Dual-polarized SAR target decomposition method based on model and learning assistance

Abstract

The invention discloses a dual-polarized SAR target decomposition method based on model and learning assistance. Belongs to the technical field of remote sensing imaging and particularly relates to the technical field of dual-polarized SAR imaging remote sensing application. The method comprises the following steps: S1, constructing a dual-polarized SAR data and scattering component pairing data set; s2, constructing a neural network fusing residual connection and dilated convolution; s3, training the convolutional neural network by using the dual-polarized SAR data and the scattering component pairing data set; and S4, performing target decomposition on the dual-polarized SAR data of other regions and other sensors by using the trained convolutional neural network. According to the method, the fast generalized polarization target decomposition result is adopted as a true value, the secondary scattering component of the oriented artificial target can be effectively estimated, the over-estimation of the volume scattering component is reduced, a new idea of model-based dual-polarization SAR target decomposition is provided, and the method has high practicability.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing imaging, and in particular relates to the technical field of dual-polarization SAR imaging remote sensing application. Background Art

[0002] Synthetic Aperture Radar (SAR) has the ability to work around the clock and in all weather conditions and is widely used in fields such as earth observation. Compared with the full-polarization SAR imaging mode, the spaceborne SAR system acquires more dual-polarization SAR data, and the dual-polarization SAR imaging mode has a larger observation width and higher resolution.

[0003] Polarimetric target decomposition is an important method for polarimetric SAR target interpretation and identification. Generalized polarimetric target decomposition is a model-based full polarimetric SAR target decomposition method. It introduces physical parameters that characterize the attitude information of the target scattering structure, establishes a refined scattering model, significantly improves the interpretation performance, and realizes polarimetric identification of multi-attitude artificial targets. However, it uses a nonlinear optimization method with high computational complexity. The fast generalized polarimetric target decomposition method realizes the inversion of model parameters by converting the nonlinear optimization process into a linear solution, but its calculation time is still relatively long. The above-mentioned polarimetric target decomposition method is designed based on full polarimetric SAR data, while there are fewer target decomposition methods based on dual-polarimetric SAR data. In view of the fact that there are relatively few full polarimetric SAR data in the current spaceborne SAR system, and there are more and more dual-polarimetric SAR data, it is urgent to establish a set of target decomposition methods for dual-polarimetric SAR data.

[0004] At present, there are few target decomposition methods for dual-polarization SAR data, especially model-based target decomposition methods. Among the existing model-based dual-polarization SAR target decomposition methods, there is currently no method that can decompose dual-polarization SAR data into secondary scattering components, odd scattering components, and volume scattering components. Moreover, these methods cannot effectively decompose the secondary scattering components of oriented artificial targets, and there is an over-estimation problem of volume scattering components. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a dual-polarization SAR target decomposition method based on a model and learning assistance, characterized in that the method comprises the following steps:

[0006] S1. Construct dual-polarization SAR data and scattering component paired data sets;

[0007] S2, construct a neural network that integrates residual connection and dilated convolution;

[0008] S3, training the convolutional neural network using dual-polarization SAR data and a scattering component paired data set;

[0009] S4. Use the trained convolutional neural network to decompose the dual-polarization SAR data of other regions and other sensors.

[0010] Furthermore, the full polarization SAR data is extracted to obtain dual-polarization SAR data, and the full polarization SAR data is subjected to rapid generalized polarization target decomposition to obtain the scattered components.

[0011] Furthermore, when constructing the dual-polarization SAR data and scattering component paired data set, the dual-polarization SAR data is divided into three types: HH-VH, VV-HV and HH-VV, and each type of dual-polarization SAR data corresponds to four scattering components, and the four scattering components are: secondary scattering component, odd scattering component, volume scattering component and spiral scattering component.

[0012] Furthermore, the neural network integrating residual connection and void convolution includes seven convolution layers, and a ReLU activation function is set after the first to sixth convolution layers, the convolution kernel sizes of all convolution layers are 3×3, the void rates of each convolution layer are 1, 2, 3, 4, 3, 2 and 1 respectively, the number of input channels of the first convolution layer is 4 and the number of output channels is 64, the number of input and output channels of the second to sixth convolution layers are both 64, the number of input channels of the seventh convolution layer is 64 and the number of output channels is 4; wherein, the output of the first layer and the output of the third layer are used as residuals, the residual output is sent to the fourth convolution layer, the output of the fourth convolution layer and the output of the sixth convolution layer are used as residuals, the residual output is sent to the seventh convolution layer and then output.

