Super-resolution method for ocean vortex chlorophyll remote sensing data

Abstract

The invention discloses a super-resolution method for ocean vortex chlorophyll remote sensing data, and belongs to the technical field of remote sensing image super-resolution. The invention aims to solve the problems of shallow feature information loss and insufficient multi-scale information utilization of the existing blind super-division network in feature fusion for ocean vortex remote sensing data. A low-resolution CHL image is input into an image blind super-resolution network model core for fitting reconstruction, a blind super-resolution network comprises a kernel prediction unit and an image restoration unit, the kernel prediction unit receives the CHL image to predict a simplified degradation kernel, and the image restoration unit receives information of the simplified degradation kernel and the CHL image for super-resolution reconstruction and outputs a super-resolution image. In the blind super-resolution network training process, a neural network with the input being the CHL image and the output being the vorticity estimation value is embedded at the tail end of an image recovery unit to assist training of the blind super-resolution network, vorticity consistency constraint can be effectively fused in the ocean vortex image super-resolution method, and therefore the reconstruction effect is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image super-resolution, and relates to a remote sensing chlorophyll image blind super-resolution method. BACKGROUND

[0002] It is of great significance to realize ocean vortex data super-resolution reconstruction by using remote sensing images in the fields of people's livelihood and national defense. Whether for vortex detection tasks or segmentation tasks, the application of higher resolution images in training will inevitably improve the task accuracy, which plays a significant role in both the risk planning of ship routes and the prediction analysis of ocean currents. However, due to the particularity of marine remote sensing data, the real ocean situation is complex, and it is difficult for existing methods to restore accurate information when performing super-resolution reconstruction.

[0003] Traditional super-resolution algorithms mostly use general image degradation models, which are not suitable for marine remote sensing data, resulting in a lack of authenticity in the reconstruction results and difficulty in meeting the needs. The existing blind super-resolution method can learn the degradation kernel from the LR data, which to some extent overcomes the problem of lack of real and effective LR-HR matching samples, but when facing marine images, it still lacks some authenticity. This is because although it has learned some degradation relationship, it has not established an explicit association between the super-resolution image and the key parameters of the ocean vortex in the original data. SUMMARY

[0004] The present application is to solve the problem of shallow feature information loss and insufficient multi-scale information utilization in the existing blind super-resolution network for marine vortex remote sensing data.

[0005] A super-resolution method for marine vortex chlorophyll remote sensing data, comprising the following steps:

[0006] inputting a low-resolution CHL image into an image blind super-resolution network model to obtain a super-resolution image for marine vortex chlorophyll remote sensing data, the CHL image being a chlorophyll concentration image; the blind super-resolution network comprising a kernel prediction unit and an image restoration unit, the kernel prediction unit receiving the CHL image to predict a simplified degradation kernel, and the image restoration unit receiving the simplified degradation kernel information and the CHL image to perform super-resolution reconstruction and output a super-resolution image.

[0007] Further, the kernel prediction unit first convolves the input LR image to obtain a first LR feature f L ′, which is then sent to a feature interweaving multi-scale enhancement subunit for processing, and the output is subjected to global average pooling to obtain a simplified degradation kernel; the LR image being a low-resolution CHL image.

[0008] Further, the feature interweaving multi-scale enhancement subunit comprises a plurality of feature interweaving multi-scale enhancement modules FIME connected in series.

[0009] The feature interweaving multi-scale enhancement module FIME first utilizes a feature concatenation layer to concatenate the first LR feature f L and the first conditional feature f c in the channel dimension to obtain f fusion , wherein the first conditional feature f c is obtained by convolution on the SR image obtained by the image restoration unit; the multi-scale spatial feature fusion submodule MFE is utilized to perform parallel depth separable convolution on f fusion , and the outputs are concatenated in the channel dimension to obtain the multi-scale feature f depth ; then the multi-scale feature f depth is input into the spatial attention module to obtain the multi-scale spatial fusion feature f s-attention , and the channel feature fusion submodule is utilized to obtain the channel fusion feature f depthwise ; and then f out is obtained as f depthwise +f fusion .

