CN113269024B - Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics - Google Patents

Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics Download PDF

Info

Publication number
CN113269024B
CN113269024B CN202110348030.1A CN202110348030A CN113269024B CN 113269024 B CN113269024 B CN 113269024B CN 202110348030 A CN202110348030 A CN 202110348030A CN 113269024 B CN113269024 B CN 113269024B
Authority
CN
China
Prior art keywords
sample
domain
target domain
samples
clustering
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
CN202110348030.1A
Other languages
Chinese (zh)
Other versions
CN113269024A (en
Inventor
汪长城
李倩
沈鹏
高晗
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.)
Central South University
Original Assignee
Central South 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 Central South University filed Critical Central South University
Priority to CN202110348030.1A priority Critical patent/CN113269024B/en
Publication of CN113269024A publication Critical patent/CN113269024A/en
Application granted granted Critical
Publication of CN113269024B publication Critical patent/CN113269024B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a non-supervision domain self-adaptive network polarization SAR terrain classification method and equipment considering polarization statistical characteristics, wherein the method comprises the following steps: selecting polarized SAR data with and without labels to calculate and generate covariance matrixes which are respectively used as a source domain data set and a target domain data set; initializing clustering centers of a source domain and a target domain by using a source domain sample; inputting the data of the source domain and the data of the target domain into the rewinding product neural networks respectively corresponding to the source domain and the target domain to obtain reconstruction characteristics; determining a pseudo label for the target domain sample by calculating the distance between the reconstruction characteristics and the clustering center, and updating the clustering center of the target domain; repeating clustering iteration until the clustering iteration converges or the maximum clustering iteration times is reached; and iteratively updating the parameters of the two rewinding product neural networks by minimizing the target function until the network parameters are converged, wherein each sample class of the target domain data set is the final ground object class. The method can efficiently classify the ground features of the polarized SAR data.

