A method for revising and improving resolution of sea surface high wind speed reanalysis products based on a bimodal U-Net

Abstract

The application belongs to the technical field of remote sensing, and relates to a sea surface high wind speed reanalysis product correction and resolution improvement method based on a bimodal U-Net. SAR and ERA5 wind speed data are acquired and preprocessed, the tropical cyclone center position of the SAR wind speed image is taken as a benchmark to correct the ERA5 wind speed image, and an ERA5 wind speed sub-image is obtained by cutting to serve as an input data set, and a SAR wind speed sub-image serves as a label; a two-stage bimodal tropical cyclone wind speed reconstruction system is constructed, a correction network is used for early correction of the ERA5 wind speed data; an attention enhancement U-Net network with bimodal input features is adopted in the reconstruction stage, and the feature map is mapped to a single channel wind speed output through two layers of convolution. The application realizes end-to-end reconstruction from low-resolution ERA5 input to high-resolution refined wind speed output, effectively restores the detailed features of the eye area and spiral structure of the tropical cyclone, and provides a reliable intelligent solution for sea surface wind speed monitoring under extreme weather.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology, specifically relating to a method for correcting and improving the resolution of high wind speed reanalysis products on the sea surface based on dual-modal U-Net. Background Technology

[0002] Tropical cyclones, as highly destructive natural disasters, pose a serious threat to the lives and property of coastal ports and residents, often resulting in significant economic losses. Accurately obtaining tropical cyclone wind speed information is crucial for ensuring maritime navigation safety, improving disaster early warning capabilities, and optimizing tropical cyclone monitoring systems. However, acquiring comprehensive and highly accurate tropical cyclone observation data remains an extremely challenging task.

[0003] With the development of remote sensing technology, satellite remote sensing data has been widely used for the dynamic monitoring and evolution analysis of tropical cyclones. Currently, commonly used sea surface wind speed data mainly includes two categories: satellite inversion products and global / regional assimilation system products. Satellite remote sensing, especially microwave remote sensing, possesses all-weather, all-time observation capabilities, but its sea surface wind products generally have low spatial resolution, making it difficult to accurately depict the fine spatial structure of tropical cyclones in complex waters such as nearshore areas. While synthetic aperture radar (SAR) can accurately measure sea surface wind speed with high spatial resolution, its wind speed products exhibit irregularities and sparsity in temporal and spatial coverage, typically failing to cover complete tropical cyclones and hindering long-term continuous observation. Although some assimilation reanalysis products have assimilated satellite observation data using statistical methods or spectral analysis to improve wind speed accuracy, due to factors such as limited atmospheric model resolution, insufficient assimilation methods, and constraints on remote sensing inversion accuracy, some operational assimilation wind speed products (such as ERA5) still exhibit significant underestimation in high-wind-speed areas. The ERA5 global reanalysis dataset, released by the European Centre for Medium-Range Weather Forecasts (ECMWF), integrates global atmospheric, oceanic, and land observation data from 1940 to the present. Through advanced data assimilation techniques, it combines historical forecasts with real-time observations, creating a spatiotemporally coherent, globally comprehensive data product. This dataset provides hourly sea surface wind speeds (u and v components) with a spatial resolution of 0.25° × 0.25°, providing an important data foundation for tropical cyclone research. However, due to its relatively coarse spatial resolution and insufficient accuracy in measuring high wind speeds, ERA5 still has limitations in supporting refined monitoring of tropical cyclones and is not yet sufficient to effectively simulate the dynamics and physical processes in the core region of tropical cyclones.

[0004] Studies have shown that high-resolution wind speed data can significantly improve the simulation and forecasting capabilities of tropical cyclones. Therefore, developing sea surface wind speed products that combine high resolution, high accuracy, and wide coverage has become an urgent need. Traditional methods for improving wind speed resolution mostly rely on interpolation algorithms, such as linear interpolation or spline interpolation. These methods can improve the spatial resolution of wind speed data to a certain extent, but they often cannot fully recover the fine structure and local features of wind speed.

