial domain X to the frequency domain K through two-dimensional discrete Fourier transform (2D-DFT), and then restored after adaptive filtering in the complex space:

[0053] in, Input SAR remote sensing image intensity: Spatial frequency coordinates; and Complex filter kernel weights; Adama accumulation; The dual-branch ConvNeXT backbone network has the following backbone structure: Figure 3 As shown, the overall approach employs two independent network architectures to nonlinearly fuse the restored frequency domain features and spatial domain features (cascaded and stitched SAR slice data):

[0054] in, Features after fusion; Spatial domain branch network processes and outputs spatial domain features; Frequency domain branch network output frequency domain characteristics; : Fusion ratio coefficient, default value is 0.8; Wind speed-guided attention module, structure as follows Figure 4 As shown, it consists of two parts: feature prediction (regression) and feature modulation (attention). The attention module is guided by wind speed to dynamically scale the gain of the feature channels. Through dimensionality reduction mapping Extract core features and calculate an intermediate scalar wind speed prediction. The predicted value of the intermediate wind speed scalar During training, the model is supervised by real wind speeds, representing its initial physical assessment of the predicted wind speed in the current area.

[0055] in, and Intermediate prediction weights and biases; Features after fusion; (·): Activation function; Then, an attention mechanism is introduced into this module:

[0056] Where A: Feature mask vector, which records the features (useful signals) that should be amplified and the features (interference or saturated signals) that should be suppressed during regression. : Dimensional upscaling; Features after fusion; Nonlinear gating; : Predicted scalar value of intermediate wind speed.

[0057] Training of the SAR high wind speed inversion model for sea surface: The dataset used in this invention is divided into training set and validation set in an 8:2 ratio.

[0058] The dataset includes SAR remote sensing image data and corresponding typhoon event SFMR data as experimental data.

[0059] The SAR remote sensing image data is Sentinel-1 GRDH-class VV VH polarized SAR remote sensing image data.

[0060] The Sentinel-1 GRDH-class VV VH polarized SAR remote sensing imagery data is in interferometric-wide imaging mode, with a spatial resolution of 10m and a radian width of 250km. It can sensitively capture minute, fragmented waves and roughness variations on the ocean surface. This data originates from the European Space Agency's Copernicus program's C-band SAR Earth observation satellite Sentinel-1, a two-satellite system providing all-weather, continuous day-and-night surface radar imagery.

[0061] SFMR data comes from NOAA (National Oceanic and Atmospheric Administration) SFMR (Stepped Frequency Microwave Radiometer) data, primarily sourced from NOAA's Hurricane Reconnaissance Program. It is typically carried by NOAA's WP-3D Orion and Gulfstream-IV hurricane reconnaissance aircraft for observing the air-sea environment within tropical cyclones. SFMR is a one-dimensional continuous observation along the aircraft's flight path, with an effective spatial resolution typically around 1–3 km and a temporal resolution generally on the order of 1 second.

[0062] This invention selects a specific sea area as the core test field for algorithm verification. The time span is from 2016 to 2020, covering a total of 16 typhoon events, specifically geographically covering latitude 36.0°N~4.78°N and longitude 57.4°W~103.8°W. This sea area has a relatively wide coverage and experiences numerous typhoons during the hurricane season. Experiments in this sea area with frequent typhoon events can effectively verify whether the model can achieve high-precision SAR inversion of high wind speeds at sea surface, and it is also an excellent experimental scenario for verifying the model's stability and generalization ability.

[0063] Repeat steps 33 and 34 to iteratively train the SAR high-wind-speed sea surface inversion model, continuously correcting the model parameters using the gradient descent algorithm until the loss function reaches a value on the validation set at the order of 1e-3 and remains stable. Figure 5 As shown, training is stopped and a trained SAR high wind speed inversion model for the sea surface is obtained, resulting in a weight file.

[0064] Hyperparameter settings for the SAR high wind speed inversion model: The AdamW optimizer is used, with an initial learning rate of 1e-5, a minimum learning rate of 1e-7, a learning rate scheduler of CosineAnnealingLR, a weight decay of 1e-6, and a learning patience epoch of 50, meaning that training stops when the validation loss and validation RMSE do not decrease for 50 consecutive epochs.

