A PolSAR urban area classification method and system taking into account radar target azimuth symmetry
By generating a polarization covariance matrix and performing multi-view and filtering processing, the electromagnetic scattering characteristics are extracted in combination with the Cloude-Pottier method, and the urban target marking is used to use the backscattering power, which solves the misclassification of urban targets caused by the assumption of orientation symmetry in the existing technology. The problem has improved the accuracy and physical explanatory nature of PolSAR urban classification.
Patent Information
- Application Number
- CN202310599702.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2043-05-25
AI Technical Summary
The existing PolSAR urban classification method assumes the azimuth symmetry of radar targets, resulting in incomplete non-azimuth symmetric target information and inaccurate scattering feature extraction, resulting in misclassification of urban targets as forests.
By generating a polarization covariance matrix, multi-view, filtering and geocoding are performed, polarization decomposition is performed using eigenvalues and eigenvectors, electromagnetic scattering feature parameters are extracted in combination with Cloude-Pottier method, and Sigmoid function is constructed using backscattering power for Urban target marking and classification.
The accuracy of the electromagnetic scattering feature extraction of radar targets and the accuracy of urban target classification are improved, misclassification is avoided, and the physical interpretability and practicality of the decomposition results are enhanced.
Smart Images

Figure CN116824221B_ABST
Abstract
Description
Technical Field
[0001] The invention discloses a PolSAR urban area classification method and system taking into account the azimuth symmetry of radar targets, which are used for PolSAR urban area target classification and belong to the technical field of full polarization synthetic aperture radar image classification. Background Art
[0002] The urban environment is closely related to human life, and the development of cities also interacts with the surrounding natural environment. Using remote sensing technology to conduct local, national, and even global urban environmental surveys and monitoring is of great significance to the development and construction of cities. Compared with optical sensors, synthetic aperture radar (SAR) has all-day and all-weather ground observation capabilities, and has different penetration and polarization response characteristics for artificial and natural targets. Therefore, using SAR for urban area classification and identification has its unique advantages.
[0003] Fully polarimetric synthetic aperture radar (PolSAR) is developed on the basis of SAR. By sending horizontal (H) and vertical (V) polarized electromagnetic waves respectively and receiving electromagnetic echoes of two polarization states simultaneously, four different polarization combinations of electromagnetic echoes (HH, HV, VH, VV) can be obtained. Fully polarimetric PolSAR can obtain richer ground object target information than SAR and has stronger radar target classification and recognition capabilities. Therefore, PolSAR can obtain higher-precision urban classification and recognition results. However, urban targets and forest targets with large azimuth angles in PolSAR data will produce strong cross-polarization scattering characteristics, which brings difficulties to the classification and recognition of the two in PolSAR data. This is also the hot spot and difficulty in the current research on the interpretation and classification of PolSAR radar target scattering characteristics.
[0004] The traditional PolSAR radar target scattering feature extraction algorithm has four main deficiencies:
[0005] 1. The polarization decomposition algorithms based on eigenvalues and eigenvectors (such as VanZyl, Cloude-Potier and Wang-Davis decomposition) all assume that the radar targets all satisfy the azimuth symmetry and convert the C in the polarization covariance matrix 12 , C 21 , C 23 , C 32 The decomposition is performed after the four elements are set to zero, resulting in incomplete information of non-azimuthally symmetrical urban targets, inaccurate scattering feature extraction, and misclassification of urban areas as forests;
[0006] 2. Both eigenvalue-based and eigenvector-based polarization decomposition algorithms identify scattering mechanisms based on eigenvalues, lacking physical guidance and resulting in low accuracy in interpreting scattering features.
[0007] 3. Model-based PolSAR decomposition methods (such as Freeman-Durden and Yamaguchi decomposition) retain the complete information in PolSAR data, but they also build models based on the assumption of azimuth symmetry, which leads to the failure to fully utilize PolSAR information during the decomposition process, and also causes insufficient accuracy in scattering feature extraction and scattering feature extraction of urban targets;
[0008] 4. The model-based PolSAR decomposition method enhances the PolSAR urban target scattering feature description capability by increasing the number of models. The increase in the number of models will increase the complexity of algorithm execution. On the other hand, the multi-component decomposition algorithm model lacks a one-to-one correspondence with specific radar targets. Although the accuracy of urban target scattering feature extraction has been improved, the physical interpretation of a large number of decomposition result parameters is poor. Summary of the invention
[0009] The purpose of the present invention is to provide a PolSAR urban area classification and recognition method and system taking into account the azimuth symmetry of radar targets, so as to solve the problem that the polarization decomposition algorithms based on eigenvalues and eigenvectors in the prior art all assume that radar targets all satisfy the azimuth symmetry and convert the C in the polarization covariance matrix into 12 , C 21 , C 23 , C 32 The decomposition is performed after the four elements are set to zero, resulting in incomplete information of non-azimuthally symmetrical urban targets, inaccurate extraction of scattering features, and misclassification of urban areas as forests.