[0005] In recent years, deep learning-based super-resolution techniques have emerged. Some deep learning methods (such as CNNs) perform well in capturing local spatial features and reconstructing low-frequency textures, but their numerous convolutional operations easily lead to the loss of high-frequency information, making it difficult to accurately reconstruct the texture structure inside tropical cyclones. Generative Adversarial Networks (GANs) can generate more realistic texture details through adversarial training, but they often face problems such as training instability and mode collapse. U-Net achieves resolution improvement by combining a multi-layer feature extraction and progressive upsampling decoding structure. Its skip connection mechanism directly transmits high-resolution features from the encoder to the corresponding layer of the decoder, thereby reducing the loss of detail caused by multiple convolutions and downsampling. However, under extreme high wind speed conditions, how to accurately model the complex nonlinear relationship between different wind speed levels remains a challenge that urgently needs to be solved.

[0006] The choice of loss function is one of the important factors affecting the training accuracy of a model. In super-resolution tasks, most models use distortion-driven loss functions (such as L1 loss) for optimization. However, such losses tend to cause the model to reconstruct pixel values ​​as intermediate values ​​among multiple possible values, resulting in overly smoothed results in high-frequency texture regions. Some studies have introduced perceptual-driven losses to enhance image realism, but this may introduce noise and artifacts. Currently, in super-resolution reconstruction of sea surface wind speed data, there is a lack of loss function designs that can effectively balance image realism and physical accuracy.

[0007] In summary, existing sea surface wind speed data often suffers from deficiencies in spatial resolution, spatiotemporal coverage, or strong wind conditions, while existing super-resolution algorithms remain limited in detail recovery (especially in tropical cyclone eye reconstruction). Furthermore, research revealed that SAR wind speeds have not yet been assimilated into the ERA5 reanalysis system. Therefore, this study proposes an algorithm based on the U-Net network architecture that integrates the advantages of ERA5 and SAR data: using ERA5 wind speeds with continuous spatiotemporal coverage as a foundation, and training with high spatial resolution SAR inversion wind speeds, this algorithm corrects the underestimation of high wind speeds in ERA5 while improving its spatial resolution and wind speed accuracy, thereby achieving high-precision reconstruction of tropical cyclone wind speeds. Summary of the Invention

[0008] The purpose of this invention is to address the problems of insufficient spatial resolution and widespread underestimation in high-wind-speed areas in reanalysis sea surface wind speed products under extreme weather conditions. This invention proposes a method for correcting and enhancing the resolution of high-wind-speed reanalysis products based on a dual-modal U-Net. The proposed resolution enhancement algorithm (WindUNet-AF) is described in this invention. 2 By fully integrating the advantages of high spatial resolution of SAR data and wide spatiotemporal coverage of ERA5 reanalysis data, the spatial resolution of ERA5 data is improved by 8 times, while significantly improving the estimation accuracy of wind speed regions, thereby completing the high-resolution and refined reconstruction of wind speed of tropical cyclones over the sea.

[0009] The technical solution of this invention is:

[0010] This invention provides a method for correcting and improving the resolution of high wind speed reanalysis products based on dual-modal U-Net, comprising the following steps:

[0011] (1) Obtain wind speed data from STAR-SAR and ECMWF-ERA5 and preprocess them. Using the location of the tropical cyclone center on the SAR wind speed image as a reference, correct the ERA5 wind speed image and then crop it into wind speed sub-images. Use the obtained ERA5 wind speed sub-images as the input low-resolution dataset and the SAR wind speed sub-images as the labels for the tropical cyclone correction network and reconstruction network.

[0012] (2) Construct a dual-stage, dual-modal tropical cyclone wind speed reconstruction system containing a correction network and a super-resolution network to reconstruct high-precision wind speed from low-resolution ERA5 wind speed data, including a correction stage and a reconstruction stage; the correction network adopts a U-Net network architecture with an attention gating mechanism for the early correction of ERA5 wind speed data; the reconstruction stage adopts an attention-enhanced U-Net network with dual-modal input features, firstly extracts features from the input corrected ERA5 and original ERA5 dual-channel wind speed data, then processes them using three feature fusion strategies, and finally enhances the feature extraction capability through the residual convolution module in the super-resolution network, and finally maps the feature map to the single-channel wind speed output through two layers of convolution;

[0013] The correction phase employs a hybrid intensity loss function, while the reconstruction phase employs an adaptive hybrid loss function.

[0014] Furthermore, the adaptive hybrid loss function consists of weighted intensity loss, gradient loss, spectral loss, and multi-scale perceptual loss, used to improve training stability and reconstruction performance, and its formula is as follows: Where, λ I Let λ be the weighted intensity loss function. G Let λ be the gradient loss function. S Let λ be the spectral loss function.P For multi-scale sensing loss function, L I , L G , L S , L P These correspond to the weights of the four loss functions. Furthermore, the hybrid intensity loss function is used to improve the accuracy of ERA5 wind speed data correction; its formula is as follows: in, The pixels in a SAR wind speed image with a resolution of 25 km. represents the pixels in the corrected ERA5 wind speed image, where n is the total number of pixels.

