CN114004998A

CN114004998A - Unsupervised polarization SAR image terrain classification method based on multi-view tensor product diffusion

Info

Publication number: CN114004998A
Application number: CN202111293497.7A
Authority: CN
Inventors: 邹焕新; 李美霖; 曹旭; 马倩; 李润林; 成飞; 贺诗甜; 魏娟; 孙丽
Original assignee: National University of Defense Technology
Current assignee: National University of Defense Technology
Priority date: 2021-11-03
Filing date: 2021-11-03
Publication date: 2022-02-01
Anticipated expiration: 2041-11-03
Also published as: CN114004998B

Abstract

The application relates to an unsupervised polarimetric SAR image terrain classification method based on multi-view tensor product diffusion. The method comprises the following steps: the method comprises the steps of segmenting a polarized SAR image to be classified by adopting a rapid superpixel segmentation method to obtain a plurality of superpixels, constructing a plurality of graph models based on the segmented plurality of superpixels by utilizing the 3 high-dimensional eigenvectors, performing multi-view tensor product operation according to the 3 graph models to obtain a plurality of multi-view tensor product maps, performing linear fusion processing according to the multi-view tensor product maps to obtain a fused multi-view tensor product map, performing diffusion according to a similarity matrix of the fused multi-view tensor product map to obtain a diffused similarity matrix, and performing spectral clustering on the diffused similarity matrix to classify the ground features in the polarized SAR image. By adopting the method, the interference of speckle noise can be reduced, and the classification precision is effectively improved.

Description

Unsupervised polarization SAR image terrain classification method based on multi-view tensor product diffusion

Technical Field

The application relates to the technical field of polarized SAR image classification, in particular to an unsupervised polarized SAR image terrain classification method based on multi-view tensor product diffusion.

Background

Synthetic Aperture Radar (SAR), as a technical means with all-weather, all-time, and high-resolution imaging capabilities in various remote sensing technologies, is gaining favor in the field of remote sensing and becomes an indispensable important branch in the technology of acquiring remote sensing information. Polarized SAR (polar synthetic aperture radar) alternately transmits and receives radar signals by adopting a horizontal polarization mode and a vertical polarization mode, and richer scattering information of a target object can be obtained by means of the penetration capability of microwaves on the ground object. Therefore, interpretation studies for polarized SAR images are particularly urgent, and geo-physical classification is the most fundamental and critical task in polarized SAR image interpretation. The crop growth monitoring system can monitor the crop growth condition, can be used for researching geological and mineral distribution, urban development and transition, mineral resource exploration, natural disaster assessment and the like, and can be widely applied to the military and civil fields.

Due to the difference of imaging modes, the polarized SAR image has different characterization modes from the optical image and also contains more speckle noise, so that the ground object classification of the polarized SAR image is a very challenging task all the time. The existing method mainly has the following problems:

(1) the existing classification method based on supervision has low automation degree and poor robustness. Generally, the supervised classification method needs to train a method model by using a data set with complete labels, needs more prior knowledge, consumes a large amount of manpower and material resources, and does not meet the requirements of practical application.

(2) The classification method based on the pixel points has large calculation amount and poor noise resistance. For a large-size remote sensing image, the calculation load of the pixel-by-pixel classification method is heavy, the regional information in the image is ignored, and the interference of speckle noise cannot be well reduced.

(3) A common way to construct high-dimensional feature vectors is feature stacking, which results in a reduced or lost classification capability of some single-view feature vectors. The polarization characteristics can objectively represent the similarity between data points in the polarized SAR image and are key elements of a polarized SAR image classification system. Generally, when a high-dimensional feature vector is constructed in a polarized SAR image classification method, extracted feature vectors are normalized and then directly added, and the method weakens the performance of single-view feature vectors and introduces classification errors.

Disclosure of Invention

In view of the above, it is necessary to provide a method for classifying surface features of unsupervised polarimetric SAR images based on multi-view tensor product diffusion, which can effectively improve the classification accuracy.

An unsupervised polarimetric SAR image terrain classification method based on multi-view tensor product diffusion, the method comprising:

obtaining a polarized SAR image to be classified, and segmenting the polarized SAR image by adopting a rapid superpixel segmentation method to obtain a plurality of superpixels;

extracting 5 representative feature vectors according to the polarized SAR image, combining the 5 feature vectors to obtain 3 high-dimensional feature vectors, and constructing and obtaining 3 graph models based on a plurality of segmented super-pixels by using the 3 high-dimensional feature vectors;

performing multi-view tensor product operation according to the 3 graph models to obtain a plurality of multi-view tensor product graphs, performing linear fusion processing according to the plurality of multi-view tensor product graphs to obtain a fused multi-view tensor product graph, and diffusing based on a similarity matrix of the fused multi-view tensor product graph to obtain a diffused similarity matrix;

and performing spectral clustering according to the diffused similarity matrix to realize the ground feature classification of the polarized SAR image.