[0065] Through this training process, the SAR high wind speed inversion model establishes a nonlinear mapping relationship between the texture features, high-dimensional frequency domain features and sea surface wind speed of Sentinel-1 GRDH-level VV VH polarization SAR remote sensing image data, thereby realizing the inversion of high wind speed on the sea surface based on SAR data.

[0066] This invention uses the Gulf of America, the Caribbean Sea, and the Sargasso Sea as experimental scenarios and acquires Sentinel-1 GRDH-class VV VH polarimetric SAR remote sensing image data and corresponding SFMR data of typhoon events.

[0067] To comprehensively evaluate the robustness and generalization ability of the model, this invention designed two ablation comparison experiments: This experiment intends to use the following core indicators to analyze the deviation between the reconstruction results and the true values: RMSE (Root Mean Square Error): the standard deviation of the error between the predicted wind speed value and the actual observed value.

[0068] MAE (Mean Absolute Error): The average absolute error between the predicted wind speed and the actual observed wind speed.

[0069] CORR (Pearson Correlation Coefficient): The similarity between the actual and predicted wind speed values.

[0070] The smaller the RMSE value, the smaller the MAE value, and the larger the CORR value, the higher the inversion accuracy of the model and the better the model performance.

[0071] Experiment 1: Under the same dataset, compare the SAR sea surface high wind speed inversion model of this invention, which integrates dual attention mechanism, with the inversion performance of models trained by other mainstream deep learning frameworks.

[0072] After obtaining the optimal sea surface wind speed inversion model under typhoon conditions, a total of 4361 sets of preprocessed SAR data and corresponding ground truth wind speed data from SFMR were acquired in the experiment. The optimal model was then used for inference, outputting the sea surface wind speed inversion values ​​for each set. Inference verification was performed on the 4361 sets of data, and statistical analysis of the results yielded the following indices for model inversion accuracy: RMSE = 0.7172 m / s, MAE = 0.4652 m / s, CORR = 0.9954. The scatter plot and error distribution are shown below. Figure 6As shown. To verify the accuracy of the inversion model implemented by the method of this invention in retrieving sea surface wind speed under typhoon conditions, it was compared with current mainstream network models such as beit3, SwinTransformer, and cs3darknet. Each model underwent necessary parameter tuning before the experiment to ensure that their performance was compared under optimal conditions.

[0073] Table 1. Inversion index results for each model on the same test set. Beit3 0.8605 0.3424 0.9933 SwinTransformer Tiny 1.0042 0.5288 0.9909 Cs3darknet 1.2112 0.7331 0.9868 Densenet 1.3598 0.9634 0.9833 Ghostnet 1.8208 1.2833 0.9701 Edgenet 2.6471 1.8745 0.9368 SAR high wind speed inversion model 0.7172 0.4652 0.9954

[0074] As shown in Table 1, the model of this invention significantly outperforms some mainstream deep learning models in retrieving sea surface wind speeds during typhoons. In terms of RMSE accuracy, the model of this invention improves upon beit3, Swin Transformer Tiny, and Cs3darknet by 16.65%, 28.58%, and 40.79%, respectively. Furthermore, the inference results of the model of this invention also outperform other networks in terms of correlation, demonstrating excellent inversion performance.

[0075] Experiment 2: Evaluate the inversion performance of the SAR sea surface high wind speed inversion model that integrates the dual attention mechanism of this invention under different wind speed ranges.

[0076] After obtaining the optimal sea surface wind speed inversion model under typhoon conditions, a total of 4361 sets of preprocessed SAR data and corresponding ground truth wind speed data for SFMR were acquired in the experiment. The optimal model was then used for inference, and the sea surface wind speed inversion values ​​for each group were output according to wind speed. Inference verification was performed on the 4361 sets of data, and statistical analysis was conducted on the results to obtain various indicators of the model inversion accuracy. Table 2 shows the inversion indices of the model of this invention in various wind speed ranges. 0~7 m / s 102 0.9724 0.5264 0.7502 0.5665 7~20 m / s 2124 0.9830 0.2826 0.4417 0.9000 20~40 m / s 1859 0.7709 0.1270 0.4846 0.9778 ≥40 m / s 276 1.4729 -0.2984 0.7242 0.9694

[0077] As shown in Table 2, the inversion indices of this invention demonstrate excellent inversion performance in the 0-40 m / s range, with an RMSE index consistently around 1 m / s. The model performs best in the medium-to-high wind speed region (7-40 m / s), maintaining a relatively good RMSE while exhibiting small BIAS error, stable MAE, and a good correlation between the predicted and actual values.