[0010] In order to achieve the above object, the technical solution adopted by the present invention is:
[0011] A PolSAR urban area classification method taking into account the azimuth symmetry of radar targets, comprising the following steps:
[0012] Step 1: Transform the polarization scattering matrix [S 2 ] Generate a polarization covariance matrix, perform multi-viewing, filtering and geocoding processing on the polarization covariance matrix in turn to obtain the final polarization covariance matrix;
[0013] Step 2: Decompose the final polarization covariance matrix using the polarization decomposition method to obtain three eigenvalues and eigenvectors, and construct three subcomponents [C 3 ] 1 , [C 3 ] 2 and [C 3 ] 3 ;
[0014] Step 3: According to the electromagnetic scattering mechanism of the radar target, the electromagnetic scattering mechanisms corresponding to the three sub-components are distinguished into single surface scattering, double even scattering and volume scattering, and the backscattering power P of each electromagnetic scattering mechanism is calculated after the distinction. single , P double and P volume ;
[0015] Step 4: Based on the eigenvalues and eigenvectors, the electromagnetic scattering characteristic parameter set of the radar target is extracted according to the Cloude-Pottier decomposition method, including polarization scattering entropy H, polarization scattering anti-entropy A, average azimuth angle α, double polarization direction angle β, surface scattering and even scattering phase difference δ, and surface scattering and volume scattering phase difference γ;
[0016] Step 5: Construct a Sigmoid function based on the backscatter power to mark the urban targets in the full polarization PolSAR image;
[0017] Step 6: Based on the labeled fully polarized PolSAR image, the electromagnetic scattering feature parameter set is used to train the machine learning method, and the fully polarized PolSAR image to be identified is classified based on the trained machine learning method to obtain the final urban target classification result.
[0018] Furthermore, in step 1, the polarization scattering matrix [S 2 ]The specific steps to generate the polarization covariance matrix are:
[0019] Step 1.1, use the orthogonal unit matrix Lexicographic to vectorize the polarization scattering matrix of the full polarization PolSAR image to obtain the polarization scattering vector;
[0020] Polarization scattering matrix of full polarization PolSAR image [S 2 ]for:
[0021]
[0022] In the formula, S HH and S VV is the co-polarization information, S HV and S VH is cross-polarization information, where HH represents transmitting and receiving horizontally polarized electromagnetic waves, VV represents transmitting and receiving vertically polarized electromagnetic waves, HV represents transmitting horizontally polarized electromagnetic waves and receiving vertically polarized electromagnetic waves, and VH represents transmitting vertically polarized electromagnetic waves and receiving horizontally polarized electromagnetic waves;
[0023] The orthogonal identity matrix Lexicographic is:
[0024]
[0025] Based on the polarization scattering matrix and the orthogonal unit matrix Lexicographic, the result of polarization scattering matrix vectorization is:
[0026] K L4 =[S HH S HV S VH S VV ] T
[0027] In the formula, T represents the transposition operation. For a single-station SAR system, according to reciprocity, S HV =S VH , the final polarization scattering vector is:
[0028]
[0029] Step 1.2: Multiply the polarization scattering vector by the conjugate transposed vector of the polarization scattering vector to obtain the polarization covariance matrix:
[0030]
[0031] In the formula, * Represents the conjugate transpose operation.
[0032] Further, in step 1, the polarization covariance matrix is subjected to multi-viewing, filtering and geocoding processing in sequence to obtain the final polarization covariance matrix in the following specific steps:
[0033] Firstly, according to the pixel size and imaging geometry of the polarization covariance matrix, multi-view processing is performed on the polarization covariance matrix in azimuth and range according to the required output spatial resolution;
[0034] Then, the refined Lee filtering algorithm is used to filter the polarization covariance matrix after multi-view processing to remove the coherent speckle noise;
[0035] Finally, the denoised polarization covariance matrix is geocoded to obtain the final polarization covariance matrix.
[0036] Furthermore, the refined Lee filtering algorithm is used to filter the polarization covariance matrix obtained by multi-view processing, where the abstract formula for removing coherent speckle noise is as follows:
[0037] I(x,y)=I 0 (x,y)·υ(x,y)
[0038] Where I(x,y) is the real PolSAR information, that is, the information of each pixel in the polarization covariance matrix obtained by multi-view processing, and I 0(x, y) is the noise-free PolSAR information, which means the target information without noise pollution, and ν(x, y) is the noise information;
[0039] Select a 3×3, 5×5 or 7×7 filter window as the local area, and use the local statistical characteristics of the polarization covariance matrix obtained by multi-view processing to perform speckle filtering. According to the local linear minimum mean square error principle, the polarization covariance matrix after denoising is:
[0040]
[0041] In the formula, is the denoised polarization covariance matrix obtained after filtering, I(x,y) is the center pixel value, which refers to the actual PolSAR measurement value and is the real PolSAR information. is the mean value in the filtering window, and b is the weighting coefficient:
[0042]
[0043] Where var(I(x,y)) represents the local neighborhood variance of I(x,y), represents the noise variance.
[0044] Further, the step 2 of decomposing the final polarization covariance matrix by the polarization decomposition method refers to decomposing the polarization covariance matrix by a matrix decomposition algorithm based on eigenvalues and eigenvectors to obtain three eigenvalues and eigenvectors respectively, and then combining the eigenvalues and eigenvectors to obtain three 3×3 sub-matrices, which constitute the three sub-components of the decomposition result [C 3 ] 1 , [C 3 ] 2 and [C 3 ] 3 , specifically:
[0045] [C 3 ]=λ 1 μ 1 μ 1 * +λ 2 μ 2 μ 2 * +λ 3 μ 3 μ 3 * =[C 3 ] 1 +[C 3 ] 2 +[C 3 ] 3
[0046] In the formula, λi is the eigenvalue, μ i is the final polarization covariance moment [C 3 ]The eigenvector obtained after decomposition is a 3D column vector, where i=1, 2, 3.