In one embodiment, the combining according to the 5 feature vectors to obtain 3 high-dimensional feature vectors, and the constructing and obtaining 3 graph models based on the segmented plurality of superpixels by using the 3 high-dimensional feature vectors specifically includes:

extracting 5 eigenvectors from each pixel point in the polarized SAR image, and combining the 5 eigenvectors extracted according to each pixel point to obtain 3 high-dimensional eigenvectors corresponding to each pixel point;

calculating the average high-dimensional feature vector of each pixel point in the same superpixel in the plurality of superpixels on 3 high-dimensional feature vectors as the feature vector of the superpixel, so that each superpixel corresponds to 3 feature vectors;

and respectively constructing graph models according to the different 3 eigenvectors to obtain 3 corresponding graph models.

In one embodiment, the 5 representative feature vectors include: yamaguchi4 eigenvectors, Krogger eigenvectors, HSI color space eigenvectors, cloud-Pottier's eigenvectors, and eigenvectors consisting of the entropy of the scattering power and the co-polarizability.

In one embodiment, the obtaining 3 high-dimensional feature vectors corresponding to the pixel points by combining the 5 feature vectors corresponding to the pixel points includes:

forming one high-dimensional characteristic vector from the Yamaguchi4 characteristic vector, the HSI color space characteristic vector, the cloud-Pottier's characteristic vector and the characteristic vector formed by the scattering power entropy and the co-polarization rate;

forming one high-dimensional characteristic vector from the Yamaguchi4 characteristic vector, the Krogger characteristic vector, the cloud-Pottier's characteristic vector and the characteristic vector formed by the scattering power entropy and the co-polarization rate;

and forming one high-dimensional feature vector from the Yamaguchi4 feature vector, the Krogger feature vector, the HSI color space feature vector, the cloud-Pottier's feature vector and the feature vector formed by the scattering power entropy and the homopolarity.

In one embodiment, the graph model is composed of a plurality of nodes and edges between two adjacent nodes; wherein each of the nodes represents a different one of the plurality of superpixels and the edges represent a similarity by superpixels at both ends of the edge.

In an embodiment, the obtaining a plurality of multi-view tensor product maps by performing the multi-view tensor product operation according to the 3 map models includes:

and selecting any two of the 3 graph models to construct a multi-view tensor product map, and performing multi-view tensor product operation according to the 3 graph models to obtain 9 multi-view tensor product maps.

In one embodiment, constructing and obtaining a multi-view volume product graph according to two graph models includes: and performing Kronecker product calculation on the two graph models to obtain a multi-apparent-tensor product graph.

In one embodiment, the fused multi-tensor product diagram is obtained by performing linear fusion processing on the basis of the similarity matrix corresponding to each multi-tensor product diagram.

In one embodiment, the similarity matrix fused with the multi-view volume product graph is diffused to obtain a diffused similarity matrix, and an efficient iteration method is adopted when the expanded similarity matrix is calculated.

In one embodiment, the plurality of super pixels are obtained by segmenting the polarized SAR image based on a fast super image segmentation method initialized by a regular hexagon.

An unsupervised polarimetric SAR image terrain classification device based on multi-view tensor product diffusion, the device comprising:

the polarized SAR image acquisition module is used for acquiring a polarized SAR image to be classified and segmenting the polarized SAR image by adopting a rapid superpixel segmentation method to obtain a plurality of superpixels;

the graph model obtaining module is used for extracting 5 representative feature vectors according to the polarized SAR image, combining the 5 representative feature vectors to obtain 3 high-dimensional feature vectors, and constructing and obtaining 3 graph models based on the segmented multiple superpixels by using the 3 high-dimensional feature vectors;

the diffused similarity matrix obtaining module is used for performing multi-view tensor product operation according to the 3 graph models to obtain a plurality of multi-view tensor product graphs, performing linear fusion processing according to the plurality of multi-view tensor product graphs to obtain a fused multi-view tensor product graph, and diffusing based on a similarity matrix of the fused multi-view tensor product graph to obtain a diffused similarity matrix;

and the ground object classification module is used for carrying out spectral clustering according to the diffused similarity matrix so as to realize ground object classification of the polarized SAR image.

A computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:

A computer-readable storage medium, on which a computer program is stored which, when executed by a processor, carries out the steps of:

The unsupervised polarimetric SAR image terrain classification method based on the multi-view tensor product diffusion is characterized in that a polarimetric SAR image to be classified is segmented by adopting a rapid superpixel segmentation method to obtain a plurality of superpixels, extracting 5 representative eigenvectors according to the polarized SAR image, combining the 5 eigenvectors to obtain 3 high-dimensional eigenvectors, constructing 3 graph models based on the segmented multiple superpixels by using the 3 high-dimensional eigenvectors, performing multi-view tensor product operation according to the 3 graph models to obtain multiple multi-view tensor product graphs, and performing linear fusion processing according to the multiple multi-tensor product maps to obtain a fused multi-tensor product map, diffusing according to a similarity matrix of the fused multi-tensor product map to obtain a diffused similarity matrix, and performing spectral clustering on the diffused similarity matrix to classify the ground features in the polarized SAR image. By adopting the method, the interference of speckle noise can be reduced, and the classification precision is effectively improved.

