Comprehensive aperture radiometer reconstruction method based on deep expansion network

By combining traditional iterative optimization algorithms and deep learning methods, using deep expansion networks to reconstruct the bright temperature image of the integrated aperture radiometer, the problems of long reconstruction time, complex parameter selection and low image accuracy in the existing technology are solved, and the problem of High precision and efficient bright temperature image reconstruction.

CN119048628APending Publication Date: 2024-11-29ZHEJIANG SCI-TECH UNIV
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411166813.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

In the prior art, the bright temperature image reconstruction method of the integrated aperture radiometer has problems such as long reconstruction time, complex selection of artificial prior parameters, Gibbs oscillation phenomenon and low image accuracy. The model-based method and deep learning method have their own problems. insufficient.

Method used

Using a comprehensive aperture radiometer reconstruction method based on deep expansion network, the traditional iterative optimization algorithm is combined with deep learning methods, the visibility function is processed through the deep learning model, and the gradient descent module and the near-end mapping module are used for iterative optimization to build the ability to Automatically update the bright temperature image reconstruction model of parameters.

Benefits of technology

It effectively improves the reconstruction accuracy, simplifies the parameter selection process, reduces Gibbs oscillation phenomenon, retains image details, and has good interpretability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048628A_ABST
    Figure CN119048628A_ABST
Patent Text Reader

Abstract

The invention discloses a synthetic aperture radiometer reconstruction method based on a deep expansion network. The method comprises the following steps: firstly, acquiring an original brightness temperature scene image sample, then acquiring an original visibility function of the original brightness temperature scene image sample by using a synthetic aperture radiometer, and processing the original visibility function V to obtain a processed visibility function; constructing a training data set; building a deep learning model and performing training; and inputting the processed visibility function # imgabs0 # to be reconstructed into the brightness temperature image reconstruction model, and outputting the reconstructed brightness temperature image of the synthetic aperture radiometer by the model. According to the method, a traditional iterative optimization method is expanded by using the deep neural network, so that the image reconstruction error can be effectively reduced, and the brightness temperature precision of inversion imaging is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a synthetic aperture radiometer inversion processing method in the field of microwave remote sensing technology, and in particular to a synthetic aperture radiometer reconstruction method based on a deep unfolding network (DGUN). Background Art

[0002] Synthetic Aperture Radiometer (SAIR) is a powerful tool for high-resolution microwave remote sensing earth observation. Different from the traditional real aperture radiometer, SAIRs uses sparse arrays and small-size antennas to replace the traditional large-aperture antennas based on synthetic aperture technology, effectively reducing the physical aperture and improving the spatial resolution. It has the advantages of no mechanical scanning and instantaneous image reconstruction, avoiding manufacturing, volume, weight and other problems.

[0003] The measurement output of SAIR is the visibility function sample in the frequency domain, and the brightness temperature distribution data in the spatial domain needs to be obtained through the brightness temperature reconstruction method. The reconstruction process from the visibility function to the brightness temperature image has been proved to be an ill-conditioned inverse problem. In order to overcome the influence of the ill-conditioned inverse problem on the reconstruction accuracy, the typical brightness temperature reconstruction method is a model-based method, which obtains a stable and unique solution by adding new restrictions or penalties to the inverse problem, such as the minimum norm regularization method (Picard B, Anterrieu E. Comparison of regularized inversion methods in synthetic aperture imaging radiometry [J]. IEEE Transactions on Geoscience and Remote Sensing, 2005, 43 (2): 218-224.), and the band-limited regularization proposed by considering the physical characteristics of the band-limited imaging instrument to obtain a suitable solution (Anterrieu EA resolving matrix approach for synthetic aperture imaging radiometers [J]. IEEE Transactions on Geoscience and Remote Sensing, 2004, 42 (8): 1649-1656.). Although the minimum norm regularization and band-limited regularization methods can effectively alleviate the ill-posedness of the inverse problem, these regularized inversion methods often rely on the selection of characteristic regularization parameters, and due to the ill-posed problem and the physical characteristics of the limited frequency band, there are still large reconstruction errors; and because only one constraint is added, the multi-scale features of the original brightness temperature image cannot be retained, and a large reconstruction error still remains. At the same time, the reconstruction time is long and the selection of prior parameters is complex.