[0047] Further, the specific steps of step 3 are:
[0048] Step 3.1, [C 3 ] 1 , [C 3 ] 2 , [C 3 ] 3 The three subcomponents are all 3×3 matrices with nine elements. Each matrix has elements in the second row and second column, namely C 1 22 , C 2 22 and C 3 22 , respectively for C 1 22 , C 2 22 and C 3 22 To judge the size, C 1 22 , C 2 22 and C 3 22 The subcomponent where the element with the largest median value is located is taken as the volume scattering component [C 3 ] volume ;
[0049] Step 3.2: The remaining two subcomponents are defined as [C 3 ] a , [C 3 ] b , [C 3 ] a and [C 3 ] b If the real part of the element in the first row and third column is positive, it is surface scattering; if it is negative, it is even scattering, that is, Re(C a 13 )>0,Re(C b 13 )<0, then [C 3 ] a is the surface scattering [C 3 ] single , and [C 3 ] b Even scattering [C 3 ] doubleOn the contrary, if Re(C a 13 )<0,Re(C b 13 )>0, then [C 3 ] a Even scattering [C 3 ] double , and [C 3 ] b is the surface scattering [C 3 ] single , where Re() is the real part operation;
[0050] Step 3.3: Based on step 3.1 and step 3.2, the final polarization decomposition result is obtained:
[0051] [C 3 ]=[C 3 ] single +[C 3 ] double +[C 3 ] volume ;
[0052] Step 3.4: Calculate the scattered power of various scattering mechanisms based on the final polarization decomposition results:
[0053]
[0054] in, * represents conjugate transpose, single represents surface scattering, double represents even scattering, volume represents volume scattering, λ com is the eigenvalue corresponding to the corresponding electromagnetic scattering mechanism, μ com is the eigenvector corresponding to the corresponding electromagnetic scattering mechanism.
[0055] Further, the specific steps in step 5 are:
[0056] Step 5.1, using the total power value as the independent variable to construct the Sigmoid function of the preliminary classification, and using the relative magnitude relationship between the even scattering and the volume scattering as the independent variable to construct the Sigmoid function of the secondary classification;
[0057] Step 5.2: Preliminary classification:
[0058] The Sigmoid function based on preliminary classification uses the strong and weak relationship of the total backscatter power to achieve preliminary classification of urban targets:
[0059]
[0060] Span=P single +P double +Pvolume
[0061] If the y value is greater than a given threshold, it is an urban target, otherwise it is a non-urban target, where Span is the total power value;
[0062] Step 5.3, secondary classification:
[0063] Based on the initial classification, the urban targets are further classified based on the Sigmoid function of the secondary classification:
[0064]
[0065]
[0066] If the y value is less than the given threshold, it is a small azimuth urban target, otherwise it is a large azimuth urban target, R volume The secondary classification coefficient constructed according to the electromagnetic scattering mechanism is added to the Sigmoid function as an independent variable;
[0067] Based on the results of secondary classification, urban targets in fully polarimetric PolSAR images are marked.
[0068] Furthermore, the machine learning methods include support vector machines, naive Bayes and decision trees.
[0069] A PolSAR urban area classification system taking into account radar target azimuth symmetry, comprising:
[0070] Polarization covariance matrix generation module: convert the polarization scattering matrix [S 2 ] Generate a polarization covariance matrix, perform multi-viewing, filtering and geocoding processing on the polarization covariance matrix in turn to obtain the final polarization covariance matrix;
[0071] Subcomponent decomposition module: The final polarization covariance matrix is decomposed using the polarization decomposition method to obtain three eigenvalues and eigenvectors, and three subcomponents are constructed from the eigenvalues and eigenvectors [C 3 ] 1 , [C 3 ] 2 and [C 3 ] 3 ;
[0072] Discrimination module: According to the electromagnetic scattering mechanism of the radar target, the electromagnetic scattering mechanism corresponding to the three sub-components is distinguished into single surface scattering, double even scattering and volume scattering, and the backscattering power P of each electromagnetic scattering mechanism is calculated after the distinction. single , P double and P volume ;
[0073] Electromagnetic scattering feature extraction module: Based on eigenvalues and eigenvectors, the electromagnetic scattering feature parameter set of the radar target is extracted according to the Cloude-Pottier decomposition method, including polarization scattering entropy H, polarization scattering anti-entropy A, average azimuth angle α, double polarization direction angle β, surface scattering and even-order scattering phase difference δ, and surface scattering and volume scattering phase difference γ;
[0074] Marking module: Construct a Sigmoid function based on the backscatter power to mark urban targets in the full polarization PolSAR image;
[0075] Classification module: Based on the labeled fully polarized PolSAR images, the machine learning method is trained using the electromagnetic scattering feature parameter set, and the fully polarized PolSAR images to be identified are classified based on the trained machine learning method to obtain the final urban target classification results.
[0076] Compared with the prior art, the advantages of the present invention are:
[0077] First, the present invention abandons the assumption of azimuth symmetry, uses the final polarization covariance matrix of the PolSAR image of the radar target (which has been processed by multi-viewing, filtering and geocoding in sequence to make the PolSAR image more detailed and clear) for polarization decomposition, and makes full use of the PolSAR backscattering information to extract the electromagnetic scattering features of the radar target at the pixel level, thereby ensuring the accuracy of the electromagnetic scattering feature extraction and classification of the radar target from the source;
[0078] Second, the decomposition algorithm based on eigenvalues and eigenvectors of the present invention is based on the mathematical method of matrix decomposition, which avoids the problem of building a complex physical scattering model, simplifies the PolSAR scattering feature (electromagnetic scattering feature) extraction process, and improves the practicality and applicability of the PolSAR scattering feature extraction algorithm;
[0079] Third, the present invention uses physical models and electromagnetic scattering mechanisms to identify decomposed subcomponents, enhances the physical interpretability of decomposition results, and further improves the accuracy of electromagnetic scattering feature extraction of radar targets;
[0080] Fourth, the present invention comprehensively utilizes the backscatter power of radar targets to construct a Sigmoid function to mark urban targets, thereby increasing the distance between radar target classes, so as to further improve the accuracy of urban target classification;
[0081] 5. The present invention uses the scattering mechanism to mark the target, which reduces the difficulty of sample selection and belongs to a semi-supervised classification method, thus improving the practicality of the method;
[0082] 6. The present invention uses the orthogonal unit matrix Lexicographic to vectorize the polarization scattering matrix to obtain the polarization scattering vector. The advantage of using the polarization scattering vector is that the polarization scattering vector directly contains the complex amplitude information of the scattering matrix, which directly corresponds to the measurement value of the system, and is convenient for interpreting the scattering characteristics of the fully polarized PolSAR image. BRIEF DESCRIPTION OF THE DRAWINGS
[0083] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0084] Figure 1 It is a schematic diagram of the framework logic of the present invention;
[0085] Figure 2 This is the true value map of the study area: the image is obtained by high-resolution optical image processing of the study area. White and gray represent urban building targets, and black represents non-urban targets;
[0086] Figure 3 for Figure 2 Urban classification results of ALOS2-PALSARL-band fully polarimetric PolSAR image. The classification method is support vector machine (SVM) method. White represents urban targets with small azimuths, gray represents urban targets with large azimuths, and black represents non-urban targets. Urban areas with large azimuths are effectively distinguished from vegetation areas. The method proposed in the present invention has high urban target classification accuracy. DETAILED DESCRIPTION
[0087] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 scope of protection of the present invention.
