A deep learning-based synthetic aperture radiometer error correction method

Abstract

A kind of comprehensive aperture radiometer error correction method based on deep learning, in order to correct the amplitude and phase error of antenna and channel and gibbs oscillation, improve the imaging quality, generate original scene brightness temperature image according to the dynamic range of actual application scene;Generate visibility function containing system error;Original scene brightness temperature image and visibility function jointly construct training data set and test data set;The deep learning network is trained using the training data set;The network correction effect is verified using the test data set.

Description

Technical Field

[0001] This invention relates to a method for error correction of integrated aperture radiometers based on deep learning, belonging to the field of microwave remote sensing and detection technology. Background Technology

[0002] Synthetic aperture technology utilizes small-aperture antennas to equivalently synthesize large-aperture antennas, which can solve problems such as limited aperture and difficulties in mechanical scanning of real-aperture microwave radiometers, and improve the spatial resolution of observations.

[0003] However, compared to real aperture radiometers, synthetic aperture systems have a wide variety of errors. System errors can lead to a decrease in the quality of synthetic aperture microwave radiation imaging. Currently used error correction methods (external source calibration method, noise injection method, etc.) cannot completely correct system antenna errors, channel errors, and Gibbs oscillation errors. Summary of the Invention

[0004] The technical problem solved by this invention is that, in the current technology, it is difficult to completely correct system antenna errors, channel errors, and Gibbs oscillation errors. Therefore, a deep learning-based integrated aperture radiometer error correction method is proposed.

[0005] The present invention solves the above-mentioned technical problem through the following technical solution:

[0006] A deep learning-based method for correcting errors in a comprehensive aperture radiometer includes:

[0007] Generate the original scene brightness and temperature image;

[0008] Error information is obtained by observing external sources using a synthetic aperture microwave radiometer system to be calibrated.

[0009] Construct a visibility function;

[0010] Construct training and testing datasets;

[0011] Build and train a deep learning network;

[0012] The training dataset is then input into the trained deep learning network for validation.

[0013] The original scene brightness temperature image is determined based on the dynamic range of the actual application scenario of the integrated aperture microwave radiometer system to be calibrated, including simulated natural scenes or actual observed natural scenes.

[0014] The error information includes antenna pattern error, channel amplitude and phase error, and Gibbs oscillation error, which are output after normalization processing.

[0015] The visibility function is generated based on the original scene brightness temperature image and error information, through a comprehensive aperture brightness temperature simulation program or a comprehensive aperture system measurement.

[0016] The two-dimensional comprehensive aperture visibility function V obtained by the comprehensive aperture brightness temperature simulation program is specifically:

[0017]

[0018] In the formula, T(ξ,η) represents the scene brightness temperature distribution, (ξ,η) represents the direction cosine, and (x,y) represents the antenna position coordinates.

[0019] The visibility function V generated by the integrated aperture system is specifically:

[0020]

[0021] In the formula, G ij For amplitude error, β ij This represents the phase error.

[0022] The total dataset is constructed based on the original scene brightness temperature image T and the visibility function V. 90% of the data is randomly selected as the training dataset for training the deep learning network, and the remaining data is used as the test dataset to test the brightness temperature image reconstruction effect of the network.

[0023] The deep learning network takes the spectral data output by the visibility function V as input and the original scene brightness temperature image Tb as the label image during training to train the deep learning network, wherein:

[0024] Deep learning networks consist of an input layer, a fully connected layer, a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer. The fully connected layer maps the input spectral data onto the two-dimensional convolutional layer. The resulting matrix is ​​then used for high-order feature extraction, dimension transformation, and image reconstruction on the two-dimensional convolutional layer to obtain the reconstructed microwave brightness temperature image. The image is then normalized by the batch normalization layer and nonlinearly processed by the ReLU activation function layer.

[0025] In the deep learning network, the weights of the network parameters in each layer are implemented using a loss function defined by the mean square error between the reconstructed microwave brightness temperature image and the original scene microwave brightness temperature image output by the deep learning network. Specifically, the loss function is:

[0026]

[0027] In the formula, m is the number of training samples, and the loss function f(U) is minimized using the standard backpropagation stochastic gradient descent (SGD) algorithm. k ;θ) is the reconstructed microwave brightness temperature image, Y k This is the original scene's microwave brightness temperature image.