Description

Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics
Technical Field
The invention belongs to the technical field of remote sensing image processing, and particularly relates to an unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics.
Background
The existing research aiming at the combination of the polarized SAR data and the neural network is not deep. The polarized SAR has unique imaging characteristics and abundant ground object structure information. However, the existing polarized SAR data processing method is the same as the optical data processing method, and the characteristics of the polarized SAR data cannot be fully exerted. And the existing polarized SAR has fewer samples and is difficult to manufacture, so that the network model is difficult to train by using a large sample. In recent years, some scholars have proposed a complex network and extracted phase polarization information of polarized SAR data. Although the application result of the polarized SAR data is not greatly improved, the data characteristics of the polarized SAR itself are combined with deep learning so as to build a network which is more in line with the characteristics of the polarized SAR data.
Disclosure of Invention
The invention discloses a method and equipment for classifying terrain features of an unsupervised domain self-adaptive network polarization SAR (synthetic aperture radar) in consideration of polarization statistical characteristics, which have a good effect on the aspect of extracting the features of polarization SAR data and can efficiently perform terrain clustering.
In order to achieve the technical purpose, the invention adopts the following technical scheme:
an unsupervised domain adaptive network polarization SAR terrain classification method considering polarization statistical characteristics comprises the following steps:
step 1, selecting polarized SAR data with sample labels to calculate and generate a covariance matrix C3I sample as source domain dataset labeled c
Figure GDA0003537094910000011
And the sample label of the source domain dataset includes Num class, c ═ 1, 2, …, Num; calculating the polarized SAR data without the sample label to generate a covariance matrix C3As the jth sample in the target domain data set
Figure GDA0003537094910000012
Step 2, calculating the clustering center of various labels c of the source domain data set
Figure GDA0003537094910000013
Clustering centers as source domain dataset initiations
Figure GDA0003537094910000014
Initializing a cluster center of a target domain data set using a cluster center of a source domain data set to
Figure GDA0003537094910000015
Step 3, inputting the source domain data set and the target domain data set to the respective corresponding convolution neural networks respectively; a softmax classifier is arranged at the output end of the rewinding product neural network corresponding to the source domain, and a sample classification result of the source domain data set is output; the method comprises the steps that a multiple convolution neural network corresponding to a target domain outputs reconstruction characteristics corresponding to samples in a target domain data set, and the reconstruction characteristics of all the samples form a reconstruction characteristic set;
step 4, for each sample in the target domain data set: all calculate their corresponding reconstruction features
Figure GDA0003537094910000021
With each cluster center
Figure GDA0003537094910000022
Determining the category of the sample by selecting the clustering center corresponding to the minimum distance among the Wishart distances; the reconstruction characteristic of the determined category c is
Figure GDA0003537094910000023
Step 5, aiming at each category in the target domain data set, calculating and updating the clustering center of the category in the target domain data set according to the reconstruction characteristics corresponding to all samples of the category in the target domain data set;
step 6, repeating the steps 4 and 5 to perform clustering iteration until the clustering iteration is converged or the maximum clustering iteration frequency is reached;
step 7, updating the parameters of the source domain rewinding and product neural network and the target domain rewinding and product neural network through the minimized target function;
and 8, returning to execute the step 3 until the parameters of the double convolution neural networks corresponding to the source domain and the target domain are converged, wherein the category of each sample of the target domain data set is the final ground object classification category.
In a more preferred embodiment, the objective function in step 7 is:
f＝losscross entropy+βDCWD
in the formula, losscross entropyCross-entropy loss for source-domain classification, DCWDFor the contrast Wishart difference between the data distributions of all the categories of the source domain and the target domain, β is a regular term coefficient and has:
Figure GDA0003537094910000024
Figure GDA0003537094910000025
Figure GDA0003537094910000026
Figure GDA0003537094910000027
Figure GDA0003537094910000031
Figure GDA0003537094910000032
wherein, mucc(y, y') is a discriminant function, CDccRepresenting the distance, CD, between data points within the same class of source and target domainscc'denotes the distance between different classes c and c' of samples of the source and target domains, and has:
CDcc＝WD，
Figure GDA0003537094910000033
wherein WD is the Wishart distance,
Figure GDA0003537094910000034
representing the source domain rewinding product neural network with the parameter theta according to the input sample
Figure GDA0003537094910000035
The predicted output label is
Figure GDA0003537094910000036
The probability value is obtained by a softmax classifier;
Figure GDA0003537094910000037
representing source domain data samples
Figure GDA0003537094910000038
The real label of (a) is,
Figure GDA0003537094910000039
representing target domain data samples
Figure GDA00035370949100000310
Clustering the obtained labels;
nsrepresenting the number of samples, n, in the source domain datasettRepresenting the number of samples in the target domain dataset;
Figure GDA00035370949100000311
a contrast Wishart difference between the same category data points representing the source domain and the target domain,
Figure GDA00035370949100000312
representing contrast Wishart differences between different categories of data points of the source domain and the target domain;
e1the sum of the following distances for all classes is indicated: the distance of the samples of the same category of the target domain and the source domain on the reconstruction feature space; e.g. of the type2The sum of the following distances for all classes is indicated: the distances of different types of samples of the target domain and the source domain on a reconstruction feature space;
Figure GDA00035370949100000313
representing source domain rewinding product neural network basis sample
Figure GDA00035370949100000314
The output of the reconstruction matrix is then processed,
Figure GDA00035370949100000315
presentation pair
Figure GDA00035370949100000316
Performing inversion operation;
Figure GDA00035370949100000317
