Polarimetric SAR Classification Method, Device and Equipment Based on Input of Complete Polarimetric Information

Abstract

The present application relates to a polarization SAR classification method, device and equipment based on the input of complete polarization information. After performing non-local mean filtering on the full-polarization SAR data to be classified, a reflection symmetry decomposition method is used for polarization grading to obtain multiple polarization features characterizing the backscattering of ground objects. Then, a non-linear normalization method is used to perform normalization processing on different types of polarization features to obtain complete polarization information. The trained convolutional neural network is used to perform accurate ground object classification based on the complete polarization information. This method improves the classification accuracy by using the complete polarization information for ground object classification and combining a non-linear normalization method for the polarization features of different ground objects.

Description

Technical Field

[0001] This application relates to the technical field of polarimetric SAR image classification, and particularly to a polarimetric SAR classification method, device, and equipment based on the input of complete polarimetric information. Background Art

[0002] Polarimetric synthetic aperture radar (PolSAR) can obtain the complete polarimetric scattering characteristics of ground objects all day and all weather, and has been applied to various remote sensing scenarios in recent years. PolSAR actively acquires the polarimetric information of surface scattering, and contains more parameters describing the electromagnetic scattering characteristics of ground objects compared with traditional single-polarization SAR. How to obtain the complete polarimetric scattering parameters that can reflect the backscattering of ground objects from PolSAR images, and how to fully and reasonably explore these polarimetric characteristics and apply them to the currently widely used deep learning algorithms is crucial for the task of classifying based on polarimetric SAR data.

[0003] Currently, PolSAR classification methods are mainly divided into three categories. 1. Based on the features decomposed from polarization, the PolSAR image is decomposed by polarization to extract the scattering characteristics of the target ground object and directly classify. Common target polarization decomposition methods include Freeman decomposition, Cloude-Potier decomposition, Huynen decomposition, etc. 2. Classify according to the statistical distribution characteristics of PolSAR data, and common algorithms include Wishart classification and other algorithms. 3. Methods using deep learning for classification. With the rapid development of deep learning methods, relevant scholars have introduced various deep learning methods into PolSAR image classification. Since the deep learning model contains multiple convolutional layers and can well extract the high-order features of ground objects in the image, better classification results have been obtained. Although a large number of scholars have used deep learning methods to classify PolSAR images and achieved good results, there are still defects. For example, some algorithms simply stack and combine the features decomposed from polarization without discrimination and input them into the network, but ignore the defects of the polarization decomposition method itself. When normalizing polarimetric features, some methods only simply normalize without considering the data distribution characteristics, and also use linear normalization for non-linear PolSAR data. When using CNN to process PolSAR images, some methods do not consider all the scattering information and various polarimetric scattering characteristics of the image, and only use a single form of incomplete polarimetric data input scheme as the network input. These problems will all lead to inaccurate classification. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a polarimetric SAR classification method, device, and equipment based on the input of complete polarimetric information that can accurately classify ground objects using complete polarimetric information.

[0005] A polarization SAR classification method based on the input of complete polarization information, the method comprising:

[0006] Obtain full-polarization SAR data to be classified, the full-polarization SAR data being data related to ground objects;

[0007] Preprocess the full-polarization SAR data to obtain complete polarization information, wherein the preprocessing includes performing non-local mean filtering on the full-polarization SAR data, and then using a reflection symmetry decomposition method for polarization grading to obtain multiple polarization features characterizing the backscattering of ground objects, and using a non-linear normalization method to perform normalization processing on different types of polarization features to obtain the complete polarization information;

[0008] Input the complete polarization information into a trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data.

[0009] In one embodiment, the full-polarization SAR data is Gaofen-3 L1A data.

[0010] In one embodiment, the multiple polarization features characterizing the backscattering of ground objects include 15 real polarization features and 6 complex polarization features.

[0011] In one embodiment, the polarization total power Span , the three main diagonal elements of the polarization coherence matrix, the volume scattering component power value , the surface scattering power value , the second scattering power value , the total power value of the second component of the reflection symmetry decomposition , the total power value of the third component of the reflection symmetry decomposition , twice the orientation angle , twice the helix angle , the power ratio of spherical scattering in the second component of the reflection symmetry decomposition x , the power ratio of spherical scattering in the third component of the reflection symmetry decomposition y , the second component of the reflection symmetry decomposition The phase of the element a , the third component of the reflection symmetry decomposition The phase of the element b ;

[0012] The 6 complex polarization features include the real part and the imaginary part of the non-main diagonal elements of the polarization coherence matrix.