[0010] Further, the channel feature fusion submodule includes a channel attention module and a point-wise convolution layer, and the multi-scale spatial fusion feature f s-attention is first input into the channel attention module to integrate channel information; the point-wise convolution uses a 1*1 convolution kernel to perform convolution on the input to output the channel fusion feature f depthwise .

[0011] Further, the image restoration unit first utilizes a convolution layer to perform convolution on the LR image to obtain the second LR feature f L ; the image restoration unit simultaneously receives the reduced degradation kernel output by the kernel prediction unit, and each element in the reduced degradation kernel is repeatedly filled into a corresponding position of a two-dimensional matrix to form a feature map as the second conditional feature f c .

[0012] Then, the output is sent into the feature interweaving multi-scale enhancement submodule for processing, and the output is sent into the PixelShuffle layer, and the output of the PixelShuffle layer is subjected to a convolution layer to obtain the predicted SR image.

[0013] Further, in the process of obtaining the second LR feature f L by the image restoration unit, two convolution layers are utilized to perform convolution.

[0014] Further, the blind super-resolution network is pre-trained, and in the training process of the blind super-resolution network, a vorticity fitting network is embedded at the end of the image restoration unit to assist the training of the blind super-resolution network; the vorticity fitting network is a neural network with CHL images as input and vorticity estimation values as output.

[0015] In the process of blind super-resolution network training, L deg loss and reconstruction vorticity consistency loss L recvor to optimize the blind super-resolution network parameters; L deg The loss is y represents the SR image, x represents the LR image, k e represents the predicted degradation kernel, and the feature interleaved multi-scale enhancement subunit output of the kernel prediction unit is obtained through a convolution layer to obtain the degradation kernel k e ; ↓ s represents down-sampling; L recvor The loss is ζ S , ζ L are the predicted vorticity value obtained by the SR vorticity fitting network and the true vorticity value calculated by the SLA data corresponding to the LR image, respectively.

[0016] Further, the vorticity fitting network is pre-trained; the loss in the training process of the vorticity fitting network is as follows:

[0017] L vortex = α × MAE + β × Gradient_Loss

[0018] Wherein, α, β are weight coefficients for controlling the contribution of MAE and Gradient_Loss to the total loss, MAE is the mean absolute error, and Gradient_Loss is the vorticity gradient loss, and its formula is:

[0019]

[0020] Wherein, y ι is the simulated vorticity value, and y is the true vorticity value.

[0021] Further, the vorticity fitting network comprises a vorticity feature extraction module and a vorticity prediction module.

[0022] The vorticity feature extraction module comprises a first convolution layer, a maximum pooling layer, a plurality of residual block groups and a flattening layer, and a one-dimensional vector is obtained through the flattening layer.

[0023] The vorticity prediction module comprises a full connection layer, a bottom layer L1, a middle layer L2 and a high layer L3; the bottom layer L1 is a full connection layer based on β spline basis function; the middle layer L2 is a full connection layer based on ReLU activation function and β spline, wherein 50% nodes use β spline and 50% use ReLU function; the high layer L3 is a full connection layer composed of β spline function; the vorticity prediction module outputs the vorticity value of the CHL image.

[0024] Further, when the blind super-resolution network processes the input image for the first time in the training process, the image restoration module first receives the LR and the initial degradation kernel, generates an SR image, and then convolves the SR to obtain the first conditional feature f c ′, as the input of the first kernel prediction module;

[0025] The initial degradation kernel is σ x , σ y , and θ are parameters of the anisotropic Gaussian kernel; (i,j) represents the position in the degradation kernel.