representing target domain convolution neural network basis samples
Figure GDA00035370949100000318
The output of the reconstruction matrix is then processed,
Figure GDA00035370949100000319
presentation pair
Figure GDA00035370949100000320
Performing inversion operation;
Figure GDA00035370949100000321
represents ntClustering label corresponding to each target domain sample
Figure GDA00035370949100000322
Z denotes a general term of a reconstruction matrix of source domain and target domain samples.
In a more preferred embodiment, before minimizing the objective function in step 7, preprocessing the sample subsets and samples of the respective types participating in the target domain data set, including:
(1) determining samples involved in the calculation of the objective function: for each sample in the various types of sample subsets, if the wishart distance between the sample and the self type clustering center exceeds a given threshold tv belonging to [0, 1], the sample does not participate in the objective function calculation of the network parameter updating process;
(2) determining the sample types participating in the calculation of the objective function: and calculating samples based on the determined participation objective function, counting the number of samples included in each type in the target domain data set, and if the number of samples included in a certain sample type is lower than a given threshold value delta, not participating in the objective function calculation of the network parameter updating process.
In a more preferred technical scheme, the calculation generation method of the covariance matrix comprises the following steps:
Figure GDA0003537094910000041
Figure GDA0003537094910000042
in the formula (I), the compound is shown in the specification,
Figure GDA0003537094910000043
for the scattering matrix corresponding to the polarized SAR data, denotes the conjugate transpose, SHHIncluding the echo power, S, of the HH polarization channelHVInvolving the echo power, S, of the HV polarisation channelVHIncluding VH polarization channel echo power, SVVIncluding the echo power number of the VV polarization channel.
In a more preferable technical scheme, the clustering iterative convergence criterion in the step 6 includes intra-class wishart dispersion and inter-class wishart distance; intra-class wishirt dispersion requirement: all samples of each type to the cluster center of the type
Figure GDA0003537094910000044
Is less than a given value
Figure GDA0003537094910000045
The wishart distance requirement between classes: the wishart distance between any two clustering centers is greater than a given value sigma, otherwise, the clustering is not converged, and iterative clustering needs to be continued; two of the cluster centers
Figure GDA0003537094910000046
And
Figure GDA0003537094910000047
the wishart distance between (1) is calculated as:
Figure GDA0003537094910000048
in the formula, | | represents a determinant of the matrix.
In a more preferred technical scheme, the calculation method of the clustering center is as follows:
Figure GDA0003537094910000049
Figure GDA00035370949100000410
in the formula (I), the compound is shown in the specification,
Figure GDA00035370949100000411
indicating the number of samples in the source domain dataset labeled c,
Figure GDA00035370949100000412
indicating the number of samples in the target domain dataset labeled c.
An unsupervised domain adaptive network polarization SAR terrain classification device considering polarization statistical characteristics comprises a processor and a memory; wherein: the memory is to store computer instructions; the processor is configured to execute the computer instructions stored in the memory, and specifically, to execute the unsupervised domain adaptive network polarimetric SAR terrain classification method considering the polarimetric statistical characteristics according to any one of the above technical solutions.
Advantageous effects
The method aims at the problems of incomplete utilization of information of the polarized SAR and combination of deep learning algorithm and less and difficult-to-manufacture polarized SAR data samples, and the polarized covariance matrix C of the method3The space structure information and the polarization information of the polarized SAR data are extracted by utilizing the complex neural network, the statistical characteristic of the polarized SAR is introduced into the constructed network, and the classification result diagram of the polarized SAR is obtained, so that the ground feature classification of the polarized SAR data is efficiently carried out, the utilization of the polarized SAR data information is greatly promoted, the application of deep learning in polarized SAR data processing is further promoted, and a new thought is provided for the processing of the polarized SAR data. In addition, the method solves the problem of difficulty in manufacturing the polarized SAR data sample to a certain extent.
Drawings
Fig. 1 is a general structural diagram of the method according to the embodiment of the present invention, which mainly includes a CCNN model and wishart distance iterative clustering process.
Fig. 2 is a diagram of a CCNN network structure according to an embodiment of the present invention.
Fig. 3 is a diagram of an iterative clustering process of a polar SAR complex Wishart according to an embodiment of the present invention. And if the initial process does not reach the clustering iteration standard, performing iterative clustering on the dotted arrow part.
Fig. 4 is a network model training process based on the domain adaptation method. The solid arrow is the network forward calculation process, and the dashed arrow is the network error back propagation process.
Detailed Description
The following describes embodiments of the present invention in detail, which are developed based on the technical solutions of the present invention, and give detailed implementation manners and specific operation procedures to further explain the technical solutions of the present invention.
The embodiment provides an unsupervised domain adaptive network polarimetric SAR terrain classification method considering polarimetric statistical characteristics, as shown in fig. 1, including:
step 1, selecting polarized SAR data with sample labels to calculate and generate a covariance matrix C3As a source domain data set DsSample i with middle label c
Figure GDA0003537094910000061
And the sample label of the source domain dataset includes Num class, c ═ 1, 2, …, Num; calculating the polarized SAR data without the sample label to generate a covariance matrix C3As a target domain data set DtSample j of (2)
Figure GDA0003537094910000062
The calculation generation mode of the covariance is as follows:
firstly, a scattering matrix S measured by a polarized SAR system is expressed
Figure GDA0003537094910000063
Then, according to the principle S that the reciprocity symmetry is satisfied under the condition of back scattering of the single station systemHV＝SVHScattering moment of polarizationThe array S is simplified as follows:
Figure GDA0003537094910000064
finally, a target vector is obtained through the scattering matrix S, and the covariance matrix C is obtained through the outer product of the target vector and the conjugate transpose vector of the target vector3：