[0004] With the development of deep learning, researchers have begun to apply deep neural networks to the field of remote sensing image reconstruction. The SAIR-CNN method (Zhang Y, Ren Y, Miao W, et al. Microwave SAIR imaging approach based on deep convolutional neural network [J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57 (12): 10376-10389.) uses a neural network to learn the mapping relationship between the visibility function and the brightness temperature image to reconstruct the brightness temperature image. The reconstruction time is fast and there is no need to manually select parameters a priori. Although the SAIR-CNN method solves the problem of complex prior parameter selection by automatically updating parameters through the network, the error of the edge details of objects in the brightness temperature image of the reconstruction result is large, and the overall image accuracy is not high, which makes it difficult to apply to the actual brightness temperature image reconstruction.

[0005] In general, the results of remote sensing image reconstruction using model-based methods and learning-based methods are not satisfactory. Model-based methods take a long time to reconstruct and require manual prior selection of parameters. At the same time, the reconstruction results have Gibbs oscillation phenomenon. Learning-based methods have low image accuracy, lack of details, and poor interpretability of the model. Therefore, a new method is needed to solve the above problems. Summary of the invention

[0006] In view of the problems existing in the background technology, the purpose of the present invention is to provide a synthetic aperture radiometer reconstruction method based on a deep unfolding network to reconstruct the brightness temperature image. The present invention combines the traditional method with the deep learning method to effectively improve the reconstruction accuracy of the synthetic aperture radiometer.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme:

[0008] 1. A synthetic aperture radiometer reconstruction method based on deep unfolding network

[0009] Step 1: Use a synthetic aperture radiometer to collect the original visibility function V corresponding to the original brightness temperature scene image sample, and process the original visibility function V to obtain the processed visibility function

[0010] Step 2: Based on the processed visibility function And the corresponding original brightness temperature scene image samples to build a training data set;

[0011] Step 3: Build a deep learning model;

[0012] Step 4: After training the deep learning model using the training data set, a brightness temperature image reconstruction model is obtained;

[0013] Step 5: Reconstruct the processed visibility function The data is input into the brightness temperature image reconstruction model, and the model outputs the reconstructed synthetic aperture radiometer brightness temperature image.

[0014] In step 3, the deep learning model includes a fully connected layer, a first convolutional layer, multiple stage blocks and a second convolutional layer; the input of the deep learning model is used as the input of the fully connected layer, the fully connected layer is sequentially connected to the second convolutional layer after passing through the first convolutional layer and multiple stage blocks, and the output of the second convolutional layer is used as the output of the deep learning model.

[0015] Each of the stage blocks contains a connected gradient descent module and a proximal mapping module.

[0016] The gradient descent module includes two residual blocks. The input of the gradient descent module is used as the input of the first residual block. The output of the first residual block is compared with the processed visibility function. After element-by-element subtraction, it is used as the input of the second residual block. The output of the second residual block is multiplied element-by-element by the trainable step size parameter μ as the residual branch output. The residual branch output is subtracted element-by-element from the input of the gradient descent module as the output of the gradient descent module.

[0017] The residual block includes a third convolutional layer, a first activation layer and a fourth convolutional layer connected in sequence, the input of the residual block is used as the input of the third convolutional layer, and the output of the fourth convolutional layer is element-by-element added to the input of the residual block as the output of the residual block.