[0026] Advantages:

[0027] The present application is aimed at the limitations of the existing ocean remote sensing image super-resolution reconstruction technology, and innovatively proposes a marine vortex chlorophyll remote sensing data super-resolution method fusing vorticity consistency constraint. The method first introduces the vorticity consistency principle, and realizes efficient reconstruction of marine vortex images by constructing a vorticity fitting module and a feature interwoven multi-scale enhancement module (FIME).

[0028] Since the current ocean remote sensing image super-resolution reconstruction cannot effectively establish the vorticity value correlation between the super-resolution image and the original image, the reconstructed image deviates from the physical properties and structural characteristics of the real marine vortex. The vorticity fitting module innovatively proposed by the present application first embeds the vorticity consistency principle into the super-resolution reconstruction process. In the training stage, only the LR image is used as the input, and the kernel prediction unit is used to learn the degradation kernel of the LR image. In the image recovery, the vorticity fitting module is used to fit the vorticity characteristics of the input remote sensing marine vortex image, to ensure the consistency of the vorticity values before and after super-resolution, so as to more accurately recover the marine vortex and other useful information, and provide high-quality image data for subsequent marine vortex detection and segmentation. The vorticity fitting module can perform good prediction performance on new and unseen data, has strong anti-overfitting ability, and to some extent provides a basis and explanation for model decision-making, and has good interpretability.

[0029] For marine vortex remote sensing data, the existing blind super-resolution network has the problems of loss of shallow feature information and insufficient utilization of multi-scale information in feature fusion, and the present application innovatively proposes a feature interwoven multi-scale enhancement module. The module effectively enhances the learning ability of the network to different scale information through the interweaving and fusion of multi-scale features. In the blind super-resolution network, the module deeply fuses the LR and SR features, and enhances the network's ability to capture key information such as marine vortex through multi-scale interweaving, significantly improving the accuracy and detail performance of the super-resolution image.

[0030] Using this method, the ocean remote sensing image can be restored more truly, and more useful information can be restored, which provides a new technical solution for the super-resolution reconstruction of the ocean remote sensing image, and can restore the ocean remote sensing image more truly. It is beneficial to subsequent image processing tasks such as detection and segmentation of ocean eddies, and is beneficial to many aspects of remote sensing applications such as ocean route planning and ocean current prediction. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 The super-resolution network structure for ocean eddy chlorophyll remote sensing data.

[0032] Figure 2 The structure diagram of the vorticity fitting network.

[0033] Figure 3 For a simulation example, the left image is an LR image, and the right image is an SR image. DETAILED DESCRIPTION

[0034] The embodiment is a super-resolution method for ocean eddy chlorophyll remote sensing data, including the following steps:

[0035] Step 1: Obtain the spatiotemporal corresponding SLA true vorticity value and CHL image multi-source remote sensing data, obtain CHL (chlorophyll concentration) data and public SLA (sea surface height anomaly) data from a public GOCI satellite platform;

[0036] The geostrophic eastward seawater velocity and the geostrophic northward seawater velocity are calculated by using the SLA and the geostrophic flow formula;

[0037] The geostrophic flow formula involves the Coriolis force caused by the earth rotation, and can be calculated by the following formula:

[0038]

[0039] Wherein, x is the longitude direction, y is the latitude direction, v is the northward velocity component (vgos), u is the eastward velocity component (ugos); g is the gravitational acceleration, about 9.81 m / s 2 ; f is the Coriolis parameter, f = 2ΩsinΦ, Ω is the earth rotation angular velocity, about 7.2921×10 -5 rad / s, Φ is the geographic latitude; h is the sea surface height anomaly (SLA);

[0040] Based on the vorticity (the curl of the velocity field) expression, the vorticity value is extracted from the SLA data; the expression of the vorticity in the two-dimensional horizontal plane is:

[0041]

[0042] Step two: input the obtained CHL image data into the vorticity fitting network (vorticity fitting module), the vorticity fitting network is a neural network, and the output is an approximate vorticity estimation value, that is, an approximate vorticity quantity; then the approximate vorticity quantity and the SLA true vorticity value are used to train the network, update the parameters, and the vorticity fitting network is as shown in Figure 2 After training, a vorticity fitting network that can obtain simulated vorticity values of the corresponding region through CHL images is obtained, and the network model parameters are frozen.