[0078] This invention utilizes a validation set composed of randomly selected samples from the sample dataset to verify the high-accuracy sea surface wind speed inversion of the proposed model under high wind speed conditions. Within the typical wind speed range of 0-20 m / s, the model maintains a stable RMSE (Recovery Mean Squared Error) and excellent overall accuracy. Even in the high wind speed range of 20-40 m / s, it achieves high inversion accuracy and robustness, with the overall RMSE fluctuation not exceeding 0.4 m / s and the MAE (Maximum Effectiveness) fluctuation not exceeding 0.4 m / s. Compared with other mainstream deep learning networks, the proposed model still achieves a relatively good level of accuracy and is significantly superior to high-performance networks such as SwinTransformer. In summary, the model's excellent performance across various wind speed ranges demonstrates that it effectively achieves in-depth analysis of the complex characteristics of the sea surface under typhoon conditions, and completes the deep decoupling of wind speed-related information from complex mixed SAR signals. Multi-dimensional comparative experiments prove the superior performance of the SAR sea surface high wind speed inversion model of this invention in solving the problem of distorted sea surface wind speed inversion caused by SAR signal saturation under typhoon conditions. Its inversion results have good performance, providing a new model architecture paradigm for sea surface wind speed inversion, and have significant scientific value and engineering application prospects.

[0079] This invention provides a SAR sea surface high wind speed inversion method that integrates a dual attention mechanism. Based on the complex texture features and high-dimensional frequency domain features of radar backscattering intensity (roughness) in SAR data, a complex nonlinear mapping model weight is constructed on the mechanism of nonlinear response of sea surface wind speed, realizing sea surface wind speed inversion under typhoon conditions. This significantly improves the problem of sea surface wind speed inversion distortion caused by SAR signal saturation under typhoon conditions and greatly enhances the inversion accuracy.

[0080] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A SAR method for high wind speed inversion over sea surface incorporating a dual attention mechanism, characterized in that, Includes the following steps: Step 1: Acquire SAR remote sensing image data under typhoon conditions; Step 2: Preprocess the SAR remote sensing image data to obtain several SAR slice data; Step 3: Input several SAR slice data into the trained SAR high wind speed inversion model. The SAR high wind speed inversion model extracts features from several SAR slice data through a weight file. The extracted features are then subjected to channel compression regression to obtain the inversion wind speed results. In step three, the construction process of the SAR sea surface high wind speed inversion model is as follows: Step 31: Perform dB conversion on the SAR slice data, resample the data, and augment the data; Step 32: Calculate the polarization ratio data for the data-enhanced SAR slice data, and concatenate and stitch the SAR slice data and polarization ratio data in the channel dimension; Polarization ratio calculation: in, SAR slice data after resampling under the VH polarization channel; SAR slice data after resampling under the VH polarization channel; Step 33: After cascading and stitching, the SAR slice data is used to invert the wind speed using the SAR sea surface wind speed model; The cascaded and stitched SAR slice data is denoted as... The wind speed was finally obtained by inversion as follows: ; in, : Model wind speed prediction; Input: SAR slice data after cascaded stitching; : Bi-branch manifold mapping operator; Prior estimation of intermediate wind speed; Dynamic gain operator; and Adama product and nonlinear gating; Eventual regression operator; Attention mechanisms; Step 34: Compare the inverted wind speed prediction values ​​with the training data. The difference from the true value is used to construct a SAR high wind speed inversion model based on a SAR sea surface high wind speed inversion model framework that integrates dual attention mechanism, by iterating the mapping weight between SAR slice data and target wind speed. In step 34, the SAR sea surface high wind speed inversion model framework integrating dual attention mechanism includes a frequency domain attention module, a dual-branch ConvNeXT backbone network, and a wind speed guided attention module.

2. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 1, characterized in that, In step two, the preprocessing of SAR remote sensing image data is as follows: First, the SAR remote sensing image data is subjected to radiometric calibration, speckle noise removal, and geographic correction in sequence. Then, the processed SAR remote sensing image data is spatiotemporally matched with the corresponding typhoon event SFMR data; Finally, the spatiotemporally matched data is sliced ​​to obtain several SAR slice data.

3. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 1, characterized in that, In step 31, dB conversion: the values ​​of each point in the SAR slice data are converted to dB. in, SAR slice data after conversion to dB format; SAR slice data; Data resampling: Resample each point of the dB SAR image data to the interval [0, 225]. in, This is SAR slice data after resampling; : Maximum value of SAR slice data in dB format; : minimum value of SAR slice data in dB format; SAR slice data in dB format; Data augmentation: Perform random horizontal flipping, vertical flipping, and 15° random rotation on the SAR slice data.

4. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 1, characterized in that, In step 34, the weight iterative optimization is achieved by the loss function formulas (3-2) to (3-4): Where TotalLoss: the total loss function; The main loss function is the loss function used to process the features after data augmentation. : Auxiliary loss function, which is the loss function that processes the characteristics of the estimated intermediate wind speed; N: Total number; : Represents the actual value and the predicted value, respectively; : Represents the actual value and the predicted value, respectively.

5. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 2, characterized in that, The frequency domain attention module maps SAR slice data from the spatial domain X to the frequency domain K through a two-dimensional discrete Fourier transform, performs adaptive filtering in the complex space, and then restores the data. in, Input SAR remote sensing image intensity: Spatial frequency coordinates; and Complex filter kernel weights; Adama accumulation; A dual-branch ConvNeXT backbone network is used for nonlinear fusion of the restored frequency domain features and spatial domain features: in, Features after fusion; Spatial domain branch network processes and outputs spatial domain features; Frequency domain branch network output frequency domain characteristics; : Fusion ratio coefficient, default value is 0.8; Wind speed-guided attention module, including feature prediction and feature modulation; used for dynamically scaling the gain of feature channels; Through dimensionality reduction mapping Extract core features and calculate an intermediate scalar wind speed prediction value. : in, and Intermediate prediction weights and biases; Features after fusion; (·): Activation function; Then, an attention mechanism is introduced into this module: Where A: feature mask vector, which records the features that should be amplified and the features that should be suppressed during regression; : Dimensional upscaling; Features after fusion; Nonlinear gating; : Predicted scalar value of intermediate wind speed.

6. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 1, characterized in that, Training of the SAR high wind speed inversion model: Repeat steps 33 and 34 to iteratively train the SAR high wind speed inversion model. Use the gradient descent algorithm to correct the model parameters until the loss function reaches the order of 1e-3 on the validation set and remains stable. Then stop training and obtain the trained SAR high wind speed inversion model and obtain the weight file.

7. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 6, characterized in that, Hyperparameter settings for the SAR high wind speed inversion model: The AdamW optimizer is used, with an initial learning rate of 1e-5, a minimum learning rate of 1e-7, a learning rate scheduler of CosineAnnealingLR, a weight decay of 1e-6, and a learning patience epoch of 50, meaning that training stops when the validation loss and validation RMSE do not decrease for 50 consecutive epochs.

8. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 2, characterized in that, Spatiotemporal matching: The matching time window for SFMR data of typhoon events is determined by 1 hour before and after the imaging time of SAR remote sensing image data. The effective sample points of SFMR data of typhoon events within the imaging range of SAR remote sensing image data are obtained by matching latitude and longitude.

9. The SAR sea surface high wind speed inversion method fused with dual attention mechanism according to claim 8, characterized in that, Slicing the spatiotemporally matched data: Based on the location of the valid sample points of the SFMR data of the typhoon event, slice the data into corresponding 224x224 pixel slices with the sample points as the center.

Images