[0018] The main body of the proximal mapping module is a U-Net structure, and the feature fusion downsampling submodule replaces the convolution layer for downsampling in the encoding unit of the U-Net structure. The feature fusion downsampling submodule includes five convolution layers, a Haar wavelet transform module and a second activation layer. The input of the fifth convolution layer is the encoding output result of the previous stage, and the input of the sixth convolution layer is the decoding output result of the previous stage. The outputs of the fifth convolution layer and the sixth convolution layer are added element by element and then used as the inputs of the seventh convolution layer and the eighth convolution layer. The output of the eighth convolution layer is multiplied element by element with the encoding feature of this stage and then added element by element with the output of the seventh convolution layer to obtain the encoding output of this stage. The encoding output of this stage is used as the input of the Haar wavelet transform module. The Haar wavelet transform module is connected to the second activation layer after the ninth convolution layer, and the output of the second activation layer is used as the output of the feature fusion downsampling submodule.

[0019] The number of the stage blocks is 7.

[0020] In step 1, the original brightness temperature scene image samples are remote sensing images of different categories.

[0021] 2. A synthetic aperture radiometer reconstruction system based on deep unfolding network

[0022] A training data set construction module, used for generating a training data set according to original brightness temperature scene image samples and original visibility functions;

[0023] A model training module is used to train the stored deep learning model using the training data set to obtain a brightness temperature image reconstruction model;

[0024] The image reconstruction module is used to reconstruct the input processed visibility function to be reconstructed by using the brightness temperature image reconstruction model to obtain a reconstructed synthetic aperture radiometer brightness temperature image.

[0025] 3. A computer device

[0026] The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the synthetic aperture radiometer reconstruction method based on a deep unfolding network when executing the computer program.

[0027] 4. A computer-readable storage medium

[0028] The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the synthetic aperture radiometer reconstruction method based on a deep unfolding network are implemented.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention proposes a synthetic aperture radiometer reconstruction method based on a deep unfolding network for SAIR reconstruction. The present invention combines the traditional iterative optimization algorithm with the deep learning method, effectively solving the problems of the traditional method, such as the complexity of manually selecting parameters, the long reconstruction time, and the existence of the Gibbs oscillation phenomenon; at the same time, compared with the deep learning method, the method of the present invention has good interpretability, and well preserves the image details, effectively improving the reconstruction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the present invention;

[0032] Figure 2 is the original brightness temperature image of the test scene;

[0033] Figure 3 It is a model structure diagram of the method of the present invention;

[0034] Figure 4It is the structure diagram of the gradient descent module;

[0035] Figure 5 It is the residual block structure diagram;

[0036] Figure 6 This is a structural diagram of the proximal mapping module;

[0037] Figure 7 This is the structural diagram of the feature fusion downsampling submodule;

[0038] Figure 8 is the reconstructed brightness temperature image of the minimum norm regularization method within the aliasing-free field of view;

[0039] Fig. 9 is the reconstructed brightness temperature image of the band-limited regularization method within the aliasing-free field of view;

[0040] Fig.10 It is the reconstructed brightness temperature image of the SAIR-CNN method within the aliasing-free field of view;

[0041] Fig.11 is a reconstructed brightness temperature image of the method of the present invention within the non-aliasing field of view;

[0042] Fig.12 Performance comparison (RMS error) of different methods at different noise levels;

[0043] Fig.13 Performance comparison of different methods at different noise levels (peak signal-to-noise ratio). DETAILED DESCRIPTION

[0044] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.

[0045] The present invention combines the traditional iterative optimization method with the deep learning method, draws on each other's strengths and overcomes each other's weaknesses, and proposes a synthetic aperture radiometer reconstruction method based on a deep unfolding network, the main process of which is: obtaining original brightness temperature scene image samples; using a synthetic aperture radiometer to collect visibility function data of the original brightness temperature scene image samples, and obtaining a visibility function after processing the visibility function data; using the visibility function to make data set samples, and dividing them into a training set and a test set; building a deep learning model; and obtaining a reconstruction result based on the preprocessed data to be reconstructed based on the trained deep learning model.