Drawings

FIG. 1 is a schematic flowchart of an unsupervised polarimetric SAR image terrain classification method based on multi-view tensor product diffusion in an embodiment;

FIG. 2 is a schematic diagram of an embodiment of an algorithm based on the method of the present application;

FIG. 3 is a simplified diagram of a multi-view tensor product diagram in one embodiment;

FIG. 4 is a schematic diagram of a Flevoland measured data image in one embodiment;

FIG. 5 is a schematic diagram of an Oberpfaffenhofen measured data image in one embodiment;

FIG. 6 is a diagram illustrating the classification results of 5 methods for polarising SAR images based on Fleviland actual measurements in one embodiment;

FIG. 7 is a diagram illustrating the evaluation results of 5 methods for polarising SAR images based on Fleviland actual measurement in one embodiment;

FIG. 8 is a diagram illustrating the classification results of 6 methods based on Oberpfaffenhofen measured data in one embodiment;

FIG. 9 is a block diagram illustrating an embodiment of a device for classifying features of unsupervised polarized SAR images based on multi-view tensor product diffusion;

FIG. 10 is a diagram showing an internal structure of a computer device according to an embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

As shown in fig. 1-2, a method for unsupervised polarimetric SAR image terrain classification based on multi-view tensor product diffusion is provided, which comprises the following steps:

step S100, obtaining a polarized SAR image to be classified, and segmenting the polarized SAR image by adopting a rapid superpixel segmentation method to obtain a plurality of superpixels;

step S110, extracting 5 representative feature vectors according to the polarized SAR image, combining the 5 feature vectors to obtain 3 high-dimensional feature vectors, and constructing and obtaining 3 graph models based on a plurality of segmented superpixels by using the 3 high-dimensional feature vectors;

step S120, performing multi-view tensor product operation according to the 3 graph models to obtain a plurality of multi-view tensor product graphs, performing linear fusion processing according to the plurality of multi-view tensor product graphs to obtain a fused multi-view tensor product graph, and diffusing according to a similarity matrix of the fused multi-view tensor product graph to obtain a diffused similarity matrix;

and S130, performing spectral clustering according to the diffused similarity matrix to realize the ground feature classification of the polarized SAR image.

In step S100, a superpixel segmentation method is adopted to perform local clustering on a series of pixels with similar positions and similar low-level features in the polarized SAR image to form superpixels, and the polarized SAR image after superpixel segmentation can obtain a plurality of superpixels. After the polarized SAR image is subjected to superpixel segmentation, the influence of speckle noise inherent in the polarized SAR image can be overcome by using local information, so that the operation efficiency of subsequent processing is improved.

In this embodiment, the plurality of superpixels are obtained by segmenting the polarized SAR image based on a regular hexagon initialization fast superpixel segmentation method, and the segmentation process sequentially includes regular hexagon initialization, local k-means clustering, updating, and post-processing, and finally the superpixels are obtained.

In step S110, the 5 feature vectors are combined to obtain 3 high-dimensional feature vectors, and then the 3 high-dimensional feature vectors are used to construct 3 graph models based on the segmented plurality of superpixels, which specifically includes: extracting 5 feature vectors of each pixel point in the polarized SAR image, combining the 5 feature vectors extracted according to the corresponding pixel points to obtain 3 high-dimensional feature vectors corresponding to the pixel points, calculating the average high-dimensional feature vector of each pixel point in the same superpixel in a plurality of superpixels on the 3 high-dimensional feature vectors as the feature vector of the superpixel, enabling each superpixel to have 3 corresponding feature vectors, and finally respectively constructing a graph model according to the different 3 feature vectors to obtain 3 corresponding graph models.

Further, the polarized SAR image includes a plurality of different surface features, and each surface feature has a different sensitivity to different polarized scattering characteristics. Therefore, a certain number of features having strong discrimination ability for different surface feature types need to be selected, and through referring to relevant documents and performing experimental verification, in this embodiment, 5 kinds of feature vectors are selected, which are respectively Yamaguchi4 feature vector, Krogager feature vector, HSI color space feature vector, cloud-Pottier's feature vector, and feature vector composed of scattering power entropy and co-polarization rate, as shown in fig. 2.

The multi-view tensor product diffusion can be effectively combined with the classification capability of data of a plurality of different visual angles, generally, a better experimental result can be obtained when the number of the visual angles is multiple, and when each single visual angle has stronger classification capability, the multi-view tensor product diffusion can fully utilize the classification capability of different visual angles to strengthen the effective diffusion of the similarity information. Therefore, in order to ensure the classification capability of each view angle, the following combination method is also adopted in the present embodiment to combine five kinds of feature vectors into 3 kinds of high-dimensional feature vectors.

Wherein, the combination mode includes: forming one high-dimensional characteristic vector from the Yamaguchi4 characteristic vector, the HSI color space characteristic vector, the cloud-Pottier's characteristic vector and the characteristic vector formed by the scattering power entropy and the co-polarization rate; forming one high-dimensional characteristic vector from the Yamaguchi4 characteristic vector, the Krogger characteristic vector, the cloud-Pottier's characteristic vector and the characteristic vector formed by the scattering power entropy and the co-polarization rate; the Yamaguchi4 eigenvector, the Krogger eigenvector, the HSI color space eigenvector, the cloud-Pottier' eigenvector, and the eigenvector composed of the scattering power entropy and the co-polarizability are combined into one high-dimensional eigenvector, as shown in FIG. 2.

