CN111680731B - Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning - Google Patents

Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning Download PDF

Info

Publication number
CN111680731B
CN111680731B CN202010486706.9A CN202010486706A CN111680731B CN 111680731 B CN111680731 B CN 111680731B CN 202010486706 A CN202010486706 A CN 202010486706A CN 111680731 B CN111680731 B CN 111680731B
Authority
CN
China
Prior art keywords
dictionary
matrix
polarized
sparse
classification
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010486706.9A
Other languages
Chinese (zh)
Other versions
CN111680731A (en
Inventor
曲延云
赖轩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xiamen University
Original Assignee
Xiamen University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xiamen University filed Critical Xiamen University
Priority to CN202010486706.9A priority Critical patent/CN111680731B/en
Publication of CN111680731A publication Critical patent/CN111680731A/en
Application granted granted Critical
Publication of CN111680731B publication Critical patent/CN111680731B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

A polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning relates to image processing. 1) Inputting a to-be-classified polarized synthetic aperture radar SAR image and a real ground object marker image thereof to obtain a polarized coherent matrix thereof; 2) Filtering the extracted coherent matrix to remove speckle noise; each category selects 5% of data as training samples, and the rest as test samples; 3) Respectively using k-means clustering to obtain a plurality of clustering centers for each type of samples, and forming an initial dictionary by taking the clustering centers as dictionary elements; 4) Calculating sparse vectors of each training sample through dictionary representation; 5) Training an SVM classifier according to the sparse vector and the label information; 6) Updating the dictionary by using the sparse vector and the classifier; 7) Returning to the step 4) if the dictionary updating amplitude is larger than 1e-5, otherwise, executing the step 8); 8) And calculating sparse vectors of each test sample through dictionary representation, and obtaining a final classification result by using an SVM classifier.