[0088] Definition 1: Polarization Decomposition
[0089] The electromagnetic echo information stored in each pixel in the PolSAR data is a combination of the electromagnetic echo information of multiple radar targets in the imaging unit. Therefore, PolSAR radar targets are usually distributed targets, and there are multiple scattering centers in each imaging unit. In order to classify and identify PolSAR radar targets more accurately, it is necessary to accurately extract the backscattering characteristics of each scattering center. The most direct and effective PolSAR scattering feature extraction algorithm is the polarization decomposition algorithm. The polarization decomposition theory believes that the scattering mechanisms of each scattering center are independent of each other, and the backscattering information is the incoherent superposition of multiple mechanisms. The polarization decomposition algorithm uses a matrix decomposition algorithm to represent the backscattering information as the sum of multiple sub-components.
[0090] Definition 2: Azimuthal Symmetry
[0091] In PolSAR data, if the ground object meets both reflection symmetry and rotational symmetry, the target is considered to be an azimuthally symmetric target, otherwise it is an asymmetric target. Generally, natural targets meet azimuth symmetry, such as water surface, bare land, grassland and forest. The backscattering characteristics of these radar targets are almost unaffected by the radar flight direction and radar observation direction. However, artificial targets almost do not meet azimuth symmetry. In particular, for rectangular urban building targets, the backscattering characteristics of these radar targets change with the change of radar flight direction and radar observation direction.
[0092] A PolSAR urban area classification method taking into account the azimuth symmetry of radar targets, comprising the following steps:
[0093] Step 1: Transform the polarization scattering matrix [S 2 ] Generate a polarization covariance matrix, perform multi-viewing, filtering and geocoding processing on the polarization covariance matrix in turn to obtain the final polarization covariance matrix;
[0094] The polarization scattering matrix [S 2 ]The specific steps to generate the polarization covariance matrix are:
[0095] Step 1.1, use the orthogonal unit matrix Lexicographic to vectorize the polarization scattering matrix of the full polarization PolSAR image to obtain the polarization scattering vector;
[0096] Polarization scattering matrix of full polarization PolSAR image [S 2 ]for:
[0097]
[0098] In the formula, S HH and S VV is the co-polarization information, SHV and S VH is cross-polarization information, where HH represents transmitting and receiving horizontally polarized electromagnetic waves, VV represents transmitting and receiving vertically polarized electromagnetic waves, HV represents transmitting horizontally polarized electromagnetic waves and receiving vertically polarized electromagnetic waves, and VH represents transmitting vertically polarized electromagnetic waves and receiving horizontally polarized electromagnetic waves;
[0099] The orthogonal identity matrix Lexicographic is:
[0100]
[0101] Based on the polarization scattering matrix and the orthogonal unit matrix Lexicographic, the result of polarization scattering matrix vectorization is:
[0102] K L4 =[S HH S HV S VH S VV ] T
[0103] In the formula, T represents the transposition operation. For a single-station SAR system, according to reciprocity, S HV =S VH , the final polarization scattering vector is:
[0104]
[0105] Step 1.2: Multiply the polarization scattering vector by the conjugate transposed vector of the polarization scattering vector to obtain the polarization covariance matrix:
[0106]
[0107] In the formula, * Represents the conjugate transpose operation.
[0108] The specific steps of multi-viewing, filtering and geocoding the polarization covariance matrix to obtain the final polarization covariance matrix are as follows:
[0109] Firstly, according to the pixel size and imaging geometry of the polarization covariance matrix, multi-view processing is performed on the polarization covariance matrix in azimuth and range according to the required output spatial resolution;
[0110] Then, the refined Lee filtering algorithm is used to filter the polarization covariance matrix after multi-view processing to remove the coherent speckle noise. The refined Lee filtering algorithm is used to filter the polarization covariance matrix obtained by multi-view processing. The abstract formula for removing the coherent speckle noise is as follows:
[0111] I(x,y)=I 0(x,y)·υ(x,y)
[0112] Where I(x,y) is the real PolSAR information, that is, the information of each pixel in the polarization covariance matrix obtained by multi-view processing, and I 0 (x, y) is the noise-free PolSAR information, which means the target information without noise pollution, and ν(x, y) is the noise information;
[0113] Select a 3×3, 5×5 or 7×7 filter window as the local area, and use the local statistical characteristics of the polarization covariance matrix obtained by multi-view processing to perform speckle filtering. According to the local linear minimum mean square error principle, the polarization covariance matrix after denoising is:
[0114]
[0115] In the formula, is the denoised polarization covariance matrix obtained after filtering, I(x,y) is the center pixel value, which refers to the actual PolSAR measurement value and is the real PolSAR information. is the mean value in the filtering window, and b is the weighting coefficient:
[0116]
[0117] Where var(I(x,y)) represents the local neighborhood variance of I(x,y), represents the noise variance.