Specifically, in order to better perform feature representation on each super pixel in the plurality of super pixels, each pixel point in the original polarized image is represented by the five feature vectors, and then the five feature vectors are combined according to the combination mode to obtain three high-dimensional feature vectors, that is, each pixel point in the polarized SAR image can be represented by the three high-dimensional feature vectors. And then calculating the average characteristic vector of each pixel point under each high-dimensional characteristic vector after processing the same superpixel, and taking the average characteristic vector obtained by corresponding each high-dimensional characteristic vector as the characteristic vector of the corresponding superpixel, namely, each superpixel can be represented by three different characteristic vectors.

On the basis of superpixel segmentation, a polarized SAR image can be described using a weighted undirected graph model G ═ V, E. Wherein, the node V ═ { V ═ V₁,v₂,...,v_MRepresents the superpixel set, edge (v)_i,v_j) E is connected to node v_iAnd v_j. The nodes represent superpixels, the edges between the nodes represent the similarity between two superpixels, and each edge has a corresponding non-negative weight w_ij(i, j ═ 1, 2.. times, M), representing neighboring node v_iAnd v_jM represents the number of superpixels obtained after over-segmentation of the polarized SAR image. Thus, in this embodiment, 3 corresponding graph models can be constructed according to 3 different feature vectors.

And constructing an original similarity matrix W by adopting a Gaussian kernel based on each model, wherein matrix elements are non-negative weights W corresponding to each edge_ijMatrix element w_ijIs calculated as follows:

in the formula (1), d_ijRepresenting a superpixel v_iAnd a super pixel v_jThe Euclidean distance between, mu is a hyper-parameter, epsilon_ijIs a scale parameter used to eliminate the scaling problem, defined as:

in equation (2), d i, N_iTo representSuper pixel v_iAnd k-Nearest Neighbors (k-Nearest Neighbors, k-NN) of the super pixel N_iThe euclidean distance between, mean di, N_iMean Euclidean distance, mean dj, N_jDefinition of (1) and mean d i, N_iThe same is true.

Since each superpixel in an image can be represented by N different types of average feature vectors, N different graph models G can be constructed_(n)＝(V_(n),E_(n)) Wherein N is more than or equal to 1 and less than or equal to N.

In the invention, 3 eigenvectors are constructed, and other eigenvectors with different quantities can be extracted and combined according to different requirements of actual application conditions. Therefore, each polarimetric SAR image can be composed of 3 map models G_(n)＝(V_(n),E_(n)) (1. ltoreq. N. ltoreq.N) and the graph model here is a single-view model, G_(n)＝(V_(n),E_(n)) Representing a single-view model constructed with the nth feature vector, N being the number of graph models constructed with different features and combinations (N being 3 in this embodiment), and W being_nThe similarity matrix corresponding to each graph model is obtained by constructing the edges in the graph models.

In step S120, multi-view tensor product diffusion is performed based on the graph model, and finally the similarity matrix after cross diffusion is obtained. In a traditional unsupervised polarized SAR image classification system, a similarity matrix is usually determined by pairwise similarity between data points (pixels or superpixels), and global feature information of a polarized SAR image is ignored. Therefore, a method of diffusing the similarity measure to the surroundings on the graph structure is proposed, and the learning of the similarity is realized by considering the relationship between each data point and the points in its neighborhood. Generally, when high-dimensional feature vectors are constructed in a polarized SAR image classification method, the extracted feature vectors are directly stacked end to end, although the method is simple and feasible and integrates various feature information, the method weakens the performance of the feature vectors with strong distinguishing capability of part of surface features and introduces classification errors. The multi-view tensor product can fully utilize complementary information among different polarization scattering characteristics, and based on the internal relation among data points (super pixels), a similarity matrix after cross diffusion with stronger classification capability is generated.

In this embodiment, the similarity information is transmitted and diffused on the multi-view tensor product diagram, and compared with the original diagram, the multi-view tensor product diagram considers the higher-order context relationship, introduces the thought of multi-view learning, and can better disclose the internal manifold structure of the data to construct a similarity matrix with stronger discrimination capability and classification capability, so that the precision of ground feature classification is improved.

Before the multi-view tensor product diffusion explanation based on the graph model is carried out, the process of diffusion based on the original graph is introduced, namely, the diffusion is directly carried out based on the original graph model, and the subsequent understanding is facilitated.