[0015] Furthermore, the weighted intensity loss improves the accuracy of the reconstructed wind speed data through dynamic weight adjustment, and its formula is as follows: in, y SAR The wind speed indicated by the label, y SR This represents the wind speed after super-resolution, where n is the total number of pixels.

[0016] The gradient loss employs the Gaussian-smoothed Sobel gradient operator to maintain spatial continuity, and its formula is as follows: in, GaussianFilter represents Gaussian filtering. Indicates the calculation of the gradient;

[0017] The spectral loss employs a radial weighting strategy in the frequency domain to ensure accurate reconstruction of the tropical cyclone eye circulation structure. The formula is as follows: in, , ; i , j These are the row and column numbers corresponding to the image's frequency domain space, respectively. ic and jc The row and column numbers of the center pixel. R max The maximum distance from the center of the image. r The distance from the image center is denoted by FFT(·), and FFT(·) is the Fourier transform.

[0018] The multi-scale perceptual loss captures feature similarity at different resolutions through three levels of average pooling: original scale, 2x downsampling, and 4x downsampling. The weight allocation is 1:2:4 to ensure the reconstruction details and quality of the image. The formula is as follows: ;

[0019] in, , , AvgPool(·) is the average pooling operation.

[0020] Furthermore, the reconstruction phase adopts an adaptive weight scheduling mechanism based on training progress, which is divided into four phases: the intensity loss of the first phase is used in the first 30% of epochs, gradient loss is gradually introduced in the second phase, i.e., 30-50% of epochs, spectral loss is gradually added in the third phase, i.e. 50-80% of epochs, and all loss functions are jointly optimized in the fourth phase, i.e. 80-100% of epochs.

[0021] Furthermore, in step (2), the dual-modal input feature attention-enhanced U-Net network used in the reconstruction stage includes a feature extraction module, a feature fusion module, and a super-resolution network module. The feature extraction module extracts features from the corrected ERA5 and the original ERA5 dual-channel wind speed data to obtain two features F1 and F2. Then, a multi-strategy feature fusion module containing channel splicing fusion, adaptive weighted fusion, and residual fusion is used to process the data, integrating the structured information and original texture details of the corrected image and the original image to prevent information loss during feature extraction and fusion. The super-resolution network adopts a progressive upsampling strategy and enhances the feature extraction capability through three residual convolution modules containing CBAM attention mechanism.

[0022] Furthermore, in step (1), the ERA5 input data is standardized using the Min-Max standardization method, with the following formula: in x , x* These are the values ​​before and after the transformation, min( x ) and max( x These represent the minimum and maximum values ​​of the data, respectively.

[0023] After data standardization, the low-resolution dataset is divided into a training set and a test set. The test set is used to test the model's wind speed reconstruction effect, and 20% of the data in the training set is used as a validation set to verify whether the model has good convergence.

[0024] The same processing method is used for the SAR image label dataset.

[0025] The beneficial effects of this invention are:

[0026] This invention first decomposes the complex wind speed reconstruction task into two stages: "correction" and "super-resolution," using a two-stage progressive architecture. The first stage focuses on correcting the systematic underestimation of high wind speeds in ERA5 data and suppressing noise interference in SAR images. The second stage achieves high-resolution reconstruction based on multimodal fusion. This phased design significantly improves the stability and generalization ability of the model training. Secondly, the algorithm design introduces a dual-modal adaptive fusion module, fusing the corrected wind speed information with the original ERA5 wind speed data through dynamic attention weights. This adaptively learns complementary features between multiple data sources, effectively reducing information loss and enhancing feature representation capabilities. Finally, a multi-stage adaptive hybrid loss function is employed in the super-resolution network. Through a dynamic weight scheduling strategy, the intensity loss, gradient loss, spectral loss, and multi-scale perception loss are progressively adjusted in stages during training. This avoids suboptimal convergence that is easily caused by traditional fixed-weight losses, significantly improving the accuracy and stability of the model in reconstructing complex wind speed structures.