[0013] Furthermore, when the convolutional neural network is trained using dual-polarization SAR data and a scattering component paired data set, the loss function is:

[0014] LoSS=||f θ (X dual )-X decomp || p ;

[0015] in,‖·‖ p represents the p-norm of the calculation matrix, p∈[0,2], f θ (·) represents the convolutional neural network, X dual represents dual-polarization SAR data, X decomp represents the scattered component.

[0016] Furthermore, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned dual-polarization SAR target decomposition method when executing the computer program.

[0017] Furthermore, the present invention also provides a computer-readable storage medium for storing computer instructions, wherein the computer instructions, when executed by a processor, implement the steps of the above-mentioned dual-polarization SAR target decomposition method.

[0018] The beneficial effects of the present invention are:

[0019] The present invention proposes a dual-polarization SAR target decomposition method based on a model and learning assistance, realizes the dual-polarization SAR target decomposition based on the model, and can accurately obtain the secondary scattering component, odd scattering component and volume scattering component. Firstly, based on the existing fully polarized SAR data, the secondary scattering component, odd scattering component, volume scattering component and spiral scattering component are calculated using a fast generalized polarization target decomposition method. Secondly, a mapping set between dual-polarization SAR data and scattering components is constructed. Then, a neural network integrating residual connection and hole convolution kernel is designed, and the constructed mapping set is used to train the neural network. Finally, the trained neural network is used to test the dual-polarization SAR data of other regions and other sensors to obtain the scattering components of these dual-polarization SAR data.

[0020] The method of the present invention adopts the result of fast generalized polarization target decomposition as the true value, can effectively estimate the secondary scattering component of the oriented artificial target, reduce the over-estimation of the volume scattering component, provides a new idea of ​​dual-polarization SAR target decomposition based on the model, and has high practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of the dual-polarization SAR target decomposition method of the present invention;

[0022] Figure 2 Figure (a) is the full polarization pseudo-color Pauli map of the ALOS-2 San Francisco region, the VV channel amplitude map and HV channel amplitude map of the VV-VH dual-polarization SAR data extracted from it are shown in Figure (b) and Figure (c), respectively, and Figure (d) is the pseudo-color map synthesized by the secondary scattering, odd scattering and volume scattering components of the ALOS-2 San Francisco region;

[0023] Figure 3 This is a diagram of the neural network structure that integrates residual connection and dilated convolution as described in the present invention;

[0024] Figure 4 In the figure, Figure (a) is a PiSAR full polarization pseudo-color Pauli image, Figure (b) is a fast generalized polarization target decomposition result image, and Figure (c) is a target decomposition result image of the method of the present invention;

[0025] Figure 5In the figure, Figures (a1)-(a3) are pseudo-color Pauli images of urban area 1, urban area 2 and forest, respectively; Figures (b1)-(b3) are fast generalized polarization target decomposition result images of urban area 1, urban area 2 and forest, respectively; Figures (c1)-(c3) are target decomposition result images of the method of the present invention of urban area 1, urban area 2 and forest, respectively. DETAILED DESCRIPTION

[0026] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0027] The model-based and learning-assisted dual-polarization SAR target decomposition method proposed in this embodiment uses a fast generalized polarization target decomposition method to calculate the scattered components of the full-polarization SAR data, and then constructs a paired data set of dual-polarization SAR data and scattered components. The designed convolutional neural network is trained using the paired data set, and finally the trained convolutional neural network is tested on the dual-polarization SAR data of other regions and other sensors to obtain the corresponding scattered components.

[0028] like Figure 1 Shown is a flow chart of the method of the present invention. The method of the present invention is introduced below in conjunction with the flow chart.

[0029] S1. Construct a paired dataset of dual-polarization SAR data and scattering components.

[0030] Extract the elements of the fully polarimetric SAR covariance matrix as a vector:

[0031]

[0032] Where X quad is the fully polarized SAR data, S HV It represents the target scattering coefficient when the radar transmits electromagnetic waves in the vertical polarization direction and receives electromagnetic waves in the horizontal polarization direction, S HH It represents the target scattering coefficient when the radar transmits electromagnetic waves in the horizontal polarization direction and receives electromagnetic waves in the horizontal polarization direction, S VV It represents the target scattering coefficient when the radar receives electromagnetic waves in the vertical polarization direction, (·) * represents the conjugate of a variable, |·| represents the calculation of the absolute value of an absolute complex number, log(·) represents the calculation of a logarithmic operation, and angle(·) represents the calculation of the phase of a complex number.

