SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet

Abstract

The present invention discloses a SAR sea surface wind speed inversion method based on multi-attention-global enhancement UNet, comprising the following steps: (1) obtaining wind speed data observed by a buoy, cleaning and preprocessing it, and making a label data set; (2) obtaining dual-polarization ground-range multi-view SAR data collected in an interferometric wide-band mode (IW); (3) preprocessing the SAR data; selecting SAR data corresponding to the buoy data information. The preprocessed SAR data set is divided into a training set, a test set, and a validation set; (4) constructing a multi-attention-global enhancement UNet model, namely a CPA-KiTUNet model, and training it to obtain a sea surface wind speed inversion model, selecting a test set to test on the trained inversion model, and obtaining an inverted wind speed result; the present invention improves the richness of the inversion data and can accurately perform wind speed inversion in complex scenarios.

Description

Technical Field

[0001] The present invention relates to the field of ocean satellite inversion technology, and in particular to a SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet. Background Art

[0002] Synthetic Aperture Radar (SAR) technology has a wide range of applications in ocean monitoring, especially in the inversion of sea surface wind speed. Since SAR can obtain ocean surface information under all-weather and all-time conditions, it has become one of the main means of wind field inversion. Traditional SAR sea surface wind speed inversion methods mainly rely on scatterometers and physical models. The geophysical model function (GMF) is established by analyzing the relationship between the backscattering cross section of the SAR image and the sea surface wind speed to invert the wind speed. In addition, the physical model requires wind direction information as an auxiliary input, which directly affects the inversion accuracy of the wind speed, and the physical model requires complex parameter calculation and selection. Due to the complexity of the marine environment, these models are easily affected by changes in the external environment and it is difficult to accurately capture the changes in wind speed under different conditions, resulting in low inversion accuracy.

[0003] In recent years, the development of deep learning technology has provided new solutions for SAR wind speed inversion. In particular, the UNet model based on convolutional neural network (CNN) can extract the feature information of SAR images to achieve sea surface wind speed inversion due to its good feature extraction ability. Convolution in UNet can extract local features, but cannot capture the long-range dependencies between different features. The jump connection simply connects the encoder and decoder, causing a semantic gap between the encoder and decoder. The SwinUNet model based on the sliding window self-attention mechanism captures more global information through local windows and hierarchical feature extraction, while avoiding the limitations of traditional convolution and enhancing the ability to extract spatial information. However, there are still some deficiencies in the integration of cross-scale features. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet, so as to solve the problems that traditional convolution is limited in extracting global information, the extraction of multi-scale features is insufficient, and the semantic gap between the encoder and the decoder exists.

[0005] Technical solution: The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet described in the present invention comprises the following steps:

[0006] (1) Obtain wind speed data observed by buoys, clean and preprocess them, and create a labeled data set;

[0007] (2) Obtain dual-polarization ground-range multi-look SAR data collected in interferometric wide-swath mode (IW);

[0008] (3) Preprocess the SAR data; select the SAR data corresponding to the buoy data information. Divide the preprocessed SAR data set into a training set, a test set, and a validation set;

[0009] (4) Construct a multi-attention-global enhancement UNet model, namely the CPA-KiTUNet model, and train it to obtain a sea surface wind speed inversion model. Select a test set to test the trained inversion model to obtain the inverted wind speed result.

[0010] Furthermore, in step (1), the preprocessing is as follows: outliers, missing values ​​and noise data are removed, and the buoy wind speed data at different heights are converted to the standard 10m height wind speed through the wind shear formula; the different heights include: 2m or 3m.

[0011] Furthermore, in step (3), the preprocessing is specifically as follows: thermal noise removal, radiation calibration, filtering, terrain correction, sea and land masking, and decibelization.

[0012] Furthermore, in step (3), the preprocessed SAR image is synthesized with VV polarization, VH polarization and incident angle as the RGB three channels. The synthesized image is cropped with a 256*256 pixel sub-image centered on the buoy coordinates, and then spatiotemporally matched with the buoy wind speed label data above j to obtain the sea surface wind speed inversion dataset.