Figure GDA0003537094910000065
Wherein denotes the conjugate transpose, SHHIncluding the echo power, S, of the HH polarization channelHVInvolving the echo power, S, of the HV polarisation channelVHIncluding VH polarization channel echo power, SVVIncluding the echo power number of the VV polarization channel.
The present invention uses covariance matrices as features of polarized SAR data for ground feature classification, which has essential advantages: the covariance matrix of the polarized SAR data follows a multivariate complex wishart distribution, while the covariance matrix of kth view of the n-view polarized covariance matrix Z
Figure GDA0003537094910000066
Wherein the vectoru(k) Is the kth single-view sample. The matrix a obeys the complex wishart distribution, so the complex wishart probability density function can be written as:
Figure GDA0003537094910000067
wherein the parameter q represents a vectoru(k) Of (c) is calculated. For reciprocity under single-station polarimetric SAR observations, q is 3.
Step 2, calculating the clustering center of various labels c of the source domain data set
Figure GDA0003537094910000071
Clustering centers as source domain dataset initiations
Figure GDA0003537094910000072
Initializing a cluster center of a target domain data set using a cluster center of a source domain data set to
Figure GDA0003537094910000073
The method for calculating the source domain clustering center comprises the following steps:
Figure GDA0003537094910000074
in the formula (I), the compound is shown in the specification,
Figure GDA0003537094910000075
representing a source domain data set DsNumber of samples with a middle label of c.
Step 3, using the source domain data set DsAnd a target domain data set DtRespectively inputting the data to the rewinding volume neural networks corresponding to the data; a softmax classifier is arranged at the output end of the rewinding product neural network corresponding to the source domain, and a sample classification result of the source domain data set is output; the method comprises the steps that a multiple convolution neural network corresponding to a target domain outputs reconstruction characteristics corresponding to samples in a target domain data set, and the reconstruction characteristics of all the samples form a reconstruction characteristic set;
a complex convolution neural network, CCNN for short, as shown in fig. 2, obtains the relationship between a training pixel and its surrounding pixels through operations such as complex convolution, complex batch normalization, and a complex activation layer, and can obtain a better effect than single sample point input. And the phase information is unique information of the polarized SAR image and is an important component for explaining the polarized SAR data. The complex input of the polarization SAR is processed by operations such as complex convolution and the like, so that the complex value information of the polarization SAR data can be better utilized to play a role in fully utilizing the polarization SAR information. And a global average pooling layer, namely a GAP layer, is used in the last layer, and Wishart distance is introduced and added into the objective function to realize clustering between classes, so that the data distribution of the same class is respectively close to and aligned, and simultaneously, the model efficiency is improved due to the occurrence of an initial clustering center.
Source domain data set DsAnd a target domain data set DtThe same as in (1)Covariance matrix C of polarimetric SAR data3Obtainable from step 1, comprising C11，......，Cij，......，C33，CijThe ij-th item features among the 9-dimensional features in the covariance matrix are represented, and these features are feature maps of 11 × 11 window size. Source domain data set DsThe reconstructed matrix Z of the samples in (1) is obtained by utilizing a GAP layer of CCNNs，′Matrix and use of Zs，′Inputting the matrix into a softmax classifier to obtain a prediction type
Figure GDA0003537094910000076
For using the cross-loss function loss at step 7cross entropyEvaluating the error of the label actually carried by the sample, and enabling the reconstruction Z to be realized by back propagation of the errors，′Matrix and source domain raw input C3The distribution of the matrices is getting closer.
The Softmax classifier is based on input features, here Zs，′A matrix to model the discrete model output for prediction. The classifier finally obtains the classification probability of classifying the sample into Num classes, judges that the sum of the class output probabilities is equal to 1, and takes the class corresponding to the maximum class classification probability as the final classification result. The calculation formula of the ith category output probability is as follows:
Figure GDA0003537094910000081
wherein, by Zs，′The matrix gets Out { Out ═ Out through the Dense layerunit1，......，Outuniti，......，OutunitNum}. Out at this time indicates that the plural network models obtain outputs (there is no relationship between the outputs, and it is difficult to compare and judge the final classification category), so the softmax classifier can be used to characterize the relative probability between different categories.
Step 4, for each sample in the target domain data set: all calculate their corresponding reconstruction features
Figure GDA0003537094910000082
With each cluster center
Figure GDA0003537094910000083
Determining the category of the sample by selecting the clustering center corresponding to the minimum distance among the Wishart distances; the reconstruction characteristic of the determined category c is
Figure GDA0003537094910000084
c＝1，2，…，Num；
The Wisharp distance is introduced according to a wissharp probability density function and is represented as WD, and the calculation method comprises the following steps:
Figure GDA0003537094910000085
the method is used for the category measurement standard in the ground feature classification. Wherein Z represents a covariance matrix Z obtained by reconstructing a target domain sample through a CCNN model, subscript l represents the ith sample of a training sample, RiThen a certain class of category cluster center is indicated.
Step 5, aiming at each category in the target domain data set, calculating and updating the clustering center of the category of the target domain data set according to the following formula according to the reconstruction characteristics corresponding to all samples of the category in the target domain data set:
Figure GDA0003537094910000086
in the formula (I), the compound is shown in the specification,
Figure GDA0003537094910000087
indicating the number of samples in the target domain dataset labeled c. The premise assumption of updating the cluster center through clustering is that the distribution of different types of data is more dispersed
And 6, repeating the steps 4 and 5 to perform clustering iteration as shown in the figure 3 until the clustering iteration is converged or the maximum clustering iteration number is reached. Specifically, the criterion of clustering iterative convergence includes intra-class wishart dispersion and inter-class wishart distance.