[0033] Dual-polarization SAR data is extracted from the full-polarization SAR data. The dual-polarization SAR data is divided into three types: HH-VH, VV-HV and HH-VV. Taking the VV-HV dual-polarization mode as an example, the dual-polarization SAR data extracted from the full-polarization SAR data is:

[0034]

[0035] Where X dual is dual-polarization SAR data, such as Figure 2 (a) shows the full polarization pseudo-color Pauli image of the ALOS-2 San Francisco region, from which the VV channel amplitude map and HV channel amplitude map of the VV-VH dual-polarization SAR data are extracted. Figure 2 (b) and Figure 2 (c) as shown.

[0036] By performing fast generalized polarization target decomposition on the fully polarized SAR data, the secondary scattering, odd scattering, volume scattering and spiral scattering components are obtained as follows:

[0037] X decomp =[R dbl R odd R vol R hel ];

[0038] Where X decomp is the set of scattered components, R dbl represents the secondary scattering component, R odd represents the odd-order scattered component, R vol represents the volume scattering component, R hel represents the spiral scattering component. Figure 2 (d) shows a pseudo-color image of the composite secondary scattering, odd scattering, and volume scattering components of the ALOS-2 San Francisco region.

[0039] The fast generalized polarimetric target decomposition is a technology for processing SAR data. For specific technical content, please refer to the paper "Polarimetric Model-Based Decomposition with Refined Double-BounceOrientation Angle and Scattering Model" (Remote Sens. 2021, 13, 3070.).

[0040] In order to facilitate the training of convolutional neural networks, X dual and X decomp The images are cropped into slices of 40×40 size and then packaged into a dataset to train the network.

[0041] S2. Construct a neural network that integrates residual connection and hole convolution.

[0042] By designing a neural network that combines residual connections and dilated convolutions θ (·) performs feature extraction on dual-polarization SAR data, where θ is a hyperparameter in the network. Figure 3 As shown, f θ (·) consists of seven convolutional layers, and ReLU activation functions are set after the first to sixth convolutional layers. The convolution kernel size of all convolutional layers is 3×3, and the number of input channels and output channels of the middle layer are both 64. The number of input channels of the first convolutional layer is 4, the number of output channels is 64, and the convolution kernel with a dilation rate of 1 is used. After the input dual-polarization SAR data passes through the first convolution and ReLU activation function, it passes through the second and third convolutional layers in turn, using convolution kernels with dilation rates of 2 and 3 respectively, where the output of the first layer and the output of the third layer are used as residuals. The residual output is sent to the fourth convolutional layer, which uses a convolution kernel with a dilation rate of 4, and then the output of the fourth convolutional layer passes through the fifth and sixth convolutional layers with dilation rates of 3 and 2, and then the output of the fourth convolutional layer and the output of the sixth convolutional layer are used as residuals. The residual output is output through the last convolution layer, where the last convolution layer only includes convolution operations, its hole rate is 1, the number of input channels is 64, and the number of output channels is 4.

[0043] S3. Training the convolutional neural network using dual-polarization SAR data and a scattering component paired data set.

[0044] The dual-polarization SAR data X dual Input convolutional neural network f θ (), the network output f θ (X dual ) and the scattered component set X decomp The loss function is

[0045] Loss=||f θ (X dual )-X decomp || p

[0046] in,‖·‖ p Represents the calculation of the p-norm of the matrix, p∈[0,2].

[0047] The network optimizer uses the Adaptive Moment Estimation (ADAM) optimizer. The optimizer parameters are set to β1 = 0.9, β2 = 0.99 and ∈ = 10 -8 , the learning rate of the network is set to 0.001.

[0048] S4. Use the trained convolutional neural network to decompose the dual-polarization SAR data of other regions and other sensors.

[0049] The beneficial effects of the method of the present invention are further illustrated by comparison.

[0050] The method of the present invention adopts the result of fast generalized polarization target decomposition as the true value, can effectively estimate the secondary scattering component of the oriented artificial target, reduce the over-estimation of the volume scattering component, provides a new idea of ​​dual-polarization SAR target decomposition based on the model, and has high practicality.

[0051] The method of the present invention uses Figure 2 (a) shows the L-band ALOS-2 San Francisco area production training set. In order to verify the practical applicability of the method described in the present invention, use Figure 4 (a) is verified by using the X-band PiSAR data shown in Figure 2.