[0013] In one embodiment, the using a non-linear normalization method to perform normalization processing on different types of polarization features includes:

[0014] For the total polarization power, after converting the total polarization power into a value in dB units, a linear normalization method is used for normalization processing;

[0015] Using the total polarization power, non-linear normalization processing is performed on the polarization characteristics related to the spherical scattering power value, the dihedral angle scattering power value, the 45° dihedral angle scattering power, as well as the volume scattering power value, the surface scattering power value, the secondary scattering power value, the second component total power value, and the third component total power value;

[0016] Using the linear normalization method, perform normalization processing on the twice the orientation angle, twice the helix angle, the phase of the second component element and the phase of the third component element.

[0017] In one embodiment, during the training process of the convolutional neural network, when constructing the training set:

[0018] On the full-polarization SAR sample data, randomly select multiple pixel points;

[0019] Taking the selected pixel points as the center according to a preset size, generate corresponding slice samples;

[0020] After preprocessing each of the slice samples to obtain complete polarization information samples, construct the training set according to the complete polarization information samples.

[0021] This application also provides a polarization SAR classification device based on the input of complete polarization information. The device includes:

[0022] A full-polarization SAR data acquisition module, which acquires the full-polarization SAR data to be classified. The full-polarization SAR data is data related to ground objects;

[0023] A preprocessing module, which is used to preprocess the full-polarization SAR data to obtain complete polarization information. Among them, the preprocessing includes performing non-local mean filtering on the full-polarization SAR data, and then using the reflection symmetry decomposition method for polarization grading to obtain multiple polarization characteristics representing the backscattering of ground objects, and using the non-linear normalization method to perform normalization processing on different types of polarization characteristics to obtain the complete polarization information;

[0024] A polarization SAR classification module, which is used to input the complete polarization information into the trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data.

[0025] A computer device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the following steps are implemented:

[0026] Obtain the full-polarization SAR data to be classified, where the full-polarization SAR data is data related to ground objects;

[0027] Preprocess the full-polarization SAR data to obtain complete polarization information. Among them, the preprocessing includes performing non-local mean filtering on the full-polarization SAR data, then using the reflection symmetry decomposition method for polarization classification to obtain multiple polarization features characterizing the backscattering of ground objects, and using the non-linear normalization method to perform normalization processing on different types of polarization features to obtain the complete polarization information;

[0028] Input the complete polarization information into the trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data.

[0029] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, the following steps are implemented:

[0030] Obtain the full-polarization SAR data to be classified, where the full-polarization SAR data is data related to ground objects;

[0031] Preprocess the full-polarization SAR data to obtain complete polarization information. Among them, the preprocessing includes performing non-local mean filtering on the full-polarization SAR data, then using the reflection symmetry decomposition method for polarization classification to obtain multiple polarization features characterizing the backscattering of ground objects, and using the non-linear normalization method to perform normalization processing on different types of polarization features to obtain the complete polarization information;

[0032] Input the complete polarization information into the trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data.

[0033] For the above polarization SAR classification method, device and equipment based on the input of complete polarization information, after performing non-local mean filtering on the full-polarization SAR data to be classified, use the reflection symmetry decomposition method for polarization classification to obtain multiple polarization features characterizing the backscattering of ground objects, and then use the non-linear normalization method to perform normalization processing on different types of polarization features to obtain complete polarization information, and use the trained convolutional neural network to perform accurate ground object classification according to the complete polarization information. This method improves the classification accuracy by using the complete polarization information for ground object classification and combining the non-linear normalization method for the polarization features of different ground objects. Description of the Drawings

[0034] Figure 1 It is a schematic flowchart of the polarization SAR classification method based on the input of complete polarization information in an embodiment;

[0035] Figure 2 Schematic block diagram of the process of using polarimetric SAR classification based on complete polarimetric information input for ground object classification in an embodiment;

[0036] Figure 3 Schematic structural diagram of a dual-branch network based on complex numbers and real numbers in an embodiment;

[0037] Figure 4 Complete fully polarimetric SAR image and corresponding ground truth image used in an experiment;