Intra-class wishirt dispersion requirement: all samples of each type to the analogous polyClass center
Figure GDA0003537094910000088
Is less than a given value
Figure GDA0003537094910000089
The wishart distance requirement between classes: the wishart distance between any two cluster centers is greater than a given value sigma, and the two cluster centers
Figure GDA00035370949100000810
And
Figure GDA00035370949100000811
the wishart distance between (1) is calculated as:
Figure GDA0003537094910000091
in the formula, | | represents a determinant of the matrix. DijThe larger the size, the higher the discrimination between the two categories. Such as DijIf the value is less than a given value sigma, the clustering is not converged, and iterative clustering needs to be continued.
And 7, updating the parameters of the source domain rewinding and product neural network and the target domain rewinding and product neural network by minimizing the target function, as shown in fig. 4.
And after each clustering iteration is terminated, training a convolutional neural network model based on domain adaptation. And calculating an objective function, minimizing the objective function f to reversely propagate and update parameters of the two complex convolution neural networks until the training iteration of the whole network reaches the maximum training times or the iterative training converges, and finishing the training of the whole model.
Before minimizing the objective function, preprocessing the sample subsets and samples of various types participating in the target domain data set, including:
(1) determining samples involved in the calculation of the objective function: for each sample in the various types of sample subsets, if the wishart distance between the sample and the self type clustering center exceeds a given threshold tv belonging to [0, 1], the sample does not participate in the objective function calculation of the network parameter updating process;
(2) determining the sample types participating in the calculation of the objective function: and calculating samples based on the determined participation objective function, counting the number of samples included in each type in the target domain data set, and if the number of samples included in a certain sample type is lower than a given threshold value delta, not participating in the objective function calculation of the network parameter updating process.
Then, the preprocessed various types of sample subsets and samples can be used to minimize the objective function by calculating the Contrast Wishart Difference (CWD), which specifically includes:
assuming that there is some judgment function:
Figure GDA0003537094910000092
in defining the distance between the classes, the distance measurement is carried out by referring to the method proposed by Lee et al. From the dispersion proposed by Lee et al, the distance CD between data samples in the same class between the source domain and the target domain can be derivedccIs defined as: CD (compact disc)ccWD; distance CD between different classes c and c' of samples of source and target domainscc′Is defined as
Figure GDA0003537094910000093
。
The judgment function only calculates the difference value of the specified categories c and c'. Thus we can define together with step 4 our contrast Wishart differences:
Figure GDA0003537094910000101
wherein the content of the first and second substances,
Figure GDA0003537094910000102
finally, the contrast Wishart differences between the data distributions for all categories of source domain, target domain can be written as:
Figure GDA0003537094910000103
the first half of the formula shows that for all classes, the sample distribution of the target domain is to be respectively drawn up in a certain layer of feature space with the sample distribution of the same class as that of the source domain. The second half of the formula indicates that for all classes, the sample distribution of a certain class c' of the target domain is to be separated from the sample distribution of all different classes c of the source domain, such that the inter-class differences are maximized.
The objective function of this embodiment based on the above contrast Wishart difference is:
Min f＝losscross entropy+βDCWD
wherein the content of the first and second substances,
Figure GDA0003537094910000104
minimizing the target function realizes updating the network parameters, including parameter updating of the CCNN of the source domain and the CCNN of the target domain. Because minimizing the objective function is an optimization process, the components of the objective function are the difference between the predicted value of the source domain sample passing through the softmax classifier and its corresponding true value and the contrast Wishart difference between the target domain sample and the source domain sample. Minimizing a target function, wherein the purpose 1 is to enable the classification of samples from a model to a source domain to be more and more accurate, a reconstruction matrix of the model can be well used as a class classification effective feature, and the effective feature can be helpful for clustering the reconstruction feature for clustering of the target domain in the whole model training process when participating in contrast wishart difference calculation; the purpose 2 is that the same category data distance between the target domain and the source domain is closer and closer, and the distance between different categories is farther and farther, so that a more accurate clustering result of the target domain is obtained, the better the clustering result is, in the process of minimizing the error through the back propagation error, the feature extracted by the network corresponding to the target domain (namely, the target domain sample reconstruction matrix) is also an effective feature for the clustering result, and the effective classification feature can enable the clustering result to be more accurate; minimizing the objective function makes the model converge and the final result is obtained. Updating the parameter θ is mainly a back propagation mechanism of model training, which mainly uses a loss function f to derive the parameter θ, and if a gradient descent method is used, the parameter can be updated by using the following formula:
Figure GDA0003537094910000111
where i represents the ith layer of the network model.
And 8, returning to execute the step 3 until the parameters of the rewinding product neural network corresponding to the source domain and the target domain are converged, terminating the training of the whole model, and determining the category of each sample of the target domain data set as the final ground object classification category.
Example 2
The embodiment provides an unsupervised domain adaptive network polarization SAR terrain classification device considering polarization statistical characteristics, which comprises a processor and a memory; wherein: the memory is to store computer instructions; the processor is configured to execute the computer instructions stored in the memory, and in particular, to perform the method of embodiment 1.
The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the claims of the present application.