[0046] In order to verify the effectiveness and superiority of the method of the present invention, numerical simulation tests were carried out on a two-dimensional synthetic aperture radiometer imaging system and compared with minimum norm regularization, band-limited regularization and SAIR-CNN. The antenna array is a 61-unit T-type array, sparsely arranged at different positions, and contains multiples of the shortest baseline Δd = 0.6λ. The specific system characteristic parameters also include a center frequency of 1.4GHz, a bandwidth of 20MHz, and an integration time τ = 1s. The original brightness temperature image of the test scene is selected as follows Figure 2 shown.

[0047] like Figure 1 As shown, this embodiment includes the following steps:

[0048] Step 1: Obtain the original brightness temperature scene image samples, which are remote sensing images of different categories. The comprehensive aperture radiometer system includes its center frequency, bandwidth, antenna array, number of antennas, and shortest baseline. Use the comprehensive aperture radiometer to collect the original visibility function V corresponding to the original brightness temperature scene image samples. The original visibility function V is solved by the following formula:

[0049] V=GH

[0050] Among them, G is the system debugging matrix, and H is the original brightness temperature scene image.

[0051] In this embodiment, if Figure 2 As shown in the figure, the resolution of a single original brightness temperature scene image is 128 × 128. The 61 antennas in the antenna array form 3661 sampling baselines, so the number of sampling points of the visibility function, including the visibility function at zero distance, is 3661. The number of sampling points and the number of pixels of the original brightness temperature image are the number of rows and columns of the system modulation matrix G, respectively, from which the visibility function V can be obtained.

[0052] After processing the original visibility function V, the processed visibility function is obtained In this embodiment, a function with a mean of zero and a variance σ is added to the original visibility function V. 2 Set to 0.01×max(V i )=730.75 Gaussian white noise e is 3661×1, and the processed visibility function is obtained

[0053] Step 2: Based on the processed visibility function And the corresponding original brightness temperature scene image samples to build a training data set. The processed visibility function As the input of the deep learning model, the corresponding original brightness temperature scene image samples are used as labels;

[0054] Step 3: Build a deep learning model;

[0055] In step 3, if Figure 3 As shown, the deep learning model includes a fully connected layer, a first convolutional layer, multiple stage blocks and a second convolutional layer; the input of the deep learning model is used as the input of the fully connected layer, the fully connected layer is sequentially connected to the second convolutional layer after passing through the first convolutional layer and multiple stage blocks, and the output of the second convolutional layer is used as the output of the deep learning model.

[0056] According to the visibility function Inversion imaging is performed by introducing constraints for regularization, that is, solving the minimum value of the inverted brightness temperature image data expression. The minimum value is the only stable solution, which is set by the following expression:

[0057]

[0058] in, represents the output of the kth iteration, R(T) is the regularization term, λ is the weighting parameter of the regularization term, Represents the two-norm square operation. The above expression is expressed as an iterative problem through the idea of ​​proximal gradient descent algorithm, and its expression is:

[0059]

[0060] where g(·) represents the data fidelity term, which ensures that the solution is consistent with the degradation process. ▽ represents the differential operator, weighted by the step size μ. represents the output of the k-1th iteration,

[0061] Further, it is decomposed into the gradient descent subproblem and the proximal mapping subproblem, and the specific expression is:

[0062]

[0063] in, represents the gradient descent operation, Indicates a near-end mapping operation. k express The output after gradient descent, μ represents the trainable step size parameter, G and G T represents the degenerate matrix and its transpose. prox(·) represents the proximal operator.

[0064] Therefore, each stage block in the present invention includes a connected gradient descent module and a proximal mapping module. In this embodiment, the number of stage blocks is 7.

