Deep learning-based polarimetric SAR image classification methods, devices, and equipment
By performing reflection symmetry decomposition and feature extraction on polarimetric SAR image data, and combining it with convolutional neural network training, the problem of insufficient classification accuracy of polarimetric SAR images in existing technologies has been solved, and more efficient target classification has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAT UNIV OF DEFENSE TECH
- Filing Date
- 2024-01-19
- Publication Date
- 2026-07-17
AI Technical Summary
Existing polarimetric SAR image classification methods are insufficient in terms of accuracy and sample labeling requirements, and the development of deep learning is relatively lagging behind, requiring a lot of human and material resources to ensure the accuracy of the model.
By acquiring the original fully polarimetric dataset, performing preprocessing, and then using the reflection symmetric decomposition algorithm to extract polarimetric scattering features, the elements in the polarimetric coherence matrix are calculated to obtain polarimetric features such as horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, and total polarization power. A training dataset is then constructed and trained using a convolutional neural network to achieve target classification of polarimetric SAR images.
It improves the accuracy of polarimetric SAR image classification, reduces the consumption of manpower and material resources, and enhances the classification effect.
Smart Images
Figure CN117911769B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of polarimetric SAR image classification technology, and in particular to a polarimetric SAR image classification method, apparatus and device based on deep learning. Background Technology
[0002] Polarimetric Synthetic Aperture Radar (PolSAR) can acquire complete polarization information of ground targets, enabling all-weather, all-day active target detection, offering significant advantages over other remote sensing methods. This method employs active microwave remote sensing, utilizing coherent measurements to acquire imaging data for Earth observation, and is a crucial component of current Earth observation technologies. PolSAR can also acquire the scattering characteristics of electromagnetic waves of arbitrary polarization along the target observation direction. PolSAR can actively transmit electromagnetic wave information to the target surface and then observe ground targets by receiving the reflected echo information.
[0003] Single-channel or multi-channel synthetic aperture radar (SAR) can provide one or more two-dimensional complex matrix data, known as single-polarization or multi-polarization SAR images. The amplitude and phase of each pixel depend on the sum of the backscattered waves from all targets within the SAR resolution cell. SAR images contain far richer ground feature information than optical images, primarily in three categories: first, the spatial geometric information provided by the "image" itself; second, the radar backscattering coefficients provided by the pixel amplitude and phase; and third, the ground feature polarimetric scattering characteristics provided by the multi-channel SAR image.
[0004] In SAR target interpretation, PolSAR land cover classification is an important research area, providing crucial information for monitoring land cover changes and serving as a reference for various decision-making processes. Currently, PolSAR classification methods mainly include those based on polarization characteristics, statistical distribution characteristics of PolSAR data, and deep learning-based classification methods.
[0005] In recent years, advancements in computer hardware technology have spurred the development of a series of excellent deep learning algorithms. Deep learning methods extract information about ground features through a certain number of network layers and utilize these extracted deep features to classify ground features in images. Compared to traditional classification methods or machine learning methods, they can make better use of the scattering characteristics inherent in ground targets. Unlike optical imagery, the development of SAR deep learning is relatively lagging. Currently, there are fewer data sources available for classification, and higher requirements are placed on the accuracy of sample annotation. This often requires extensive domain knowledge and significant human and material resources to ensure the accuracy of the model. Summary of the Invention
[0006] Therefore, it is necessary to provide a deep learning-based polarimetric SAR image classification method, apparatus, and device that can improve classification accuracy in response to the above-mentioned technical problems.
[0007] A deep learning-based polarimetric SAR image classification method, the method comprising:
[0008] Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each of the fully polarimetric SAR image data as sample data;
[0009] After preprocessing the sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features.
[0010] The horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and circular polarization basis are calculated based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0011] Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different types of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data.
[0012] The slice samples and corresponding ground truth labels in the training dataset are input into the target classification model, and the model is trained to obtain the trained target classification model.