[0013] Furthermore, step (4) is as follows: the CPA-KiTUNet inversion model consists of an encoder, a skip connection, a bottleneck and a decoder; the encoder includes a residual module, a pyramid segmentation attention module and a maximum pooling; the skip connection consists of a context anchor attention module; the bottleneck includes a KAN-based Vision Transformer module; the decoder includes a transposed convolution, a residual model and a pyramid segmentation attention module; the pyramid segmentation attention is introduced in the encoder stage to extract feature information of different scales; the context anchor attention is introduced at the skip connection to extract context information; the transposed convolution and pyramid segmentation attention are introduced in the decoder stage; the KAN-based Vision Transformer module is introduced in the bottleneck between the encoder and the decoder to extract global information.

[0014] Furthermore, the encoder is as follows: a pyramid segmentation attention mechanism is adopted to enhance the model to extract SAR data of information of different scales; the encoder extracts features of different scales in the image through multiple residual convolution layers; the encoder introduces a pyramid segmentation attention module after each residual convolution layer; the residual convolution includes a convolution with a 1×1 convolution kernel, two groups of convolution layers with 3×3 convolution kernels and a ReLU activation function; the original image is respectively subjected to 1×1 convolution and 3×3 convolution groups for residual summation and then input into the pyramid segmentation attention module; among them, the pyramid segmentation attention module includes a segmentation and connection module SPC, a squeeze and excitation module SE and a Softmax layer; firstly, the feature map of H×W×C is multi-scale segmented into S groups by the segmentation and connection module SPC in the channel dimension, and each group has C / S units. Channel, each set of features is convolved at different scales to obtain feature information of different scales, and the processed multiple sets of multi-scale features are spliced ​​again in the channel dimension; the feature map after the SPC module is then passed through the squeeze and excitation module SE using global average pooling to obtain the channel statistics of the input feature map, and then the attention weight of each channel is generated through the fully connected layer and activation function; then the original attention weight output by the SE module is normalized through Softmax to obtain the attention weight result; then the recalibrated weight and the corresponding feature pattern element are dot-multiplied to obtain the feature map after multi-scale feature information attention weighting; finally, the feature map obtained by the maximum pooling with a step size of 2 and a kernel size of 2×2 is downsampled layer by layer to obtain a feature map with a size of half that of the input feature map.

[0015] Furthermore, the skip connection is as follows: the features after residual convolution and PSA module in the encoder are input into the context anchor attention CAA module; the CAA module calculates the attention coefficient through average pooling, 1x1 convolution, horizontal and vertical convolution and Sigmoid activation function; the input features are first compressed by the average pooling layer for global information, and then the number of channels is adjusted by the 1x1 convolution module in turn, and the horizontal and vertical convolution modules are convolved in the horizontal and vertical directions respectively, and finally the attention coefficient is generated by the Sigmoid activation function; the generated attention coefficient is multiplied by the input feature to obtain the enhanced feature map, and passed to the decoder.

[0016] Furthermore, the bottleneck is as follows: the feature map output by the last layer of the encoder is input into the bottleneck, and after residual convolution and pyramid segmentation attention, the feature map is input into KAN-ViT; Among them, the MLP layer in ViT is replaced by KAN. The KAN-ViT module contains a position encoding layer Embedding, a multi-head self-attention layer MHSA and a KAN module. The details are as follows: The Embedding layer serializes the feature map and adds position encoding. The position-encoded feature sequence is input into 5 groups of Transformer encoders. Each group of Transformer contains a multi-head self-attention layer and a KAN module. In the multi-head self-attention layer, each Patch embedding is mapped to a query Q, a key K, and a value V vector to calculate the global dependency between patches. Eight self-attention heads are used to extract different feature patterns in SAR data through the interaction of multi-scale features to obtain global information. After multi-head self-attention, the feature sequence enters KAN. KAN inputs the input features into the basic linear layer, calculates the basic weights, and then calculates the spline weights through the B-spilne basis function. The calculated spline basis function and the spline weights are linearly transformed to form a spline output. Finally, the basic output and the spline output are added to obtain the final feature map output.