Graph-theory based diffusion processes can reveal the inherent geometric relationships between data points. The simplest understanding of graph-based diffusion is the product W of the graph similarity matrix^t(t is the number of iterations). From here it is known that the diffusion of the graph is based on a similarity matrix. However, this classical diffusion process is limited by the number of iterations t. If the row and column of the similarity matrix W are both smaller than 1, where "row and column" represents the sum of elements in each row in the matrix, the classical diffusion process based on W will eventually converge to a 0 matrix, and therefore, the setting of the number of iterations becomes crucial. In order to reduce the influence of the iteration number t on the diffusion process, a diffusion form for weighting the similarity matrix W may be adopted, as shown in the following formula:

in equation (3), in order to avoid convergence of the diffusion process to a 0 matrix, it is necessary to ensure that W is a non-negative matrix and satisfies

(M is the number of super pixels), namely, both rows and columns are less than 1; for any i, if any

Then formula (3) is noneAnd (4) converging the method. According to the matrix principle, the matrix w satisfying the above conditions_ijMay be generated by a random matrix transformation. After the condition is satisfied, equation (3) can converge to:

in the formula (4), I₁Is an identity matrix with dimensions identical to W.

In this embodiment, the constructed three graph models are transformed into a plurality of multi-view tensor product graphs by multi-view tensor product operation, and then expanded. The process of obtaining a plurality of multi-volume images comprises the following steps: and selecting any two of the 3 graph models to construct a multi-view tensor product graph, and performing multi-view tensor product operation according to the 3 graph models to obtain 9 multi-view tensor product graphs. And performing linear fusion processing on the similarity matrixes corresponding to the 9 multi-tensor product maps to obtain a fused multi-tensor product map, and finally obtaining a similarity matrix after cross diffusion based on the fused multi-tensor product map.

Specifically, a multi-view tensor product diagram (i.e., a multi-view tensor product diagram) is adopted

Is defined as:

multi-view field tensor product diagram

Each node in the graph is respectively constructed by a single-view model G from different feature vectors_kAnd G_lIs made up of one node in the group,

the similarity matrix is defined as

Wherein,

representing the Kronecker product. In particular, for α, β, i, j ═ 1, 2.

Therefore, if

Then

A simplified example thereof is shown in fig. 3.

Compared with other multi-view learning methods based on graph theory, in the method, two Kronecker product operations are performed on N graph models constructed based on the feature vectors, so that N × N multi-view tensor product graphs are obtained instead of N, and therefore in the embodiment, 9 multi-view tensor product graphs can be obtained finally.

Compared with the original image, each node in the image contains richer information, and a single-view model constructed based on different feature vectors captures higher-order similarity information, so that the internal relation between data points is effectively disclosed.

After a plurality of multi-view tensor product maps are obtained, a fused multi-view tensor product map is obtained based on linear fusion processing, and expansion is carried out based on a similar feature matrix of the multi-view tensor product map.

Among the graph model fusion methods applied to multi-view clustering based on graph theory, linear addition of a single graph model is the simplest and easiest processing mode. Therefore, after obtaining N × N multi-view tensor product maps, in this embodiment, the similarity matrix corresponding to each multi-view tensor product map is subjected to linear fusion processing to implement a diffusion process based on the multi-view tensor product maps, and the obtained fused multi-view tensor product map has the following formula:

according to the diffusion process, the diffusion process based on the multi-view tensor product diagram is similar to the formula (3), as shown in the following formula:

similar to the convergence shown in equation (4), the diffusion process shown in equation (8) can also converge according to the following equation:

in formula (9), I₂Is prepared by reacting with

And (4) identity matrix. However, as previously mentioned, to avoid convergence of the diffusion process to a 0 matrix, it must be ensured that

Both rows and columns of (c) are less than 1.

In the process of diffusion based on the similarity matrix expressed by the formula (9), the tensor product diagram is generated due to the multi-view angles

Corresponding similarity matrix

The dimensionality is very high, the calculation efficiency of the diffusion and convergence processes is very low, and the similarity matrix obtained after diffusion

With the original G_(n)Corresponding similarity matrix W_nAre also not uniform. Therefore, in this embodiment, it is desirable to be able to use the original image G_(n)The similarity information is learned by using the multi-view tensor product, i.e. the similarity matrix W after diffusion^*Dimension is M × M, and similarity information of multiple views is combined:

in equation (10), vec is an operator, which functions to convert each column of a matrix into a column vector by stacking, vec^-1To convert a column vector into the inverse of a matrix.

After calculation according to the formula (10), the similarity matrix after cross diffusion can be obtained, but since the multiview tensor product map has a higher order than the original map, the calculation of the formula (10) consumes more calculation amount and storage space when some larger maps are processed. Therefore, in order to improve the diffusion efficiency of the multi-view tensor product map, a new iteration method based on the original image, namely a c-TPGD high-efficiency iteration method is adopted, which is equivalent to the diffusion process based on the multi-view tensor product map, but saves a large amount of calculation time and storage space, and the calculation process is as follows:

first, define:

repeating the iteration of equation (11) until Q converges, so as to obtain the similarity matrix after cross diffusion:

for ease of understanding, the present invention will be briefly described below with respect to the proof concept of equation (12). First, Q is added^(t ⁺¹⁾When the system is unfolded, the following steps are carried out:

due to the fact that

(M is the number of super pixels), i.e. W_(k)Both rows and columns of (1) are less than 1, thus having

Therefore, the method comprises the following steps:

the following identity is known:

thus, it is possible to obtain:

wherein,

the following was demonstrated:

by combining equations (13) through (17), we can obtain:

thus, it is finally obtained:

therefore, the new iterative method based on the original graph is equivalent to the diffusion process based on the multi-view tensor product graph, but can save a large amount of computing time and storage space. The obtained new cross-diffused similarity matrix contains higher-order internal information between data points, and generally, when the iteration number is 20, the classification capability of the diffused similarity matrix can reach a relatively stable level, so that the iteration number t is set to be 20 in the experiment of the method.