[0118] Finally, the denoised polarization covariance matrix is geocoded to obtain the final polarization covariance matrix.
[0119] Step 2: Decompose the final polarization covariance matrix using the polarization decomposition method to obtain three eigenvalues and eigenvectors, and construct three subcomponents [C 3 ] 1 , [C 3 ] 2 and [C 3 ] 3 ; Using the polarization decomposition method to decompose the final polarization covariance matrix means that the matrix decomposition algorithm based on eigenvalues and eigenvectors is used to decompose the polarization covariance matrix to obtain three eigenvalues and eigenvectors respectively, and then the eigenvalues and eigenvectors are combined to obtain three 3×3 sub-matrices, which constitute the three sub-components of the decomposition result [C 3 ] 1 , [C 3 ] 2 and [C 3 ] 3 , specifically:
[0120]
[0121] In the formula, λ i is the eigenvalue, μ i is the final polarization covariance moment [C 3 ]The eigenvector obtained after decomposition is a 3D column vector, where i=1, 2, 3.
[0122] Step 3: According to the electromagnetic scattering mechanism of the radar target, the electromagnetic scattering mechanisms corresponding to the three sub-components are distinguished into single surface scattering, double even scattering and volume scattering, and the backscattering power P of each electromagnetic scattering mechanism is calculated after the distinction. single , P double and P volume ;
[0123] The specific steps are:
[0124] Step 3.1, [C 3 ] 1 , [C 3 ] 2 , [C 3 ] 3 The three subcomponents are all 3×3 matrices with nine elements. Each matrix has elements in the second row and second column, namely C 1 22 , C 2 22 and C 3 22 , respectively for C 1 22 , C 2 22 and C 3 22 To judge the size, C 1 22 , C 2 22 and C 3 22 The subcomponent where the element with the largest median value is located is taken as the volume scattering component [C 3 ] volume ;
[0125] Step 3.2: The remaining two subcomponents are defined as [C 3 ] a , [C 3 ] b , according to the elements in the first row and third column of each matrix, that is, C a 13 and C b 13The real part sign of is used to distinguish the electromagnetic scattering mechanism. The even scattering in the full polarization PolSAR image comes from the dihedral ground object target. The dihedral target will cause the V polarization electromagnetic wave to have a 180° phase deflection, but will not affect the H polarization electromagnetic wave phase. The surface scattering comes from the crown surface or the ground, and does not change the electromagnetic wave phase. Therefore, in the full polarization PolSAR image, if the real part value of the element in the first row and the third column is positive, it is surface scattering, and if it is negative, it is even scattering, that is, Re(C a 13 )>0,Re(C b 13 )<0, then [C 3 ] a is the surface scattering [C 3 ] single , and [C 3 ] b Even scattering [C 3 ] double On the contrary, if Re(C a 13 )<0,Re(C b 13 )>0, then [C 3 ] a Even scattering [C 3 ] double , and [C 3 ] b is the surface scattering [C 3 ] single , where Re() is the real part operation;
[0126] Step 3.3: Based on step 3.1 and step 3.2, the final polarization decomposition result is obtained:
[0127] [C 3 ]=[C 3 ] single +[C 3 ] double +[C 3 ] volume ;
[0128] Step 3.4: Calculate the scattered power of various scattering mechanisms based on the final polarization decomposition results:
[0129]
[0130] in, * represents conjugate transpose, single represents surface scattering, double represents even scattering, volume represents volume scattering, λ com is the eigenvalue corresponding to the corresponding electromagnetic scattering mechanism, μ comis the eigenvector corresponding to the corresponding electromagnetic scattering mechanism.
[0131] Step 4: Based on the eigenvalues and eigenvectors, the Cloude-Pottier decomposition method is used to extract the electromagnetic scattering characteristic parameter set of the radar target, including polarization scattering entropy H, polarization scattering anti-entropy A, average azimuth angle α, double polarization direction angle β, surface scattering and even scattering phase difference δ, and surface scattering and volume scattering phase difference γ;
[0132] Step 5: Construct a Sigmoid function based on the backscatter power to mark the urban targets in the full polarization PolSAR image;
[0133] The specific steps are:
[0134] Step 5.1, using the total power value as the independent variable to construct the Sigmoid function of the preliminary classification, and using the relative magnitude relationship between the even scattering and the volume scattering as the independent variable to construct the Sigmoid function of the secondary classification;
[0135] Step 5.2: Preliminary classification:
[0136] Targets on the water surface or bare land generate strong surface scattered echoes due to mirror reflection, while vegetation targets scatter and absorb electromagnetic waves, which also weakens the intensity of electromagnetic echoes. In contrast, urban targets generate strong electromagnetic echo information. Therefore, the Sigmoid function based on preliminary classification uses the strong and weak relationship of the total backscattered power to achieve preliminary classification of urban targets:
[0137]
[0138] Span=P single +P double +P volume
[0139] If the y value is greater than 0.6, it is an urban target, otherwise it is a non-urban target, where Span is the total power value;
[0140] Step 5.3, secondary classification:
[0141] When the wall of the urban target is parallel to the flight direction of the radar, strong even-order scattering echo and surface scattering echo are generated, and no scattering echo is generated. When the wall of the urban target is not parallel to the flight direction of the radar, strong body scattering echo and surface scattering echo are generated, and no even-order scattering echo is generated. Therefore, on the basis of the initial classification, the relative size relationship between even-order scattering and body scattering is used to realize the further classification of urban targets with polarization azimuth and non-polarization azimuth, that is, the further classification of urban targets based on the Sigmoid function of secondary classification:
[0142]
[0143]
[0144] If the y value is less than 0.6, it is a small azimuth urban target; otherwise, it is a large azimuth urban target. volume The secondary classification coefficient constructed according to the electromagnetic scattering mechanism is added to the Sigmoid function as an independent variable;
[0145] Based on the results of secondary classification, urban targets in fully polarimetric PolSAR images are marked.