[0065] like Figure 4 As shown, the gradient descent module includes two residual blocks. The input of the gradient descent module is used as the input of the first residual block. The output of the first residual block is compared with the processed visibility function After element-by-element subtraction, it is used as the input of the second residual block. The output of the second residual block is multiplied element-by-element by the trainable step size parameter μ as the residual branch output. The residual branch output is subtracted element-by-element from the input of the gradient descent module as the output of the gradient descent module. Figure 5 As shown in the figure, the residual block includes the third convolutional layer, the first activation layer and the fourth convolutional layer connected in sequence. The input of the residual block is used as the input of the third convolutional layer, and the output of the fourth convolutional layer is added element by element to the input of the residual block as the output of the residual block. The first activation layer uses the ReLU activation function. The gradient descent module is used to simulate the gradient descent step in the PGD method. Two independent residual blocks are used to simulate the matrix G and its transpose G T , the trainable step size μ is used to improve robustness. Therefore, the gradient descent module can generate the following gradient descent operation, the specific expression is:

[0066]

[0067] in, and Represents the matrix G and its transpose G used to simulate T The residual block.

[0068] like Figure 6 As shown, the main body of the proximal mapping module is a U-Net structure. U-Net is a U-shaped symmetric convolutional neural network, which includes a coding unit and a decoding unit. The coding unit is composed of a channel attention module, a first residual block, a feature fusion downsampling submodule and a convolution layer. The channel attention module includes an adaptive average pooling layer, a convolution layer, a ReLU layer and a sigmoid layer. The channel attention module is used to selectively emphasize useful features in the visibility function and suppress meaningless features. Each feature fusion downsampling submodule and the upsampling layer are connected through a convolution layer. The present invention replaces the convolution layer for downsampling in the coding unit of the U-Net structure with the feature fusion downsampling submodule, such as Figure 7As shown in FIG. 1 , the feature fusion downsampling submodule includes five convolutional layers, a Haar wavelet transform module and a second activation layer. The input of the fifth convolutional layer is the encoding output result of the corresponding scale of the previous stage, and the input of the sixth convolutional layer is the decoding output result of the corresponding scale of the previous stage. The outputs of the fifth and sixth convolutional layers are added element by element and then used as the inputs of the seventh and eighth convolutional layers. The output of the eighth convolutional layer is multiplied element by element with the encoding feature of this stage (i.e., the output of the previous residual block) and then added element by element with the output of the seventh convolutional layer to obtain one of the encoding outputs of this stage. The encoding output of this stage is used as the input of the Haar wavelet transform module. The Haar wavelet transform module is connected to the second activation layer after the ninth convolutional layer. The output of the second activation layer is used as the output of the feature fusion downsampling submodule, recorded as the downsampling output of this stage and used as the input of the next residual block. The second activation layer uses the ReLU activation function. Considering the information loss problem existing in the traditional method, an inter-stage path is designed in the feature fusion downsampling submodule of each stage to transmit useful information from the encoding unit and the decoding unit at different stages. At the same time, since the encoding unit and the decoding unit use a skip connection method, the inter-level information can also be transmitted to the decoding unit, which retains the feature information to a large extent. Specifically, in the feature fusion downsampling submodule, the features of the encoding unit and the decoding unit in the previous stage will be transmitted to the current stage. They are first input into two independent 1×1 convolutions and merged by element-by-element addition to obtain the fusion result of the features of the previous stage (nth order). then, Will be input into the seventh convolutional layer and the eighth convolutional layer respectively to obtain two mapping parameters and Encoding features at this stage will with After element-wise multiplication, Add element by element to get is the encoding output of this stage. Downsampling is performed to obtain features with reduced resolution The downsampled output of this stage is the input of the decoding unit. The present invention abandons the traditional downsampling operation such as pooling and adopts Haar wavelet transform to reduce the spatial resolution of the feature map. Specifically, Haar wavelet transform decomposes the feature map into four components: low frequency (A), horizontal (H), vertical (V) and high frequency (D), and extracts features through the convolution layer to obtain the information-rich features after downsampling. It is worth noting that the resolution of each component is reduced to half of the original, while the number of channels of the feature map is increased by four times. In other words, the Haar wavelet transform encodes part of the information from the spatial dimension to the channel dimension without losing any information.

[0069] The decoding unit consists of a convolutional layer, an upsampling layer, a second residual block and a self-attention module. In the present invention, bilinear upsampling and convolutional layers are used for upsampling operations, and the supervised attention module consists of a convolutional layer, which extracts clean features and then inputs them into the next stage through subspace projection.