[0013] A polarimetric SAR image to be classified is acquired, and the polarimetric SAR image is input into the trained target classification model to classify the targets in the polarimetric SAR image.
[0014] In one embodiment, the preprocessing of each of the sample data includes:
[0015] After radiometric correction, the sample data is then processed by nonlocal uniform value filtering to obtain the preprocessed sample data.
[0016] In one embodiment, the polarization coherence matrix is the polarization coherence matrix corresponding to a three-channel polarization SAR radar system, and the equivalent polarization power characteristics of the three non-main diagonal elements in the polarization coherence matrix.
[0017] In one embodiment, the horizontal polarization corresponding to the linear polarization basis includes: the polarization coherence matrix T 12 T13 and T 23 The horizontal polarization power characteristics of the element;
[0018] The vertical polarization corresponding to the linear polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 Vertical polarization power characteristics corresponding to the element;
[0019] The following formulas were used to calculate the results:
[0020]
[0021]
[0022]
[0023]
[0024]
[0025]
[0026] In the above formula, T ij Re represents the coherence coefficient between the i-th and j-th channels in the polarization coherence matrix, Re() represents the real part, S represents the scattering characteristic, and the subscripts H and V represent the horizontal and vertical directions, respectively. ij ) represents T ij Horizontal polarization, V(T) ij ) represents T ij Vertical polarization.
[0027] In one embodiment, the left-hand circular polarization corresponding to the circular polarization basis includes: a polarization coherence matrix T 12 T 13 and T 23 The left-handed circular polarization power characteristics of the element;
[0028] The right-hand circular polarization corresponding to the circular polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 The right-hand circular polarization power characteristics of the element;
[0029] The following formulas were used to calculate the results:
[0030]
[0031]
[0032]
[0033]
[0034]
[0035]
[0036] In the above formula, T ij R(T) represents the coherence coefficient between the i-th and j-th channels in the polarization coherence matrix, Im() represents the imaginary part, S represents the scattering characteristic, and the subscripts RR and LL represent the right-handed and left-handed helical bodies, respectively. ij ) represents T ij Right-hand circular polarization, L(T) ij ) represents T ij Left-handed circular polarization.
[0037] In one embodiment, before obtaining multiple slice samples based on the polarization feature data corresponding to the sample data, the polarization feature data corresponding to each sample data is also normalized.
[0038] In one embodiment, the target classification model is a convolutional neural network, including: AlexNet neural network and VGG16 neural network.
[0039] In one embodiment, the target classification model is trained using a cross-entropy loss function until it converges.
[0040] This application also provides a deep learning-based polarimetric SAR image classification device, the device comprising:
[0041] The raw data acquisition module is used to acquire the raw fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and uses each of the fully polarimetric SAR image data as sample data;
[0042] The polarization coherence matrix acquisition module is used to preprocess each of the sample data, decompose the preprocessed sample data using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and represent the polarization coherence matrix based on the polarization scattering features.
[0043] The polarization feature data acquisition module is used to calculate the horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and the circular polarization basis, respectively, based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0044] The training dataset construction module is used to obtain slice samples of multiple different categories of targets in the fully polarimetric SAR image data based on the polarimetric feature data corresponding to each sample data and a window of a preset size, and to construct the training dataset based on the slice samples obtained from all sample data.
[0045] The target classification model training module is used to input the slice samples and corresponding ground truth labels from the training dataset into the target classification model, train it, and obtain the trained target classification model.
[0046] The target classification module is used to acquire the polarimetric SAR image to be classified, and input the polarimetric SAR image into the trained target classification model to classify the targets in the polarimetric SAR image.
[0047] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:
[0048] Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each of the fully polarimetric SAR image data as sample data;
[0049] After preprocessing the sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features.
[0050] The horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and circular polarization basis are calculated based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0051] Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different types of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data.
[0052] The slice samples and corresponding ground truth labels in the training dataset are input into the target classification model, and the model is trained to obtain the trained target classification model.