[0043] First, the low-resolution CHL image data obtained is used to train the vorticity fitting module, and the module that can fit the true vorticity value of the CHL image is trained. In the application, the vorticity fitting module in the blind super-resolution network is inputted with the SR image obtained by super-resolution, and the vorticity value fitted to the SR image is outputted.

[0044] As shown in Figure 2 , the vorticity fitting network includes a vorticity feature extraction module and a vorticity prediction module.

[0045] The vorticity feature extraction module includes a first convolutional layer, a maximum pooling layer, a plurality of residual block groups, and a flattening layer. In this embodiment, four residual block groups are provided, namely residual block group 1, residual block group 2, residual block group 3, and residual block group 4.

[0046] The first convolutional layer is a convolutional layer with a convolution kernel size of 7x7 and a step size of 2, which preliminarily extracts features and reduces the resolution through a large convolution kernel. The output is inputted into the maximum pooling layer.

[0047] The maximum pooling layer is a layer with a size of 3x3 and a step size of 2. The output is inputted into the residual block group 1.

[0048] The residual block group 1 is composed of two residual blocks, each of which contains two convolutional layers with a convolution kernel size of 3x3. The output is inputted into the residual block group 2.

[0049] The residual block group 2 is composed of two residual blocks, the first of which performs downsampling with a convolution kernel size of 3x3 and a step size of 2. The output is inputted into the residual block group 3.

[0050] The residual block group 3 is composed of two residual blocks, the first of which performs downsampling with a convolution kernel size of 3x3 and a step size of 2. The output is inputted into the residual block group 4.

[0051] The residual block group 4 is composed of two residual blocks with a convolution kernel size of 3x3 and a step size of 2. The output is converted into a one-dimensional vector through the flattening layer.

[0052] The vorticity prediction module comprises a full connection layer, a bottom layer L1, a middle layer L2 and a high layer L3, and the number of neurons of each layer is reduced by 30%-50% of the previous layer according to an exponential decay rule;

[0053] The bottom layer L1 is a full connection layer based on a beta spline basis function, the basis function is 3-order, and the grid number is 7, which is used to capture high-frequency details, and the output is input into L2;

[0054] The middle layer L2 is a full connection layer based on a ReLU activation function and a beta spline, wherein 50% of the nodes use the beta spline, and 50% use the ReLU function, and the output is input into L3;

[0055] The high layer L3 is a full connection layer composed of beta spline functions, the spline function is 5-order, and the grid number is 11, which is used to increase smoothness, and the output is the vorticity value of the obtained CHL image; the high layer L3 is initially provided with 5 uniformly distributed beta spline basis functions, and during training, dynamic grid refinement is used, and when the variance of the activation value of a certain interval exceeds a threshold value, a new node is automatically inserted;

[0056] When training the vorticity fitting network capable of simulating the vorticity value of the corresponding region, in order to minimize the difference between the simulated vorticity value and the real vorticity value as much as possible, a vorticity fitting loss is proposed, which is constrained by calculating the vorticity fitting loss of the simulated vorticity value and the real vorticity value of SLA, and the loss L vortex which can be expressed as:

[0057] L vortex =α×MAE+β×Gradient_Loss

[0058] Wherein, α, β are weight coefficients for controlling the contribution of MAE and Gradient_Loss to the total loss, MAE is the mean absolute error, and Gradient_Loss is the vorticity gradient loss, and its formula is:

[0059]

[0060] Where y ι is the simulated vorticity value, and y is the real vorticity value;

[0061] Step three: input the CHL image data obtained in step one into the image blind super-resolution network to train the network, the blind super-resolution network comprises a kernel prediction unit and an image recovery unit; the kernel prediction unit receives the CHL image as a LR image (low resolution image) to predict the degradation kernel, and the image recovery unit receives the degradation kernel information predicted by the kernel prediction unit and the CHL image to perform super-resolution reconstruction, and outputs the predicted SR image.