[0038] Figure 5 Schematic diagram of the distribution of training, validation, and test samples in an experiment, Figure 5 (a), Figure 5 (b), Figure 5 (c) and Figure 5 (d) are fully polarimetric SAR images obtained at different times;

[0039] Figure 6 Classification result graph of using the proposed method on AlexNet in an experiment. Among them, Figure 6 (a) represents T Schematic diagram of the classification results of the main diagonal elements and the correlation coefficients of non-main diagonal elements of the matrix, Figure 6 (b) represents T Schematic diagram of the classification results of all elements of the matrix and the total polarization power, Figure 6 (c) represents the schematic diagram of the classification results using the method proposed in this paper, Figure 6 (d) represents the ground truth map.

[0040] Figure 7 Classification result graph of using the proposed method on VGG16 in an experiment. Among them, Figure 7 (a) represents T Schematic diagram of the classification results of the main diagonal elements and the correlation coefficients of non-main diagonal elements of the matrix, Figure 7 (b) represents T Schematic diagram of the classification results of all elements of the matrix and the total polarization power, Figure 7 (c) represents the schematic diagram of the classification results using the method proposed in this paper, Figure 7 (d) represents the ground truth map.

[0041] Figure 8 Schematic block diagram of the structure of a polarimetric SAR classification device based on complete polarimetric information input in an embodiment;

[0042] Figure 9 Internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0043] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0044] In the prior art, when using a neural network for the classification of polarimetric SAR images, some algorithms simply stack and combine the features decomposed from various polarizations without discrimination and then input them into the network, without considering the defects existing in the polarimetric decomposition method itself. And there are also some methods that simply normalize the polarimetric features without considering the characteristics of the data distribution itself. Even some linear normalization methods are applied to non-linear PolSAR data. These preprocessing processes before inputting the data into the neural network will all lead to difficulties in convergence during the subsequent neural network training and inaccurate classification of the trained neural network. In the present application, as Figure 1 shown, a polarimetric SAR classification method based on the input of complete polarimetric information is provided, which specifically includes the following steps:

[0045] Step S100, obtain the full-polarimetric SAR data to be classified, and the full-polarimetric SAR data is data related to ground objects.

[0046] Step S110, preprocess the full-polarimetric SAR data to obtain complete polarimetric information. Among them, the preprocessing includes performing non-local mean filtering on the full-polarimetric SAR data, and then using the reflection symmetry decomposition method for polarimetric classification to obtain multiple polarimetric features characterizing the backscattering of ground objects, and using a non-linear normalization method to perform normalization processing on different types of polarimetric features to obtain complete polarimetric information.

[0047] Step S120, input the complete polarimetric information into the trained convolutional neural network to obtain the ground object classification result in the full-polarimetric SAR data.

[0048] In this method, based on the complete polarimetric decomposition algorithm, by extracting features at the input end and using the non-linear normalization algorithm, the classification of PolSAR images is performed through a classical convolutional neural network, which further improves the classification accuracy and obtains an input scheme for the deep learning method of complete polarimetric information.

[0049] In step S100, the method in the present application classifies according to the types of ground objects, that is, the full-polarimetric SAR data is the entire image detected from a high place with the ground as the target, and its purpose is to classify different ground objects on the ground.

[0050] Furthermore, the full-polarization SAR data is the GF-3 L1A level data, which is the full-polarization raw data. In remote sensing data processing, L1A level data generally undergoes preliminary radiometric correction and other processing, but there may still be problems such as noise, and further filtering and other processing are required to improve data quality. Therefore, in the subsequent preprocessing process, non-local means filtering is performed after obtaining the L1A level data, which targets the L1A level data itself for subsequent processing.

[0051] In step S110, first, non-local means filtering is performed on the full-polarization SAR data to eliminate possible noise problems, and then a polarization decomposition algorithm that can completely extract the polarization information of the ground objects, that is, the reflection symmetry decomposition method, is used to perform polarization feature processing on the processed data to obtain various polarization features that can completely represent the scattering information of different ground objects in the full-polarization SAR data.

[0052] In this embodiment, the multiple polarization features representing the backscattering of the ground objects include 15 real polarization features and 6 complex polarization features. Among them, there are power value type polarization features, angle type polarization features, phase type polarization features, power ratio type polarization features, and complex polarization feature types.