[0053] A polarimetric SAR image to be classified is acquired, and the polarimetric SAR image is input into the trained target classification model to classify the targets in the polarimetric SAR image.
[0054] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0055] Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each of the fully polarimetric SAR image data as sample data;
[0056] After preprocessing the sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features.
[0057] The horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and circular polarization basis are calculated based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0058] Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different types of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data.
[0059] The slice samples and corresponding ground truth labels in the training dataset are input into the target classification model, and the model is trained to obtain the trained target classification model.
[0060] A polarimetric SAR image to be classified is acquired, and the polarimetric SAR image is input into the trained target classification model to classify the targets in the polarimetric SAR image.
[0061] The aforementioned deep learning-based polarimetric SAR image classification method, apparatus, and device preprocess all sample data in the original fully polarimetric dataset, then decompose it using a reflection symmetric decomposition algorithm to obtain polarimetric scattering features. The polarimetric coherence matrix is represented based on these features, and the horizontal, vertical, left-handed, and right-handed circular polarizations corresponding to the linear and circular polarization bases are calculated from the elements in the matrix. The polarization, total polarization power, and elements located on the diagonal of the polarimetric coherence matrix are used as the polarimetric feature data for the corresponding sample data. A training dataset is generated based on this polarimetric feature data to train the target classification model. The trained target classification model is then used to classify targets in the complete polarimetric SAR image. This method can accurately classify targets in polarimetric SAR images. Attached Figure Description
[0062] Figure 1 This is a flowchart illustrating a deep learning-based polarimetric SAR image classification method in one embodiment.
[0063] Figure 2 This is a schematic diagram of the normalized distribution histogram of the 16 polarization characteristic parameters proposed in this method in one embodiment;
[0064] Figure 3 This is a flowchart illustrating the process of training a target classification model in one embodiment;
[0065] Figure 4 This is a schematic diagram of the classification results corresponding to the proposed data input scheme on the AlexNet network in one embodiment;
[0066] Figure 5 This is a schematic diagram of the classification results corresponding to the data input scheme proposed on the VGG16 network in one embodiment.
[0067] Figure 6 This is a structural block diagram of a deep learning-based polarimetric SAR image classification device in one embodiment;
[0068] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0070] In this application, as Figure 1 As shown, a deep learning-based polarimetric SAR image classification method is provided, including the following steps:
[0071] Step S100: Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each fully polarimetric SAR image data as sample data.
[0072] Step S110: After preprocessing each sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features.
[0073] Step S120: Calculate the horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear and circular polarization bases, respectively, based on the elements in the polarization coherence matrix. Use the horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and elements located on the diagonal of the polarization coherence matrix as the polarization feature data of the corresponding sample data.
[0074] Step S130: Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different categories of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data.
[0075] Step S140: Input the slice samples and corresponding ground truth labels from the training dataset into the target classification model and train it to obtain the trained target classification model.
[0076] Step S150: Obtain the polarimetric SAR image to be classified, and input the polarimetric SAR image into the trained target classification model to classify the targets in the polarimetric SAR image.
[0077] In this application, addressing the issue of limited training samples available for target classification in polarimetric SAR images using neural networks, a relatively complete data input scheme is explored, starting with the input data at the deep learning input end and based on polarization power information. Scattering features are extracted from preprocessed PolSAR images using reflection symmetry decomposition. Then, the polarization coherence matrix is represented based on these scattering polarization features, and multiple polarization features are extracted again. This allows the deep neural network to be trained using multiple polarization features from each original polarimetric image, enabling the deep neural network to improve target classification accuracy by learning the polarization features corresponding to different target categories.
[0078] In step S100, the original fully polarimetric dataset includes multiple original polarimetric SAR images. The dataset can be obtained through various means, including measured data and simulated data.