[0017] Furthermore, the decoder is specifically as follows: the feature map output at the bottleneck is input into the decoder, and the decoder contains multiple layers of upsampling and feature splicing; the upsampling is performed through 2×2 transposed convolution to expand the spatial dimension of the feature map, and then feature splicing is performed with the feature map transmitted by the jump connection, and then it passes through the same residual convolution and pyramid segmentation attention module as the encoder, and is transmitted layer by layer to the next upsampling module.

[0018] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, any one of the SAR sea surface wind speed inversion methods based on multi-attention-global enhanced UNet is implemented.

[0019] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: By replacing the convolutional blocks in the traditional UNet with the residual module and the pyramid segmentation attention (PSA) module, the present method can effectively extract multi-scale information of wind speed from the input features. PSA applies the attention mechanism at different scales, which enhances the model's extraction of features of different scales, enables the model to focus on global and local features at the same time, and significantly improves the ability to understand complex scenes. The introduction of KAN-ViT enables the model to better capture long-range dependencies, thereby enhancing the feature representation capability, while replacing the MLP module in the traditional VisionTransformer with a more explanatory KAN module; the contextual anchor attention (CAA) module is introduced on the jump connection, and the contextual information in the horizontal and vertical directions is effectively extracted through the combination of average pooling and directional convolution; the CAA module uses bidirectional convolution and attention mechanisms to significantly enhance the expressiveness of the feature map and achieve efficient feature enhancement and information transfer. This structural design can help the model better capture spatial dependencies, thereby improving the accuracy of feature extraction and enhancing the accuracy of wind speed inversion; intelligent inversion is achieved through SAR and buoy label data. This process suppresses the noise of SAR images and improves the richness of inversion data, enabling the model to accurately invert wind speed in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is the overall flow chart of the present invention;

[0021] Figure 2 It is a flow chart of SAR image preprocessing of the present invention;

[0022] Figure 3 This is the CPA-KiTUNet model structure diagram of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0024] like Figure 1 As shown, an embodiment of the present invention provides a SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet, comprising the following steps:

[0025] (1) Buoy data preprocessing: The acquired buoy data has a time resolution of 1 hour and includes sea surface wind speeds at 2m and 3m. By checking the maximum, minimum and median values ​​in each set of data, outliers are eliminated. Through gradient testing, based on the continuity of ocean observation data within a certain time range, the change values ​​of observation elements adjacent in time or position should be within a certain range, otherwise the data is considered abnormal. The specific method is: assuming that the current observation value is, the previous correct value adjacent in time is , the test formula is as follows:

[0026] ;

[0027] Among them, Vi represents the current observation value, Representation and The previous correct value adjacent in time, Hg represents the gradient test parameter, Take 20m / s.

[0028] Through the peak test, based on the ocean observation elements in a certain time range

[0029] If there is a peak that is significantly different from the surrounding observations, the data is considered abnormal. The specific method is: Assume that the current observation is , the first correct observation adjacent to its time is and , the test formula is as follows:

[0030] ;

[0031] in, represents the current observation value, Representation and The previous correct value adjacent in time, Representation and The next correct value in time, is the peak test parameter, Take 20m / s.

[0032] After removing abnormal data according to the above two methods, the wind profile algorithm is used to unify the 2m and 3m high wind speeds to the 10m high wind speed on the sea surface. The wind profile formula is as follows:

[0033] ;

[0034] in, Height from ground Wind speed at m; Height from ground Wind speed; is the ground condition factor. is the wind speed height conversion coefficient. When U>7m / s, Take 0.022, z take 2m, Take 1.36, z take 3m, Take 1.24; when U≤7m / s, Take 0.023, z take 2m, Take 1.23, z take 3m, Take 1.17;

[0035] Finally, the preprocessed buoy wind speed at a height of 10m above the sea surface is obtained as the label data set.