[0070] Step 4: After training the deep learning model using the training data set, a brightness temperature image reconstruction model is obtained;

[0071] The specific steps of step 4 are:

[0072] During the deep learning model training phase, the data set samples are divided into training data samples and test data samples at a ratio of 9:1, and the data samples are preprocessed and input into the neural network. The preprocessing process specifically includes: separating the visibility function in the data sample into real and imaginary parts, and then splicing the separated real part of the visibility function and the imaginary part of the visibility function back and forth in the first dimension. The L2 norm is used to calculate the error between the reconstructed result and the original brightness temperature image during the training process. Furthermore, adaptive gradient descent is used to minimize the error, and then backpropagation is used to update the network weights and biases. The data input batch size is set to 16; the learning rate is set to 0.0001; considering the depth of the model, the present invention adopts a warm-up strategy to gradually increase the learning rate.

[0073] Figure 8 , Fig. 9 , Fig.10 and Fig.11 The figure shows the reconstruction of one sample of the test data preprocessed in step 5 in the aliasing-free field of view using the minimum norm regularization method, band-limited regularization method, SAIR-CNN method, and the method of the present invention. It is worth noting that the reconstruction results of the minimum norm regularization and band-limited regularization have truncation of the sampling of the observed target in the spatial frequency domain due to the limited number of sampling baselines of the system, resulting in obvious Gibbs phenomenon in the imaging, obvious oscillations at the edge of the image, and a long reconstruction time. Compared with the SAIR-CNN method, the method of the present invention can produce smoother and more accurate reconstruction results with clearer detail information.

[0074] For the quantitative comparison of the method of the present invention with the minimum norm regularization, band-limited regularization method and SAIR-CNN method, the present invention measures by calculating the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) under different methods. The specific calculation formula is as follows:

[0075]

[0076] In the formula, are brightness temperature images reconstructed by different methods, H is the original brightness temperature image, J and L are and the length and width of H. The RMSE values ​​of minimum norm regularization, band-limited regularization, SAIR-CNN and the proposed method are 17.08K, 15.94K, 15.37K and 10.6K respectively. The PSNRs are 23.48dB, 24.08dB, 24.40dB and 27.62dB respectively. Their reconstruction times are 36.697s, 22.986s, 0.008s and 0.084s respectively.

[0077] In order to further verify the inversion performance of the method of the present invention under different noise levels, the sensitivity of the inversion method to error and noise interference is quantitatively analyzed. 2 0.01×max(V i )=730.75 as the spacing, and tested adding a mean of zero to the visibility function and setting the variance to 0.01×max(V i )=730.75~0.10×max(V i )=7307.5 Gaussian white noise. The inversion results of the method of the present invention are compared with the minimum norm regularization method, the band-limited regularization method and SAIR-CNN respectively. Fig.12 and Fig.13 The average performance of the minimum norm regularization, band-limited regularization, SAIR-CNN method and the method of the present invention under different noise levels are shown respectively. The results show that compared with the minimum norm regularization, band-limited regularization method and SAIR-CNN method, the method of the present invention has stronger robustness to measurement errors and noise interference. It can be found that the method of the present invention can better improve the accuracy of reconstructed brightness temperature distribution.

[0078] The present invention also proposes a synthetic aperture radiometer reconstruction system based on a deep unfolding network, comprising:

[0079] A training data set construction module, used for generating a training data set according to original brightness temperature scene image samples and original visibility functions;

[0080] A model training module is used to train the stored deep learning model using the training data set to obtain a brightness temperature image reconstruction model;

[0081] The image reconstruction module is used to reconstruct the input processed visibility function to be reconstructed by using the brightness temperature image reconstruction model to obtain a reconstructed synthetic aperture radiometer brightness temperature image.