[0079] In this embodiment, the original fully polarimetric dataset is Gaofen-3 L1A level fully polarimetric data. "Gaofen-3" refers to a satellite, officially named High-Resolution Earth Observation Satellite 03 (GF-3). Therefore, all sample data in the original fully polarimetric dataset are synthetic aperture radar images with polarimetric information acquired by the Gaofen-3 satellite.
[0080] In step S110, the preprocessing of each sample data includes: after radiometric correction of the sample data, nonlocal uniform value filtering is used to remove the influence of inherent speckle noise in the data to obtain preprocessed sample data.
[0081] In this embodiment, when performing radiometric correction on sample data from the L1A class of Gaofen-3 fully polarized data, the relevant data description is referred to.
[0082] Furthermore, an affine symmetric decomposition algorithm capable of complete decomposition is applied to the preprocessed sample data to obtain the different scattering components corresponding to each sample data. A polarization correlation matrix is then constructed based on the obtained physical parameters. In step S120, the polarization characteristics corresponding to linear and circular polarization are obtained by calculating the elements in the polarization correlation matrix.
[0083] In this embodiment, the polarization coherence matrix is the polarization coherence matrix corresponding to the three-channel polarization SAR radar system, which is used to describe the equivalent polarization power characteristics of the three non-main diagonal elements in the polarization coherence matrix.
[0084] In this embodiment, the horizontal polarization corresponding to the linear polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 The horizontal polarization power characteristics corresponding to the element. The vertical polarization corresponding to the linear polarization basis includes: the polarization coherence matrix T. 12 T 13 and T 23 Vertical polarization power characteristics corresponding to the element.
[0085] Specifically, the following formulas are used for calculation:
[0086]
[0087]
[0088]
[0089]
[0090]
[0091]
[0092] In formulas (1) to (6), T ij Re represents the coherence coefficient between the i-th and j-th channels in the polarization coherence matrix, Re() represents the real part, S represents the scattering characteristic, and the subscripts H and V represent the horizontal and vertical directions, respectively. ij ) represents Tij Horizontal polarization, V(T) ij ) represents T ij Vertical polarization.
[0093] In this embodiment, the left-hand circular polarization corresponding to the circular polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 The left-hand circular polarization power characteristic corresponding to the element. The right-hand circular polarization corresponding to the circular polarization basis includes: the polarization coherence matrix T. 12 T 13 and T 23 The right-hand circular polarization power characteristics corresponding to the element.
[0094] Specifically, the following formulas are used for calculation:
[0095]
[0096]
[0097]
[0098]
[0099]
[0100]
[0101] In formulas (7) to (12), T ij R(T) represents the coherence coefficient between the i-th and j-th channels in the polarization coherence matrix, Im() represents the imaginary part, S represents the scattering characteristic, and the subscripts RR and LL represent the right-handed and left-handed helical bodies, respectively. ij ) represents T ij Right-hand circular polarization, L(T) ij ) represents T ij Left-handed circular polarization.
[0102] Next, the derivation process of the polarization expressions of the above formulas (1) to (12) will be explained.
[0103] There are two main types of polarization used in target radar antennas: linear polarization and circular polarization. Typical linear polarization methods include horizontal polarization (H) and vertical polarization (V), while circular polarization methods include left-hand circular polarization (L) and right-hand circular polarization (R).