[0036] (2) SAR image data acquisition: The present invention uses dual-polarization (VV+VH) ground-range multi-look (GRD) SAR images from Sentinel-1 A and Sentinel-1 B in interferometric wide-swath mode (IW) working mode. The SAR image data volume in IW mode is the largest, and has good resolution and swath width. Dual-polarization (VV+VH) images not only improve the sensitivity to the scattering characteristics of the sea surface, making the analysis of complex sea conditions more accurate, but also enhance the recognition of changes in sea surface roughness through the cross-polarization (VH) channel, thereby improving the robustness of the inversion. GRD images are processed by multi-look, with lower noise and higher image quality.

[0037] (3) SAR image data preprocessing is extremely important for model inversion. Effective preprocessing can greatly suppress the noise content of SAR images, enhance the information utilization of SAR images, and improve the inversion accuracy. Figure 2 As shown, this method uses SNAP software for preprocessing, and the preprocessing process used is as follows: thermal noise removal, radiation calibration, filtering, terrain correction, sea and land masking, and decibel processing. It includes the following steps:

[0038] (31) Thermal noise removal: When the SAR satellite is working, it will generate a lot of internal heat when transmitting electromagnetic wave signals to the ground, which will affect the accuracy of the radar backscatter signal. Therefore, it is necessary to remove thermal noise from the image. This step can effectively eliminate system noise, improve the image signal-to-noise ratio, and make the subsequent wind speed inversion process more accurate.

[0039] (32) Radiometric calibration: The relationship between the SAR image and the backscatter coefficient of the ground target is established through absolute calibration, and the grayscale value of the image is converted into the normalized radar backscatter coefficient ( ) for radiometric calibration. Through radiometric calibration, the pixel values ​​of the original SAR image are converted into scattering coefficients with clear physical meanings, so as to better characterize the scattering characteristics of the sea surface. The formula is as follows:

[0040] ;

[0041] in, is the radar backscatter coefficient, i and j are the pixel coordinates in the image. is the grayscale value, that is, the amplitude value. is the calibration parameter. All are provided in the SAR image data files.

[0042] (33) Filtering: The speckle noise in SAR images is caused by the coherence of radar signals and appears as randomly distributed bright and dark spots in the image. This noise interferes with subsequent data analysis, especially in areas with low signal-to-noise ratio (such as the ocean surface). To reduce the speckle noise in the image, the Lee filtering method is applied. Lee filtering suppresses the speckle noise by weighted mean filtering, and the weights are calculated based on the variance distribution of pixel values ​​in the local window. The basic idea of ​​Lee filtering is to reduce the impact of noise by reducing the local mean difference. The formula is as follows:

[0043]

[0044] in, is the gray value after filtering, is the input grayscale value, is the average value of the pixels in the moving window. K is the weight.

[0045] (34) Terrain correction: Since SAR images may be geometrically distorted due to the influence of terrain undulations, terrain correction is required to ensure that the images are aligned with the real geographic coordinates. Especially in nearshore areas, terrain correction can greatly improve the accuracy of wind speed inversion. This method is based on the SRTM 3Sec GeoTiff digital elevation model and uses the three-dimensional information of the surface to calculate the real geographic location of each pixel in the SAR image. The orthorectification process projects the pixels into a plane coordinate system through geometric transformation, thereby ensuring that the image is accurately aligned with the geographic location.

[0046] (35) Land and sea mask: The backscatter characteristics of SAR images in land and ocean areas are significantly different. The complex terrain features on land will produce different scattering patterns, which can easily introduce errors. The land and sea mask can be used to exclude these land areas, ensuring that the model focuses on the inversion of sea surface wind speed.