In step S130, a similarity matrix after cross diffusion is obtained based on the above calculation, and a result of the ground feature classification in the polarized SAR image can be obtained by performing spectral clustering.

In particular, the clustering method is receiving more and more attention because it can obtain better clustering result on the (ground feature) feature space with any shape and has a more perfect mathematical framework. The spectral clustering method is based on the feature decomposition of the similarity matrix and further utilizes a k-means method for clustering, in this embodiment, a new similarity matrix generated by tensor product diffusion is used as the input of the spectral clustering method for feature decomposition, so as to obtain the final classification result.

To evaluate the performance of the method, multiple sets of experiments were developed based on multiple data and analyzed in detail. The experimental part is described in detail below:

in the experiment, the first actually measured polarized SAR image is a partial area of a Flevoland test area shot by AirSAR, the area is 300 × 270 pixels, the true value map is shown in fig. 4(a), and the corresponding Pauli-RGB image and the generated superpixel map are respectively shown in fig. 4(b) and (c). The second actually measured polarized SAR image is an L-band image shot by the ESAR, the shooting area is located in the Oberpfaffenhofen test area, the size of the image is 700 × 1000 pixels, the true value image is shown in fig. 5(a), and the Pauli-RGB image and the generated super-pixel image are respectively shown in fig. 5(b) and (c). The actually measured polarized SAR image mainly comprises 3 types of ground objects: woodland, open area 1 and open area 2.

The organization of this experimental part is as follows. In order to verify the effectiveness of the classification method in the application, 5 methods of comparison experiments are carried out on the basis of a Flevoland actually-measured polarized SAR image. The 5 methods involved in the comparison include: an Unsupervised k-means pixel-based wished Classification method (Unsupervised k-means wished Classification algorithm based on Pixels), an Unsupervised k-means wished Classification algorithm based on Superpixels (Unsupervised k-means wished Classification algorithm based on Superpixels, UKWC-S), a pixel-level Unsupervised Classification method based on Scattering power entropy and co-polarization (Unsupervised Classification based on Scattering power entropy and co-polarization, UCSC-P), a superpixel-level Unsupervised Classification method based on Scattering power entropy and co-polarization (Unsupervised Classification based on Scattering power entropy and co-polarization).

Then, a comparative analysis experiment of 6 methods is carried out on the basis of the Oberpfaffenhofen actually-measured polarized SAR image, which comprises the following steps: the method comprises the following steps of unsupervised classification based on single vision tensor product diffusion (TPGD method), UCSC-S method, UKWC-S method, superpixel-level Wishart classification based on geodesic distance (GDWC-S method), superpixel-level Wishart classification based on polarization decomposition (CPWC-S method) and the method (c-TPGD) of the invention.

In order to ensure the fairness of the comparison experiments, the category number in all the experiments is manually given in advance according to the priori knowledge, and the experiment parameters of the comparison method are set according to the optimal parameters of the invention of the corresponding theory. In the experiment, 4 evaluation metrics were taken to evaluate effectiveness: the method includes a User Accuracy (UA, which indicates the proportion of pixels with the measurement type of i-th class among all pixels classified into the i-th class), a drawing Accuracy (Producer Accuracy, PA, which indicates the proportion of pixels with the correct classification of i-th class among all pixels with the measurement type of i-th class), an Overall Accuracy (OA, which indicates the proportion of pixels with the correct classification among all samples), and a Kappa coefficient (K, which integrates UA and PA to evaluate the Accuracy of classified images).

Comparative experiment based on Flevoland actually measured polarized SAR data:

to further evaluate the performance of the c-TPGD method, 5 unsupervised classification methods including UKWC-P, UKWC-S, UCSC-P, UCSC-S and the method of the present invention were performed on the actually measured polarized SAR image, and the experimental results are shown in FIG. 6, wherein FIG. 6(a) is the experimental result of UKWC-P, FIG. 6(b) is the experimental result of UKWC-S, FIG. 6(c) is the experimental result of UCSC-P, FIG. 6(d) is the experimental result of UCSC-S, and FIG. 6(e) is the experimental result of the method. Since these 5 methods are unsupervised, the labels of the classification results are random, and for easy observation and comparison, the final label is re-labeled according to the truth diagram, but this does not affect the result of the unsupervised classification, and some classes of some classification results may be mistakenly classified into one class, and the re-labeling principle is to maximize the OA value.