Claims (7)

1. An unsupervised domain adaptive network polarization SAR terrain classification method considering polarization statistical characteristics is characterized by comprising the following steps:
step 1, selecting polarized SAR data with sample labels to calculate and generate a covariance matrix C3I sample as source domain dataset labeled c
Figure FDA0003537094900000011
And the sample label of the source domain dataset includes Num class, c ═ 1, 2, …, Num; calculating the polarized SAR data without the sample label to generate a covariance matrix C3As the jth sample in the target domain data set
Figure FDA0003537094900000012
Step 2, calculating the clustering center of various labels c of the source domain data set
Figure FDA0003537094900000013
Clustering centers as source domain dataset initiations
Figure FDA0003537094900000014
Initializing a cluster center of a target domain data set using a cluster center of a source domain data set to
Figure FDA0003537094900000015
Step 3, inputting the source domain data set and the target domain data set to the respective corresponding convolution neural networks respectively; a softmax classifier is arranged at the output end of the rewinding product neural network corresponding to the source domain, and a sample classification result of the source domain data set is output; the method comprises the steps that a multiple convolution neural network corresponding to a target domain outputs reconstruction characteristics corresponding to samples in a target domain data set, and the reconstruction characteristics of all the samples form a reconstruction characteristic set;
step 4, for each sample in the target domain data set: all calculate their corresponding reconstruction features
Figure FDA0003537094900000016
With each cluster center
Figure FDA0003537094900000017
Determining the category of the sample by selecting the clustering center corresponding to the minimum distance among the Wishart distances; the reconstruction characteristic of the determined category c is
Figure FDA0003537094900000018
Step 5, aiming at each category in the target domain data set, calculating and updating the clustering center of the category in the target domain data set according to the reconstruction characteristics corresponding to all samples of the category in the target domain data set;
step 6, repeating the steps 4 and 5 to perform clustering iteration until the clustering iteration is converged or the maximum clustering iteration frequency is reached;
step 7, updating the parameters of the source domain rewinding and product neural network and the target domain rewinding and product neural network through the minimized target function;
and 8, returning to execute the step 3 until the parameters of the double convolution neural networks corresponding to the source domain and the target domain are converged, wherein the category of each sample of the target domain data set is the final ground object classification category.
2. The method of claim 1, wherein the objective function of step 7 is:
f＝losscross entropy+βDCWD
in the formula, losscross entropyCross-entropy loss for source-domain classification, DCWDFor the contrast Wishart difference between the data distributions of all the categories of the source domain and the target domain, β is a regular term coefficient and has:
Figure FDA0003537094900000021
Figure FDA0003537094900000022
Figure FDA0003537094900000023
Figure FDA0003537094900000024
Figure FDA0003537094900000025
Figure FDA0003537094900000026
wherein, mucc(y, y') is a discriminant function, CDccRepresenting the distance, CD, between data points within the same class of source and target domainscc′Represents the distance between different classes c and c' of samples of the source domain and the target domain, and has:
CDcc＝WD，
Figure FDA0003537094900000027
wherein WD is the Wishart distance,
Figure FDA0003537094900000028
representing the source domain rewinding product neural network with the parameter theta according to the input sample
Figure FDA0003537094900000029
The predicted output label is
Figure FDA00035370949000000210
The probability value is obtained by a softmax classifier;
Figure FDA00035370949000000211
representing source domain data samples
Figure FDA00035370949000000212
The real label of (a) is,
Figure FDA00035370949000000213
representing target domain data samples
Figure FDA00035370949000000214
Clustering the obtained labels;
nsrepresenting the number of samples, n, in the source domain datasettRepresenting the number of samples in the target domain dataset;
Figure FDA00035370949000000215
a contrast Wishart difference between the same category data points representing the source domain and the target domain,
Figure FDA00035370949000000216
representing contrast Wishart differences between different categories of data points of the source domain and the target domain;
e1the sum of the following distances for all classes is indicated: the distance of the samples of the same category of the target domain and the source domain on the reconstruction feature space; e.g. of the type2The sum of the following distances for all classes is indicated: the distances of different types of samples of the target domain and the source domain on a reconstruction feature space;
Figure FDA0003537094900000031
representing source domain rewinding product neural network basis sample
Figure FDA0003537094900000032
The output of the reconstruction matrix is then processed,
Figure FDA0003537094900000033
presentation pair
Figure FDA0003537094900000034
Performing inversion operation;
Figure FDA0003537094900000035
representing target domain convolution neural network basis samples
Figure FDA0003537094900000036
The output of the reconstruction matrix is then processed,
Figure FDA0003537094900000037
presentation pair
Figure FDA0003537094900000038
Performing inversion operation;
Figure FDA0003537094900000039
represents ntClustering label corresponding to each target domain sample
Figure FDA00035370949000000310
Z denotes a general term of a reconstruction matrix of source domain and target domain samples.
3. The method of claim 1, wherein preprocessing each type of sample subset and sample participating in the target domain dataset before minimizing the objective function in step 7 comprises:
(1) determining samples involved in the calculation of the objective function: for each sample in the various types of sample subsets, if the wishart distance between the sample and the self type clustering center exceeds a given threshold tv belonging to [0, 1], the sample does not participate in the objective function calculation of the network parameter updating process;
(2) determining the sample types participating in the calculation of the objective function: and calculating samples based on the determined participation objective function, counting the number of samples included in each type in the target domain data set, and if the number of samples included in a certain sample type is lower than a given threshold value delta, not participating in the objective function calculation of the network parameter updating process.
4. The method of claim 1, wherein the covariance matrix is computed by:
Figure FDA00035370949000000311
Figure FDA00035370949000000312
in the formula (I), the compound is shown in the specification,
Figure FDA00035370949000000313
for the scattering matrix corresponding to the polarized SAR data, denotes the conjugate transpose, SHHIncluding the echo power, S, of the HH polarization channelHVInvolving the echo power, S, of the HV polarisation channelVHIncluding VH polarization channel echo power, SVVIncluding the echo power number of the VV polarization channel.
5. The method of claim 1, wherein the criterion for convergence of clustering iteration in step 6 includes intra-class wishart dispersion and inter-class wishart distance; intra-class wishirt dispersion requirement: all samples of each type to the cluster center of the type
Figure FDA00035370949000000314
Is less than a given value
Figure FDA00035370949000000315
The wishart distance requirement between classes: the wishart distance between any two clustering centers is greater than a given value sigma, otherwise, the clustering is not converged, and iterative clustering needs to be continued; two of the cluster centers
Figure FDA00035370949000000316
And
Figure FDA00035370949000000317
the wishart distance between (1) is calculated as:
Figure FDA0003537094900000041
in the formula, | | represents a determinant of the matrix.
6. The method of claim 1, wherein the cluster center is calculated by:
Figure FDA0003537094900000042
Figure FDA0003537094900000043
in the formula (I), the compound is shown in the specification,
Figure FDA0003537094900000044
indicating the number of samples in the source domain dataset labeled c,
Figure FDA0003537094900000045
indicating the number of samples in the target domain dataset labeled c.
7. An unsupervised domain adaptive network polarization SAR terrain classification device considering polarization statistical characteristics is characterized by comprising a processor and a memory; wherein: the memory is to store computer instructions; the processor is configured to execute the computer instructions stored by the memory, in particular to perform the method according to any one of claims 1 to 6.
CN202110348030.1A 2021-03-31 2021-03-31 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics Active CN113269024B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110348030.1A CN113269024B (en) 2021-03-31 2021-03-31 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110348030.1A CN113269024B (en) 2021-03-31 2021-03-31 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics

Publications (2)

Publication Number Publication Date
CN113269024A CN113269024A (en) 2021-08-17
CN113269024B true CN113269024B (en) 2022-04-26

Family

ID=77227907

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110348030.1A Active CN113269024B (en) 2021-03-31 2021-03-31 Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics

Country Status (1)

Country Link
CN (1) CN113269024B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113902913A (en) * 2021-08-31 2022-01-07 际络科技(上海)有限公司 Image semantic segmentation method and device

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102360496A (en) * 2011-10-19 2012-02-22 西安电子科技大学 SAR (synthetic aperture radar) image segmentation method based on GMM (gaussian mixture model) parameter transferring and clustering
CN102999761A (en) * 2012-10-25 2013-03-27 西安电子科技大学 Method for classifying polarimetric SAR (synthetic aperture radar) images on the basis of Cloude decomposition and K-wishart distribution
CN108388927A (en) * 2018-03-26 2018-08-10 西安电子科技大学 Small sample polarization SAR terrain classification method based on the twin network of depth convolution
CN110728324A (en) * 2019-10-12 2020-01-24 西安电子科技大学 Depth complex value full convolution neural network-based polarimetric SAR image classification method
CN111709315A (en) * 2020-05-27 2020-09-25 西安交通大学 Underwater acoustic target radiation noise identification method based on field adaptation
CN111931814A (en) * 2020-07-03 2020-11-13 浙江大学 Unsupervised anti-domain adaptation method based on intra-class structure compactness constraint