[0047] (36) Decibel conversion: The backscatter coefficient after radiation calibration is converted to a logarithmic scale (dB) for decibel processing. The logarithmic transformation compresses large intensity differences into a smaller numerical range, while expanding the numerical distribution in the low-intensity range, making the details of the weak backscattering area more obvious, so that the subsequent feature extraction and modeling process can better process the data on a logarithmic scale. This processing step helps to enhance the recognition ability of weak scattering signals, thereby improving the sensitivity of the model to low wind speed ranges. The decibel conversion formula is as follows:

[0048] ;

[0049] in is the backscatter coefficient on a logarithmic scale, is the linear scale backscatter coefficient.

[0050] (37) Cropping and channel synthesis: The SAR image is centered on the buoy and the image sub-block of size 256×256 is cropped. The VV polarization, VH polarization and incident angle images are used as the R, G and B channels of the RGB image for RGB image synthesis. Cropping centered on the buoy ensures that the sea surface area around the buoy is accurately covered in the image, and the wind speed characteristics around the buoy have a direct impact on the inverted wind speed. The 256×256 cropping size provides sufficient resolution while covering a suitable range around the buoy, so that SAR details such as the scattering pattern and wave characteristics around the buoy are well preserved. These features directly reflect the changes in local sea surface wind speed and are very important for fine feature extraction in model training. The VV and VH polarization channels capture the horizontal and vertical scattering characteristics respectively, while the incident angle channel provides additional geometric information. The three are synthesized into an RGB image, allowing the model to use these three types of information for analysis at the same time.

[0051] (38) Constructing a dataset: The cropped and synthesized RGB images are time- and space-matched with the corresponding central buoys they cover; and the high sea condition wind speed data with buoy wind speeds above 10 m / s are screened out to construct a sea surface wind speed inversion dataset. The time range of the inversion dataset is from January 16, 2015 to June 3, 2023.

[0052] (4) Construct a multi-attention global enhancement network model, namely the CPA-KiTUNet model, and train it; use the training set to train the sea surface wind speed prediction data, and select the model with the best inversion accuracy in the validation set to test it on the test set. Figure 3 As shown in the figure, the details are as follows: The CPA-KiTUNet inversion model adopts the basic architecture of UNet's encoding, skip connection and decoding. In the encoder stage, pyramid segmentation attention is introduced to extract information of different scales; context anchor point attention is introduced at the skip connection to extract context information; transposed convolution and pyramid segmentation attention are introduced at the decoder stage. VisionTransformer, which replaces MLP with KAN, is introduced at the bottleneck between the encoder and decoder to enhance global information extraction. It includes the following steps:

[0053] (41) Encoding: A pyramid segmentation attention mechanism is used in the encoder stage to enhance the model's ability to extract information of different scales and overcome the limitations of traditional convolutional neural networks in capturing global information. The encoder includes multiple residual convolutional layers, which extract features of different scales in SAR images by deepening the convolution operation layer by layer. In order to capture feature information of different scales, the encoder introduces a pyramid segmentation attention module after each residual convolutional layer. The residual convolution consists of a convolution with a 1×1 convolution kernel, two groups of convolutional layers with 3×3 convolution kernels and a ReLU activation function; the original image is summed up after the residuals of the 1×1 convolution and the 3×3 convolution groups are respectively input into the pyramid segmentation attention module. The pyramid segmentation attention module includes a segmentation and connection module (SPC), a squeeze and excitation module (SE) and a Softmax layer. First, the H×W×C feature map is split into S groups in multi-scale by the segmentation and connection module (SPC) in the channel dimension, each group has C / S channels, and each group of features is grouped and convolved at different scales to obtain feature information at different scales. The processed multiple groups of multi-scale features are spliced ​​again in the channel dimension. The formula is as follows:

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] in, is the input feature; for The sub-blocks after segmentation; They are The corresponding group convolution kernel size and group size. represents the feature map obtained by the i-th sub-block, Represents the feature map after multi-scale fusion.