Description

Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning
Technical Field
The invention relates to image processing, in particular to a polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning, which can be used for polarized SAR image ground feature classification.
Background
In the last decades, polarized synthetic aperture radar (PolSAR) has attracted attention from researchers due to its wide range of applications as an important branch of remote sensing research. One of the most important applications of PolSAR is image classification, where each pixel is assigned to one terrain type. The classification map may be used directly in an application or as input to further processing steps such as disaster management, site interpretation, environmental monitoring, agriculture, etc. However, the POLSAR image classification is considered a difficult task due to its complex representation and strong speckle interference. Although many approaches have been proposed, the POLAR image classification problem remains an active topic of research.
In general, a PolSAR that transmits and receives electromagnetic waves in different states can provide more detailed information, represented as a quadrupolar matrix, rather than a single value in a monopolar SAR. In order to intuitively express implicit features about geometric and physical attribute information, such as statistics, scattering, texture, spatial and color information, a number of useful decomposition algorithms have been proposed.
CloudePattier et al (S.R.Cloude and E.Pottier, "An entropy based classification scheme for land applications of polarimetric sar," IEEE Transactions on Geoscience and Remote Sensing, vol.35, no.1, pp.68-78, jan 1997.) use a three-level Bernoulli statistical model to generate estimates of the average target scattering matrix parameters from the data. Yamaguchi et al extended the Freeman and Durden introduced three-component decomposition method, which was extended to a four-component decomposition algorithm that handled the general scattering case. However, each proposed decomposition algorithm cannot represent the perfection of the original data, since some information is lost during the decomposition. On the other hand, stacked high-dimensional polarization feature vectors remain the most important information, which is superfluous, as these polarization features can be decomposed from other algorithms and can be represented in other ways. Therefore, it is necessary to extract appropriate low-dimensional features to improve classification performance. Various polarization features were selected by y.wang et al (y.wang, et al, "Hierarchical polarimetric sar image classification based on feature selection and genetic algorithm," in 2014 12th International Conference on Signal Processing (ICSP), oct 2014, pp.764-768.) to construct a high-dimensional polarization manifold that was mapped into the most compact low-dimensional structure by a dimension-reduction algorithm.
Disclosure of Invention
The invention aims to provide a polarized SAR image supervision and classification method based on geometric sense discrimination dictionary learning, which aims to overcome the defects in the prior art and improve model robustness and classification accuracy.
The invention comprises the following steps:
1) Inputting a to-be-classified polarized synthetic aperture radar SAR image and a real ground object marker image thereof to obtain a polarized coherent matrix T;
2) Filtering the extracted coherent matrix to remove speckle noise and obtain a filtered coherent matrix; each category selects 5% of data as training samples, and the rest as test samples;
3) Respectively using k-means clustering to obtain a plurality of clustering centers for each type of samples, and forming an initial dictionary by taking the clustering centers as dictionary elements;
4) Calculating sparse vectors of each training sample through dictionary representation;
5) Training an SVM classifier according to the sparse vector and the label information;
6) Updating the dictionary by using the sparse vector and the classifier;
7) Returning to the step 4) if the dictionary updating amplitude is larger than 1e-5, otherwise, directly executing the step 8);
8) And calculating sparse vectors of each test sample through dictionary representation, and obtaining a final classification result by using an SVM classifier.
In step 2), filtering the extracted coherent matrix to remove speckle noise, wherein the obtained filtered coherent matrix is based on the polarized coherent matrix T obtained in step 1), filtering the extracted coherent matrix by using a Box Car filtering method with a window size of 7×7, reducing the speckle effect of the PolSAR image, removing the speckle noise, and filtering the pixel points with the trace of the matrix smaller than 1e-5 as outliers.
In step 3), the specific steps of forming the initial dictionary may be:
for each class, according to the training sample number of each class as n, respectively performing k-means clustering, wherein the distance measurement between each sample and the clustering center is defined as KL divergence distance, and sample X 1 ，X 2 The KL divergence of (2) is defined as:
Figure BDA0002519375350000021
where tr (X) denotes the trace of the matrix and size (X, 1) denotes the number of rows of the matrix.
And (3) carrying out loop iteration until the cluster center does not move any more, obtaining a cluster center of one category, and taking the cluster centers obtained by calculating all the categories as dictionary elements to form an initial dictionary.
In step 4), the dictionary uses the following dictionary model:
Figure BDA0002519375350000031
wherein,,
Figure BDA0002519375350000032
representation dictionary X i Is the ith coherent matrix X, z i X represents i In dictionary->
Figure BDA0002519375350000033
Sparse representation on->
Figure BDA0002519375350000034
Sparsity representing a sparse vector matrix Z, +.>
Figure BDA0002519375350000035
Representation dictionary->
Figure BDA0002519375350000036
L (Z, y, ω, b) represents the sparsity of the sparse vector Z for the label y classification penalty in an SVM classifier with a hyper-parameter ω, b, wherein:
Figure BDA0002519375350000037
wherein θ is a weight coefficient.
When dictionary
Figure BDA0002519375350000038
Fixed with the SVM classifier parameters omega, b, for a given data matrix X i ∈H d + The minimum value of the sparse vector z can be expressed as the following sub-problem:
Figure BDA0002519375350000039
for z j Obtaining the partial derivative:
Figure BDA00025193753500000310
wherein,,
Figure BDA00025193753500000311
in step 5), the specific method for training the SVM classifier according to the sparse vector and the label information may be: the dictionary may be fixed when the sparse vector z and the SVM classifier parameters ω, b are both fixed
Figure BDA00025193753500000312
Expressed as the following non-convex optimization problem:
Figure BDA00025193753500000313
the method adopts the Riemann Conjugate Gradient (CG) method to derive the Riemann conjugate gradient, and the Riemann conjugate gradient is used for nonlinear function Θ (B i ),
Figure BDA00025193753500000314
Dictionary element of step k+1->
Figure BDA00025193753500000315
Expressed as:
Figure BDA00025193753500000316
wherein, gamma k Is the step size found by an efficient linear search method, while the direction of descent ζ (k) The definition is as follows:
Figure BDA0002519375350000041
wherein grad represents the gradient function, map Φ X (Y) defines a vector transfer of two points, for two points X, Y ε T on the Riemann manifold P M，
Figure BDA0002519375350000042
Figure BDA0002519375350000043
Figure BDA0002519375350000044
Where Θ (B) is the euclidean gradient of Θ (B), and the dictionary element of the j th generation is defined as:
Figure BDA0002519375350000045
order the
Figure BDA0002519375350000046
Derivative->
Figure BDA0002519375350000047
Can be approximated as +.>
Figure BDA0002519375350000048
In step 8), the dictionary model used is:
Figure BDA0002519375350000049
the class to which the sparse vector z belongs is calculated using an SVM classifier:
Figure BDA00025193753500000410
and predicting the category of each pixel point in turn, so that the classification of the planning image can be realized.
The method starts from the three-dimensional tensor dictionary learning optimization, considers the category information and the inherent Riemann geometric information of the data, so that the calculated dictionary is more suitable for the data set, the robustness of the model is improved, and the classification precision is effectively improved.
The invention constructs a three-dimensional polarization feature tensor dictionary by utilizing a polarization coherent matrix extracted from polarization data, obtains low-dimensional feature vectors according to a dictionary representation method, classifies the low-dimensional feature vectors by using a maximum interval support vector machine, updates the tensor dictionary according to a trained classifier and the feature vectors, and continuously loops until the dictionary update amplitude is smaller than a threshold value, thereby obtaining a proper dictionary and classifying SAR images. The method and the device find out proper dictionary transformation on the basis of not damaging the three-dimensional polarization characteristic tensor structure, remove redundancy among characteristic quantities, avoid the problem of dimension disasters, well improve the classification effect, improve the efficiency and the robustness of the algorithm and can be applied to classification of various complex terrains.
Compared with the prior art, the invention has the following outstanding advantages and technical effects:
1. the invention utilizes a novel geometric perception discriminant dictionary learning framework to carry out PolSAR image classification, each data point is expressed as a nonnegative linear combination of HPD atoms from a learning dictionary, and original data is characterized as sparse coding coefficients by coding category information and inherent Riemann geometric information, and the original data is converted into sparse representation without destroying three-dimensional polarization characteristic tensors.
2. According to the method, the proposed objective function is solved by using an alternate direction multiplier method, and the dictionary obtained by alternate iteration is more robust compared with the initial clustering dictionary. Experiments prove that the method used by the invention has good effect on the currently disclosed labeled data set and is superior to other currently best methods.
Drawings
FIG. 1 is a general flow chart of the present invention;