[0027] The WindUNet-AF provided by this invention 2 The algorithm achieves end-to-end reconstruction from low-resolution ERA5 input to high-resolution refined wind speed output, effectively restoring the detailed features of the eye region and spiral structure of tropical cyclones, and providing a reliable intelligent solution for sea surface wind speed monitoring under extreme weather conditions. Attached Figure Description

[0028] Figure 1 This is a flowchart of the ERA5 sea surface wind speed correction and super-resolution reconstruction algorithm (WindUNet-AF²) based on dual-stage dual-modal U-Net in this invention.

[0029] Figure 2 The WindUNet-AF² tropical cyclone wind speed reconstruction model provided by this invention.

[0030] Figure 3 A schematic diagram for improving the PPM module.

[0031] Figure 4 The effects of different algorithms on the wind speed reconstruction results of tropical cyclones are shown. Among them, (a) is the input low-resolution ERA5 wind speed data, (e) is the high-resolution SAR wind speed data, (b), (c), and (d) are the results of this algorithm (WindUNet-AF² algorithm), the direct reconstruction method, and the cubic spline interpolation method, respectively, and (f), (g), and (h) are the relative errors between the reconstruction results of the corresponding algorithms and the SAR wind speed data, respectively. Detailed Implementation

[0032] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] To further understand the present invention, it will be further described in conjunction with the accompanying drawings and embodiments.

[0034] Example 1

[0035] This invention provides a method for correcting and improving the resolution of high wind speed reanalysis products based on dual-modal U-Net, comprising the following steps:

[0036] Step 1: Acquire SAR wind speed data from the Center for Satellite Applications and Research (STAR) containing tropical cyclones and fifth-generation reanalysis wind speed data (ECMWF-ERA5) from the European Centre for Medium-Range Weather Forecasts (ECMWF-ERA5). ERA5 wind speed data is used as input, while SAR wind speed data is used for labeling and model accuracy evaluation. A 15×15 Lee filter window is used to denoise the SAR wind speed images to eliminate the influence of speckle noise. To ensure data timeliness, linear interpolation is performed on the ERA5 wind speed data based on the SAR imaging time to generate data with the same timestamp for spatiotemporal matching.

[0037] SAR wind speed data were downsampled and projected onto a standardized grid. The original SAR wind speed data had a resolution of 500 m. Cubic spline interpolation was used to downsample the data, reducing the resolution to 3.125 km (0.03°, 8 times the ERA5 resolution), which was then used as the label for the super-resolution network.

[0038] Because of discrepancies in imaging time and resolution between the two datasets, it is necessary to calibrate their tropical cyclone center locations. First, the locations of the tropical cyclone centers (i.e., the latitude and longitude coordinates of the pixels containing the minimum wind speed within the tropical cyclone eye) are extracted from the SAR and ERA5 wind speed images. Then, the latitude and longitude offsets lat_offset and lon_offset of the two tropical cyclone center locations are calculated respectively. These two offsets are then used to correct the latitude and longitude coordinates of the ERA5 wind speed data.

[0039] The SAR wind speed image downsampled by 8 times is cropped into a 64×64 pixel sub-image, with a slider step of 5 pixels in both the row and column directions. Using this 64×64 pixel sub-image as a baseline, the corresponding region of the ERA5 wind speed sub-image is extracted. The cropped ERA5 wind speed sub-image should then be 8×8 pixels. If the pixel sizes do not match, the SAR or ERA5 wind speed sub-image can be adjusted appropriately to ensure data consistency.

[0040] After cropping the SAR and ERA5 wind speed sub-images, the data is filtered and outliers are removed. The ERA5 wind speed sub-images form an n×8×8 three-dimensional matrix, serving as the input low-resolution dataset. The SAR wind speed sub-images form an n×64×64 three-dimensional matrix, serving as the labels for super-resolution reconstruction. n represents the number of successfully matched sub-images.

[0041] For a SAR wind speed dataset of size n×64×64, cubic spline interpolation was used to downsample it to 25km (0.25°, consistent with ERA5 resolution) and used as labels for the tropical cyclone correction network.

[0042] Step 2: Construct a product correction and resolution enhancement algorithm for high wind speed reanalysis at sea surface based on bimodal U-Net (WindUNet-AF) using the Python programming language. 2 The specific steps are as follows:

[0043] (1) Data standardization

[0044] When the input parameters differ significantly in value, directly using the raw values ​​for analysis will emphasize the role of parameters with higher values ​​in the model, relatively weakening the role of parameters with lower values. Therefore, to ensure the reliability of the results, it is necessary to standardize the raw input data. This invention uses the Min-Max standardization method, i.e. (1)

[0045] in x , x * Represents the values ​​before and after the transformation, min( x ) and max( x ) represent the minimum and maximum values ​​of the data, respectively.