[0059] The squeeze and excitation module (SE) uses global average pooling to obtain the channel statistics of the input feature map, and then generates the attention weight of each channel through the fully connected layer and activation functions (such as ReLU and Sigmoid). The original attention weight output by the SE module is then normalized by Softmax to obtain the attention weight. The recalibrated weight is then multiplied by the corresponding feature pattern element to obtain the feature map after the multi-scale feature information attention weighting. The formula is as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] in, Representing feature maps of different scales Extracted channel attention weights; Represents the entire multi-scale channel attention weight vector; Represents the result of recalibration of the channel attention information weights; Represents the result of dot multiplication of the recalibrated weights and the input feature map; Represents a multi-scale channel attention weighted feature map.

[0066] Finally, the feature map is downsampled by maximum pooling with a stride of 2 and a kernel size of 2×2, so that the size of the feature map becomes half of the input feature map.

[0067] (42) Skip connection: The features after residual convolution and PSA module in the encoder are input into the contextual anchor attention (CAA) module. The CAA module contains an average pooling layer, a 1x1 convolution module, horizontal and vertical convolution modules, and a sigmoid activation function. The input features are first compressed globally through a 7x7 average pooling layer with a step size of 1 and a padding of 3; then the number of channels is adjusted through a 1x1 convolution.

[0068] The average pooling formula is as follows:

[0069] ;

[0070] in, Represents the feature map output by the i-th encoder.

[0071] Then, feature extraction is performed through convolution in the horizontal direction (1×h_kernel_size) and the vertical direction (v_kernel_size×1).

[0072] The horizontal convolution formula is as follows:

[0073] ;

[0074] in is the average pooling convolution output; The convolution kernel size is Horizontal convolution; Represents the horizontal feature extraction output.

[0075] The vertical convolution formula is as follows:

[0076] ;

[0077] in Represents the horizontal feature extraction output; The convolution kernel size is Vertical convolution; Represents the vertical feature extraction output.

[0078] Finally, the attention coefficient is generated through 1x1 convolution and Sigmoid activation function. The generated attention coefficient is multiplied by the input feature and connected to the decoder. The formula is as follows:

[0079] ;

[0080] ;

[0081] ;

[0082] in The spatial feature output is extracted through horizontal and vertical convolution; Represents the output feature map of 1×1 convolution adjusting the number of channels; For the Sutra Calculate the output attention coefficient, is the original feature map and The dot product outputs the final feature map.

[0083] (43) Bottleneck: The feature map output by the last layer of the encoder is input into the bottleneck. After residual convolution and pyramid segmentation attention, the feature map is input into the KAN-Vision Transformer (KAN-ViT). In order to further improve the expressive power of the model, the MLP layer in ViT is replaced by Kolmogorov-Arnold Networks (KAN). The KAN-ViT module contains a position encoding layer (Embedding), a multi-head self-attention layer (MHSA) and a KAN module. The details are as follows: The Embedding layer serializes the feature map and adds position encoding. The feature sequence after position encoding is input into 5 groups of Transformer encoders. The formula is as follows:

[0084] ;

[0085] in represents the i-th image block of the serialized feature map, represents the linear projection matrix, represents the position encoding matrix.

[0086] Each group of Transformers contains a multi-head self-attention layer and a KAN module. In the multi-head self-attention layer, each patch embedding is mapped to a query (Q), key (K), and value (V) vector to calculate the global dependency between patches. Eight self-attention heads are used to extract different feature patterns in SAR images through the interaction of multi-scale features to obtain global information. After multi-head self-attention, the feature sequence enters the Kolmogorov-Arnold Networks (KAN) to replace the MLP layer in the traditional Transformer; KAN calculates the basic weights of the input features through the basic linear layer, and then calculates the spline weights through the B-spilne basis function, and generates rich feature representations using the local information of the input features. These spline basis functions can capture the changing characteristics of the input features at different locations. The calculated spline basis functions and spline weights are linearly transformed to form spline outputs, which can capture the subtle changes between features. Finally, the basic output and the spline output are added to obtain the final feature map output. This output combines linear and nonlinear transformations to ensure the diversity and richness of the model when extracting information. The formula is as follows:

[0087] ;

[0088] ;

[0089] ;