FIG. 2 is a graph of the classification results of the Flevoland1989 dataset using the present invention and four methods available. In fig. 2, (a) is a visual image of the raw data decomposed by Pauli; (b) is an example of the color to which each category corresponds; (c) is the result of a genuine label; (d) Wishare-ML results; (e) LE-NDR results; (f) ND-KSVD is the result; (g) is the result obtained by RSC-SVM; (h) the results obtained by the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described with reference to the following examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. On the contrary, the invention is intended to cover any alternatives, modifications, equivalents, and variations as may be included within the spirit and scope of the invention as defined by the appended claims. Further, in the following detailed description of the present invention, certain specific details are set forth in order to provide a thorough understanding of the present invention. The present invention will be fully understood by those skilled in the art without the details described herein.
The embodiment of the invention comprises the following steps:
(1) Inputting a polarized SAR image to be classified to obtain a polarization coherence matrix T thereof;
(2) Based on a polarization coherence matrix T in the polarized SAR image, reducing the speckle effect of the PolSAR image by using a BoxCar filter with a window size of 7 multiplied by 7, and filtering outliers with the trace of the matrix smaller than 1 e-5; randomly selecting 5% of data from each class as a training set, and the rest as a test sample set;
(3) For each class of data, performing k-means unsupervised clustering to obtain multiple clustering centers serving as dictionary elements to form an initial dictionary
Figure BDA0002519375350000051
(4) For each sample in the training set, minimizing dictionary representations
Figure BDA0002519375350000061
Obtain the corresponding sparse representation z i ；
(5) Using sparse representation z i Minimizing classification loss by tag information y
Figure BDA0002519375350000062
Training SVM scoreA class device;
(6) Fixed sparse representation z i Learning a new dictionary with a classifier by minimizing discriminative dictionary
Figure BDA0002519375350000063
Figure BDA0002519375350000064
(7) Fixed dictionary
Figure BDA0002519375350000065
Learning new sparse representation z with classifier by minimizing discriminant dictionary i new ：
Figure BDA0002519375350000066
(8) Setting update dictionary
Figure BDA0002519375350000067
Judging whether the iteration termination condition of dictionary change is satisfied
Figure BDA0002519375350000068
Where tr is the trace of the matrix, +.>
Figure BDA0002519375350000069
Representing a dictionary generated in the previous iteration, wherein k is the iteration number, epsilon is a set value, and represents convergence accuracy;
(9) If the convergence condition is not satisfied, returning to the step (5) to obtain a new sparse representation z i new Substitute original sparse representation z i Performing iterative loop steps (5) to (7) until convergence conditions are met, and obtaining an optimized dictionary
Figure BDA00025193753500000610
And a corresponding SVM classifier;
(10) For a pair ofIn test samples, representation by minimizing dictionary
Figure BDA00025193753500000611
And obtaining a corresponding sparse vector, and classifying by using an SVM classifier to obtain a polarized image classification result graph based on geometric perception discriminant dictionary learning.
Specific examples are given below.
Referring to fig. 1, the steps of the present embodiment are as follows:
step 1, inputting a polarized SAR image to be classified to obtain a polarized coherent matrix T of the SAR image.
Referring to fig. 2, the polarized SAR image is a fleveland 1989 farmland chart, the types of the ground objects to be classified include beans, forests, potatoes, alfalfa, wheat, bare land, beet, rape, peas, grasses and water, different colors in the chart represent different types of the ground objects, and the true ground object category label chart is shown in fig. 2 (c).
The invention realizes semi-supervised ground object classification of the polarized SAR image, and the actual classification effect of the invention is verified by a classification experiment of 11 class ground objects in the image.
And 2, obtaining a polarization characteristic vector based on a polarization coherence matrix T in the polarization SAR image, reducing the speckle effect of the PolSAR image by using a BoxCar filter, and filtering outliers with the trace of the matrix smaller than 1 e-5.
(2a) The polarization coherence matrix of each pixel point of the polarized SAR image is represented by a matrix with the dimension of 3×3:
Figure BDA0002519375350000071
wherein T is ij ＝T ji * I+.j, superscript x denotes complex conjugate.
(2b) The polarization coherence matrix contains all polarization information of the polarized SAR data, has the capability of expressing the characteristics of the polarized SAR data, but noise exists in the original data, and uses the Box Car filter with the window size of 7 multiplied by 7 to reduce the speckle effect, and uses the space information to realize the polarization coherence matrix X of the pixel points of the ith row and the jth column in the image i,j The union of the plurality of pixel-polarized coherence matrices, denoted as their surrounding neighborhood, can be expressed as:
Figure BDA0002519375350000072
(2c) Further screening out invalid data, e.g. not meeting symmetry positive definite or trace less than 10 -5 Is a data of (a) a data of (b).
And 3, randomly selecting a training sample set. And randomly selecting 5% of data from each class of polarized SAR images to be classified as a training set, and the rest as a test sample set.
Step 4, constructing an initial dictionary
Figure BDA0002519375350000073
For each class, according to the training sample number of n, respectively performing k-means clustering,
(4a) K sample points are randomly selected initially to be used as a clustering center A k Wherein
Figure BDA0002519375350000074