As can be seen from fig. 6(a) and (c), the experimental results of the UKWC-P and UCSC-P methods are greatly affected by speckle noise, and particularly, the UCSC-P method has a Kappa coefficient of only 0.3195 due to interference from speckle noise, although it has a good classification ability for each class based on visual observation. This indicates that the classification method based on the pixel points cannot well resist the interference of speckle noise inherent in the polarized SAR image. Fig. 6(b) and (d) show the classification results of the two methods based on the super-pixel, respectively, and it can be clearly seen that compared with the classification result based on the pixel, the method can better overcome the influence caused by speckle noise, and the classification result reduces a lot of "speckle" phenomena, which indicates that the classification method based on the super-pixel can effectively improve the classification accuracy while overcoming the interference of the speckle noise.

Some superpixels in FIGS. 6(b) and (d) are misclassified, probably because both UKWC-S and UCSC-S methods use insufficient features to describe all terrain features. It can be clearly seen that the experimental result of fig. 6(e) is better than that of other examples, because the method of the present invention selects and combines 3 effective eigenvectors, performs diffusion of similarity information based on a multi-view tensor product diagram, and effectively combines the classification capability between polarization scattering features, so that the similarity matrix after cross diffusion has stronger discrimination capability on ground objects.

To further quantitatively evaluate these 6 methods, evaluation was performed using UA, PA, OA, and K4 evaluation metrics, and the corresponding results are shown in fig. 7, where fig. 7(a) is the user accuracy of 5 methods, fig. 7(b) is the charting accuracy of 5 methods, fig. 7(c) is the overall accuracy of 5 methods, and fig. 7(d) is the Kappa coefficient of 5 methods. As can be seen from the data in the figures, the c-TPGD method is generally higher than other methods for 6 terrain categories of mapping accuracy versus user accuracy. The overall accuracy and Kappa coefficient of the classification results of the 5 methods are shown in fig. 7(c) and (d), respectively, and it can be seen that the overall accuracy and Kappa coefficient of the classification results of the method of the present invention are the highest. Meanwhile, the instability of the UCSC method can be seen, and the selected characteristics are single, so that the classification requirement of complex actually-measured polarization SAR data can not be well met. The method can embody strong ground feature distinguishing capability on the complex measured data, further shows that the multi-view tensor product diagram can better reveal the internal relation of the data, effectively fuses the similarity information of a plurality of views, and accordingly improves the classification precision of the ground features.

A comparative experiment based on Oberpfaffenhofen actually-measured polarized SAR data is as follows:

the results of the 6 methods of comparison experiments based on measured data of Oberpfaffenhofen are shown in FIG. 8, in which FIG. 8(a) is the S1-TPGD method, FIG. 8(b) is the S2-TPGD method, FIG. 8(c) is the S3-TPGD method, FIG. 8(d) is the UCSC-S method, FIG. 8(e) is the UKWC-S method, FIG. 8(f) is the GDWC-S method, FIG. 8(g) is the CPWC-S method, FIG. 8(h) is the c-TPGD method, and Table 1 shows the PA values, Kappa coefficients, and OA values of the 6 methods. Since the TPGD method is the tensor product diffusion of the single-view eigenvectors, the 3 original similarity matrices S1, S2, and S3 shown in fig. 2 are used as the input of the single-view tensor product diffusion to obtain the final classification result, and are referred to as the S1-TPGD, S2-TPGD, and S3-TPGD methods, respectively.

Table 16 methods 3 evaluation measurement results based on Oberpfaffenhofen measured data

As shown in the area B in fig. 8(d), the UCSC-S algorithm misclassifies a large number of open areas 2 into open areas 1, while part of feature edges and homogeneous areas in fig. 8(e) exhibit incomplete fracture, which indicates that the discrimination capability of the feature may be affected by not extracting more polarization scattering features. There are many isolated small regions in fig. 8(g), which indicates that the classification result of the CPWC-S algorithm is severely affected by speckle noise inherent in the polarized SAR image, as shown in region a, and the algorithm cannot classify the open region 2 in region B. The OA values of the GDWC-S algorithm can reach 80.74%, which are all higher than the 4 algorithms mentioned above, again proving the importance of finding the shortest distance between data points (superpixels). However, looking at region a in fig. 8(c), there is still much noise, mainly due to the phenomenon that it is not classified in conjunction with many valid polarimetric SAR features.

In fig. 8, (a), (b), (c), and (e) all adopt a tensor product diffusion-based method, but the TPGD algorithm only considers single-view feature information, and thus the classification effect is not ideal. The algorithm combines various typical polarization SAR characteristics, performs diffusion of similarity information based on a multi-tensor product diagram, and effectively combines the classification capability among polarization scattering characteristics, so that a similarity matrix after cross diffusion has stronger discrimination capability on ground objects.

In the method for classifying the ground features of the unsupervised polarized SAR image based on the multi-view tensor product diffusion, the unsupervised classification method is adopted, a large amount of manpower and material resources are not needed to be consumed, a complete data set is labeled, the automation degree is high, and the actual requirements are met. The method also adopts a multi-view tensor product diffusion technology, so that the intrinsic similarity information of the data can be effectively mined, the interference of speckle noise is reduced, and the classification precision is improved. The method provides an unsupervised polarimetric SAR image terrain classification framework based on multi-view tensor product diffusion, the framework has universality, and the purpose of terrain classification can be achieved by combining with other methods for measuring similarity. I.e. the method is migratable.