[0104] When a polarimetric radar uses a linear polarization base, this method first transmits horizontally polarized electromagnetic waves, which are then received by horizontal and vertical antennas respectively. Next, it transmits vertically polarized electromagnetic waves, which are also received by horizontal and vertical antennas respectively. In monostatic radar systems, backscatter alignment convention (BSA) is typically used, and the transmitting and receiving antennas employ a single coordinate system. In this system, the Z-axis points from the antenna to the target, the X-axis is horizontal to the ground, and the Y-axis, along with the XZ-axis, forms a right-handed coordinate system pointing towards the sky. This coordinate system corresponds well to both horizontal (H) and vertical (V) polarization bases. The Sinclair scattering matrix in this case can be simplified as:
[0105]
[0106] Using Pauli base ψ P The Sinclair matrix can be vectorized to obtain:
[0107]
[0108] Sinclair matrix in ψ P The scattering vector below is:
[0109]
[0110] For a monostatic polarization radar, under the condition that the reciprocity theorem is satisfied, S HV =S VH Formula (15) can be written as:
[0111]
[0112] Therefore, in single-station, single-view polarimetric SAR data, the polarimetric scattering characteristics of the target under the Pauli basis represent the polarimetric coherence matrix, which is then expressed as:
[0113]
[0114] In formula (17), * represents conjugate, T represents transpose, and <·> represents set average, so we can get formula (1) and formula (2). Similarly, we can get formula (3) to formula (6).
[0115] Under circularly polarized bases, the scattering matrix under circularly polarized bases can be defined in the same way as follows:
[0116]
[0117] For a monostation radar, S is satisfied when the reciprocity theorem is true. LR =S RLElectromagnetic waves can be converted between linearly polarized and circularly polarized bases, which allows the scattering matrix to also be converted between the two bases. The conversion formula is expressed as:
[0118]
[0119] The corresponding transformation formulas between the circularly polarized basis and the Pauli vector are as follows:
[0120]
[0121] Then we have:
[0122]
[0123] Based on the above formula, we can obtain:
[0124]
[0125]
[0126] From formulas (22) and (23), it can be seen that the transformation from a horizontal to a vertical polarization basis to a circular polarization basis can also be considered as a process of performing a certain correlation decomposition on the scattering matrix, that is, decomposing the Sinclair matrix into a combination of a plane, a left-handed helix, and a right-handed helix. LR S RR S LL These correspond to the phase and power magnitude of each component, respectively. Therefore, formulas (7) to (12) can be obtained.
[0127] In this embodiment, based on the coherent system between channels in the polarization coherence matrix and formulas (1) to (12), H(T) is calculated. 12 ), H(T 13 ), H(T 23 ), L(T 12 ), L(T 13 ), L(T 23 ), V(T 12 ), V(T 13 ), V(T 23 ), R(T 12 ), R(T 13 ) and R(T 23 )12 polarization characteristics, and at the same time, extract the incoherence coefficients T of the three channels on the diagonal from the polarization coherence matrix. 11 T 22 T 33 And the total polarization power P0, for a total of 16 polarization characteristic data.
[0128] In this embodiment, before obtaining multiple slice samples based on the polarization feature data corresponding to the sample data, the polarization feature data corresponding to each sample data is also normalized. For example... Figure 2 As shown, this is a normalized histogram of the distribution of 16 polarization feature data, from which... Figure 2 (a) to Figure 2 (p) P0 and T are represented in alphabetical order. 11 T 22 T 33 H(T) 12 ), H(T 13 ), H(T 23 ), L(T 12 ), L(T 13 ), L(T 23 ), V(T 12 ), V(T 13 ), V(T 23 ), R(T 12 ), R(T 13 ), R(T 23 The polarization characteristic distribution histogram of ).
[0129] In this embodiment, through the above steps, 16 polarization features can be obtained from each sample data in the original fully polarimetric dataset. In fact, each polarization feature is two-dimensional, so polarization features can be understood as polarization feature data, which are 16 polarization feature images representing different polarization feature categories for each polarimetric SAR image.
[0130] Because polarimetric SAR data contains rich scattering features of ground objects, its spatial texture features, statistical characteristics, and polarimetric decomposition parameters can distinguish different ground object information to a certain extent. Therefore, in this embodiment, the traditional classification method with pixels as the smallest unit is abandoned. Instead, a window of a certain size is used as the basic unit. By extracting different numbers of polarimetric features, the rich polarimetric feature information of polarimetric SAR is fully utilized to improve feature utilization. When constructing the training sample set of the neural network, each sample is an image cube of size m×n×c, where m×n is the window size and c is the number of polarimetric features, i.e., 16. The minimum values of m and n are mainly to ensure that the image cube can still output features after passing through the convolution layers of the neural network model.