[0053] Specifically, the 15 real polarization scattering features include: total polarization power Span , the three main diagonal elements of the polarization coherence matrix, the power value of the volume scattering component , the power value of the surface scattering , the power value of the double scattering , the total power value of the second component of the reflection symmetry decomposition , the total power value of the third component of the reflection symmetry decomposition , twice the orientation angle , twice the helix angle , the power ratio of the spherical scattering in the second component of the reflection symmetry decomposition x , the power ratio of the spherical scattering in the third component of the reflection symmetry decomposition y , the second component of the reflection symmetry decomposition the phase of the element a , the third component of the reflection symmetry decomposition the phase of the element b .

[0054] Specifically, the 6 complex polarization features include: the real part and the imaginary part of the non-main diagonal elements of the polarization coherence matrix.

[0055] In this embodiment, in order to meet the requirements of the neural network input end, it is necessary to normalize the polarization feature parameters. Considering that the polarization SAR data is essentially non-linear in nature, a non-linear normalization method needs to be adopted to increase the distinguishability of the polarization feature parameters of different ground objects and normalize the distribution range of the polarization features to 0-1.

[0056] Furthermore, the non-linear normalization method for normalizing different types of polarization features includes: for the total polarization power, after converting the total polarization power into a value in dB units, a linear normalization method is used for normalization. Using the total polarization power, the polarization features related to the spherical scattering power value, the dihedral scattering power value, the 45° dihedral scattering power, as well as the volume scattering power value, the surface scattering power value, the secondary scattering power value, the total power value of the second component, and the total power value of the third component, are normalized. Using the linear normalization method, the twice of the orientation angle, the twice of the helix angle, the phase of the second component element and the phase of the third component element are normalized.

[0057] Specifically, before normalizing each polarization feature, it is necessary to process the total polarization power generated by polarization decomposition. Span In matrix, the total polarization power Span value is the sum of the diagonal values of the matrix, that is:

[0058] .

[0059] Next, in order to better represent Span , Span is converted into a quantity in dB units, using the following formula:

[0060] .

[0061] At this time, is a linear variable, and a linear normalization method is used for this variable. In one embodiment, the maximum-minimum normalization method can be adopted, expressed as:

[0062] .

[0063] In the above formula, represents the normalized physical quantity, represents the physical quantity to be processed, , are respectively the maximum and minimum values of the value range of the physical quantity to be processed.

[0064] Furthermore, since , therefore, for processing the main diagonal elements in the matrix, use (where i, j represents the row and column numbers of the matrix) to achieve the normalization of the values of each element in the matrix.

[0065] In this embodiment, for the volume scattering power value , the surface scattering power value , the secondary scattering power value , the total power value of the second component , the total power value of the third component , these physical quantities are all smaller than the total polarization power Span value. Therefore, for the processing of these physical quantities, divide by Span value to achieve normalization.

[0066] Furthermore, the value range of twice the orientation angle is (-π / 2, π / 2]. Twice the helix angle φ , the value range is [-π / 4, π / 4], the phase of the second component T 12 element a , the third component T 12 element b phase

[0067] Furthermore, since the power ratio of spherical scattering in the second component x , the power ratio of spherical scattering in the third component y， these two physical quantities have a value range of [0,1], so there is no need to normalize these two physical quantities anymore.

[0068] For the 6 complex polarization characteristics, including the real and imaginary parts of the non-main diagonal elements of the polarization coherence matrix, use and method to achieve normalization, where Re() represents the real part of the complex polarization characteristic, Im () represents the imaginary part of the complex polarization characteristic.

[0069] As Figure 2 shown, it is a schematic block diagram of the process of using polarization SAR classification based on complete polarization information input for ground object classification.

[0070] In this embodiment, during the training of the convolutional neural network, when constructing the training data: on the full-polarization SAR sample data, multiple pixel points are randomly selected, and the selected pixel points are used as the center according to a preset size to generate corresponding slice samples. After preprocessing each slice sample, a complete polarization information sample is obtained, and a training set is constructed based on the complete polarization information sample. Among them, in the training set, the ground object category of each slice sample is also included as the true label.