[0090] Among them, MSA represents the multi-head self-attention mechanism. represents normalization; Represents the multi-head self-attention output feature map; represents the KAN-ViT output of the i-th layer; Represents the function matrix corresponding to the i-th KAN layer

[0091] (44) Decoding: The feature map output at the bottleneck is input to the decoder, which contains multiple layers of upsampling and feature concatenation. Upsampling is performed through 2×2 transposed convolution to expand the spatial dimension of the feature map, and then feature concatenation is performed with the feature map transmitted by the jump connection. After that, it is transmitted layer by layer to the next upsampling module after passing through the same residual convolution and pyramid segmentation attention module as the encoder.

[0092] (45) The training set is used for training to obtain the sea surface wind speed inversion model, and the test set is selected to test the trained inversion model. The test results are compared with the buoy wind speed labels to obtain the inversion accuracy evaluation. The relevant evaluation indicators are as follows: two evaluation indicators, mean-square error (MSE) and mean absolute error (MAE), are selected, and the formulas of each indicator are as follows:

[0093] ;

[0094] ;

[0095] in, is the model prediction value, is the buoy wind speed value, and n is the number of samples.

[0096] In order to further analyze the inversion capability of the proposed CPA-KiTUNet model, the present invention selected SwinUNet and TransUNet for comparison with this model. SAR images of the southeastern coastal waters in the past 10 years with wind speeds above 10m / s were selected for training and the training set, validation set and test set were divided into 8:1:1 for inversion. The results are shown in Table 1.

[0097] Table 1 Experimental results of each model on the test set

[0098] ;

[0099] The results show that under high sea conditions, the accuracy of sea surface wind speed inverted from SAR image data by the present invention is significantly better than other models, and the inversion performance is the best.

[0100] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the computer program implements any one of the SAR sea surface wind speed inversion methods based on multiple attention-global enhanced UNet.

Claims

1. A SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet, characterized in that: The following steps are involved: (1) Obtain wind speed data observed by buoys, clean and preprocess them, and create a labeled data set; (2) Obtain dual-polarization ground-range multi-look SAR data collected in interferometric wide-swath mode (IW); (3) Preprocessing of SAR data; Select the SAR data corresponding to the buoy data information; The preprocessed SAR data set is divided into training set, test set and validation set; (4) Construct a multi-attention-global enhancement UNet model, namely the CPA-KiTUNet model, and train it to obtain a sea surface wind speed inversion model. Select a test set to test the trained inversion model to obtain the inverted wind speed result. The CPA-KiTUNet inversion model consists of an encoder, skip connections, bottlenecks, and a decoder; the encoder includes a residual module, a pyramid segmentation attention module, and a maximum pooling module; The skip connection consists of a contextual anchor attention module; the bottleneck includes a KAN-based Vision Transformer module; the decoder includes a transposed convolution, a residual model, and a pyramid segmentation attention module; Pyramid segmentation attention is introduced in the encoder stage to extract feature information of different scales; context anchor attention is introduced at the jump connection to extract context information; transposed convolution and pyramid segmentation attention are introduced in the decoder stage; and the KAN-based Vision Transformer module is introduced at the bottleneck between the encoder and decoder to extract global information.

2. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 1 is characterized in that: In step (1), the preprocessing is as follows: outliers, missing values ​​and noise data are removed, and the buoy wind speed data at different heights are converted to the standard 10m height wind speed through the wind shear formula; the different heights include: 2m or 3m.

3. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 1 is characterized in that: In step (3), the preprocessing is as follows: thermal noise removal, radiation calibration, filtering, terrain correction, sea and land masking, and decibel conversion.

4. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 3 is characterized in that: In step (3), the preprocessed SAR image is synthesized with VV polarization, VH polarization and incident angle as the RGB three channels respectively. After the synthesized image, a sub-image of 256*256 pixels is cropped with the buoy coordinates as the center, and the sub-image is temporally and spatially matched with the buoy wind speed label data under higher sea conditions of more than 10 m / s to obtain the sea surface wind speed inversion dataset.

5. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 1 is characterized in that: The encoder is as follows: a pyramid segmentation attention mechanism is used to enhance the model to extract SAR data of information at different scales; the encoder extracts features of different scales in the image through multiple residual convolution layers; the encoder introduces a pyramid segmentation attention module after each residual convolution layer; the residual convolution includes a convolution with a 1×1 convolution kernel, two groups of convolution layers with 3×3 convolution kernels and a ReLU activation function; the original image is summed up after the residuals of the 1×1 convolution and the 3×3 convolution groups are respectively input into the pyramid segmentation attention module; among them, the pyramid segmentation attention module includes a segmentation and connection module SPC, a squeeze and excitation module SE and a Softmax layer; first, the H×W×C feature map is multi-scale segmented into S groups by the segmentation and connection module SPC in the channel dimension, and each group has C / S channels. , each set of features is convolved at different scales to obtain feature information at different scales, and the processed sets of multi-scale features are spliced ​​again in the channel dimension; the feature map after the SPC module is then passed through the squeeze and excitation module SE using global average pooling to obtain the channel statistics of the input feature map, and then the attention weight of each channel is generated through the fully connected layer and activation function; then the original attention weight output by the SE module is normalized by Softmax to obtain the attention weight result; then the recalibrated weights and the corresponding feature pattern elements are dot-multiplied to obtain the feature map after the multi-scale feature information attention weighting; finally, the feature map obtained by the maximum pooling with a step size of 2 and a kernel size of 2×2 is downsampled layer by layer to obtain a feature map with a size of half that of the input feature map.

6. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 1 is characterized in that: The skip connection is as follows: the features after residual convolution and PSA module in the encoder are input into the context anchor attention CAA module; the CAA module calculates the attention coefficient through average pooling, 1x1 convolution, horizontal and vertical convolution and Sigmoid activation function; the input features are first compressed by the average pooling layer for global information, and then the number of channels is adjusted by the 1x1 convolution module in turn, and the horizontal and vertical convolution modules are convolved in the horizontal and vertical directions respectively, and finally the attention coefficient is generated by the Sigmoid activation function; the generated attention coefficient is multiplied by the input feature to obtain the enhanced feature map, which is passed to the decoder.

7. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 1 is characterized in that: The bottleneck is as follows: the feature map output by the last layer of the encoder is input into the bottleneck. After residual convolution and pyramid segmentation attention, the feature map is input into KAN-ViT. Among them, the MLP layer in ViT is replaced by KAN. The KAN-ViT module contains a position encoding layer Embedding, a multi-head self-attention layer MHSA and a KAN module. Specifically, the Embedding layer serializes the feature map and adds position encoding. The position-encoded feature sequence is input into 5 groups of Transformer encoders. Each group of Transformer contains a multi-head self-attention layer and a KAN module. In the multi-head self-attention layer, each Patch embedding is mapped to a query Q, a key K, and a value V vector to calculate the global dependency between patches. Eight self-attention heads are used to extract different feature patterns in SAR data through the interaction of multi-scale features to obtain global information. After multi-head self-attention, the feature sequence enters KAN. KAN inputs the input features into the basic linear layer, calculates the basic weights, and then calculates the spline weights through the B-spilne basis function. The calculated spline basis function and the spline weights are linearly transformed to form the spline output. Finally, the basic output and the spline output are added to obtain the final feature map output.

8. The SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet according to claim 1 is characterized in that: The decoder is as follows: the feature map output at the bottleneck is input into the decoder, which contains multi-layer upsampling and feature concatenation; The upsampling is performed through 2×2 transposed convolution to expand the spatial dimension of the feature map, and then the feature map transmitted by the jump connection is concatenated. It is then transmitted layer by layer to the next upsampling module through the same residual convolution and pyramid segmentation attention module as the encoder.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the SAR sea surface wind speed inversion method based on multiple attention-global enhanced UNet is implemented according to any one of claims 1 to 8.

Images