[0028] The deep learning network is trained iteratively a specified number of times using a loss function. When the output of the loss function reaches a stable standard, the current trained deep learning network is saved and verified using a test dataset. The current output two-dimensional integrated aperture microwave brightness temperature image is obtained, and the spatial and frequency domain analyses of the network-reconstructed image are performed to verify the training effect.

[0029] The trained deep learning network, after being validated, is used to perform error correction on the integrated aperture microwave radiometer system to be calibrated.

[0030] The advantages of this invention compared to the prior art are:

[0031] This invention provides a deep learning-based method for correcting errors in integrated aperture radiometers. It generates an original scene brightness temperature image based on the dynamic range of the actual application scenario; generates a visibility function that includes system errors; the original scene brightness temperature image and the visibility function together construct training and testing datasets; trains a deep learning network using the training dataset; and verifies the network's correction effect using the testing dataset. This method solves the problems of traditional methods requiring additional calibration networks to correct system antenna and channel errors, and the inability to suppress Gibbs oscillation errors. It simplifies the system hardware configuration and improves detection accuracy. Attached Figure Description

[0032] Figure 1 A schematic diagram of a synthetic aperture radiometer provided for the invention;

[0033] Figure 2 A schematic diagram of signal reception provided for the invention;

[0034] Figure 3 A schematic diagram of a deep learning network structure provided for the invention;

[0035] Figure 4 A flowchart of a deep learning-based integrated aperture radiometer error correction method provided for the invention;

[0036] Figure 5 A schematic diagram of a 24-cell rectangular array used for testing in this invention;

[0037] Figure 6 A schematic diagram of an uncorrected noise source image provided for the invention;

[0038] Figure 7 A schematic diagram of the corrected noise source image provided for the invention;

[0039] Figure 8 A schematic diagram of an uncorrected stepped spread source image provided for the invention;

[0040] Figure 9A stepped source image corrected for the schematic diagram provided for the invention; Detailed Implementation

[0041] A deep learning-based method for correcting errors in an integrated aperture radiometer is proposed. To correct the amplitude and phase errors and Gibbs oscillations of the antenna and channel, and improve imaging quality, the method generates an original scene brightness temperature image based on the dynamic range of the actual application scenario; generates a visibility function that includes systematic errors; the original scene brightness temperature image and the visibility function are used to construct training and testing datasets; a deep learning network is trained using the training dataset; and the network correction effect is verified using the testing dataset.

[0042] The specific steps are as follows:

[0043] Generate the original scene brightness and temperature image;

[0044] Error information is obtained by observing external sources using a synthetic aperture microwave radiometer system to be calibrated;

[0045] Construct a visibility function;

[0046] Construct training and testing datasets;

[0047] Build and train a deep learning network;

[0048] The training dataset is then input into the trained deep learning network for validation.

[0049] The original scene brightness temperature image is determined based on the dynamic range of the actual application scenario of the integrated aperture microwave radiometer system to be calibrated, including simulated natural scenes or actual observed natural scenes.

[0050] Error information includes antenna pattern error, channel amplitude and phase error, and Gibbs oscillation error, which are output after normalization.

[0051] The visibility function is generated based on the original scene brightness temperature image and error information, through a comprehensive aperture brightness temperature simulation program or a comprehensive aperture system measurement.

[0052] The total dataset is constructed based on the original scene brightness temperature image T and the visibility function V. 90% of the data is randomly selected as the training dataset for training the deep learning network, and the remaining data is used as the test dataset to test the brightness temperature image reconstruction effect of the network.

[0053] The deep learning network is trained using the spectral data output by the visibility function V as input and the original scene brightness temperature image Tb as the label image during training.

[0054] The deep learning network includes an input layer, a fully connected layer, a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer. The fully connected layer maps the input spectral data onto the two-dimensional convolutional layer. The resulting matrix is ​​then used for high-order feature extraction, dimension transformation, and image reconstruction on the two-dimensional convolutional layer to obtain the reconstructed microwave brightness temperature image. The image is then normalized by the batch normalization layer and nonlinearly processed by the ReLU activation function layer.