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016097890A1 (en) * 2014-12-15 2016-06-23 Airbus Group Singapore Pte. Ltd. Automated method for selecting training areas of sea clutter and detecting ship targets in polarimetric synthetic aperture radar imagery

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102360496A (en) * 2011-10-19 2012-02-22 西安电子科技大学 SAR (synthetic aperture radar) image segmentation method based on GMM (gaussian mixture model) parameter transferring and clustering
CN102999761A (en) * 2012-10-25 2013-03-27 西安电子科技大学 Method for classifying polarimetric SAR (synthetic aperture radar) images on the basis of Cloude decomposition and K-wishart distribution
CN108388927A (en) * 2018-03-26 2018-08-10 西安电子科技大学 Small sample polarization SAR terrain classification method based on the twin network of depth convolution
CN110728324A (en) * 2019-10-12 2020-01-24 西安电子科技大学 Depth complex value full convolution neural network-based polarimetric SAR image classification method
CN111709315A (en) * 2020-05-27 2020-09-25 西安交通大学 Underwater acoustic target radiation noise identification method based on field adaptation
CN111931814A (en) * 2020-07-03 2020-11-13 浙江大学 Unsupervised anti-domain adaptation method based on intra-class structure compactness constraint

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
Complex-Valued Convolutional Neural Network and Its Application in Polarimetric SAR Image Classification;Zhimian Zhang et al;《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》;20170918;第55卷(第12期);第7177-7188页 *
PolSAR image classification via a novel semi-supervised recurrent complex-valued convolution neural network;Wen Xie et al;《Neurocomputing》;20200507;第388卷;第255-268页 *
基于半监督学习的SVM-Wishart极化SAR图像分类方法;滑文强等;《雷达学报》;20150228;第4卷(第1期);第93-98页 *
基于极化SAR数据的地物要素提取;肖可可等;《现代测绘》;20120531;第35卷(第3期);第7-10页 *
基于深度学习与Wishart距离相似性度量的极化SAR分类;李德博;《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》;20200215;第31-59页 *
小样本PolSAR图像地物分类技术研究;刘振;《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》;20190215;第35-77页 *
深度学习在SAR目标识别与地物分类中的应用;徐丰等;《雷达学报》;20170430;第6卷(第2期);第136-148页 *

Also Published As

Publication number Publication date
CN113269024A (en) 2021-08-17

Similar Documents

Publication Publication Date Title
Zhang et al. Discovering new intents with deep aligned clustering
CN113378632B (en) Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method
CN111126488B (en) Dual-attention-based image recognition method
Yang et al. CNN-based polarimetric decomposition feature selection for PolSAR image classification
CN111723674B (en) Remote sensing image scene classification method based on Markov chain Monte Carlo and variation deduction and semi-Bayesian deep learning
CN108805167B (en) Sparse depth confidence network image classification method based on Laplace function constraint
CN114488140B (en) Small sample radar one-dimensional image target recognition method based on deep migration learning
CN110599461B (en) Thalamus function partitioning method based on subspace feature learning
CN112464004A (en) Multi-view depth generation image clustering method
CN112434628B (en) Small sample image classification method based on active learning and collaborative representation
CN113326731A (en) Cross-domain pedestrian re-identification algorithm based on momentum network guidance
CN109359525B (en) Polarized SAR image classification method based on sparse low-rank discrimination spectral clustering
CN113889192B (en) Single-cell RNA-seq data clustering method based on deep noise reduction self-encoder
Franchi et al. Encoding the latent posterior of bayesian neural networks for uncertainty quantification
CN113269024B (en) Unsupervised domain self-adaptive network polarization SAR terrain classification method and device considering polarization statistical characteristics
CN113971440A (en) Unsupervised radar signal sorting method based on deep clustering
CN114819143A (en) Model compression method suitable for communication network field maintenance
CN112749633A (en) Separate and reconstructed individual radiation source identification method
CN109948589B (en) Facial expression recognition method based on quantum depth belief network
CN112115806A (en) Remote sensing image scene accurate classification method based on Dual-ResNet small sample learning
Sohail et al. Classification of ultrasound medical images using distance based feature selection and fuzzy-SVM
CN111680731B (en) Polarized SAR image supervision and classification method based on geometric perception discriminant dictionary learning
CN113627240A (en) Unmanned aerial vehicle tree species identification method based on improved SSD learning model
CN117034060A (en) AE-RCNN-based flood classification intelligent forecasting method
CN115601578A (en) Multi-view clustering method and system based on self-walking learning and view weighting

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