[0146] Step 6: Based on the labeled full-polarization PolSAR image, the electromagnetic scattering feature parameter set is used to train the machine learning method, and the full-polarization PolSAR image to be identified is classified based on the trained machine learning method to obtain the final urban target classification result. The machine learning method includes support vector machine, naive Bayes and decision tree.
[0147] The present invention firstly improves the PolSAR decomposition algorithm based on the characteristics of non-azimuth symmetry of urban targets and azimuth symmetry of forest targets, and uses the azimuth symmetry information of radar targets to improve the accuracy of PolSAR radar target scattering feature extraction. Then, according to the different penetration of electromagnetic waves into urban and forest targets, the PolSAR electromagnetic scattering feature parameters are used to construct a Sigmoid function to preliminarily identify urban targets and complete the unsupervised training sample selection. Finally, according to the different scattering mechanisms of different radar targets, machine learning methods (SVM, K-means or Wishart, etc.) are used for training and prediction to obtain the final classification results of PolSAR urban targets.
Claims
1. A PolSAR urban area classification method taking into account the azimuth symmetry of radar targets, characterized in that: Follow these steps: Step 1: Generate a polarization covariance matrix from the polarization scattering matrix [S2] of the full polarization PolSAR image, and perform multi-view, filtering and geocoding processing on the polarization covariance matrix in turn to obtain the final polarization covariance matrix; Step 2: Decompose the final polarization covariance matrix using the polarization decomposition method to obtain three eigenvalues and eigenvectors, and construct three subcomponents [C3]1, [C3]2, and [C3]3 from the eigenvalues and eigenvectors; Step 3: According to the electromagnetic scattering mechanism of the radar target, the electromagnetic scattering mechanisms corresponding to the three sub-components are distinguished into single surface scattering, double even scattering and volume scattering, and the backscattering power P of each electromagnetic scattering mechanism is calculated after the distinction. single , P double and P volume ; Step 4: Based on the eigenvalues and eigenvectors, the electromagnetic scattering characteristic parameter set of the radar target is extracted according to the Cloude-Pottier decomposition method, including polarization scattering entropy H, polarization scattering anti-entropy A, average azimuth angle α, double polarization direction angle β, surface scattering and even scattering phase difference δ, and surface scattering and volume scattering phase difference γ; Step 5: Construct a Sigmoid function based on the backscatter power to mark the urban targets in the full polarization PolSAR image; Step 6: Based on the labeled full-polarization PolSAR image, the electromagnetic scattering feature parameter set is used to train the machine learning method, and the full-polarization PolSAR image to be identified is classified based on the trained machine learning method to obtain the final urban target classification result; In step 1, the specific steps of generating a polarization covariance matrix from the polarization scattering matrix [S2] of the fully polarized PolSAR image are as follows: Step 1.1, use the orthogonal unit matrix Lexicographic to vectorize the polarization scattering matrix of the full polarization PolSAR image to obtain the polarization scattering vector; The polarization scattering matrix [S2] of the fully polarimetric PolSAR image is: In the formula, S HH and S VV is the co-polarization information, S HV and S VH is cross-polarization information, where HH represents transmitting and receiving horizontally polarized electromagnetic waves, VV represents transmitting and receiving vertically polarized electromagnetic waves, HV represents transmitting horizontally polarized electromagnetic waves and receiving vertically polarized electromagnetic waves, and VH represents transmitting vertically polarized electromagnetic waves and receiving horizontally polarized electromagnetic waves; The orthogonal identity matrix Lexicographic is: Based on the polarization scattering matrix and the orthogonal unit matrix Lexicographic, the result of polarization scattering matrix vectorization is: K L4 =[S HH S HV S VH S VV ] T In the formula, T represents the transposition operation. For a single-station SAR system, according to reciprocity, S HV =S VH , the final polarization scattering vector is: Step 1.2: Multiply the polarization scattering vector by the conjugate transposed vector of the polarization scattering vector to obtain the polarization covariance matrix: In the formula, * represents the conjugate transpose operation; In step 1, the polarization covariance matrix is subjected to multi-viewing, filtering and geocoding processing in sequence to obtain the final polarization covariance matrix in the following specific steps: Firstly, according to the pixel size and imaging geometry of the polarization covariance matrix, multi-view processing is performed on the polarization covariance matrix in azimuth and range according to the required output spatial resolution; Then, the refined Lee filtering algorithm is used to filter the polarization covariance matrix after multi-view processing to remove the coherent speckle noise; Finally, the denoised polarization covariance matrix is geocoded to obtain the final polarization covariance matrix; The refined Lee filtering algorithm is used to filter the polarization covariance matrix obtained by multi-view processing, where the abstract formula for removing coherent speckle noise is as follows: I(x,y)=I0(x,y)·υ(x,y) Where I(x, y) is the real PolSAR information, that is, the information of each pixel in the polarization covariance matrix obtained by multi-view processing, I0(x, y) is the noise-free PolSAR information, that is, the target information without noise pollution, and ν(x, y) is the noise information; Select a 3×3, 5×5 or 7×7 filter window as the local area, and use the local statistical characteristics of the polarization covariance matrix obtained by multi-view processing to perform speckle filtering. According to the local linear minimum mean square error principle, the polarization covariance matrix after denoising is: In the formula, is the denoised polarization covariance matrix obtained after filtering, I(x,y) is the center pixel value, which refers to the actual PolSAR measurement value and is the real PolSAR information. is the mean value in the