[0046] (2) Division of training set, validation set and test set

[0047] The input features of the n×8×8 low-resolution dataset after data standardization in (1) are randomly divided into training and test sets in an 8:2 ratio. The test set is used to test the wind speed reconstruction effect of the model. 20% of the data in the training set is used as the validation set to verify whether the model has good convergence. The n×64×64 high-resolution SAR image label dataset is processed in the same way.

[0048] (3) Model structure

[0049] The proposed algorithm (WindUNet-AF²) for correcting and upgrading high-wind-speed reanalysis products based on a bimodal U-Net is a two-stage, bimodal tropical cyclone wind speed reconstruction system comprising a correction network and a super-resolution network. It is used to reconstruct high-precision wind speeds from low-resolution ERA5 wind speed data. The network framework is as follows: Figure 1 As shown.

[0050] (3.1) Correction stage

[0051] The correction network employs a U-Net architecture with an attention gate mechanism, with both input and output being 8×8 pixel images, specifically designed for pre-correction processing of ERA5 wind speed data. The encoder consists of three downsampling blocks, while the decoder reconstructs features through transposed convolutions and skip connections. An innovative attention gate mechanism is introduced at each skip connection, dynamically weighting skip features with learnable attention coefficients to enhance the model's feature selection capability for important regions.

[0052] In the correction phase, a hybrid loss function combining mean squared error (MSE) and mean absolute error (MAE) was used, as shown in the formula below. This loss function can fully combine the advantages of both, ensuring rapid convergence of the model and avoiding extreme values ​​from tending to the mean, thereby improving the accuracy of ERA5 wind speed data correction. (2)

[0053] in, The pixels in a SAR wind speed image at a resolution of 25 km. represents the pixels in the corrected ERA5 wind speed image, where n is the total number of pixels.

[0054] (3.2) Restructuring Phase

[0055] In the reconstruction phase, an attention-enhanced U-Net network with dual-modal input features was designed, comprising a feature extraction module, a feature fusion module, and a super-resolution network module. For the dual-channel input data of the corrected ERA5 and the original ERA5, feature extraction was first performed to obtain two features, F1 and F2. Then, a multi-strategy feature fusion module, including "channel concatenation fusion," "adaptive weighted fusion," and "residual fusion," was used for processing. The results of the three fusion strategies were concatenated and then subjected to 1×1 convolution for dimensionality reduction, effectively integrating the structured information and original texture details of the corrected and original images, preventing information loss during feature extraction and re-fusion.

[0056] The super-resolution network employs a progressive upsampling strategy, upsampling the image from 8×8 to 64×64 using three residual convolutional modules incorporating a CBAM attention mechanism. The structure of the residual convolutional modules is shown below. Figure 1 Compared to traditional convolutional methods, residual convolution can more fully preserve upper-layer information and avoid the loss of shallow information during transmission. The CBAM attention mechanism combines spatial and channel attention mechanisms, enhancing feature extraction capabilities.

[0057] The encoder-decoder architecture comprises three downsampling layers and three upsampling layers, each containing a residual convolutional module. This architecture combines residual connections and attention mechanisms to enhance feature extraction capabilities and training stability. In the decoding stage, a spatial attention mechanism and an improved pyramid pooling module (PPM) are introduced to further enhance feature extraction. The spatial attention module learns the spatial weight distribution by concatenating max-pooling and average-pooling features, helping the model focus more on the key eye region and maintain the spatial continuity of wind speed. The improved PPM module captures contextual information from different receptive fields through multi-scale pooling, such as... Figure 2 As shown, a residual structure is introduced based on the traditional PPM module, and the learning ability of key features, such as low-frequency features of tropical cyclone circulation structure, is further enhanced through a channel attention mechanism. Finally, the network maps the feature map to a single-channel wind speed output through two layers of convolution.

[0058] (3.3) Adaptive hybrid loss function

[0059] To overcome the limitations of traditional single loss functions, an adaptive hybrid loss function is used in the reconstruction stage. This loss function consists of four key components: "weighted intensity loss," "gradient loss," "spectral loss," and "multi-scale perceptual loss," as shown in the following formula: (3)

[0060] Where, λ I , λ G , λ S , λ PThese correspond to four loss functions respectively. L I , L G , L S , L P The weights correspond to the four loss functions and change with the number of training rounds, as detailed in Table 1.