[0062] In the first processing, the image restoration module receives the LR and the initial degradation kernel, generates the SR image, and then convolves the SR to obtain the first conditional feature f c ′ as the input of the first kernel prediction module. The initial structure of the degradation kernel adopts an anisotropic Gaussian kernel, which can selectively blur according to the direction and shape of the edge, and is helpful to retain important edge features in the image and is closer to the degradation condition of the remote sensing ocean vortex image. The initial degradation kernel can be expressed as:

[0063]

[0064] where σ x , σ y , and θ are parameters of the anisotropic Gaussian kernel, used to simulate the blurring of the edge region; (i,j) represents the position in the degradation kernel.

[0065] The kernel prediction unit includes a convolution layer, a feature interweaving multi-scale enhancement subunit, and a global average pooling layer;

[0066] The convolution layer convolves the input LR image using a convolution kernel with a size of 3 to obtain the first LR feature f L ′;

[0067] The feature interweaving multi-scale enhancement subunit includes a plurality of feature interweaving multi-scale enhancement modules FIME (Feature Interweaving Multi-scale Enhancer) connected in series. The feature interweaving multi-scale enhancement module FIME includes a feature concatenation layer, a multi-scale spatial feature fusion sub-module MFE, and a channel feature fusion sub-module.

[0068] The feature concatenation layer concatenates the input first LR feature f L ′ and the first conditional feature f c ′ in the channel dimension, and the concatenation output of f L ′ and f c ′ is a concatenation feature f fusion ; wherein the concatenation feature f fusion can be expressed as:

[0069] f fusion =concat(f L ′,f c ′)

[0070] wherein the first conditional feature f c ′ is obtained by convolving the SR image obtained by the image restoration unit using a convolution kernel with a size of 9;

[0071] The multi-scale feature extraction submodule MFE includes three parallel depth-separable convolution layers, which are convolution kernels of 3×3, 5×5 and 7×7 sizes respectively. The depth-separable convolution is to extract the input feature map f fusion Convolution is performed on each channel of , using an independent convolution kernel for each channel without mixing between channels. The outputs of the three parallel depth-separable convolution layers are spliced ​​in the channel dimension to obtain the multi-scale feature f depth ; Multi-scale features f depth Input into the spatial attention module to obtain the multi-scale spatial fusion feature f s-attention ;

[0072] The channel feature fusion submodule includes a channel attention module and a point-by-point convolution layer, and a multi-scale spatial fusion feature f s-attention First, the input is sent to the channel attention module to integrate the channel information, and the point-by-point convolution is performed using a 1×1 convolution kernel to convolve the input and output the channel fusion feature f depthwise ;

[0073] The final output of FIME is f out =f depthwise +f fusion ;

[0074] During the training process, the output f out After the convolution kernel of size 3, the degradation kernel k is obtained e , used to calculate loss;

[0075] The global average pooling layer performs global average pooling on the final output of the feature interleaved multi-scale enhancement sub-unit to obtain a vector as a simplified degradation kernel.

[0076] The image restoration unit includes a convolution layer, a feature interleaved multi-scale enhancement subunit, and a PixelShuffle layer.

[0077] The image restoration unit convolves the LR image with two convolution kernels of size 3 and outputs the second LR feature f L The image restoration unit simultaneously receives the simplified degraded kernel output by the kernel prediction unit, and repeatedly fills each element in the simplified degraded kernel into the corresponding position of a two-dimensional matrix to form a feature map as the second conditional feature f c ;

[0078] It is then sent to the feature interleaved multi-scale enhancement subunit for processing, and the output is sent to the PixelShuffle layer. The output of the PixelShuffle layer passes through one convolutional layer to obtain the predicted SR image.