[0052] As shown in Table 1, the time required for target decomposition of PiSAR data by different methods is compared. It can be seen that the fast generalized polarimetric target decomposition takes 426.10 seconds for full-polarization SAR data, while the method of the present invention takes 8.82 seconds for dual-polarization SAR data, which greatly reduces the target decomposition time and has strong practicality.

[0053] Table 1:

[0054] Fast generalized polarimetric target decomposition Method of the present invention 426.10s 8.82s

[0055] like Figure 4 (a) shows the PiSAR full polarization pseudo-color Pauli image. Figure 4 (b) and Figure 4 (c) shows the fast generalized polarimetric target decomposition and the target decomposition results of the method of the present invention, wherein red represents secondary scattering components, green represents volume scattering components, and blue represents odd scattering components;

[0056] It can be seen that both the fast generalized polarimetric target decomposition method and the method of the present invention can accurately identify artificial targets such as buildings as secondary scattering components, and the method of the present invention can accurately identify forest areas as volume scattering components.

[0057] from Figure 4 In (a), three areas, urban area 1, urban area 2 and forest, are selected for comparative analysis of target decomposition results. Urban area 1 has no orientation, while urban area 2 has orientation. Therefore, urban area 2 has Figure 4 The Pauli diagram in (a) is shown in green. Figure 5The figure shows the target decomposition results of different methods in three areas: urban area 1, urban area 2 and forest, where red represents secondary scattering components, green represents volume scattering components, and blue represents odd scattering components. It can be seen that the target decomposition results of the method of the present invention in urban area 1 and urban area 2 are close to the fast generalized polarization target decomposition results, and in the forest area, the volume scattering components of the decomposition results of the method of the present invention are dominant, while the odd scattering components of the fast generalized polarization target decomposition results are dominant.

[0058] As shown in Table 2, the quantitative decomposition results (%) obtained by different target decomposition methods in different areas are compared. The method of the present invention can accurately estimate the secondary scattering components of artificial targets in urban areas 1 and 2, and reduce the over-estimation of the volume scattering components. For forest areas, the method of the present invention can accurately estimate the volume scattering components, while the fast generalized polarization target method under-estimates the volume scattering components and over-estimates the odd-order scattering components. Figure 5 It appears in blue in (b3).

[0059] Table 2:

[0060]

Claims

1. A dual-polarization SAR target decomposition method based on a model and learning assistance, characterized in that: The method comprises the following steps: S1. Construct dual-polarization SAR data and scattering component paired data sets; S2, construct a neural network that integrates residual connection and dilated convolution; S3, training the convolutional neural network using dual-polarization SAR data and a scattering component paired data set; S4. Use the trained convolutional neural network to decompose the dual-polarization SAR data of other regions and other sensors.

2. The dual-polarization SAR target decomposition method according to claim 1, characterized in that: The full polarization SAR data is extracted to obtain the dual-polarization SAR data, and the full polarization SAR data is subjected to rapid generalized polarization target decomposition to obtain the scattered components.

3. The dual-polarization SAR target decomposition method according to claim 2, characterized in that: When constructing the dual-polarization SAR data and scattering component paired data set, the dual-polarization SAR data is divided into three types: HH-VH, VV-HV and HH-VV, each type of dual-polarization SAR data corresponds to four scattering components, and the four scattering components are: secondary scattering component, odd scattering component, volume scattering component and spiral scattering component.

4. The dual-polarization SAR target decomposition method according to claim 1, characterized in that: The neural network integrating residual connection and void convolution includes seven convolution layers, and ReLU activation functions are set after the first to sixth convolution layers, the convolution kernel sizes of all convolution layers are 3×3, the void rates of each convolution layer are 1, 2, 3, 4, 3, 2 and 1 respectively, the number of input channels of the first convolution layer is 4 and the number of output channels is 64, the number of input and output channels of the second to sixth convolution layers are both 64, the number of input channels of the seventh convolution layer is 64 and the number of output channels is 4; wherein, the output of the first layer and the output of the third layer are used as residuals, the residual output is sent to the fourth convolution layer, the output of the fourth convolution layer and the output of the sixth convolution layer are used as residuals, the residual output is sent to the seventh convolution layer and then output.

5. The dual-polarization SAR target decomposition method according to claim 1, characterized in that: When the convolutional neural network is trained using dual-polarization SAR data and a paired data set of scattered components, the loss function is: Loss=||f θ (X dual )-X decomp || p ; in,‖·‖ p represents the p-norm of the calculation matrix, p∈[0,2], fθ(·) represents the convolutional neural network, X dual represents dual-polarization SAR data, X decomp represents the scattered component.

6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images