[0055] In deep learning networks, the weights of the network parameters in each layer are implemented using a loss function defined by the mean square error between the reconstructed microwave brightness temperature image and the original scene microwave brightness temperature image output by the deep learning network.

[0056] The deep learning network is trained iteratively a specified number of times using a loss function. When the output of the loss function reaches a stable standard, the current trained deep learning network is saved and validated using a test dataset. The current output two-dimensional integrated aperture microwave brightness temperature image is obtained, and the spatial and frequency domain analyses of the network-reconstructed image are performed to verify the training effect.

[0057] The trained deep learning network, after being validated, is used to perform error correction on the integrated aperture microwave radiometer system to be calibrated.

[0058] The following description, in conjunction with the accompanying drawings and preferred embodiments, provides further details:

[0059] In the current embodiment, the signal reception schematic diagram of the deep learning-based integrated aperture radiometer error correction method is shown below. Figure 2 As shown, the integrated aperture radiometer system is as follows: Figure 1 As shown, the synthetic aperture radiometer includes an antenna array, a receiving channel array, an A / D array, and a correlator. The antenna array receives radiated signals from the observed scene. The receiving channel array comprises multiple receiving channels, each corresponding to one antenna element. The receiving channels down-convert, filter, and amplify the signals received by the antenna elements. The A / D array corresponds one-to-one with the receiving channel array, converting the analog signals in the receiving channels into digital signals. The correlator correlates the signals converted by the A / D array pairwise. The output of the correlation is a visibility function.

[0060] like Figure 4 As shown, the deep learning-based integrated aperture radiometer error correction method specifically includes the following steps:

[0061] 1. Generate the original scene brightness and temperature image

[0062] There are two methods to generate the original scene brightness temperature image. The first method is to simulate the natural scene to generate the original scene brightness temperature image. This is done by collecting color images of different scenes, converting them into grayscale images, and mapping the values ​​to 2.73k to 300k, which is then used as the original scene brightness temperature image Tb. The second method is to actually observe the natural scene and obtain the actual brightness temperature of the natural scene through auxiliary means such as thermistors, buoys, and weather balloons, which is then used as the original scene brightness temperature image Tb.

[0063] 2. Using external sources to measure the error of the actual system

[0064] The normalized output of the actual system after observing external sources is

[0065] V ij =G ij ·exp(-jβ ij )

[0066] The amplitude of the output is the amplitude error G. ij The output phase is the phase error β. ij .

[0067] 3. Generate visibility function

[0068] There are two methods to generate the visibility function. The first is to use a two-dimensional synthetic aperture brightness temperature simulation program to simulate and obtain the simulated two-dimensional synthetic aperture visibility function V. The second is to obtain the measured visibility function by actually observing the natural scene through a synthetic aperture system.

[0069] The ideal visibility function expression is:

[0070]

[0071] Taking into account antenna and channel amplitude and phase errors, the expression for the visibility function is:

[0072]

[0073] 4. Construct the dataset

[0074] A dataset was constructed using the original scene brightness temperature image T and the visibility function V. 90% of the data was randomly selected as the training dataset for training the network, including the two-dimensional synthetic aperture visibility function V and the corresponding original scene brightness temperature image T. 10% of the data was used as the test dataset to test the brightness temperature image reconstruction effect of the network.

[0075] 5. Training deep learning networks

[0076] 5.1 The network structure mainly includes one input layer, one fully connected layer, seven two-dimensional convolutional layers, a batch normalization layer, and a ReLU activation function layer, as follows: Figure 3 As shown.

[0077] 5.2 The two-dimensional synthetic aperture visibility function V is used as input, and the original scene brightness temperature image Tb is used as the label image during training to jointly train the deep learning network. Since each point in the visibility function includes both imaginary and real parts, a fully connected layer is used to map the input spectral data onto the convolutional layer. This layer mainly consists of a hyperbolic tangent activation function (tanh activation function, defined as...). The fully connected layer (FC1 layer) is implemented. An N*N visibility function (complex number) is shaped into an N^2*1 real-valued vector, then activated by the hyperbolic tangent function, and reshaped into an N*N matrix in preparation for convolution processing.

[0078] 5.3 Following this are two-dimensional convolutional layers (CONV1 to CONV7) with rectified nonlinear activation. The first two convolutional layers are responsible for extracting high-order features from the spectral data, while the last five layers are responsible for dimensionality transformation and image reconstruction. The extracted high-order feature information is used to reconstruct the brightness temperature image, obtaining the reconstructed microwave brightness temperature image. In subsequent convolutional layers (CONV1 to CONV6), the output f(U) after network execution is processed... i ) is defined as:

[0079] f(X i ) = ReLU[BN(W i *X i +B i )],i=0,...,n

[0080] In the formula: * represents the convolution operation; BN represents the batch normalization operation; ReLU is a non-linear operation unit performed after the BN layer, defined as ReLU(X) = max(0,X); n is the number of convolutional layers; W i and B i X represents the weights and biases of the i-th convolutional kernel, respectively; i X1 is the image patch extracted from the i-th layer; X2 is the input spectral data of size N*N mapped by fully connected layers. After passing through multiple convolutional networks, the desired image is reconstructed from the spectral matrix, thus obtaining the reconstructed microwave brightness temperature image.

[0081] 5.4 A batch normalization (BN) layer is added after the convolutional layer to perform batch normalization before applying the ReLU function. The BN layer can accelerate convergence, solve the internal covariate shift problem that occurs when training deep neural networks, and improve generalization ability. It also helps to improve the accuracy of the network training model. The ReLU layer sets the output of some neurons to 0, reducing the interdependence between parameters and mitigating overfitting.

[0082] 5.5 To reduce the difference between the reconstructed brightness temperature image and the original scene microwave brightness temperature image T, it is necessary to adjust the weights of the network parameters θ={W1,...W i ,B1,…B i The process involves continuous optimization and adjustment, which involves minimizing the network to reconstruct the brightness temperature image f(U). k ;θ) and the original scene microwave brightness temperature image Y k The mean squared error between them is used to define the network's loss function as follows:

[0083]

[0084] In the formula: m is the number of training samples, and the loss function is minimized using the standard backpropagation stochastic gradient descent (SGD) algorithm.

[0085] 5.6 After multiple training iterations, when the loss function tends to stabilize, save the final trained network parameters.

[0086] 6. Verify the network reconstruction effect

[0087] The visibility function V from the test dataset is input into the trained deep learning network to obtain a two-dimensional composite aperture microwave brightness temperature image reconstructed by the network. Spatial and frequency domain analyses are then performed on the reconstructed image to verify the network reconstruction effect.

[0088] In this embodiment, such as Figure 5 The image shows the array used in the simulation, which is a 24-element rectangular array.

[0089] The specific steps are as follows:

[0090] (1) For a 24-element rectangular array, the integrated aperture system collects the radiation signal of the noise source scene;

[0091] (2) Uncorrected brightness temperature image of noise sources, such as Figure 6 As shown;

[0092] (3) Input the visibility function of the uncorrected noise source radiation signal into the network to obtain the reconstructed brightness temperature image, such as... Figure 7 As shown.

[0093] Based on the experimental results, Figure 6 The image is an uncorrected brightness temperature image of a noise source, resulting in distortion and making it impossible to distinguish the noise source. Figure 7 The image shows the brightness temperature of the noise source after correction. The noise source has a clear outline, a pure background, and minimal Gibbs oscillation. The amplitude and phase errors of the antenna and channel, as well as the Gibbs oscillation, can be corrected using a deep learning-based integrated aperture radiometer error correction method, resulting in a high-quality image.

[0094] Example 2: Brightness-temperature image of a stepped spread source scene

[0095] In this embodiment, a simulation verification is performed on the error correction method for a comprehensive aperture radiometer based on deep learning.

[0096] The specific steps are as follows:

[0097] (1) For a 24-element rectangular array, the integrated aperture system collects the radiation signal of the noise source scene;

[0098] (2) Uncorrected brightness temperature image of noise sources, such as Figure 8 As shown;

[0099] (3) Input the visibility function of the uncorrected noise source radiation signal into the network to obtain the reconstructed brightness temperature image, such as... Figure 9 As shown.