[0071] Specifically, when training the convolutional neural network, the samples selected from the training set are randomly input in batches. The convolutional neural network is used to calculate the probability that the sample belongs to all candidate label categories, compare with the true ground object label, and output the prediction result. According to the number of true labels predicted by the model and the number of misclassified ones, the accuracy of the training samples and the validation samples is calculated, and whether the training model is moderately fitted is determined through the accuracies of the two. Whenever the accuracy on the validation set improves, the weights at this time are retained, and the weights that make the validation set accuracy the highest are stored when the training ends. The model with the highest validation set accuracy is saved, and finally the data with the highest accuracy is obtained, and this model is used for the test of subsequent data.

[0072] In one of the embodiments, the batch data size is 64, the initial learning rate is 0.1, it decays once every 10 epochs, the decay coefficient is 0.1, the weight coefficient is 0.9, and the weight coefficient is 0.0005.

[0073] In this embodiment, in the convolutional neural network, the eigenvalue size can be obtained through convolution according to the complete polarization information. The fully connected layer and the Softmax function are used to judge the category to which the ground object belongs. Then the judged result is filled into an empty matrix with the same size as the predicted image to obtain the classification result of the entire image.

[0074] In this embodiment, the convolutional neural network can adopt the AlexNet convolutional neural network, the VGG16 convolutional neural network, and the dual-branch network to process real and complex polarization features respectively.

[0075] As Figure 3 shown, in this method, a structure of a dual-branch network based on complex numbers and real numbers is also proposed to process real and complex polarization features respectively.

[0076] Specifically, the real-number branch network is built based on the ResNet50 residual structure and the CBAM attention mechanism. This branch network consists of a convolutional module (ConvBlock) and three residual modules (RCBlock) that integrate the CBAM attention mechanism. First, ConvBlock preliminarily extracts the features of the input real-number parameter features, and the features extracted by the ConvBlock convolutional block are denoted as , , and then after passing through the first RCBlock convolutional block, the feature , . Similarly, the feature passing through the second RCBlock convolutional block is , , and finally the feature extracted by the real number branch network (the feature extracted by the 4th RCBlock). Among them, represents element-wise multiplication. During the multiplication process, the channel attention mechanism broadcasts along the spatial dimension, is the feature after passing through the CBAM attention mechanism.

[0077] Specifically, the complex number branch network includes three complex number blocks with different convolutional kernel sizes in parallel. Each complex number block includes two complex number convolutional layers (CCD) with convolutional kernel sizes of 3, 7, and 11 respectively, a complex number batch normalization layer (CBN), a complex number activation function layer (CRL), and a complex number pooling layer (CMP). Specifically, the extracted feature map is input into the complex number layers with three different sizes of convolutional kernels, and then the extracted features are summed additively. As the input feature of the next complex number block.

[0078] Finally, a fusion classification network is used to fuse the features extracted by the complex number branch network and the features extracted by the real number branch network, and then a fully connected layer is used to perform ground object classification based on the fused features. Among them, the size adjustment is achieved by using different sizes of convolutional kernels in the pooling layer.

[0079] In this paper, experiments are also carried out on the AlexNet convolutional neural network and the VGG16 convolutional neural network respectively to prove the effectiveness of this method.

[0080] As Figure 4 shown, it is the complete fully polarized SAR image and the corresponding ground truth image used in the experiment.

[0081] As Figure 5 shown, it is the schematic diagram of the training, validation and test sample distributions. Figures (a), (b), (c) and (d) are fully polarized SAR images obtained at different times. In the experiment, polarized images at different times are used to highlight the universality of the proposed method.

[0082] As Figure 6 shown, it is the result map of classification using this method on the AlexNet convolutional neural network. It can be seen from this figure that the method proposed in this paper achieves the best overall classification effect and the classification effect between ground objects.

[0083] As Figure 7As shown in the figure, it is the result graph after classification using the method proposed in this paper on the VGG16 convolutional neural network. It can be seen from this figure that the method proposed in this paper achieves the best overall classification effect and the classification effect between different ground objects.

[0084] In the above polarization SAR classification method based on the input of complete polarization information, for the polarization SAR classification method, device and equipment based on the input of complete polarization information, after performing non-local mean filtering on the full polarization SAR data to be classified, the polarization decomposition method is used for polarization classification to obtain multiple polarization features characterizing the backscattering of ground objects, and then the non-linear normalization method is used to perform normalization processing on different types of polarization features to obtain complete polarization information. The trained convolutional neural network is used to perform accurate ground object classification based on the complete polarization information. By using the complete polarization information for ground object classification and combining the non-linear normalization method for the polarization features of different ground objects, the accuracy of classification is improved. This method is based on the reflection symmetry decomposition algorithm that can fully extract polarization features, extracts the complete backscattering information of ground objects, and solves the problem of incomplete information input in the deep learning network.