It should be understood that, although the steps in the flowchart of fig. 1 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least a portion of the steps in fig. 1 may include multiple sub-steps or multiple stages that are not necessarily performed at the same time, but may be performed at different times, and the order of performance of the sub-steps or stages is not necessarily sequential, but may be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

In one embodiment, as shown in fig. 9, there is provided an unsupervised polarized SAR image terrain classification device based on multi-view tensor product diffusion, including: a polarized SAR image obtaining module 200, a graph model obtaining module 210, a similarity matrix after diffusion obtaining module 220 and a ground feature classification module 230, wherein:

the polarized SAR image acquisition module 200 is used for acquiring a polarized SAR image to be classified and segmenting the polarized SAR image by adopting a fast superpixel segmentation method to obtain a plurality of superpixels;

a graph model obtaining module 210, configured to extract 5 representative feature vectors according to the polarized SAR image, combine the 5 representative feature vectors to obtain 3 high-dimensional feature vectors, and construct 3 graph models based on the segmented plurality of superpixels by using the 3 high-dimensional feature vectors;

a diffused similarity matrix obtaining module 220, configured to perform multi-view tensor product operation according to the 3 graph models to obtain multiple multi-view tensor product graphs, perform linear fusion processing according to the multiple multi-view tensor product graphs to obtain a fused multi-view tensor product graph, and perform diffusion based on a similarity matrix of the fused multi-view tensor product graph to obtain a diffused similarity matrix;

and the surface feature classification module 230 is configured to perform spectral clustering according to the diffused similarity matrix to realize surface feature classification on the polarized SAR image.

The specific definition of the unsupervised polarized SAR image surface feature classification device based on the multi-view tensor product diffusion can refer to the definition of the unsupervised polarized SAR image surface feature classification method based on the multi-view tensor product diffusion in the above, and is not described herein again. The modules in the device for classifying the terrestrial objects based on the multi-view tensor product diffusion unsupervised polarized SAR image can be completely or partially realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. 10. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile 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 operation of an operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to realize an unsupervised polarimetric SAR image terrain classification method based on multi-view tensor product diffusion. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.

Those skilled in the art will appreciate that the architecture shown in fig. 10 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided, comprising a memory and a processor, the memory having a computer program stored therein, the processor implementing the following steps when executing the computer program:

In one embodiment, a computer-readable storage medium is provided, having a computer program stored thereon, which when executed by a processor, performs the steps of:

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in the embodiments provided herein may include non-volatile and/or volatile memory, among others. 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 a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. The unsupervised polarized SAR image terrain classification method based on the multi-view tensor product diffusion is characterized by comprising the following steps of:

2. The unsupervised polarimetric SAR image terrain classification method according to claim 1, wherein said combining according to the 5 feature vectors to obtain 3 high-dimensional feature vectors, and constructing and obtaining 3 map models based on the segmented plurality of superpixels using the 3 high-dimensional feature vectors specifically comprises:

calculating the average feature vector of 3 high-dimensional feature vectors of all pixel points in the same superpixel in the plurality of superpixels as the feature vector of the superpixel, so that each superpixel corresponds to 3 feature vectors;

3. The unsupervised polarimetric SAR image terrain classification method of claim 2, wherein said 5 representative feature vectors comprise: yamaguchi4 eigenvectors, Krogger eigenvectors, HSI color space eigenvectors, cloud-Pottier's eigenvectors, and eigenvectors consisting of the entropy of the scattering power and the co-polarizability.

4. The unsupervised polarimetric SAR image terrain classification method of claim 3, wherein said combining according to the 5 eigenvectors corresponding to each pixel point to obtain the 3 high-dimensional eigenvectors corresponding to each pixel point comprises:

5. The unsupervised polarimetric SAR image terrain classification method according to any of claims 1-4, characterized in that the graph model consists of a plurality of nodes and edges between two adjacent nodes; wherein each node represents a different superpixel in the superpixel segmented image, and the edge represents the similarity of superpixels at both ends of the edge.

6. The unsupervised polarimetric SAR image terrain classification method of claim 5, wherein said performing a multi-view tensor product operation according to 3 of said graph models to obtain a plurality of multi-view tensor product graphs comprises:

7. The unsupervised polarimetric SAR image terrain classification method of claim 6, characterized in that constructing and obtaining a multi-volume-of-view map according to two map models comprises: and performing Kronecker product calculation on the two graph models to obtain a multi-apparent-tensor product graph.

8. The unsupervised polarimetric SAR image terrain classification method according to claim 7, characterized in that the fused multi-tensor product map is obtained by performing linear fusion processing based on a similarity matrix corresponding to each multi-tensor product map.

9. The unsupervised polarimetric SAR image terrain classification method according to claim 8, wherein a similarity matrix based on the fused multi-view volume product map is diffused to obtain a diffused similarity matrix, wherein an efficient iteration method is adopted when the expanded similarity matrix is calculated.

10. The unsupervised polarimetric SAR image terrain classification method according to claim 1, characterized in that the plurality of super pixels are obtained by segmenting the polarimetric SAR image based on a fast super image segmentation method initialized by regular hexagons.