[0079] Step four: based on the training process of blind super-resolution network, the accuracy of the degradation kernel and the SR image is gradually improved by iterative optimization, the kernel prediction unit receives the SR output by the image restoration unit as the conditional feature for kernel prediction, and the image restoration unit receives the predicted degradation kernel of the kernel prediction unit as the conditional feature for super-resolution reconstruction.

[0080] In the training process of blind super-resolution network, the above trained vorticity fitting network is embedded at the end of the image restoration unit, the SR image is input into the vorticity fitting network, the predicted vorticity value of the SR image is obtained, and the L recvor loss is used to optimize the network parameters;

[0081] When training the blind super-resolution network, L deg loss is used to update the degradation kernel and network parameters, and L deg loss can be expressed as:

[0082]

[0083] Where y represents the SR image, x represents the LR image, k e represents the predicted degradation kernel Kernel estimated , s downsampling;

[0084] In order to minimize the difference between the vorticity value of the SR image and the original LR image, the difference between the vorticity value of the SR image and the real vorticity value of the LR image is calculated to make the reconstruction vorticity consistency loss, and the loss L recvor can be expressed as:

[0085]

[0086] Where ζ S and ζ L are the predicted vorticity value of the SR image through the vorticity fitting module and the real vorticity value calculated by the SLA data corresponding to the CHL image (LR), respectively.

[0087] L deg loss and reconstruction vorticity consistency loss L recvor are used to optimize the blind super-resolution network parameters. When the loss function meets the requirements, the kernel prediction unit outputs the final predicted degradation kernel, and the image restoration unit generates the SR image, completing the super-resolution of the marine vortex remote sensing image.

[0088] Step five: using the trained blind super-resolution network (without the vorticity fitting network embedded at the end during training) to realize the super-resolution of marine vortex chlorophyll remote sensing data, as shown in Figure 3 , where the left image is the LR image, and the right image is the SR image.

[0089] The above examples of the present application are only to illustrate the calculation model and calculation process of the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art, and it is impossible to enumerate all the embodiments here. Any obvious changes or variations derived from the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A super-resolution method for ocean eddy chlorophyll remote sensing data, characterized in that: The following steps are involved: The low-resolution CHL image is input into the image blind super-resolution network model to obtain a super-resolution image for ocean eddy chlorophyll remote sensing data, where the CHL image is a chlorophyll concentration image. The blind super-resolution network includes a kernel prediction unit and an image restoration unit. The kernel prediction unit receives the CHL image to predict the simplified degraded kernel, and the image restoration unit receives the simplified degraded kernel information and the CHL image for super-resolution reconstruction to output a super-resolution image.

2. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 1, characterized in that: The kernel prediction unit first convolves the input LR image to obtain the first LR feature f L ′, and then sent to the feature interleaved multi-scale enhancement sub-unit for processing, and the output is further subjected to global average pooling to obtain a simplified degradation kernel; the LR image is a low-resolution CHL image.

3. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 2, characterized in that: The feature interleaving multi-scale enhancement subunit includes a plurality of serially connected feature interleaving multi-scale enhancement modules FIME; The feature interleaving multi-scale enhancement module FIME first uses the feature concatenation layer to transform the first LR feature f L ′ and the first conditional feature f c ′ is concatenated in the channel dimension to obtain f fusion , the first conditional feature f c ′ is the SR image obtained by the image restoration unit after convolution; the multi-scale spatial feature fusion submodule MFE is used to fusion Perform parallel depth-wise separable convolution and then concatenate the outputs in the channel dimension to obtain multi-scale features f depth ; Then the multi-scale feature f depth Input into the spatial attention module to obtain the multi-scale spatial fusion feature f s-attention , and then obtain the channel fusion feature f through the channel feature fusion submodule depthwise ; and then get f out =f depthwise +f fusion .

4. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 3, characterized in that: The channel feature fusion submodule includes a channel attention module and a point-by-point convolution layer, and a multi-scale spatial fusion feature f s-attention First, the input is sent to the channel attention module to integrate the channel information; the point-by-point convolution uses a 1×1 convolution kernel to convolve the input and output the channel fusion feature f depthwise .

5. A super-resolution method for ocean eddy chlorophyll remote sensing data according to any one of claims 2 to 4, characterized in that: The image restoration unit first uses the convolution layer to convolve the LR image to obtain the second LR feature f L The image restoration unit simultaneously receives the simplified degraded kernel output by the kernel prediction unit, and repeatedly fills each element in the simplified degraded kernel into the corresponding position of a two-dimensional matrix to form a feature map as the second conditional feature f c ; It is then sent to the feature interleaved multi-scale enhancement subunit for processing, and the output is sent to the PixelShuffle layer. The output of the PixelShuffle layer passes through one convolutional layer to obtain the predicted SR image.

6. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 5, characterized in that: The image restoration unit first uses the convolution layer to convolve the LR image to obtain the second LR feature f L In the process of convolution, two convolution layers are used for convolution.

7. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 5, characterized in that: The blind super-resolution network is pre-trained. During the training of the blind super-resolution network, an eddy fitting network is embedded at the end of the image restoration unit to assist in the training of the blind super-resolution network. The eddy fitting network is a neural network whose input is a CHL image and whose output is an eddy estimation value. In the process of blind super-resolution network training, L deg Loss and reconstruction vorticity consistency loss L recvor To optimize the blind super-resolution network parameters; L deg The loss is y represents the SR image, x represents the LR image, k e Represents the predicted degradation kernel, the feature interleaved multi-scale enhancement sub-unit output of the kernel prediction unit passes through a convolution layer to obtain the degradation kernel k e ↓ s represents downsampling; L recvor The loss is ζ S ,ζ L They are the predicted vorticity value obtained by the SR vorticity fitting network and the real vorticity value calculated by the SLA data corresponding to the LR image.

8. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 7, characterized in that: The eddy fitting network is pre-trained; the loss during the training of the eddy fitting network is as follows: L vortex =α×MAE+β×Gradient_Loss Among them, α and β are the weight coefficients that control the contribution of MAE and Gradient_Loss to the total loss, MAE is the mean absolute error, and Gradient_Loss is the vorticity gradient loss. The formula is: Among them, y ι is the simulated vorticity value, and y is the real vorticity value.

9. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 7, characterized in that: The vorticity fitting network includes a vorticity feature extraction module and a vorticity prediction module; The vorticity feature extraction module includes a first convolution layer, a maximum pooling layer, multiple residual block groups and a flattening layer, and a one-dimensional vector is obtained through the flattening layer; The vorticity prediction module includes a fully connected layer, a bottom layer L1, a middle layer L2 and a high layer L3; the bottom layer L1 is a fully connected layer based on a β-spline basis function; the middle layer L2 is a fully connected layer based on a ReLU activation function and a β-spline, wherein 50% of the nodes use β-splines and 50% use ReLU functions; the high layer L3 is a fully connected layer composed of β-spline functions; the vorticity prediction module outputs the vorticity value of the CHL image.

10. The super-resolution method for ocean eddy chlorophyll remote sensing data according to claim 7, characterized in that: When the input image is processed for the first time during the training of the blind super-resolution network, the image restoration module first receives the LR and the initial degradation kernel to generate the SR image, and then convolves the SR to obtain the first conditional feature f c ′, as the input of the first kernel prediction module; The initial degenerate kernel is σ x , σ y , θ are the parameters of the anisotropic Gaussian kernel; (i, j) represents the position in the degenerate kernel.

Images