[0061] (3.3.1) Weighted intensity loss: The problem of "regression to the mean" in the regression is solved by dynamically adjusting the weights, and higher weights are given to low and high wind speed areas to improve the accuracy of the reconstructed wind speed data; (4)

[0062] (5)

[0063] Among them, y SAR The wind speed indicated by the label, y SR This represents the wind speed after super-resolution, where n is the total number of pixels.

[0064] (3.3.2) Gradient loss: The Gaussian smoothing Sobel gradient operator is used to maintain spatial continuity and has good robustness to noise; (6) (7)

[0065] in,

[0066] GaussianFilter

[0067] Indicates Gaussian filtering. This indicates that the gradient is calculated.

[0068] (3.3.3) Spectral loss: A radial weighting strategy is adopted in the frequency domain space, with the weight of the central low-frequency region being 1.0 and the weight of the edge high-frequency region being reduced to 0.2, to ensure accurate reconstruction of the circulation structure of the tropical cyclone eye region; (8)

[0069] (9)

[0070] (10)

[0071] in, i , j These are the row and column numbers corresponding to the image's frequency domain space, respectively. ic and jc The row and column numbers of the center pixel. R max The maximum distance from the center of the image. ris the distance from the center of the image. FFT(·) is the Fourier transform.

[0072] (3.3.4) Multi-scale perceptual loss: Three-level average pooling (original scale, 2x downsampling, 4x downsampling) is used to capture feature similarity at different resolutions, with a weight distribution of 1:2:4 to ensure the reconstruction details and quality of the image. (11)

[0073] (12)

[0074] (13) (14)

[0075] AvgPool(·) is the average pooling operation.

[0076] To optimize the training process, an adaptive weight scheduling mechanism based on training progress was designed. This mechanism divides the training into four stages:

[0077] The first phase (0-30% epoch) focuses primarily on intensity loss to ensure the accuracy of wind speeds after reconstruction.

[0078] In the second stage (30-50% epoch), gradient loss is gradually introduced to enhance the smoothness and continuity of the image;

[0079] In the third stage (50-80% epoch), spectral loss is gradually added to improve the reconstruction effect of tropical cyclone structure;

[0080] In the fourth stage (80-100% epoch), all loss functions are optimized collaboratively.

[0081] This progressive loss weight adjustment strategy effectively balances the learning focus of different training stages, ensuring a smooth transition from simple to complex and from local to global training processes. It avoids the problem of overfitting to a single loss term in early training and significantly improves training stability and reconstruction performance.

[0082] Table 1. Weights of the loss function at each stage in the adaptive weight scheduling mechanism.

[0083]

[0084] * p = epoch / total_epochs ∈ [0,1]

[0085] The two-stage optimizer uses Adam (Adaptive Moment Estimation), with learning rates set to 1e-3 and 1e-4 respectively, which helps stabilize training.

[0086] (6) Accuracy assessment

[0087] The selected accuracy evaluation indices are Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Peak Signal-to-Noise Ratio (PSNR). RMSE is used to comprehensively measure the difference between the generated image and the original image and is a key indicator of wind speed reconstruction accuracy. The calculation formula is as follows: (15)

[0088] MAE can represent the ratio of the deviation between the reconstructed wind speed and the SAR data to the true value. The larger the error, the more dispersed the deviation distribution between the two data. The calculation formula is: (16)

[0089] PSNR represents the signal-to-noise ratio. A higher value indicates that the image with improved model resolution is closer to the real image. The calculation formula is as follows: (17)

[0090] in, SR and HR These represent the reconstructed ERA5 wind speed and SAR wind speed, respectively.

[0091] Example 2

[0092] This application example uses the SAR sea surface wind speed product provided by the STAR dataset and the ERA5 reanalysis wind speed product released by ECMWF as examples. Figure 3 The following describes the sea surface high wind speed reanalysis product correction and resolution enhancement algorithm (WindUNet-AF²) based on dual-modal U-Net.

[0093] First, install the Python programming software on the user terminal, and you will need to have the following environment packages: Python-3.8.20, tensorflow-GPU-2.6.0, and keras-2.6.0. Tensorflow and keras are open source software libraries for deep learning. Tensorflow is used to build, train, and deploy various complex neural network models and supports multiple programming languages. Keras, as its high-level API, provides a concise interface for quickly implementing common network models.

[0094] I. Accuracy Analysis of Model Reconstruction

[0095] The experiment matched 21,833 image pairs of low-resolution ERA5 wind speed data with high-resolution SAR wind speed data, which were proportionally divided into a training set (90%, 15,719 pairs) and a test set (10%, 2,184 pairs). An additional 20% (3,930 pairs) of the training set was used as a validation set. Figure 4As shown in (b), the WindUNet-AF² model can effectively reconstruct wind speeds in both the low-wind-speed region surrounding tropical cyclones and the eye region of tropical cyclones. This indicates that the WindUNet-AF² algorithm not only improves the spatial resolution of sea surface wind speed images and corrects ERA5 wind speed values, but also generates textures and structures similar to those in the eye region of tropical cyclones in SAR images, significantly improving the quality of wind speed data products. Quantitative evaluation results show that the WindUNet-AF² model achieved good reconstruction results on the test set: compared with SAR wind speeds, the RMSE is 2.29 m / s, the MAE is 1.41 m / s, and the PNSR is 32.30.

[0096] II. Model Structure Analysis

[0097] The WindUNet-AF² model innovatively employs a dual-branch network architecture. Through a two-stage design of a "wind speed correction network" and a "super-resolution network" with dual-modal input, and using an end-to-end progressive training method, it achieves accurate reconstruction from 8×8 low-resolution input to 64×64 high-resolution wind speed output. To verify the improvement effect of dual-stage and dual-modal input on reconstruction performance, the algorithm of this invention (…) Figure 4 (b) Compare with the following two methods respectively:

[0098] (1) Direct reconstruction method, that is, directly super-resolution (corresponding to) the ERA5 data. Figure 4 (c)

[0099] (2) Traditional cubic spline interpolation method Figure 4 (d)

[0100] Compared to traditional cubic spline interpolation algorithms, deep learning-based algorithms can more accurately reconstruct the eye region and wind speed structure of tropical cyclones. However, when the eye region of a tropical cyclone is small, the model's ability to capture detailed changes is limited: directly super-resolution processing of ERA5 data makes it difficult to reconstruct the low-wind-speed contours within the eye region, and high-wind-speed areas are underestimated. By introducing a correction network and inputting both the original ERA5 data and the corrected wind speeds (using the WindUNet-AF² algorithm of this invention), the reconstruction error in high-wind-speed areas is reduced, the structure of the tropical cyclone eye region is reconstructed almost completely, and the reconstruction error within the eye region is significantly reduced.

[0101] Table 2 lists the quantitative comparison results of the three reconstruction algorithms on the test set. It can be seen that, compared with the cubic spline interpolation method, the WindUNet-AF² method improves the overall quality of the wind speed reconstruction image by enhancing the attention to image texture details, reducing RMSE by 3.83 m / s, MAE by 2.51 m / s, and PSNR by 7.06.

[0102] Table 2. Accuracy comparison of the proposed algorithm, the direct reconstruction algorithm, and the cubic spline interpolation method on the test set.

[0103]

[0104] III. Loss Function Analysis

[0105] This invention improves the design of the loss function, proposing a multi-stage adaptive loss framework suitable for tropical cyclone wind speed reconstruction tasks, and achieving progressive optimization through dynamic weight scheduling. This framework introduces weighted intensity loss, gradient loss, spectral loss, and multi-scale perceptual loss, respectively constraining the reconstruction quality of the model from multiple dimensions such as intensity distribution, gradient preservation, spectral characteristics, and multi-scale spatial structure, significantly improving the stability of model training and final performance.

[0106] Table 3 compares the accuracy performance of the WindUNet-AF² model on the test set using the multi-stage adaptive loss function (the algorithm of this invention) and using only the weighted intensity loss function. The results show that the multi-stage adaptive hybrid loss framework can avoid the suboptimal convergence problem easily caused by the traditional fixed weight loss, effectively reduce the reconstruction error, and improve the accuracy of wind speed reconstruction.

[0107] Table 3. Accuracy comparison of models using the multi-stage adaptive loss function (the algorithm of this invention) and the weighted intensity loss function on the test set.

[0108]

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, alterations, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reanalysis product revision and resolution enhancement of sea surface high wind speed based on a bimodal U-Net, characterized in that, Includes the following steps: (1) Obtain wind speed data from STAR-SAR and ECMWF-ERA5 and preprocess them. Using the location of the tropical cyclone center on the SAR wind speed image as a reference, correct the ERA5 wind speed image and then crop it into wind speed sub-images. Use the obtained ERA5 wind speed sub-images as the input low-resolution dataset and the SAR wind speed sub-images as the labels for the tropical cyclone correction network and reconstruction network. (2) Construct a dual-stage, dual-modal tropical cyclone wind speed reconstruction system containing a correction network and a super-resolution network to reconstruct high-precision wind speed from low-resolution ERA5 wind speed data, including a correction stage and a reconstruction stage; the correction network adopts a U-Net network architecture with an attention gating mechanism for the early correction of ERA5 wind speed data; the reconstruction stage adopts an attention-enhanced U-Net network with dual-modal input features, firstly extracts features from the input corrected ERA5 and original ERA5 dual-channel wind speed data, then processes them using three feature fusion strategies, and finally enhances the feature extraction capability through the residual convolution module in the super-resolution network, and finally maps the feature map to the single-channel wind speed output through two layers of convolution; The correction phase employs a hybrid intensity loss function, which is used to improve the accuracy of ERA5 wind speed data correction. The formula for this function is as follows: ； in, The pixels in a SAR wind speed image at a resolution of 25 km. represents the pixels in the corrected ERA5 wind speed image, where n is the total number of pixels; The reconstruction stage employs an adaptive hybrid loss function, which consists of weighted intensity loss, gradient loss, spectral loss, and multi-scale perceptual loss, and its formula is as follows: ； Where, λ I Let λ be the weighted intensity loss function. G Let λ be the gradient loss function. S Let λ be the spectral loss function. P For multi-scale sensing loss function, L I , L G , L S , L P These correspond to the weights of the four loss functions.

2. The method according to claim 1, characterized in that, The weighted intensity loss improves the accuracy of the reconstructed wind speed data through dynamic weight adjustment, and its formula is as follows: ； in, , The label indicates the wind speed. This represents the wind speed after super-resolution, where n is the total number of pixels. The gradient loss employs the Gaussian-smoothed Sobel gradient operator to maintain spatial continuity, and its formula is as follows: ； in, , , Indicates Gaussian filtering. Indicates the calculation of the gradient; The spectral loss employs a radial weighting strategy in the frequency domain to ensure accurate reconstruction of the tropical cyclone eye circulation structure. The formula is as follows: ； in, , ; i , j These are the row and column numbers corresponding to the image's frequency domain space, respectively. ic and jc The row and column numbers of the center pixel. The maximum distance from the center of the image. r The distance from the image center is denoted by FFT(·), and FFT(·) is the Fourier transform. The multi-scale perceptual loss captures feature similarity at different resolutions through three levels of average pooling: original scale, 2x downsampling, and 4x downsampling. The weight allocation is 1:2:4 to ensure the reconstruction details and quality of the image. The formula is as follows: ； in, , , , This is for average pooling operations.

3. The method according to claim 1, characterized in that, The reconstruction phase adopts an adaptive weight scheduling mechanism based on training progress, which is divided into four phases: the first phase focuses on intensity loss in the first 30% of epochs; the second phase gradually introduces gradient loss in the 30-50% of epochs; the third phase gradually adds spectral loss in the 50-80% of epochs; and the fourth phase optimizes all loss functions in the 80-100% of epochs.

4. The method according to claim 1, characterized in that, In step (2), the dual-modal input feature attention-enhanced U-Net network used in the reconstruction stage includes a feature extraction module, a feature fusion module, and a super-resolution network module. The feature extraction module extracts features from the dual-channel wind speed data of the corrected ERA5 and the original ERA5 to obtain two features F1 and F2. Then, a multi-strategy feature fusion module containing channel splicing fusion, adaptive weighted fusion, and residual fusion is used to process the data, integrating the structured information and original texture details of the corrected image and the original image to prevent information loss during feature extraction and fusion. The super-resolution network adopts a progressive upsampling strategy and enhances the feature extraction capability through three residual convolution modules containing CBAM attention mechanism.

5. The method according to claim 1, characterized in that, In step (1), the ERA5 input data is standardized using the Min-Max standardization method, and the formula is: ； in x , These are the values ​​before and after the transformation, min( x ) and max( x These represent the minimum and maximum values ​​of the data, respectively. After data standardization, the low-resolution dataset is divided into a training set and a test set. The test set is used to test the model's wind speed reconstruction effect, and 20% of the data in the training set is used as a validation set to verify whether the model has good convergence. The same processing method is used for the SAR image label dataset.

Images