filtering window, and b is the weighting coefficient: Where var(I(x,y)) represents the local neighborhood variance of I(x,y), represents the noise variance; Decomposing the final polarization covariance matrix by the polarization decomposition method in step 2 means decomposing the polarization covariance matrix by a matrix decomposition algorithm based on eigenvalues and eigenvectors to obtain three eigenvalues and eigenvectors respectively, and then combining the eigenvalues and eigenvectors to obtain three 3×3 sub-matrices, which constitute the three sub-components [C3]1, [C3]2 and [C3]3 of the decomposition result, specifically: In the formula, λ i is the eigenvalue, μ i is the eigenvector obtained after decomposition of the final polarization covariance moment [C3], i.e., a 3D column vector, where i = 1, 2, 3; The specific steps of step 3 are: Step 3.1, [C3]1, [C3]2, [C3]3 are all 3×3 matrices with nine elements. Each matrix has an element in the second row and second column, which are C 1 22 , C 2 22 and C 3 22 , respectively for C 1 22 , C 2 22 and C 3 22 To judge the size, C 1 22 , C 2 22 and C 3 22 The subcomponent of the element with the largest median value is taken as the volume scattering component [C3] volume ; Step 3.2: The remaining two subcomponents are defined as [C3] a , [C3] b , [C3] a and [C3] b If the real part of the element in the first row and third column is positive, it is surface scattering; if it is negative, it is even scattering, that is, Re(C a 13 )>0,Re(C b 13 )<0, then [C3] a is surface scattering [C3] single , and [C3] b Even scattering [C3] double On the contrary, if Re(C a 13 )<0,Re(C b 13 )>0, then [C3] a Even scattering [C3] double , and [C3] b is surface scattering [C3] single , where Re() is the real part operation; Step 3.3: Based on step 3.1 and step 3.2, the final polarization decomposition result is obtained: [C3]=[C3] single +[C3] double +[C3] volume ; Step 3.4: Calculate the scattered power of various scattering mechanisms based on the final polarization decomposition results: in, * represents conjugate transpose, single represents surface scattering, double represents even scattering, volume represents volume scattering, λ com is the eigenvalue corresponding to the corresponding electromagnetic scattering mechanism, μ com is the eigenvector corresponding to the corresponding electromagnetic scattering mechanism; The specific steps in step 5 are: Step 5.1, using the total power value as the independent variable to construct the Sigmoid function of the preliminary classification, and using the relative magnitude relationship between the even scattering and the volume scattering as the independent variable to construct the Sigmoid function of the secondary classification; Step 5.2: Preliminary classification: The Sigmoid function based on preliminary classification uses the strong and weak relationship of the total backscatter power to achieve preliminary classification of urban targets: Span=P single +P double +P volume If the y value is greater than a given threshold, it is an urban target, otherwise it is a non-urban target, where Span is the total power value; Step 5.3, secondary classification: Based on the initial classification, the urban targets are further classified based on the Sigmoid function of the secondary classification: If the y value is less than the given threshold, it is a small azimuth urban target, otherwise it is a large azimuth urban target, R volume The secondary classification coefficient constructed according to the electromagnetic scattering mechanism is added to the Sigmoid function as an independent variable; Based on the results of secondary classification, urban targets in fully polarimetric PolSAR images are marked.
2. The PolSAR urban area classification method taking into account the azimuth symmetry of radar targets according to claim 1, characterized in that: The machine learning methods include support vector machines, naive Bayes and decision trees.
3. A PolSAR urban area classification system taking into account radar target azimuth symmetry, characterized in that: include: Polarization covariance matrix generation module: The polarization scattering matrix [S2] of the full polarization PolSAR image is used to generate the polarization covariance matrix, and the polarization covariance matrix is processed by multi-viewing, filtering and geocoding in sequence to obtain the final polarization covariance matrix; Sub-component decomposition module: The final polarization covariance matrix is decomposed using the polarization decomposition method to obtain three eigenvalues and eigenvectors, and three sub-components [C3]1, [C3]2 and [C3]3 are constructed from the eigenvalues and eigenvectors; Discrimination module: According to the electromagnetic scattering mechanism of the radar target, the electromagnetic scattering mechanism corresponding to the three sub-components is distinguished into single surface scattering, double even scattering and volume scattering, and the backscattering power P of each electromagnetic scattering mechanism is calculated after the distinction. single , P double and P volume ; Electromagnetic scattering feature extraction module: Based on eigenvalues and eigenvectors, the electromagnetic scattering feature parameter set of the radar target is extracted according to the Cloude-Pottier decomposition method, including polarization scattering entropy H, polarization scattering anti-entropy A, average azimuth angle α, double polarization direction angle β, surface scattering and even-order scattering phase difference δ, and surface scattering and volume scattering phase difference γ; Marking module: Construct a Sigmoid function based on the backscatter power to mark urban targets in the full polarization PolSAR image; Classification module: Based on the labeled fully polarized PolSAR images, the machine learning method is trained using the electromagnetic scattering feature parameter set, and the fully polarized PolSAR images to be identified are classified based on the trained machine learning method to obtain the final urban target classification results; In the polarization covariance matrix generation module, the specific steps of generating the polarization covariance matrix from the polarization scattering matrix [S2] of the full polarization PolSAR image are as follows: Step 1.1, use the orthogonal unit matrix Lexicographic to vectorize the polarization scattering matrix of the full polarization PolSAR image to obtain the polarization scattering vector; The polarization scattering matrix [S2] of the fully polarimetric PolSAR image is: In the formula, S HH and S VV is the co-polarization information, S HV and S VH is cross-polarization information, where HH