(4b) Calculating the KL divergence distance between each sample and the cluster center, and classifying the KL divergence distance as the cluster center A closest to the sample k Sample X 1 ，X 2 The KL divergence of (2) is defined as:
Figure BDA0002519375350000075
where tr (X) denotes the trace of the matrix and size (X, 1) denotes the number of rows of the matrix.
(4c) For each cluster center A i I=1, 2,..k, calculating the mean of the samples attributed to this cluster center, defining the sample closest to the mean as the new cluster center a i new ；
(4d) If the number of the changed cluster centers is not more than 10 compared with the cluster centers obtained last time, ending iteration;otherwise use new cluster center A i new Substitute for the original clustering center A i And returning to step (4 b);
stacking the cluster centers obtained by each category to obtain an initial dictionary
Figure BDA0002519375350000081
Where m is the sum of the number of cluster centers for all categories.
Step 5, calculating an initial sparse vector Z (0) 。
Computing an initial sparse vector Z by minimizing dictionary representations on Riemann space (0) The expression is as follows:
Figure BDA0002519375350000082
wherein the method comprises the steps of
Figure BDA0002519375350000083
For the Riemann geodesic distance, for two samples X, Y in Riemann space, the geodesic distance is calculated by:
Figure BDA0002519375350000084
and sparse vectors satisfy
Figure BDA0002519375350000085
When the fixed dictionary is the initial dictionary
Figure BDA0002519375350000086
When the objective function is derived, the method comprises the following steps:
Figure BDA0002519375350000087
wherein,,
Figure BDA0002519375350000088
and 6, training an SVM classifier.
The invention distinguishes through using the maximum interval criterion, introduces a discriminant function S (z, y) E R, which measures the correctness of the association between the encoding vector z and the class label y, and the optimization function can be described as:
Figure BDA0002519375350000089
Figure BDA00025193753500000810
ξ i ≥0,i＝1,...,n
wherein,,
Figure BDA00025193753500000811
y for each tag i Is z of (2) i To ensure a correct association of S (z i ,y i ) Greater than the error associated S (z i Y), wherein y+.y i . R (S) is a regularization term used to constrain the complexity of the function S. Following standard SVM derivation, a relaxation variable ζ is introduced i To constrain the cases that may violate. With reference to Support vector guided dictionary learning, by setting S (z i ,y i )＝y i (ω T z i +b),/>
Figure BDA00025193753500000812
And->
Figure BDA00025193753500000813
The discrimination term of the classification problem becomes:
Figure BDA0002519375350000091
and 7, updating the dictionary.
Constructing a geometric perception based discriminant dictionary, wherein the expression is as follows:
Figure BDA0002519375350000092
solving by using an alternate direction iteration method, and fixing the sparse vector Z and the classifier super-parameters omega and b, wherein the above method can be expressed as follows:
Figure BDA0002519375350000093
the invention solves the above problem by using the Riemann conjugate gradient method, and compared with other first-order methods (such as the steepest descent method and the trust zone method), the method is more stable and has higher speed. For the nonlinear function Θ (B i ),
Figure BDA0002519375350000094
The Riemann conjugate gradient method uses the following recursion at step k+1:
Figure BDA0002519375350000095
wherein,,
Figure BDA0002519375350000096
Figure BDA0002519375350000097
Figure BDA0002519375350000098
wherein grad represents the gradient function, map Φ A (B) Vector transfer defining two points, B εT for A P M，
Figure BDA0002519375350000099
And 8, updating the sparse vector.
Constructing a geometric perception based discriminant dictionary, wherein the expression is as follows:
Figure BDA00025193753500000910
solving by using alternate direction iteration method and fixing dictionary
Figure BDA00025193753500000911
And classifier super-parameters ω, b, the above formula can be expressed as:
Figure BDA00025193753500000912
the invention adopts the Riemann conjugate gradient method to derive the following formula:
Figure BDA0002519375350000101
wherein beta is j ＝y i c (ω c T z i +b c )-1，
Figure BDA0002519375350000102
Step 9, setting convergence conditions based on geometric perception discrimination dictionary learning alternate iteration to obtain an optimized dictionary
Figure BDA0002519375350000103
And corresponding classifier super-parameters omega, b.
(9a) Setting a geometric perception discriminating dictionary
Figure BDA0002519375350000104
Judging whether the iteration termination condition is satisfied:
Figure BDA0002519375350000105
where Tr is the trace of the matrix,
Figure BDA0002519375350000106
dictionary created for the current iteration->
Figure BDA0002519375350000107
Is selected from the group consisting of the (i) th element,
Figure BDA0002519375350000108
dictionary representing previous iteration generation>
Figure BDA0002519375350000109
And k is the iteration number, and ζ is the set value to represent convergence accuracy.
(9b) If the convergence condition is not satisfied and the iteration number is less than 50, returning to the step 6, and updating the generated dictionary with the current time
Figure BDA00025193753500001010
Substitute initial dictionary->
Figure BDA00025193753500001011
Current update generated sparse vector Z (k) Instead of the original sparse vector Z (0) Performing iterative loop steps 6-8 until convergence conditions are met to obtain an optimized dictionary ∈F>
Figure BDA00025193753500001012