[0131] Furthermore, each polarimetric SAR image contains multiple target categories. To better train the target classification model, based on 16 polarimetric feature images representing different polarimetric feature categories, the polarimetric feature images are cropped according to the known target category location coordinates and a window of a preset size, resulting in slice samples corresponding to the region where each target category is located. Each slice sample can be understood as consisting of 16 two-dimensional data points, i.e., an image cube.
[0132] In step S140, multiple experiments were conducted to obtain the training parameters of the target classification model. The experiments showed that the batch size of the data used in the experiment was 64, the initial learning rate was 0.1, the learning rate decayed once every 10 epochs with a decay coefficient of 0.1, the weight coefficient was 0.9, the weight decay coefficient was 0.0005, and the weight initialization method used was the He Kaiming initialization method.
[0133] When training the target classification model, training samples are randomly input in batches.
[0134] In this embodiment, the target classification model is a convolutional neural network. The probability of the sample belonging to the candidate label category is calculated based on the network, compared with the real land cover label, and the prediction result is output. The convolutional neural network can be, but is not limited to, AlexNet neural network and VGG16 neural network.
[0135] In this embodiment, the cross-entropy loss function is used to train the target classification model until it converges.
[0136] Finally, in step S150, when using the trained target classification model for target classification, the complex polarimetric SAR image is directly input into the trained target classification model for classification tasks, eliminating the need for retraining from scratch and reducing unnecessary human and material consumption.
[0137] In this paper, the effectiveness of the proposed method is also demonstrated through experiments.
[0138] like Figure 3 The diagram shown is a flowchart of the method proposed in this application.
[0139] In the experiment, the designed input scheme for the training data, which proposed 16 polarization features as input parameters, is shown in Table 1, and the same model was trained accordingly.
[0140] Table 1. List of three polarization data input schemes
[0141]
[0142] like Figure 4As shown, the classification results are obtained after training the AlexNet network using the three different training data input schemes described above. Among them, Figure 4 (a) represents a pseudo-color composite image (converted to a black and white view). Figure 4 (b) represents the truth graph. Figure 4 (c) shows a schematic diagram of the classification results corresponding to the proposed data input scheme. For example... Figure 5 As shown, the classification results are obtained after training the VGG16 network using the three different training data input schemes described above. Among them, Figure 5 (a) represents a pseudo-color composite image (converted to a black and white view). Figure 5 (b) represents the truth graph. Figure 5 (c) shows a schematic diagram of the classification results corresponding to the proposed data input scheme 1.
[0143] Comparing the classification images of the proposed data input scheme, the proposed scheme is closer to the ground truth image, and the classification effect is significantly improved. In the aforementioned deep learning-based polarimetric SAR image classification method, a reflection-symmetric decomposition algorithm capable of complete decomposition is employed. Compared to classic polarimetric decomposition algorithms, this algorithm yields more polarimetric features. Furthermore, the reflection-symmetric decomposition algorithm is a model-based incoherent polarimetric decomposition method without polarimetric information loss, exhibiting excellent performance. All three decomposed polarimetric features satisfy the mirror symmetry assumption, and the original polarimetric coherence matrix can be completely reconstructed from the decomposition results, making it a complete decomposition algorithm. Due to the large number of decomposed polarimetric features, it offers more choices. In this method, through rigorous mathematical derivation, equivalent polarimetric power is obtained. Surface scattering, volume scattering, and secondary scattering polarimetric features, along with the total polarimetric power, are used as the basic data input scheme. Based on this, and according to existing research, the data input scheme is further designed using the polarimetric coherence matrix and the correlation coefficients between the off-diagonal elements of this matrix. Finally, based on rigorous mathematical derivation, using the horizontal, vertical, left-handed, and right-handed polarization characteristics under linear and circular polarization bases, as well as the three main diagonal elements of the non-polarization coherence matrix and the total polarization power, each polarization characteristic in this data input scheme has a clear physical meaning.
[0144] Finally, experiments demonstrate that the data input scheme proposed in this method increases the amount of input information to a certain extent compared to the other two comparison schemes. Under the same neural network conditions, the final PolSAR image classification accuracy is improved to a certain extent.
[0145] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0146] In one embodiment, such as Figure 6 As shown, a deep learning-based polarimetric SAR image classification device is provided, comprising: a raw data acquisition module 200, a polarimetric coherence matrix acquisition module 210, a polarimetric feature data acquisition module 220, a training dataset construction module 230, a target classification model training module 240, and a target classification module 250, wherein:
[0147] The raw data acquisition module 200 is used to acquire the raw fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and uses each of the fully polarimetric SAR image data as sample data;
[0148] The polarization coherence matrix acquisition module 210 is used to preprocess each of the sample data, decompose the preprocessed sample data using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and represent the polarization coherence matrix based on the polarization scattering features.
[0149] The polarization feature data acquisition module 220 is used to calculate the horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and the circular polarization basis, respectively, based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0150] The training dataset construction module 230 is used to obtain slice samples of multiple different types of targets in the fully polarimetric SAR image data based on the polarimetric feature data corresponding to each sample data and a window of a preset size, and to construct a training dataset based on the slice samples obtained from all sample data.
[0151] The target classification model training module 240 is used to input the slice samples and corresponding ground truth labels from the training dataset into the target classification model, train it, and obtain a trained target classification model.
[0152] The target classification module 250 is used to acquire the polarimetric SAR image to be classified, and input the polarimetric SAR image into the trained target classification model to classify the targets in the polarimetric SAR image.
[0153] Specific limitations regarding the deep learning-based polarimetric SAR image classification device can be found in the limitations of the deep learning-based polarimetric SAR image classification method described above, and will not be repeated here. Each module in the aforementioned deep learning-based polarimetric SAR image classification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0154] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a deep learning-based polarimetric SAR image classification method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0155] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0156] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:
[0157] Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each of the fully polarimetric SAR image data as sample data;
[0158] After preprocessing the sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features.
[0159] The horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and circular polarization basis are calculated based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0160] Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different types of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data.
[0161] The slice samples and corresponding ground truth labels in the training dataset are input into the target classification model, and the model is trained to obtain the trained target classification model.
[0162] A polarimetric SAR image to be classified is acquired, and the polarimetric SAR image is input into the trained target classification model to classify the targets in the polarimetric SAR image.
[0163] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0164] Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each of the fully polarimetric SAR image data as sample data;
[0165] After preprocessing the sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features.
[0166] The horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and circular polarization basis are calculated based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data.
[0167] Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different types of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data.
[0168] The slice samples and corresponding ground truth labels in the training dataset are input into the target classification model, and the model is trained to obtain the trained target classification model.
[0169] A polarimetric SAR image to be classified is acquired, and the polarimetric SAR image is input into the trained target classification model to classify the targets in the polarimetric SAR image.
[0170] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0171] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0172] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A deep learning-based polarimetric SAR image classification method, characterized in that, The method includes: Obtain the original fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and use each of the fully polarimetric SAR image data as sample data; After preprocessing the sample data, the preprocessed sample data is decomposed using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and the polarization coherence matrix is represented based on the polarization scattering features. The horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and circular polarization basis are calculated based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data. Based on the polarization feature data corresponding to each sample data and the window of a preset size, slice samples of multiple different types of targets in the fully polarimetric SAR image data are obtained, and a training dataset is constructed based on the slice samples obtained from all sample data. The slice samples and corresponding ground truth labels in the training dataset are input into the target classification model, and the model is trained to obtain the trained target classification model. A polarimetric SAR image to be classified is acquired, and the polarimetric SAR image is input into the trained target classification model to classify the targets in the polarimetric SAR image.
2. The polarimetric SAR image classification method according to claim 1, characterized in that, The preprocessing of the sample data includes: After radiometric correction, the sample data is then processed by nonlocal uniform value filtering to obtain the preprocessed sample data.
3. The polarimetric SAR image classification method according to claim 2, characterized in that, The polarization coherence matrix is the polarization coherence matrix corresponding to the three-channel polarization SAR radar system, used to describe the equivalent polarization power characteristics of the three non-main diagonal elements in the polarization coherence matrix.
4. The polarimetric SAR image classification method according to claim 3, characterized in that, The horizontal polarization corresponding to the linear polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 The horizontal polarization power characteristics of the element; The vertical polarization corresponding to the linear polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 Vertical polarization power characteristics corresponding to the element; The following formulas were used to calculate the results: In the above formula, T ij Re represents the coherence coefficient between the i-th and j-th channels in the polarization coherence matrix, Re() represents the real part, S represents the scattering characteristic, and the subscripts H and V represent the horizontal and vertical directions, respectively. ij ) represents T ij Horizontal polarization, V(T) ij ) represents T ij Vertical polarization.
5. The polarimetric SAR image classification method according to claim 4, characterized in that, The left-hand circular polarization corresponding to the circular polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 The left-handed circular polarization power characteristics of the element; The right-hand circular polarization corresponding to the circular polarization basis includes: the polarization coherence matrix T 12 T 13 and T 23 The right-hand circular polarization power characteristics of the element; The following formulas were used to calculate the results: In the above formula, T ij Represents the (i, j) element in the polarization coherence matrix, Im() represents the imaginary part, S represents the scattering characteristic, and subscripts RR and LL represent right-handed and left-handed helical bodies, respectively. ij ) represents T ij Right-hand circular polarization, L(T) ij ) represents T ij Left-handed circular polarization.
6. The polarimetric SAR image classification method according to claim 5, characterized in that, Before obtaining multiple slice samples based on the polarization feature data corresponding to the sample data, the polarization feature data corresponding to each sample data is also normalized.
7. The polarimetric SAR image classification method according to claim 6, characterized in that, The target classification model is a convolutional neural network, including: AlexNet neural network and VGG16 neural network.
8. The polarimetric SAR image classification method according to claim 7, characterized in that, When training the target classification model, the cross-entropy loss function is used to train it until it converges.
9. A deep learning-based polarimetric SAR image classification device, characterized in that, The device includes: The raw data acquisition module is used to acquire the raw fully polarimetric dataset, which includes multiple fully polarimetric SAR image data, and uses each of the fully polarimetric SAR image data as sample data; The polarization coherence matrix acquisition module is used to preprocess each of the sample data, decompose the preprocessed sample data using a reflection symmetric decomposition algorithm to obtain polarization scattering features, and represent the polarization coherence matrix based on the polarization scattering features. The polarization feature data acquisition module is used to calculate the horizontal polarization, vertical polarization, left-hand circular polarization, and right-hand circular polarization corresponding to the linear polarization basis and the circular polarization basis, respectively, based on the elements in the polarization coherence matrix. The horizontal polarization, vertical polarization, left-hand circular polarization, right-hand circular polarization, total polarization power, and the elements located on the diagonal of the polarization coherence matrix are used as the polarization feature data corresponding to the sample data. The training dataset construction module is used to obtain slice samples of multiple different categories of targets in the fully polarimetric SAR image data based on the polarimetric feature data corresponding to each sample data and a window of a preset size, and to construct the training dataset based on the slice samples obtained from all sample data. The target classification model training module is used to input the slice samples and corresponding ground truth labels from the training dataset into the target classification model, train it, and obtain the trained target classification model. The target classification module is used to acquire the polarimetric SAR image to be classified, and input the polarimetric SAR image into the trained target classification model to classify the targets in the polarimetric SAR image.
10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.