represents transmitting and receiving horizontally polarized electromagnetic waves, VV represents transmitting and receiving vertically polarized electromagnetic waves, HV represents transmitting horizontally polarized electromagnetic waves and receiving vertically polarized electromagnetic waves, and VH represents transmitting vertically polarized electromagnetic waves and receiving horizontally polarized electromagnetic waves; The orthogonal identity matrix Lexicographic is: Based on the polarization scattering matrix and the orthogonal unit matrix Lexicographic, the result of polarization scattering matrix vectorization is: K L4 =[S HH S HV S VH S VV ] T In the formula, T represents the transposition operation. For a single-station SAR system, according to reciprocity, S HV =S VH , the final polarization scattering vector is: Step 1.2: Multiply the polarization scattering vector by the conjugate transposed vector of the polarization scattering vector to obtain the polarization covariance matrix: In the formula, * represents the conjugate transpose operation; In the polarization covariance matrix generation module, the polarization covariance matrix is subjected to multi-viewing, filtering and geocoding processing in sequence to obtain the final polarization covariance matrix in the following specific steps: Firstly, according to the pixel size and imaging geometry of the polarization covariance matrix, multi-view processing is performed on the polarization covariance matrix in azimuth and range according to the required output spatial resolution; Then, the refined Lee filtering algorithm is used to filter the polarization covariance matrix after multi-view processing to remove the coherent speckle noise; Finally, the denoised polarization covariance matrix is geocoded to obtain the final polarization covariance matrix; The refined Lee filtering algorithm is used to filter the polarization covariance matrix obtained by multi-view processing, where the abstract formula for removing coherent speckle noise is as follows: I(x,y)=I0(x,y)·υ(x,y) Where I(x, y) is the real PolSAR information, that is, the information of each pixel in the polarization covariance matrix obtained by multi-view processing, I0(x, y) is the noise-free PolSAR information, that is, the target information without noise pollution, and ν(x, y) is the noise information; Select a 3×3, 5×5 or 7×7 filter window as the local area, and use the local statistical characteristics of the polarization covariance matrix obtained by multi-view processing to perform speckle filtering. According to the local linear minimum mean square error principle, the polarization covariance matrix after denoising is: In the formula, is the denoised polarization covariance matrix obtained after filtering, I(x,y) is the center pixel value, which refers to the actual PolSAR measurement value and is the real PolSAR information. is the mean value in the filtering window, and b is the weighting coefficient: Where var(I(x,y)) represents the local neighborhood variance of I(x,y), represents the noise variance; Decomposing the final polarization covariance matrix by using the polarization decomposition method in the sub-component decomposition module means that the polarization covariance matrix is decomposed by a matrix decomposition algorithm based on eigenvalues and eigenvectors to obtain three eigenvalues and eigenvectors respectively, and then three 3×3 sub-matrices are obtained by combining the eigenvalues and eigenvectors to form three sub-components [C3]1, [C3]2 and [C3]3 of the decomposition result, specifically: In the formula, λ i is the eigenvalue, μ i is the eigenvector obtained after decomposition of the final polarization covariance moment [C3], i.e., a 3D column vector, where i = 1, 2, 3; The specific implementation steps of the discrimination module are: Step 3.1, [C3]1, [C3]2, [C3]3 are all 3×3 matrices with nine elements. Each matrix has an element in the second row and second column, which are C 1 22 , C 2 22 and C 3 22 , respectively for C 1 22 , C 2 22 and C 3 22 To judge the size, C 1 22 , C 2 22 and C 3 22 The subcomponent of the element with the largest median value is taken as the volume scattering component [C3] volume ; Step 3.2: The remaining two subcomponents are defined as [C3] a , [C3] b , [C3] a and [C3] b If the real part of the element in the first row and third column is positive, it is surface scattering; if it is negative, it is even scattering, that is, Re(C a 13 )>0,Re(C b 13 )<0, then [C3] a is surface scattering [C3] single , and [C3] b Even scattering [C3] double On the contrary, if Re(C a 13 )<0,Re(C b 13 )>0, then [C3] a Even scattering [C3] double , and [C3] b is surface scattering [C3] single , where Re() is the real part operation; Step 3.3: Based on step 3.1 and step 3.2, the final polarization decomposition result is obtained: [C3]=[C3] single +[C3] double +[C3] volume ; Step 3.4: Calculate the scattered power of various scattering mechanisms based on the final polarization decomposition results: in, * represents conjugate transpose, single represents surface scattering, double represents even scattering, volume represents volume scattering, λ com is the eigenvalue corresponding to the corresponding electromagnetic scattering mechanism, μ com is the eigenvector corresponding to the corresponding electromagnetic scattering mechanism; The specific implementation steps in the marking module are: Step 5.1, using the total power value as the independent variable to construct the Sigmoid function of the preliminary classification, and using the relative magnitude relationship between the even scattering and the volume scattering as the independent variable to construct the Sigmoid function of the secondary classification; Step 5.2: Preliminary classification: The Sigmoid function based on preliminary classification uses the strong and weak relationship of the total backscatter power to achieve preliminary classification of urban targets: Span=P single +P double +P volume If the y value is greater than a given threshold, it is an urban target, otherwise it is a non-urban target, where Span is the total power value; Step 5.3, secondary classification: Based on the initial classification, the urban targets are further classified based on the Sigmoid function of the secondary classification: If the y value is less than the given threshold, it is a small azimuth urban target, otherwise it is a large azimuth urban target, R volume The secondary classification coefficient constructed according to the electromagnetic scattering mechanism is added to the Sigmoid function as an independent variable; Based on the results of secondary classification, urban targets in fully polarimetric PolSAR images are marked.