[0082] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A synthetic aperture radiometer reconstruction method based on deep unfolding network, characterized in that: The following steps are included: Step 1: Use the synthetic aperture radiometer to collect the original visibility function V corresponding to the original brightness temperature scene image sample, and process the original visibility function V to obtain the processed visibility function Step 2: Based on the processed visibility function And the corresponding original brightness temperature scene image samples to build a training data set; Step 3: Build a deep learning model; Step 4: After training the deep learning model using the training data set, a brightness temperature image reconstruction model is obtained; Step 5: Reconstruct the processed visibility function The data is input into the brightness temperature image reconstruction model, and the model outputs the reconstructed synthetic aperture radiometer brightness temperature image.

2. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 1, characterized in that: In step 3, the deep learning model includes a fully connected layer, a first convolutional layer, multiple stage blocks and a second convolutional layer; the input of the deep learning model is used as the input of the fully connected layer, the fully connected layer is sequentially connected to the second convolutional layer after passing through the first convolutional layer and multiple stage blocks, and the output of the second convolutional layer is used as the output of the deep learning model.

3. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 2, characterized in that: Each of the stage blocks contains a connected gradient descent module and a proximal mapping module.

4. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 3, characterized in that: The gradient descent module includes two residual blocks. The input of the gradient descent module is used as the input of the first residual block. The output of the first residual block is compared with the processed visibility function. After element-by-element subtraction, it is used as the input of the second residual block. The output of the second residual block is multiplied element-by-element by the trainable step size parameter μ as the residual branch output. The residual branch output is subtracted element-by-element from the input of the gradient descent module as the output of the gradient descent module.

5. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 4, characterized in that: The residual block includes a third convolutional layer, a first activation layer and a fourth convolutional layer connected in sequence, the input of the residual block is used as the input of the third convolutional layer, and the output of the fourth convolutional layer is element-by-element added to the input of the residual block as the output of the residual block.

6. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 3, characterized in that: The main body of the proximal mapping module is a U-Net structure, and the feature fusion downsampling submodule replaces the convolution layer for downsampling in the encoding unit of the U-Net structure. The feature fusion downsampling submodule includes five convolution layers, a Haar wavelet transform module and a second activation layer. The input of the fifth convolution layer is the encoding output result of the previous stage, and the input of the sixth convolution layer is the decoding output result of the previous stage. The outputs of the fifth convolution layer and the sixth convolution layer are added element by element and then used as the inputs of the seventh convolution layer and the eighth convolution layer. The output of the eighth convolution layer is multiplied element by element with the encoding feature of this stage and then added element by element with the output of the seventh convolution layer to obtain the encoding output of this stage. The encoding output of this stage is used as the input of the Haar wavelet transform module. The Haar wavelet transform module is connected to the second activation layer after the ninth convolution layer, and the output of the second activation layer is used as the output of the feature fusion downsampling submodule.

7. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 2, characterized in that: The number of the stage blocks is 7.

8. The method for reconstructing a synthetic aperture radiometer based on a deep unfolding network according to claim 1, characterized in that: In step 1, the original brightness temperature scene image samples are remote sensing images of different categories.

9. A synthetic aperture radiometer reconstruction system based on a deep unfolding network, characterized in that: include: A training data set construction module, used for generating a training data set according to original brightness temperature scene image samples and original visibility functions; A model training module is used to train the stored deep learning model using the training data set to obtain a brightness temperature image reconstruction model; The image reconstruction module is used to reconstruct the input processed visibility function to be reconstructed by using the brightness temperature image reconstruction model to obtain a reconstructed synthetic aperture radiometer brightness temperature image.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Citation Information

Cited By

  • Error correction method and system for synthetic aperture radiometer, terminal equipment and medium

    CN120352841A

  • Comprehensive aperture radio array rapid imaging method based on deep learning

    CN121120838A

  • A Deep Learning-Based Fast Imaging Method for Integrated Aperture Radio Arrays

    CN121120838B

  • Satellite-borne synthetic aperture microwave radiometer sea surface bright temperature difference reconstruction method and system based on double-region fitting

    CN122194142A

  • Self-supervised radio interference imaging method and system

    CN122435092A