The classifier super parameters omega and b are corresponding to the classifier super parameters;
and step 10, carrying out dictionary representation on the test sample, and importing the test sample into a classifier for classification.
The specific substeps are as follows: building a Riemann sparse dictionary model, wherein the expression is as follows:
Figure BDA00025193753500001013
fixed optimization dictionary
Figure BDA00025193753500001014
And 5, solving the above formula to obtain a dictionary-converted sparse vector Z of the test sample. And (3) inputting the obtained sparse vector Z into the classifier in the step (9) to obtain a polarized image classification result graph based on the geometric perception discrimination dictionary.
The effect of the present invention can be further illustrated by the following simulation experiment.
(1) Simulation conditions
Laboratory desktop parameters: CPU is Inter (R) Core (TM) i7-6800k, main frequency is 3.40GHz, internal memory is 4G, operating system is Win7 bit system, experimental platform is Matlab2014b.
In a simulation experiment, the method of the invention is compared with the existing latest Wishare-ML, LE-NDR, ND-KSVD and RSC-SVM methods on Flevoland 1989; wherein:
the corresponding reference to Wishare-ML is "L.J.Du and J.S. Lee," Polarimetric sar image classification based on target decomposition theorem and complex Wishart distribution, "in IGARSS'96.1996International Geoscience and Remote Sensing Symposium,vol.1,May 1996,pp.439-4471 vol.1"
Corresponding references to LE-NDR are "H.Liu, Y.Wang, S.Yang, S.Wang, J.Feng, and L.Jiao," Large polarimetric sar data semi-supervised classification with spatial-anchor graph, "IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing, vol.9, no.4, pp.1439-1458,2016"
The corresponding reference to ND-KSVD is "W.Xie, L.Jiao, and J.Zhao," Polsar image classification via d-KSVD and act-domain features extraction, "IEEE Geoscience & Remote Sensing Letters, vol.13, no.2, pp.227-231,2016.".
The corresponding references for RSC-SVM are "N.Zhong, T.Yan, W.Yang and g.s.xia," A supervised classification approach for polsar images based on covariance matrix sparse coding, "in IEEE International Conference on Signal Processing,2017, pp.213-216.
(2) Emulation content
Experiment comparison of the method of the present invention with the above-described most recently method at present, using classification accuracy and tag prediction consistency as evaluation criteria on the fleveland 1989 dataset, fig. 2 and table 1 are comparison results. Wherein fig. 2 (a) is a visual image of the raw data subjected to Pauli decomposition; FIG. 2 (b) is an example of the colors to which each category corresponds; FIG. 2 (c) is the result of a real tag; FIG. 2 (d) shows the result of Wishare-ML; FIG. 2 (e) shows the results obtained for LE-NDR; FIG. 2 (f) shows the result of ND-KSVD; FIG. 2 (g) shows the result of RSC-SVM; FIG. 2 (h) shows the results obtained in the present invention.
Fig. 2 shows the visual classification results of all algorithms on the fleveland image, and it can be seen that both Wishart-ML and K-SVD have significant classification errors. For example, in FIG. 2 (d) of the Wishare-ML classification, middle and bottom grown wheat is mistaken for canola, and the bare land on the left is also classified as water. For FIG. 2 (f), classified by ND-KSVD, almost all grasses were misclassified. LE-NDR also misidentifies peas when alfalfa is planted in the bottom of peas, like the Wishare-ML method, while ND-KSVD misidentifies it as wheat. The RSC-SVM and the proposed method are generally correct in most places, whereas the proposed method achieves a high degree of accuracy in almost all types of earth coverage. However, there are many erroneous classification point distributions randomly among the correct blocks, which can be easily modified by morphological opening and closing operations.
As can be seen from Table 1, the process of the present invention gave the highest AC and Kappa values among all five processes. Compared to RSC-SVM, which also uses Riemann sparse coding, the proposed method exceeds 3.0% in AC and 4.1% in Kappa. For ND-KSVD using sparse coding in Euclidean space, AC and Kappa were improved by 13.5% and 16.0%, respectively.
TABLE 1
Figure BDA0002519375350000121
In summary, the invention provides a method for classifying a polarized SAR supervised image by utilizing geometric perception and discriminant dictionary learning, which mainly solves the problems that HPD high-dimensional matrix data is difficult to classify in Euclidean space and common discriminant dictionary is difficult to convert high-dimensional data. The main implementation steps are as follows: 1) And obtaining a filtered polarization coherence matrix. 2) An initial dictionary is constructed. 3) A sparse vector is calculated for each sample. 4) And training an SVM classifier. 5) Updating the discriminant dictionary, and tuning to the step 3) if the dictionary updating amplitude is larger than the threshold value. 6) And calculating sparse vectors of each test sample through dictionary representation, and obtaining a final classification result by using an SVM classifier. The method converts the three-dimensional polarization characteristic tensor into sparse representation under the condition that the three-dimensional polarization characteristic tensor is not required to be destroyed, the robustness of the model is enhanced by the alternative iterative updating of the discriminant dictionary, the classification precision is effectively improved, and the method is suitable for classifying the ground features of the polarized SAR image.