[0085] It should be understood that although Figure 1 the steps in the flowchart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps in

[0086] In one embodiment, as Figure 8 shown, a polarization SAR classification device based on the input of complete polarization information is provided, including: a full polarization SAR data acquisition module 200, a preprocessing module 210, and a polarization SAR classification module 220, where:

[0087] The full polarization SAR data acquisition module 200 acquires the full polarization SAR data to be classified, and the full polarization SAR data is data related to ground objects;

[0088] A preprocessing module 210 is configured to preprocess the full-polarization SAR data to obtain complete polarization information. The preprocessing includes performing non-local mean filtering on the full-polarization SAR data, and then using a reflection symmetry decomposition method to perform polarization classification to obtain multiple polarization features characterizing the backscattering of the ground object. A non-linear normalization method is used to perform normalization processing on different types of polarization features to obtain the complete polarization information.

[0089] A polarization SAR classification module 220 is configured to input the complete polarization information into a trained convolutional neural network to obtain a ground object classification result in the full-polarization SAR data.

[0090] For the specific limitations of the polarization SAR classification device based on the input of complete polarization information, reference can be made to the limitations of the polarization SAR classification method based on the input of complete polarization information in the foregoing text, which will not be elaborated herein. Each module in the above-mentioned polarization SAR classification device based on the input of complete polarization information can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor of the computer device in the form of hardware or be independent of it, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0091] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 9 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program stored in the non-volatile storage medium to run. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a polarization SAR classification method based on the input of complete polarization information. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0092] Those skilled in the art can understand that Figure 9 the structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0093] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0094] Obtain full-polarization SAR data to be classified, where the full-polarization SAR data is data related to ground objects;

[0095] Preprocess the full-polarization SAR data to obtain complete polarization information. Wherein, the preprocessing includes performing non-local mean filtering on the full-polarization SAR data, then using a reflection symmetry decomposition method for polarization classification to obtain multiple polarization features characterizing the backscattering of ground objects, and using a non-linear normalization method to perform normalization processing on different types of polarization features to obtain the complete polarization information;

[0096] Input the complete polarization information into a trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data.

[0097] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0098] Obtain full-polarization SAR data to be classified, where the full-polarization SAR data is data related to ground objects;

[0099] Preprocess the full-polarization SAR data to obtain complete polarization information. Wherein, the preprocessing includes performing non-local mean filtering on the full-polarization SAR data, then using a reflection symmetry decomposition method for polarization classification to obtain multiple polarization features characterizing the backscattering of ground objects, and using a non-linear normalization method to perform normalization processing on different types of polarization features to obtain the complete polarization information;

[0100] Input the complete polarization information into a trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data.

[0101] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. 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 many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0102] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered as the scope described in this specification.

[0103] The above-described embodiments merely represent several implementation manners of this application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of the method of this application should be subject to the appended claims.

Claims

1. A polarization SAR classification method based on the input of complete polarization information, characterized in that The method includes: Obtaining fully polarimetric SAR data to be classified, where the fully polarimetric SAR data is data related to ground objects; Preprocessing the fully polarimetric SAR data to obtain complete polarization information. Among them, the preprocessing includes performing non-local mean filtering on the fully polarimetric SAR data and then using the reflection symmetry decomposition method for polarization decomposition to extract multiple polarization features characterizing the backscattering of ground objects. The multiple polarization features include 15 real polarization features: total polarization power, three main diagonal elements of the polarization coherence matrix, volume scattering component power value, surface scattering power value, second scattering power value, total power value of the second component of reflection symmetry decomposition, total power value of the third component of reflection symmetry decomposition, twice the orientation angle, twice the helix angle, power ratio of spherical scattering in the second component of reflection symmetry decomposition, power ratio of spherical scattering in the third component of reflection symmetry decomposition, phase of the elements in the second component of reflection symmetry decomposition, phase of the elements in the third component of reflection symmetry decomposition, and 6 complex polarization features: real and imaginary parts of the non-main diagonal elements of the polarization coherence matrix; The non - linear normalization method is adopted to normalize different types of polarization features to obtain the complete polarization information. Among them, for the total polarization power, after converting the total polarization power into a value in dB units, a linear normalization method is used for normalization. Using the total polarization power, non - linear normalization is performed on the polarization features related to the spherical scattering power value, the dihedral angle scattering power value, the 45° dihedral angle scattering power, as well as the volume scattering power value, the surface scattering power value, the secondary scattering power value, the second component total power value, and the third component total power value. Using the linear normalization method, normalization is performed on the twice - oriented angle, the twice - helical angle, the phase of the second - component element, and the phase of the third - component element. For the 6 complex polarization features, including the real and imaginary parts of the non - diagonal elements of the polarization coherence matrix, and ways are used to achieve normalization, where Re() represents the real part of the complex polarization feature, and Im() represents the imaginary part of the complex polarization feature; Inputting the complete polarization information into a trained convolutional neural network to obtain the ground object classification result in the fully polarimetric SAR data. The trained convolutional neural network includes a complex and real dual-branch network that processes real and complex polarization features respectively. Among them, the real branch network is built based on the ResNet50 residual structure and the CBAM attention mechanism, and the complex branch network includes three complex blocks with different convolutional kernel sizes in parallel.

2. The polarization SAR classification method based on the input of complete polarization information according to claim 1, wherein The fully polarimetric SAR data is GF-3 L1A data.

3. The polarization SAR classification method based on the input of complete polarization information according to claim 2, wherein During the training process of the convolutional neural network, when constructing the training set: Randomly select multiple pixel points from the fully polarimetric SAR sample data; Generate corresponding slice samples with the selected pixel points as the center according to a preset size; Preprocess each of the slice samples to obtain complete polarization information samples, and construct the training set according to the complete polarization information samples.

4. A polarimetric SAR classification device based on the input of complete polarization information, characterized in that, The device includes: A fully polarimetric SAR data acquisition module that obtains fully polarimetric SAR data to be classified, where the fully polarimetric SAR data is data related to ground objects; A preprocessing module for preprocessing the full-polarization SAR data to obtain complete polarization information. The preprocessing includes performing non-local mean filtering on the full-polarization SAR data and then using a reflection symmetry decomposition method for polarization decomposition to obtain multiple polarization features characterizing the backscattering of the ground objects. The multiple polarization features include 15 real polarization features: total polarization power, three main diagonal elements of the polarization coherence matrix, volume scattering component power value, surface scattering power value, second scattering power value, total power value of the second component of the reflection symmetry decomposition, total power value of the third component of the reflection symmetry decomposition, twice the orientation angle, twice the helix angle, power ratio of spherical scattering in the second component of the reflection symmetry decomposition, power ratio of spherical scattering in the third component of the reflection symmetry decomposition, phase of the elements of the second component of the reflection symmetry decomposition, phase of the elements of the third component of the reflection symmetry decomposition, and 6 complex polarization features: real and imaginary parts of the non-main diagonal elements of the polarization coherence matrix; The polarization feature processing module is used to perform normalization processing on different types of polarization features by using a non-linear normalization method to obtain the complete polarization information. Among them, for the total polarization power, after converting the total polarization power into a value in dB units, a linear normalization method is used for normalization processing. By using the total polarization power, non-linear normalization processing is performed on the polarization features related to the spherical scattering power value, the dihedral angle scattering power value, the 45° dihedral angle scattering power, as well as the volume scattering power value, the surface scattering power value, the secondary scattering power value, the second component total power value, and the third component total power value. By using the linear normalization method, normalization processing is performed on the twice the orientation angle, twice the helix angle, the phase of the second component element, and the phase of the third component element. For the 6 complex polarization features, including the real and imaginary parts of the non-principal diagonal elements of the polarization coherence matrix, and are used to achieve normalization, where Re() represents the real part of the complex polarization feature, and Im() represents the imaginary part of the complex polarization feature; A polarization SAR classification module for inputting the complete polarization information into a trained convolutional neural network to obtain the ground object classification result in the full-polarization SAR data. The trained convolutional neural network includes a complex and real dual-branch network that separately processes real and complex polarization features. The real-branch network is built based on the ResNet50 residual structure and the CBAM attention mechanism, and the complex-branch network includes three complex blocks with different convolutional kernel sizes in parallel.

5. A computer device, comprising a memory and a processor, the memory storing 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 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 3.

Images