[0100] Based on the experimental results, Figure 8 The image is an uncorrected brightness temperature image of the source, resulting in distortion and indistinguishability of the source. Figure 9 The corrected image of the exposed source brightness temperature shows a clear exposed source outline, a pure background, and minimal Gibbs oscillation. The deep learning-based integrated aperture radiometer error correction method can correct the amplitude and phase errors of the antenna and channel, as well as the Gibbs oscillation, resulting in a high-quality image.

[0101] This embodiment can correct the amplitude and phase errors of the antenna and channel, as well as Gibbs oscillations, thereby improving imaging quality. This solves the problem that existing technologies cannot completely correct system antenna errors, channel errors, and Gibbs oscillation errors. Furthermore, this method can replace the function of the system calibration network, simplifying the system hardware structure.

[0102] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

[0103] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A method for error correction of a synthetic aperture radiometer based on deep learning, characterized in that... include: Generate the original scene brightness and temperature image; Error information is obtained by observing external sources using a synthetic aperture microwave radiometer system to be calibrated. Construct a visibility function; Construct training and testing datasets; Build and train a deep learning network; Input the training dataset into the trained deep learning network for validation; The original scene brightness temperature image is determined based on the dynamic range of the actual application scenario of the integrated aperture microwave radiometer system to be calibrated, including simulated natural scenes or actual observed natural scenes. The error information includes antenna pattern error, channel amplitude and phase error, and Gibbs oscillation error, which are output after normalization. The visibility function is based on the original scene brightness temperature image and error information, and is generated by a comprehensive aperture brightness temperature simulation program or by actual measurement of a comprehensive aperture system. The two-dimensional comprehensive aperture visibility function V obtained by the comprehensive aperture brightness temperature simulation program is specifically: In the formula, For scene brightness temperature distribution, The direction cosine, These are the antenna position coordinates; The visibility function V generated by the integrated aperture system is specifically: In the formula, For amplitude error, This is the phase error; The total dataset is constructed based on the original scene brightness temperature image T and the visibility function V. 90% of the data is randomly selected as the training dataset for training the deep learning network, and the remaining data is used as the test dataset to test the brightness temperature image reconstruction effect of the network.

2. The method for error correction of a synthetic aperture radiometer based on deep learning according to claim 1, characterized in that: The deep learning network takes the spectral data output by the visibility function V as input and the original scene brightness temperature image Tb as the label image during training to train the deep learning network, wherein: Deep learning networks consist of an input layer, a fully connected layer, a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer. The fully connected layer maps the input spectral data onto the two-dimensional convolutional layer. The resulting matrix is ​​then used for high-order feature extraction, dimension transformation, and image reconstruction on the two-dimensional convolutional layer to obtain the reconstructed microwave brightness temperature image. The image is then normalized by the batch normalization layer and nonlinearly processed by the ReLU activation function layer.

3. The method for error correction of a synthetic aperture radiometer based on deep learning according to claim 2, characterized in that: In the deep learning network, the weights of the network parameters in each layer are implemented using a loss function defined by the mean square error between the reconstructed microwave brightness temperature image and the original scene microwave brightness temperature image output by the deep learning network. Specifically, the loss function is: In the formula, m Given the number of training samples, the standard backpropagation stochastic gradient descent (SGD) algorithm is used to minimize the loss function. The image shows the reconstructed microwave brightness temperature. This is the original scene's microwave brightness temperature image.

4. The method for error correction of a synthetic aperture radiometer based on deep learning according to claim 3, characterized in that: The deep learning network is trained iteratively a specified number of times using a loss function. When the output of the loss function reaches a stable standard, the current trained deep learning network is saved and verified using a test dataset. The current output two-dimensional integrated aperture microwave brightness temperature image is obtained, and the spatial and frequency domain analyses of the network-reconstructed image are performed to verify the training effect. The trained deep learning network, after being validated, is used to perform error correction on the integrated aperture microwave radiometer system to be calibrated.

Images