Claims (4)

1. The polarized SAR image supervision and classification method based on geometric perception discrimination dictionary learning is characterized by comprising the following steps of:
1) Inputting a to-be-classified polarized synthetic aperture radar SAR image and a real ground object marker image thereof to obtain a polarized coherent matrix T;
2) Filtering the extracted coherent matrix to remove speckle noise and obtain a filtered coherent matrix; each category selects 5% of data as training samples, and the rest as test samples;
3) Respectively using k-means clustering to obtain a plurality of clustering centers for each type of samples, and forming an initial dictionary by taking the clustering centers as dictionary elements;
4) Calculating sparse vectors of each training sample through dictionary representation;
the dictionary model used for calculating the sparse vector of each training sample through dictionary representation is as follows:
Figure FDA0004190387800000011
wherein,,
Figure FDA0004190387800000012
representation dictionary X i Is the ith coherent matrix X, z i X represents i In dictionary->
Figure FDA0004190387800000013
Sparse representation on->
Figure FDA0004190387800000014
Sparsity representing a sparse vector matrix Z, +.>
Figure FDA0004190387800000015
Representation dictionary->
Figure FDA0004190387800000016
L (Z, y, ω, b) represents the sparsity of the sparse vector Z for the label y classification penalty in an SVM classifier with a hyper-parameter ω, b, wherein:
Figure FDA0004190387800000017
wherein θ is a weight coefficient;
when dictionary
Figure FDA0004190387800000018
Fixed with the SVM classifier parameters omega, b, for a given data matrix X i ∈H d + The minimum value of z is expressed as the following sub-problem:
Figure FDA0004190387800000019
for z j Obtaining the partial derivative:
Figure FDA00041903878000000110
wherein,,
Figure FDA00041903878000000111
5) Training an SVM classifier according to the sparse vector and the label information, wherein the specific method comprises the following steps: when the sparse vector z and the SVM classifier parameters omega and b are fixed, the dictionary is formed
Figure FDA00041903878000000112
Expressed as the following non-convex optimization problem:
Figure FDA0004190387800000021
the method adopts the Riemann conjugate gradient method to derive the gradient, and the gradient is used for nonlinear function theta (B i ),
Figure FDA0004190387800000022
Dictionary element of step k+1->
Figure FDA0004190387800000023
Expressed as:
Figure FDA0004190387800000024
wherein, gamma k Is the step size found by an efficient linear search method, while the direction of descent ζ (k) The definition is as follows:
Figure FDA0004190387800000025
wherein grad represents the gradient function, map Φ X (Y) defines a vector transfer of two points, for two points X, Y ε T on the Riemann manifold P M；
Figure FDA0004190387800000026
Figure FDA0004190387800000027
Figure FDA0004190387800000028
Where Θ (B) is the euclidean gradient of Θ (B), and the dictionary element of the j th generation is defined as:
Figure FDA0004190387800000029
order the
Figure FDA00041903878000000210
Derivative->
Figure FDA00041903878000000211
Is approximated as +.>
Figure FDA00041903878000000212
6) Updating the dictionary by using the sparse vector and the classifier;
7) Returning to the step 4) if the dictionary updating amplitude is larger than 1e-5, otherwise, directly executing the step 8);
8) And calculating sparse vectors of each test sample through dictionary representation, and obtaining a final classification result by using an SVM classifier.
2. The polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning as set forth in claim 1, characterized in that in step 2), the filtering is performed on the extracted coherent matrix to remove speckle noise, and the obtained filtered coherent matrix is based on the polarized coherent matrix T obtained in step 1), and a BoxCar filtering method with a window size of 7 x 7 is adopted to filter the extracted coherent matrix, so as to reduce speckle effect of the PolSAR image, remove speckle noise, and filter outliers with a trace less than 1 e-5.
3. The polarized SAR image supervised classification method based on geometric sense discriminating dictionary learning of claim 1, wherein in step 3), the specific steps of constructing the initial dictionary are:
for each class, according to the training sample number of each class as n, respectively performing k-means clustering, wherein the distance measurement between each sample and the clustering center is defined as KL divergence distance, and sample X 1 ，X 2 The KL divergence of (2) is defined as:
Figure FDA0004190387800000031
where tr (X) denotes the trace of the matrix and size (X, 1) denotes the number of rows of the matrix;
and (3) carrying out loop iteration until the cluster center does not move any more, obtaining a cluster center of one category, and taking the cluster centers obtained by calculating all the categories as dictionary elements to form an initial dictionary.
4. The method for supervised classification of polarized SAR images based on geometric sense discriminating dictionary learning according to claim 1 wherein in step 8), said calculating sparse vectors of each test sample represented by a dictionary uses a dictionary model of:
Figure FDA0004190387800000032
the class to which the sparse vector z belongs is calculated using an SVM classifier:
Figure FDA0004190387800000033
and predicting the category of each pixel point in turn, namely realizing the classification of the planning image.
CN202010486706.9A 2020-06-01 2020-06-01 Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning Active CN111680731B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010486706.9A CN111680731B (en) 2020-06-01 2020-06-01 Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010486706.9A CN111680731B (en) 2020-06-01 2020-06-01 Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning

Publications (2)

Publication Number Publication Date
CN111680731A CN111680731A (en) 2020-09-18
CN111680731B true CN111680731B (en) 2023-06-27

Family

ID=72453022

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010486706.9A Active CN111680731B (en) 2020-06-01 2020-06-01 Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning

Country Status (1)

Country Link
CN (1) CN111680731B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113381827B (en) * 2021-06-09 2022-04-26 厦门大学 Electromagnetic signal spectrum sensing method based on deep clustering network
CN113643284B (en) * 2021-09-09 2023-08-15 西南交通大学 Polarized synthetic aperture radar image ship detection method based on convolutional neural network

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103714353A (en) * 2014-01-09 2014-04-09 西安电子科技大学 Polarization SAR image classification method based on vision prior model
CN104123555A (en) * 2014-02-24 2014-10-29 西安电子科技大学 Super-pixel polarimetric SAR land feature classification method based on sparse representation
CN108446716A (en) * 2018-02-07 2018-08-24 武汉大学 Based on FCN the PolSAR image classification methods merged are indicated with sparse-low-rank subspace

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110717354B (en) * 2018-07-11 2023-05-12 哈尔滨工业大学 Super-pixel classification method based on semi-supervised K-SVD and multi-scale sparse representation

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103714353A (en) * 2014-01-09 2014-04-09 西安电子科技大学 Polarization SAR image classification method based on vision prior model
CN104123555A (en) * 2014-02-24 2014-10-29 西安电子科技大学 Super-pixel polarimetric SAR land feature classification method based on sparse representation
CN108446716A (en) * 2018-02-07 2018-08-24 武汉大学 Based on FCN the PolSAR image classification methods merged are indicated with sparse-low-rank subspace

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
采用联合域字典稀疏表示的极化SAR图像分类;刘璐;《华中科技大学学报》;20160223;全文 *

Also Published As

Publication number Publication date
CN111680731A (en) 2020-09-18

Similar Documents

Publication Publication Date Title
Tang et al. Unsupervised feature selection via latent representation learning and manifold regularization
Ghasedi Dizaji et al. Deep clustering via joint convolutional autoencoder embedding and relative entropy minimization
Mrabah et al. Deep clustering with a dynamic autoencoder: From reconstruction towards centroids construction
Farag et al. A unified framework for map estimation in remote sensing image segmentation
Okori et al. Machine learning classification technique for famine prediction
CN109657611B (en) Adaptive image regularization non-negative matrix decomposition method for face recognition
CN111680731B (en) Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning
Ding et al. Unsupervised feature selection via adaptive hypergraph regularized latent representation learning
Dehshibi et al. Cubic norm and kernel-based bi-directional PCA: toward age-aware facial kinship verification
Liu et al. Terrain classification based on spatial multi-attribute graph using polarimetric SAR data
Ma et al. Spectral correlation-based diverse band selection for hyperspectral image classification
Masuyama et al. Adaptive resonance theory-based topological clustering with a divisive hierarchical structure capable of continual learning
Hsieh et al. Adaptive structural co-regularization for unsupervised multi-view feature selection
Liu et al. Adaboost-based SVDD for anomaly detection with dictionary learning
Zeng et al. Autoencoders for discovering manifold dimension and coordinates in data from complex dynamical systems
Ngadi et al. Uniformed two local binary pattern combined with neighboring support vector classifier for classification
Zhao et al. Multiple kernel fuzzy clustering with unsupervised random forests kernel and matrix-induced regularization
Ghasemi et al. AFDL: a new adaptive fuzzy dictionary learning for medical image classification
Karlik The positive effects of fuzzy c-means clustering on supervised learning classifiers
Yi et al. Inner product regularized nonnegative self representation for image classification and clustering
Yang et al. Robust landmark graph-based clustering for high-dimensional data
Bhatt et al. Coarse Clustering and Classification of Images with CNN Features for Participatory Sensing in Agriculture.
CN114419382A (en) Method and system for embedding picture of unsupervised multi-view image
Lu et al. Flower classification based on single petal image and machine learning methods
Kang et al. Kernel optimisation for KPCA based on Gaussianity estimation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant